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/Utils/BuildLibCalls.h" 36 #include <cassert> 37 #include <cstddef> 38 #include <cstdint> 39 #include <utility> 40 41 using namespace llvm; 42 using namespace PatternMatch; 43 44 #define DEBUG_TYPE "instcombine" 45 46 /// The specific integer value is used in a context where it is known to be 47 /// non-zero. If this allows us to simplify the computation, do so and return 48 /// the new operand, otherwise return null. 49 static Value *simplifyValueKnownNonZero(Value *V, InstCombiner &IC, 50 Instruction &CxtI) { 51 // If V has multiple uses, then we would have to do more analysis to determine 52 // if this is safe. For example, the use could be in dynamically unreached 53 // code. 54 if (!V->hasOneUse()) return nullptr; 55 56 bool MadeChange = false; 57 58 // ((1 << A) >>u B) --> (1 << (A-B)) 59 // Because V cannot be zero, we know that B is less than A. 60 Value *A = nullptr, *B = nullptr, *One = nullptr; 61 if (match(V, m_LShr(m_OneUse(m_Shl(m_Value(One), m_Value(A))), m_Value(B))) && 62 match(One, m_One())) { 63 A = IC.Builder.CreateSub(A, B); 64 return IC.Builder.CreateShl(One, A); 65 } 66 67 // (PowerOfTwo >>u B) --> isExact since shifting out the result would make it 68 // inexact. Similarly for <<. 69 BinaryOperator *I = dyn_cast<BinaryOperator>(V); 70 if (I && I->isLogicalShift() && 71 IC.isKnownToBeAPowerOfTwo(I->getOperand(0), false, 0, &CxtI)) { 72 // We know that this is an exact/nuw shift and that the input is a 73 // non-zero context as well. 74 if (Value *V2 = simplifyValueKnownNonZero(I->getOperand(0), IC, CxtI)) { 75 I->setOperand(0, V2); 76 MadeChange = true; 77 } 78 79 if (I->getOpcode() == Instruction::LShr && !I->isExact()) { 80 I->setIsExact(); 81 MadeChange = true; 82 } 83 84 if (I->getOpcode() == Instruction::Shl && !I->hasNoUnsignedWrap()) { 85 I->setHasNoUnsignedWrap(); 86 MadeChange = true; 87 } 88 } 89 90 // TODO: Lots more we could do here: 91 // If V is a phi node, we can call this on each of its operands. 92 // "select cond, X, 0" can simplify to "X". 93 94 return MadeChange ? V : nullptr; 95 } 96 97 /// A helper routine of InstCombiner::visitMul(). 98 /// 99 /// If C is a scalar/vector of known powers of 2, then this function returns 100 /// a new scalar/vector obtained from logBase2 of C. 101 /// Return a null pointer otherwise. 102 static Constant *getLogBase2(Type *Ty, Constant *C) { 103 const APInt *IVal; 104 if (match(C, m_APInt(IVal)) && IVal->isPowerOf2()) 105 return ConstantInt::get(Ty, IVal->logBase2()); 106 107 if (!Ty->isVectorTy()) 108 return nullptr; 109 110 SmallVector<Constant *, 4> Elts; 111 for (unsigned I = 0, E = Ty->getVectorNumElements(); I != E; ++I) { 112 Constant *Elt = C->getAggregateElement(I); 113 if (!Elt) 114 return nullptr; 115 if (isa<UndefValue>(Elt)) { 116 Elts.push_back(UndefValue::get(Ty->getScalarType())); 117 continue; 118 } 119 if (!match(Elt, m_APInt(IVal)) || !IVal->isPowerOf2()) 120 return nullptr; 121 Elts.push_back(ConstantInt::get(Ty->getScalarType(), IVal->logBase2())); 122 } 123 124 return ConstantVector::get(Elts); 125 } 126 127 // TODO: This is a specific form of a much more general pattern. 128 // We could detect a select with any binop identity constant, or we 129 // could use SimplifyBinOp to see if either arm of the select reduces. 130 // But that needs to be done carefully and/or while removing potential 131 // reverse canonicalizations as in InstCombiner::foldSelectIntoOp(). 132 static Value *foldMulSelectToNegate(BinaryOperator &I, 133 InstCombiner::BuilderTy &Builder) { 134 Value *Cond, *OtherOp; 135 136 // mul (select Cond, 1, -1), OtherOp --> select Cond, OtherOp, -OtherOp 137 // mul OtherOp, (select Cond, 1, -1) --> select Cond, OtherOp, -OtherOp 138 if (match(&I, m_c_Mul(m_OneUse(m_Select(m_Value(Cond), m_One(), m_AllOnes())), 139 m_Value(OtherOp)))) 140 return Builder.CreateSelect(Cond, OtherOp, Builder.CreateNeg(OtherOp)); 141 142 // mul (select Cond, -1, 1), OtherOp --> select Cond, -OtherOp, OtherOp 143 // mul OtherOp, (select Cond, -1, 1) --> select Cond, -OtherOp, OtherOp 144 if (match(&I, m_c_Mul(m_OneUse(m_Select(m_Value(Cond), m_AllOnes(), m_One())), 145 m_Value(OtherOp)))) 146 return Builder.CreateSelect(Cond, Builder.CreateNeg(OtherOp), OtherOp); 147 148 // fmul (select Cond, 1.0, -1.0), OtherOp --> select Cond, OtherOp, -OtherOp 149 // fmul OtherOp, (select Cond, 1.0, -1.0) --> select Cond, OtherOp, -OtherOp 150 if (match(&I, m_c_FMul(m_OneUse(m_Select(m_Value(Cond), m_SpecificFP(1.0), 151 m_SpecificFP(-1.0))), 152 m_Value(OtherOp)))) { 153 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); 154 Builder.setFastMathFlags(I.getFastMathFlags()); 155 return Builder.CreateSelect(Cond, OtherOp, Builder.CreateFNeg(OtherOp)); 156 } 157 158 // fmul (select Cond, -1.0, 1.0), OtherOp --> select Cond, -OtherOp, OtherOp 159 // fmul OtherOp, (select Cond, -1.0, 1.0) --> select Cond, -OtherOp, OtherOp 160 if (match(&I, m_c_FMul(m_OneUse(m_Select(m_Value(Cond), m_SpecificFP(-1.0), 161 m_SpecificFP(1.0))), 162 m_Value(OtherOp)))) { 163 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); 164 Builder.setFastMathFlags(I.getFastMathFlags()); 165 return Builder.CreateSelect(Cond, Builder.CreateFNeg(OtherOp), OtherOp); 166 } 167 168 return nullptr; 169 } 170 171 Instruction *InstCombiner::visitMul(BinaryOperator &I) { 172 if (Value *V = SimplifyMulInst(I.getOperand(0), I.getOperand(1), 173 SQ.getWithInstruction(&I))) 174 return replaceInstUsesWith(I, V); 175 176 if (SimplifyAssociativeOrCommutative(I)) 177 return &I; 178 179 if (Instruction *X = foldVectorBinop(I)) 180 return X; 181 182 if (Value *V = SimplifyUsingDistributiveLaws(I)) 183 return replaceInstUsesWith(I, V); 184 185 // X * -1 == 0 - X 186 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 187 if (match(Op1, m_AllOnes())) { 188 BinaryOperator *BO = BinaryOperator::CreateNeg(Op0, I.getName()); 189 if (I.hasNoSignedWrap()) 190 BO->setHasNoSignedWrap(); 191 return BO; 192 } 193 194 // Also allow combining multiply instructions on vectors. 195 { 196 Value *NewOp; 197 Constant *C1, *C2; 198 const APInt *IVal; 199 if (match(&I, m_Mul(m_Shl(m_Value(NewOp), m_Constant(C2)), 200 m_Constant(C1))) && 201 match(C1, m_APInt(IVal))) { 202 // ((X << C2)*C1) == (X * (C1 << C2)) 203 Constant *Shl = ConstantExpr::getShl(C1, C2); 204 BinaryOperator *Mul = cast<BinaryOperator>(I.getOperand(0)); 205 BinaryOperator *BO = BinaryOperator::CreateMul(NewOp, Shl); 206 if (I.hasNoUnsignedWrap() && Mul->hasNoUnsignedWrap()) 207 BO->setHasNoUnsignedWrap(); 208 if (I.hasNoSignedWrap() && Mul->hasNoSignedWrap() && 209 Shl->isNotMinSignedValue()) 210 BO->setHasNoSignedWrap(); 211 return BO; 212 } 213 214 if (match(&I, m_Mul(m_Value(NewOp), m_Constant(C1)))) { 215 // Replace X*(2^C) with X << C, where C is either a scalar or a vector. 216 if (Constant *NewCst = getLogBase2(NewOp->getType(), C1)) { 217 BinaryOperator *Shl = BinaryOperator::CreateShl(NewOp, NewCst); 218 219 if (I.hasNoUnsignedWrap()) 220 Shl->setHasNoUnsignedWrap(); 221 if (I.hasNoSignedWrap()) { 222 const APInt *V; 223 if (match(NewCst, m_APInt(V)) && *V != V->getBitWidth() - 1) 224 Shl->setHasNoSignedWrap(); 225 } 226 227 return Shl; 228 } 229 } 230 } 231 232 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 233 // (Y - X) * (-(2**n)) -> (X - Y) * (2**n), for positive nonzero n 234 // (Y + const) * (-(2**n)) -> (-constY) * (2**n), for positive nonzero n 235 // The "* (2**n)" thus becomes a potential shifting opportunity. 236 { 237 const APInt & Val = CI->getValue(); 238 const APInt &PosVal = Val.abs(); 239 if (Val.isNegative() && PosVal.isPowerOf2()) { 240 Value *X = nullptr, *Y = nullptr; 241 if (Op0->hasOneUse()) { 242 ConstantInt *C1; 243 Value *Sub = nullptr; 244 if (match(Op0, m_Sub(m_Value(Y), m_Value(X)))) 245 Sub = Builder.CreateSub(X, Y, "suba"); 246 else if (match(Op0, m_Add(m_Value(Y), m_ConstantInt(C1)))) 247 Sub = Builder.CreateSub(Builder.CreateNeg(C1), Y, "subc"); 248 if (Sub) 249 return 250 BinaryOperator::CreateMul(Sub, 251 ConstantInt::get(Y->getType(), PosVal)); 252 } 253 } 254 } 255 } 256 257 if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I)) 258 return FoldedMul; 259 260 if (Value *FoldedMul = foldMulSelectToNegate(I, Builder)) 261 return replaceInstUsesWith(I, FoldedMul); 262 263 // Simplify mul instructions with a constant RHS. 264 if (isa<Constant>(Op1)) { 265 // Canonicalize (X+C1)*CI -> X*CI+C1*CI. 266 Value *X; 267 Constant *C1; 268 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_Constant(C1))))) { 269 Value *Mul = Builder.CreateMul(C1, Op1); 270 // Only go forward with the transform if C1*CI simplifies to a tidier 271 // constant. 