| 1 | //===- InstCombineMulDivRem.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 visit functions for mul, fmul, sdiv, udiv, fdiv, |
| 10 | // srem, urem, frem. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "InstCombineInternal.h" |
| 15 | #include "llvm/ADT/APInt.h" |
| 16 | #include "llvm/ADT/SmallPtrSet.h" |
| 17 | #include "llvm/ADT/SmallVector.h" |
| 18 | #include "llvm/Analysis/InstructionSimplify.h" |
| 19 | #include "llvm/Analysis/ValueTracking.h" |
| 20 | #include "llvm/IR/BasicBlock.h" |
| 21 | #include "llvm/IR/Constant.h" |
| 22 | #include "llvm/IR/Constants.h" |
| 23 | #include "llvm/IR/InstrTypes.h" |
| 24 | #include "llvm/IR/Instruction.h" |
| 25 | #include "llvm/IR/Instructions.h" |
| 26 | #include "llvm/IR/IntrinsicInst.h" |
| 27 | #include "llvm/IR/Intrinsics.h" |
| 28 | #include "llvm/IR/Operator.h" |
| 29 | #include "llvm/IR/PatternMatch.h" |
| 30 | #include "llvm/IR/Type.h" |
| 31 | #include "llvm/IR/Value.h" |
| 32 | #include "llvm/Support/Casting.h" |
| 33 | #include "llvm/Support/ErrorHandling.h" |
| 34 | #include "llvm/Transforms/InstCombine/InstCombiner.h" |
| 35 | #include "llvm/Transforms/Utils/BuildLibCalls.h" |
| 36 | #include <cassert> |
| 37 | |
| 38 | #define DEBUG_TYPE "instcombine" |
| 39 | #include "llvm/Transforms/Utils/InstructionWorklist.h" |
| 40 | |
| 41 | using namespace llvm; |
| 42 | using namespace PatternMatch; |
| 43 | |
| 44 | namespace llvm { |
| 45 | extern cl::opt<bool> ProfcheckDisableMetadataFixes; |
| 46 | } |
| 47 | |
| 48 | /// The specific integer value is used in a context where it is known to be |
| 49 | /// non-zero. If this allows us to simplify the computation, do so and return |
| 50 | /// the new operand, otherwise return null. |
| 51 | static Value *simplifyValueKnownNonZero(Value *V, InstCombinerImpl &IC, |
| 52 | Instruction &CxtI) { |
| 53 | // If V has multiple uses, then we would have to do more analysis to determine |
| 54 | // if this is safe. For example, the use could be in dynamically unreached |
| 55 | // code. |
| 56 | if (!V->hasOneUse()) return nullptr; |
| 57 | |
| 58 | bool MadeChange = false; |
| 59 | |
| 60 | // ((1 << A) >>u B) --> (1 << (A-B)) |
| 61 | // Because V cannot be zero, we know that B is less than A. |
| 62 | Value *A = nullptr, *B = nullptr, *One = nullptr; |
| 63 | if (match(V, P: m_LShr(L: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: One), R: m_Value(V&: A))), R: m_Value(V&: B))) && |
| 64 | match(V: One, P: m_One())) { |
| 65 | A = IC.Builder.CreateSub(LHS: A, RHS: B); |
| 66 | return IC.Builder.CreateShl(LHS: One, RHS: A); |
| 67 | } |
| 68 | |
| 69 | // (PowerOfTwo >>u B) --> isExact since shifting out the result would make it |
| 70 | // inexact. Similarly for <<. |
| 71 | BinaryOperator *I = dyn_cast<BinaryOperator>(Val: V); |
| 72 | if (I && I->isLogicalShift() && |
| 73 | IC.isKnownToBeAPowerOfTwo(V: I->getOperand(i_nocapture: 0), OrZero: false, CxtI: &CxtI)) { |
| 74 | // We know that this is an exact/nuw shift and that the input is a |
| 75 | // non-zero context as well. |
| 76 | { |
| 77 | IRBuilderBase::InsertPointGuard Guard(IC.Builder); |
| 78 | IC.Builder.SetInsertPoint(I); |
| 79 | if (Value *V2 = simplifyValueKnownNonZero(V: I->getOperand(i_nocapture: 0), IC, CxtI)) { |
| 80 | IC.replaceOperand(I&: *I, OpNum: 0, V: V2); |
| 81 | MadeChange = true; |
| 82 | } |
| 83 | } |
| 84 | |
| 85 | if (I->getOpcode() == Instruction::LShr && !I->isExact()) { |
| 86 | I->setIsExact(); |
| 87 | MadeChange = true; |
| 88 | } |
| 89 | |
| 90 | if (I->getOpcode() == Instruction::Shl && !I->hasNoUnsignedWrap()) { |
| 91 | I->setHasNoUnsignedWrap(); |
| 92 | MadeChange = true; |
| 93 | } |
| 94 | } |
| 95 | |
| 96 | // TODO: Lots more we could do here: |
| 97 | // If V is a phi node, we can call this on each of its operands. |
| 98 | // "select cond, X, 0" can simplify to "X". |
| 99 | |
| 100 | return MadeChange ? V : nullptr; |
| 101 | } |
| 102 | |
| 103 | // TODO: This is a specific form of a much more general pattern. |
| 104 | // We could detect a select with any binop identity constant, or we |
| 105 | // could use SimplifyBinOp to see if either arm of the select reduces. |
| 106 | // But that needs to be done carefully and/or while removing potential |
| 107 | // reverse canonicalizations as in InstCombiner::foldSelectIntoOp(). |
| 108 | static Value *foldMulSelectToNegate(BinaryOperator &I, |
| 109 | InstCombiner::BuilderTy &Builder) { |
| 110 | Value *Cond, *OtherOp; |
| 111 | |
| 112 | // mul (select Cond, 1, -1), OtherOp --> select Cond, OtherOp, -OtherOp |
| 113 | // mul OtherOp, (select Cond, 1, -1) --> select Cond, OtherOp, -OtherOp |
| 114 | if (match(V: &I, P: m_c_Mul(L: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_One(), R: m_AllOnes())), |
| 115 | R: m_Value(V&: OtherOp)))) { |
| 116 | bool HasAnyNoWrap = I.hasNoSignedWrap() || I.hasNoUnsignedWrap(); |
| 117 | Value *Neg = Builder.CreateNeg(V: OtherOp, Name: "" , HasNSW: HasAnyNoWrap); |
| 118 | return Builder.CreateSelect(C: Cond, True: OtherOp, False: Neg); |
| 119 | } |
| 120 | // mul (select Cond, -1, 1), OtherOp --> select Cond, -OtherOp, OtherOp |
| 121 | // mul OtherOp, (select Cond, -1, 1) --> select Cond, -OtherOp, OtherOp |
| 122 | if (match(V: &I, P: m_c_Mul(L: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_AllOnes(), R: m_One())), |
| 123 | R: m_Value(V&: OtherOp)))) { |
| 124 | bool HasAnyNoWrap = I.hasNoSignedWrap() || I.hasNoUnsignedWrap(); |
| 125 | Value *Neg = Builder.CreateNeg(V: OtherOp, Name: "" , HasNSW: HasAnyNoWrap); |
| 126 | return Builder.CreateSelect(C: Cond, True: Neg, False: OtherOp); |
| 127 | } |
| 128 | |
| 129 | // fmul (select Cond, 1.0, -1.0), OtherOp --> select Cond, OtherOp, -OtherOp |
| 130 | // fmul OtherOp, (select Cond, 1.0, -1.0) --> select Cond, OtherOp, -OtherOp |
| 131 | if (match(V: &I, P: m_c_FMul(L: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_SpecificFP(V: 1.0), |
| 132 | R: m_SpecificFP(V: -1.0))), |
| 133 | R: m_Value(V&: OtherOp)))) |
| 134 | return Builder.CreateSelectFMF(C: Cond, True: OtherOp, |
| 135 | False: Builder.CreateFNegFMF(V: OtherOp, FMFSource: &I), FMFSource: &I); |
| 136 | |
| 137 | // fmul (select Cond, -1.0, 1.0), OtherOp --> select Cond, -OtherOp, OtherOp |
| 138 | // fmul OtherOp, (select Cond, -1.0, 1.0) --> select Cond, -OtherOp, OtherOp |
| 139 | if (match(V: &I, P: m_c_FMul(L: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_SpecificFP(V: -1.0), |
| 140 | R: m_SpecificFP(V: 1.0))), |
| 141 | R: m_Value(V&: OtherOp)))) |
| 142 | return Builder.CreateSelectFMF(C: Cond, True: Builder.CreateFNegFMF(V: OtherOp, FMFSource: &I), |
| 143 | False: OtherOp, FMFSource: &I); |
| 144 | |
| 145 | return nullptr; |
| 146 | } |
| 147 | |
| 148 | /// Reduce integer multiplication patterns that contain a (+/-1 << Z) factor. |
| 149 | /// Callers are expected to call this twice to handle commuted patterns. |
| 150 | static Value *foldMulShl1(BinaryOperator &Mul, bool CommuteOperands, |
| 151 | InstCombiner::BuilderTy &Builder) { |
| 152 | Value *X = Mul.getOperand(i_nocapture: 0), *Y = Mul.getOperand(i_nocapture: 1); |
| 153 | if (CommuteOperands) |
| 154 | std::swap(a&: X, b&: Y); |
| 155 | |
| 156 | const bool HasNSW = Mul.hasNoSignedWrap(); |
| 157 | const bool HasNUW = Mul.hasNoUnsignedWrap(); |
| 158 | |
| 159 | // X * (1 << Z) --> X << Z |
| 160 | Value *Z; |
| 161 | if (match(V: Y, P: m_Shl(L: m_One(), R: m_Value(V&: Z)))) { |
| 162 | bool PropagateNSW = HasNSW && cast<ShlOperator>(Val: Y)->hasNoSignedWrap(); |
| 163 | return Builder.CreateShl(LHS: X, RHS: Z, Name: Mul.getName(), HasNUW, HasNSW: PropagateNSW); |
| 164 | } |
| 165 | |
| 166 | // Similar to above, but an increment of the shifted value becomes an add: |
| 167 | // X * ((1 << Z) + 1) --> (X * (1 << Z)) + X --> (X << Z) + X |
| 168 | // This increases uses of X, so it may require a freeze, but that is still |
| 169 | // expected to be an improvement because it removes the multiply. |
| 170 | BinaryOperator *Shift; |
| 171 | if (match(V: Y, P: m_OneUse(SubPattern: m_Add(L: m_BinOp(I&: Shift), R: m_One()))) && |
| 172 | match(V: Shift, P: m_OneUse(SubPattern: m_Shl(L: m_One(), R: m_Value(V&: Z))))) { |
| 173 | bool PropagateNSW = HasNSW && Shift->hasNoSignedWrap(); |
| 174 | Value *FrX = X; |
| 175 | if (!isGuaranteedNotToBeUndef(V: X)) |
| 176 | FrX = Builder.CreateFreeze(V: X, Name: X->getName() + ".fr" ); |
| 177 | Value *Shl = Builder.CreateShl(LHS: FrX, RHS: Z, Name: "mulshl" , HasNUW, HasNSW: PropagateNSW); |
| 178 | return Builder.CreateAdd(LHS: Shl, RHS: FrX, Name: Mul.getName(), HasNUW, HasNSW: PropagateNSW); |
| 179 | } |
| 180 | |
| 181 | // Similar to above, but a decrement of the shifted value is disguised as |
| 182 | // 'not' and becomes a sub: |
| 183 | // X * (~(-1 << Z)) --> X * ((1 << Z) - 1) --> (X << Z) - X |
| 184 | // This increases uses of X, so it may require a freeze, but that is still |
| 185 | // expected to be an improvement because it removes the multiply. |
| 186 | if (match(V: Y, P: m_OneUse(SubPattern: m_Not(V: m_OneUse(SubPattern: m_Shl(L: m_AllOnes(), R: m_Value(V&: Z))))))) { |
| 187 | Value *FrX = X; |
| 188 | if (!isGuaranteedNotToBeUndef(V: X)) |
| 189 | FrX = Builder.CreateFreeze(V: X, Name: X->getName() + ".fr" ); |
| 190 | Value *Shl = Builder.CreateShl(LHS: FrX, RHS: Z, Name: "mulshl" ); |
| 191 | return Builder.CreateSub(LHS: Shl, RHS: FrX, Name: Mul.getName()); |
| 192 | } |
| 193 | |
| 194 | return nullptr; |
| 195 | } |
| 196 | |
| 197 | Instruction *InstCombinerImpl::visitMul(BinaryOperator &I) { |
| 198 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 199 | if (Value *V = |
| 200 | simplifyMulInst(LHS: Op0, RHS: Op1, IsNSW: I.hasNoSignedWrap(), IsNUW: I.hasNoUnsignedWrap(), |
| 201 | Q: SQ.getWithInstruction(I: &I))) |
| 202 | return replaceInstUsesWith(I, V); |
| 203 | |
| 204 | if (SimplifyAssociativeOrCommutative(I)) |
| 205 | return &I; |
| 206 | |
| 207 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
| 208 | return X; |
| 209 | |
| 210 | if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I)) |
| 211 | return Phi; |
| 212 | |
| 213 | if (Value *V = foldUsingDistributiveLaws(I)) |
| 214 | return replaceInstUsesWith(I, V); |
| 215 | |
| 216 | Type *Ty = I.getType(); |
| 217 | const unsigned BitWidth = Ty->getScalarSizeInBits(); |
| 218 | const bool HasNSW = I.hasNoSignedWrap(); |
| 219 | const bool HasNUW = I.hasNoUnsignedWrap(); |
| 220 | |
| 221 | // X * -1 --> 0 - X |
| 222 | if (match(V: Op1, P: m_AllOnes())) { |
| 223 | return HasNSW ? BinaryOperator::CreateNSWNeg(Op: Op0) |
| 224 | : BinaryOperator::CreateNeg(Op: Op0); |
| 225 | } |
| 226 | |
| 227 | // Also allow combining multiply instructions on vectors. |
| 228 | { |
| 229 | Value *NewOp; |
| 230 | Constant *C1, *C2; |
| 231 | const APInt *IVal; |
| 232 | if (match(V: &I, P: m_Mul(L: m_Shl(L: m_Value(V&: NewOp), R: m_ImmConstant(C&: C2)), |
| 233 | R: m_ImmConstant(C&: C1))) && |
| 234 | match(V: C1, P: m_APInt(Res&: IVal))) { |
| 235 | // ((X << C2)*C1) == (X * (C1 << C2)) |
| 236 | Constant *Shl = |
| 237 | ConstantFoldBinaryOpOperands(Opcode: Instruction::Shl, LHS: C1, RHS: C2, DL); |
| 238 | assert(Shl && "Constant folding of immediate constants failed" ); |
| 239 | BinaryOperator *Mul = cast<BinaryOperator>(Val: I.getOperand(i_nocapture: 0)); |
| 240 | BinaryOperator *BO = BinaryOperator::CreateMul(V1: NewOp, V2: Shl); |
| 241 | if (HasNUW && Mul->hasNoUnsignedWrap()) |
| 242 | BO->setHasNoUnsignedWrap(); |
| 243 | if (HasNSW && Mul->hasNoSignedWrap() && Shl->isNotMinSignedValue()) |
| 244 | BO->setHasNoSignedWrap(); |
| 245 | return BO; |
| 246 | } |
| 247 | |
| 248 | if (match(V: &I, P: m_Mul(L: m_Value(V&: NewOp), R: m_Constant(C&: C1)))) { |
| 249 | // Replace X*(2^C) with X << C, where C is either a scalar or a vector. |
| 250 | if (Constant *NewCst = ConstantExpr::getExactLogBase2(C: C1)) { |
| 251 | BinaryOperator *Shl = BinaryOperator::CreateShl(V1: NewOp, V2: NewCst); |
| 252 | |
| 253 | if (HasNUW) |
| 254 | Shl->setHasNoUnsignedWrap(); |
| 255 | if (HasNSW) { |
| 256 | const APInt *V; |
| 257 | if (match(V: NewCst, P: m_APInt(Res&: V)) && *V != V->getBitWidth() - 1) |
| 258 | Shl->setHasNoSignedWrap(); |
| 259 | } |
| 260 | |
| 261 | return Shl; |
| 262 | } |
| 263 | } |
| 264 | } |
| 265 | |
| 266 | // mul (shr exact X, N), (2^N + 1) -> add (X, shr exact (X, N)) |
| 267 | { |
| 268 | Value *NewOp; |
| 269 | const APInt *ShiftC; |
| 270 | const APInt *MulAP; |
| 271 | if (BitWidth > 2 && |
| 272 | match(V: &I, P: m_Mul(L: m_Exact(SubPattern: m_Shr(L: m_Value(V&: NewOp), R: m_APInt(Res&: ShiftC))), |
| 273 | R: m_APInt(Res&: MulAP))) && |
| 274 | (*MulAP - 1).isPowerOf2() && *ShiftC == MulAP->logBase2()) { |
| 275 | Value *BinOp = Op0; |
| 276 | BinaryOperator *OpBO = cast<BinaryOperator>(Val: Op0); |
| 277 | |
| 278 | // mul nuw (ashr exact X, N) -> add nuw (X, lshr exact (X, N)) |
| 279 | if (HasNUW && OpBO->getOpcode() == Instruction::AShr && OpBO->hasOneUse()) |
| 280 | BinOp = Builder.CreateLShr(LHS: NewOp, RHS: ConstantInt::get(Ty, V: *ShiftC), Name: "" , |
| 281 | /*isExact=*/true); |
| 282 | |
| 283 | auto *NewAdd = BinaryOperator::CreateAdd(V1: NewOp, V2: BinOp); |
| 284 | if (HasNSW && (HasNUW || OpBO->getOpcode() == Instruction::LShr || |
| 285 | ShiftC->getZExtValue() < BitWidth - 1)) |
| 286 | NewAdd->setHasNoSignedWrap(true); |
| 287 | |
| 288 | NewAdd->setHasNoUnsignedWrap(HasNUW); |
| 289 | return NewAdd; |
| 290 | } |
| 291 | } |
| 292 | |
| 293 | if (Op0->hasOneUse() && match(V: Op1, P: m_NegatedPower2())) { |
| 294 | // Interpret X * (-1<<C) as (-X) * (1<<C) and try to sink the negation. |
| 295 | // The "* (1<<C)" thus becomes a potential shifting opportunity. |
| 296 | if (Value *NegOp0 = |
| 297 | Negator::Negate(/*IsNegation*/ LHSIsZero: true, IsNSW: HasNSW, Root: Op0, IC&: *this)) { |
| 298 | auto *Op1C = cast<Constant>(Val: Op1); |
| 299 | return replaceInstUsesWith( |
| 300 | I, V: Builder.CreateMul(LHS: NegOp0, RHS: ConstantExpr::getNeg(C: Op1C), Name: "" , |
| 301 | /*HasNUW=*/false, |
| 302 | HasNSW: HasNSW && Op1C->isNotMinSignedValue())); |
| 303 | } |
| 304 | |
| 305 | // Try to convert multiply of extended operand to narrow negate and shift |
| 306 | // for better analysis. |
| 307 | // This is valid if the shift amount (trailing zeros in the multiplier |
| 308 | // constant) clears more high bits than the bitwidth difference between |
| 309 | // source and destination types: |
| 310 | // ({z/s}ext X) * (-1<<C) --> (zext (-X)) << C |
| 311 | const APInt *NegPow2C; |
| 312 | Value *X; |
| 313 | if (match(V: Op0, P: m_ZExtOrSExt(Op: m_Value(V&: X))) && |
| 314 | match(V: Op1, P: m_APIntAllowPoison(Res&: NegPow2C))) { |
| 315 | unsigned SrcWidth = X->getType()->getScalarSizeInBits(); |
| 316 | unsigned ShiftAmt = NegPow2C->countr_zero(); |
| 317 | if (ShiftAmt >= BitWidth - SrcWidth) { |
| 318 | Value *N = Builder.CreateNeg(V: X, Name: X->getName() + ".neg" ); |
| 319 | Value *Z = Builder.CreateZExt(V: N, DestTy: Ty, Name: N->getName() + ".z" ); |
| 320 | return BinaryOperator::CreateShl(V1: Z, V2: ConstantInt::get(Ty, V: ShiftAmt)); |
| 321 | } |
| 322 | } |
| 323 | } |
| 324 | |
| 325 | if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I)) |
| 326 | return FoldedMul; |
| 327 | |
| 328 | if (Instruction *FoldedLogic = foldBinOpSelectBinOp(Op&: I)) |
| 329 | return FoldedLogic; |
| 330 | |
| 331 | if (Value *FoldedMul = foldMulSelectToNegate(I, Builder)) |
| 332 | return replaceInstUsesWith(I, V: FoldedMul); |
| 333 | |
| 334 | // (shl X, C1)*(select cond, C2, C3)--> X * (select cond, C2<<C1, C3<<C1) |
| 335 | // (mul X, C1)*(select cond, C2, C3)--> X * (select cond, C2*C1, C3*C1) |
| 336 | // (Includes commuted forms) |
| 337 | |
| 338 | { |
| 339 | Value *NewOp, *Cond, *OtherValue; |
| 340 | Constant *C1, *C2, *C3; |
| 341 | |
| 342 | if (match(V: &I, P: m_c_Mul(L: m_OneUse(SubPattern: m_Value(V&: OtherValue)), |
| 343 | R: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_ImmConstant(C&: C2), |
