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/APFloat.h" 16 #include "llvm/ADT/APInt.h" 17 #include "llvm/ADT/SmallVector.h" 18 #include "llvm/Analysis/InstructionSimplify.h" 19 #include "llvm/IR/BasicBlock.h" 20 #include "llvm/IR/Constant.h" 21 #include "llvm/IR/Constants.h" 22 #include "llvm/IR/InstrTypes.h" 23 #include "llvm/IR/Instruction.h" 24 #include "llvm/IR/Instructions.h" 25 #include "llvm/IR/IntrinsicInst.h" 26 #include "llvm/IR/Intrinsics.h" 27 #include "llvm/IR/Operator.h" 28 #include "llvm/IR/PatternMatch.h" 29 #include "llvm/IR/Type.h" 30 #include "llvm/IR/Value.h" 31 #include "llvm/Support/Casting.h" 32 #include "llvm/Support/ErrorHandling.h" 33 #include "llvm/Support/KnownBits.h" 34 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h" 35 #include "llvm/Transforms/InstCombine/InstCombiner.h" 36 #include "llvm/Transforms/Utils/BuildLibCalls.h" 37 #include <cassert> 38 #include <cstddef> 39 #include <cstdint> 40 #include <utility> 41 42 using namespace llvm; 43 using namespace PatternMatch; 44 45 #define DEBUG_TYPE "instcombine" 46 47 /// The specific integer value is used in a context where it is known to be 48 /// non-zero. If this allows us to simplify the computation, do so and return 49 /// the new operand, otherwise return null. 50 static Value *simplifyValueKnownNonZero(Value *V, InstCombinerImpl &IC, 51 Instruction &CxtI) { 52 // If V has multiple uses, then we would have to do more analysis to determine 53 // if this is safe. For example, the use could be in dynamically unreached 54 // code. 55 if (!V->hasOneUse()) return nullptr; 56 57 bool MadeChange = false; 58 59 // ((1 << A) >>u B) --> (1 << (A-B)) 60 // Because V cannot be zero, we know that B is less than A. 61 Value *A = nullptr, *B = nullptr, *One = nullptr; 62 if (match(V, m_LShr(m_OneUse(m_Shl(m_Value(One), m_Value(A))), m_Value(B))) && 63 match(One, m_One())) { 64 A = IC.Builder.CreateSub(A, B); 65 return IC.Builder.CreateShl(One, A); 66 } 67 68 // (PowerOfTwo >>u B) --> isExact since shifting out the result would make it 69 // inexact. Similarly for <<. 70 BinaryOperator *I = dyn_cast<BinaryOperator>(V); 71 if (I && I->isLogicalShift() && 72 IC.isKnownToBeAPowerOfTwo(I->getOperand(0), false, 0, &CxtI)) { 73 // We know that this is an exact/nuw shift and that the input is a 74 // non-zero context as well. 75 if (Value *V2 = simplifyValueKnownNonZero(I->getOperand(0), IC, CxtI)) { 76 IC.replaceOperand(*I, 0, V2); 77 MadeChange = true; 78 } 79 80 if (I->getOpcode() == Instruction::LShr && !I->isExact()) { 81 I->setIsExact(); 82 MadeChange = true; 83 } 84 85 if (I->getOpcode() == Instruction::Shl && !I->hasNoUnsignedWrap()) { 86 I->setHasNoUnsignedWrap(); 87 MadeChange = true; 88 } 89 } 90 91 // TODO: Lots more we could do here: 92 // If V is a phi node, we can call this on each of its operands. 93 // "select cond, X, 0" can simplify to "X". 94 95 return MadeChange ? V : nullptr; 96 } 97 98 // TODO: This is a specific form of a much more general pattern. 99 // We could detect a select with any binop identity constant, or we 100 // could use SimplifyBinOp to see if either arm of the select reduces. 101 // But that needs to be done carefully and/or while removing potential 102 // reverse canonicalizations as in InstCombiner::foldSelectIntoOp(). 103 static Value *foldMulSelectToNegate(BinaryOperator &I, 104 InstCombiner::BuilderTy &Builder) { 105 Value *Cond, *OtherOp; 106 107 // mul (select Cond, 1, -1), OtherOp --> select Cond, OtherOp, -OtherOp 108 // mul OtherOp, (select Cond, 1, -1) --> select Cond, OtherOp, -OtherOp 109 if (match(&I, m_c_Mul(m_OneUse(m_Select(m_Value(Cond), m_One(), m_AllOnes())), 110 m_Value(OtherOp)))) 111 return Builder.CreateSelect(Cond, OtherOp, Builder.CreateNeg(OtherOp)); 112 113 // mul (select Cond, -1, 1), OtherOp --> select Cond, -OtherOp, OtherOp 114 // mul OtherOp, (select Cond, -1, 1) --> select Cond, -OtherOp, OtherOp 115 if (match(&I, m_c_Mul(m_OneUse(m_Select(m_Value(Cond), m_AllOnes(), m_One())), 116 m_Value(OtherOp)))) 117 return Builder.CreateSelect(Cond, Builder.CreateNeg(OtherOp), OtherOp); 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 (Value *V = SimplifyUsingDistributiveLaws(I)) 154 return replaceInstUsesWith(I, V); 155 156 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 157 unsigned BitWidth = I.getType()->getScalarSizeInBits(); 158 159 // X * -1 == 0 - X 160 if (match(Op1, m_AllOnes())) { 161 BinaryOperator *BO = BinaryOperator::CreateNeg(Op0, I.getName()); 162 if (I.hasNoSignedWrap()) 163 BO->setHasNoSignedWrap(); 164 return BO; 165 } 166 167 // Also allow combining multiply instructions on vectors. 168 { 169 Value *NewOp; 170 Constant *C1, *C2; 171 const APInt *IVal; 172 if (match(&I, m_Mul(m_Shl(m_Value(NewOp), m_Constant(C2)), 173 m_Constant(C1))) && 174 match(C1, m_APInt(IVal))) { 175 // ((X << C2)*C1) == (X * (C1 << C2)) 176 Constant *Shl = ConstantExpr::getShl(C1, C2); 177 BinaryOperator *Mul = cast<BinaryOperator>(I.getOperand(0)); 178 BinaryOperator *BO = BinaryOperator::CreateMul(NewOp, Shl); 179 if (I.hasNoUnsignedWrap() && Mul->hasNoUnsignedWrap()) 180 BO->setHasNoUnsignedWrap(); 181 if (I.hasNoSignedWrap() && Mul->hasNoSignedWrap() && 182 Shl->isNotMinSignedValue()) 183 BO->setHasNoSignedWrap(); 184 return BO; 185 } 186 187 if (match(&I, m_Mul(m_Value(NewOp), m_Constant(C1)))) { 188 // Replace X*(2^C) with X << C, where C is either a scalar or a vector. 189 if (Constant *NewCst = ConstantExpr::getExactLogBase2(C1)) { 190 BinaryOperator *Shl = BinaryOperator::CreateShl(NewOp, NewCst); 191 192 if (I.hasNoUnsignedWrap()) 193 Shl->setHasNoUnsignedWrap(); 194 if (I.hasNoSignedWrap()) { 195 const APInt *V; 196 if (match(NewCst, m_APInt(V)) && *V != V->getBitWidth() - 1) 197 Shl->setHasNoSignedWrap(); 198 } 199 200 return Shl; 201 } 202 } 203 } 204 205 if (Op0->hasOneUse() && match(Op1, m_NegatedPower2())) { 206 // Interpret X * (-1<<C) as (-X) * (1<<C) and try to sink the negation. 207 // The "* (1<<C)" thus becomes a potential shifting opportunity. 208 if (Value *NegOp0 = Negator::Negate(/*IsNegation*/ true, Op0, *this)) 209 return BinaryOperator::CreateMul( 210 NegOp0, ConstantExpr::getNeg(cast<Constant>(Op1)), I.getName()); 211 } 212 213 if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I)) 214 return FoldedMul; 215 216 if (Value *FoldedMul = foldMulSelectToNegate(I, Builder)) 217 return replaceInstUsesWith(I, FoldedMul); 218 219 // Simplify mul instructions with a constant RHS. 220 if (isa<Constant>(Op1)) { 221 // Canonicalize (X+C1)*CI -> X*CI+C1*CI. 222 Value *X; 223 Constant *C1; 224 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_Constant(C1))))) { 225 Value *Mul = Builder.CreateMul(C1, Op1); 226 // Only go forward with the transform if C1*CI simplifies to a tidier 227 // constant. 