272 if (!match(Mul, m_Mul(m_Value(), m_Value()))) 273 return BinaryOperator::CreateAdd(Builder.CreateMul(X, Op1), Mul); 274 } 275 } 276 277 // -X * C --> X * -C 278 Value *X, *Y; 279 Constant *Op1C; 280 if (match(Op0, m_Neg(m_Value(X))) && match(Op1, m_Constant(Op1C))) 281 return BinaryOperator::CreateMul(X, ConstantExpr::getNeg(Op1C)); 282 283 // -X * -Y --> X * Y 284 if (match(Op0, m_Neg(m_Value(X))) && match(Op1, m_Neg(m_Value(Y)))) { 285 auto *NewMul = BinaryOperator::CreateMul(X, Y); 286 if (I.hasNoSignedWrap() && 287 cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap() && 288 cast<OverflowingBinaryOperator>(Op1)->hasNoSignedWrap()) 289 NewMul->setHasNoSignedWrap(); 290 return NewMul; 291 } 292 293 // -X * Y --> -(X * Y) 294 // X * -Y --> -(X * Y) 295 if (match(&I, m_c_Mul(m_OneUse(m_Neg(m_Value(X))), m_Value(Y)))) 296 return BinaryOperator::CreateNeg(Builder.CreateMul(X, Y)); 297 298 // (X / Y) * Y = X - (X % Y) 299 // (X / Y) * -Y = (X % Y) - X 300 { 301 Value *Y = Op1; 302 BinaryOperator *Div = dyn_cast<BinaryOperator>(Op0); 303 if (!Div || (Div->getOpcode() != Instruction::UDiv && 304 Div->getOpcode() != Instruction::SDiv)) { 305 Y = Op0; 306 Div = dyn_cast<BinaryOperator>(Op1); 307 } 308 Value *Neg = dyn_castNegVal(Y); 309 if (Div && Div->hasOneUse() && 310 (Div->getOperand(1) == Y || Div->getOperand(1) == Neg) && 311 (Div->getOpcode() == Instruction::UDiv || 312 Div->getOpcode() == Instruction::SDiv)) { 313 Value *X = Div->getOperand(0), *DivOp1 = Div->getOperand(1); 314 315 // If the division is exact, X % Y is zero, so we end up with X or -X. 316 if (Div->isExact()) { 317 if (DivOp1 == Y) 318 return replaceInstUsesWith(I, X); 319 return BinaryOperator::CreateNeg(X); 320 } 321 322 auto RemOpc = Div->getOpcode() == Instruction::UDiv ? Instruction::URem 323 : Instruction::SRem; 324 Value *Rem = Builder.CreateBinOp(RemOpc, X, DivOp1); 325 if (DivOp1 == Y) 326 return BinaryOperator::CreateSub(X, Rem); 327 return BinaryOperator::CreateSub(Rem, X); 328 } 329 } 330 331 /// i1 mul -> i1 and. 332 if (I.getType()->isIntOrIntVectorTy(1)) 333 return BinaryOperator::CreateAnd(Op0, Op1); 334 335 // X*(1 << Y) --> X << Y 336 // (1 << Y)*X --> X << Y 337 { 338 Value *Y; 339 BinaryOperator *BO = nullptr; 340 bool ShlNSW = false; 341 if (match(Op0, m_Shl(m_One(), m_Value(Y)))) { 342 BO = BinaryOperator::CreateShl(Op1, Y); 343 ShlNSW = cast<ShlOperator>(Op0)->hasNoSignedWrap(); 344 } else if (match(Op1, m_Shl(m_One(), m_Value(Y)))) { 345 BO = BinaryOperator::CreateShl(Op0, Y); 346 ShlNSW = cast<ShlOperator>(Op1)->hasNoSignedWrap(); 347 } 348 if (BO) { 349 if (I.hasNoUnsignedWrap()) 350 BO->setHasNoUnsignedWrap(); 351 if (I.hasNoSignedWrap() && ShlNSW) 352 BO->setHasNoSignedWrap(); 353 return BO; 354 } 355 } 356 357 // (bool X) * Y --> X ? Y : 0 358 // Y * (bool X) --> X ? Y : 0 359 if (match(Op0, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) 360 return SelectInst::Create(X, Op1, ConstantInt::get(I.getType(), 0)); 361 if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) 362 return SelectInst::Create(X, Op0, ConstantInt::get(I.getType(), 0)); 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 if (Instruction *Ext = narrowMathIfNoOverflow(I)) 377 return Ext; 378 379 bool Changed = false; 380 if (!I.hasNoSignedWrap() && willNotOverflowSignedMul(Op0, Op1, I)) { 381 Changed = true; 382 I.setHasNoSignedWrap(true); 383 } 384 385 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedMul(Op0, Op1, I)) { 386 Changed = true; 387 I.setHasNoUnsignedWrap(true); 388 } 389 390 return Changed ? &I : nullptr; 391 } 392 393 Instruction *InstCombiner::visitFMul(BinaryOperator &I) { 394 if (Value *V = SimplifyFMulInst(I.getOperand(0), I.getOperand(1), 395 I.getFastMathFlags(), 396 SQ.getWithInstruction(&I))) 397 return replaceInstUsesWith(I, V); 398 399 if (SimplifyAssociativeOrCommutative(I)) 400 return &I; 401 402 if (Instruction *X = foldVectorBinop(I)) 403 return X; 404 405 if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I)) 406 return FoldedMul; 407 408 if (Value *FoldedMul = foldMulSelectToNegate(I, Builder)) 409 return replaceInstUsesWith(I, FoldedMul); 410 411 // X * -1.0 --> -X 412 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 413 if (match(Op1, m_SpecificFP(-1.0))) 414 return BinaryOperator::CreateFNegFMF(Op0, &I); 415 416 // -X * -Y --> X * Y 417 Value *X, *Y; 418 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y)))) 419 return BinaryOperator::CreateFMulFMF(X, Y, &I); 420 421 // -X * C --> X * -C 422 Constant *C; 423 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_Constant(C))) 424 return BinaryOperator::CreateFMulFMF(X, ConstantExpr::getFNeg(C), &I); 425 426 // fabs(X) * fabs(X) -> X * X 427 if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::fabs>(m_Value(X)))) 428 return BinaryOperator::CreateFMulFMF(X, X, &I); 429 430 // (select A, B, C) * (select A, D, E) --> select A, (B*D), (C*E) 431 if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1)) 432 return replaceInstUsesWith(I, V); 433 434 if (I.hasAllowReassoc()) { 435 // Reassociate constant RHS with another constant to form constant 436 // expression. 437 if (match(Op1, m_Constant(C)) && C->isFiniteNonZeroFP()) { 438 Constant *C1; 439 if (match(Op0, m_OneUse(m_FDiv(m_Constant(C1), m_Value(X))))) { 440 // (C1 / X) * C --> (C * C1) / X 441 Constant *CC1 = ConstantExpr::getFMul(C, C1); 442 if (CC1->isNormalFP()) 443 return BinaryOperator::CreateFDivFMF(CC1, X, &I); 444 } 445 if (match(Op0, m_FDiv(m_Value(X), m_Constant(C1)))) { 446 // (X / C1) * C --> X * (C / C1) 447 Constant *CDivC1 = ConstantExpr::getFDiv(C, C1); 448 if (CDivC1->isNormalFP()) 449 return BinaryOperator::CreateFMulFMF(X, CDivC1, &I); 450 451 // If the constant was a denormal, try reassociating differently. 452 // (X / C1) * C --> X / (C1 / C) 453 Constant *C1DivC = ConstantExpr::getFDiv(C1, C); 454 if (Op0->hasOneUse() && C1DivC->isNormalFP()) 455 return BinaryOperator::CreateFDivFMF(X, C1DivC, &I); 456 } 457 458 // We do not need to match 'fadd C, X' and 'fsub X, C' because they are 459 // canonicalized to 'fadd X, C'. Distributing the multiply may allow 460 // further folds and (X * C) + C2 is 'fma'. 461 if (match(Op0, m_OneUse(m_FAdd(m_Value(X), m_Constant(C1))))) { 462 // (X + C1) * C --> (X * C) + (C * C1) 463 Constant *CC1 = ConstantExpr::getFMul(C, C1); 464 Value *XC = Builder.CreateFMulFMF(X, C, &I); 465 return BinaryOperator::CreateFAddFMF(XC, CC1, &I); 466 } 467 if (match(Op0, m_OneUse(m_FSub(m_Constant(C1), m_Value(X))))) { 468 // (C1 - X) * C --> (C * C1) - (X * C) 469 Constant *CC1 = ConstantExpr::getFMul(C, C1); 470 Value *XC = Builder.CreateFMulFMF(X, C, &I); 471 return BinaryOperator::CreateFSubFMF(CC1, XC, &I); 472 } 473 } 474 475 Value *Z; 476 if (match(&I, m_c_FMul(m_OneUse(m_FDiv(m_Value(X), m_Value(Y))), 477 m_Value(Z)))) { 478 // Sink division: (X / Y) * Z --> (X * Z) / Y 479 Value *NewFMul = Builder.CreateFMulFMF(X, Z, &I); 480 return BinaryOperator::CreateFDivFMF(NewFMul, Y, &I); 481 } 482 483 // sqrt(X) * sqrt(Y) -> sqrt(X * Y) 484 // nnan disallows the possibility of returning a number if both operands are 485 // negative (in that case, we should return NaN). 486 if (I.hasNoNaNs() && 487 match(Op0, m_OneUse(m_Intrinsic<Intrinsic::sqrt>(m_Value(X)))) && 488 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::sqrt>(m_Value(Y))))) { 489 Value *XY = Builder.CreateFMulFMF(X, Y, &I); 490 Value *Sqrt = Builder.CreateUnaryIntrinsic(Intrinsic::sqrt, XY, &I); 491 return replaceInstUsesWith(I, Sqrt); 492 } 493 494 // Like the similar transform in instsimplify, this requires 'nsz' because 495 // sqrt(-0.0) = -0.0, and -0.0 * -0.0 does not simplify to -0.0. 