| 344 | R: m_ImmConstant(C&: C3))))) && |
| 345 | (match(V: OtherValue, P: m_Mul(L: m_Value(V&: NewOp), R: m_ImmConstant(C&: C1))) || |
| 346 | match(V: OtherValue, P: m_Shl(L: m_Value(V&: NewOp), R: m_ImmConstant(C&: C1))))) { |
| 347 | |
| 348 | auto *OtherInst = cast<OverflowingBinaryOperator>(Val: OtherValue); |
| 349 | auto Opc = OtherInst->getOpcode(); |
| 350 | |
| 351 | Constant *NewTV = ConstantFoldBinaryOpOperands(Opcode: Opc, LHS: C2, RHS: C1, DL); |
| 352 | Constant *NewFV = ConstantFoldBinaryOpOperands(Opcode: Opc, LHS: C3, RHS: C1, DL); |
| 353 | |
| 354 | if (NewTV && NewFV) { |
| 355 | Value *NewSel = Builder.CreateSelect(C: Cond, True: NewTV, False: NewFV); |
| 356 | BinaryOperator *BO = BinaryOperator::CreateMul(V1: NewOp, V2: NewSel); |
| 357 | |
| 358 | if (HasNUW && OtherInst->hasNoUnsignedWrap()) |
| 359 | BO->setHasNoUnsignedWrap(); |
| 360 | if (HasNSW && OtherInst->hasNoSignedWrap() && |
| 361 | NewTV->isNotMinSignedValue() && NewFV->isNotMinSignedValue()) |
| 362 | BO->setHasNoSignedWrap(); |
| 363 | |
| 364 | return BO; |
| 365 | } |
| 366 | } |
| 367 | } |
| 368 | |
| 369 | // Simplify mul instructions with a constant RHS. |
| 370 | Constant *MulC; |
| 371 | if (match(V: Op1, P: m_ImmConstant(C&: MulC))) { |
| 372 | // Canonicalize (X+C1)*MulC -> X*MulC+C1*MulC. |
| 373 | // Canonicalize (X|C1)*MulC -> X*MulC+C1*MulC. |
| 374 | Value *X; |
| 375 | Constant *C1; |
| 376 | if (match(V: Op0, P: m_OneUse(SubPattern: m_AddLike(L: m_Value(V&: X), R: m_ImmConstant(C&: C1))))) { |
| 377 | // C1*MulC simplifies to a tidier constant. |
| 378 | Value *NewC = Builder.CreateMul(LHS: C1, RHS: MulC); |
| 379 | auto *BOp0 = cast<BinaryOperator>(Val: Op0); |
| 380 | bool Op0NUW = |
| 381 | (BOp0->getOpcode() == Instruction::Or || BOp0->hasNoUnsignedWrap()); |
| 382 | Value *NewMul = Builder.CreateMul(LHS: X, RHS: MulC); |
| 383 | auto *BO = BinaryOperator::CreateAdd(V1: NewMul, V2: NewC); |
| 384 | if (HasNUW && Op0NUW) { |
| 385 | // If NewMulBO is constant we also can set BO to nuw. |
| 386 | if (auto *NewMulBO = dyn_cast<BinaryOperator>(Val: NewMul)) |
| 387 | NewMulBO->setHasNoUnsignedWrap(); |
| 388 | BO->setHasNoUnsignedWrap(); |
| 389 | } |
| 390 | return BO; |
| 391 | } |
| 392 | } |
| 393 | |
| 394 | // abs(X) * abs(X) -> X * X |
| 395 | Value *X; |
| 396 | if (Op0 == Op1 && match(V: Op0, P: m_Intrinsic<Intrinsic::abs>(Op0: m_Value(V&: X)))) |
| 397 | return BinaryOperator::CreateMul(V1: X, V2: X); |
| 398 | |
| 399 | { |
| 400 | Value *Y; |
| 401 | // abs(X) * abs(Y) -> abs(X * Y) |
| 402 | if (I.hasNoSignedWrap() && |
| 403 | match(V: Op0, |
| 404 | P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::abs>(Op0: m_Value(V&: X), Op1: m_One()))) && |
| 405 | match(V: Op1, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::abs>(Op0: m_Value(V&: Y), Op1: m_One())))) |
| 406 | return replaceInstUsesWith( |
| 407 | I, V: Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, |
| 408 | LHS: Builder.CreateNSWMul(LHS: X, RHS: Y), |
| 409 | RHS: Builder.getTrue())); |
| 410 | } |
| 411 | |
| 412 | // -X * C --> X * -C |
| 413 | Value *Y; |
| 414 | Constant *Op1C; |
| 415 | if (match(V: Op0, P: m_Neg(V: m_Value(V&: X))) && match(V: Op1, P: m_Constant(C&: Op1C))) |
| 416 | return BinaryOperator::CreateMul(V1: X, V2: ConstantExpr::getNeg(C: Op1C)); |
| 417 | |
| 418 | // -X * -Y --> X * Y |
| 419 | if (match(V: Op0, P: m_Neg(V: m_Value(V&: X))) && match(V: Op1, P: m_Neg(V: m_Value(V&: Y)))) { |
| 420 | auto *NewMul = BinaryOperator::CreateMul(V1: X, V2: Y); |
| 421 | if (HasNSW && cast<OverflowingBinaryOperator>(Val: Op0)->hasNoSignedWrap() && |
| 422 | cast<OverflowingBinaryOperator>(Val: Op1)->hasNoSignedWrap()) |
| 423 | NewMul->setHasNoSignedWrap(); |
| 424 | return NewMul; |
| 425 | } |
| 426 | |
| 427 | // -X * Y --> -(X * Y) |
| 428 | // X * -Y --> -(X * Y) |
| 429 | if (match(V: &I, P: m_c_Mul(L: m_OneUse(SubPattern: m_Neg(V: m_Value(V&: X))), R: m_Value(V&: Y)))) |
| 430 | return BinaryOperator::CreateNeg(Op: Builder.CreateMul(LHS: X, RHS: Y)); |
| 431 | |
| 432 | // (-X * Y) * -X --> (X * Y) * X |
| 433 | // (-X << Y) * -X --> (X << Y) * X |
| 434 | if (match(V: Op1, P: m_Neg(V: m_Value(V&: X)))) { |
| 435 | if (Value *NegOp0 = Negator::Negate(LHSIsZero: false, /*IsNSW*/ false, Root: Op0, IC&: *this)) |
| 436 | return BinaryOperator::CreateMul(V1: NegOp0, V2: X); |
| 437 | } |
| 438 | |
| 439 | if (Op0->hasOneUse()) { |
| 440 | // (mul (div exact X, C0), C1) |
| 441 | // -> (div exact X, C0 / C1) |
| 442 | // iff C0 % C1 == 0 and X / (C0 / C1) doesn't create UB. |
| 443 | const APInt *C1; |
| 444 | auto UDivCheck = [&C1](const APInt &C) { return C.urem(RHS: *C1).isZero(); }; |
| 445 | auto SDivCheck = [&C1](const APInt &C) { |
| 446 | APInt Quot, Rem; |
| 447 | APInt::sdivrem(LHS: C, RHS: *C1, Quotient&: Quot, Remainder&: Rem); |
| 448 | return Rem.isZero() && !Quot.isAllOnes(); |
| 449 | }; |
| 450 | if (match(V: Op1, P: m_APInt(Res&: C1)) && |
| 451 | (match(V: Op0, P: m_Exact(SubPattern: m_UDiv(L: m_Value(V&: X), R: m_CheckedInt(CheckFn: UDivCheck)))) || |
| 452 | match(V: Op0, P: m_Exact(SubPattern: m_SDiv(L: m_Value(V&: X), R: m_CheckedInt(CheckFn: SDivCheck)))))) { |
| 453 | auto BOpc = cast<BinaryOperator>(Val: Op0)->getOpcode(); |
| 454 | return BinaryOperator::CreateExact( |
| 455 | Opc: BOpc, V1: X, |
| 456 | V2: Builder.CreateBinOp(Opc: BOpc, LHS: cast<BinaryOperator>(Val: Op0)->getOperand(i_nocapture: 1), |
| 457 | RHS: Op1)); |
| 458 | } |
| 459 | } |
| 460 | |
| 461 | // (X / Y) * Y = X - (X % Y) |
| 462 | // (X / Y) * -Y = (X % Y) - X |
| 463 | { |
| 464 | Value *Y = Op1; |
| 465 | BinaryOperator *Div = dyn_cast<BinaryOperator>(Val: Op0); |
| 466 | if (!Div || (Div->getOpcode() != Instruction::UDiv && |
| 467 | Div->getOpcode() != Instruction::SDiv)) { |
| 468 | Y = Op0; |
| 469 | Div = dyn_cast<BinaryOperator>(Val: Op1); |
| 470 | } |
| 471 | Value *Neg = dyn_castNegVal(V: Y); |
| 472 | if (Div && Div->hasOneUse() && |
| 473 | (Div->getOperand(i_nocapture: 1) == Y || Div->getOperand(i_nocapture: 1) == Neg) && |
| 474 | (Div->getOpcode() == Instruction::UDiv || |
| 475 | Div->getOpcode() == Instruction::SDiv)) { |
| 476 | Value *X = Div->getOperand(i_nocapture: 0), *DivOp1 = Div->getOperand(i_nocapture: 1); |
| 477 | |
| 478 | // If the division is exact, X % Y is zero, so we end up with X or -X. |
| 479 | if (Div->isExact()) { |
| 480 | if (DivOp1 == Y) |
| 481 | return replaceInstUsesWith(I, V: X); |
| 482 | return BinaryOperator::CreateNeg(Op: X); |
| 483 | } |
| 484 | |
| 485 | auto RemOpc = Div->getOpcode() == Instruction::UDiv ? Instruction::URem |
| 486 | : Instruction::SRem; |
| 487 | // X must be frozen because we are increasing its number of uses. |
| 488 | Value *XFreeze = X; |
| 489 | if (!isGuaranteedNotToBeUndef(V: X)) |
| 490 | XFreeze = Builder.CreateFreeze(V: X, Name: X->getName() + ".fr" ); |
| 491 | Value *Rem = Builder.CreateBinOp(Opc: RemOpc, LHS: XFreeze, RHS: DivOp1); |
| 492 | if (DivOp1 == Y) |
| 493 | return BinaryOperator::CreateSub(V1: XFreeze, V2: Rem); |
| 494 | return BinaryOperator::CreateSub(V1: Rem, V2: XFreeze); |
| 495 | } |
| 496 | } |
| 497 | |
| 498 | // Fold the following two scenarios: |
| 499 | // 1) i1 mul -> i1 and. |
| 500 | // 2) X * Y --> X & Y, iff X, Y can be only {0,1}. |
| 501 | // Note: We could use known bits to generalize this and related patterns with |
| 502 | // shifts/truncs |
| 503 | if (Ty->isIntOrIntVectorTy(BitWidth: 1) || |
| 504 | (match(V: Op0, P: m_And(L: m_Value(), R: m_One())) && |
| 505 | match(V: Op1, P: m_And(L: m_Value(), R: m_One())))) |
| 506 | return BinaryOperator::CreateAnd(V1: Op0, V2: Op1); |
| 507 | |
| 508 | if (Value *R = foldMulShl1(Mul&: I, /* CommuteOperands */ false, Builder)) |
| 509 | return replaceInstUsesWith(I, V: R); |
| 510 | if (Value *R = foldMulShl1(Mul&: I, /* CommuteOperands */ true, Builder)) |
| 511 | return replaceInstUsesWith(I, V: R); |
| 512 | |
| 513 | // (zext bool X) * (zext bool Y) --> zext (and X, Y) |
| 514 | // (sext bool X) * (sext bool Y) --> zext (and X, Y) |
| 515 | // Note: -1 * -1 == 1 * 1 == 1 (if the extends match, the result is the same) |
| 516 | if (((match(V: Op0, P: m_ZExt(Op: m_Value(V&: X))) && match(V: Op1, P: m_ZExt(Op: m_Value(V&: Y)))) || |
| 517 | (match(V: Op0, P: m_SExt(Op: m_Value(V&: X))) && match(V: Op1, P: m_SExt(Op: m_Value(V&: Y))))) && |
| 518 | X->getType()->isIntOrIntVectorTy(BitWidth: 1) && X->getType() == Y->getType() && |
| 519 | (Op0->hasOneUse() || Op1->hasOneUse() || X == Y)) { |
| 520 | Value *And = Builder.CreateAnd(LHS: X, RHS: Y, Name: "mulbool" ); |
| 521 | return CastInst::Create(Instruction::ZExt, S: And, Ty); |
| 522 | } |
| 523 | // (sext bool X) * (zext bool Y) --> sext (and X, Y) |
| 524 | // (zext bool X) * (sext bool Y) --> sext (and X, Y) |
| 525 | // Note: -1 * 1 == 1 * -1 == -1 |
| 526 | if (((match(V: Op0, P: m_SExt(Op: m_Value(V&: X))) && match(V: Op1, P: m_ZExt(Op: m_Value(V&: Y)))) || |
| 527 | (match(V: Op0, P: m_ZExt(Op: m_Value(V&: X))) && match(V: Op1, P: m_SExt(Op: m_Value(V&: Y))))) && |
| 528 | X->getType()->isIntOrIntVectorTy(BitWidth: 1) && X->getType() == Y->getType() && |
| 529 | (Op0->hasOneUse() || Op1->hasOneUse())) { |
| 530 | Value *And = Builder.CreateAnd(LHS: X, RHS: Y, Name: "mulbool" ); |
| 531 | return CastInst::Create(Instruction::SExt, S: And, Ty); |
| 532 | } |
| 533 | |
| 534 | // (zext bool X) * Y --> X ? Y : 0 |
| 535 | // Y * (zext bool X) --> X ? Y : 0 |
| 536 | if (match(V: Op0, P: m_ZExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1)) |
| 537 | return createSelectInstWithUnknownProfile(C: X, S1: Op1, |
| 538 | S2: ConstantInt::getNullValue(Ty)); |
| 539 | if (match(V: Op1, P: m_ZExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1)) |
| 540 | return createSelectInstWithUnknownProfile(C: X, S1: Op0, |
| 541 | S2: ConstantInt::getNullValue(Ty)); |
| 542 | |
| 543 | // mul (sext X), Y -> select X, -Y, 0 |
| 544 | // mul Y, (sext X) -> select X, -Y, 0 |
| 545 | if (match(V: &I, P: m_c_Mul(L: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: X))), R: m_Value(V&: Y))) && |
| 546 | X->getType()->isIntOrIntVectorTy(BitWidth: 1)) |
| 547 | return createSelectInstWithUnknownProfile( |
| 548 | C: X, S1: Builder.CreateNeg(V: Y, Name: "" , HasNSW: I.hasNoSignedWrap()), |
| 549 | S2: ConstantInt::getNullValue(Ty: Op0->getType())); |
| 550 | |
| 551 | Constant *ImmC; |
| 552 | if (match(V: Op1, P: m_ImmConstant(C&: ImmC))) { |
| 553 | // (sext bool X) * C --> X ? -C : 0 |
| 554 | if (match(V: Op0, P: m_SExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
| 555 | Constant *NegC = ConstantExpr::getNeg(C: ImmC); |
| 556 | return createSelectInstWithUnknownProfile(C: X, S1: NegC, |
| 557 | S2: ConstantInt::getNullValue(Ty)); |
| 558 | } |
| 559 | |
| 560 | // (ashr i32 X, 31) * C --> (X < 0) ? -C : 0 |
| 561 | const APInt *C; |
| 562 | if (match(V: Op0, P: m_OneUse(SubPattern: m_AShr(L: m_Value(V&: X), R: m_APInt(Res&: C)))) && |
| 563 | *C == C->getBitWidth() - 1) { |
| 564 | Constant *NegC = ConstantExpr::getNeg(C: ImmC); |
| 565 | Value *IsNeg = Builder.CreateIsNeg(Arg: X, Name: "isneg" ); |
| 566 | return createSelectInstWithUnknownProfile(C: IsNeg, S1: NegC, |
| 567 | S2: ConstantInt::getNullValue(Ty)); |
| 568 | } |
| 569 | } |
| 570 | |
| 571 | // (lshr X, 31) * Y --> (X < 0) ? Y : 0 |
| 572 | // TODO: We are not checking one-use because the elimination of the multiply |
| 573 | // is better for analysis? |
| 574 | const APInt *C; |
| 575 | if (match(V: &I, P: m_c_BinOp(L: m_LShr(L: m_Value(V&: X), R: m_APInt(Res&: C)), R: m_Value(V&: Y))) && |
| 576 | *C == C->getBitWidth() - 1) { |
| 577 | Value *IsNeg = Builder.CreateIsNeg(Arg: X, Name: "isneg" ); |
| 578 | return createSelectInstWithUnknownProfile(C: IsNeg, S1: Y, |
| 579 | S2: ConstantInt::getNullValue(Ty)); |
| 580 | } |
| 581 | |
| 582 | // (and X, 1) * Y --> (trunc X) ? Y : 0 |
| 583 | if (match(V: &I, P: m_c_BinOp(L: m_OneUse(SubPattern: m_And(L: m_Value(V&: X), R: m_One())), R: m_Value(V&: Y)))) { |
| 584 | Value *Tr = Builder.CreateTrunc(V: X, DestTy: CmpInst::makeCmpResultType(opnd_type: Ty)); |
| 585 | return createSelectInstWithUnknownProfile(C: Tr, S1: Y, |
| 586 | S2: ConstantInt::getNullValue(Ty)); |
| 587 | } |
| 588 | |
| 589 | // ((ashr X, 31) | 1) * X --> abs(X) |
| 590 | // X * ((ashr X, 31) | 1) --> abs(X) |
| 591 | if (match(V: &I, P: m_c_BinOp(L: m_Or(L: m_AShr(L: m_Value(V&: X), |
| 592 | R: m_SpecificIntAllowPoison(V: BitWidth - 1)), |
| 593 | R: m_One()), |
| 594 | R: m_Deferred(V: X)))) { |
| 595 | Value *Abs = Builder.CreateBinaryIntrinsic( |
| 596 | ID: Intrinsic::abs, LHS: X, RHS: ConstantInt::getBool(Context&: I.getContext(), V: HasNSW)); |
| 597 | Abs->takeName(V: &I); |
| 598 | return replaceInstUsesWith(I, V: Abs); |
| 599 | } |
| 600 | |
| 601 | if (Instruction *Ext = narrowMathIfNoOverflow(I)) |
| 602 | return Ext; |
| 603 | |
| 604 | if (Instruction *Res = foldBinOpOfSelectAndCastOfSelectCondition(I)) |
| 605 | return Res; |
| 606 | |
| 607 | // (mul Op0 Op1): |
| 608 | // if Log2(Op0) folds away -> |
| 609 | // (shl Op1, Log2(Op0)) |
| 610 | // if Log2(Op1) folds away -> |
| 611 | // (shl Op0, Log2(Op1)) |
| 612 | if (Value *Res = tryGetLog2(Op: Op0, /*AssumeNonZero=*/false)) { |
| 613 | BinaryOperator *Shl = BinaryOperator::CreateShl(V1: Op1, V2: Res); |
| 614 | // We can only propegate nuw flag. |
| 615 | Shl->setHasNoUnsignedWrap(HasNUW); |
| 616 | return Shl; |
| 617 | } |
| 618 | if (Value *Res = tryGetLog2(Op: Op1, /*AssumeNonZero=*/false)) { |
| 619 | BinaryOperator *Shl = BinaryOperator::CreateShl(V1: Op0, V2: Res); |
| 620 | // We can only propegate nuw flag. |
| 621 | Shl->setHasNoUnsignedWrap(HasNUW); |
| 622 | return Shl; |
| 623 | } |
| 624 | |
| 625 | bool Changed = false; |
| 626 | if (!HasNSW && willNotOverflowSignedMul(LHS: Op0, RHS: Op1, CxtI: I)) { |
| 627 | Changed = true; |
| 628 | I.setHasNoSignedWrap(true); |
| 629 | } |
| 630 | |
| 631 | if (!HasNUW && willNotOverflowUnsignedMul(LHS: Op0, RHS: Op1, CxtI: I, IsNSW: I.hasNoSignedWrap())) { |
| 632 | Changed = true; |
| 633 | I.setHasNoUnsignedWrap(true); |
| 634 | } |
| 635 | |
| 636 | return Changed ? &I : nullptr; |
| 637 | } |
| 638 | |
| 639 | Instruction *InstCombinerImpl::foldFPSignBitOps(BinaryOperator &I) { |
| 640 | BinaryOperator::BinaryOps Opcode = I.getOpcode(); |
| 641 | assert((Opcode == Instruction::FMul || Opcode == Instruction::FDiv) && |
| 642 | "Expected fmul or fdiv" ); |
| 643 | |
| 644 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 645 | Value *X, *Y; |
| 646 | |
| 647 | // -X * -Y --> X * Y |
| 648 | // -X / -Y --> X / Y |
| 649 | if (match(V: Op0, P: m_FNeg(X: m_Value(V&: X))) && match(V: Op1, P: m_FNeg(X: m_Value(V&: Y)))) |
| 650 | return BinaryOperator::CreateWithCopiedFlags(Opc: Opcode, V1: X, V2: Y, CopyO: &I); |
| 651 | |
| 652 | // fabs(X) * fabs(X) -> X * X |
| 653 | // fabs(X) / fabs(X) -> X / X |
| 654 | if (Op0 == Op1 && match(V: Op0, P: m_FAbs(Op0: m_Value(V&: X)))) |
| 655 | return BinaryOperator::CreateWithCopiedFlags(Opc: Opcode, V1: X, V2: X, CopyO: &I); |
| 656 | |
| 657 | // fabs(X) * fabs(Y) --> fabs(X * Y) |
| 658 | // fabs(X) / fabs(Y) --> fabs(X / Y) |
| 659 | if (match(V: Op0, P: m_FAbs(Op0: m_Value(V&: X))) && match(V: Op1, P: m_FAbs(Op0: m_Value(V&: Y))) && |
| 660 | (Op0->hasOneUse() || Op1->hasOneUse())) { |
| 661 | Value *XY = Builder.CreateBinOpFMF(Opc: Opcode, LHS: X, RHS: Y, FMFSource: &I); |
| 662 | Value *Fabs = Builder.CreateFAbs(V: XY, FMFSource: &I, Name: I.getName()); |
| 663 | return replaceInstUsesWith(I, V: Fabs); |
| 664 | } |
| 665 | |
| 666 | return nullptr; |
| 667 | } |
| 668 | |
| 669 | Instruction *InstCombinerImpl::foldPowiReassoc(BinaryOperator &I) { |
| 670 | auto createPowiExpr = [](BinaryOperator &I, InstCombinerImpl &IC, Value *X, |
| 671 | Value *Y, Value *Z) { |
| 672 | InstCombiner::BuilderTy &Builder = IC.Builder; |
| 673 | Value *YZ = Builder.CreateNSWAdd(LHS: Y, RHS: Z); |
| 674 | Value *NewPow = Builder.CreateIntrinsic( |