228 if (!match(Mul, m_Mul(m_Value(), m_Value()))) 229 return BinaryOperator::CreateAdd(Builder.CreateMul(X, Op1), Mul); 230 } 231 } 232 233 // abs(X) * abs(X) -> X * X 234 // nabs(X) * nabs(X) -> X * X 235 if (Op0 == Op1) { 236 Value *X, *Y; 237 SelectPatternFlavor SPF = matchSelectPattern(Op0, X, Y).Flavor; 238 if (SPF == SPF_ABS || SPF == SPF_NABS) 239 return BinaryOperator::CreateMul(X, X); 240 241 if (match(Op0, m_Intrinsic<Intrinsic::abs>(m_Value(X)))) 242 return BinaryOperator::CreateMul(X, X); 243 } 244 245 // -X * C --> X * -C 246 Value *X, *Y; 247 Constant *Op1C; 248 if (match(Op0, m_Neg(m_Value(X))) && match(Op1, m_Constant(Op1C))) 249 return BinaryOperator::CreateMul(X, ConstantExpr::getNeg(Op1C)); 250 251 // -X * -Y --> X * Y 252 if (match(Op0, m_Neg(m_Value(X))) && match(Op1, m_Neg(m_Value(Y)))) { 253 auto *NewMul = BinaryOperator::CreateMul(X, Y); 254 if (I.hasNoSignedWrap() && 255 cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap() && 256 cast<OverflowingBinaryOperator>(Op1)->hasNoSignedWrap()) 257 NewMul->setHasNoSignedWrap(); 258 return NewMul; 259 } 260 261 // -X * Y --> -(X * Y) 262 // X * -Y --> -(X * Y) 263 if (match(&I, m_c_Mul(m_OneUse(m_Neg(m_Value(X))), m_Value(Y)))) 264 return BinaryOperator::CreateNeg(Builder.CreateMul(X, Y)); 265 266 // (X / Y) * Y = X - (X % Y) 267 // (X / Y) * -Y = (X % Y) - X 268 { 269 Value *Y = Op1; 270 BinaryOperator *Div = dyn_cast<BinaryOperator>(Op0); 271 if (!Div || (Div->getOpcode() != Instruction::UDiv && 272 Div->getOpcode() != Instruction::SDiv)) { 273 Y = Op0; 274 Div = dyn_cast<BinaryOperator>(Op1); 275 } 276 Value *Neg = dyn_castNegVal(Y); 277 if (Div && Div->hasOneUse() && 278 (Div->getOperand(1) == Y || Div->getOperand(1) == Neg) && 279 (Div->getOpcode() == Instruction::UDiv || 280 Div->getOpcode() == Instruction::SDiv)) { 281 Value *X = Div->getOperand(0), *DivOp1 = Div->getOperand(1); 282 283 // If the division is exact, X % Y is zero, so we end up with X or -X. 284 if (Div->isExact()) { 285 if (DivOp1 == Y) 286 return replaceInstUsesWith(I, X); 287 return BinaryOperator::CreateNeg(X); 288 } 289 290 auto RemOpc = Div->getOpcode() == Instruction::UDiv ? Instruction::URem 291 : Instruction::SRem; 292 Value *Rem = Builder.CreateBinOp(RemOpc, X, DivOp1); 293 if (DivOp1 == Y) 294 return BinaryOperator::CreateSub(X, Rem); 295 return BinaryOperator::CreateSub(Rem, X); 296 } 297 } 298 299 /// i1 mul -> i1 and. 300 if (I.getType()->isIntOrIntVectorTy(1)) 301 return BinaryOperator::CreateAnd(Op0, Op1); 302 303 // X*(1 << Y) --> X << Y 304 // (1 << Y)*X --> X << Y 305 { 306 Value *Y; 307 BinaryOperator *BO = nullptr; 308 bool ShlNSW = false; 309 if (match(Op0, m_Shl(m_One(), m_Value(Y)))) { 310 BO = BinaryOperator::CreateShl(Op1, Y); 311 ShlNSW = cast<ShlOperator>(Op0)->hasNoSignedWrap(); 312 } else if (match(Op1, m_Shl(m_One(), m_Value(Y)))) { 313 BO = BinaryOperator::CreateShl(Op0, Y); 314 ShlNSW = cast<ShlOperator>(Op1)->hasNoSignedWrap(); 315 } 316 if (BO) { 317 if (I.hasNoUnsignedWrap()) 318 BO->setHasNoUnsignedWrap(); 319 if (I.hasNoSignedWrap() && ShlNSW) 320 BO->setHasNoSignedWrap(); 321 return BO; 322 } 323 } 324 325 // (zext bool X) * (zext bool Y) --> zext (and X, Y) 326 // (sext bool X) * (sext bool Y) --> zext (and X, Y) 327 // Note: -1 * -1 == 1 * 1 == 1 (if the extends match, the result is the same) 328 if (((match(Op0, m_ZExt(m_Value(X))) && match(Op1, m_ZExt(m_Value(Y)))) || 329 (match(Op0, m_SExt(m_Value(X))) && match(Op1, m_SExt(m_Value(Y))))) && 330 X->getType()->isIntOrIntVectorTy(1) && X->getType() == Y->getType() && 331 (Op0->hasOneUse() || Op1->hasOneUse())) { 332 Value *And = Builder.CreateAnd(X, Y, "mulbool"); 333 return CastInst::Create(Instruction::ZExt, And, I.getType()); 334 } 335 // (sext bool X) * (zext bool Y) --> sext (and X, Y) 336 // (zext bool X) * (sext bool Y) --> sext (and X, Y) 337 // Note: -1 * 1 == 1 * -1 == -1 338 if (((match(Op0, m_SExt(m_Value(X))) && match(Op1, m_ZExt(m_Value(Y)))) || 339 (match(Op0, m_ZExt(m_Value(X))) && match(Op1, m_SExt(m_Value(Y))))) && 340 X->getType()->isIntOrIntVectorTy(1) && X->getType() == Y->getType() && 341 (Op0->hasOneUse() || Op1->hasOneUse())) { 342 Value *And = Builder.CreateAnd(X, Y, "mulbool"); 343 return CastInst::Create(Instruction::SExt, And, I.getType()); 344 } 345 346 // (bool X) * Y --> X ? Y : 0 347 // Y * (bool X) --> X ? Y : 0 348 if (match(Op0, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) 349 return SelectInst::Create(X, Op1, ConstantInt::get(I.getType(), 0)); 350 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) 351 return SelectInst::Create(X, Op0, ConstantInt::get(I.getType(), 0)); 352 353 // (lshr X, 31) * Y --> (ashr X, 31) & Y 354 // Y * (lshr X, 31) --> (ashr X, 31) & Y 355 // TODO: We are not checking one-use because the elimination of the multiply 356 // is better for analysis? 357 // TODO: Should we canonicalize to '(X < 0) ? Y : 0' instead? That would be 358 // more similar to what we're doing above. 359 const APInt *C; 360 if (match(Op0, m_LShr(m_Value(X), m_APInt(C))) && *C == C->getBitWidth() - 1) 361 return BinaryOperator::CreateAnd(Builder.CreateAShr(X, *C), Op1); 362 if (match(Op1, m_LShr(m_Value(X), m_APInt(C))) && *C == C->getBitWidth() - 1) 363 return BinaryOperator::CreateAnd(Builder.CreateAShr(X, *C), Op0); 364 365 // ((ashr X, 31) | 1) * X --> abs(X) 366 // X * ((ashr X, 31) | 1) --> abs(X) 367 if (match(&I, m_c_BinOp(m_Or(m_AShr(m_Value(X), 368 m_SpecificIntAllowUndef(BitWidth - 1)), 369 m_One()), 370 m_Deferred(X)))) { 371 Value *Abs = Builder.CreateBinaryIntrinsic( 372 Intrinsic::abs, X, 373 ConstantInt::getBool(I.getContext(), I.hasNoSignedWrap())); 374 Abs->takeName(&I); 375 return replaceInstUsesWith(I, Abs); 376 } 377 378 if (Instruction *Ext = narrowMathIfNoOverflow(I)) 379 return Ext; 380 381 bool Changed = false; 382 if (!I.hasNoSignedWrap() && willNotOverflowSignedMul(Op0, Op1, I)) { 383 Changed = true; 384 I.setHasNoSignedWrap(true); 385 } 386 387 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedMul(Op0, Op1, I)) { 388 Changed = true; 389 I.setHasNoUnsignedWrap(true); 390 } 391 392 return Changed ? &I : nullptr; 393 } 394 395 Instruction *InstCombinerImpl::foldFPSignBitOps(BinaryOperator &I) { 396 BinaryOperator::BinaryOps Opcode = I.getOpcode(); 397 assert((Opcode == Instruction::FMul || Opcode == Instruction::FDiv) && 398 "Expected fmul or fdiv"); 399 400 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 401 Value *X, *Y; 402 403 // -X * -Y --> X * Y 404 // -X / -Y --> X / Y 405 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y)))) 406 return BinaryOperator::CreateWithCopiedFlags(Opcode, X, Y, &I); 407 408 // fabs(X) * fabs(X) -> X * X 409 // fabs(X) / fabs(X) -> X / X 410 if (Op0 == Op1 && match(Op0, m_FAbs(m_Value(X)))) 411 return BinaryOperator::CreateWithCopiedFlags(Opcode, X, X, &I); 412 413 // fabs(X) * fabs(Y) --> fabs(X * Y) 414 // fabs(X) / fabs(Y) --> fabs(X / Y) 415 if (match(Op0, m_FAbs(m_Value(X))) && match(Op1, m_FAbs(m_Value(Y))) && 416 (Op0->hasOneUse() || Op1->hasOneUse())) { 417 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); 418 Builder.setFastMathFlags(I.getFastMathFlags()); 419 Value *XY = Builder.CreateBinOp(Opcode, X, Y); 420 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, XY); 421 Fabs->takeName(&I); 422 return replaceInstUsesWith(I, Fabs); 423 } 424 425 return nullptr; 426 } 427 428 Instruction *InstCombinerImpl::visitFMul(BinaryOperator &I) { 429 if (Value *V = SimplifyFMulInst(I.getOperand(0), I.getOperand(1), 430 I.getFastMathFlags(), 431 SQ.getWithInstruction(&I))) 432 return replaceInstUsesWith(I, V); 433 434 if (SimplifyAssociativeOrCommutative(I)) 435 return &I; 436 437 if (Instruction *X = foldVectorBinop(I)) 438 return X; 439 440 if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I)) 441 return FoldedMul; 442 443 if (Value *FoldedMul = foldMulSelectToNegate(I, Builder)) 444 return replaceInstUsesWith(I, FoldedMul); 445 446 if (Instruction *R = foldFPSignBitOps(I)) 447 return R; 448 449 // X * -1.0 --> -X 450 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 451 if (match(Op1, m_SpecificFP(-1.0))) 452 return UnaryOperator::CreateFNegFMF(Op0, &I); 453 454 // -X * C --> X * -C 455 Value *X, *Y; 456 Constant *C; 457 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_Constant(C))) 458 return BinaryOperator::CreateFMulFMF(X, ConstantExpr::getFNeg(C), &I); 459 460 // (select A, B, C) * (select A, D, E) --> select A, (B*D), (C*E) 461 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1)) 462 return replaceInstUsesWith(I, V); 463 464 if (I.hasAllowReassoc()) { 465 // Reassociate constant RHS with another constant to form constant 466 // expression. 