496 if (I.hasNoNaNs() && I.hasNoSignedZeros() && Op0 == Op1 && 497 Op0->hasNUses(2)) { 498 // Peek through fdiv to find squaring of square root: 499 // (X / sqrt(Y)) * (X / sqrt(Y)) --> (X * X) / Y 500 if (match(Op0, m_FDiv(m_Value(X), 501 m_Intrinsic<Intrinsic::sqrt>(m_Value(Y))))) { 502 Value *XX = Builder.CreateFMulFMF(X, X, &I); 503 return BinaryOperator::CreateFDivFMF(XX, Y, &I); 504 } 505 // (sqrt(Y) / X) * (sqrt(Y) / X) --> Y / (X * X) 506 if (match(Op0, m_FDiv(m_Intrinsic<Intrinsic::sqrt>(m_Value(Y)), 507 m_Value(X)))) { 508 Value *XX = Builder.CreateFMulFMF(X, X, &I); 509 return BinaryOperator::CreateFDivFMF(Y, XX, &I); 510 } 511 } 512 513 // exp(X) * exp(Y) -> exp(X + Y) 514 // Match as long as at least one of exp has only one use. 515 if (match(Op0, m_Intrinsic<Intrinsic::exp>(m_Value(X))) && 516 match(Op1, m_Intrinsic<Intrinsic::exp>(m_Value(Y))) && 517 (Op0->hasOneUse() || Op1->hasOneUse())) { 518 Value *XY = Builder.CreateFAddFMF(X, Y, &I); 519 Value *Exp = Builder.CreateUnaryIntrinsic(Intrinsic::exp, XY, &I); 520 return replaceInstUsesWith(I, Exp); 521 } 522 523 // exp2(X) * exp2(Y) -> exp2(X + Y) 524 // Match as long as at least one of exp2 has only one use. 525 if (match(Op0, m_Intrinsic<Intrinsic::exp2>(m_Value(X))) && 526 match(Op1, m_Intrinsic<Intrinsic::exp2>(m_Value(Y))) && 527 (Op0->hasOneUse() || Op1->hasOneUse())) { 528 Value *XY = Builder.CreateFAddFMF(X, Y, &I); 529 Value *Exp2 = Builder.CreateUnaryIntrinsic(Intrinsic::exp2, XY, &I); 530 return replaceInstUsesWith(I, Exp2); 531 } 532 533 // (X*Y) * X => (X*X) * Y where Y != X 534 // The purpose is two-fold: 535 // 1) to form a power expression (of X). 536 // 2) potentially shorten the critical path: After transformation, the 537 // latency of the instruction Y is amortized by the expression of X*X, 538 // and therefore Y is in a "less critical" position compared to what it 539 // was before the transformation. 540 if (match(Op0, m_OneUse(m_c_FMul(m_Specific(Op1), m_Value(Y)))) && 541 Op1 != Y) { 542 Value *XX = Builder.CreateFMulFMF(Op1, Op1, &I); 543 return BinaryOperator::CreateFMulFMF(XX, Y, &I); 544 } 545 if (match(Op1, m_OneUse(m_c_FMul(m_Specific(Op0), m_Value(Y)))) && 546 Op0 != Y) { 547 Value *XX = Builder.CreateFMulFMF(Op0, Op0, &I); 548 return BinaryOperator::CreateFMulFMF(XX, Y, &I); 549 } 550 } 551 552 // log2(X * 0.5) * Y = log2(X) * Y - Y 553 if (I.isFast()) { 554 IntrinsicInst *Log2 = nullptr; 555 if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::log2>( 556 m_OneUse(m_FMul(m_Value(X), m_SpecificFP(0.5))))))) { 557 Log2 = cast<IntrinsicInst>(Op0); 558 Y = Op1; 559 } 560 if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::log2>( 561 m_OneUse(m_FMul(m_Value(X), m_SpecificFP(0.5))))))) { 562 Log2 = cast<IntrinsicInst>(Op1); 563 Y = Op0; 564 } 565 if (Log2) { 566 Log2->setArgOperand(0, X); 567 Log2->copyFastMathFlags(&I); 568 Value *LogXTimesY = Builder.CreateFMulFMF(Log2, Y, &I); 569 return BinaryOperator::CreateFSubFMF(LogXTimesY, Y, &I); 570 } 571 } 572 573 return nullptr; 574 } 575 576 /// Fold a divide or remainder with a select instruction divisor when one of the 577 /// select operands is zero. In that case, we can use the other select operand 578 /// because div/rem by zero is undefined. 579 bool InstCombiner::simplifyDivRemOfSelectWithZeroOp(BinaryOperator &I) { 580 SelectInst *SI = dyn_cast<SelectInst>(I.getOperand(1)); 581 if (!SI) 582 return false; 583 584 int NonNullOperand; 585 if (match(SI->getTrueValue(), m_Zero())) 586 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y 587 NonNullOperand = 2; 588 else if (match(SI->getFalseValue(), m_Zero())) 589 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y 590 NonNullOperand = 1; 591 else 592 return false; 593 594 // Change the div/rem to use 'Y' instead of the select. 595 I.setOperand(1, SI->getOperand(NonNullOperand)); 596 597 // Okay, we know we replace the operand of the div/rem with 'Y' with no 598 // problem. However, the select, or the condition of the select may have 599 // multiple uses. Based on our knowledge that the operand must be non-zero, 600 // propagate the known value for the select into other uses of it, and 601 // propagate a known value of the condition into its other users. 