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/SmallVector.h" 17 #include "llvm/Analysis/InstructionSimplify.h" 18 #include "llvm/IR/BasicBlock.h" 19 #include "llvm/IR/Constant.h" 20 #include "llvm/IR/Constants.h" 21 #include "llvm/IR/InstrTypes.h" 22 #include "llvm/IR/Instruction.h" 23 #include "llvm/IR/Instructions.h" 24 #include "llvm/IR/IntrinsicInst.h" 25 #include "llvm/IR/Intrinsics.h" 26 #include "llvm/IR/Operator.h" 27 #include "llvm/IR/PatternMatch.h" 28 #include "llvm/IR/Type.h" 29 #include "llvm/IR/Value.h" 30 #include "llvm/Support/Casting.h" 31 #include "llvm/Support/ErrorHandling.h" 32 #include "llvm/Transforms/InstCombine/InstCombiner.h" 33 #include "llvm/Transforms/Utils/BuildLibCalls.h" 34 #include <cassert> 35 36 #define DEBUG_TYPE "instcombine" 37 #include "llvm/Transforms/Utils/InstructionWorklist.h" 38 39 using namespace llvm; 40 using namespace PatternMatch; 41 42 /// The specific integer value is used in a context where it is known to be 43 /// non-zero. If this allows us to simplify the computation, do so and return 44 /// the new operand, otherwise return null. 45 static Value *simplifyValueKnownNonZero(Value *V, InstCombinerImpl &IC, 46 Instruction &CxtI) { 47 // If V has multiple uses, then we would have to do more analysis to determine 48 // if this is safe. For example, the use could be in dynamically unreached 49 // code. 50 if (!V->hasOneUse()) return nullptr; 51 52 bool MadeChange = false; 53 54 // ((1 << A) >>u B) --> (1 << (A-B)) 55 // Because V cannot be zero, we know that B is less than A. 56 Value *A = nullptr, *B = nullptr, *One = nullptr; 57 if (match(V, m_LShr(m_OneUse(m_Shl(m_Value(One), m_Value(A))), m_Value(B))) && 58 match(One, m_One())) { 59 A = IC.Builder.CreateSub(A, B); 60 return IC.Builder.CreateShl(One, A); 61 } 62 63 // (PowerOfTwo >>u B) --> isExact since shifting out the result would make it 64 // inexact. Similarly for <<. 65 BinaryOperator *I = dyn_cast<BinaryOperator>(V); 66 if (I && I->isLogicalShift() && 67 IC.isKnownToBeAPowerOfTwo(I->getOperand(0), false, 0, &CxtI)) { 68 // We know that this is an exact/nuw shift and that the input is a 69 // non-zero context as well. 70 if (Value *V2 = simplifyValueKnownNonZero(I->getOperand(0), IC, CxtI)) { 71 IC.replaceOperand(*I, 0, V2); 72 MadeChange = true; 73 } 74 75 if (I->getOpcode() == Instruction::LShr && !I->isExact()) { 76 I->setIsExact(); 77 MadeChange = true; 78 } 79 80 if (I->getOpcode() == Instruction::Shl && !I->hasNoUnsignedWrap()) { 81 I->setHasNoUnsignedWrap(); 82 MadeChange = true; 83 } 84 } 85 86 // TODO: Lots more we could do here: 87 // If V is a phi node, we can call this on each of its operands. 88 // "select cond, X, 0" can simplify to "X". 89 90 return MadeChange ? V : nullptr; 91 } 92 93 // TODO: This is a specific form of a much more general pattern. 94 // We could detect a select with any binop identity constant, or we 95 // could use SimplifyBinOp to see if either arm of the select reduces. 96 // But that needs to be done carefully and/or while removing potential 97 // reverse canonicalizations as in InstCombiner::foldSelectIntoOp(). 98 static Value *foldMulSelectToNegate(BinaryOperator &I, 99 InstCombiner::BuilderTy &Builder) { 100 Value *Cond, *OtherOp; 101 102 // mul (select Cond, 1, -1), OtherOp --> select Cond, OtherOp, -OtherOp 103 // mul OtherOp, (select Cond, 1, -1) --> select Cond, OtherOp, -OtherOp 104 if (match(&I, m_c_Mul(m_OneUse(m_Select(m_Value(Cond), m_One(), m_AllOnes())), 105 m_Value(OtherOp)))) { 106 bool HasAnyNoWrap = I.hasNoSignedWrap() || I.hasNoUnsignedWrap(); 107 Value *Neg = Builder.CreateNeg(OtherOp, "", false, HasAnyNoWrap); 108 return Builder.CreateSelect(Cond, OtherOp, Neg); 109 } 110 // mul (select Cond, -1, 1), OtherOp --> select Cond, -OtherOp, OtherOp 111 // mul OtherOp, (select Cond, -1, 1) --> select Cond, -OtherOp, OtherOp 112 if (match(&I, m_c_Mul(m_OneUse(m_Select(m_Value(Cond), m_AllOnes(), m_One())), 113 m_Value(OtherOp)))) { 114 bool HasAnyNoWrap = I.hasNoSignedWrap() || I.hasNoUnsignedWrap(); 115 Value *Neg = Builder.CreateNeg(OtherOp, "", false, HasAnyNoWrap); 116 return Builder.CreateSelect(Cond, Neg, OtherOp); 117 } 118 119 // fmul (select Cond, 1.0, -1.0), OtherOp --> select Cond, OtherOp, -OtherOp 120 // fmul OtherOp, (select Cond, 1.0, -1.0) --> select Cond, OtherOp, -OtherOp 121 if (match(&I, m_c_FMul(m_OneUse(m_Select(m_Value(Cond), m_SpecificFP(1.0), 122 m_SpecificFP(-1.0))), 123 m_Value(OtherOp)))) { 124 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); 125 Builder.setFastMathFlags(I.getFastMathFlags()); 126 return Builder.CreateSelect(Cond, OtherOp, Builder.CreateFNeg(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(&I, m_c_FMul(m_OneUse(m_Select(m_Value(Cond), m_SpecificFP(-1.0), 132 m_SpecificFP(1.0))), 133 m_Value(OtherOp)))) { 134 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); 135 Builder.setFastMathFlags(I.getFastMathFlags()); 136 return Builder.CreateSelect(Cond, Builder.CreateFNeg(OtherOp), OtherOp); 137 } 138 139 return nullptr; 140 } 141 142 Instruction *InstCombinerImpl::visitMul(BinaryOperator &I) { 143 if (Value *V = SimplifyMulInst(I.getOperand(0), I.getOperand(1), 144 SQ.getWithInstruction(&I))) 145 return replaceInstUsesWith(I, V); 146 147 if (SimplifyAssociativeOrCommutative(I)) 148 return &I; 149 150 if (Instruction *X = foldVectorBinop(I)) 151 return X; 152 153 if (Instruction *Phi = foldBinopWithPhiOperands(I)) 154 return Phi; 155 156 if (Value *V = SimplifyUsingDistributiveLaws(I)) 157 return replaceInstUsesWith(I, V); 158 159 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 160 unsigned BitWidth = I.getType()->getScalarSizeInBits(); 161 162 // X * -1 == 0 - X 163 if (match(Op1, m_AllOnes())) { 164 BinaryOperator *BO = BinaryOperator::CreateNeg(Op0, I.getName()); 165 if (I.hasNoSignedWrap()) 166 BO->setHasNoSignedWrap(); 167 return BO; 168 } 169 170 // Also allow combining multiply instructions on vectors. 171 { 172 Value *NewOp; 173 Constant *C1, *C2; 174 const APInt *IVal; 175 if (match(&I, m_Mul(m_Shl(m_Value(NewOp), m_Constant(C2)), 176 m_Constant(C1))) && 177 match(C1, m_APInt(IVal))) { 178 // ((X << C2)*C1) == (X * (C1 << C2)) 179 Constant *Shl = ConstantExpr::getShl(C1, C2); 180 BinaryOperator *Mul = cast<BinaryOperator>(I.getOperand(0)); 181 BinaryOperator *BO = BinaryOperator::CreateMul(NewOp, Shl); 182 if (I.hasNoUnsignedWrap() && Mul->hasNoUnsignedWrap()) 183 BO->setHasNoUnsignedWrap(); 184 if (I.hasNoSignedWrap() && Mul->hasNoSignedWrap() && 185 Shl->isNotMinSignedValue()) 186 BO->setHasNoSignedWrap(); 187 return BO; 188 } 189 190 if (match(&I, m_Mul(m_Value(NewOp), m_Constant(C1)))) { 191 // Replace X*(2^C) with X << C, where C is either a scalar or a vector. 192 if (Constant *NewCst = ConstantExpr::getExactLogBase2(C1)) { 193 BinaryOperator *Shl = BinaryOperator::CreateShl(NewOp, NewCst); 194 195 if (I.hasNoUnsignedWrap()) 196 Shl->setHasNoUnsignedWrap(); 197 if (I.hasNoSignedWrap()) { 198 const APInt *V; 199 if (match(NewCst, m_APInt(V)) && *V != V->getBitWidth() - 1) 200 Shl->setHasNoSignedWrap(); 201 } 202 203 return Shl; 204 } 205 } 206 } 207 208 if (Op0->hasOneUse() && match(Op1, m_NegatedPower2())) { 209 // Interpret X * (-1<<C) as (-X) * (1<<C) and try to sink the negation. 210 // The "* (1<<C)" thus becomes a potential shifting opportunity. 