467 if (match(Op1, m_Constant(C)) && C->isFiniteNonZeroFP()) { 468 Constant *C1; 469 if (match(Op0, m_OneUse(m_FDiv(m_Constant(C1), m_Value(X))))) { 470 // (C1 / X) * C --> (C * C1) / X 471 Constant *CC1 = ConstantExpr::getFMul(C, C1); 472 if (CC1->isNormalFP()) 473 return BinaryOperator::CreateFDivFMF(CC1, X, &I); 474 } 475 if (match(Op0, m_FDiv(m_Value(X), m_Constant(C1)))) { 476 // (X / C1) * C --> X * (C / C1) 477 Constant *CDivC1 = ConstantExpr::getFDiv(C, C1); 478 if (CDivC1->isNormalFP()) 479 return BinaryOperator::CreateFMulFMF(X, CDivC1, &I); 480 481 // If the constant was a denormal, try reassociating differently. 482 // (X / C1) * C --> X / (C1 / C) 483 Constant *C1DivC = ConstantExpr::getFDiv(C1, C); 484 if (Op0->hasOneUse() && C1DivC->isNormalFP()) 485 return BinaryOperator::CreateFDivFMF(X, C1DivC, &I); 486 } 487 488 // We do not need to match 'fadd C, X' and 'fsub X, C' because they are 489 // canonicalized to 'fadd X, C'. Distributing the multiply may allow 490 // further folds and (X * C) + C2 is 'fma'. 491 if (match(Op0, m_OneUse(m_FAdd(m_Value(X), m_Constant(C1))))) { 492 // (X + C1) * C --> (X * C) + (C * C1) 493 Constant *CC1 = ConstantExpr::getFMul(C, C1); 494 Value *XC = Builder.CreateFMulFMF(X, C, &I); 495 return BinaryOperator::CreateFAddFMF(XC, CC1, &I); 496 } 497 if (match(Op0, m_OneUse(m_FSub(m_Constant(C1), m_Value(X))))) { 498 // (C1 - X) * C --> (C * C1) - (X * C) 499 Constant *CC1 = ConstantExpr::getFMul(C, C1); 500 Value *XC = Builder.CreateFMulFMF(X, C, &I); 501 return BinaryOperator::CreateFSubFMF(CC1, XC, &I); 502 } 503 } 504 505 Value *Z; 506 if (match(&I, m_c_FMul(m_OneUse(m_FDiv(m_Value(X), m_Value(Y))), 507 m_Value(Z)))) { 508 // Sink division: (X / Y) * Z --> (X * Z) / Y 509 Value *NewFMul = Builder.CreateFMulFMF(X, Z, &I); 510 return BinaryOperator::CreateFDivFMF(NewFMul, Y, &I); 511 } 512 513 // sqrt(X) * sqrt(Y) -> sqrt(X * Y) 514 // nnan disallows the possibility of returning a number if both operands are 515 // negative (in that case, we should return NaN). 516 if (I.hasNoNaNs() && 517 match(Op0, m_OneUse(m_Intrinsic<Intrinsic::sqrt>(m_Value(X)))) && 518 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::sqrt>(m_Value(Y))))) { 519 Value *XY = Builder.CreateFMulFMF(X, Y, &I); 520 Value *Sqrt = Builder.CreateUnaryIntrinsic(Intrinsic::sqrt, XY, &I); 521 return replaceInstUsesWith(I, Sqrt); 522 } 523 524 // The following transforms are done irrespective of the number of uses 525 // for the expression "1.0/sqrt(X)". 526 // 1) 1.0/sqrt(X) * X -> X/sqrt(X) 527 // 2) X * 1.0/sqrt(X) -> X/sqrt(X) 528 // We always expect the backend to reduce X/sqrt(X) to sqrt(X), if it 529 // has the necessary (reassoc) fast-math-flags. 530 if (I.hasNoSignedZeros() && 531 match(Op0, (m_FDiv(m_SpecificFP(1.0), m_Value(Y)))) && 532 match(Y, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) && Op1 == X) 533 return BinaryOperator::CreateFDivFMF(X, Y, &I); 534 if (I.hasNoSignedZeros() && 535 match(Op1, (m_FDiv(m_SpecificFP(1.0), m_Value(Y)))) && 536 match(Y, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) && Op0 == X) 537 return BinaryOperator::CreateFDivFMF(X, Y, &I); 538 539 // Like the similar transform in instsimplify, this requires 'nsz' because 540 // sqrt(-0.0) = -0.0, and -0.0 * -0.0 does not simplify to -0.0. 541 if (I.hasNoNaNs() && I.hasNoSignedZeros() && Op0 == Op1 && 542 Op0->hasNUses(2)) { 543 // Peek through fdiv to find squaring of square root: 544 // (X / sqrt(Y)) * (X / sqrt(Y)) --> (X * X) / Y 545 if (match(Op0, m_FDiv(m_Value(X), 546 m_Intrinsic<Intrinsic::sqrt>(m_Value(Y))))) { 547 Value *XX = Builder.CreateFMulFMF(X, X, &I); 548 return BinaryOperator::CreateFDivFMF(XX, Y, &I); 549 } 550 // (sqrt(Y) / X) * (sqrt(Y) / X) --> Y / (X * X) 551 if (match(Op0, m_FDiv(m_Intrinsic<Intrinsic::sqrt>(m_Value(Y)), 552 m_Value(X)))) { 553 Value *XX = Builder.CreateFMulFMF(X, X, &I); 554 return BinaryOperator::CreateFDivFMF(Y, XX, &I); 555 } 556 } 557 558 // exp(X) * exp(Y) -> exp(X + Y) 559 // Match as long as at least one of exp has only one use. 560 if (match(Op0, m_Intrinsic<Intrinsic::exp>(m_Value(X))) && 561 match(Op1, m_Intrinsic<Intrinsic::exp>(m_Value(Y))) && 562 (Op0->hasOneUse() || Op1->hasOneUse())) { 563 Value *XY = Builder.CreateFAddFMF(X, Y, &I); 564 Value *Exp = Builder.CreateUnaryIntrinsic(Intrinsic::exp, XY, &I); 565 return replaceInstUsesWith(I, Exp); 566 } 567 568 // exp2(X) * exp2(Y) -> exp2(X + Y) 569 // Match as long as at least one of exp2 has only one use. 570 if (match(Op0, m_Intrinsic<Intrinsic::exp2>(m_Value(X))) && 571 match(Op1, m_Intrinsic<Intrinsic::exp2>(m_Value(Y))) && 572 (Op0->hasOneUse() || Op1->hasOneUse())) { 573 Value *XY = Builder.CreateFAddFMF(X, Y, &I); 574 Value *Exp2 = Builder.CreateUnaryIntrinsic(Intrinsic::exp2, XY, &I); 575 return replaceInstUsesWith(I, Exp2); 576 } 577 578 // (X*Y) * X => (X*X) * Y where Y != X 579 // The purpose is two-fold: 580 // 1) to form a power expression (of X). 581 // 2) potentially shorten the critical path: After transformation, the 582 // latency of the instruction Y is amortized by the expression of X*X, 583 // and therefore Y is in a "less critical" position compared to what it 584 // was before the transformation. 585 if (match(Op0, m_OneUse(m_c_FMul(m_Specific(Op1), m_Value(Y)))) && 586 Op1 != Y) { 587 Value *XX = Builder.CreateFMulFMF(Op1, Op1, &I); 588 return BinaryOperator::CreateFMulFMF(XX, Y, &I); 589 } 590 if (match(Op1, m_OneUse(m_c_FMul(m_Specific(Op0), m_Value(Y)))) && 591 Op0 != Y) { 592 Value *XX = Builder.CreateFMulFMF(Op0, Op0, &I); 593 return BinaryOperator::CreateFMulFMF(XX, Y, &I); 594 } 595 } 596 597 // log2(X * 0.5) * Y = log2(X) * Y - Y 598 if (I.isFast()) { 599 IntrinsicInst *Log2 = nullptr; 600 if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::log2>( 601 m_OneUse(m_FMul(m_Value(X), m_SpecificFP(0.5))))))) { 602 Log2 = cast<IntrinsicInst>(Op0); 603 Y = Op1; 604 } 605 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::log2>( 606 m_OneUse(m_FMul(m_Value(X), m_SpecificFP(0.5))))))) { 607 Log2 = cast<IntrinsicInst>(Op1); 608 Y = Op0; 609 } 610 if (Log2) { 611 Value *Log2 = Builder.CreateUnaryIntrinsic(Intrinsic::log2, X, &I); 612 Value *LogXTimesY = Builder.CreateFMulFMF(Log2, Y, &I); 613 return BinaryOperator::CreateFSubFMF(LogXTimesY, Y, &I); 614 } 615 } 616 617 return nullptr; 618 } 619 620 /// Fold a divide or remainder with a select instruction divisor when one of the 621 /// select operands is zero. In that case, we can use the other select operand 622 /// because div/rem by zero is undefined. 