| 675 | ID: Intrinsic::powi, OverloadTypes: {X->getType(), YZ->getType()}, Args: {X, YZ}, FMFSource: &I); |
| 676 | |
| 677 | return NewPow; |
| 678 | }; |
| 679 | |
| 680 | Value *X, *Y, *Z; |
| 681 | unsigned Opcode = I.getOpcode(); |
| 682 | assert((Opcode == Instruction::FMul || Opcode == Instruction::FDiv) && |
| 683 | "Unexpected opcode" ); |
| 684 | |
| 685 | // powi(X, Y) * X --> powi(X, Y+1) |
| 686 | // X * powi(X, Y) --> powi(X, Y+1) |
| 687 | if (match(V: &I, P: m_c_FMul(L: m_OneUse(SubPattern: m_AllowReassoc(SubPattern: m_Intrinsic<Intrinsic::powi>( |
| 688 | Op0: m_Value(V&: X), Op1: m_Value(V&: Y)))), |
| 689 | R: m_Deferred(V: X)))) { |
| 690 | Constant *One = ConstantInt::get(Ty: Y->getType(), V: 1); |
| 691 | if (willNotOverflowSignedAdd(LHS: Y, RHS: One, CxtI: I)) { |
| 692 | Value *NewPow = createPowiExpr(I, *this, X, Y, One); |
| 693 | return replaceInstUsesWith(I, V: NewPow); |
| 694 | } |
| 695 | } |
| 696 | |
| 697 | // powi(x, y) * powi(x, z) -> powi(x, y + z) |
| 698 | Value *Op0 = I.getOperand(i_nocapture: 0); |
| 699 | Value *Op1 = I.getOperand(i_nocapture: 1); |
| 700 | if (Opcode == Instruction::FMul && I.isOnlyUserOfAnyOperand() && |
| 701 | match(V: Op0, P: m_AllowReassoc( |
| 702 | SubPattern: m_Intrinsic<Intrinsic::powi>(Op0: m_Value(V&: X), Op1: m_Value(V&: Y)))) && |
| 703 | match(V: Op1, P: m_AllowReassoc(SubPattern: m_Intrinsic<Intrinsic::powi>(Op0: m_Specific(V: X), |
| 704 | Op1: m_Value(V&: Z)))) && |
| 705 | Y->getType() == Z->getType() && willNotOverflowSignedAdd(LHS: Y, RHS: Z, CxtI: I)) { |
| 706 | Value *NewPow = createPowiExpr(I, *this, X, Y, Z); |
| 707 | return replaceInstUsesWith(I, V: NewPow); |
| 708 | } |
| 709 | |
| 710 | if (Opcode == Instruction::FDiv && I.hasAllowReassoc() && I.hasNoNaNs()) { |
| 711 | // powi(X, Y) / X --> powi(X, Y-1) |
| 712 | // This is legal when (Y - 1) can't wraparound, in which case reassoc and |
| 713 | // nnan are required. |
| 714 | // TODO: Multi-use may be also better off creating Powi(x,y-1) |
| 715 | if (match(V: Op0, P: m_OneUse(SubPattern: m_AllowReassoc(SubPattern: m_Intrinsic<Intrinsic::powi>( |
| 716 | Op0: m_Specific(V: Op1), Op1: m_Value(V&: Y))))) && |
| 717 | willNotOverflowSignedSub(LHS: Y, RHS: ConstantInt::get(Ty: Y->getType(), V: 1), CxtI: I)) { |
| 718 | Constant *NegOne = ConstantInt::getAllOnesValue(Ty: Y->getType()); |
| 719 | Value *NewPow = createPowiExpr(I, *this, Op1, Y, NegOne); |
| 720 | return replaceInstUsesWith(I, V: NewPow); |
| 721 | } |
| 722 | |
| 723 | // powi(X, Y) / (X * Z) --> powi(X, Y-1) / Z |
| 724 | // This is legal when (Y - 1) can't wraparound, in which case reassoc and |
| 725 | // nnan are required. |
| 726 | // TODO: Multi-use may be also better off creating Powi(x,y-1) |
| 727 | if (match(V: Op0, P: m_OneUse(SubPattern: m_AllowReassoc(SubPattern: m_Intrinsic<Intrinsic::powi>( |
| 728 | Op0: m_Value(V&: X), Op1: m_Value(V&: Y))))) && |
| 729 | match(V: Op1, P: m_AllowReassoc(SubPattern: m_c_FMul(L: m_Specific(V: X), R: m_Value(V&: Z)))) && |
| 730 | willNotOverflowSignedSub(LHS: Y, RHS: ConstantInt::get(Ty: Y->getType(), V: 1), CxtI: I)) { |
| 731 | Constant *NegOne = ConstantInt::getAllOnesValue(Ty: Y->getType()); |
| 732 | auto *NewPow = createPowiExpr(I, *this, X, Y, NegOne); |
| 733 | return BinaryOperator::CreateFDivFMF(V1: NewPow, V2: Z, FMFSource: &I); |
| 734 | } |
| 735 | } |
| 736 | |
| 737 | return nullptr; |
| 738 | } |
| 739 | |
| 740 | // If we have the following pattern, |
| 741 | // X = 1.0/sqrt(a) |
| 742 | // R1 = X * X |
| 743 | // R2 = a/sqrt(a) |
| 744 | // then this method collects all the instructions that match R1 and R2. |
| 745 | static bool getFSqrtDivOptPattern(Instruction *Div, |
| 746 | SmallPtrSetImpl<Instruction *> &R1, |
| 747 | SmallPtrSetImpl<Instruction *> &R2) { |
| 748 | Value *A; |
| 749 | if (match(V: Div, P: m_FDiv(L: m_FPOne(), R: m_Sqrt(Op0: m_Value(V&: A)))) || |
| 750 | match(V: Div, P: m_FDiv(L: m_SpecificFP(V: -1.0), R: m_Sqrt(Op0: m_Value(V&: A))))) { |
| 751 | for (User *U : Div->users()) { |
| 752 | Instruction *I = cast<Instruction>(Val: U); |
| 753 | if (match(V: I, P: m_FMul(L: m_Specific(V: Div), R: m_Specific(V: Div)))) |
| 754 | R1.insert(Ptr: I); |
| 755 | } |
| 756 | |
| 757 | CallInst *CI = cast<CallInst>(Val: Div->getOperand(i: 1)); |
| 758 | for (User *U : CI->users()) { |
| 759 | Instruction *I = cast<Instruction>(Val: U); |
| 760 | if (match(V: I, P: m_FDiv(L: m_Specific(V: A), R: m_Sqrt(Op0: m_Specific(V: A))))) |
| 761 | R2.insert(Ptr: I); |
| 762 | } |
| 763 | } |
| 764 | return !R1.empty() && !R2.empty(); |
| 765 | } |
| 766 | |
| 767 | // Check legality for transforming |
| 768 | // x = 1.0/sqrt(a) |
| 769 | // r1 = x * x; |
| 770 | // r2 = a/sqrt(a); |
| 771 | // |
| 772 | // TO |
| 773 | // |
| 774 | // r1 = 1/a |
| 775 | // r2 = sqrt(a) |
| 776 | // x = r1 * r2 |
| 777 | // This transform works only when 'a' is known positive. |
| 778 | static bool isFSqrtDivToFMulLegal(Instruction *X, |
| 779 | SmallPtrSetImpl<Instruction *> &R1, |
| 780 | SmallPtrSetImpl<Instruction *> &R2) { |
| 781 | // Check if the required pattern for the transformation exists. |
| 782 | if (!getFSqrtDivOptPattern(Div: X, R1, R2)) |
| 783 | return false; |
| 784 | |
| 785 | BasicBlock *BBx = X->getParent(); |
| 786 | BasicBlock *BBr1 = (*R1.begin())->getParent(); |
| 787 | BasicBlock *BBr2 = (*R2.begin())->getParent(); |
| 788 | |
| 789 | CallInst *FSqrt = cast<CallInst>(Val: X->getOperand(i: 1)); |
| 790 | if (!FSqrt->hasAllowReassoc() || !FSqrt->hasNoNaNs() || |
| 791 | !FSqrt->hasNoSignedZeros() || !FSqrt->hasNoInfs()) |
| 792 | return false; |
| 793 | |
| 794 | // We change x = 1/sqrt(a) to x = sqrt(a) * 1/a . This change isn't allowed |
| 795 | // by recip fp as it is strictly meant to transform ops of type a/b to |
| 796 | // a * 1/b. So, this can be considered as algebraic rewrite and reassoc flag |
| 797 | // has been used(rather abused)in the past for algebraic rewrites. |
| 798 | if (!X->hasAllowReassoc() || !X->hasAllowReciprocal() || !X->hasNoInfs()) |
| 799 | return false; |
| 800 | |
| 801 | // Check the constraints on X, R1 and R2 combined. |
| 802 | // fdiv instruction and one of the multiplications must reside in the same |
| 803 | // block. If not, the optimized code may execute more ops than before and |
| 804 | // this may hamper the performance. |
| 805 | if (BBx != BBr1 && BBx != BBr2) |
| 806 | return false; |
| 807 | |
| 808 | // Check the constraints on instructions in R1. |
| 809 | if (any_of(Range&: R1, P: [BBr1](Instruction *I) { |
| 810 | // When you have multiple instructions residing in R1 and R2 |
| 811 | // respectively, it's difficult to generate combinations of (R1,R2) and |
| 812 | // then check if we have the required pattern. So, for now, just be |
| 813 | // conservative. |
| 814 | return (I->getParent() != BBr1 || !I->hasAllowReassoc()); |
| 815 | })) |
| 816 | return false; |
| 817 | |
| 818 | // Check the constraints on instructions in R2. |
| 819 | return all_of(Range&: R2, P: [BBr2](Instruction *I) { |
| 820 | // When you have multiple instructions residing in R1 and R2 |
| 821 | // respectively, it's difficult to generate combination of (R1,R2) and |
| 822 | // then check if we have the required pattern. So, for now, just be |
| 823 | // conservative. |
| 824 | return (I->getParent() == BBr2 && I->hasAllowReassoc()); |
| 825 | }); |
| 826 | } |
| 827 | |
| 828 | Instruction *InstCombinerImpl::foldFMulReassoc(BinaryOperator &I) { |
| 829 | Value *Op0 = I.getOperand(i_nocapture: 0); |
| 830 | Value *Op1 = I.getOperand(i_nocapture: 1); |
| 831 | Value *X, *Y; |
| 832 | Constant *C; |
| 833 | BinaryOperator *Op0BinOp; |
| 834 | |
| 835 | // Reassociate constant RHS with another constant to form constant |
| 836 | // expression. |
| 837 | if (match(V: Op1, P: m_Constant(C)) && C->isFiniteNonZeroFP() && |
| 838 | match(V: Op0, P: m_AllowReassoc(SubPattern: m_BinOp(I&: Op0BinOp)))) { |
| 839 | // Everything in this scope folds I with Op0, intersecting their FMF. |
| 840 | FastMathFlags FMF = I.getFastMathFlags() & Op0BinOp->getFastMathFlags(); |
| 841 | Constant *C1; |
| 842 | if (match(V: Op0, P: m_OneUse(SubPattern: m_FDiv(L: m_Constant(C&: C1), R: m_Value(V&: X))))) { |
| 843 | // (C1 / X) * C --> (C * C1) / X |
| 844 | Constant *CC1 = |
| 845 | ConstantFoldBinaryOpOperands(Opcode: Instruction::FMul, LHS: C, RHS: C1, DL); |
| 846 | if (CC1 && CC1->isNormalFP()) |
| 847 | return BinaryOperator::CreateFDivFMF(V1: CC1, V2: X, FMF); |
| 848 | } |
| 849 | if (match(V: Op0, P: m_FDiv(L: m_Value(V&: X), R: m_Constant(C&: C1)))) { |
| 850 | // FIXME: This seems like it should also be checking for arcp |
| 851 | // (X / C1) * C --> X * (C / C1) |
| 852 | Constant *CDivC1 = |
| 853 | ConstantFoldBinaryOpOperands(Opcode: Instruction::FDiv, LHS: C, RHS: C1, DL); |
| 854 | if (CDivC1 && CDivC1->isNormalFP()) |
| 855 | return BinaryOperator::CreateFMulFMF(V1: X, V2: CDivC1, FMF); |
| 856 | |
| 857 | // If the constant was a denormal, try reassociating differently. |
| 858 | // (X / C1) * C --> X / (C1 / C) |
| 859 | Constant *C1DivC = |
| 860 | ConstantFoldBinaryOpOperands(Opcode: Instruction::FDiv, LHS: C1, RHS: C, DL); |
| 861 | if (C1DivC && Op0->hasOneUse() && C1DivC->isNormalFP()) |
| 862 | return BinaryOperator::CreateFDivFMF(V1: X, V2: C1DivC, FMF); |
| 863 | } |
| 864 | |
| 865 | // We do not need to match 'fadd C, X' and 'fsub X, C' because they are |
| 866 | // canonicalized to 'fadd X, C'. Distributing the multiply may allow |
| 867 | // further folds and (X * C) + C2 is 'fma'. |
| 868 | if (match(V: Op0, P: m_OneUse(SubPattern: m_FAdd(L: m_Value(V&: X), R: m_Constant(C&: C1))))) { |
| 869 | // (X + C1) * C --> (X * C) + (C * C1) |
| 870 | if (Constant *CC1 = |
| 871 | ConstantFoldBinaryOpOperands(Opcode: Instruction::FMul, LHS: C, RHS: C1, DL)) { |
| 872 | Value *XC = Builder.CreateFMulFMF(L: X, R: C, FMFSource: FMF); |
| 873 | return BinaryOperator::CreateFAddFMF(V1: XC, V2: CC1, FMF); |
| 874 | } |
| 875 | } |
| 876 | if (match(V: Op0, P: m_OneUse(SubPattern: m_FSub(L: m_Constant(C&: C1), R: m_Value(V&: X))))) { |
| 877 | // (C1 - X) * C --> (C * C1) - (X * C) |
| 878 | if (Constant *CC1 = |
| 879 | ConstantFoldBinaryOpOperands(Opcode: Instruction::FMul, LHS: C, RHS: C1, DL)) { |
| 880 | Value *XC = Builder.CreateFMulFMF(L: X, R: C, FMFSource: FMF); |
| 881 | return BinaryOperator::CreateFSubFMF(V1: CC1, V2: XC, FMF); |
| 882 | } |
| 883 | } |
| 884 | } |
| 885 | |
| 886 | Value *Z; |
| 887 | if (match(V: &I, |
| 888 | P: m_c_FMul(L: m_AllowReassoc(SubPattern: m_OneUse(SubPattern: m_FDiv(L: m_Value(V&: X), R: m_Value(V&: Y)))), |
| 889 | R: m_Value(V&: Z)))) { |
| 890 | BinaryOperator *DivOp = cast<BinaryOperator>(Val: ((Z == Op0) ? Op1 : Op0)); |
| 891 | FastMathFlags FMF = I.getFastMathFlags() & DivOp->getFastMathFlags(); |
| 892 | if (FMF.allowReassoc()) { |
| 893 | // Sink division: (X / Y) * Z --> (X * Z) / Y |
| 894 | auto *NewFMul = Builder.CreateFMulFMF(L: X, R: Z, FMFSource: FMF); |
| 895 | return BinaryOperator::CreateFDivFMF(V1: NewFMul, V2: Y, FMF); |
| 896 | } |
| 897 | } |
| 898 | |
| 899 | // sqrt(X) * sqrt(Y) -> sqrt(X * Y) |
| 900 | // nnan disallows the possibility of returning a number if both operands are |
| 901 | // negative (in that case, we should return NaN). |
| 902 | if (I.hasNoNaNs() && match(V: Op0, P: m_OneUse(SubPattern: m_Sqrt(Op0: m_Value(V&: X)))) && |
| 903 | match(V: Op1, P: m_OneUse(SubPattern: m_Sqrt(Op0: m_Value(V&: Y))))) { |
| 904 | Value *XY = Builder.CreateFMulFMF(L: X, R: Y, FMFSource: &I); |
| 905 | Value *Sqrt = Builder.CreateUnaryIntrinsic(ID: Intrinsic::sqrt, Op: XY, FMFSource: &I); |
| 906 | return replaceInstUsesWith(I, V: Sqrt); |
| 907 | } |
| 908 | |
| 909 | // The following transforms are done irrespective of the number of uses |
| 910 | // for the expression "1.0/sqrt(X)". |
| 911 | // 1) 1.0/sqrt(X) * X -> X/sqrt(X) |
| 912 | // 2) X * 1.0/sqrt(X) -> X/sqrt(X) |
| 913 | // We always expect the backend to reduce X/sqrt(X) to sqrt(X), if it |
| 914 | // has the necessary (reassoc) fast-math-flags. |
| 915 | if (I.hasNoSignedZeros() && |
| 916 | match(V: Op0, P: (m_FDiv(L: m_SpecificFP(V: 1.0), R: m_Value(V&: Y)))) && |
| 917 | match(V: Y, P: m_Sqrt(Op0: m_Value(V&: X))) && Op1 == X) |
| 918 | return BinaryOperator::CreateFDivFMF(V1: X, V2: Y, FMFSource: &I); |
| 919 | if (I.hasNoSignedZeros() && |
| 920 | match(V: Op1, P: (m_FDiv(L: m_SpecificFP(V: 1.0), R: m_Value(V&: Y)))) && |
| 921 | match(V: Y, P: m_Sqrt(Op0: m_Value(V&: X))) && Op0 == X) |
| 922 | return BinaryOperator::CreateFDivFMF(V1: X, V2: Y, FMFSource: &I); |
| 923 | |
| 924 | // Like the similar transform in instsimplify, this requires 'nsz' because |
| 925 | // sqrt(-0.0) = -0.0, and -0.0 * -0.0 does not simplify to -0.0. |
| 926 | if (I.hasNoNaNs() && I.hasNoSignedZeros() && Op0 == Op1 && Op0->hasNUses(N: 2)) { |
| 927 | // Peek through fdiv to find squaring of square root: |
| 928 | // (X / sqrt(Y)) * (X / sqrt(Y)) --> (X * X) / Y |
| 929 | if (match(V: Op0, P: m_FDiv(L: m_Value(V&: X), R: m_Sqrt(Op0: m_Value(V&: Y))))) { |
| 930 | Value *XX = Builder.CreateFMulFMF(L: X, R: X, FMFSource: &I); |
| 931 | return BinaryOperator::CreateFDivFMF(V1: XX, V2: Y, FMFSource: &I); |
| 932 | } |
| 933 | // (sqrt(Y) / X) * (sqrt(Y) / X) --> Y / (X * X) |
| 934 | if (match(V: Op0, P: m_FDiv(L: m_Sqrt(Op0: m_Value(V&: Y)), R: m_Value(V&: X)))) { |
| 935 | Value *XX = Builder.CreateFMulFMF(L: X, R: X, FMFSource: &I); |
| 936 | return BinaryOperator::CreateFDivFMF(V1: Y, V2: XX, FMFSource: &I); |
| 937 | } |
| 938 | } |
| 939 | |
| 940 | // pow(X, Y) * X --> pow(X, Y+1) |
| 941 | // X * pow(X, Y) --> pow(X, Y+1) |
| 942 | if (match(V: &I, P: m_c_FMul(L: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::pow>(Op0: m_Value(V&: X), |
| 943 | Op1: m_Value(V&: Y))), |
| 944 | R: m_Deferred(V: X)))) { |
| 945 | Value *Y1 = Builder.CreateFAddFMF(L: Y, R: ConstantFP::get(Ty: I.getType(), V: 1.0), FMFSource: &I); |
| 946 | Value *Pow = Builder.CreateBinaryIntrinsic(ID: Intrinsic::pow, LHS: X, RHS: Y1, FMFSource: &I); |
| 947 | return replaceInstUsesWith(I, V: Pow); |
| 948 | } |
| 949 | |
| 950 | if (Instruction *FoldedPowi = foldPowiReassoc(I)) |
| 951 | return FoldedPowi; |
| 952 | |
| 953 | if (I.isOnlyUserOfAnyOperand()) { |
| 954 | // pow(X, Y) * pow(X, Z) -> pow(X, Y + Z) |
| 955 | if (match(V: Op0, P: m_Intrinsic<Intrinsic::pow>(Op0: m_Value(V&: X), Op1: m_Value(V&: Y))) && |
| 956 | match(V: Op1, P: m_Intrinsic<Intrinsic::pow>(Op0: m_Specific(V: X), Op1: m_Value(V&: Z)))) { |
| 957 | auto *YZ = Builder.CreateFAddFMF(L: Y, R: Z, FMFSource: &I); |
| 958 | auto *NewPow = Builder.CreateBinaryIntrinsic(ID: Intrinsic::pow, LHS: X, RHS: YZ, FMFSource: &I); |
| 959 | return replaceInstUsesWith(I, V: NewPow); |
| 960 | } |
| 961 | // pow(X, Y) * pow(Z, Y) -> pow(X * Z, Y) |
| 962 | if (match(V: Op0, P: m_Intrinsic<Intrinsic::pow>(Op0: m_Value(V&: X), Op1: m_Value(V&: Y))) && |
| 963 | match(V: Op1, P: m_Intrinsic<Intrinsic::pow>(Op0: m_Value(V&: Z), Op1: m_Specific(V: Y)))) { |
| 964 | auto *XZ = Builder.CreateFMulFMF(L: X, R: Z, FMFSource: &I); |
| 965 | auto *NewPow = Builder.CreateBinaryIntrinsic(ID: Intrinsic::pow, LHS: XZ, RHS: Y, FMFSource: &I); |
| 966 | return replaceInstUsesWith(I, V: NewPow); |
| 967 | } |
| 968 | |
| 969 | // exp(X) * exp(Y) -> exp(X + Y) |
| 970 | if (match(V: Op0, P: m_Intrinsic<Intrinsic::exp>(Op0: m_Value(V&: X))) && |
| 971 | match(V: Op1, P: m_Intrinsic<Intrinsic::exp>(Op0: m_Value(V&: Y)))) { |
| 972 | Value *XY = Builder.CreateFAddFMF(L: X, R: Y, FMFSource: &I); |
| 973 | Value *Exp = Builder.CreateUnaryIntrinsic(ID: Intrinsic::exp, Op: XY, FMFSource: &I); |
| 974 | return replaceInstUsesWith(I, V: Exp); |
| 975 | } |
| 976 | |
| 977 | // exp2(X) * exp2(Y) -> exp2(X + Y) |
| 978 | if (match(V: Op0, P: m_Intrinsic<Intrinsic::exp2>(Op0: m_Value(V&: X))) && |
| 979 | match(V: Op1, P: m_Intrinsic<Intrinsic::exp2>(Op0: m_Value(V&: Y)))) { |
| 980 | Value *XY = Builder.CreateFAddFMF(L: X, R: Y, FMFSource: &I); |