602 603 // If the select and condition only have a single use, don't bother with this, 604 // early exit. 605 Value *SelectCond = SI->getCondition(); 606 if (SI->use_empty() && SelectCond->hasOneUse()) 607 return true; 608 609 // Scan the current block backward, looking for other uses of SI. 610 BasicBlock::iterator BBI = I.getIterator(), BBFront = I.getParent()->begin(); 611 Type *CondTy = SelectCond->getType(); 612 while (BBI != BBFront) { 613 --BBI; 614 // If we found an instruction that we can't assume will return, so 615 // information from below it cannot be propagated above it. 616 if (!isGuaranteedToTransferExecutionToSuccessor(&*BBI)) 617 break; 618 619 // Replace uses of the select or its condition with the known values. 620 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end(); 621 I != E; ++I) { 622 if (*I == SI) { 623 *I = SI->getOperand(NonNullOperand); 624 Worklist.Add(&*BBI); 625 } else if (*I == SelectCond) { 626 *I = NonNullOperand == 1 ? ConstantInt::getTrue(CondTy) 627 : ConstantInt::getFalse(CondTy); 628 Worklist.Add(&*BBI); 629 } 630 } 631 632 // If we past the instruction, quit looking for it. 633 if (&*BBI == SI) 634 SI = nullptr; 635 if (&*BBI == SelectCond) 636 SelectCond = nullptr; 637 638 // If we ran out of things to eliminate, break out of the loop. 639 if (!SelectCond && !SI) 640 break; 641 642 } 643 return true; 644 } 645 646 /// True if the multiply can not be expressed in an int this size. 647 static bool multiplyOverflows(const APInt &C1, const APInt &C2, APInt &Product, 648 bool IsSigned) { 649 bool Overflow; 650 Product = IsSigned ? C1.smul_ov(C2, Overflow) : C1.umul_ov(C2, Overflow); 651 return Overflow; 652 } 653 654 /// True if C1 is a multiple of C2. Quotient contains C1/C2. 655 static bool isMultiple(const APInt &C1, const APInt &C2, APInt &Quotient, 656 bool IsSigned) { 657 assert(C1.getBitWidth() == C2.getBitWidth() && "Constant widths not equal"); 658 659 // Bail if we will divide by zero. 660 if (C2.isNullValue()) 661 return false; 662 663 // Bail if we would divide INT_MIN by -1. 664 if (IsSigned && C1.isMinSignedValue() && C2.isAllOnesValue()) 665 return false; 666 667 APInt Remainder(C1.getBitWidth(), /*val=*/0ULL, IsSigned); 668 if (IsSigned) 669 APInt::sdivrem(C1, C2, Quotient, Remainder); 670 else 671 APInt::udivrem(C1, C2, Quotient, Remainder); 672 673 return Remainder.isMinValue(); 674 } 675 676 /// This function implements the transforms common to both integer division 677 /// instructions (udiv and sdiv). It is called by the visitors to those integer 678 /// division instructions. 679 /// Common integer divide transforms 680 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) { 681 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 682 bool IsSigned = I.getOpcode() == Instruction::SDiv; 683 Type *Ty = I.getType(); 684 685 // The RHS is known non-zero. 686 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I)) { 687 I.setOperand(1, V); 688 return &I; 689 } 690 691 // Handle cases involving: [su]div X, (select Cond, Y, Z) 692 // This does not apply for fdiv. 693 if (simplifyDivRemOfSelectWithZeroOp(I)) 694 return &I; 695 696 const APInt *C2; 697 if (match(Op1, m_APInt(C2))) { 698 Value *X; 699 const APInt *C1; 700 701 // (X / C1) / C2 -> X / (C1*C2) 702 if ((IsSigned && match(Op0, m_SDiv(m_Value(X), m_APInt(C1)))) || 703 (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_APInt(C1))))) { 704 APInt Product(C1->getBitWidth(), /*val=*/0ULL, IsSigned); 705 if (!multiplyOverflows(*C1, *C2, Product, IsSigned)) 706 return BinaryOperator::Create(I.getOpcode(), X, 707 ConstantInt::get(Ty, Product)); 708 } 709 710 if ((IsSigned && match(Op0, m_NSWMul(m_Value(X), m_APInt(C1)))) || 711 (!IsSigned && match(Op0, m_NUWMul(m_Value(X), m_APInt(C1))))) { 712 APInt Quotient(C1->getBitWidth(), /*val=*/0ULL, IsSigned); 713 714 // (X * C1) / C2 -> X / (C2 / C1) if C2 is a multiple of C1. 715 if (isMultiple(*C2, *C1, Quotient, IsSigned)) { 716 auto *NewDiv = BinaryOperator::Create(I.getOpcode(), X, 717 ConstantInt::get(Ty, Quotient)); 718 NewDiv->setIsExact(I.isExact()); 719 return NewDiv; 720 } 721 722 // (X * C1) / C2 -> X * (C1 / C2) if C1 is a multiple of C2. 