623 bool InstCombinerImpl::simplifyDivRemOfSelectWithZeroOp(BinaryOperator &I) { 624 SelectInst *SI = dyn_cast<SelectInst>(I.getOperand(1)); 625 if (!SI) 626 return false; 627 628 int NonNullOperand; 629 if (match(SI->getTrueValue(), m_Zero())) 630 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y 631 NonNullOperand = 2; 632 else if (match(SI->getFalseValue(), m_Zero())) 633 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y 634 NonNullOperand = 1; 635 else 636 return false; 637 638 // Change the div/rem to use 'Y' instead of the select. 639 replaceOperand(I, 1, SI->getOperand(NonNullOperand)); 640 641 // Okay, we know we replace the operand of the div/rem with 'Y' with no 642 // problem. However, the select, or the condition of the select may have 643 // multiple uses. Based on our knowledge that the operand must be non-zero, 644 // propagate the known value for the select into other uses of it, and 645 // propagate a known value of the condition into its other users. 646 647 // If the select and condition only have a single use, don't bother with this, 648 // early exit. 649 Value *SelectCond = SI->getCondition(); 650 if (SI->use_empty() && SelectCond->hasOneUse()) 651 return true; 652 653 // Scan the current block backward, looking for other uses of SI. 654 BasicBlock::iterator BBI = I.getIterator(), BBFront = I.getParent()->begin(); 655 Type *CondTy = SelectCond->getType(); 656 while (BBI != BBFront) { 657 --BBI; 658 // If we found an instruction that we can't assume will return, so 659 // information from below it cannot be propagated above it. 660 if (!isGuaranteedToTransferExecutionToSuccessor(&*BBI)) 661 break; 662 663 // Replace uses of the select or its condition with the known values. 664 for (Use &Op : BBI->operands()) { 665 if (Op == SI) { 666 replaceUse(Op, SI->getOperand(NonNullOperand)); 667 Worklist.push(&*BBI); 668 } else if (Op == SelectCond) { 669 replaceUse(Op, NonNullOperand == 1 ? ConstantInt::getTrue(CondTy) 670 : ConstantInt::getFalse(CondTy)); 671 Worklist.push(&*BBI); 672 } 673 } 674 675 // If we past the instruction, quit looking for it. 676 if (&*BBI == SI) 677 SI = nullptr; 678 if (&*BBI == SelectCond) 679 SelectCond = nullptr; 680 681 // If we ran out of things to eliminate, break out of the loop. 682 if (!SelectCond && !SI) 683 break; 684 685 } 686 return true; 687 } 688 689 /// True if the multiply can not be expressed in an int this size. 690 static bool multiplyOverflows(const APInt &C1, const APInt &C2, APInt &Product, 691 bool IsSigned) { 692 bool Overflow; 693 Product = IsSigned ? C1.smul_ov(C2, Overflow) : C1.umul_ov(C2, Overflow); 694 return Overflow; 695 } 696 697 /// True if C1 is a multiple of C2. Quotient contains C1/C2. 698 static bool isMultiple(const APInt &C1, const APInt &C2, APInt &Quotient, 699 bool IsSigned) { 700 assert(C1.getBitWidth() == C2.getBitWidth() && "Constant widths not equal"); 701 702 // Bail if we will divide by zero. 703 if (C2.isNullValue()) 704 return false; 705 706 // Bail if we would divide INT_MIN by -1. 707 if (IsSigned && C1.isMinSignedValue() && C2.isAllOnesValue()) 708 return false; 709 710 APInt Remainder(C1.getBitWidth(), /*val=*/0ULL, IsSigned); 711 if (IsSigned) 712 APInt::sdivrem(C1, C2, Quotient, Remainder); 713 else 714 APInt::udivrem(C1, C2, Quotient, Remainder); 715 716 return Remainder.isMinValue(); 717 } 718 719 /// This function implements the transforms common to both integer division 720 /// instructions (udiv and sdiv). It is called by the visitors to those integer 721 /// division instructions. 722 /// Common integer divide transforms 723 Instruction *InstCombinerImpl::commonIDivTransforms(BinaryOperator &I) { 724 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 725 bool IsSigned = I.getOpcode() == Instruction::SDiv; 726 Type *Ty = I.getType(); 727 728 // The RHS is known non-zero. 729 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I)) 730 return replaceOperand(I, 1, V); 731 732 // Handle cases involving: [su]div X, (select Cond, Y, Z) 733 // This does not apply for fdiv. 734 if (simplifyDivRemOfSelectWithZeroOp(I)) 735 return &I; 736 737 const APInt *C2; 738 if (match(Op1, m_APInt(C2))) { 739 Value *X; 740 const APInt *C1; 741 742 // (X / C1) / C2 -> X / (C1*C2) 743 if ((IsSigned && match(Op0, m_SDiv(m_Value(X), m_APInt(C1)))) || 744 (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_APInt(C1))))) { 745 APInt Product(C1->getBitWidth(), /*val=*/0ULL, IsSigned); 746 if (!multiplyOverflows(*C1, *C2, Product, IsSigned)) 747 return BinaryOperator::Create(I.getOpcode(), X, 748 ConstantInt::get(Ty, Product)); 749 } 750 751 if ((IsSigned && match(Op0, m_NSWMul(m_Value(X), m_APInt(C1)))) || 752 (!IsSigned && match(Op0, m_NUWMul(m_Value(X), m_APInt(C1))))) { 753 APInt Quotient(C1->getBitWidth(), /*val=*/0ULL, IsSigned); 754 755 // (X * C1) / C2 -> X / (C2 / C1) if C2 is a multiple of C1. 756 if (isMultiple(*C2, *C1, Quotient, IsSigned)) { 757 auto *NewDiv = BinaryOperator::Create(I.getOpcode(), X, 758 ConstantInt::get(Ty, Quotient)); 759 NewDiv->setIsExact(I.isExact()); 760 return NewDiv; 761 } 762 763 // (X * C1) / C2 -> X * (C1 / C2) if C1 is a multiple of C2. 764 if (isMultiple(*C1, *C2, Quotient, IsSigned)) { 765 auto *Mul = BinaryOperator::Create(Instruction::Mul, X, 766 ConstantInt::get(Ty, Quotient)); 767 auto *OBO = cast<OverflowingBinaryOperator>(Op0); 768 Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap()); 769 Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap()); 770 return Mul; 771 } 772 } 773 774 if ((IsSigned && match(Op0, m_NSWShl(m_Value(X), m_APInt(C1))) && 775 *C1 != C1->getBitWidth() - 1) || 776 (!IsSigned && match(Op0, m_NUWShl(m_Value(X), m_APInt(C1))))) { 777 APInt Quotient(C1->getBitWidth(), /*val=*/0ULL, IsSigned); 778 APInt C1Shifted = APInt::getOneBitSet( 779 C1->getBitWidth(), static_cast<unsigned>(C1->getLimitedValue())); 780 781 // (X << C1) / C2 -> X / (C2 >> C1) if C2 is a multiple of 1 << C1. 782 if (isMultiple(*C2, C1Shifted, Quotient, IsSigned)) { 783 auto *BO = BinaryOperator::Create(I.getOpcode(), X, 784 ConstantInt::get(Ty, Quotient)); 785 BO->setIsExact(I.isExact()); 786 return BO; 787 } 788 789 // (X << C1) / C2 -> X * ((1 << C1) / C2) if 1 << C1 is a multiple of C2. 790 if (isMultiple(C1Shifted, *C2, Quotient, IsSigned)) { 791 auto *Mul = BinaryOperator::Create(Instruction::Mul, X, 792 ConstantInt::get(Ty, Quotient)); 793 auto *OBO = cast<OverflowingBinaryOperator>(Op0); 794 Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap()); 795 Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap()); 796 return Mul; 797 } 798 } 799 800 if (!C2->isNullValue()) // avoid X udiv 0 801 if (Instruction *FoldedDiv = foldBinOpIntoSelectOrPhi(I)) 802 return FoldedDiv; 803 } 804 805 if (match(Op0, m_One())) { 806 assert(!Ty->isIntOrIntVectorTy(1) && "i1 divide not removed?"); 807 if (IsSigned) { 808 // If Op1 is 0 then it's undefined behaviour, if Op1 is 1 then the 809 // result is one, if Op1 is -1 then the result is minus one, otherwise 810 // it's zero. 811 Value *Inc = Builder.CreateAdd(Op1, Op0); 812 Value *Cmp = Builder.CreateICmpULT(Inc, ConstantInt::get(Ty, 3)); 813 return SelectInst::Create(Cmp, Op1, ConstantInt::get(Ty, 0)); 814 } else { 815 // If Op1 is 0 then it's undefined behaviour. If Op1 is 1 then the 816 // result is one, otherwise it's zero. 817 return new ZExtInst(Builder.CreateICmpEQ(Op1, Op0), Ty); 818 } 819 } 820 821 // See if we can fold away this div instruction. 