211 if (Value *NegOp0 = Negator::Negate(/*IsNegation*/ true, Op0, *this)) 212 return BinaryOperator::CreateMul( 213 NegOp0, ConstantExpr::getNeg(cast<Constant>(Op1)), I.getName()); 214 } 215 216 if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I)) 217 return FoldedMul; 218 219 if (Value *FoldedMul = foldMulSelectToNegate(I, Builder)) 220 return replaceInstUsesWith(I, FoldedMul); 221 222 // Simplify mul instructions with a constant RHS. 223 if (isa<Constant>(Op1)) { 224 // Canonicalize (X+C1)*CI -> X*CI+C1*CI. 225 Value *X; 226 Constant *C1; 227 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_Constant(C1))))) { 228 Value *Mul = Builder.CreateMul(C1, Op1); 229 // Only go forward with the transform if C1*CI simplifies to a tidier 230 // constant. 231 if (!match(Mul, m_Mul(m_Value(), m_Value()))) 232 return BinaryOperator::CreateAdd(Builder.CreateMul(X, Op1), Mul); 233 } 234 } 235 236 // abs(X) * abs(X) -> X * X 237 // nabs(X) * nabs(X) -> X * X 238 if (Op0 == Op1) { 239 Value *X, *Y; 240 SelectPatternFlavor SPF = matchSelectPattern(Op0, X, Y).Flavor; 241 if (SPF == SPF_ABS || SPF == SPF_NABS) 242 return BinaryOperator::CreateMul(X, X); 243 244 if (match(Op0, m_Intrinsic<Intrinsic::abs>(m_Value(X)))) 245 return BinaryOperator::CreateMul(X, X); 246 } 247 248 // -X * C --> X * -C 249 Value *X, *Y; 250 Constant *Op1C; 251 if (match(Op0, m_Neg(m_Value(X))) && match(Op1, m_Constant(Op1C))) 252 return BinaryOperator::CreateMul(X, ConstantExpr::getNeg(Op1C)); 253 254 // -X * -Y --> X * Y 255 if (match(Op0, m_Neg(m_Value(X))) && match(Op1, m_Neg(m_Value(Y)))) { 256 auto *NewMul = BinaryOperator::CreateMul(X, Y); 257 if (I.hasNoSignedWrap() && 258 cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap() && 259 cast<OverflowingBinaryOperator>(Op1)->hasNoSignedWrap()) 260 NewMul->setHasNoSignedWrap(); 261 return NewMul; 262 } 263 264 // -X * Y --> -(X * Y) 265 // X * -Y --> -(X * Y) 266 if (match(&I, m_c_Mul(m_OneUse(m_Neg(m_Value(X))), m_Value(Y)))) 267 return BinaryOperator::CreateNeg(Builder.CreateMul(X, Y)); 268 269 // (X / Y) * Y = X - (X % Y) 270 // (X / Y) * -Y = (X % Y) - X 271 { 272 Value *Y = Op1; 273 BinaryOperator *Div = dyn_cast<BinaryOperator>(Op0); 274 if (!Div || (Div->getOpcode() != Instruction::UDiv && 275 Div->getOpcode() != Instruction::SDiv)) { 276 Y = Op0; 277 Div = dyn_cast<BinaryOperator>(Op1); 278 } 279 Value *Neg = dyn_castNegVal(Y); 280 if (Div && Div->hasOneUse() && 281 (Div->getOperand(1) == Y || Div->getOperand(1) == Neg) && 282 (Div->getOpcode() == Instruction::UDiv || 283 Div->getOpcode() == Instruction::SDiv)) { 284 Value *X = Div->getOperand(0), *DivOp1 = Div->getOperand(1); 285 286 // If the division is exact, X % Y is zero, so we end up with X or -X. 287 if (Div->isExact()) { 288 if (DivOp1 == Y) 289 return replaceInstUsesWith(I, X); 290 return BinaryOperator::CreateNeg(X); 291 } 292 293 auto RemOpc = Div->getOpcode() == Instruction::UDiv ? Instruction::URem 294 : Instruction::SRem; 295 Value *Rem = Builder.CreateBinOp(RemOpc, X, DivOp1); 296 if (DivOp1 == Y) 297 return BinaryOperator::CreateSub(X, Rem); 298 return BinaryOperator::CreateSub(Rem, X); 299 } 300 } 301 302 /// i1 mul -> i1 and. 303 if (I.getType()->isIntOrIntVectorTy(1)) 304 return BinaryOperator::CreateAnd(Op0, Op1); 305 306 // X*(1 << Y) --> X << Y 307 // (1 << Y)*X --> X << Y 308 { 309 Value *Y; 310 BinaryOperator *BO = nullptr; 311 bool ShlNSW = false; 312 if (match(Op0, m_Shl(m_One(), m_Value(Y)))) { 313 BO = BinaryOperator::CreateShl(Op1, Y); 314 ShlNSW = cast<ShlOperator>(Op0)->hasNoSignedWrap(); 315 } else if (match(Op1, m_Shl(m_One(), m_Value(Y)))) { 316 BO = BinaryOperator::CreateShl(Op0, Y); 317 ShlNSW = cast<ShlOperator>(Op1)->hasNoSignedWrap(); 318 } 319 if (BO) { 320 if (I.hasNoUnsignedWrap()) 321 BO->setHasNoUnsignedWrap(); 322 if (I.hasNoSignedWrap() && ShlNSW) 323 BO->setHasNoSignedWrap(); 324 return BO; 325 } 326 } 327 328 // (zext bool X) * (zext bool Y) --> zext (and X, Y) 329 // (sext bool X) * (sext bool Y) --> zext (and X, Y) 330 // Note: -1 * -1 == 1 * 1 == 1 (if the extends match, the result is the same) 331 if (((match(Op0, m_ZExt(m_Value(X))) && match(Op1, m_ZExt(m_Value(Y)))) || 332 (match(Op0, m_SExt(m_Value(X))) && match(Op1, m_SExt(m_Value(Y))))) && 333 X->getType()->isIntOrIntVectorTy(1) && X->getType() == Y->getType() && 334 (Op0->hasOneUse() || Op1->hasOneUse() || X == Y)) { 335 Value *And = Builder.CreateAnd(X, Y, "mulbool"); 336 return CastInst::Create(Instruction::ZExt, And, I.getType()); 337 } 338 // (sext bool X) * (zext bool Y) --> sext (and X, Y) 339 // (zext bool X) * (sext bool Y) --> sext (and X, Y) 340 // Note: -1 * 1 == 1 * -1 == -1 341 if (((match(Op0, m_SExt(m_Value(X))) && match(Op1, m_ZExt(m_Value(Y)))) || 342 (match(Op0, m_ZExt(m_Value(X))) && match(Op1, m_SExt(m_Value(Y))))) && 343 X->getType()->isIntOrIntVectorTy(1) && X->getType() == Y->getType() && 344 (Op0->hasOneUse() || Op1->hasOneUse())) { 345 Value *And = Builder.CreateAnd(X, Y, "mulbool"); 346 return CastInst::Create(Instruction::SExt, And, I.getType()); 347 } 348 349 // (zext bool X) * Y --> X ? Y : 0 350 // Y * (zext bool X) --> X ? Y : 0 351 if (match(Op0, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) 352 return SelectInst::Create(X, Op1, ConstantInt::get(I.getType(), 0)); 353 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) 354 return SelectInst::Create(X, Op0, ConstantInt::get(I.getType(), 0)); 355 356 // (sext bool X) * C --> X ? -C : 0 357 Constant *ImmC; 358 if (match(Op0, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1) && 359 match(Op1, m_ImmConstant(ImmC))) { 360 Constant *NegC = ConstantExpr::getNeg(ImmC); 361 return SelectInst::Create(X, NegC, ConstantInt::getNullValue(I.getType())); 362 } 363 364 // (lshr X, 31) * Y --> (ashr X, 31) & Y 365 // Y * (lshr X, 31) --> (ashr X, 31) & Y 366 // TODO: We are not checking one-use because the elimination of the multiply 367 // is better for analysis? 368 // TODO: Should we canonicalize to '(X < 0) ? Y : 0' instead? That would be 369 // more similar to what we're doing above. 370 const APInt *C; 371 if (match(Op0, m_LShr(m_Value(X), m_APInt(C))) && *C == C->getBitWidth() - 1) 372 return BinaryOperator::CreateAnd(Builder.CreateAShr(X, *C), Op1); 373 if (match(Op1, m_LShr(m_Value(X), m_APInt(C))) && *C == C->getBitWidth() - 1) 374 return BinaryOperator::CreateAnd(Builder.CreateAShr(X, *C), Op0); 375 376 // ((ashr X, 31) | 1) * X --> abs(X) 377 // X * ((ashr X, 31) | 1) --> abs(X) 378 if (match(&I, m_c_BinOp(m_Or(m_AShr(m_Value(X), 379 m_SpecificIntAllowUndef(BitWidth - 1)), 380 m_One()), 381 m_Deferred(X)))) { 382 Value *Abs = Builder.CreateBinaryIntrinsic( 383 Intrinsic::abs, X, 384 ConstantInt::getBool(I.getContext(), I.hasNoSignedWrap())); 385 Abs->takeName(&I); 386 return replaceInstUsesWith(I, Abs); 387 } 388 389 if (Instruction *Ext = narrowMathIfNoOverflow(I)) 390 return Ext; 391 392 bool Changed = false; 393 if (!I.hasNoSignedWrap() && willNotOverflowSignedMul(Op0, Op1, I)) { 394 Changed = true; 395 I.setHasNoSignedWrap(true); 396 } 397 398 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedMul(Op0, Op1, I)) { 399 Changed = true; 400 I.setHasNoUnsignedWrap(true); 401 } 402 403 return Changed ? &I : nullptr; 404 } 405 406 Instruction *InstCombinerImpl::foldFPSignBitOps(BinaryOperator &I) { 407 BinaryOperator::BinaryOps Opcode = I.getOpcode(); 408 assert((Opcode == Instruction::FMul || Opcode == Instruction::FDiv) && 409 "Expected fmul or fdiv"); 410 411 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 412 Value *X, *Y; 413 414 // -X * -Y --> X * Y 415 // -X / -Y --> X / Y 416 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y)))) 417 return BinaryOperator::CreateWithCopiedFlags(Opcode, X, Y, &I); 418 419 // fabs(X) * fabs(X) -> X * X 420 // fabs(X) / fabs(X) -> X / X 421 if (Op0 == Op1 && match(Op0, m_FAbs(m_Value(X)))) 422 return BinaryOperator::CreateWithCopiedFlags(Opcode, X, X, &I); 423 424 // fabs(X) * fabs(Y) --> fabs(X * Y) 425 // fabs(X) / fabs(Y) --> fabs(X / Y) 426 if (match(Op0, m_FAbs(m_Value(X))) && match(Op1, m_FAbs(m_Value(Y))) && 427 (Op0->hasOneUse() || Op1->hasOneUse())) { 428 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); 429 Builder.setFastMathFlags(I.getFastMathFlags()); 430 Value *XY = Builder.CreateBinOp(Opcode, X, Y); 431 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, XY); 432 Fabs->takeName(&I); 433 return replaceInstUsesWith(I, Fabs); 434 } 435 436 return nullptr; 437 } 438 439 Instruction *InstCombinerImpl::visitFMul(BinaryOperator &I) { 440 if (Value *V = SimplifyFMulInst(I.getOperand(0), I.getOperand(1), 441 I.getFastMathFlags(), 442 SQ.getWithInstruction(&I))) 443 return replaceInstUsesWith(I, V); 444 445 if (SimplifyAssociativeOrCommutative(I)) 446 return &I; 447 448 if (Instruction *X = foldVectorBinop(I)) 449 return X; 450 451 if (Instruction *Phi = foldBinopWithPhiOperands(I)) 452 return Phi; 453 454 if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I)) 455 return FoldedMul; 456 457 if (Value *FoldedMul = foldMulSelectToNegate(I, Builder)) 458 return replaceInstUsesWith(I, FoldedMul); 459 460 if (Instruction *R = foldFPSignBitOps(I)) 461 return R; 462 463 // X * -1.0 --> -X 464 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 465 if (match(Op1, m_SpecificFP(-1.0))) 466 return UnaryOperator::CreateFNegFMF(Op0, &I); 467 468 // -X * C --> X * -C 469 Value *X, *Y; 470 Constant *C; 471 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_Constant(C))) 472 return BinaryOperator::CreateFMulFMF(X, ConstantExpr::getFNeg(C), &I); 473 474 // (select A, B, C) * (select A, D, E) --> select A, (B*D), (C*E) 475 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1)) 476 return replaceInstUsesWith(I, V); 477 478 if (I.hasAllowReassoc()) { 479 // Reassociate constant RHS with another constant to form constant 480 // expression. 481 if (match(Op1, m_Constant(C)) && C->isFiniteNonZeroFP()) { 482 Constant *C1; 483 if (match(Op0, m_OneUse(m_FDiv(m_Constant(C1), m_Value(X))))) { 484 // (C1 / X) * C --> (C * C1) / X 485 Constant *CC1 = ConstantExpr::getFMul(C, C1); 486 if (CC1->isNormalFP()) 487 return BinaryOperator::CreateFDivFMF(CC1, X, &I); 488 } 489 if (match(Op0, m_FDiv(m_Value(X), m_Constant(C1)))) { 490 // (X / C1) * C --> X * (C / C1) 491 Constant *CDivC1 = ConstantExpr::getFDiv(C, C1); 492 if (CDivC1->isNormalFP()) 493 return BinaryOperator::CreateFMulFMF(X, CDivC1, &I); 494 495 // If the constant was a denormal, try reassociating differently. 496 // (X / C1) * C --> X / (C1 / C) 497 Constant *C1DivC = ConstantExpr::getFDiv(C1, C); 498 if (Op0->hasOneUse() && C1DivC->isNormalFP()) 499 return BinaryOperator::CreateFDivFMF(X, C1DivC, &I); 500 } 501 502 // We do not need to match 'fadd C, X' and 'fsub X, C' because they are 503 // canonicalized to 'fadd X, C'. Distributing the multiply may allow 504 // further folds and (X * C) + C2 is 'fma'. 505 if (match(Op0, m_OneUse(m_FAdd(m_Value(X), m_Constant(C1))))) { 506 // (X + C1) * C --> (X * C) + (C * C1) 507 Constant *CC1 = ConstantExpr::getFMul(C, C1); 508 Value *XC = Builder.CreateFMulFMF(X, C, &I); 509 return BinaryOperator::CreateFAddFMF(XC, CC1, &I); 510 } 511 if (match(Op0, m_OneUse(m_FSub(m_Constant(C1), m_Value(X))))) { 512 // (C1 - X) * C --> (C * C1) - (X * C) 513 Constant *CC1 = ConstantExpr::getFMul(C, C1); 514 Value *XC = Builder.CreateFMulFMF(X, C, &I); 515 return BinaryOperator::CreateFSubFMF(CC1, XC, &I); 516 } 517 } 518 519 Value *Z; 520 if (match(&I, m_c_FMul(m_OneUse(m_FDiv(m_Value(X), m_Value(Y))), 521 m_Value(Z)))) { 522 // Sink division: (X / Y) * Z --> (X * Z) / Y 523 Value *NewFMul = Builder.CreateFMulFMF(X, Z, &I); 524 return BinaryOperator::CreateFDivFMF(NewFMul, Y, &I); 525 } 526 527 // sqrt(X) * sqrt(Y) -> sqrt(X * Y) 528 // nnan disallows the possibility of returning a number if both operands are 529 // negative (in that case, we should return NaN). 530 if (I.hasNoNaNs() && 531 match(Op0, m_OneUse(m_Intrinsic<Intrinsic::sqrt>(m_Value(X)))) && 532 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::sqrt>(m_Value(Y))))) { 533 Value *XY = Builder.CreateFMulFMF(X, Y, &I); 534 Value *Sqrt = Builder.CreateUnaryIntrinsic(Intrinsic::sqrt, XY, &I); 535 return replaceInstUsesWith(I, Sqrt); 536 } 537 538 // The following transforms are done irrespective of the number of uses 539 // for the expression "1.0/sqrt(X)". 540 // 1) 1.0/sqrt(X) * X -> X/sqrt(X) 541 // 2) X * 1.0/sqrt(X) -> X/sqrt(X) 542 // We always expect the backend to reduce X/sqrt(X) to sqrt(X), if it 543 // has the necessary (reassoc) fast-math-flags. 544 if (I.hasNoSignedZeros() && 545 match(Op0, (m_FDiv(m_SpecificFP(1.0), m_Value(Y)))) && 546 match(Y, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) && Op1 == X) 547 return BinaryOperator::CreateFDivFMF(X, Y, &I); 548 if (I.hasNoSignedZeros() && 549 match(Op1, (m_FDiv(m_SpecificFP(1.0), m_Value(Y)))) && 550 match(Y, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) && Op0 == X) 551 return BinaryOperator::CreateFDivFMF(X, Y, &I); 552 553 // Like the similar transform in instsimplify, this requires 'nsz' because 554 // sqrt(-0.0) = -0.0, and -0.0 * -0.0 does not simplify to -0.0. 555 if (I.hasNoNaNs() && I.hasNoSignedZeros() && Op0 == Op1 && 556 Op0->hasNUses(2)) { 557 // Peek through fdiv to find squaring of square root: 558 // (X / sqrt(Y)) * (X / sqrt(Y)) --> (X * X) / Y 559 if (match(Op0, m_FDiv(m_Value(X), 560 m_Intrinsic<Intrinsic::sqrt>(m_Value(Y))))) { 561 Value *XX = Builder.CreateFMulFMF(X, X, &I); 562 return BinaryOperator::CreateFDivFMF(XX, Y, &I); 563 } 564 // (sqrt(Y) / X) * (sqrt(Y) / X) --> Y / (X * X) 565 if (match(Op0, m_FDiv(m_Intrinsic<Intrinsic::sqrt>(m_Value(Y)), 566 m_Value(X)))) { 567 Value *XX = Builder.CreateFMulFMF(X, X, &I); 568 return BinaryOperator::CreateFDivFMF(Y, XX, &I); 569 } 570 } 571 572 if (I.isOnlyUserOfAnyOperand()) { 573 // pow(x, y) * pow(x, z) -> pow(x, y + z) 574 if (match(Op0, m_Intrinsic<Intrinsic::pow>(m_Value(X), m_Value(Y))) && 575 match(Op1, m_Intrinsic<Intrinsic::pow>(m_Specific(X), m_Value(Z)))) { 576 auto *YZ = Builder.CreateFAddFMF(Y, Z, &I); 577 auto *NewPow = Builder.CreateBinaryIntrinsic(Intrinsic::pow, X, YZ, &I); 578 return replaceInstUsesWith(I, NewPow); 579 } 580 581 // powi(x, y) * powi(x, z) -> powi(x, y + z) 582 if (match(Op0, m_Intrinsic<Intrinsic::powi>(m_Value(X), m_Value(Y))) && 583 match(Op1, m_Intrinsic<Intrinsic::powi>(m_Specific(X), m_Value(Z))) && 584 Y->getType() == Z->getType()) { 585 auto *YZ = Builder.CreateAdd(Y, Z); 586 auto *NewPow = Builder.CreateIntrinsic( 587 Intrinsic::powi, {X->getType(), YZ->getType()}, {X, YZ}, &I); 588 return replaceInstUsesWith(I, NewPow); 589 } 590 591 // exp(X) * exp(Y) -> exp(X + Y) 592 if (match(Op0, m_Intrinsic<Intrinsic::exp>(m_Value(X))) && 593 match(Op1, m_Intrinsic<Intrinsic::exp>(m_Value(Y)))) { 594 Value *XY = Builder.CreateFAddFMF(X, Y, &I); 595 Value *Exp = Builder.CreateUnaryIntrinsic(Intrinsic::exp, XY, &I); 596 return replaceInstUsesWith(I, Exp); 597 } 598 599 // exp2(X) * exp2(Y) -> exp2(X + Y) 600 if (match(Op0, m_Intrinsic<Intrinsic::exp2>(m_Value(X))) && 601 match(Op1, m_Intrinsic<Intrinsic::exp2>(m_Value(Y)))) { 602 Value *XY = Builder.CreateFAddFMF(X, Y, &I); 603 Value *Exp2 = Builder.CreateUnaryIntrinsic(Intrinsic::exp2, XY, &I); 604 return replaceInstUsesWith(I, Exp2); 605 } 606 } 607 608 // (X*Y) * X => (X*X) * Y where Y != X 609 // The purpose is two-fold: 610 // 1) to form a power expression (of X). 