| 981 | Value *Exp2 = Builder.CreateUnaryIntrinsic(ID: Intrinsic::exp2, Op: XY, FMFSource: &I); |
| 982 | return replaceInstUsesWith(I, V: Exp2); |
| 983 | } |
| 984 | } |
| 985 | |
| 986 | // (X*Y) * X => (X*X) * Y where Y != X |
| 987 | // The purpose is two-fold: |
| 988 | // 1) to form a power expression (of X). |
| 989 | // 2) potentially shorten the critical path: After transformation, the |
| 990 | // latency of the instruction Y is amortized by the expression of X*X, |
| 991 | // and therefore Y is in a "less critical" position compared to what it |
| 992 | // was before the transformation. |
| 993 | if (match(V: Op0, P: m_OneUse(SubPattern: m_c_FMul(L: m_Specific(V: Op1), R: m_Value(V&: Y)))) && Op1 != Y) { |
| 994 | Value *XX = Builder.CreateFMulFMF(L: Op1, R: Op1, FMFSource: &I); |
| 995 | return BinaryOperator::CreateFMulFMF(V1: XX, V2: Y, FMFSource: &I); |
| 996 | } |
| 997 | if (match(V: Op1, P: m_OneUse(SubPattern: m_c_FMul(L: m_Specific(V: Op0), R: m_Value(V&: Y)))) && Op0 != Y) { |
| 998 | Value *XX = Builder.CreateFMulFMF(L: Op0, R: Op0, FMFSource: &I); |
| 999 | return BinaryOperator::CreateFMulFMF(V1: XX, V2: Y, FMFSource: &I); |
| 1000 | } |
| 1001 | |
| 1002 | return nullptr; |
| 1003 | } |
| 1004 | |
| 1005 | Instruction *InstCombinerImpl::visitFMul(BinaryOperator &I) { |
| 1006 | if (Value *V = simplifyFMulInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), |
| 1007 | FMF: I.getFastMathFlags(), |
| 1008 | Q: SQ.getWithInstruction(I: &I))) |
| 1009 | return replaceInstUsesWith(I, V); |
| 1010 | |
| 1011 | if (SimplifyAssociativeOrCommutative(I)) |
| 1012 | return &I; |
| 1013 | |
| 1014 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
| 1015 | return X; |
| 1016 | |
| 1017 | if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I)) |
| 1018 | return Phi; |
| 1019 | |
| 1020 | if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I)) |
| 1021 | return FoldedMul; |
| 1022 | |
| 1023 | if (Value *FoldedMul = foldMulSelectToNegate(I, Builder)) |
| 1024 | return replaceInstUsesWith(I, V: FoldedMul); |
| 1025 | |
| 1026 | if (Instruction *R = foldFPSignBitOps(I)) |
| 1027 | return R; |
| 1028 | |
| 1029 | if (Instruction *R = foldFBinOpOfIntCasts(I)) |
| 1030 | return R; |
| 1031 | |
| 1032 | // X * -1.0 --> -X |
| 1033 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 1034 | if (match(V: Op1, P: m_SpecificFP(V: -1.0))) |
| 1035 | return UnaryOperator::CreateFNegFMF(Op: Op0, FMFSource: &I); |
| 1036 | |
| 1037 | // -X * C --> X * -C |
| 1038 | Value *X, *Y; |
| 1039 | Constant *C; |
| 1040 | if (match(V: Op0, P: m_FNeg(X: m_Value(V&: X))) && match(V: Op1, P: m_Constant(C))) |
| 1041 | if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)) |
| 1042 | return BinaryOperator::CreateFMulFMF(V1: X, V2: NegC, FMFSource: &I); |
| 1043 | |
| 1044 | if (I.hasNoNaNs() && I.hasNoSignedZeros()) { |
| 1045 | // (uitofp bool X) * Y --> X ? Y : 0 |
| 1046 | // Y * (uitofp bool X) --> X ? Y : 0 |
| 1047 | // Note INF * 0 is NaN. |
| 1048 | if (match(V: Op0, P: m_UIToFP(Op: m_Value(V&: X))) && |
| 1049 | X->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
| 1050 | auto *SI = createSelectInstWithUnknownProfile( |
| 1051 | C: X, S1: Op1, S2: ConstantFP::get(Ty: I.getType(), V: 0.0)); |
| 1052 | SI->copyFastMathFlags(FMF: I.getFastMathFlags()); |
| 1053 | return SI; |
| 1054 | } |
| 1055 | if (match(V: Op1, P: m_UIToFP(Op: m_Value(V&: X))) && |
| 1056 | X->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
| 1057 | auto *SI = createSelectInstWithUnknownProfile( |
| 1058 | C: X, S1: Op0, S2: ConstantFP::get(Ty: I.getType(), V: 0.0)); |
| 1059 | SI->copyFastMathFlags(FMF: I.getFastMathFlags()); |
| 1060 | return SI; |
| 1061 | } |
| 1062 | } |
| 1063 | |
| 1064 | // (select A, B, C) * (select A, D, E) --> select A, (B*D), (C*E) |
| 1065 | if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS: Op0, RHS: Op1)) |
| 1066 | return replaceInstUsesWith(I, V); |
| 1067 | |
| 1068 | if (I.hasAllowReassoc()) |
| 1069 | if (Instruction *FoldedMul = foldFMulReassoc(I)) |
| 1070 | return FoldedMul; |
| 1071 | |
| 1072 | // log2(X * 0.5) * Y = log2(X) * Y - Y |
| 1073 | if (I.isFast()) { |
| 1074 | IntrinsicInst *Log2 = nullptr; |
| 1075 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::log2>( |
| 1076 | Op0: m_OneUse(SubPattern: m_FMul(L: m_Value(V&: X), R: m_SpecificFP(V: 0.5))))))) { |
| 1077 | Log2 = cast<IntrinsicInst>(Val: Op0); |
| 1078 | Y = Op1; |
| 1079 | } |
| 1080 | if (match(V: Op1, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::log2>( |
| 1081 | Op0: m_OneUse(SubPattern: m_FMul(L: m_Value(V&: X), R: m_SpecificFP(V: 0.5))))))) { |
| 1082 | Log2 = cast<IntrinsicInst>(Val: Op1); |
| 1083 | Y = Op0; |
| 1084 | } |
| 1085 | if (Log2) { |
| 1086 | Value *Log2 = Builder.CreateUnaryIntrinsic(ID: Intrinsic::log2, Op: X, FMFSource: &I); |
| 1087 | Value *LogXTimesY = Builder.CreateFMulFMF(L: Log2, R: Y, FMFSource: &I); |
| 1088 | return BinaryOperator::CreateFSubFMF(V1: LogXTimesY, V2: Y, FMFSource: &I); |
| 1089 | } |
| 1090 | } |
| 1091 | |
| 1092 | // Simplify FMUL recurrences starting with 0.0 to 0.0 if nnan and nsz are set. |
| 1093 | // Given a phi node with entry value as 0 and it used in fmul operation, |
| 1094 | // we can replace fmul with 0 safely and eleminate loop operation. |
| 1095 | PHINode *PN = nullptr; |
| 1096 | Value *Start = nullptr, *Step = nullptr; |
| 1097 | if (matchSimpleRecurrence(I: &I, P&: PN, Start, Step) && I.hasNoNaNs() && |
| 1098 | I.hasNoSignedZeros() && match(V: Start, P: m_Zero())) |
| 1099 | return replaceInstUsesWith(I, V: Start); |
| 1100 | |
| 1101 | // minimum(X, Y) * maximum(X, Y) => X * Y. |
| 1102 | if (match(V: &I, |
| 1103 | P: m_c_FMul(L: m_Intrinsic<Intrinsic::maximum>(Op0: m_Value(V&: X), Op1: m_Value(V&: Y)), |
| 1104 | R: m_c_Intrinsic<Intrinsic::minimum>(Op0: m_Deferred(V: X), |
| 1105 | Op1: m_Deferred(V: Y))))) { |
| 1106 | BinaryOperator *Result = BinaryOperator::CreateFMulFMF(V1: X, V2: Y, FMFSource: &I); |
| 1107 | // We cannot preserve ninf if nnan flag is not set. |
| 1108 | // If X is NaN and Y is Inf then in original program we had NaN * NaN, |
| 1109 | // while in optimized version NaN * Inf and this is a poison with ninf flag. |
| 1110 | if (!Result->hasNoNaNs()) |
| 1111 | Result->setHasNoInfs(false); |
| 1112 | return Result; |
| 1113 | } |
| 1114 | |
| 1115 | // tan(X) * cos(X) -> sin(X) |
| 1116 | if (I.hasAllowContract() && |
| 1117 | match(V: &I, |
| 1118 | P: m_c_FMul(L: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::tan>(Op0: m_Value(V&: X))), |
| 1119 | R: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::cos>(Op0: m_Deferred(V: X)))))) { |
| 1120 | Value *Sin = Builder.CreateUnaryIntrinsic(ID: Intrinsic::sin, Op: X, FMFSource: &I); |
| 1121 | if (auto *Metadata = I.getMetadata(KindID: LLVMContext::MD_fpmath)) |
| 1122 | if (auto *SinI = dyn_cast<Instruction>(Val: Sin)) |
| 1123 | SinI->setMetadata(KindID: LLVMContext::MD_fpmath, Node: Metadata); |
| 1124 | return replaceInstUsesWith(I, V: Sin); |
| 1125 | } |
| 1126 | |
| 1127 | // X * ldexp(1.0, Y) -> ldexp(X, Y) |
| 1128 | if (match(V: &I, P: m_AllowReassoc(SubPattern: m_c_FMul( |
| 1129 | L: m_Value(V&: X), |
| 1130 | R: m_AllowReassoc(SubPattern: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::ldexp>( |
| 1131 | Op0: m_FPOne(), Op1: m_Value(V&: Y)))))))) |
| 1132 | return replaceInstUsesWith( |
| 1133 | I, V: Builder.CreateIntrinsic(ID: Intrinsic::ldexp, |
| 1134 | OverloadTypes: {X->getType(), Y->getType()}, Args: {X, Y}, FMFSource: &I)); |
| 1135 | |
| 1136 | if (SimplifyDemandedInstructionFPClass(Inst&: I)) |
| 1137 | return &I; |
| 1138 | |
| 1139 | return nullptr; |
| 1140 | } |
| 1141 | |
| 1142 | /// Fold a divide or remainder with a select instruction divisor when one of the |
| 1143 | /// select operands is zero. In that case, we can use the other select operand |
| 1144 | /// because div/rem by zero is undefined. |
| 1145 | bool InstCombinerImpl::simplifyDivRemOfSelectWithZeroOp(BinaryOperator &I) { |
| 1146 | SelectInst *SI = dyn_cast<SelectInst>(Val: I.getOperand(i_nocapture: 1)); |
| 1147 | if (!SI) |
| 1148 | return false; |
| 1149 | |
| 1150 | int NonNullOperand; |
| 1151 | if (match(V: SI->getTrueValue(), P: m_Zero())) |
| 1152 | // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y |
| 1153 | NonNullOperand = 2; |
| 1154 | else if (match(V: SI->getFalseValue(), P: m_Zero())) |
| 1155 | // div/rem X, (Cond ? Y : 0) -> div/rem X, Y |
| 1156 | NonNullOperand = 1; |
| 1157 | else |
| 1158 | return false; |
| 1159 | |
| 1160 | // Change the div/rem to use 'Y' instead of the select. |
| 1161 | replaceOperand(I, OpNum: 1, V: SI->getOperand(i_nocapture: NonNullOperand)); |
| 1162 | |
| 1163 | // Okay, we know we replace the operand of the div/rem with 'Y' with no |
| 1164 | // problem. However, the select, or the condition of the select may have |
| 1165 | // multiple uses. Based on our knowledge that the operand must be non-zero, |
| 1166 | // propagate the known value for the select into other uses of it, and |
| 1167 | // propagate a known value of the condition into its other users. |
| 1168 | |
| 1169 | // If the select and condition only have a single use, don't bother with this, |
| 1170 | // early exit. |
| 1171 | Value *SelectCond = SI->getCondition(); |
| 1172 | if (SI->use_empty() && SelectCond->hasOneUse()) |
| 1173 | return true; |
| 1174 | |
| 1175 | // Scan the current block backward, looking for other uses of SI. |
| 1176 | BasicBlock::iterator BBI = I.getIterator(), BBFront = I.getParent()->begin(); |
| 1177 | Type *CondTy = SelectCond->getType(); |
| 1178 | while (BBI != BBFront) { |
| 1179 | --BBI; |
| 1180 | // If we found an instruction that we can't assume will return, so |
| 1181 | // information from below it cannot be propagated above it. |
| 1182 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &*BBI)) |
| 1183 | break; |
| 1184 | |
| 1185 | // Replace uses of the select or its condition with the known values. |
| 1186 | for (Use &Op : BBI->operands()) { |
| 1187 | if (Op == SI) { |
| 1188 | replaceUse(U&: Op, NewValue: SI->getOperand(i_nocapture: NonNullOperand)); |
| 1189 | Worklist.push(I: &*BBI); |
| 1190 | } else if (Op == SelectCond) { |
| 1191 | replaceUse(U&: Op, NewValue: NonNullOperand == 1 ? ConstantInt::getTrue(Ty: CondTy) |
| 1192 | : ConstantInt::getFalse(Ty: CondTy)); |
| 1193 | Worklist.push(I: &*BBI); |
| 1194 | } |
| 1195 | } |
| 1196 | |
| 1197 | // If we past the instruction, quit looking for it. |
| 1198 | if (&*BBI == SI) |
| 1199 | SI = nullptr; |
| 1200 | if (&*BBI == SelectCond) |
| 1201 | SelectCond = nullptr; |
| 1202 | |
| 1203 | // If we ran out of things to eliminate, break out of the loop. |
| 1204 | if (!SelectCond && !SI) |
| 1205 | break; |
| 1206 | |
| 1207 | } |
| 1208 | return true; |
| 1209 | } |
| 1210 | |
| 1211 | /// True if the multiply can not be expressed in an int this size. |
| 1212 | static bool multiplyOverflows(const APInt &C1, const APInt &C2, APInt &Product, |
| 1213 | bool IsSigned) { |
| 1214 | bool Overflow; |
| 1215 | Product = IsSigned ? C1.smul_ov(RHS: C2, Overflow) : C1.umul_ov(RHS: C2, Overflow); |
| 1216 | return Overflow; |
| 1217 | } |
| 1218 | |
| 1219 | /// True if C1 is a multiple of C2. Quotient contains C1/C2. |
| 1220 | static bool isMultiple(const APInt &C1, const APInt &C2, APInt &Quotient, |
| 1221 | bool IsSigned) { |
| 1222 | assert(C1.getBitWidth() == C2.getBitWidth() && "Constant widths not equal" ); |
| 1223 | |
| 1224 | // Bail if we will divide by zero. |
| 1225 | if (C2.isZero()) |
| 1226 | return false; |
| 1227 | |
| 1228 | // Bail if we would divide INT_MIN by -1. |
| 1229 | if (IsSigned && C1.isMinSignedValue() && C2.isAllOnes()) |
| 1230 | return false; |
| 1231 | |
| 1232 | APInt Remainder(C1.getBitWidth(), /*val=*/0ULL, IsSigned); |
| 1233 | if (IsSigned) |
| 1234 | APInt::sdivrem(LHS: C1, RHS: C2, Quotient, Remainder); |
| 1235 | else |
| 1236 | APInt::udivrem(LHS: C1, RHS: C2, Quotient, Remainder); |
| 1237 | |
| 1238 | return Remainder.isMinValue(); |
| 1239 | } |
| 1240 | |
| 1241 | static Value *foldIDivShl(BinaryOperator &I, InstCombiner::BuilderTy &Builder) { |
| 1242 | assert((I.getOpcode() == Instruction::SDiv || |
| 1243 | I.getOpcode() == Instruction::UDiv) && |
| 1244 | "Expected integer divide" ); |
| 1245 | |
| 1246 | bool IsSigned = I.getOpcode() == Instruction::SDiv; |
| 1247 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 1248 | Type *Ty = I.getType(); |
| 1249 | |
| 1250 | Value *X, *Y, *Z; |
| 1251 | |
| 1252 | // With appropriate no-wrap constraints, remove a common factor in the |
| 1253 | // dividend and divisor that is disguised as a left-shifted value. |
| 1254 | if (match(V: Op1, P: m_Shl(L: m_Value(V&: X), R: m_Value(V&: Z))) && |
| 1255 | match(V: Op0, P: m_c_Mul(L: m_Specific(V: X), R: m_Value(V&: Y)))) { |
| 1256 | // Both operands must have the matching no-wrap for this kind of division. |
| 1257 | auto *Mul = cast<OverflowingBinaryOperator>(Val: Op0); |
| 1258 | auto *Shl = cast<OverflowingBinaryOperator>(Val: Op1); |
| 1259 | bool HasNUW = Mul->hasNoUnsignedWrap() && Shl->hasNoUnsignedWrap(); |
| 1260 | bool HasNSW = Mul->hasNoSignedWrap() && Shl->hasNoSignedWrap(); |
| 1261 | |
| 1262 | // (X * Y) u/ (X << Z) --> Y u>> Z |
| 1263 | if (!IsSigned && HasNUW) |
| 1264 | return Builder.CreateLShr(LHS: Y, RHS: Z, Name: "" , isExact: I.isExact()); |
| 1265 | |
| 1266 | // (X * Y) s/ (X << Z) --> Y s/ (1 << Z) |
| 1267 | if (IsSigned && HasNSW && (Op0->hasOneUse() || Op1->hasOneUse())) { |
| 1268 | Value *Shl = Builder.CreateShl(LHS: ConstantInt::get(Ty, V: 1), RHS: Z); |
| 1269 | return Builder.CreateSDiv(LHS: Y, RHS: Shl, Name: "" , isExact: I.isExact()); |
| 1270 | } |
| 1271 | } |
| 1272 | |
| 1273 | // With appropriate no-wrap constraints, remove a common factor in the |
| 1274 | // dividend and divisor that is disguised as a left-shift amount. |
| 1275 | if (match(V: Op0, P: m_Shl(L: m_Value(V&: X), R: m_Value(V&: Z))) && |
| 1276 | match(V: Op1, P: m_Shl(L: m_Value(V&: Y), R: m_Specific(V: Z)))) { |
| 1277 | auto *Shl0 = cast<OverflowingBinaryOperator>(Val: Op0); |
| 1278 | auto *Shl1 = cast<OverflowingBinaryOperator>(Val: Op1); |
| 1279 | |
| 1280 | // For unsigned div, we need 'nuw' on both shifts or |
| 1281 | // 'nsw' on both shifts + 'nuw' on the dividend. |
| 1282 | // (X << Z) / (Y << Z) --> X / Y |
| 1283 | if (!IsSigned && |
| 1284 | ((Shl0->hasNoUnsignedWrap() && Shl1->hasNoUnsignedWrap()) || |
| 1285 | (Shl0->hasNoUnsignedWrap() && Shl0->hasNoSignedWrap() && |
| 1286 | Shl1->hasNoSignedWrap()))) |
| 1287 | return Builder.CreateUDiv(LHS: X, RHS: Y, Name: "" , isExact: I.isExact()); |
| 1288 | |
| 1289 | // For signed div, we need 'nsw' on both shifts + 'nuw' on the divisor. |
| 1290 | // (X << Z) / (Y << Z) --> X / Y |
| 1291 | if (IsSigned && Shl0->hasNoSignedWrap() && Shl1->hasNoSignedWrap() && |
| 1292 | Shl1->hasNoUnsignedWrap()) |
| 1293 | return Builder.CreateSDiv(LHS: X, RHS: Y, Name: "" , isExact: I.isExact()); |
| 1294 | } |
| 1295 | |
| 1296 | // If X << Y and X << Z does not overflow, then: |
| 1297 | // (X << Y) / (X << Z) -> (1 << Y) / (1 << Z) -> 1 << Y >> Z |
| 1298 | if (match(V: Op0, P: m_Shl(L: m_Value(V&: X), R: m_Value(V&: Y))) && |
| 1299 | match(V: Op1, P: m_Shl(L: m_Specific(V: X), R: m_Value(V&: Z)))) { |
| 1300 | auto *Shl0 = cast<OverflowingBinaryOperator>(Val: Op0); |
| 1301 | auto *Shl1 = cast<OverflowingBinaryOperator>(Val: Op1); |
| 1302 | |
| 1303 | if (IsSigned ? (Shl0->hasNoSignedWrap() && Shl1->hasNoSignedWrap()) |
| 1304 | : (Shl0->hasNoUnsignedWrap() && Shl1->hasNoUnsignedWrap())) { |
| 1305 | Constant *One = ConstantInt::get(Ty: X->getType(), V: 1); |
| 1306 | // Only preserve the nsw flag if dividend has nsw |
| 1307 | // or divisor has nsw and operator is sdiv. |
| 1308 | Value *Dividend = Builder.CreateShl( |
| 1309 | LHS: One, RHS: Y, Name: "shl.dividend" , |
| 1310 | /*HasNUW=*/true, |
| 1311 | /*HasNSW=*/ |