822 if (SimplifyDemandedInstructionBits(I)) 823 return &I; 824 825 // (X - (X rem Y)) / Y -> X / Y; usually originates as ((X / Y) * Y) / Y 826 Value *X, *Z; 827 if (match(Op0, m_Sub(m_Value(X), m_Value(Z)))) // (X - Z) / Y; Y = Op1 828 if ((IsSigned && match(Z, m_SRem(m_Specific(X), m_Specific(Op1)))) || 829 (!IsSigned && match(Z, m_URem(m_Specific(X), m_Specific(Op1))))) 830 return BinaryOperator::Create(I.getOpcode(), X, Op1); 831 832 // (X << Y) / X -> 1 << Y 833 Value *Y; 834 if (IsSigned && match(Op0, m_NSWShl(m_Specific(Op1), m_Value(Y)))) 835 return BinaryOperator::CreateNSWShl(ConstantInt::get(Ty, 1), Y); 836 if (!IsSigned && match(Op0, m_NUWShl(m_Specific(Op1), m_Value(Y)))) 837 return BinaryOperator::CreateNUWShl(ConstantInt::get(Ty, 1), Y); 838 839 // X / (X * Y) -> 1 / Y if the multiplication does not overflow. 840 if (match(Op1, m_c_Mul(m_Specific(Op0), m_Value(Y)))) { 841 bool HasNSW = cast<OverflowingBinaryOperator>(Op1)->hasNoSignedWrap(); 842 bool HasNUW = cast<OverflowingBinaryOperator>(Op1)->hasNoUnsignedWrap(); 843 if ((IsSigned && HasNSW) || (!IsSigned && HasNUW)) { 844 replaceOperand(I, 0, ConstantInt::get(Ty, 1)); 845 replaceOperand(I, 1, Y); 846 return &I; 847 } 848 } 849 850 return nullptr; 851 } 852 853 static const unsigned MaxDepth = 6; 854 855 namespace { 856 857 using FoldUDivOperandCb = Instruction *(*)(Value *Op0, Value *Op1, 858 const BinaryOperator &I, 859 InstCombinerImpl &IC); 860 861 /// Used to maintain state for visitUDivOperand(). 862 struct UDivFoldAction { 863 /// Informs visitUDiv() how to fold this operand. This can be zero if this 864 /// action joins two actions together. 865 FoldUDivOperandCb FoldAction; 866 867 /// Which operand to fold. 868 Value *OperandToFold; 869 870 union { 871 /// The instruction returned when FoldAction is invoked. 872 Instruction *FoldResult; 873 874 /// Stores the LHS action index if this action joins two actions together. 875 size_t SelectLHSIdx; 876 }; 877 878 UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand) 879 : FoldAction(FA), OperandToFold(InputOperand), FoldResult(nullptr) {} 880 UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand, size_t SLHS) 881 : FoldAction(FA), OperandToFold(InputOperand), SelectLHSIdx(SLHS) {} 882 }; 883 884 } // end anonymous namespace 885 886 // X udiv 2^C -> X >> C 887 static Instruction *foldUDivPow2Cst(Value *Op0, Value *Op1, 888 const BinaryOperator &I, 889 InstCombinerImpl &IC) { 890 Constant *C1 = ConstantExpr::getExactLogBase2(cast<Constant>(Op1)); 891 if (!C1) 892 llvm_unreachable("Failed to constant fold udiv -> logbase2"); 893 BinaryOperator *LShr = BinaryOperator::CreateLShr(Op0, C1); 894 if (I.isExact()) 895 LShr->setIsExact(); 896 return LShr; 897 } 898 899 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2) 900 // X udiv (zext (C1 << N)), where C1 is "1<<C2" --> X >> (N+C2) 901 static Instruction *foldUDivShl(Value *Op0, Value *Op1, const BinaryOperator &I, 902 InstCombinerImpl &IC) { 903 Value *ShiftLeft; 904 if (!match(Op1, m_ZExt(m_Value(ShiftLeft)))) 905 ShiftLeft = Op1; 906 907 Constant *CI; 908 Value *N; 909 if (!match(ShiftLeft, m_Shl(m_Constant(CI), m_Value(N)))) 910 llvm_unreachable("match should never fail here!"); 911 Constant *Log2Base = ConstantExpr::getExactLogBase2(CI); 912 if (!Log2Base) 913 llvm_unreachable("getLogBase2 should never fail here!"); 914 N = IC.Builder.CreateAdd(N, Log2Base); 915 if (Op1 != ShiftLeft) 916 N = IC.Builder.CreateZExt(N, Op1->getType()); 917 BinaryOperator *LShr = BinaryOperator::CreateLShr(Op0, N); 918 if (I.isExact()) 919 LShr->setIsExact(); 920 return LShr; 921 } 922 923 // Recursively visits the possible right hand operands of a udiv 924 // instruction, seeing through select instructions, to determine if we can 925 // replace the udiv with something simpler. If we find that an operand is not 926 // able to simplify the udiv, we abort the entire transformation. 927 static size_t visitUDivOperand(Value *Op0, Value *Op1, const BinaryOperator &I, 928 SmallVectorImpl<UDivFoldAction> &Actions, 929 unsigned Depth = 0) { 930 // FIXME: assert that Op1 isn't/doesn't contain undef. 931 932 // Check to see if this is an unsigned division with an exact power of 2, 933 // if so, convert to a right shift. 934 if (match(Op1, m_Power2())) { 935 Actions.push_back(UDivFoldAction(foldUDivPow2Cst, Op1)); 936 return Actions.size(); 937 } 938 939 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2) 940 if (match(Op1, m_Shl(m_Power2(), m_Value())) || 941 match(Op1, m_ZExt(m_Shl(m_Power2(), m_Value())))) { 942 Actions.push_back(UDivFoldAction(foldUDivShl, Op1)); 943 return Actions.size(); 944 } 945 946 // The remaining tests are all recursive, so bail out if we hit the limit. 947 if (Depth++ == MaxDepth) 948 return 0; 949 950 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 951 // FIXME: missed optimization: if one of the hands of select is/contains 952 // undef, just directly pick the other one. 953 // FIXME: can both hands contain undef? 954 if (size_t LHSIdx = 955 visitUDivOperand(Op0, SI->getOperand(1), I, Actions, Depth)) 956 if (visitUDivOperand(Op0, SI->getOperand(2), I, Actions, Depth)) { 957 Actions.push_back(UDivFoldAction(nullptr, Op1, LHSIdx - 1)); 958 return Actions.size(); 959 } 960 961 return 0; 962 } 963 964 /// If we have zero-extended operands of an unsigned div or rem, we may be able 965 /// to narrow the operation (sink the zext below the math). 966 static Instruction *narrowUDivURem(BinaryOperator &I, 967 InstCombiner::BuilderTy &Builder) { 968 Instruction::BinaryOps Opcode = I.getOpcode(); 969 Value *N = I.getOperand(0); 970 Value *D = I.getOperand(1); 971 Type *Ty = I.getType(); 972 Value *X, *Y; 973 if (match(N, m_ZExt(m_Value(X))) && match(D, m_ZExt(m_Value(Y))) && 974 X->getType() == Y->getType() && (N->hasOneUse() || D->hasOneUse())) { 975 // udiv (zext X), (zext Y) --> zext (udiv X, Y) 976 // urem (zext X), (zext Y) --> zext (urem X, Y) 977 Value *NarrowOp = Builder.CreateBinOp(Opcode, X, Y); 978 return new ZExtInst(NarrowOp, Ty); 979 } 980 981 Constant *C; 982 if ((match(N, m_OneUse(m_ZExt(m_Value(X)))) && match(D, m_Constant(C))) || 983 (match(D, m_OneUse(m_ZExt(m_Value(X)))) && match(N, m_Constant(C)))) { 984 // If the constant is the same in the smaller type, use the narrow version. 985 Constant *TruncC = ConstantExpr::getTrunc(C, X->getType()); 986 if (ConstantExpr::getZExt(TruncC, Ty) != C) 987 return nullptr; 988 989 // udiv (zext X), C --> zext (udiv X, C') 990 // urem (zext X), C --> zext (urem X, C') 991 // udiv C, (zext X) --> zext (udiv C', X) 992 // urem C, (zext X) --> zext (urem C', X) 993 Value *NarrowOp = isa<Constant>(D) ? Builder.CreateBinOp(Opcode, X, TruncC) 994 : Builder.CreateBinOp(Opcode, TruncC, X); 995 return new ZExtInst(NarrowOp, Ty); 996 } 997 998 return nullptr; 999 } 1000 1001 Instruction *InstCombinerImpl::visitUDiv(BinaryOperator &I) { 1002 if (Value *V = SimplifyUDivInst(I.getOperand(0), I.getOperand(1), 1003 SQ.getWithInstruction(&I))) 1004 return replaceInstUsesWith(I, V); 1005 1006 if (Instruction *X = foldVectorBinop(I)) 1007 return X; 1008 1009 // Handle the integer div common cases 1010 if (Instruction *Common = commonIDivTransforms(I)) 1011 return Common; 1012 1013 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1014 Value *X; 1015 const APInt *C1, *C2; 1016 if (match(Op0, m_LShr(m_Value(X), m_APInt(C1))) && match(Op1, m_APInt(C2))) { 1017 // (X lshr C1) udiv C2 --> X udiv (C2 << C1) 1018 bool Overflow; 1019 APInt C2ShlC1 = C2->ushl_ov(*C1, Overflow); 1020 if (!Overflow) { 1021 bool IsExact = I.isExact() && match(Op0, m_Exact(m_Value())); 1022 BinaryOperator *BO = BinaryOperator::CreateUDiv( 1023 X, ConstantInt::get(X->getType(), C2ShlC1)); 1024 if (IsExact) 1025 BO->setIsExact(); 1026 return BO; 1027 } 1028 } 1029 1030 // Op0 / C where C is large (negative) --> zext (Op0 >= C) 1031 // TODO: Could use isKnownNegative() to handle non-constant values. 