611 // 2) potentially shorten the critical path: After transformation, the 612 // latency of the instruction Y is amortized by the expression of X*X, 613 // and therefore Y is in a "less critical" position compared to what it 614 // was before the transformation. 615 if (match(Op0, m_OneUse(m_c_FMul(m_Specific(Op1), m_Value(Y)))) && 616 Op1 != Y) { 617 Value *XX = Builder.CreateFMulFMF(Op1, Op1, &I); 618 return BinaryOperator::CreateFMulFMF(XX, Y, &I); 619 } 620 if (match(Op1, m_OneUse(m_c_FMul(m_Specific(Op0), m_Value(Y)))) && 621 Op0 != Y) { 622 Value *XX = Builder.CreateFMulFMF(Op0, Op0, &I); 623 return BinaryOperator::CreateFMulFMF(XX, Y, &I); 624 } 625 } 626 627 // log2(X * 0.5) * Y = log2(X) * Y - Y 628 if (I.isFast()) { 629 IntrinsicInst *Log2 = nullptr; 630 if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::log2>( 631 m_OneUse(m_FMul(m_Value(X), m_SpecificFP(0.5))))))) { 632 Log2 = cast<IntrinsicInst>(Op0); 633 Y = Op1; 634 } 635 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::log2>( 636 m_OneUse(m_FMul(m_Value(X), m_SpecificFP(0.5))))))) { 637 Log2 = cast<IntrinsicInst>(Op1); 638 Y = Op0; 639 } 640 if (Log2) { 641 Value *Log2 = Builder.CreateUnaryIntrinsic(Intrinsic::log2, X, &I); 642 Value *LogXTimesY = Builder.CreateFMulFMF(Log2, Y, &I); 643 return BinaryOperator::CreateFSubFMF(LogXTimesY, Y, &I); 644 } 645 } 646 647 return nullptr; 648 } 649 650 /// Fold a divide or remainder with a select instruction divisor when one of the 651 /// select operands is zero. In that case, we can use the other select operand 652 /// because div/rem by zero is undefined. 653 bool InstCombinerImpl::simplifyDivRemOfSelectWithZeroOp(BinaryOperator &I) { 654 SelectInst *SI = dyn_cast<SelectInst>(I.getOperand(1)); 655 if (!SI) 656 return false; 657 658 int NonNullOperand; 659 if (match(SI->getTrueValue(), m_Zero())) 660 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y 661 NonNullOperand = 2; 662 else if (match(SI->getFalseValue(), m_Zero())) 663 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y 664 NonNullOperand = 1; 665 else 666 return false; 667 668 // Change the div/rem to use 'Y' instead of the select. 669 replaceOperand(I, 1, SI->getOperand(NonNullOperand)); 670 671 // Okay, we know we replace the operand of the div/rem with 'Y' with no 672 // problem. However, the select, or the condition of the select may have 673 // multiple uses. Based on our knowledge that the operand must be non-zero, 674 // propagate the known value for the select into other uses of it, and 675 // propagate a known value of the condition into its other users. 676 677 // If the select and condition only have a single use, don't bother with this, 678 // early exit. 679 Value *SelectCond = SI->getCondition(); 680 if (SI->use_empty() && SelectCond->hasOneUse()) 681 return true; 682 683 // Scan the current block backward, looking for other uses of SI. 684 BasicBlock::iterator BBI = I.getIterator(), BBFront = I.getParent()->begin(); 685 Type *CondTy = SelectCond->getType(); 686 while (BBI != BBFront) { 687 --BBI; 688 // If we found an instruction that we can't assume will return, so 689 // information from below it cannot be propagated above it. 690 if (!isGuaranteedToTransferExecutionToSuccessor(&*BBI)) 691 break; 692 693 // Replace uses of the select or its condition with the known values. 694 for (Use &Op : BBI->operands()) { 695 if (Op == SI) { 696 replaceUse(Op, SI->getOperand(NonNullOperand)); 697 Worklist.push(&*BBI); 698 } else if (Op == SelectCond) { 699 replaceUse(Op, NonNullOperand == 1 ? ConstantInt::getTrue(CondTy) 700 : ConstantInt::getFalse(CondTy)); 701 Worklist.push(&*BBI); 702 } 703 } 704 705 // If we past the instruction, quit looking for it. 706 if (&*BBI == SI) 707 SI = nullptr; 708 if (&*BBI == SelectCond) 709 SelectCond = nullptr; 710 711 // If we ran out of things to eliminate, break out of the loop. 712 if (!SelectCond && !SI) 713 break; 714 715 } 716 return true; 717 } 718 719 /// True if the multiply can not be expressed in an int this size. 720 static bool multiplyOverflows(const APInt &C1, const APInt &C2, APInt &Product, 721 bool IsSigned) { 722 bool Overflow; 723 Product = IsSigned ? C1.smul_ov(C2, Overflow) : C1.umul_ov(C2, Overflow); 724 return Overflow; 725 } 726 727 /// True if C1 is a multiple of C2. Quotient contains C1/C2. 728 static bool isMultiple(const APInt &C1, const APInt &C2, APInt &Quotient, 729 bool IsSigned) { 730 assert(C1.getBitWidth() == C2.getBitWidth() && "Constant widths not equal"); 731 732 // Bail if we will divide by zero. 733 if (C2.isZero()) 734 return false; 735 736 // Bail if we would divide INT_MIN by -1. 737 if (IsSigned && C1.isMinSignedValue() && C2.isAllOnes()) 738 return false; 739 740 APInt Remainder(C1.getBitWidth(), /*val=*/0ULL, IsSigned); 741 if (IsSigned) 742 APInt::sdivrem(C1, C2, Quotient, Remainder); 743 else 744 APInt::udivrem(C1, C2, Quotient, Remainder); 745 746 return Remainder.isMinValue(); 747 } 748 749 /// This function implements the transforms common to both integer division 750 /// instructions (udiv and sdiv). It is called by the visitors to those integer 751 /// division instructions. 752 /// Common integer divide transforms 753 Instruction *InstCombinerImpl::commonIDivTransforms(BinaryOperator &I) { 754 if (Instruction *Phi = foldBinopWithPhiOperands(I)) 755 return Phi; 756 757 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 758 bool IsSigned = I.getOpcode() == Instruction::SDiv; 759 Type *Ty = I.getType(); 760 761 // The RHS is known non-zero. 762 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I)) 763 return replaceOperand(I, 1, V); 764 765 // Handle cases involving: [su]div X, (select Cond, Y, Z) 766 // This does not apply for fdiv. 767 if (simplifyDivRemOfSelectWithZeroOp(I)) 768 return &I; 769 770 // If the divisor is a select-of-constants, try to constant fold all div ops: 771 // C / (select Cond, TrueC, FalseC) --> select Cond, (C / TrueC), (C / FalseC) 772 // TODO: Adapt simplifyDivRemOfSelectWithZeroOp to allow this and other folds. 773 if (match(Op0, m_ImmConstant()) && 774 match(Op1, m_Select(m_Value(), m_ImmConstant(), m_ImmConstant()))) { 775 if (Instruction *R = FoldOpIntoSelect(I, cast<SelectInst>(Op1))) 776 return R; 777 } 778 779 const APInt *C2; 780 if (match(Op1, m_APInt(C2))) { 781 Value *X; 782 const APInt *C1; 783 784 // (X / C1) / C2 -> X / (C1*C2) 785 if ((IsSigned && match(Op0, m_SDiv(m_Value(X), m_APInt(C1)))) || 786 (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_APInt(C1))))) { 787 APInt Product(C1->getBitWidth(), /*val=*/0ULL, IsSigned); 788 if (!multiplyOverflows(*C1, *C2, Product, IsSigned)) 789 return BinaryOperator::Create(I.getOpcode(), X, 790 ConstantInt::get(Ty, Product)); 791 } 792 793 if ((IsSigned && match(Op0, m_NSWMul(m_Value(X), m_APInt(C1)))) || 794 (!IsSigned && match(Op0, m_NUWMul(m_Value(X), m_APInt(C1))))) { 795 APInt Quotient(C1->getBitWidth(), /*val=*/0ULL, IsSigned); 796 797 // (X * C1) / C2 -> X / (C2 / C1) if C2 is a multiple of C1. 798 if (isMultiple(*C2, *C1, Quotient, IsSigned)) { 799 auto *NewDiv = BinaryOperator::Create(I.getOpcode(), X, 800 ConstantInt::get(Ty, Quotient)); 801 NewDiv->setIsExact(I.isExact()); 802 return NewDiv; 803 } 804 805 // (X * C1) / C2 -> X * (C1 / C2) if C1 is a multiple of C2. 