| 1312 | IsSigned ? (Shl0->hasNoUnsignedWrap() || Shl1->hasNoUnsignedWrap()) |
| 1313 | : Shl0->hasNoSignedWrap()); |
| 1314 | return Builder.CreateLShr(LHS: Dividend, RHS: Z, Name: "" , isExact: I.isExact()); |
| 1315 | } |
| 1316 | } |
| 1317 | |
| 1318 | return nullptr; |
| 1319 | } |
| 1320 | |
| 1321 | /// Common integer divide/remainder transforms |
| 1322 | Instruction *InstCombinerImpl::commonIDivRemTransforms(BinaryOperator &I) { |
| 1323 | assert(I.isIntDivRem() && "Unexpected instruction" ); |
| 1324 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 1325 | |
| 1326 | // If any element of a constant divisor fixed width vector is zero or undef |
| 1327 | // the behavior is undefined and we can fold the whole op to poison. |
| 1328 | if (match(V: Op1, P: m_ContainsMatchingVectorElement( |
| 1329 | SubPattern: m_CombineOr(Ps: m_Zero(), Ps: m_UndefValue())))) { |
| 1330 | return replaceInstUsesWith(I, V: PoisonValue::get(T: I.getType())); |
| 1331 | } |
| 1332 | |
| 1333 | if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I)) |
| 1334 | return Phi; |
| 1335 | |
| 1336 | // The RHS is known non-zero. |
| 1337 | if (Value *V = simplifyValueKnownNonZero(V: I.getOperand(i_nocapture: 1), IC&: *this, CxtI&: I)) |
| 1338 | return replaceOperand(I, OpNum: 1, V); |
| 1339 | |
| 1340 | // Handle cases involving: div/rem X, (select Cond, Y, Z) |
| 1341 | if (simplifyDivRemOfSelectWithZeroOp(I)) |
| 1342 | return &I; |
| 1343 | |
| 1344 | // If the divisor is a select-of-constants, try to constant fold all div ops: |
| 1345 | // C div/rem (select Cond, TrueC, FalseC) --> select Cond, (C div/rem TrueC), |
| 1346 | // (C div/rem FalseC) |
| 1347 | // TODO: Adapt simplifyDivRemOfSelectWithZeroOp to allow this and other folds. |
| 1348 | if (match(V: Op0, P: m_ImmConstant()) && |
| 1349 | match(V: Op1, P: m_Select(C: m_Value(), L: m_ImmConstant(), R: m_ImmConstant()))) { |
| 1350 | if (Instruction *R = FoldOpIntoSelect(Op&: I, SI: cast<SelectInst>(Val: Op1), |
| 1351 | /*FoldWithMultiUse*/ true)) |
| 1352 | return R; |
| 1353 | } |
| 1354 | |
| 1355 | return nullptr; |
| 1356 | } |
| 1357 | |
| 1358 | /// This function implements the transforms common to both integer division |
| 1359 | /// instructions (udiv and sdiv). It is called by the visitors to those integer |
| 1360 | /// division instructions. |
| 1361 | /// Common integer divide transforms |
| 1362 | Instruction *InstCombinerImpl::commonIDivTransforms(BinaryOperator &I) { |
| 1363 | if (Instruction *Res = commonIDivRemTransforms(I)) |
| 1364 | return Res; |
| 1365 | |
| 1366 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 1367 | bool IsSigned = I.getOpcode() == Instruction::SDiv; |
| 1368 | Type *Ty = I.getType(); |
| 1369 | |
| 1370 | const APInt *C2; |
| 1371 | if (match(V: Op1, P: m_APInt(Res&: C2))) { |
| 1372 | Value *X; |
| 1373 | const APInt *C1; |
| 1374 | |
| 1375 | // (X / C1) / C2 -> X / (C1*C2) |
| 1376 | if ((IsSigned && match(V: Op0, P: m_SDiv(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) || |
| 1377 | (!IsSigned && match(V: Op0, P: m_UDiv(L: m_Value(V&: X), R: m_APInt(Res&: C1))))) { |
| 1378 | APInt Product(C1->getBitWidth(), /*val=*/0ULL, IsSigned); |
| 1379 | if (!multiplyOverflows(C1: *C1, C2: *C2, Product, IsSigned)) |
| 1380 | return BinaryOperator::Create(Op: I.getOpcode(), S1: X, |
| 1381 | S2: ConstantInt::get(Ty, V: Product)); |
| 1382 | } |
| 1383 | |
| 1384 | APInt Quotient(C2->getBitWidth(), /*val=*/0ULL, IsSigned); |
| 1385 | if ((IsSigned && match(V: Op0, P: m_NSWMul(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) || |
| 1386 | (!IsSigned && match(V: Op0, P: m_NUWMul(L: m_Value(V&: X), R: m_APInt(Res&: C1))))) { |
| 1387 | |
| 1388 | // (X * C1) / C2 -> X / (C2 / C1) if C2 is a multiple of C1. |
| 1389 | if (isMultiple(C1: *C2, C2: *C1, Quotient, IsSigned)) { |
| 1390 | auto *NewDiv = BinaryOperator::Create(Op: I.getOpcode(), S1: X, |
| 1391 | S2: ConstantInt::get(Ty, V: Quotient)); |
| 1392 | NewDiv->setIsExact(I.isExact()); |
| 1393 | return NewDiv; |
| 1394 | } |
| 1395 | |
| 1396 | // (X * C1) / C2 -> X * (C1 / C2) if C1 is a multiple of C2. |
| 1397 | if (isMultiple(C1: *C1, C2: *C2, Quotient, IsSigned)) { |
| 1398 | auto *Mul = BinaryOperator::Create(Op: Instruction::Mul, S1: X, |
| 1399 | S2: ConstantInt::get(Ty, V: Quotient)); |
| 1400 | auto *OBO = cast<OverflowingBinaryOperator>(Val: Op0); |
| 1401 | Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap()); |
| 1402 | Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap()); |
| 1403 | return Mul; |
| 1404 | } |
| 1405 | } |
| 1406 | |
| 1407 | if ((IsSigned && match(V: Op0, P: m_NSWShl(L: m_Value(V&: X), R: m_APInt(Res&: C1))) && |
| 1408 | C1->ult(RHS: C1->getBitWidth() - 1)) || |
| 1409 | (!IsSigned && match(V: Op0, P: m_NUWShl(L: m_Value(V&: X), R: m_APInt(Res&: C1))) && |
| 1410 | C1->ult(RHS: C1->getBitWidth()))) { |
| 1411 | APInt C1Shifted = APInt::getOneBitSet( |
| 1412 | numBits: C1->getBitWidth(), BitNo: static_cast<unsigned>(C1->getZExtValue())); |
| 1413 | |
| 1414 | // (X << C1) / C2 -> X / (C2 >> C1) if C2 is a multiple of 1 << C1. |
| 1415 | if (isMultiple(C1: *C2, C2: C1Shifted, Quotient, IsSigned)) { |
| 1416 | auto *BO = BinaryOperator::Create(Op: I.getOpcode(), S1: X, |
| 1417 | S2: ConstantInt::get(Ty, V: Quotient)); |
| 1418 | BO->setIsExact(I.isExact()); |
| 1419 | return BO; |
| 1420 | } |
| 1421 | |
| 1422 | // (X << C1) / C2 -> X * ((1 << C1) / C2) if 1 << C1 is a multiple of C2. |
| 1423 | if (isMultiple(C1: C1Shifted, C2: *C2, Quotient, IsSigned)) { |
| 1424 | auto *Mul = BinaryOperator::Create(Op: Instruction::Mul, S1: X, |
| 1425 | S2: ConstantInt::get(Ty, V: Quotient)); |
| 1426 | auto *OBO = cast<OverflowingBinaryOperator>(Val: Op0); |
| 1427 | Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap()); |
| 1428 | Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap()); |
| 1429 | return Mul; |
| 1430 | } |
| 1431 | } |
| 1432 | |
| 1433 | // Distribute div over add to eliminate a matching div/mul pair: |
| 1434 | // ((X * C2) + C1) / C2 --> X + C1/C2 |
| 1435 | // We need a multiple of the divisor for a signed add constant, but |
| 1436 | // unsigned is fine with any constant pair. |
| 1437 | if (IsSigned && |
| 1438 | match(V: Op0, P: m_NSWAddLike(L: m_NSWMul(L: m_Value(V&: X), R: m_SpecificInt(V: *C2)), |
| 1439 | R: m_APInt(Res&: C1))) && |
| 1440 | isMultiple(C1: *C1, C2: *C2, Quotient, IsSigned)) { |
| 1441 | return BinaryOperator::CreateNSWAdd(V1: X, V2: ConstantInt::get(Ty, V: Quotient)); |
| 1442 | } |
| 1443 | if (!IsSigned && |
| 1444 | match(V: Op0, P: m_NUWAddLike(L: m_NUWMul(L: m_Value(V&: X), R: m_SpecificInt(V: *C2)), |
| 1445 | R: m_APInt(Res&: C1)))) { |
| 1446 | return BinaryOperator::CreateNUWAdd(V1: X, |
| 1447 | V2: ConstantInt::get(Ty, V: C1->udiv(RHS: *C2))); |
| 1448 | } |
| 1449 | |
| 1450 | if (!C2->isZero()) // avoid X udiv 0 |
| 1451 | if (Instruction *FoldedDiv = foldBinOpIntoSelectOrPhi(I)) |
| 1452 | return FoldedDiv; |
| 1453 | } |
| 1454 | |
| 1455 | if (match(V: Op0, P: m_One())) { |
| 1456 | assert(!Ty->isIntOrIntVectorTy(1) && "i1 divide not removed?" ); |
| 1457 | if (IsSigned) { |
| 1458 | // 1 / 0 --> undef ; 1 / 1 --> 1 ; 1 / -1 --> -1 ; 1 / anything else --> 0 |
| 1459 | // (Op1 + 1) u< 3 ? Op1 : 0 |
| 1460 | // Op1 must be frozen because we are increasing its number of uses. |
| 1461 | Value *F1 = Op1; |
| 1462 | if (!isGuaranteedNotToBeUndef(V: Op1)) |
| 1463 | F1 = Builder.CreateFreeze(V: Op1, Name: Op1->getName() + ".fr" ); |
| 1464 | Value *Inc = Builder.CreateAdd(LHS: F1, RHS: Op0); |
| 1465 | Value *Cmp = Builder.CreateICmpULT(LHS: Inc, RHS: ConstantInt::get(Ty, V: 3)); |
| 1466 | return createSelectInstWithUnknownProfile(C: Cmp, S1: F1, |
| 1467 | S2: ConstantInt::get(Ty, V: 0)); |
| 1468 | } else { |
| 1469 | // If Op1 is 0 then it's undefined behaviour. If Op1 is 1 then the |
| 1470 | // result is one, otherwise it's zero. |
| 1471 | return new ZExtInst(Builder.CreateICmpEQ(LHS: Op1, RHS: Op0), Ty); |
| 1472 | } |
| 1473 | } |
| 1474 | |
| 1475 | // See if we can fold away this div instruction. |
| 1476 | if (SimplifyDemandedInstructionBits(Inst&: I)) |
| 1477 | return &I; |
| 1478 | |
| 1479 | // (X - (X rem Y)) / Y -> X / Y; usually originates as ((X / Y) * Y) / Y |
| 1480 | Value *X, *Z; |
| 1481 | if (match(V: Op0, P: m_Sub(L: m_Value(V&: X), R: m_Value(V&: Z)))) // (X - Z) / Y; Y = Op1 |
| 1482 | if ((IsSigned && match(V: Z, P: m_SRem(L: m_Specific(V: X), R: m_Specific(V: Op1)))) || |
| 1483 | (!IsSigned && match(V: Z, P: m_URem(L: m_Specific(V: X), R: m_Specific(V: Op1))))) |
| 1484 | return BinaryOperator::Create(Op: I.getOpcode(), S1: X, S2: Op1); |
| 1485 | |
| 1486 | // (X << Y) / X -> 1 << Y |
| 1487 | Value *Y; |
| 1488 | if (IsSigned && match(V: Op0, P: m_NSWShl(L: m_Specific(V: Op1), R: m_Value(V&: Y)))) |
| 1489 | return BinaryOperator::CreateNSWShl(V1: ConstantInt::get(Ty, V: 1), V2: Y); |
| 1490 | if (!IsSigned && match(V: Op0, P: m_NUWShl(L: m_Specific(V: Op1), R: m_Value(V&: Y)))) |
| 1491 | return BinaryOperator::CreateNUWShl(V1: ConstantInt::get(Ty, V: 1), V2: Y); |
| 1492 | |
| 1493 | // X / (X * Y) -> 1 / Y if the multiplication does not overflow. |
| 1494 | if (match(V: Op1, P: m_c_Mul(L: m_Specific(V: Op0), R: m_Value(V&: Y)))) { |
| 1495 | bool HasNSW = cast<OverflowingBinaryOperator>(Val: Op1)->hasNoSignedWrap(); |
| 1496 | bool HasNUW = cast<OverflowingBinaryOperator>(Val: Op1)->hasNoUnsignedWrap(); |
| 1497 | if ((IsSigned && HasNSW) || (!IsSigned && HasNUW)) { |
| 1498 | replaceOperand(I, OpNum: 0, V: ConstantInt::get(Ty, V: 1)); |
| 1499 | replaceOperand(I, OpNum: 1, V: Y); |
| 1500 | return &I; |
| 1501 | } |
| 1502 | } |
| 1503 | |
| 1504 | // (X << Z) / (X * Y) -> (1 << Z) / Y |
| 1505 | // TODO: Handle sdiv. |
| 1506 | if (!IsSigned && Op1->hasOneUse() && |
| 1507 | match(V: Op0, P: m_NUWShl(L: m_Value(V&: X), R: m_Value(V&: Z))) && |
| 1508 | match(V: Op1, P: m_c_Mul(L: m_Specific(V: X), R: m_Value(V&: Y)))) |
| 1509 | if (cast<OverflowingBinaryOperator>(Val: Op1)->hasNoUnsignedWrap()) { |
| 1510 | Instruction *NewDiv = BinaryOperator::CreateUDiv( |
| 1511 | V1: Builder.CreateShl(LHS: ConstantInt::get(Ty, V: 1), RHS: Z, Name: "" , /*NUW*/ HasNUW: true), V2: Y); |
| 1512 | NewDiv->setIsExact(I.isExact()); |
| 1513 | return NewDiv; |
| 1514 | } |
| 1515 | |
| 1516 | if (Value *R = foldIDivShl(I, Builder)) |
| 1517 | return replaceInstUsesWith(I, V: R); |
| 1518 | |
| 1519 | // With the appropriate no-wrap constraint, remove a multiply by the divisor |
| 1520 | // after peeking through another divide: |
| 1521 | // ((Op1 * X) / Y) / Op1 --> X / Y |
| 1522 | if (match(V: Op0, P: m_BinOp(Opcode: I.getOpcode(), L: m_c_Mul(L: m_Specific(V: Op1), R: m_Value(V&: X)), |
| 1523 | R: m_Value(V&: Y)))) { |
| 1524 | auto *InnerDiv = cast<PossiblyExactOperator>(Val: Op0); |
| 1525 | auto *Mul = cast<OverflowingBinaryOperator>(Val: InnerDiv->getOperand(i_nocapture: 0)); |
| 1526 | Instruction *NewDiv = nullptr; |
| 1527 | if (!IsSigned && Mul->hasNoUnsignedWrap()) |
| 1528 | NewDiv = BinaryOperator::CreateUDiv(V1: X, V2: Y); |
| 1529 | else if (IsSigned && Mul->hasNoSignedWrap()) |
| 1530 | NewDiv = BinaryOperator::CreateSDiv(V1: X, V2: Y); |
| 1531 | |
| 1532 | // Exact propagates only if both of the original divides are exact. |
| 1533 | if (NewDiv) { |
| 1534 | NewDiv->setIsExact(I.isExact() && InnerDiv->isExact()); |
| 1535 | return NewDiv; |
| 1536 | } |
| 1537 | } |
| 1538 | |
| 1539 | // X / (select Cond, 1, Y) --> select Cond, X, (X / Y) |
| 1540 | // X / (select Cond, Y, 1) --> select Cond, (X / Y), X |
| 1541 | // Division by 1 is a no-op, so we sink the division into the non-1 arm. |
| 1542 | // For sdiv, limit Y to constant to avoid signed overflow concern. |
| 1543 | { |
| 1544 | Value *Cond, *DivY; |
| 1545 | const APInt *C; |
| 1546 | auto IsSafeDivisor = [&](Value *V) { |
| 1547 | if (IsSigned) |
| 1548 | return match(V, P: m_APInt(Res&: C)) && !C->isZero() && !C->isAllOnes(); |
| 1549 | return isKnownNonZero(V, Q: SQ.getWithInstruction(I: &I)) && |
| 1550 | isGuaranteedNotToBePoison(V, AC: SQ.AC, CtxI: &I, DT: SQ.DT); |
| 1551 | }; |
| 1552 | if (match(V: Op1, P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_One(), R: m_Value(V&: DivY)))) && |
| 1553 | IsSafeDivisor(DivY)) { |
| 1554 | Value *NewDiv = |
| 1555 | Builder.CreateExactBinOp(Opc: I.getOpcode(), LHS: Op0, RHS: DivY, IsExact: I.isExact()); |
| 1556 | return SelectInst::Create(C: Cond, S1: Op0, S2: NewDiv, NameStr: "" , InsertBefore: nullptr, |
| 1557 | MDFrom: cast<SelectInst>(Val: Op1)); |
| 1558 | } |
| 1559 | if (match(V: Op1, P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: DivY), R: m_One()))) && |
| 1560 | IsSafeDivisor(DivY)) { |
| 1561 | Value *NewDiv = |
| 1562 | Builder.CreateExactBinOp(Opc: I.getOpcode(), LHS: Op0, RHS: DivY, IsExact: I.isExact()); |
| 1563 | return SelectInst::Create(C: Cond, S1: NewDiv, S2: Op0, NameStr: "" , InsertBefore: nullptr, |
| 1564 | MDFrom: cast<SelectInst>(Val: Op1)); |
| 1565 | } |
| 1566 | } |
| 1567 | |
| 1568 | // (X * Y) / (X * Z) --> Y / Z (and commuted variants) |
| 1569 | if (match(V: Op0, P: m_Mul(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
| 1570 | auto OB0HasNSW = cast<OverflowingBinaryOperator>(Val: Op0)->hasNoSignedWrap(); |
| 1571 | auto OB0HasNUW = cast<OverflowingBinaryOperator>(Val: Op0)->hasNoUnsignedWrap(); |
| 1572 | |
| 1573 | auto CreateDivOrNull = [&](Value *A, Value *B) -> Instruction * { |
| 1574 | auto OB1HasNSW = cast<OverflowingBinaryOperator>(Val: Op1)->hasNoSignedWrap(); |
| 1575 | auto OB1HasNUW = |
| 1576 | cast<OverflowingBinaryOperator>(Val: Op1)->hasNoUnsignedWrap(); |
| 1577 | const APInt *C1, *C2; |
| 1578 | if (IsSigned && OB0HasNSW) { |
| 1579 | if (OB1HasNSW && match(V: B, P: m_APInt(Res&: C1)) && !C1->isAllOnes()) |
| 1580 | return BinaryOperator::CreateSDiv(V1: A, V2: B); |
| 1581 | } |
| 1582 | if (!IsSigned && OB0HasNUW) { |
| 1583 | if (OB1HasNUW) |
| 1584 | return BinaryOperator::CreateUDiv(V1: A, V2: B); |
| 1585 | if (match(V: A, P: m_APInt(Res&: C1)) && match(V: B, P: m_APInt(Res&: C2)) && C2->ule(RHS: *C1)) |
| 1586 | return BinaryOperator::CreateUDiv(V1: A, V2: B); |
| 1587 | } |
| 1588 | return nullptr; |
| 1589 | }; |
| 1590 | |
| 1591 | if (match(V: Op1, P: m_c_Mul(L: m_Specific(V: X), R: m_Value(V&: Z)))) { |
| 1592 | if (auto *Val = CreateDivOrNull(Y, Z)) |
| 1593 | return Val; |
| 1594 | } |
| 1595 | if (match(V: Op1, P: m_c_Mul(L: m_Specific(V: Y), R: m_Value(V&: Z)))) { |
| 1596 | if (auto *Val = CreateDivOrNull(X, Z)) |
| 1597 | return Val; |
| 1598 | } |
| 1599 | } |
| 1600 | return nullptr; |
| 1601 | } |
| 1602 | |
| 1603 | Value *InstCombinerImpl::takeLog2(Value *Op, unsigned Depth, bool AssumeNonZero, |
| 1604 | bool DoFold) { |
| 1605 | auto IfFold = [DoFold](function_ref<Value *()> Fn) { |
| 1606 | if (!DoFold) |
| 1607 | return reinterpret_cast<Value *>(-1); |
| 1608 | return Fn(); |
| 1609 | }; |
| 1610 | |
| 1611 | // FIXME: assert that Op1 isn't/doesn't contain undef. |
| 1612 | |
| 1613 | // log2(2^C) -> C |
| 1614 | if (match(V: Op, P: m_Power2())) |
| 1615 | return IfFold([&]() { |
| 1616 | Constant *C = ConstantExpr::getExactLogBase2(C: cast<Constant>(Val: Op)); |
| 1617 | if (!C) |
| 1618 | llvm_unreachable("Failed to constant fold udiv -> logbase2" ); |
| 1619 | return C; |
| 1620 | }); |
| 1621 | |
| 1622 | // The remaining tests are all recursive, so bail out if we hit the limit. |
| 1623 | if (Depth++ == MaxAnalysisRecursionDepth) |
| 1624 | return nullptr; |
| 1625 | |
| 1626 | // log2(zext X) -> zext log2(X) |
| 1627 | // FIXME: Require one use? |
| 1628 | Value *X, *Y; |
| 1629 | if (match(V: Op, P: m_ZExt(Op: m_Value(V&: X)))) |
| 1630 | if (Value *LogX = takeLog2(Op: X, Depth, AssumeNonZero, DoFold)) |
| 1631 | return IfFold([&]() { return Builder.CreateZExt(V: LogX, DestTy: Op->getType()); }); |
| 1632 | |
| 1633 | // log2(trunc x) -> trunc log2(X) |
| 1634 | // FIXME: Require one use? |
| 1635 | if (match(V: Op, P: m_Trunc(Op: m_Value(V&: X)))) { |