1032 Type *Ty = I.getType(); 1033 if (match(Op1, m_Negative())) { 1034 Value *Cmp = Builder.CreateICmpUGE(Op0, Op1); 1035 return CastInst::CreateZExtOrBitCast(Cmp, Ty); 1036 } 1037 // Op0 / (sext i1 X) --> zext (Op0 == -1) (if X is 0, the div is undefined) 1038 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) { 1039 Value *Cmp = Builder.CreateICmpEQ(Op0, ConstantInt::getAllOnesValue(Ty)); 1040 return CastInst::CreateZExtOrBitCast(Cmp, Ty); 1041 } 1042 1043 if (Instruction *NarrowDiv = narrowUDivURem(I, Builder)) 1044 return NarrowDiv; 1045 1046 // If the udiv operands are non-overflowing multiplies with a common operand, 1047 // then eliminate the common factor: 1048 // (A * B) / (A * X) --> B / X (and commuted variants) 1049 // TODO: The code would be reduced if we had m_c_NUWMul pattern matching. 1050 // TODO: If -reassociation handled this generally, we could remove this. 1051 Value *A, *B; 1052 if (match(Op0, m_NUWMul(m_Value(A), m_Value(B)))) { 1053 if (match(Op1, m_NUWMul(m_Specific(A), m_Value(X))) || 1054 match(Op1, m_NUWMul(m_Value(X), m_Specific(A)))) 1055 return BinaryOperator::CreateUDiv(B, X); 1056 if (match(Op1, m_NUWMul(m_Specific(B), m_Value(X))) || 1057 match(Op1, m_NUWMul(m_Value(X), m_Specific(B)))) 1058 return BinaryOperator::CreateUDiv(A, X); 1059 } 1060 1061 // (LHS udiv (select (select (...)))) -> (LHS >> (select (select (...)))) 1062 SmallVector<UDivFoldAction, 6> UDivActions; 1063 if (visitUDivOperand(Op0, Op1, I, UDivActions)) 1064 for (unsigned i = 0, e = UDivActions.size(); i != e; ++i) { 1065 FoldUDivOperandCb Action = UDivActions[i].FoldAction; 1066 Value *ActionOp1 = UDivActions[i].OperandToFold; 1067 Instruction *Inst; 1068 if (Action) 1069 Inst = Action(Op0, ActionOp1, I, *this); 1070 else { 1071 // This action joins two actions together. The RHS of this action is 1072 // simply the last action we processed, we saved the LHS action index in 1073 // the joining action. 1074 size_t SelectRHSIdx = i - 1; 1075 Value *SelectRHS = UDivActions[SelectRHSIdx].FoldResult; 1076 size_t SelectLHSIdx = UDivActions[i].SelectLHSIdx; 1077 Value *SelectLHS = UDivActions[SelectLHSIdx].FoldResult; 1078 Inst = SelectInst::Create(cast<SelectInst>(ActionOp1)->getCondition(), 1079 SelectLHS, SelectRHS); 1080 } 1081 1082 // If this is the last action to process, return it to the InstCombiner. 1083 // Otherwise, we insert it before the UDiv and record it so that we may 1084 // use it as part of a joining action (i.e., a SelectInst). 1085 if (e - i != 1) { 1086 Inst->insertBefore(&I); 1087 UDivActions[i].FoldResult = Inst; 1088 } else 1089 return Inst; 1090 } 1091 1092 return nullptr; 1093 } 1094 1095 Instruction *InstCombinerImpl::visitSDiv(BinaryOperator &I) { 1096 if (Value *V = SimplifySDivInst(I.getOperand(0), I.getOperand(1), 1097 SQ.getWithInstruction(&I))) 1098 return replaceInstUsesWith(I, V); 1099 1100 if (Instruction *X = foldVectorBinop(I)) 1101 return X; 1102 1103 // Handle the integer div common cases 1104 if (Instruction *Common = commonIDivTransforms(I)) 1105 return Common; 1106 1107 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1108 Type *Ty = I.getType(); 1109 Value *X; 1110 // sdiv Op0, -1 --> -Op0 1111 // sdiv Op0, (sext i1 X) --> -Op0 (because if X is 0, the op is undefined) 1112 if (match(Op1, m_AllOnes()) || 1113 (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))) 1114 return BinaryOperator::CreateNeg(Op0); 1115 1116 // X / INT_MIN --> X == INT_MIN 1117 if (match(Op1, m_SignMask())) 1118 return new ZExtInst(Builder.CreateICmpEQ(Op0, Op1), Ty); 1119 1120 // sdiv exact X, 1<<C --> ashr exact X, C iff 1<<C is non-negative 1121 // sdiv exact X, -1<<C --> -(ashr exact X, C) 1122 if (I.isExact() && ((match(Op1, m_Power2()) && match(Op1, m_NonNegative())) || 1123 match(Op1, m_NegatedPower2()))) { 1124 bool DivisorWasNegative = match(Op1, m_NegatedPower2()); 1125 if (DivisorWasNegative) 1126 Op1 = ConstantExpr::getNeg(cast<Constant>(Op1)); 1127 auto *AShr = BinaryOperator::CreateExactAShr( 1128 Op0, ConstantExpr::getExactLogBase2(cast<Constant>(Op1)), I.getName()); 1129 if (!DivisorWasNegative) 1130 return AShr; 1131 Builder.Insert(AShr); 1132 AShr->setName(I.getName() + ".neg"); 1133 return BinaryOperator::CreateNeg(AShr, I.getName()); 1134 } 1135 1136 const APInt *Op1C; 1137 if (match(Op1, m_APInt(Op1C))) { 1138 // If the dividend is sign-extended and the constant divisor is small enough 1139 // to fit in the source type, shrink the division to the narrower type: 1140 // (sext X) sdiv C --> sext (X sdiv C) 1141 Value *Op0Src; 1142 if (match(Op0, m_OneUse(m_SExt(m_Value(Op0Src)))) && 1143 Op0Src->getType()->getScalarSizeInBits() >= Op1C->getMinSignedBits()) { 1144 1145 // In the general case, we need to make sure that the dividend is not the 1146 // minimum signed value because dividing that by -1 is UB. But here, we 1147 // know that the -1 divisor case is already handled above. 1148 1149 Constant *NarrowDivisor = 1150 ConstantExpr::getTrunc(cast<Constant>(Op1), Op0Src->getType()); 1151 Value *NarrowOp = Builder.CreateSDiv(Op0Src, NarrowDivisor); 1152 return new SExtInst(NarrowOp, Ty); 1153 } 1154 1155 // -X / C --> X / -C (if the negation doesn't overflow). 1156 // TODO: This could be enhanced to handle arbitrary vector constants by 1157 // checking if all elements are not the min-signed-val. 1158 if (!Op1C->isMinSignedValue() && 1159 match(Op0, m_NSWSub(m_Zero(), m_Value(X)))) { 1160 Constant *NegC = ConstantInt::get(Ty, -(*Op1C)); 1161 Instruction *BO = BinaryOperator::CreateSDiv(X, NegC); 1162 BO->setIsExact(I.isExact()); 1163 return BO; 1164 } 1165 } 1166 1167 // -X / Y --> -(X / Y) 1168 Value *Y; 1169 if (match(&I, m_SDiv(m_OneUse(m_NSWSub(m_Zero(), m_Value(X))), m_Value(Y)))) 1170 return BinaryOperator::CreateNSWNeg( 1171 Builder.CreateSDiv(X, Y, I.getName(), I.isExact())); 1172 1173 // abs(X) / X --> X > -1 ? 1 : -1 1174 // X / abs(X) --> X > -1 ? 1 : -1 1175 if (match(&I, m_c_BinOp( 1176 m_OneUse(m_Intrinsic<Intrinsic::abs>(m_Value(X), m_One())), 1177 m_Deferred(X)))) { 1178 Constant *NegOne = ConstantInt::getAllOnesValue(Ty); 1179 Value *Cond = Builder.CreateICmpSGT(X, NegOne); 1180 return SelectInst::Create(Cond, ConstantInt::get(Ty, 1), NegOne); 1181 } 1182 1183 // If the sign bits of both operands are zero (i.e. we can prove they are 1184 // unsigned inputs), turn this into a udiv. 1185 APInt Mask(APInt::getSignMask(Ty->getScalarSizeInBits())); 1186 if (MaskedValueIsZero(Op0, Mask, 0, &I)) { 1187 if (MaskedValueIsZero(Op1, Mask, 0, &I)) { 1188 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set 1189 auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName()); 1190 BO->setIsExact(I.isExact()); 1191 return BO; 1192 } 1193 1194 if (match(Op1, m_NegatedPower2())) { 1195 // X sdiv (-(1 << C)) -> -(X sdiv (1 << C)) -> 1196 // -> -(X udiv (1 << C)) -> -(X u>> C) 1197 return BinaryOperator::CreateNeg(Builder.Insert(foldUDivPow2Cst( 1198 Op0, ConstantExpr::getNeg(cast<Constant>(Op1)), I, *this))); 1199 } 1200 1201 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, &I)) { 1202 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y) 1203 // Safe because the only negative value (1 << Y) can take on is 1204 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have 1205 // the sign bit set. 1206 auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName()); 1207 BO->setIsExact(I.isExact()); 1208 return BO; 1209 } 1210 } 1211 1212 return nullptr; 1213 } 1214 1215 /// Remove negation and try to convert division into multiplication. 