723 if (isMultiple(*C1, *C2, Quotient, IsSigned)) { 724 auto *Mul = BinaryOperator::Create(Instruction::Mul, X, 725 ConstantInt::get(Ty, Quotient)); 726 auto *OBO = cast<OverflowingBinaryOperator>(Op0); 727 Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap()); 728 Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap()); 729 return Mul; 730 } 731 } 732 733 if ((IsSigned && match(Op0, m_NSWShl(m_Value(X), m_APInt(C1))) && 734 *C1 != C1->getBitWidth() - 1) || 735 (!IsSigned && match(Op0, m_NUWShl(m_Value(X), m_APInt(C1))))) { 736 APInt Quotient(C1->getBitWidth(), /*val=*/0ULL, IsSigned); 737 APInt C1Shifted = APInt::getOneBitSet( 738 C1->getBitWidth(), static_cast<unsigned>(C1->getLimitedValue())); 739 740 // (X << C1) / C2 -> X / (C2 >> C1) if C2 is a multiple of 1 << C1. 741 if (isMultiple(*C2, C1Shifted, Quotient, IsSigned)) { 742 auto *BO = BinaryOperator::Create(I.getOpcode(), X, 743 ConstantInt::get(Ty, Quotient)); 744 BO->setIsExact(I.isExact()); 745 return BO; 746 } 747 748 // (X << C1) / C2 -> X * ((1 << C1) / C2) if 1 << C1 is a multiple of C2. 749 if (isMultiple(C1Shifted, *C2, Quotient, IsSigned)) { 750 auto *Mul = BinaryOperator::Create(Instruction::Mul, X, 751 ConstantInt::get(Ty, Quotient)); 752 auto *OBO = cast<OverflowingBinaryOperator>(Op0); 753 Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap()); 754 Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap()); 755 return Mul; 756 } 757 } 758 759 if (!C2->isNullValue()) // avoid X udiv 0 760 if (Instruction *FoldedDiv = foldBinOpIntoSelectOrPhi(I)) 761 return FoldedDiv; 762 } 763 764 if (match(Op0, m_One())) { 765 assert(!Ty->isIntOrIntVectorTy(1) && "i1 divide not removed?"); 766 if (IsSigned) { 767 // If Op1 is 0 then it's undefined behaviour, if Op1 is 1 then the 768 // result is one, if Op1 is -1 then the result is minus one, otherwise 769 // it's zero. 770 Value *Inc = Builder.CreateAdd(Op1, Op0); 771 Value *Cmp = Builder.CreateICmpULT(Inc, ConstantInt::get(Ty, 3)); 772 return SelectInst::Create(Cmp, Op1, ConstantInt::get(Ty, 0)); 773 } else { 774 // If Op1 is 0 then it's undefined behaviour. If Op1 is 1 then the 775 // result is one, otherwise it's zero. 776 return new ZExtInst(Builder.CreateICmpEQ(Op1, Op0), Ty); 777 } 778 } 779 780 // See if we can fold away this div instruction. 781 if (SimplifyDemandedInstructionBits(I)) 782 return &I; 783 784 // (X - (X rem Y)) / Y -> X / Y; usually originates as ((X / Y) * Y) / Y 785 Value *X, *Z; 786 if (match(Op0, m_Sub(m_Value(X), m_Value(Z)))) // (X - Z) / Y; Y = Op1 787 if ((IsSigned && match(Z, m_SRem(m_Specific(X), m_Specific(Op1)))) || 788 (!IsSigned && match(Z, m_URem(m_Specific(X), m_Specific(Op1))))) 789 return BinaryOperator::Create(I.getOpcode(), X, Op1); 790 791 // (X << Y) / X -> 1 << Y 792 Value *Y; 793 if (IsSigned && match(Op0, m_NSWShl(m_Specific(Op1), m_Value(Y)))) 794 return BinaryOperator::CreateNSWShl(ConstantInt::get(Ty, 1), Y); 795 if (!IsSigned && match(Op0, m_NUWShl(m_Specific(Op1), m_Value(Y)))) 796 return BinaryOperator::CreateNUWShl(ConstantInt::get(Ty, 1), Y); 797 798 // X / (X * Y) -> 1 / Y if the multiplication does not overflow. 799 if (match(Op1, m_c_Mul(m_Specific(Op0), m_Value(Y)))) { 800 bool HasNSW = cast<OverflowingBinaryOperator>(Op1)->hasNoSignedWrap(); 801 bool HasNUW = cast<OverflowingBinaryOperator>(Op1)->hasNoUnsignedWrap(); 802 if ((IsSigned && HasNSW) || (!IsSigned && HasNUW)) { 803 I.setOperand(0, ConstantInt::get(Ty, 1)); 804 I.setOperand(1, Y); 805 return &I; 806 } 807 } 808 809 return nullptr; 810 } 811 812 static const unsigned MaxDepth = 6; 813 814 namespace { 815 816 using FoldUDivOperandCb = Instruction *(*)(Value *Op0, Value *Op1, 817 const BinaryOperator &I, 818 InstCombiner &IC); 819 820 /// Used to maintain state for visitUDivOperand(). 821 struct UDivFoldAction { 822 /// Informs visitUDiv() how to fold this operand. This can be zero if this 823 /// action joins two actions together. 824 FoldUDivOperandCb FoldAction; 825 826 /// Which operand to fold. 827 Value *OperandToFold; 828 829 union { 830 /// The instruction returned when FoldAction is invoked. 