806 if (isMultiple(*C1, *C2, Quotient, IsSigned)) { 807 auto *Mul = BinaryOperator::Create(Instruction::Mul, X, 808 ConstantInt::get(Ty, Quotient)); 809 auto *OBO = cast<OverflowingBinaryOperator>(Op0); 810 Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap()); 811 Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap()); 812 return Mul; 813 } 814 } 815 816 if ((IsSigned && match(Op0, m_NSWShl(m_Value(X), m_APInt(C1))) && 817 C1->ult(C1->getBitWidth() - 1)) || 818 (!IsSigned && match(Op0, m_NUWShl(m_Value(X), m_APInt(C1))) && 819 C1->ult(C1->getBitWidth()))) { 820 APInt Quotient(C1->getBitWidth(), /*val=*/0ULL, IsSigned); 821 APInt C1Shifted = APInt::getOneBitSet( 822 C1->getBitWidth(), static_cast<unsigned>(C1->getZExtValue())); 823 824 // (X << C1) / C2 -> X / (C2 >> C1) if C2 is a multiple of 1 << C1. 825 if (isMultiple(*C2, C1Shifted, Quotient, IsSigned)) { 826 auto *BO = BinaryOperator::Create(I.getOpcode(), X, 827 ConstantInt::get(Ty, Quotient)); 828 BO->setIsExact(I.isExact()); 829 return BO; 830 } 831 832 // (X << C1) / C2 -> X * ((1 << C1) / C2) if 1 << C1 is a multiple of C2. 833 if (isMultiple(C1Shifted, *C2, Quotient, IsSigned)) { 834 auto *Mul = BinaryOperator::Create(Instruction::Mul, X, 835 ConstantInt::get(Ty, Quotient)); 836 auto *OBO = cast<OverflowingBinaryOperator>(Op0); 837 Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap()); 838 Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap()); 839 return Mul; 840 } 841 } 842 843 if (!C2->isZero()) // avoid X udiv 0 844 if (Instruction *FoldedDiv = foldBinOpIntoSelectOrPhi(I)) 845 return FoldedDiv; 846 } 847 848 if (match(Op0, m_One())) { 849 assert(!Ty->isIntOrIntVectorTy(1) && "i1 divide not removed?"); 850 if (IsSigned) { 851 // If Op1 is 0 then it's undefined behaviour, if Op1 is 1 then the 852 // result is one, if Op1 is -1 then the result is minus one, otherwise 853 // it's zero. 854 Value *Inc = Builder.CreateAdd(Op1, Op0); 855 Value *Cmp = Builder.CreateICmpULT(Inc, ConstantInt::get(Ty, 3)); 856 return SelectInst::Create(Cmp, Op1, ConstantInt::get(Ty, 0)); 857 } else { 858 // If Op1 is 0 then it's undefined behaviour. If Op1 is 1 then the 859 // result is one, otherwise it's zero. 860 return new ZExtInst(Builder.CreateICmpEQ(Op1, Op0), Ty); 861 } 862 } 863 864 // See if we can fold away this div instruction. 865 if (SimplifyDemandedInstructionBits(I)) 866 return &I; 867 868 // (X - (X rem Y)) / Y -> X / Y; usually originates as ((X / Y) * Y) / Y 869 Value *X, *Z; 870 if (match(Op0, m_Sub(m_Value(X), m_Value(Z)))) // (X - Z) / Y; Y = Op1 871 if ((IsSigned && match(Z, m_SRem(m_Specific(X), m_Specific(Op1)))) || 872 (!IsSigned && match(Z, m_URem(m_Specific(X), m_Specific(Op1))))) 873 return BinaryOperator::Create(I.getOpcode(), X, Op1); 874 875 // (X << Y) / X -> 1 << Y 876 Value *Y; 877 if (IsSigned && match(Op0, m_NSWShl(m_Specific(Op1), m_Value(Y)))) 878 return BinaryOperator::CreateNSWShl(ConstantInt::get(Ty, 1), Y); 879 if (!IsSigned && match(Op0, m_NUWShl(m_Specific(Op1), m_Value(Y)))) 880 return BinaryOperator::CreateNUWShl(ConstantInt::get(Ty, 1), Y); 881 882 // X / (X * Y) -> 1 / Y if the multiplication does not overflow. 883 if (match(Op1, m_c_Mul(m_Specific(Op0), m_Value(Y)))) { 884 bool HasNSW = cast<OverflowingBinaryOperator>(Op1)->hasNoSignedWrap(); 885 bool HasNUW = cast<OverflowingBinaryOperator>(Op1)->hasNoUnsignedWrap(); 886 if ((IsSigned && HasNSW) || (!IsSigned && HasNUW)) { 887 replaceOperand(I, 0, ConstantInt::get(Ty, 1)); 888 replaceOperand(I, 1, Y); 889 return &I; 890 } 891 } 892 893 return nullptr; 894 } 895 896 static const unsigned MaxDepth = 6; 897 898 // Take the exact integer log2 of the value. If DoFold is true, create the 899 // actual instructions, otherwise return a non-null dummy value. Return nullptr 900 // on failure. 901 static Value *takeLog2(IRBuilderBase &Builder, Value *Op, unsigned Depth, 902 bool DoFold) { 903 auto IfFold = [DoFold](function_ref<Value *()> Fn) { 904 if (!DoFold) 905 return reinterpret_cast<Value *>(-1); 906 return Fn(); 907 }; 908 909 // FIXME: assert that Op1 isn't/doesn't contain undef. 910 911 // log2(2^C) -> C 912 if (match(Op, m_Power2())) 913 return IfFold([&]() { 914 Constant *C = ConstantExpr::getExactLogBase2(cast<Constant>(Op)); 915 if (!C) 916 llvm_unreachable("Failed to constant fold udiv -> logbase2"); 917 return C; 918 }); 919 920 // The remaining tests are all recursive, so bail out if we hit the limit. 921 if (Depth++ == MaxDepth) 922 return nullptr; 923 924 // log2(zext X) -> zext log2(X) 925 // FIXME: Require one use? 926 Value *X, *Y; 927 if (match(Op, m_ZExt(m_Value(X)))) 928 if (Value *LogX = takeLog2(Builder, X, Depth, DoFold)) 929 return IfFold([&]() { return Builder.CreateZExt(LogX, Op->getType()); }); 930 931 // log2(X << Y) -> log2(X) + Y 932 // FIXME: Require one use unless X is 1? 933 if (match(Op, m_Shl(m_Value(X), m_Value(Y)))) 934 if (Value *LogX = takeLog2(Builder, X, Depth, DoFold)) 935 return IfFold([&]() { return Builder.CreateAdd(LogX, Y); }); 936 937 // log2(Cond ? X : Y) -> Cond ? log2(X) : log2(Y) 938 // FIXME: missed optimization: if one of the hands of select is/contains 939 // undef, just directly pick the other one. 940 // FIXME: can both hands contain undef? 941 // FIXME: Require one use? 942 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) 943 if (Value *LogX = takeLog2(Builder, SI->getOperand(1), Depth, DoFold)) 944 if (Value *LogY = takeLog2(Builder, SI->getOperand(2), Depth, DoFold)) 945 return IfFold([&]() { 946 return Builder.CreateSelect(SI->getOperand(0), LogX, LogY); 947 }); 948 949 // log2(umin(X, Y)) -> umin(log2(X), log2(Y)) 950 // log2(umax(X, Y)) -> umax(log2(X), log2(Y)) 951 auto *MinMax = dyn_cast<MinMaxIntrinsic>(Op); 952 if (MinMax && MinMax->hasOneUse() && !MinMax->isSigned()) 953 if (Value *LogX = takeLog2(Builder, MinMax->getLHS(), Depth, DoFold)) 954 if (Value *LogY = takeLog2(Builder, MinMax->getRHS(), Depth, DoFold)) 955 return IfFold([&]() { 956 return Builder.CreateBinaryIntrinsic( 957 MinMax->getIntrinsicID(), LogX, LogY); 958 }); 959 960 return nullptr; 961 } 962 963 /// If we have zero-extended operands of an unsigned div or rem, we may be able 964 /// to narrow the operation (sink the zext below the math). 965 static Instruction *narrowUDivURem(BinaryOperator &I, 966 InstCombiner::BuilderTy &Builder) { 967 Instruction::BinaryOps Opcode = I.getOpcode(); 968 Value *N = I.getOperand(0); 969 Value *D = I.getOperand(1); 970 Type *Ty = I.getType(); 971 Value *X, *Y; 972 if (match(N, m_ZExt(m_Value(X))) && match(D, m_ZExt(m_Value(Y))) && 973 X->getType() == Y->getType() && (N->hasOneUse() || D->hasOneUse())) { 974 // udiv (zext X), (zext Y) --> zext (udiv X, Y) 975 // urem (zext X), (zext Y) --> zext (urem X, Y) 976 Value *NarrowOp = Builder.CreateBinOp(Opcode, X, Y); 977 return new ZExtInst(NarrowOp, Ty); 978 } 979 980 Constant *C; 981 if ((match(N, m_OneUse(m_ZExt(m_Value(X)))) && match(D, m_Constant(C))) || 982 (match(D, m_OneUse(m_ZExt(m_Value(X)))) && match(N, m_Constant(C)))) { 983 // If the constant is the same in the smaller type, use the narrow version. 