1216 static Instruction *foldFDivConstantDivisor(BinaryOperator &I) { 1217 Constant *C; 1218 if (!match(I.getOperand(1), m_Constant(C))) 1219 return nullptr; 1220 1221 // -X / C --> X / -C 1222 Value *X; 1223 if (match(I.getOperand(0), m_FNeg(m_Value(X)))) 1224 return BinaryOperator::CreateFDivFMF(X, ConstantExpr::getFNeg(C), &I); 1225 1226 // If the constant divisor has an exact inverse, this is always safe. If not, 1227 // then we can still create a reciprocal if fast-math-flags allow it and the 1228 // constant is a regular number (not zero, infinite, or denormal). 1229 if (!(C->hasExactInverseFP() || (I.hasAllowReciprocal() && C->isNormalFP()))) 1230 return nullptr; 1231 1232 // Disallow denormal constants because we don't know what would happen 1233 // on all targets. 1234 // TODO: Use Intrinsic::canonicalize or let function attributes tell us that 1235 // denorms are flushed? 1236 auto *RecipC = ConstantExpr::getFDiv(ConstantFP::get(I.getType(), 1.0), C); 1237 if (!RecipC->isNormalFP()) 1238 return nullptr; 1239 1240 // X / C --> X * (1 / C) 1241 return BinaryOperator::CreateFMulFMF(I.getOperand(0), RecipC, &I); 1242 } 1243 1244 /// Remove negation and try to reassociate constant math. 1245 static Instruction *foldFDivConstantDividend(BinaryOperator &I) { 1246 Constant *C; 1247 if (!match(I.getOperand(0), m_Constant(C))) 1248 return nullptr; 1249 1250 // C / -X --> -C / X 1251 Value *X; 1252 if (match(I.getOperand(1), m_FNeg(m_Value(X)))) 1253 return BinaryOperator::CreateFDivFMF(ConstantExpr::getFNeg(C), X, &I); 1254 1255 if (!I.hasAllowReassoc() || !I.hasAllowReciprocal()) 1256 return nullptr; 1257 1258 // Try to reassociate C / X expressions where X includes another constant. 1259 Constant *C2, *NewC = nullptr; 1260 if (match(I.getOperand(1), m_FMul(m_Value(X), m_Constant(C2)))) { 1261 // C / (X * C2) --> (C / C2) / X 1262 NewC = ConstantExpr::getFDiv(C, C2); 1263 } else if (match(I.getOperand(1), m_FDiv(m_Value(X), m_Constant(C2)))) { 1264 // C / (X / C2) --> (C * C2) / X 1265 NewC = ConstantExpr::getFMul(C, C2); 1266 } 1267 // Disallow denormal constants because we don't know what would happen 1268 // on all targets. 1269 // TODO: Use Intrinsic::canonicalize or let function attributes tell us that 1270 // denorms are flushed? 1271 if (!NewC || !NewC->isNormalFP()) 1272 return nullptr; 1273 1274 return BinaryOperator::CreateFDivFMF(NewC, X, &I); 1275 } 1276 1277 /// Negate the exponent of pow/exp to fold division-by-pow() into multiply. 1278 static Instruction *foldFDivPowDivisor(BinaryOperator &I, 1279 InstCombiner::BuilderTy &Builder) { 1280 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1281 auto *II = dyn_cast<IntrinsicInst>(Op1); 1282 if (!II || !II->hasOneUse() || !I.hasAllowReassoc() || 1283 !I.hasAllowReciprocal()) 1284 return nullptr; 1285 1286 // Z / pow(X, Y) --> Z * pow(X, -Y) 1287 // Z / exp{2}(Y) --> Z * exp{2}(-Y) 1288 // In the general case, this creates an extra instruction, but fmul allows 1289 // for better canonicalization and optimization than fdiv. 1290 Intrinsic::ID IID = II->getIntrinsicID(); 1291 SmallVector<Value *> Args; 1292 switch (IID) { 1293 case Intrinsic::pow: 1294 Args.push_back(II->getArgOperand(0)); 1295 Args.push_back(Builder.CreateFNegFMF(II->getArgOperand(1), &I)); 1296 break; 1297 case Intrinsic::powi: 1298 // Require 'ninf' assuming that makes powi(X, -INT_MIN) acceptable. 1299 // That is, X ** (huge negative number) is 0.0, ~1.0, or INF and so 1300 // dividing by that is INF, ~1.0, or 0.0. Code that uses powi allows 1301 // non-standard results, so this corner case should be acceptable if the 1302 // code rules out INF values. 1303 if (!I.hasNoInfs()) 1304 return nullptr; 1305 Args.push_back(II->getArgOperand(0)); 1306 Args.push_back(Builder.CreateNeg(II->getArgOperand(1))); 1307 break; 1308 case Intrinsic::exp: 1309 case Intrinsic::exp2: 1310 Args.push_back(Builder.CreateFNegFMF(II->getArgOperand(0), &I)); 1311 break; 1312 default: 1313 return nullptr; 1314 } 1315 Value *Pow = Builder.CreateIntrinsic(IID, I.getType(), Args, &I); 1316 return BinaryOperator::CreateFMulFMF(Op0, Pow, &I); 1317 } 1318 1319 Instruction *InstCombinerImpl::visitFDiv(BinaryOperator &I) { 1320 if (Value *V = SimplifyFDivInst(I.getOperand(0), I.getOperand(1), 1321 I.getFastMathFlags(), 1322 SQ.getWithInstruction(&I))) 1323 return replaceInstUsesWith(I, V); 1324 1325 if (Instruction *X = foldVectorBinop(I)) 1326 return X; 1327 1328 if (Instruction *R = foldFDivConstantDivisor(I)) 1329 return R; 1330 1331 if (Instruction *R = foldFDivConstantDividend(I)) 1332 return R; 1333 1334 if (Instruction *R = foldFPSignBitOps(I)) 1335 return R; 1336 1337 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1338 if (isa<Constant>(Op0)) 1339 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 1340 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1341 return R; 1342 1343 if (isa<Constant>(Op1)) 1344 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 1345 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1346 return R; 1347 1348 if (I.hasAllowReassoc() && I.hasAllowReciprocal()) { 1349 Value *X, *Y; 1350 if (match(Op0, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))) && 1351 (!isa<Constant>(Y) || !isa<Constant>(Op1))) { 1352 // (X / Y) / Z => X / (Y * Z) 1353 Value *YZ = Builder.CreateFMulFMF(Y, Op1, &I); 1354 return BinaryOperator::CreateFDivFMF(X, YZ, &I); 1355 } 1356 if (match(Op1, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))) && 1357 (!isa<Constant>(Y) || !isa<Constant>(Op0))) { 1358 // Z / (X / Y) => (Y * Z) / X 1359 Value *YZ = Builder.CreateFMulFMF(Y, Op0, &I); 1360 return BinaryOperator::CreateFDivFMF(YZ, X, &I); 1361 } 1362 // Z / (1.0 / Y) => (Y * Z) 1363 // 1364 // This is a special case of Z / (X / Y) => (Y * Z) / X, with X = 1.0. The 1365 // m_OneUse check is avoided because even in the case of the multiple uses 1366 // for 1.0/Y, the number of instructions remain the same and a division is 1367 // replaced by a multiplication. 1368 if (match(Op1, m_FDiv(m_SpecificFP(1.0), m_Value(Y)))) 1369 return BinaryOperator::CreateFMulFMF(Y, Op0, &I); 1370 } 1371 1372 if (I.hasAllowReassoc() && Op0->hasOneUse() && Op1->hasOneUse()) { 1373 // sin(X) / cos(X) -> tan(X) 1374 // cos(X) / sin(X) -> 1/tan(X) (cotangent) 1375 Value *X; 1376 bool IsTan = match(Op0, m_Intrinsic<Intrinsic::sin>(m_Value(X))) && 1377 match(Op1, m_Intrinsic<Intrinsic::cos>(m_Specific(X))); 1378 bool IsCot = 1379 !IsTan && match(Op0, m_Intrinsic<Intrinsic::cos>(m_Value(X))) && 1380 match(Op1, m_Intrinsic<Intrinsic::sin>(m_Specific(X))); 1381 1382 if ((IsTan || IsCot) && 1383 hasFloatFn(&TLI, I.getType(), LibFunc_tan, LibFunc_tanf, LibFunc_tanl)) { 1384 IRBuilder<> B(&I); 1385 IRBuilder<>::FastMathFlagGuard FMFGuard(B); 1386 B.setFastMathFlags(I.getFastMathFlags()); 1387 AttributeList Attrs = 1388 cast<CallBase>(Op0)->getCalledFunction()->getAttributes(); 1389 Value *Res = emitUnaryFloatFnCall(X, &TLI, LibFunc_tan, LibFunc_tanf, 1390 LibFunc_tanl, B, Attrs); 1391 if (IsCot) 1392 Res = B.CreateFDiv(ConstantFP::get(I.getType(), 1.0), Res); 1393 return replaceInstUsesWith(I, Res); 1394 } 1395 } 1396 1397 // X / (X * Y) --> 1.0 / Y 1398 // Reassociate to (X / X -> 1.0) is legal when NaNs are not allowed. 