| 1636 | auto *TI = cast<TruncInst>(Val: Op); |
| 1637 | if (AssumeNonZero || TI->hasNoUnsignedWrap()) |
| 1638 | if (Value *LogX = takeLog2(Op: X, Depth, AssumeNonZero, DoFold)) |
| 1639 | return IfFold([&]() { |
| 1640 | return Builder.CreateTrunc(V: LogX, DestTy: Op->getType(), Name: "" , |
| 1641 | /*IsNUW=*/TI->hasNoUnsignedWrap()); |
| 1642 | }); |
| 1643 | } |
| 1644 | |
| 1645 | // log2(X << Y) -> log2(X) + Y |
| 1646 | // FIXME: Require one use unless X is 1? |
| 1647 | if (match(V: Op, P: m_Shl(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
| 1648 | auto *BO = cast<OverflowingBinaryOperator>(Val: Op); |
| 1649 | // nuw will be set if the `shl` is trivially non-zero. |
| 1650 | if (AssumeNonZero || BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap()) |
| 1651 | if (Value *LogX = takeLog2(Op: X, Depth, AssumeNonZero, DoFold)) |
| 1652 | return IfFold([&]() { return Builder.CreateAdd(LHS: LogX, RHS: Y); }); |
| 1653 | } |
| 1654 | |
| 1655 | // log2(X >>u Y) -> log2(X) - Y |
| 1656 | // FIXME: Require one use? |
| 1657 | if (match(V: Op, P: m_LShr(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
| 1658 | auto *PEO = cast<PossiblyExactOperator>(Val: Op); |
| 1659 | if (AssumeNonZero || PEO->isExact()) |
| 1660 | if (Value *LogX = takeLog2(Op: X, Depth, AssumeNonZero, DoFold)) |
| 1661 | return IfFold([&]() { return Builder.CreateSub(LHS: LogX, RHS: Y); }); |
| 1662 | } |
| 1663 | |
| 1664 | // log2(X & Y) -> either log2(X) or log2(Y) |
| 1665 | // This requires `AssumeNonZero` as `X & Y` may be zero when X != Y. |
| 1666 | if (AssumeNonZero && match(V: Op, P: m_And(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
| 1667 | if (Value *LogX = takeLog2(Op: X, Depth, AssumeNonZero, DoFold)) |
| 1668 | return IfFold([&]() { return LogX; }); |
| 1669 | if (Value *LogY = takeLog2(Op: Y, Depth, AssumeNonZero, DoFold)) |
| 1670 | return IfFold([&]() { return LogY; }); |
| 1671 | } |
| 1672 | |
| 1673 | // log2(Cond ? X : Y) -> Cond ? log2(X) : log2(Y) |
| 1674 | // FIXME: Require one use? |
| 1675 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op)) |
| 1676 | if (Value *LogX = takeLog2(Op: SI->getOperand(i_nocapture: 1), Depth, AssumeNonZero, DoFold)) |
| 1677 | if (Value *LogY = |
| 1678 | takeLog2(Op: SI->getOperand(i_nocapture: 2), Depth, AssumeNonZero, DoFold)) |
| 1679 | return IfFold([&]() { |
| 1680 | return Builder.CreateSelect(C: SI->getOperand(i_nocapture: 0), True: LogX, False: LogY, Name: "" , |
| 1681 | MDFrom: ProfcheckDisableMetadataFixes ? nullptr |
| 1682 | : SI); |
| 1683 | }); |
| 1684 | |
| 1685 | // log2(umin(X, Y)) -> umin(log2(X), log2(Y)) |
| 1686 | // log2(umax(X, Y)) -> umax(log2(X), log2(Y)) |
| 1687 | auto *MinMax = dyn_cast<MinMaxIntrinsic>(Val: Op); |
| 1688 | if (MinMax && MinMax->hasOneUse() && !MinMax->isSigned()) { |
| 1689 | // Use AssumeNonZero as false here. Otherwise we can hit case where |
| 1690 | // log2(umax(X, Y)) != umax(log2(X), log2(Y)) (because overflow). |
| 1691 | if (Value *LogX = takeLog2(Op: MinMax->getLHS(), Depth, |
| 1692 | /*AssumeNonZero*/ false, DoFold)) |
| 1693 | if (Value *LogY = takeLog2(Op: MinMax->getRHS(), Depth, |
| 1694 | /*AssumeNonZero*/ false, DoFold)) |
| 1695 | return IfFold([&]() { |
| 1696 | return Builder.CreateBinaryIntrinsic(ID: MinMax->getIntrinsicID(), LHS: LogX, |
| 1697 | RHS: LogY); |
| 1698 | }); |
| 1699 | } |
| 1700 | |
| 1701 | return nullptr; |
| 1702 | } |
| 1703 | |
| 1704 | /// If we have zero-extended operands of an unsigned div or rem, we may be able |
| 1705 | /// to narrow the operation (sink the zext below the math). |
| 1706 | static Instruction *narrowUDivURem(BinaryOperator &I, |
| 1707 | InstCombinerImpl &IC) { |
| 1708 | Instruction::BinaryOps Opcode = I.getOpcode(); |
| 1709 | Value *N = I.getOperand(i_nocapture: 0); |
| 1710 | Value *D = I.getOperand(i_nocapture: 1); |
| 1711 | Type *Ty = I.getType(); |
| 1712 | Value *X, *Y; |
| 1713 | if (match(V: N, P: m_ZExt(Op: m_Value(V&: X))) && match(V: D, P: m_ZExt(Op: m_Value(V&: Y))) && |
| 1714 | X->getType() == Y->getType() && (N->hasOneUse() || D->hasOneUse())) { |
| 1715 | // udiv (zext X), (zext Y) --> zext (udiv X, Y) |
| 1716 | // urem (zext X), (zext Y) --> zext (urem X, Y) |
| 1717 | Value *NarrowOp = IC.Builder.CreateBinOp(Opc: Opcode, LHS: X, RHS: Y); |
| 1718 | return new ZExtInst(NarrowOp, Ty); |
| 1719 | } |
| 1720 | |
| 1721 | Constant *C; |
| 1722 | auto &DL = IC.getDataLayout(); |
| 1723 | if (isa<Instruction>(Val: N) && match(V: N, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X)))) && |
| 1724 | match(V: D, P: m_Constant(C))) { |
| 1725 | // If the constant is the same in the smaller type, use the narrow version. |
| 1726 | Constant *TruncC = getLosslessUnsignedTrunc(C, DestTy: X->getType(), DL); |
| 1727 | if (!TruncC) |
| 1728 | return nullptr; |
| 1729 | |
| 1730 | // udiv (zext X), C --> zext (udiv X, C') |
| 1731 | // urem (zext X), C --> zext (urem X, C') |
| 1732 | return new ZExtInst(IC.Builder.CreateBinOp(Opc: Opcode, LHS: X, RHS: TruncC), Ty); |
| 1733 | } |
| 1734 | if (isa<Instruction>(Val: D) && match(V: D, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X)))) && |
| 1735 | match(V: N, P: m_Constant(C))) { |
| 1736 | // If the constant is the same in the smaller type, use the narrow version. |
| 1737 | Constant *TruncC = getLosslessUnsignedTrunc(C, DestTy: X->getType(), DL); |
| 1738 | if (!TruncC) |
| 1739 | return nullptr; |
| 1740 | |
| 1741 | // udiv C, (zext X) --> zext (udiv C', X) |
| 1742 | // urem C, (zext X) --> zext (urem C', X) |
| 1743 | return new ZExtInst(IC.Builder.CreateBinOp(Opc: Opcode, LHS: TruncC, RHS: X), Ty); |
| 1744 | } |
| 1745 | |
| 1746 | return nullptr; |
| 1747 | } |
| 1748 | |
| 1749 | Instruction *InstCombinerImpl::visitUDiv(BinaryOperator &I) { |
| 1750 | if (Value *V = simplifyUDivInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), IsExact: I.isExact(), |
| 1751 | Q: SQ.getWithInstruction(I: &I))) |
| 1752 | return replaceInstUsesWith(I, V); |
| 1753 | |
| 1754 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
| 1755 | return X; |
| 1756 | |
| 1757 | // Handle the integer div common cases |
| 1758 | if (Instruction *Common = commonIDivTransforms(I)) |
| 1759 | return Common; |
| 1760 | |
| 1761 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 1762 | Value *X; |
| 1763 | const APInt *C1, *C2; |
| 1764 | if (match(V: Op0, P: m_LShr(L: m_Value(V&: X), R: m_APInt(Res&: C1))) && match(V: Op1, P: m_APInt(Res&: C2))) { |
| 1765 | // (X lshr C1) udiv C2 --> X udiv (C2 << C1) |
| 1766 | bool Overflow; |
| 1767 | APInt C2ShlC1 = C2->ushl_ov(Amt: *C1, Overflow); |
| 1768 | if (!Overflow) { |
| 1769 | bool IsExact = I.isExact() && match(V: Op0, P: m_Exact(SubPattern: m_Value())); |
| 1770 | BinaryOperator *BO = BinaryOperator::CreateUDiv( |
| 1771 | V1: X, V2: ConstantInt::get(Ty: X->getType(), V: C2ShlC1)); |
| 1772 | if (IsExact) |
| 1773 | BO->setIsExact(); |
| 1774 | return BO; |
| 1775 | } |
| 1776 | } |
| 1777 | |
| 1778 | // Op0 / C where C is large (negative) --> zext (Op0 >= C) |
| 1779 | // TODO: Could use isKnownNegative() to handle non-constant values. |
| 1780 | Type *Ty = I.getType(); |
| 1781 | if (match(V: Op1, P: m_Negative())) { |
| 1782 | Value *Cmp = Builder.CreateICmpUGE(LHS: Op0, RHS: Op1); |
| 1783 | return CastInst::CreateZExtOrBitCast(S: Cmp, Ty); |
| 1784 | } |
| 1785 | // Op0 / (sext i1 X) --> zext (Op0 == -1) (if X is 0, the div is undefined) |
| 1786 | if (match(V: Op1, P: m_SExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
| 1787 | Value *Cmp = Builder.CreateICmpEQ(LHS: Op0, RHS: ConstantInt::getAllOnesValue(Ty)); |
| 1788 | return CastInst::CreateZExtOrBitCast(S: Cmp, Ty); |
| 1789 | } |
| 1790 | |
| 1791 | if (Instruction *NarrowDiv = narrowUDivURem(I, IC&: *this)) |
| 1792 | return NarrowDiv; |
| 1793 | |
| 1794 | Value *A, *B; |
| 1795 | |
| 1796 | // Look through a right-shift to find the common factor: |
| 1797 | // ((Op1 *nuw A) >> B) / Op1 --> A >> B |
| 1798 | if (match(V: Op0, P: m_LShr(L: m_NUWMul(L: m_Specific(V: Op1), R: m_Value(V&: A)), R: m_Value(V&: B))) || |
| 1799 | match(V: Op0, P: m_LShr(L: m_NUWMul(L: m_Value(V&: A), R: m_Specific(V: Op1)), R: m_Value(V&: B)))) { |
| 1800 | Instruction *Lshr = BinaryOperator::CreateLShr(V1: A, V2: B); |
| 1801 | if (I.isExact() && cast<PossiblyExactOperator>(Val: Op0)->isExact()) |
| 1802 | Lshr->setIsExact(); |
| 1803 | return Lshr; |
| 1804 | } |
| 1805 | |
| 1806 | auto GetShiftableDenom = [&](Value *Denom) -> Value * { |
| 1807 | // Op0 udiv Op1 -> Op0 lshr log2(Op1), if log2() folds away. |
| 1808 | if (Value *Log2 = tryGetLog2(Op: Op1, /*AssumeNonZero=*/true)) |
| 1809 | return Log2; |
| 1810 | |
| 1811 | // Op0 udiv Op1 -> Op0 lshr cttz(Op1), if Op1 is a power of 2. |
| 1812 | if (isKnownToBeAPowerOfTwo(V: Denom, /*OrZero=*/true, CxtI: &I)) |
| 1813 | // This will increase instruction count but it's okay |
| 1814 | // since bitwise operations are substantially faster than |
| 1815 | // division. |
| 1816 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::cttz, LHS: Denom, |
| 1817 | RHS: Builder.getTrue()); |
| 1818 | |
| 1819 | return nullptr; |
| 1820 | }; |
| 1821 | |
| 1822 | if (auto *Res = GetShiftableDenom(Op1)) |
| 1823 | return replaceInstUsesWith( |
| 1824 | I, V: Builder.CreateLShr(LHS: Op0, RHS: Res, Name: I.getName(), isExact: I.isExact())); |
| 1825 | |
| 1826 | return nullptr; |
| 1827 | } |
| 1828 | |
| 1829 | Instruction *InstCombinerImpl::visitSDiv(BinaryOperator &I) { |
| 1830 | if (Value *V = simplifySDivInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), IsExact: I.isExact(), |
| 1831 | Q: SQ.getWithInstruction(I: &I))) |
| 1832 | return replaceInstUsesWith(I, V); |
| 1833 | |
| 1834 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
| 1835 | return X; |
| 1836 | |
| 1837 | // Handle the integer div common cases |
| 1838 | if (Instruction *Common = commonIDivTransforms(I)) |
| 1839 | return Common; |
| 1840 | |
| 1841 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 1842 | Type *Ty = I.getType(); |
| 1843 | Value *X; |
| 1844 | // sdiv Op0, -1 --> -Op0 |
| 1845 | // sdiv Op0, (sext i1 X) --> -Op0 (because if X is 0, the op is undefined) |
| 1846 | if (match(V: Op1, P: m_AllOnes()) || |
| 1847 | (match(V: Op1, P: m_SExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1))) |
| 1848 | return BinaryOperator::CreateNSWNeg(Op: Op0); |
| 1849 | |
| 1850 | // X / INT_MIN --> X == INT_MIN |
| 1851 | if (match(V: Op1, P: m_SignMask())) |
| 1852 | return new ZExtInst(Builder.CreateICmpEQ(LHS: Op0, RHS: Op1), Ty); |
| 1853 | |
| 1854 | if (I.isExact()) { |
| 1855 | // sdiv exact X, 1<<C --> ashr exact X, C iff 1<<C is non-negative |
| 1856 | if (match(V: Op1, P: m_Power2()) && match(V: Op1, P: m_NonNegative())) { |
| 1857 | Constant *C = ConstantExpr::getExactLogBase2(C: cast<Constant>(Val: Op1)); |
| 1858 | return BinaryOperator::CreateExactAShr(V1: Op0, V2: C); |
| 1859 | } |
| 1860 | |
| 1861 | // sdiv exact X, (1<<ShAmt) --> ashr exact X, ShAmt (if shl is non-negative) |
| 1862 | Value *ShAmt; |
| 1863 | if (match(V: Op1, P: m_NSWShl(L: m_One(), R: m_Value(V&: ShAmt)))) |
| 1864 | return BinaryOperator::CreateExactAShr(V1: Op0, V2: ShAmt); |
| 1865 | |
| 1866 | // sdiv exact X, -1<<C --> -(ashr exact X, C) |
| 1867 | if (match(V: Op1, P: m_NegatedPower2())) { |
| 1868 | Constant *NegPow2C = ConstantExpr::getNeg(C: cast<Constant>(Val: Op1)); |
| 1869 | Constant *C = ConstantExpr::getExactLogBase2(C: NegPow2C); |
| 1870 | Value *Ashr = Builder.CreateAShr(LHS: Op0, RHS: C, Name: I.getName() + ".neg" , isExact: true); |
| 1871 | return BinaryOperator::CreateNSWNeg(Op: Ashr); |
| 1872 | } |
| 1873 | } |
| 1874 | |
| 1875 | const APInt *Op1C; |
| 1876 | if (match(V: Op1, P: m_APInt(Res&: Op1C))) { |
| 1877 | // If the dividend is sign-extended and the constant divisor is small enough |
| 1878 | // to fit in the source type, shrink the division to the narrower type: |
| 1879 | // (sext X) sdiv C --> sext (X sdiv C) |
| 1880 | Value *Op0Src; |
| 1881 | if (match(V: Op0, P: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: Op0Src)))) && |
| 1882 | Op0Src->getType()->getScalarSizeInBits() >= |
| 1883 | Op1C->getSignificantBits()) { |
| 1884 | |
| 1885 | // In the general case, we need to make sure that the dividend is not the |
| 1886 | // minimum signed value because dividing that by -1 is UB. But here, we |
| 1887 | // know that the -1 divisor case is already handled above. |
| 1888 | |
| 1889 | Constant *NarrowDivisor = |
| 1890 | ConstantExpr::getTrunc(C: cast<Constant>(Val: Op1), Ty: Op0Src->getType()); |
| 1891 | Value *NarrowOp = Builder.CreateSDiv(LHS: Op0Src, RHS: NarrowDivisor); |
| 1892 | return new SExtInst(NarrowOp, Ty); |
| 1893 | } |
| 1894 | |
| 1895 | // -X / C --> X / -C (if the negation doesn't overflow). |
| 1896 | // TODO: This could be enhanced to handle arbitrary vector constants by |
| 1897 | // checking if all elements are not the min-signed-val. |
| 1898 | if (!Op1C->isMinSignedValue() && match(V: Op0, P: m_NSWNeg(V: m_Value(V&: X)))) { |
| 1899 | Constant *NegC = ConstantInt::get(Ty, V: -(*Op1C)); |
| 1900 | Instruction *BO = BinaryOperator::CreateSDiv(V1: X, V2: NegC); |
| 1901 | BO->setIsExact(I.isExact()); |
| 1902 | return BO; |
| 1903 | } |
| 1904 | } |
| 1905 | |
| 1906 | // -X / Y --> -(X / Y) |
| 1907 | Value *Y; |
| 1908 | if (match(V: &I, P: m_SDiv(L: m_OneUse(SubPattern: m_NSWNeg(V: m_Value(V&: X))), R: m_Value(V&: Y)))) |
| 1909 | return BinaryOperator::CreateNSWNeg( |
| 1910 | Op: Builder.CreateSDiv(LHS: X, RHS: Y, Name: I.getName(), isExact: I.isExact())); |
| 1911 | |
| 1912 | // abs(X) / X --> X > -1 ? 1 : -1 |
| 1913 | // X / abs(X) --> X > -1 ? 1 : -1 |
| 1914 | if (match(V: &I, P: m_c_BinOp( |
| 1915 | L: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::abs>(Op0: m_Value(V&: X), Op1: m_One())), |
| 1916 | R: m_Deferred(V: X)))) { |
| 1917 | Value *Cond = Builder.CreateIsNotNeg(Arg: X); |
| 1918 | return createSelectInstWithUnknownProfile(C: Cond, S1: ConstantInt::get(Ty, V: 1), |
| 1919 | S2: ConstantInt::getAllOnesValue(Ty)); |
| 1920 | } |
| 1921 | |
| 1922 | KnownBits KnownDividend = computeKnownBits(V: Op0, CxtI: &I); |
| 1923 | if (!I.isExact() && |
| 1924 | (match(V: Op1, P: m_Power2(V&: Op1C)) || match(V: Op1, P: m_NegatedPower2(V&: Op1C))) && |
| 1925 | KnownDividend.countMinTrailingZeros() >= Op1C->countr_zero()) { |
| 1926 | I.setIsExact(); |
| 1927 | return &I; |
| 1928 | } |
| 1929 | |
| 1930 | if (KnownDividend.isNonNegative()) { |
| 1931 | // If both operands are unsigned, turn this into a udiv. |
| 1932 | if (isKnownNonNegative(V: Op1, SQ: SQ.getWithInstruction(I: &I))) { |
| 1933 | auto *BO = BinaryOperator::CreateUDiv(V1: Op0, V2: Op1, Name: I.getName()); |
| 1934 | BO->setIsExact(I.isExact()); |
| 1935 | return BO; |
| 1936 | } |
| 1937 | |
| 1938 | if (match(V: Op1, P: m_NegatedPower2())) { |
| 1939 | // X sdiv (-(1 << C)) -> -(X sdiv (1 << C)) -> |
| 1940 | // -> -(X udiv (1 << C)) -> -(X u>> C) |
| 1941 | Constant *CNegLog2 = ConstantExpr::getExactLogBase2( |
| 1942 | C: ConstantExpr::getNeg(C: cast<Constant>(Val: Op1))); |
| 1943 | Value *Shr = Builder.CreateLShr(LHS: Op0, RHS: CNegLog2, Name: I.getName(), isExact: I.isExact()); |
| 1944 | return BinaryOperator::CreateNeg(Op: Shr); |
| 1945 | } |
| 1946 | |
| 1947 | if (isKnownToBeAPowerOfTwo(V: Op1, /*OrZero*/ true, CxtI: &I)) { |
| 1948 | // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y) |
| 1949 | // Safe because the only negative value (1 << Y) can take on is |
| 1950 | // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have |
| 1951 | // the sign bit set. |
| 1952 | auto *BO = BinaryOperator::CreateUDiv(V1: Op0, V2: Op1, Name: I.getName()); |
| 1953 | BO->setIsExact(I.isExact()); |
| 1954 | return BO; |
| 1955 | } |
| 1956 | } |
| 1957 | |
| 1958 | // -X / X --> X == INT_MIN ? 1 : -1 |
| 1959 | if (isKnownNegation(X: Op0, Y: Op1)) { |
| 1960 | APInt MinVal = APInt::getSignedMinValue(numBits: Ty->getScalarSizeInBits()); |