984 Constant *TruncC = ConstantExpr::getTrunc(C, X->getType()); 985 if (ConstantExpr::getZExt(TruncC, Ty) != C) 986 return nullptr; 987 988 // udiv (zext X), C --> zext (udiv X, C') 989 // urem (zext X), C --> zext (urem X, C') 990 // udiv C, (zext X) --> zext (udiv C', X) 991 // urem C, (zext X) --> zext (urem C', X) 992 Value *NarrowOp = isa<Constant>(D) ? Builder.CreateBinOp(Opcode, X, TruncC) 993 : Builder.CreateBinOp(Opcode, TruncC, X); 994 return new ZExtInst(NarrowOp, Ty); 995 } 996 997 return nullptr; 998 } 999 1000 Instruction *InstCombinerImpl::visitUDiv(BinaryOperator &I) { 1001 if (Value *V = SimplifyUDivInst(I.getOperand(0), I.getOperand(1), 1002 SQ.getWithInstruction(&I))) 1003 return replaceInstUsesWith(I, V); 1004 1005 if (Instruction *X = foldVectorBinop(I)) 1006 return X; 1007 1008 // Handle the integer div common cases 1009 if (Instruction *Common = commonIDivTransforms(I)) 1010 return Common; 1011 1012 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1013 Value *X; 1014 const APInt *C1, *C2; 1015 if (match(Op0, m_LShr(m_Value(X), m_APInt(C1))) && match(Op1, m_APInt(C2))) { 1016 // (X lshr C1) udiv C2 --> X udiv (C2 << C1) 1017 bool Overflow; 1018 APInt C2ShlC1 = C2->ushl_ov(*C1, Overflow); 1019 if (!Overflow) { 1020 bool IsExact = I.isExact() && match(Op0, m_Exact(m_Value())); 1021 BinaryOperator *BO = BinaryOperator::CreateUDiv( 1022 X, ConstantInt::get(X->getType(), C2ShlC1)); 1023 if (IsExact) 1024 BO->setIsExact(); 1025 return BO; 1026 } 1027 } 1028 1029 // Op0 / C where C is large (negative) --> zext (Op0 >= C) 1030 // TODO: Could use isKnownNegative() to handle non-constant values. 1031 Type *Ty = I.getType(); 1032 if (match(Op1, m_Negative())) { 1033 Value *Cmp = Builder.CreateICmpUGE(Op0, Op1); 1034 return CastInst::CreateZExtOrBitCast(Cmp, Ty); 1035 } 1036 // Op0 / (sext i1 X) --> zext (Op0 == -1) (if X is 0, the div is undefined) 1037 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) { 1038 Value *Cmp = Builder.CreateICmpEQ(Op0, ConstantInt::getAllOnesValue(Ty)); 1039 return CastInst::CreateZExtOrBitCast(Cmp, Ty); 1040 } 1041 1042 if (Instruction *NarrowDiv = narrowUDivURem(I, Builder)) 1043 return NarrowDiv; 1044 1045 // If the udiv operands are non-overflowing multiplies with a common operand, 1046 // then eliminate the common factor: 1047 // (A * B) / (A * X) --> B / X (and commuted variants) 1048 // TODO: The code would be reduced if we had m_c_NUWMul pattern matching. 1049 // TODO: If -reassociation handled this generally, we could remove this. 1050 Value *A, *B; 1051 if (match(Op0, m_NUWMul(m_Value(A), m_Value(B)))) { 1052 if (match(Op1, m_NUWMul(m_Specific(A), m_Value(X))) || 1053 match(Op1, m_NUWMul(m_Value(X), m_Specific(A)))) 1054 return BinaryOperator::CreateUDiv(B, X); 1055 if (match(Op1, m_NUWMul(m_Specific(B), m_Value(X))) || 1056 match(Op1, m_NUWMul(m_Value(X), m_Specific(B)))) 1057 return BinaryOperator::CreateUDiv(A, X); 1058 } 1059 1060 // Op1 udiv Op2 -> Op1 lshr log2(Op2), if log2() folds away. 1061 if (takeLog2(Builder, Op1, /*Depth*/0, /*DoFold*/false)) { 1062 Value *Res = takeLog2(Builder, Op1, /*Depth*/0, /*DoFold*/true); 1063 return replaceInstUsesWith( 1064 I, Builder.CreateLShr(Op0, Res, I.getName(), I.isExact())); 1065 } 1066 1067 return nullptr; 1068 } 1069 1070 Instruction *InstCombinerImpl::visitSDiv(BinaryOperator &I) { 1071 if (Value *V = SimplifySDivInst(I.getOperand(0), I.getOperand(1), 1072 SQ.getWithInstruction(&I))) 1073 return replaceInstUsesWith(I, V); 1074 1075 if (Instruction *X = foldVectorBinop(I)) 1076 return X; 1077 1078 // Handle the integer div common cases 1079 if (Instruction *Common = commonIDivTransforms(I)) 1080 return Common; 1081 1082 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1083 Type *Ty = I.getType(); 1084 Value *X; 1085 // sdiv Op0, -1 --> -Op0 1086 // sdiv Op0, (sext i1 X) --> -Op0 (because if X is 0, the op is undefined) 1087 if (match(Op1, m_AllOnes()) || 1088 (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))) 1089 return BinaryOperator::CreateNeg(Op0); 1090 1091 // X / INT_MIN --> X == INT_MIN 1092 if (match(Op1, m_SignMask())) 1093 return new ZExtInst(Builder.CreateICmpEQ(Op0, Op1), Ty); 1094 1095 // sdiv exact X, 1<<C --> ashr exact X, C iff 1<<C is non-negative 1096 // sdiv exact X, -1<<C --> -(ashr exact X, C) 1097 if (I.isExact() && ((match(Op1, m_Power2()) && match(Op1, m_NonNegative())) || 1098 match(Op1, m_NegatedPower2()))) { 1099 bool DivisorWasNegative = match(Op1, m_NegatedPower2()); 1100 if (DivisorWasNegative) 1101 Op1 = ConstantExpr::getNeg(cast<Constant>(Op1)); 1102 auto *AShr = BinaryOperator::CreateExactAShr( 1103 Op0, ConstantExpr::getExactLogBase2(cast<Constant>(Op1)), I.getName()); 1104 if (!DivisorWasNegative) 1105 return AShr; 1106 Builder.Insert(AShr); 1107 AShr->setName(I.getName() + ".neg"); 1108 return BinaryOperator::CreateNeg(AShr, I.getName()); 1109 } 1110 1111 const APInt *Op1C; 1112 if (match(Op1, m_APInt(Op1C))) { 1113 // If the dividend is sign-extended and the constant divisor is small enough 1114 // to fit in the source type, shrink the division to the narrower type: 1115 // (sext X) sdiv C --> sext (X sdiv C) 1116 Value *Op0Src; 1117 if (match(Op0, m_OneUse(m_SExt(m_Value(Op0Src)))) && 1118 Op0Src->getType()->getScalarSizeInBits() >= Op1C->getMinSignedBits()) { 1119 1120 // In the general case, we need to make sure that the dividend is not the 1121 // minimum signed value because dividing that by -1 is UB. But here, we 1122 // know that the -1 divisor case is already handled above. 1123 1124 Constant *NarrowDivisor = 1125 ConstantExpr::getTrunc(cast<Constant>(Op1), Op0Src->getType()); 1126 Value *NarrowOp = Builder.CreateSDiv(Op0Src, NarrowDivisor); 1127 return new SExtInst(NarrowOp, Ty); 1128 } 1129 1130 // -X / C --> X / -C (if the negation doesn't overflow). 1131 // TODO: This could be enhanced to handle arbitrary vector constants by 1132 // checking if all elements are not the min-signed-val. 1133 if (!Op1C->isMinSignedValue() && 1134 match(Op0, m_NSWSub(m_Zero(), m_Value(X)))) { 1135 Constant *NegC = ConstantInt::get(Ty, -(*Op1C)); 1136 Instruction *BO = BinaryOperator::CreateSDiv(X, NegC); 1137 BO->setIsExact(I.isExact()); 1138 return BO; 1139 } 1140 } 1141 1142 // -X / Y --> -(X / Y) 1143 Value *Y; 1144 if (match(&I, m_SDiv(m_OneUse(m_NSWSub(m_Zero(), m_Value(X))), m_Value(Y)))) 1145 return BinaryOperator::CreateNSWNeg( 1146 Builder.CreateSDiv(X, Y, I.getName(), I.isExact())); 1147 1148 // abs(X) / X --> X > -1 ? 1 : -1 1149 // X / abs(X) --> X > -1 ? 1 : -1 1150 if (match(&I, m_c_BinOp( 1151 m_OneUse(m_Intrinsic<Intrinsic::abs>(m_Value(X), m_One())), 1152 m_Deferred(X)))) { 1153 Constant *NegOne = ConstantInt::getAllOnesValue(Ty); 1154 Value *Cond = Builder.CreateICmpSGT(X, NegOne); 1155 return SelectInst::Create(Cond, ConstantInt::get(Ty, 1), NegOne); 1156 } 1157 1158 // If the sign bits of both operands are zero (i.e. we can prove they are 1159 // unsigned inputs), turn this into a udiv. 1160 APInt Mask(APInt::getSignMask(Ty->getScalarSizeInBits())); 1161 if (MaskedValueIsZero(Op0, Mask, 0, &I)) { 1162 if (MaskedValueIsZero(Op1, Mask, 0, &I)) { 1163 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set 1164 auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName()); 1165 BO->setIsExact(I.isExact()); 1166 return BO; 1167 } 1168 1169 if (match(Op1, m_NegatedPower2())) { 1170 // X sdiv (-(1 << C)) -> -(X sdiv (1 << C)) -> 1171 // -> -(X udiv (1 << C)) -> -(X u>> C) 1172 Constant *CNegLog2 = ConstantExpr::getExactLogBase2( 1173 ConstantExpr::getNeg(cast<Constant>(Op1))); 1174 Value *Shr = Builder.CreateLShr(Op0, CNegLog2, I.getName(), I.isExact()); 1175 return BinaryOperator::CreateNeg(Shr); 1176 } 1177 1178 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, &I)) { 1179 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y) 1180 // Safe because the only negative value (1 << Y) can take on is 1181 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have 1182 // the sign bit set. 1183 auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName()); 1184 BO->setIsExact(I.isExact()); 1185 return BO; 1186 } 1187 } 1188 1189 return nullptr; 1190 } 1191 1192 /// Remove negation and try to convert division into multiplication. 