1399 // We can ignore the possibility that X is infinity because INF/INF is NaN. 1400 Value *X, *Y; 1401 if (I.hasNoNaNs() && I.hasAllowReassoc() && 1402 match(Op1, m_c_FMul(m_Specific(Op0), m_Value(Y)))) { 1403 replaceOperand(I, 0, ConstantFP::get(I.getType(), 1.0)); 1404 replaceOperand(I, 1, Y); 1405 return &I; 1406 } 1407 1408 // X / fabs(X) -> copysign(1.0, X) 1409 // fabs(X) / X -> copysign(1.0, X) 1410 if (I.hasNoNaNs() && I.hasNoInfs() && 1411 (match(&I, m_FDiv(m_Value(X), m_FAbs(m_Deferred(X)))) || 1412 match(&I, m_FDiv(m_FAbs(m_Value(X)), m_Deferred(X))))) { 1413 Value *V = Builder.CreateBinaryIntrinsic( 1414 Intrinsic::copysign, ConstantFP::get(I.getType(), 1.0), X, &I); 1415 return replaceInstUsesWith(I, V); 1416 } 1417 1418 if (Instruction *Mul = foldFDivPowDivisor(I, Builder)) 1419 return Mul; 1420 1421 return nullptr; 1422 } 1423 1424 /// This function implements the transforms common to both integer remainder 1425 /// instructions (urem and srem). It is called by the visitors to those integer 1426 /// remainder instructions. 1427 /// Common integer remainder transforms 1428 Instruction *InstCombinerImpl::commonIRemTransforms(BinaryOperator &I) { 1429 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1430 1431 // The RHS is known non-zero. 1432 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I)) 1433 return replaceOperand(I, 1, V); 1434 1435 // Handle cases involving: rem X, (select Cond, Y, Z) 1436 if (simplifyDivRemOfSelectWithZeroOp(I)) 1437 return &I; 1438 1439 if (isa<Constant>(Op1)) { 1440 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) { 1441 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) { 1442 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1443 return R; 1444 } else if (auto *PN = dyn_cast<PHINode>(Op0I)) { 1445 const APInt *Op1Int; 1446 if (match(Op1, m_APInt(Op1Int)) && !Op1Int->isMinValue() && 1447 (I.getOpcode() == Instruction::URem || 1448 !Op1Int->isMinSignedValue())) { 1449 // foldOpIntoPhi will speculate instructions to the end of the PHI's 1450 // predecessor blocks, so do this only if we know the srem or urem 1451 // will not fault. 1452 if (Instruction *NV = foldOpIntoPhi(I, PN)) 1453 return NV; 1454 } 1455 } 1456 1457 // See if we can fold away this rem instruction. 1458 if (SimplifyDemandedInstructionBits(I)) 1459 return &I; 1460 } 1461 } 1462 1463 return nullptr; 1464 } 1465 1466 Instruction *InstCombinerImpl::visitURem(BinaryOperator &I) { 1467 if (Value *V = SimplifyURemInst(I.getOperand(0), I.getOperand(1), 1468 SQ.getWithInstruction(&I))) 1469 return replaceInstUsesWith(I, V); 1470 1471 if (Instruction *X = foldVectorBinop(I)) 1472 return X; 1473 1474 if (Instruction *common = commonIRemTransforms(I)) 1475 return common; 1476 1477 if (Instruction *NarrowRem = narrowUDivURem(I, Builder)) 1478 return NarrowRem; 1479 1480 // X urem Y -> X and Y-1, where Y is a power of 2, 1481 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1482 Type *Ty = I.getType(); 1483 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, &I)) { 1484 // This may increase instruction count, we don't enforce that Y is a 1485 // constant. 1486 Constant *N1 = Constant::getAllOnesValue(Ty); 1487 Value *Add = Builder.CreateAdd(Op1, N1); 1488 return BinaryOperator::CreateAnd(Op0, Add); 1489 } 1490 1491 // 1 urem X -> zext(X != 1) 1492 if (match(Op0, m_One())) { 1493 Value *Cmp = Builder.CreateICmpNE(Op1, ConstantInt::get(Ty, 1)); 1494 return CastInst::CreateZExtOrBitCast(Cmp, Ty); 1495 } 1496 1497 // X urem C -> X < C ? X : X - C, where C >= signbit. 1498 if (match(Op1, m_Negative())) { 1499 Value *Cmp = Builder.CreateICmpULT(Op0, Op1); 1500 Value *Sub = Builder.CreateSub(Op0, Op1); 1501 return SelectInst::Create(Cmp, Op0, Sub); 1502 } 1503 1504 // If the divisor is a sext of a boolean, then the divisor must be max 1505 // unsigned value (-1). Therefore, the remainder is Op0 unless Op0 is also 1506 // max unsigned value. In that case, the remainder is 0: 1507 // urem Op0, (sext i1 X) --> (Op0 == -1) ? 0 : Op0 1508 Value *X; 1509 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) { 1510 Value *Cmp = Builder.CreateICmpEQ(Op0, ConstantInt::getAllOnesValue(Ty)); 1511 return SelectInst::Create(Cmp, ConstantInt::getNullValue(Ty), Op0); 1512 } 1513 1514 return nullptr; 1515 } 1516 1517 Instruction *InstCombinerImpl::visitSRem(BinaryOperator &I) { 1518 if (Value *V = SimplifySRemInst(I.getOperand(0), I.getOperand(1), 1519 SQ.getWithInstruction(&I))) 1520 return replaceInstUsesWith(I, V); 1521 1522 if (Instruction *X = foldVectorBinop(I)) 1523 return X; 1524 1525 // Handle the integer rem common cases 1526 if (Instruction *Common = commonIRemTransforms(I)) 1527 return Common; 1528 1529 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1530 { 1531 const APInt *Y; 1532 // X % -Y -> X % Y 1533 if (match(Op1, m_Negative(Y)) && !Y->isMinSignedValue()) 1534 return replaceOperand(I, 1, ConstantInt::get(I.getType(), -*Y)); 1535 } 1536 1537 // -X srem Y --> -(X srem Y) 1538 Value *X, *Y; 1539 if (match(&I, m_SRem(m_OneUse(m_NSWSub(m_Zero(), m_Value(X))), m_Value(Y)))) 1540 return BinaryOperator::CreateNSWNeg(Builder.CreateSRem(X, Y)); 1541 1542 // If the sign bits of both operands are zero (i.e. we can prove they are 1543 // unsigned inputs), turn this into a urem. 1544 APInt Mask(APInt::getSignMask(I.getType()->getScalarSizeInBits())); 1545 if (MaskedValueIsZero(Op1, Mask, 0, &I) && 1546 MaskedValueIsZero(Op0, Mask, 0, &I)) { 1547 // X srem Y -> X urem Y, iff X and Y don't have sign bit set 1548 return BinaryOperator::CreateURem(Op0, Op1, I.getName()); 1549 } 1550 1551 // If it's a constant vector, flip any negative values positive. 1552 if (isa<ConstantVector>(Op1) || isa<ConstantDataVector>(Op1)) { 1553 Constant *C = cast<Constant>(Op1); 1554 unsigned VWidth = cast<FixedVectorType>(C->getType())->getNumElements(); 1555 1556 bool hasNegative = false; 1557 bool hasMissing = false; 1558 for (unsigned i = 0; i != VWidth; ++i) { 1559 Constant *Elt = C->getAggregateElement(i); 1560 if (!Elt) { 1561 hasMissing = true; 1562 break; 1563 } 1564 1565 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elt)) 1566 if (RHS->isNegative()) 1567 hasNegative = true; 1568 } 1569 1570 if (hasNegative && !hasMissing) { 1571 SmallVector<Constant *, 16> Elts(VWidth); 1572 for (unsigned i = 0; i != VWidth; ++i) { 1573 Elts[i] = C->getAggregateElement(i); // Handle undef, etc. 1574 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elts[i])) { 1575 if (RHS->isNegative()) 1576 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS)); 1577 } 1578 } 1579 1580 Constant *NewRHSV = ConstantVector::get(Elts); 1581 if (NewRHSV != C) // Don't loop on -MININT 1582 return replaceOperand(I, 1, NewRHSV); 1583 } 1584 } 1585 1586 return nullptr; 1587 } 1588 1589 Instruction *InstCombinerImpl::visitFRem(BinaryOperator &I) { 1590 if (Value *V = SimplifyFRemInst(I.getOperand(0), I.getOperand(1), 1591 I.getFastMathFlags(), 1592 SQ.getWithInstruction(&I))) 1593 return replaceInstUsesWith(I, V); 1594 1595 if (Instruction *X = foldVectorBinop(I)) 1596 return X; 1597 1598 return nullptr; 1599 } 1600