| 1961 | Value *Cond = Builder.CreateICmpEQ(LHS: Op0, RHS: ConstantInt::get(Ty, V: MinVal)); |
| 1962 | return createSelectInstWithUnknownProfile(C: Cond, S1: ConstantInt::get(Ty, V: 1), |
| 1963 | S2: ConstantInt::getAllOnesValue(Ty)); |
| 1964 | } |
| 1965 | return nullptr; |
| 1966 | } |
| 1967 | |
| 1968 | /// Remove negation and try to convert division into multiplication. |
| 1969 | Instruction *InstCombinerImpl::foldFDivConstantDivisor(BinaryOperator &I) { |
| 1970 | Constant *C; |
| 1971 | if (!match(V: I.getOperand(i_nocapture: 1), P: m_Constant(C))) |
| 1972 | return nullptr; |
| 1973 | |
| 1974 | // -X / C --> X / -C |
| 1975 | Value *X; |
| 1976 | const DataLayout &DL = I.getDataLayout(); |
| 1977 | if (match(V: I.getOperand(i_nocapture: 0), P: m_FNeg(X: m_Value(V&: X)))) |
| 1978 | if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)) |
| 1979 | return BinaryOperator::CreateFDivFMF(V1: X, V2: NegC, FMFSource: &I); |
| 1980 | |
| 1981 | // nnan X / +0.0 -> copysign(inf, X) |
| 1982 | // nnan nsz X / -0.0 -> copysign(inf, X) |
| 1983 | if (I.hasNoNaNs() && |
| 1984 | (match(V: I.getOperand(i_nocapture: 1), P: m_PosZeroFP()) || |
| 1985 | (I.hasNoSignedZeros() && match(V: I.getOperand(i_nocapture: 1), P: m_AnyZeroFP())))) { |
| 1986 | IRBuilder<> B(&I); |
| 1987 | Value *CopySign = B.CreateIntrinsic( |
| 1988 | ID: Intrinsic::copysign, OverloadTypes: {C->getType()}, |
| 1989 | Args: {ConstantFP::getInfinity(Ty: I.getType()), I.getOperand(i_nocapture: 0)}, FMFSource: &I); |
| 1990 | CopySign->takeName(V: &I); |
| 1991 | return replaceInstUsesWith(I, V: CopySign); |
| 1992 | } |
| 1993 | |
| 1994 | // If the constant divisor has an exact inverse, this is always safe. If not, |
| 1995 | // then we can still create a reciprocal if fast-math-flags allow it and the |
| 1996 | // constant is a regular number (not zero, infinite, or denormal). |
| 1997 | if (!(C->hasExactInverseFP() || (I.hasAllowReciprocal() && C->isNormalFP()))) |
| 1998 | return nullptr; |
| 1999 | |
| 2000 | // Disallow denormal constants because we don't know what would happen |
| 2001 | // on all targets. |
| 2002 | // TODO: Use Intrinsic::canonicalize or let function attributes tell us that |
| 2003 | // denorms are flushed? |
| 2004 | auto *RecipC = ConstantFoldBinaryOpOperands( |
| 2005 | Opcode: Instruction::FDiv, LHS: ConstantFP::get(Ty: I.getType(), V: 1.0), RHS: C, DL); |
| 2006 | if (!RecipC || !RecipC->isNormalFP()) |
| 2007 | return nullptr; |
| 2008 | |
| 2009 | // X / C --> X * (1 / C) |
| 2010 | return BinaryOperator::CreateFMulFMF(V1: I.getOperand(i_nocapture: 0), V2: RecipC, FMFSource: &I); |
| 2011 | } |
| 2012 | |
| 2013 | /// Remove negation and try to reassociate constant math. |
| 2014 | static Instruction *foldFDivConstantDividend(BinaryOperator &I) { |
| 2015 | Constant *C; |
| 2016 | if (!match(V: I.getOperand(i_nocapture: 0), P: m_Constant(C))) |
| 2017 | return nullptr; |
| 2018 | |
| 2019 | // C / -X --> -C / X |
| 2020 | Value *X; |
| 2021 | const DataLayout &DL = I.getDataLayout(); |
| 2022 | if (match(V: I.getOperand(i_nocapture: 1), P: m_FNeg(X: m_Value(V&: X)))) |
| 2023 | if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)) |
| 2024 | return BinaryOperator::CreateFDivFMF(V1: NegC, V2: X, FMFSource: &I); |
| 2025 | |
| 2026 | if (!I.hasAllowReassoc() || !I.hasAllowReciprocal()) |
| 2027 | return nullptr; |
| 2028 | |
| 2029 | // Try to reassociate C / X expressions where X includes another constant. |
| 2030 | Constant *C2, *NewC = nullptr; |
| 2031 | if (match(V: I.getOperand(i_nocapture: 1), P: m_FMul(L: m_Value(V&: X), R: m_Constant(C&: C2)))) { |
| 2032 | // C / (X * C2) --> (C / C2) / X |
| 2033 | NewC = ConstantFoldBinaryOpOperands(Opcode: Instruction::FDiv, LHS: C, RHS: C2, DL); |
| 2034 | } else if (match(V: I.getOperand(i_nocapture: 1), P: m_FDiv(L: m_Value(V&: X), R: m_Constant(C&: C2)))) { |
| 2035 | // C / (X / C2) --> (C * C2) / X |
| 2036 | NewC = ConstantFoldBinaryOpOperands(Opcode: Instruction::FMul, LHS: C, RHS: C2, DL); |
| 2037 | } |
| 2038 | // Disallow denormal constants because we don't know what would happen |
| 2039 | // on all targets. |
| 2040 | // TODO: Use Intrinsic::canonicalize or let function attributes tell us that |
| 2041 | // denorms are flushed? |
| 2042 | if (!NewC || !NewC->isNormalFP()) |
| 2043 | return nullptr; |
| 2044 | |
| 2045 | return BinaryOperator::CreateFDivFMF(V1: NewC, V2: X, FMFSource: &I); |
| 2046 | } |
| 2047 | |
| 2048 | /// Negate the exponent of pow/exp to fold division-by-pow() into multiply. |
| 2049 | static Instruction *foldFDivPowDivisor(BinaryOperator &I, |
| 2050 | InstCombiner::BuilderTy &Builder) { |
| 2051 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 2052 | auto *II = dyn_cast<IntrinsicInst>(Val: Op1); |
| 2053 | if (!II || !II->hasOneUse() || !I.hasAllowReassoc() || |
| 2054 | !I.hasAllowReciprocal()) |
| 2055 | return nullptr; |
| 2056 | |
| 2057 | // Z / pow(X, Y) --> Z * pow(X, -Y) |
| 2058 | // Z / exp{2}(Y) --> Z * exp{2}(-Y) |
| 2059 | // In the general case, this creates an extra instruction, but fmul allows |
| 2060 | // for better canonicalization and optimization than fdiv. |
| 2061 | Intrinsic::ID IID = II->getIntrinsicID(); |
| 2062 | SmallVector<Value *> Args; |
| 2063 | switch (IID) { |
| 2064 | case Intrinsic::pow: |
| 2065 | Args.push_back(Elt: II->getArgOperand(i: 0)); |
| 2066 | Args.push_back(Elt: Builder.CreateFNegFMF(V: II->getArgOperand(i: 1), FMFSource: &I)); |
| 2067 | break; |
| 2068 | case Intrinsic::powi: { |
| 2069 | // Require 'ninf' assuming that makes powi(X, -INT_MIN) acceptable. |
| 2070 | // That is, X ** (huge negative number) is 0.0, ~1.0, or INF and so |
| 2071 | // dividing by that is INF, ~1.0, or 0.0. Code that uses powi allows |
| 2072 | // non-standard results, so this corner case should be acceptable if the |
| 2073 | // code rules out INF values. |
| 2074 | if (!I.hasNoInfs()) |
| 2075 | return nullptr; |
| 2076 | Args.push_back(Elt: II->getArgOperand(i: 0)); |
| 2077 | Args.push_back(Elt: Builder.CreateNeg(V: II->getArgOperand(i: 1))); |
| 2078 | Type *Tys[] = {I.getType(), II->getArgOperand(i: 1)->getType()}; |
| 2079 | Value *Pow = Builder.CreateIntrinsic(ID: IID, OverloadTypes: Tys, Args, FMFSource: &I); |
| 2080 | return BinaryOperator::CreateFMulFMF(V1: Op0, V2: Pow, FMFSource: &I); |
| 2081 | } |
| 2082 | case Intrinsic::exp: |
| 2083 | case Intrinsic::exp2: |
| 2084 | Args.push_back(Elt: Builder.CreateFNegFMF(V: II->getArgOperand(i: 0), FMFSource: &I)); |
| 2085 | break; |
| 2086 | default: |
| 2087 | return nullptr; |
| 2088 | } |
| 2089 | Value *Pow = Builder.CreateIntrinsic(ID: IID, OverloadTypes: I.getType(), Args, FMFSource: &I); |
| 2090 | return BinaryOperator::CreateFMulFMF(V1: Op0, V2: Pow, FMFSource: &I); |
| 2091 | } |
| 2092 | |
| 2093 | /// Convert div to mul if we have an sqrt divisor iff sqrt's operand is a fdiv |
| 2094 | /// instruction. |
| 2095 | static Instruction *foldFDivSqrtDivisor(BinaryOperator &I, |
| 2096 | InstCombiner::BuilderTy &Builder) { |
| 2097 | // X / sqrt(Y / Z) --> X * sqrt(Z / Y) |
| 2098 | if (!I.hasAllowReassoc() || !I.hasAllowReciprocal()) |
| 2099 | return nullptr; |
| 2100 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 2101 | auto *II = dyn_cast<IntrinsicInst>(Val: Op1); |
| 2102 | if (!II || II->getIntrinsicID() != Intrinsic::sqrt || !II->hasOneUse() || |
| 2103 | !II->hasAllowReassoc() || !II->hasAllowReciprocal()) |
| 2104 | return nullptr; |
| 2105 | |
| 2106 | Value *Y, *Z; |
| 2107 | auto *DivOp = dyn_cast<Instruction>(Val: II->getOperand(i_nocapture: 0)); |
| 2108 | if (!DivOp) |
| 2109 | return nullptr; |
| 2110 | if (!match(V: DivOp, P: m_FDiv(L: m_Value(V&: Y), R: m_Value(V&: Z)))) |
| 2111 | return nullptr; |
| 2112 | if (!DivOp->hasAllowReassoc() || !I.hasAllowReciprocal() || |
| 2113 | !DivOp->hasOneUse()) |
| 2114 | return nullptr; |
| 2115 | Value *SwapDiv = Builder.CreateFDivFMF(L: Z, R: Y, FMFSource: DivOp); |
| 2116 | Value *NewSqrt = |
| 2117 | Builder.CreateUnaryIntrinsic(ID: II->getIntrinsicID(), Op: SwapDiv, FMFSource: II); |
| 2118 | return BinaryOperator::CreateFMulFMF(V1: Op0, V2: NewSqrt, FMFSource: &I); |
| 2119 | } |
| 2120 | |
| 2121 | // Change |
| 2122 | // X = 1/sqrt(a) |
| 2123 | // R1 = X * X |
| 2124 | // R2 = a * X |
| 2125 | // |
| 2126 | // TO |
| 2127 | // |
| 2128 | // FDiv = 1/a |
| 2129 | // FSqrt = sqrt(a) |
| 2130 | // FMul = FDiv * FSqrt |
| 2131 | // Replace Uses Of R1 With FDiv |
| 2132 | // Replace Uses Of R2 With FSqrt |
| 2133 | // Replace Uses Of X With FMul |
| 2134 | static Instruction * |
| 2135 | convertFSqrtDivIntoFMul(CallInst *CI, Instruction *X, |
| 2136 | const SmallPtrSetImpl<Instruction *> &R1, |
| 2137 | const SmallPtrSetImpl<Instruction *> &R2, |
| 2138 | InstCombiner::BuilderTy &B, InstCombinerImpl *IC) { |
| 2139 | |
| 2140 | B.SetInsertPoint(X); |
| 2141 | |
| 2142 | // Have an instruction that is representative of all of instructions in R1 and |
| 2143 | // get the most common fpmath metadata and fast-math flags on it. |
| 2144 | Value *SqrtOp = CI->getArgOperand(i: 0); |
| 2145 | auto *FDiv = cast<Instruction>( |
| 2146 | Val: B.CreateFDiv(L: ConstantFP::get(Ty: X->getType(), V: 1.0), R: SqrtOp)); |
| 2147 | auto *R1FPMathMDNode = (*R1.begin())->getMetadata(KindID: LLVMContext::MD_fpmath); |
| 2148 | FastMathFlags R1FMF = (*R1.begin())->getFastMathFlags(); // Common FMF |
| 2149 | for (Instruction *I : R1) { |
| 2150 | R1FPMathMDNode = MDNode::getMostGenericFPMath( |
| 2151 | A: R1FPMathMDNode, B: I->getMetadata(KindID: LLVMContext::MD_fpmath)); |
| 2152 | R1FMF &= I->getFastMathFlags(); |
| 2153 | IC->replaceInstUsesWith(I&: *I, V: FDiv); |
| 2154 | IC->eraseInstFromFunction(I&: *I); |
| 2155 | } |
| 2156 | FDiv->setMetadata(KindID: LLVMContext::MD_fpmath, Node: R1FPMathMDNode); |
| 2157 | FDiv->copyFastMathFlags(FMF: R1FMF); |
| 2158 | |
| 2159 | // Have a single sqrt call instruction that is representative of all of |
| 2160 | // instructions in R2 and get the most common fpmath metadata and fast-math |
| 2161 | // flags on it. |
| 2162 | auto *FSqrt = cast<CallInst>(Val: CI->clone()); |
| 2163 | FSqrt->insertBefore(InsertPos: CI->getIterator()); |
| 2164 | auto *R2FPMathMDNode = (*R2.begin())->getMetadata(KindID: LLVMContext::MD_fpmath); |
| 2165 | FastMathFlags R2FMF = (*R2.begin())->getFastMathFlags(); // Common FMF |
| 2166 | for (Instruction *I : R2) { |
| 2167 | R2FPMathMDNode = MDNode::getMostGenericFPMath( |
| 2168 | A: R2FPMathMDNode, B: I->getMetadata(KindID: LLVMContext::MD_fpmath)); |
| 2169 | R2FMF &= I->getFastMathFlags(); |
| 2170 | IC->replaceInstUsesWith(I&: *I, V: FSqrt); |
| 2171 | IC->eraseInstFromFunction(I&: *I); |
| 2172 | } |
| 2173 | FSqrt->setMetadata(KindID: LLVMContext::MD_fpmath, Node: R2FPMathMDNode); |
| 2174 | FSqrt->copyFastMathFlags(FMF: R2FMF); |
| 2175 | |
| 2176 | Instruction *FMul; |
| 2177 | // If X = -1/sqrt(a) initially,then FMul = -(FDiv * FSqrt) |
| 2178 | if (match(V: X, P: m_FDiv(L: m_SpecificFP(V: -1.0), R: m_Specific(V: CI)))) { |
| 2179 | Value *Mul = B.CreateFMul(L: FDiv, R: FSqrt); |
| 2180 | FMul = cast<Instruction>(Val: B.CreateFNeg(V: Mul)); |
| 2181 | } else |
| 2182 | FMul = cast<Instruction>(Val: B.CreateFMul(L: FDiv, R: FSqrt)); |
| 2183 | FMul->copyMetadata(SrcInst: *X); |
| 2184 | FMul->copyFastMathFlags(FMF: FastMathFlags::intersectRewrite(LHS: R1FMF, RHS: R2FMF) | |
| 2185 | FastMathFlags::unionValue(LHS: R1FMF, RHS: R2FMF)); |
| 2186 | return IC->replaceInstUsesWith(I&: *X, V: FMul); |
| 2187 | } |
| 2188 | |
| 2189 | Instruction *InstCombinerImpl::visitFDiv(BinaryOperator &I) { |
| 2190 | Module *M = I.getModule(); |
| 2191 | |
| 2192 | if (Value *V = simplifyFDivInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), |
| 2193 | FMF: I.getFastMathFlags(), |
| 2194 | Q: SQ.getWithInstruction(I: &I))) |
| 2195 | return replaceInstUsesWith(I, V); |
| 2196 | |
| 2197 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
| 2198 | return X; |
| 2199 | |
| 2200 | if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I)) |
| 2201 | return Phi; |
| 2202 | |
| 2203 | if (Instruction *R = foldFDivConstantDivisor(I)) |
| 2204 | return R; |
| 2205 | |
| 2206 | if (Instruction *R = foldFDivConstantDividend(I)) |
| 2207 | return R; |
| 2208 | |
| 2209 | if (Instruction *R = foldFPSignBitOps(I)) |
| 2210 | return R; |
| 2211 | |
| 2212 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 2213 | |
| 2214 | // Convert |
| 2215 | // x = 1.0/sqrt(a) |
| 2216 | // r1 = x * x; |
| 2217 | // r2 = a/sqrt(a); |
| 2218 | // |
| 2219 | // TO |
| 2220 | // |
| 2221 | // r1 = 1/a |
| 2222 | // r2 = sqrt(a) |
| 2223 | // x = r1 * r2 |
| 2224 | SmallPtrSet<Instruction *, 2> R1, R2; |
| 2225 | if (isFSqrtDivToFMulLegal(X: &I, R1, R2)) { |
| 2226 | CallInst *CI = cast<CallInst>(Val: I.getOperand(i_nocapture: 1)); |
| 2227 | if (Instruction *D = convertFSqrtDivIntoFMul(CI, X: &I, R1, R2, B&: Builder, IC: this)) |
| 2228 | return D; |
| 2229 | } |
| 2230 | |
| 2231 | if (isa<Constant>(Val: Op0)) |
| 2232 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op1)) |
| 2233 | if (Instruction *R = FoldOpIntoSelect(Op&: I, SI)) |
| 2234 | return R; |
| 2235 | |
| 2236 | if (isa<Constant>(Val: Op1)) |
| 2237 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op0)) |
| 2238 | if (Instruction *R = FoldOpIntoSelect(Op&: I, SI)) |
| 2239 | return R; |
| 2240 | |
| 2241 | if (I.hasAllowReassoc() && I.hasAllowReciprocal()) { |
| 2242 | Value *X, *Y; |
| 2243 | if (match(V: Op0, P: m_OneUse(SubPattern: m_FDiv(L: m_Value(V&: X), R: m_Value(V&: Y)))) && |
| 2244 | (!isa<Constant>(Val: Y) || !isa<Constant>(Val: Op1))) { |
| 2245 | // (X / Y) / Z => X / (Y * Z) |
| 2246 | Value *YZ = Builder.CreateFMulFMF(L: Y, R: Op1, FMFSource: &I); |
| 2247 | return BinaryOperator::CreateFDivFMF(V1: X, V2: YZ, FMFSource: &I); |
| 2248 | } |
| 2249 | if (match(V: Op1, P: m_OneUse(SubPattern: m_FDiv(L: m_Value(V&: X), R: m_Value(V&: Y)))) && |
| 2250 | (!isa<Constant>(Val: Y) || !isa<Constant>(Val: Op0))) { |
| 2251 | // Z / (X / Y) => (Y * Z) / X |
| 2252 | Value *YZ = Builder.CreateFMulFMF(L: Y, R: Op0, FMFSource: &I); |
| 2253 | return BinaryOperator::CreateFDivFMF(V1: YZ, V2: X, FMFSource: &I); |
| 2254 | } |
| 2255 | // Z / (1.0 / Y) => (Y * Z) |
| 2256 | // |
| 2257 | // This is a special case of Z / (X / Y) => (Y * Z) / X, with X = 1.0. The |
| 2258 | // m_OneUse check is avoided because even in the case of the multiple uses |
| 2259 | // for 1.0/Y, the number of instructions remain the same and a division is |
| 2260 | // replaced by a multiplication. |
| 2261 | if (match(V: Op1, P: m_FDiv(L: m_SpecificFP(V: 1.0), R: m_Value(V&: Y)))) |
| 2262 | return BinaryOperator::CreateFMulFMF(V1: Y, V2: Op0, FMFSource: &I); |
| 2263 | } |
| 2264 | |
| 2265 | if (I.hasAllowReassoc() && Op0->hasOneUse() && Op1->hasOneUse()) { |
| 2266 | // sin(X) / cos(X) -> tan(X) |
| 2267 | // cos(X) / sin(X) -> 1/tan(X) (cotangent) |
| 2268 | Value *X; |
| 2269 | bool IsTan = match(V: Op0, P: m_Intrinsic<Intrinsic::sin>(Op0: m_Value(V&: X))) && |
| 2270 | match(V: Op1, P: m_Intrinsic<Intrinsic::cos>(Op0: m_Specific(V: X))); |
| 2271 | bool IsCot = |
| 2272 | !IsTan && match(V: Op0, P: m_Intrinsic<Intrinsic::cos>(Op0: m_Value(V&: X))) && |
| 2273 | match(V: Op1, P: m_Intrinsic<Intrinsic::sin>(Op0: m_Specific(V: X))); |
| 2274 | |
| 2275 | if ((IsTan || IsCot) && hasFloatFn(M, TLI: &TLI, Ty: I.getType(), DoubleFn: LibFunc_tan, |
| 2276 | FloatFn: LibFunc_tanf, LongDoubleFn: LibFunc_tanl)) { |
| 2277 | IRBuilder<> B(&I); |
| 2278 | IRBuilder<>::FastMathFlagGuard FMFGuard(B); |
| 2279 | B.setFastMathFlags(I.getFastMathFlags()); |
| 2280 | AttributeList Attrs = |
| 2281 | cast<CallBase>(Val: Op0)->getCalledFunction()->getAttributes(); |
| 2282 | Value *Res = emitUnaryFloatFnCall(Op: X, TLI: &TLI, DoubleFn: LibFunc_tan, FloatFn: LibFunc_tanf, |
| 2283 | LongDoubleFn: LibFunc_tanl, B, Attrs); |
| 2284 | if (IsCot) |
| 2285 | Res = B.CreateFDiv(L: ConstantFP::get(Ty: I.getType(), V: 1.0), R: Res); |