1193 static Instruction *foldFDivConstantDivisor(BinaryOperator &I) { 1194 Constant *C; 1195 if (!match(I.getOperand(1), m_Constant(C))) 1196 return nullptr; 1197 1198 // -X / C --> X / -C 1199 Value *X; 1200 if (match(I.getOperand(0), m_FNeg(m_Value(X)))) 1201 return BinaryOperator::CreateFDivFMF(X, ConstantExpr::getFNeg(C), &I); 1202 1203 // If the constant divisor has an exact inverse, this is always safe. If not, 1204 // then we can still create a reciprocal if fast-math-flags allow it and the 1205 // constant is a regular number (not zero, infinite, or denormal). 1206 if (!(C->hasExactInverseFP() || (I.hasAllowReciprocal() && C->isNormalFP()))) 1207 return nullptr; 1208 1209 // Disallow denormal constants because we don't know what would happen 1210 // on all targets. 1211 // TODO: Use Intrinsic::canonicalize or let function attributes tell us that 1212 // denorms are flushed? 1213 auto *RecipC = ConstantExpr::getFDiv(ConstantFP::get(I.getType(), 1.0), C); 1214 if (!RecipC->isNormalFP()) 1215 return nullptr; 1216 1217 // X / C --> X * (1 / C) 1218 return BinaryOperator::CreateFMulFMF(I.getOperand(0), RecipC, &I); 1219 } 1220 1221 /// Remove negation and try to reassociate constant math. 1222 static Instruction *foldFDivConstantDividend(BinaryOperator &I) { 1223 Constant *C; 1224 if (!match(I.getOperand(0), m_Constant(C))) 1225 return nullptr; 1226 1227 // C / -X --> -C / X 1228 Value *X; 1229 if (match(I.getOperand(1), m_FNeg(m_Value(X)))) 1230 return BinaryOperator::CreateFDivFMF(ConstantExpr::getFNeg(C), X, &I); 1231 1232 if (!I.hasAllowReassoc() || !I.hasAllowReciprocal()) 1233 return nullptr; 1234 1235 // Try to reassociate C / X expressions where X includes another constant. 1236 Constant *C2, *NewC = nullptr; 1237 if (match(I.getOperand(1), m_FMul(m_Value(X), m_Constant(C2)))) { 1238 // C / (X * C2) --> (C / C2) / X 1239 NewC = ConstantExpr::getFDiv(C, C2); 1240 } else if (match(I.getOperand(1), m_FDiv(m_Value(X), m_Constant(C2)))) { 1241 // C / (X / C2) --> (C * C2) / X 1242 NewC = ConstantExpr::getFMul(C, C2); 1243 } 1244 // Disallow denormal constants because we don't know what would happen 1245 // on all targets. 1246 // TODO: Use Intrinsic::canonicalize or let function attributes tell us that 1247 // denorms are flushed? 1248 if (!NewC || !NewC->isNormalFP()) 1249 return nullptr; 1250 1251 return BinaryOperator::CreateFDivFMF(NewC, X, &I); 1252 } 1253 1254 /// Negate the exponent of pow/exp to fold division-by-pow() into multiply. 1255 static Instruction *foldFDivPowDivisor(BinaryOperator &I, 1256 InstCombiner::BuilderTy &Builder) { 1257 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1258 auto *II = dyn_cast<IntrinsicInst>(Op1); 1259 if (!II || !II->hasOneUse() || !I.hasAllowReassoc() || 1260 !I.hasAllowReciprocal()) 1261 return nullptr; 1262 1263 // Z / pow(X, Y) --> Z * pow(X, -Y) 1264 // Z / exp{2}(Y) --> Z * exp{2}(-Y) 1265 // In the general case, this creates an extra instruction, but fmul allows 1266 // for better canonicalization and optimization than fdiv. 1267 Intrinsic::ID IID = II->getIntrinsicID(); 1268 SmallVector<Value *> Args; 1269 switch (IID) { 1270 case Intrinsic::pow: 1271 Args.push_back(II->getArgOperand(0)); 1272 Args.push_back(Builder.CreateFNegFMF(II->getArgOperand(1), &I)); 1273 break; 1274 case Intrinsic::powi: { 1275 // Require 'ninf' assuming that makes powi(X, -INT_MIN) acceptable. 1276 // That is, X ** (huge negative number) is 0.0, ~1.0, or INF and so 1277 // dividing by that is INF, ~1.0, or 0.0. Code that uses powi allows 1278 // non-standard results, so this corner case should be acceptable if the 1279 // code rules out INF values. 1280 if (!I.hasNoInfs()) 1281 return nullptr; 1282 Args.push_back(II->getArgOperand(0)); 1283 Args.push_back(Builder.CreateNeg(II->getArgOperand(1))); 1284 Type *Tys[] = {I.getType(), II->getArgOperand(1)->getType()}; 1285 Value *Pow = Builder.CreateIntrinsic(IID, Tys, Args, &I); 1286 return BinaryOperator::CreateFMulFMF(Op0, Pow, &I); 1287 } 1288 case Intrinsic::exp: 1289 case Intrinsic::exp2: 1290 Args.push_back(Builder.CreateFNegFMF(II->getArgOperand(0), &I)); 1291 break; 1292 default: 1293 return nullptr; 1294 } 1295 Value *Pow = Builder.CreateIntrinsic(IID, I.getType(), Args, &I); 1296 return BinaryOperator::CreateFMulFMF(Op0, Pow, &I); 1297 } 1298 1299 Instruction *InstCombinerImpl::visitFDiv(BinaryOperator &I) { 1300 if (Value *V = SimplifyFDivInst(I.getOperand(0), I.getOperand(1), 1301 I.getFastMathFlags(), 1302 SQ.getWithInstruction(&I))) 1303 return replaceInstUsesWith(I, V); 1304 1305 if (Instruction *X = foldVectorBinop(I)) 1306 return X; 1307 1308 if (Instruction *Phi = foldBinopWithPhiOperands(I)) 1309 return Phi; 1310 1311 if (Instruction *R = foldFDivConstantDivisor(I)) 1312 return R; 1313 1314 if (Instruction *R = foldFDivConstantDividend(I)) 1315 return R; 1316 1317 if (Instruction *R = foldFPSignBitOps(I)) 1318 return R; 1319 1320 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1321 if (isa<Constant>(Op0)) 1322 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 1323 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1324 return R; 1325 1326 if (isa<Constant>(Op1)) 1327 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 1328 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1329 return R; 1330 1331 if (I.hasAllowReassoc() && I.hasAllowReciprocal()) { 1332 Value *X, *Y; 1333 if (match(Op0, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))) && 1334 (!isa<Constant>(Y) || !isa<Constant>(Op1))) { 1335 // (X / Y) / Z => X / (Y * Z) 1336 Value *YZ = Builder.CreateFMulFMF(Y, Op1, &I); 1337 return BinaryOperator::CreateFDivFMF(X, YZ, &I); 1338 } 1339 if (match(Op1, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))) && 1340 (!isa<Constant>(Y) || !isa<Constant>(Op0))) { 1341 // Z / (X / Y) => (Y * Z) / X 1342 Value *YZ = Builder.CreateFMulFMF(Y, Op0, &I); 1343 return BinaryOperator::CreateFDivFMF(YZ, X, &I); 1344 } 1345 // Z / (1.0 / Y) => (Y * Z) 1346 // 1347 // This is a special case of Z / (X / Y) => (Y * Z) / X, with X = 1.0. The 1348 // m_OneUse check is avoided because even in the case of the multiple uses 1349 // for 1.0/Y, the number of instructions remain the same and a division is 1350 // replaced by a multiplication. 1351 if (match(Op1, m_FDiv(m_SpecificFP(1.0), m_Value(Y)))) 1352 return BinaryOperator::CreateFMulFMF(Y, Op0, &I); 1353 } 1354 1355 if (I.hasAllowReassoc() && Op0->hasOneUse() && Op1->hasOneUse()) { 1356 // sin(X) / cos(X) -> tan(X) 1357 // cos(X) / sin(X) -> 1/tan(X) (cotangent) 1358 Value *X; 1359 bool IsTan = match(Op0, m_Intrinsic<Intrinsic::sin>(m_Value(X))) && 1360 match(Op1, m_Intrinsic<Intrinsic::cos>(m_Specific(X))); 1361 bool IsCot = 1362 !IsTan && match(Op0, m_Intrinsic<Intrinsic::cos>(m_Value(X))) && 1363 match(Op1, m_Intrinsic<Intrinsic::sin>(m_Specific(X))); 1364 1365 if ((IsTan || IsCot) && 1366 hasFloatFn(&TLI, I.getType(), LibFunc_tan, LibFunc_tanf, LibFunc_tanl)) { 1367 IRBuilder<> B(&I); 1368 IRBuilder<>::FastMathFlagGuard FMFGuard(B); 1369 B.setFastMathFlags(I.getFastMathFlags()); 1370 AttributeList Attrs = 1371 cast<CallBase>(Op0)->getCalledFunction()->getAttributes(); 1372 Value *Res = emitUnaryFloatFnCall(X, &TLI, LibFunc_tan, LibFunc_tanf, 1373 LibFunc_tanl, B, Attrs); 1374 if (IsCot) 1375 Res = B.CreateFDiv(ConstantFP::get(I.getType(), 1.0), Res); 1376 return replaceInstUsesWith(I, Res); 1377 } 1378 } 1379 1380 // X / (X * Y) --> 1.0 / Y 1381 // Reassociate to (X / X -> 1.0) is legal when NaNs are not allowed. 1382 // We can ignore the possibility that X is infinity because INF/INF is NaN. 1383 Value *X, *Y; 1384 if (I.hasNoNaNs() && I.hasAllowReassoc() && 1385 match(Op1, m_c_FMul(m_Specific(Op0), m_Value(Y)))) { 1386 replaceOperand(I, 0, ConstantFP::get(I.getType(), 1.0)); 1387 replaceOperand(I, 1, Y); 1388 return &I; 1389 } 1390 1391 // X / fabs(X) -> copysign(1.0, X) 1392 // fabs(X) / X -> copysign(1.0, X) 1393 if (I.hasNoNaNs() && I.hasNoInfs() && 1394 (match(&I, m_FDiv(m_Value(X), m_FAbs(m_Deferred(X)))) || 1395 match(&I, m_FDiv(m_FAbs(m_Value(X)), m_Deferred(X))))) { 1396 Value *V = Builder.CreateBinaryIntrinsic( 1397 Intrinsic::copysign, ConstantFP::get(I.getType(), 1.0), X, &I); 1398 return replaceInstUsesWith(I, V); 1399 } 1400 1401 if (Instruction *Mul = foldFDivPowDivisor(I, Builder)) 1402 return Mul; 1403 1404 return nullptr; 1405 } 1406 1407 /// This function implements the transforms common to both integer remainder 1408 /// instructions (urem and srem). It is called by the visitors to those integer 1409 /// remainder instructions. 1410 /// Common integer remainder transforms 1411 Instruction *InstCombinerImpl::commonIRemTransforms(BinaryOperator &I) { 1412 if (Instruction *Phi = foldBinopWithPhiOperands(I)) 1413 return Phi; 1414 1415 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1416 1417 // The RHS is known non-zero. 1418 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I)) 1419 return replaceOperand(I, 1, V); 1420 1421 // Handle cases involving: rem X, (select Cond, Y, Z) 1422 if (simplifyDivRemOfSelectWithZeroOp(I)) 1423 return &I; 1424 1425 // If the divisor is a select-of-constants, try to constant fold all rem ops: 1426 // C % (select Cond, TrueC, FalseC) --> select Cond, (C % TrueC), (C % FalseC) 1427 // TODO: Adapt simplifyDivRemOfSelectWithZeroOp to allow this and other folds. 1428 if (match(Op0, m_ImmConstant()) && 1429 match(Op1, m_Select(m_Value(), m_ImmConstant(), m_ImmConstant()))) { 1430 if (Instruction *R = FoldOpIntoSelect(I, cast<SelectInst>(Op1))) 1431 return R; 1432 } 1433 1434 if (isa<Constant>(Op1)) { 1435 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) { 1436 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) { 1437 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1438 return R; 1439 } else if (auto *PN = dyn_cast<PHINode>(Op0I)) { 1440 const APInt *Op1Int; 1441 if (match(Op1, m_APInt(Op1Int)) && !Op1Int->isMinValue() && 1442 (I.getOpcode() == Instruction::URem || 1443 !Op1Int->isMinSignedValue())) { 1444 // foldOpIntoPhi will speculate instructions to the end of the PHI's 1445 // predecessor blocks, so do this only if we know the srem or urem 1446 // will not fault. 1447 if (Instruction *NV = foldOpIntoPhi(I, PN)) 1448 return NV; 1449 } 1450 } 1451 1452 // See if we can fold away this rem instruction. 1453 if (SimplifyDemandedInstructionBits(I)) 1454 return &I; 1455 } 1456 } 1457 1458 return nullptr; 1459 } 1460 1461 Instruction *InstCombinerImpl::visitURem(BinaryOperator &I) { 1462 if (Value *V = SimplifyURemInst(I.getOperand(0), I.getOperand(1), 1463 SQ.getWithInstruction(&I))) 1464 return replaceInstUsesWith(I, V); 1465 1466 if (Instruction *X = foldVectorBinop(I)) 1467 return X; 1468 1469 if (Instruction *common = commonIRemTransforms(I)) 1470 return common; 1471 1472 if (Instruction *NarrowRem = narrowUDivURem(I, Builder)) 1473 return NarrowRem; 1474 1475 // X urem Y -> X and Y-1, where Y is a power of 2, 1476 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1477 Type *Ty = I.getType(); 1478 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, &I)) { 1479 // This may increase instruction count, we don't enforce that Y is a 1480 // constant. 1481 Constant *N1 = Constant::getAllOnesValue(Ty); 1482 Value *Add = Builder.CreateAdd(Op1, N1); 1483 return BinaryOperator::CreateAnd(Op0, Add); 1484 } 1485 1486 // 1 urem X -> zext(X != 1) 1487 if (match(Op0, m_One())) { 1488 Value *Cmp = Builder.CreateICmpNE(Op1, ConstantInt::get(Ty, 1)); 1489 return CastInst::CreateZExtOrBitCast(Cmp, Ty); 1490 } 1491 1492 // X urem C -> X < C ? X : X - C, where C >= signbit. 1493 if (match(Op1, m_Negative())) { 1494 Value *Cmp = Builder.CreateICmpULT(Op0, Op1); 1495 Value *Sub = Builder.CreateSub(Op0, Op1); 1496 return SelectInst::Create(Cmp, Op0, Sub); 1497 } 1498 1499 // If the divisor is a sext of a boolean, then the divisor must be max 1500 // unsigned value (-1). Therefore, the remainder is Op0 unless Op0 is also 1501 // max unsigned value. In that case, the remainder is 0: 1502 // urem Op0, (sext i1 X) --> (Op0 == -1) ? 0 : Op0 1503 Value *X; 1504 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) { 1505 Value *Cmp = Builder.CreateICmpEQ(Op0, ConstantInt::getAllOnesValue(Ty)); 1506 return SelectInst::Create(Cmp, ConstantInt::getNullValue(Ty), Op0); 1507 } 1508 1509 return nullptr; 1510 } 1511 1512 Instruction *InstCombinerImpl::visitSRem(BinaryOperator &I) { 1513 if (Value *V = SimplifySRemInst(I.getOperand(0), I.getOperand(1), 1514 SQ.getWithInstruction(&I))) 1515 return replaceInstUsesWith(I, V); 1516 1517 if (Instruction *X = foldVectorBinop(I)) 1518 return X; 1519 1520 // Handle the integer rem common cases 1521 if (Instruction *Common = commonIRemTransforms(I)) 1522 return Common; 1523 1524 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1525 { 1526 const APInt *Y; 1527 // X % -Y -> X % Y 1528 if (match(Op1, m_Negative(Y)) && !Y->isMinSignedValue()) 1529 return replaceOperand(I, 1, ConstantInt::get(I.getType(), -*Y)); 1530 } 1531 1532 // -X srem Y --> -(X srem Y) 1533 Value *X, *Y; 1534 if (match(&I, m_SRem(m_OneUse(m_NSWSub(m_Zero(), m_Value(X))), m_Value(Y)))) 1535 return BinaryOperator::CreateNSWNeg(Builder.CreateSRem(X, Y)); 1536 1537 // If the sign bits of both operands are zero (i.e. we can prove they are 1538 // unsigned inputs), turn this into a urem. 1539 APInt Mask(APInt::getSignMask(I.getType()->getScalarSizeInBits())); 1540 if (MaskedValueIsZero(Op1, Mask, 0, &I) && 1541 MaskedValueIsZero(Op0, Mask, 0, &I)) { 1542 // X srem Y -> X urem Y, iff X and Y don't have sign bit set 1543 return BinaryOperator::CreateURem(Op0, Op1, I.getName()); 1544 } 1545 1546 // If it's a constant vector, flip any negative values positive. 1547 if (isa<ConstantVector>(Op1) || isa<ConstantDataVector>(Op1)) { 1548 Constant *C = cast<Constant>(Op1); 1549 unsigned VWidth = cast<FixedVectorType>(C->getType())->getNumElements(); 1550 1551 bool hasNegative = false; 1552 bool hasMissing = false; 1553 for (unsigned i = 0; i != VWidth; ++i) { 1554 Constant *Elt = C->getAggregateElement(i); 1555 if (!Elt) { 1556 hasMissing = true; 1557 break; 1558 } 1559 1560 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elt)) 1561 if (RHS->isNegative()) 1562 hasNegative = true; 1563 } 1564 1565 if (hasNegative && !hasMissing) { 1566 SmallVector<Constant *, 16> Elts(VWidth); 1567 for (unsigned i = 0; i != VWidth; ++i) { 1568 Elts[i] = C->getAggregateElement(i); // Handle undef, etc. 1569 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elts[i])) { 1570 if (RHS->isNegative()) 1571 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS)); 1572 } 1573 } 1574 1575 Constant *NewRHSV = ConstantVector::get(Elts); 1576 if (NewRHSV != C) // Don't loop on -MININT 1577 return replaceOperand(I, 1, NewRHSV); 1578 } 1579 } 1580 1581 return nullptr; 1582 } 1583 1584 Instruction *InstCombinerImpl::visitFRem(BinaryOperator &I) { 1585 if (Value *V = SimplifyFRemInst(I.getOperand(0), I.getOperand(1), 1586 I.getFastMathFlags(), 1587 SQ.getWithInstruction(&I))) 1588 return replaceInstUsesWith(I, V); 1589 1590 if (Instruction *X = foldVectorBinop(I)) 1591 return X; 1592 1593 if (Instruction *Phi = foldBinopWithPhiOperands(I)) 1594 return Phi; 1595 1596 return nullptr; 1597 } 1598