| 2286 | return replaceInstUsesWith(I, V: Res); |
| 2287 | } |
| 2288 | } |
| 2289 | |
| 2290 | // X / (X * Y) --> 1.0 / Y |
| 2291 | // Reassociate to (X / X -> 1.0) is legal when NaNs are not allowed. |
| 2292 | // We can ignore the possibility that X is infinity because INF/INF is NaN. |
| 2293 | Value *X, *Y; |
| 2294 | if (I.hasNoNaNs() && I.hasAllowReassoc() && |
| 2295 | match(V: Op1, P: m_c_FMul(L: m_Specific(V: Op0), R: m_Value(V&: Y)))) { |
| 2296 | replaceOperand(I, OpNum: 0, V: ConstantFP::get(Ty: I.getType(), V: 1.0)); |
| 2297 | replaceOperand(I, OpNum: 1, V: Y); |
| 2298 | return &I; |
| 2299 | } |
| 2300 | |
| 2301 | // X / fabs(X) -> copysign(1.0, X) |
| 2302 | // fabs(X) / X -> copysign(1.0, X) |
| 2303 | if (I.hasNoNaNs() && I.hasNoInfs() && |
| 2304 | (match(V: &I, P: m_FDiv(L: m_Value(V&: X), R: m_FAbs(Op0: m_Deferred(V: X)))) || |
| 2305 | match(V: &I, P: m_FDiv(L: m_FAbs(Op0: m_Value(V&: X)), R: m_Deferred(V: X))))) { |
| 2306 | Value *V = Builder.CreateBinaryIntrinsic( |
| 2307 | ID: Intrinsic::copysign, LHS: ConstantFP::get(Ty: I.getType(), V: 1.0), RHS: X, FMFSource: &I); |
| 2308 | return replaceInstUsesWith(I, V); |
| 2309 | } |
| 2310 | |
| 2311 | if (Instruction *Mul = foldFDivPowDivisor(I, Builder)) |
| 2312 | return Mul; |
| 2313 | |
| 2314 | if (Instruction *Mul = foldFDivSqrtDivisor(I, Builder)) |
| 2315 | return Mul; |
| 2316 | |
| 2317 | // pow(X, Y) / X --> pow(X, Y-1) |
| 2318 | if (I.hasAllowReassoc() && |
| 2319 | match(V: Op0, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::pow>(Op0: m_Specific(V: Op1), |
| 2320 | Op1: m_Value(V&: Y))))) { |
| 2321 | Value *Y1 = |
| 2322 | Builder.CreateFAddFMF(L: Y, R: ConstantFP::get(Ty: I.getType(), V: -1.0), FMFSource: &I); |
| 2323 | Value *Pow = Builder.CreateBinaryIntrinsic(ID: Intrinsic::pow, LHS: Op1, RHS: Y1, FMFSource: &I); |
| 2324 | return replaceInstUsesWith(I, V: Pow); |
| 2325 | } |
| 2326 | |
| 2327 | if (Instruction *FoldedPowi = foldPowiReassoc(I)) |
| 2328 | return FoldedPowi; |
| 2329 | |
| 2330 | return nullptr; |
| 2331 | } |
| 2332 | |
| 2333 | // Variety of transform for: |
| 2334 | // (urem/srem (mul X, Y), (mul X, Z)) |
| 2335 | // (urem/srem (shl X, Y), (shl X, Z)) |
| 2336 | // (urem/srem (shl Y, X), (shl Z, X)) |
| 2337 | // NB: The shift cases are really just extensions of the mul case. We treat |
| 2338 | // shift as Val * (1 << Amt). |
| 2339 | static Instruction *simplifyIRemMulShl(BinaryOperator &I, |
| 2340 | InstCombinerImpl &IC) { |
| 2341 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1), *X = nullptr; |
| 2342 | APInt Y, Z; |
| 2343 | bool ShiftByX = false; |
| 2344 | |
| 2345 | // If V is not nullptr, it will be matched using m_Specific. |
| 2346 | auto MatchShiftOrMulXC = [](Value *Op, Value *&V, APInt &C, |
| 2347 | bool &PreserveNSW) -> bool { |
| 2348 | const APInt *Tmp = nullptr; |
| 2349 | if ((!V && match(V: Op, P: m_Mul(L: m_Value(V), R: m_APInt(Res&: Tmp)))) || |
| 2350 | (V && match(V: Op, P: m_Mul(L: m_Specific(V), R: m_APInt(Res&: Tmp))))) |
| 2351 | C = *Tmp; |
| 2352 | else if ((!V && match(V: Op, P: m_Shl(L: m_Value(V), R: m_APInt(Res&: Tmp)))) || |
| 2353 | (V && match(V: Op, P: m_Shl(L: m_Specific(V), R: m_APInt(Res&: Tmp))))) { |
| 2354 | C = APInt(Tmp->getBitWidth(), 1) << *Tmp; |
| 2355 | // We cannot preserve NSW when shifting by BW - 1. |
| 2356 | PreserveNSW = Tmp->ult(RHS: Tmp->getBitWidth() - 1); |
| 2357 | } |
| 2358 | if (Tmp != nullptr) |
| 2359 | return true; |
| 2360 | |
| 2361 | // Reset `V` so we don't start with specific value on next match attempt. |
| 2362 | V = nullptr; |
| 2363 | return false; |
| 2364 | }; |
| 2365 | |
| 2366 | auto MatchShiftCX = [](Value *Op, APInt &C, Value *&V) -> bool { |
| 2367 | const APInt *Tmp = nullptr; |
| 2368 | if ((!V && match(V: Op, P: m_Shl(L: m_APInt(Res&: Tmp), R: m_Value(V)))) || |
| 2369 | (V && match(V: Op, P: m_Shl(L: m_APInt(Res&: Tmp), R: m_Specific(V))))) { |
| 2370 | C = *Tmp; |
| 2371 | return true; |
| 2372 | } |
| 2373 | |
| 2374 | // Reset `V` so we don't start with specific value on next match attempt. |
| 2375 | V = nullptr; |
| 2376 | return false; |
| 2377 | }; |
| 2378 | |
| 2379 | bool Op0PreserveNSW = true, Op1PreserveNSW = true; |
| 2380 | if (MatchShiftOrMulXC(Op0, X, Y, Op0PreserveNSW) && |
| 2381 | MatchShiftOrMulXC(Op1, X, Z, Op1PreserveNSW)) { |
| 2382 | // pass |
| 2383 | } else if (MatchShiftCX(Op0, Y, X) && MatchShiftCX(Op1, Z, X)) { |
| 2384 | ShiftByX = true; |
| 2385 | } else { |
| 2386 | return nullptr; |
| 2387 | } |
| 2388 | |
| 2389 | bool IsSRem = I.getOpcode() == Instruction::SRem; |
| 2390 | |
| 2391 | OverflowingBinaryOperator *BO0 = cast<OverflowingBinaryOperator>(Val: Op0); |
| 2392 | // TODO: We may be able to deduce more about nsw/nuw of BO0/BO1 based on Y >= |
| 2393 | // Z or Z >= Y. |
| 2394 | bool BO0HasNSW = Op0PreserveNSW && BO0->hasNoSignedWrap(); |
| 2395 | bool BO0HasNUW = BO0->hasNoUnsignedWrap(); |
| 2396 | bool BO0NoWrap = IsSRem ? BO0HasNSW : BO0HasNUW; |
| 2397 | |
| 2398 | APInt RemYZ = IsSRem ? Y.srem(RHS: Z) : Y.urem(RHS: Z); |
| 2399 | // (rem (mul nuw/nsw X, Y), (mul X, Z)) |
| 2400 | // if (rem Y, Z) == 0 |
| 2401 | // -> 0 |
| 2402 | if (RemYZ.isZero() && BO0NoWrap) |
| 2403 | return IC.replaceInstUsesWith(I, V: ConstantInt::getNullValue(Ty: I.getType())); |
| 2404 | |
| 2405 | // Helper function to emit either (RemSimplificationC << X) or |
| 2406 | // (RemSimplificationC * X) depending on whether we matched Op0/Op1 as |
| 2407 | // (shl V, X) or (mul V, X) respectively. |
| 2408 | auto CreateMulOrShift = |
| 2409 | [&](const APInt &RemSimplificationC) -> BinaryOperator * { |
| 2410 | Value *RemSimplification = |
| 2411 | ConstantInt::get(Ty: I.getType(), V: RemSimplificationC); |
| 2412 | return ShiftByX ? BinaryOperator::CreateShl(V1: RemSimplification, V2: X) |
| 2413 | : BinaryOperator::CreateMul(V1: X, V2: RemSimplification); |
| 2414 | }; |
| 2415 | |
| 2416 | OverflowingBinaryOperator *BO1 = cast<OverflowingBinaryOperator>(Val: Op1); |
| 2417 | bool BO1HasNSW = Op1PreserveNSW && BO1->hasNoSignedWrap(); |
| 2418 | bool BO1HasNUW = BO1->hasNoUnsignedWrap(); |
| 2419 | bool BO1NoWrap = IsSRem ? BO1HasNSW : BO1HasNUW; |
| 2420 | // (rem (mul X, Y), (mul nuw/nsw X, Z)) |
| 2421 | // if (rem Y, Z) == Y |
| 2422 | // -> (mul nuw/nsw X, Y) |
| 2423 | if (RemYZ == Y && BO1NoWrap) { |
| 2424 | BinaryOperator *BO = CreateMulOrShift(Y); |
| 2425 | // Copy any overflow flags from Op0. |
| 2426 | BO->setHasNoSignedWrap(IsSRem || BO0HasNSW); |
| 2427 | BO->setHasNoUnsignedWrap(!IsSRem || BO0HasNUW); |
| 2428 | return BO; |
| 2429 | } |
| 2430 | |
| 2431 | // (rem (mul nuw/nsw X, Y), (mul {nsw} X, Z)) |
| 2432 | // if Y >= Z |
| 2433 | // -> (mul {nuw} nsw X, (rem Y, Z)) |
| 2434 | if (Y.uge(RHS: Z) && (IsSRem ? (BO0HasNSW && BO1HasNSW) : BO0HasNUW)) { |
| 2435 | BinaryOperator *BO = CreateMulOrShift(RemYZ); |
| 2436 | BO->setHasNoSignedWrap(); |
| 2437 | BO->setHasNoUnsignedWrap(BO0HasNUW); |
| 2438 | return BO; |
| 2439 | } |
| 2440 | |
| 2441 | return nullptr; |
| 2442 | } |
| 2443 | |
| 2444 | /// This function implements the transforms common to both integer remainder |
| 2445 | /// instructions (urem and srem). It is called by the visitors to those integer |
| 2446 | /// remainder instructions. |
| 2447 | /// Common integer remainder transforms |
| 2448 | Instruction *InstCombinerImpl::commonIRemTransforms(BinaryOperator &I) { |
| 2449 | if (Instruction *Res = commonIDivRemTransforms(I)) |
| 2450 | return Res; |
| 2451 | |
| 2452 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 2453 | |
| 2454 | if (isa<Constant>(Val: Op1)) { |
| 2455 | if (Instruction *Op0I = dyn_cast<Instruction>(Val: Op0)) { |
| 2456 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op0I)) { |
| 2457 | if (Instruction *R = FoldOpIntoSelect(Op&: I, SI)) |
| 2458 | return R; |
| 2459 | } else if (auto *PN = dyn_cast<PHINode>(Val: Op0I)) { |
| 2460 | const APInt *Op1Int; |
| 2461 | if (match(V: Op1, P: m_APInt(Res&: Op1Int)) && !Op1Int->isMinValue() && |
| 2462 | (I.getOpcode() == Instruction::URem || |
| 2463 | !Op1Int->isMinSignedValue())) { |
| 2464 | // foldOpIntoPhi will speculate instructions to the end of the PHI's |
| 2465 | // predecessor blocks, so do this only if we know the srem or urem |
| 2466 | // will not fault. |
| 2467 | if (Instruction *NV = foldOpIntoPhi(I, PN)) |
| 2468 | return NV; |
| 2469 | } |
| 2470 | } |
| 2471 | |
| 2472 | // See if we can fold away this rem instruction. |
| 2473 | if (SimplifyDemandedInstructionBits(Inst&: I)) |
| 2474 | return &I; |
| 2475 | } |
| 2476 | } |
| 2477 | |
| 2478 | if (Instruction *R = simplifyIRemMulShl(I, IC&: *this)) |
| 2479 | return R; |
| 2480 | |
| 2481 | return nullptr; |
| 2482 | } |
| 2483 | |
| 2484 | Instruction *InstCombinerImpl::visitURem(BinaryOperator &I) { |
| 2485 | if (Value *V = simplifyURemInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), |
| 2486 | Q: SQ.getWithInstruction(I: &I))) |
| 2487 | return replaceInstUsesWith(I, V); |
| 2488 | |
| 2489 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
| 2490 | return X; |
| 2491 | |
| 2492 | if (Instruction *common = commonIRemTransforms(I)) |
| 2493 | return common; |
| 2494 | |
| 2495 | if (Instruction *NarrowRem = narrowUDivURem(I, IC&: *this)) |
| 2496 | return NarrowRem; |
| 2497 | |
| 2498 | // X urem Y -> X and Y-1, where Y is a power of 2, |
| 2499 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 2500 | Type *Ty = I.getType(); |
| 2501 | if (isKnownToBeAPowerOfTwo(V: Op1, /*OrZero*/ true, CxtI: &I)) { |
| 2502 | // This may increase instruction count, we don't enforce that Y is a |
| 2503 | // constant. |
| 2504 | Constant *N1 = Constant::getAllOnesValue(Ty); |
| 2505 | Value *Add = Builder.CreateAdd(LHS: Op1, RHS: N1); |
| 2506 | return BinaryOperator::CreateAnd(V1: Op0, V2: Add); |
| 2507 | } |
| 2508 | |
| 2509 | // 1 urem X -> zext(X != 1) |
| 2510 | if (match(V: Op0, P: m_One())) { |
| 2511 | Value *Cmp = Builder.CreateICmpNE(LHS: Op1, RHS: ConstantInt::get(Ty, V: 1)); |
| 2512 | return CastInst::CreateZExtOrBitCast(S: Cmp, Ty); |
| 2513 | } |
| 2514 | |
| 2515 | // Op0 urem C -> Op0 < C ? Op0 : Op0 - C, where C >= signbit. |
| 2516 | // Op0 must be frozen because we are increasing its number of uses. |
| 2517 | if (match(V: Op1, P: m_Negative())) { |
| 2518 | Value *F0 = Op0; |
| 2519 | if (!isGuaranteedNotToBeUndef(V: Op0)) |
| 2520 | F0 = Builder.CreateFreeze(V: Op0, Name: Op0->getName() + ".fr" ); |
| 2521 | Value *Cmp = Builder.CreateICmpULT(LHS: F0, RHS: Op1); |
| 2522 | Value *Sub = Builder.CreateSub(LHS: F0, RHS: Op1); |
| 2523 | return createSelectInstWithUnknownProfile(C: Cmp, S1: F0, S2: Sub); |
| 2524 | } |
| 2525 | |
| 2526 | // If the divisor is a sext of a boolean, then the divisor must be max |
| 2527 | // unsigned value (-1). Therefore, the remainder is Op0 unless Op0 is also |
| 2528 | // max unsigned value. In that case, the remainder is 0: |
| 2529 | // urem Op0, (sext i1 X) --> (Op0 == -1) ? 0 : Op0 |
| 2530 | Value *X; |
| 2531 | if (match(V: Op1, P: m_SExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
| 2532 | Value *FrozenOp0 = Op0; |
| 2533 | if (!isGuaranteedNotToBeUndef(V: Op0)) |
| 2534 | FrozenOp0 = Builder.CreateFreeze(V: Op0, Name: Op0->getName() + ".frozen" ); |
| 2535 | Value *Cmp = |
| 2536 | Builder.CreateICmpEQ(LHS: FrozenOp0, RHS: ConstantInt::getAllOnesValue(Ty)); |
| 2537 | return createSelectInstWithUnknownProfile( |
| 2538 | C: Cmp, S1: ConstantInt::getNullValue(Ty), S2: FrozenOp0); |
| 2539 | } |
| 2540 | |
| 2541 | // For "(X + 1) % Op1" and if (X u< Op1) => (X + 1) == Op1 ? 0 : X + 1 . |
| 2542 | if (match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_One()))) { |
| 2543 | Value *Val = |
| 2544 | simplifyICmpInst(Pred: ICmpInst::ICMP_ULT, LHS: X, RHS: Op1, Q: SQ.getWithInstruction(I: &I)); |
| 2545 | if (Val && match(V: Val, P: m_One())) { |
| 2546 | Value *FrozenOp0 = Op0; |
| 2547 | if (!isGuaranteedNotToBeUndef(V: Op0)) |
| 2548 | FrozenOp0 = Builder.CreateFreeze(V: Op0, Name: Op0->getName() + ".frozen" ); |
| 2549 | Value *Cmp = Builder.CreateICmpEQ(LHS: FrozenOp0, RHS: Op1); |
| 2550 | return createSelectInstWithUnknownProfile( |
| 2551 | C: Cmp, S1: ConstantInt::getNullValue(Ty), S2: FrozenOp0); |
| 2552 | } |
| 2553 | } |
| 2554 | |
| 2555 | return nullptr; |
| 2556 | } |
| 2557 | |
| 2558 | Instruction *InstCombinerImpl::visitSRem(BinaryOperator &I) { |
| 2559 | if (Value *V = simplifySRemInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), |
| 2560 | Q: SQ.getWithInstruction(I: &I))) |
| 2561 | return replaceInstUsesWith(I, V); |
| 2562 | |
| 2563 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
| 2564 | return X; |
| 2565 | |
| 2566 | // Handle the integer rem common cases |
| 2567 | if (Instruction *Common = commonIRemTransforms(I)) |
| 2568 | return Common; |
| 2569 | |
| 2570 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
| 2571 | { |
| 2572 | const APInt *Y; |
| 2573 | // X % -Y -> X % Y |
| 2574 | if (match(V: Op1, P: m_Negative(V&: Y)) && !Y->isMinSignedValue()) |
| 2575 | return replaceOperand(I, OpNum: 1, V: ConstantInt::get(Ty: I.getType(), V: -*Y)); |
| 2576 | } |
| 2577 | |
| 2578 | // -X srem Y --> -(X srem Y) |
| 2579 | Value *X, *Y; |
| 2580 | if (match(V: &I, P: m_SRem(L: m_OneUse(SubPattern: m_NSWNeg(V: m_Value(V&: X))), R: m_Value(V&: Y)))) |
| 2581 | return BinaryOperator::CreateNSWNeg(Op: Builder.CreateSRem(LHS: X, RHS: Y)); |
| 2582 | |
| 2583 | // If the sign bits of both operands are zero (i.e. we can prove they are |
| 2584 | // unsigned inputs), turn this into a urem. |
| 2585 | APInt Mask(APInt::getSignMask(BitWidth: I.getType()->getScalarSizeInBits())); |
| 2586 | if (MaskedValueIsZero(V: Op1, Mask, CxtI: &I) && MaskedValueIsZero(V: Op0, Mask, CxtI: &I)) { |
| 2587 | // X srem Y -> X urem Y, iff X and Y don't have sign bit set |
| 2588 | return BinaryOperator::CreateURem(V1: Op0, V2: Op1, Name: I.getName()); |
| 2589 | } |
| 2590 | |
| 2591 | // If it's a constant vector, flip any negative values positive. |
| 2592 | if (isa<ConstantVector>(Val: Op1) || isa<ConstantDataVector>(Val: Op1)) { |
| 2593 | Constant *C = cast<Constant>(Val: Op1); |
| 2594 | unsigned VWidth = cast<FixedVectorType>(Val: C->getType())->getNumElements(); |
| 2595 | |
| 2596 | bool hasNegative = false; |
| 2597 | bool hasMissing = false; |
| 2598 | for (unsigned i = 0; i != VWidth; ++i) { |
| 2599 | Constant *Elt = C->getAggregateElement(Elt: i); |
| 2600 | if (!Elt) { |
| 2601 | hasMissing = true; |
| 2602 | break; |
| 2603 | } |
| 2604 | |
| 2605 | if (ConstantInt *RHS = dyn_cast<ConstantInt>(Val: Elt)) |
| 2606 | if (RHS->isNegative()) |
| 2607 | hasNegative = true; |
| 2608 | } |
| 2609 | |
| 2610 | if (hasNegative && !hasMissing) { |
| 2611 | SmallVector<Constant *, 16> Elts(VWidth); |
| 2612 | for (unsigned i = 0; i != VWidth; ++i) { |
| 2613 | Elts[i] = C->getAggregateElement(Elt: i); // Handle undef, etc. |
| 2614 | if (ConstantInt *RHS = dyn_cast<ConstantInt>(Val: Elts[i])) { |
| 2615 | if (RHS->isNegative()) |
| 2616 | Elts[i] = cast<ConstantInt>(Val: ConstantExpr::getNeg(C: RHS)); |
| 2617 | } |
| 2618 | } |
| 2619 | |
| 2620 | Constant *NewRHSV = ConstantVector::get(V: Elts); |
| 2621 | if (NewRHSV != C) // Don't loop on -MININT |
| 2622 | return replaceOperand(I, OpNum: 1, V: NewRHSV); |
| 2623 | } |
| 2624 | } |
| 2625 | |
| 2626 | return nullptr; |
| 2627 | } |
| 2628 | |
| 2629 | Instruction *InstCombinerImpl::visitFRem(BinaryOperator &I) { |
| 2630 | if (Value *V = simplifyFRemInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), |
| 2631 | FMF: I.getFastMathFlags(), |
| 2632 | Q: SQ.getWithInstruction(I: &I))) |
| 2633 | return replaceInstUsesWith(I, V); |
| 2634 | |
| 2635 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
| 2636 | return X; |
| 2637 | |
| 2638 | if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I)) |
| 2639 | return Phi; |
| 2640 | |
| 2641 | return nullptr; |
| 2642 | } |
| 2643 | |