831 Instruction *FoldResult; 832 833 /// Stores the LHS action index if this action joins two actions together. 834 size_t SelectLHSIdx; 835 }; 836 837 UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand) 838 : FoldAction(FA), OperandToFold(InputOperand), FoldResult(nullptr) {} 839 UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand, size_t SLHS) 840 : FoldAction(FA), OperandToFold(InputOperand), SelectLHSIdx(SLHS) {} 841 }; 842 843 } // end anonymous namespace 844 845 // X udiv 2^C -> X >> C 846 static Instruction *foldUDivPow2Cst(Value *Op0, Value *Op1, 847 const BinaryOperator &I, InstCombiner &IC) { 848 Constant *C1 = getLogBase2(Op0->getType(), cast<Constant>(Op1)); 849 if (!C1) 850 llvm_unreachable("Failed to constant fold udiv -> logbase2"); 851 BinaryOperator *LShr = BinaryOperator::CreateLShr(Op0, C1); 852 if (I.isExact()) 853 LShr->setIsExact(); 854 return LShr; 855 } 856 857 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2) 858 // X udiv (zext (C1 << N)), where C1 is "1<<C2" --> X >> (N+C2) 859 static Instruction *foldUDivShl(Value *Op0, Value *Op1, const BinaryOperator &I, 860 InstCombiner &IC) { 861 Value *ShiftLeft; 862 if (!match(Op1, m_ZExt(m_Value(ShiftLeft)))) 863 ShiftLeft = Op1; 864 865 Constant *CI; 866 Value *N; 867 if (!match(ShiftLeft, m_Shl(m_Constant(CI), m_Value(N)))) 868 llvm_unreachable("match should never fail here!"); 869 Constant *Log2Base = getLogBase2(N->getType(), CI); 870 if (!Log2Base) 871 llvm_unreachable("getLogBase2 should never fail here!"); 872 N = IC.Builder.CreateAdd(N, Log2Base); 873 if (Op1 != ShiftLeft) 874 N = IC.Builder.CreateZExt(N, Op1->getType()); 875 BinaryOperator *LShr = BinaryOperator::CreateLShr(Op0, N); 876 if (I.isExact()) 877 LShr->setIsExact(); 878 return LShr; 879 } 880 881 // Recursively visits the possible right hand operands of a udiv 882 // instruction, seeing through select instructions, to determine if we can 883 // replace the udiv with something simpler. If we find that an operand is not 884 // able to simplify the udiv, we abort the entire transformation. 885 static size_t visitUDivOperand(Value *Op0, Value *Op1, const BinaryOperator &I, 886 SmallVectorImpl<UDivFoldAction> &Actions, 887 unsigned Depth = 0) { 888 // Check to see if this is an unsigned division with an exact power of 2, 889 // if so, convert to a right shift. 890 if (match(Op1, m_Power2())) { 891 Actions.push_back(UDivFoldAction(foldUDivPow2Cst, Op1)); 892 return Actions.size(); 893 } 894 895 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2) 896 if (match(Op1, m_Shl(m_Power2(), m_Value())) || 897 match(Op1, m_ZExt(m_Shl(m_Power2(), m_Value())))) { 898 Actions.push_back(UDivFoldAction(foldUDivShl, Op1)); 899 return Actions.size(); 900 } 901 902 // The remaining tests are all recursive, so bail out if we hit the limit. 903 if (Depth++ == MaxDepth) 904 return 0; 905 906 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 907 if (size_t LHSIdx = 908 visitUDivOperand(Op0, SI->getOperand(1), I, Actions, Depth)) 909 if (visitUDivOperand(Op0, SI->getOperand(2), I, Actions, Depth)) { 910 Actions.push_back(UDivFoldAction(nullptr, Op1, LHSIdx - 1)); 911 return Actions.size(); 912 } 913 914 return 0; 915 } 916 917 /// If we have zero-extended operands of an unsigned div or rem, we may be able 918 /// to narrow the operation (sink the zext below the math). 919 static Instruction *narrowUDivURem(BinaryOperator &I, 920 InstCombiner::BuilderTy &Builder) { 921 Instruction::BinaryOps Opcode = I.getOpcode(); 922 Value *N = I.getOperand(0); 923 Value *D = I.getOperand(1); 924 Type *Ty = I.getType(); 925 Value *X, *Y; 926 if (match(N, m_ZExt(m_Value(X))) && match(D, m_ZExt(m_Value(Y))) && 927 X->getType() == Y->getType() && (N->hasOneUse() || D->hasOneUse())) { 928 // udiv (zext X), (zext Y) --> zext (udiv X, Y) 929 // urem (zext X), (zext Y) --> zext (urem X, Y) 930 Value *NarrowOp = Builder.CreateBinOp(Opcode, X, Y); 931 return new ZExtInst(NarrowOp, Ty); 932 } 933 934 Constant *C; 935 if ((match(N, m_OneUse(m_ZExt(m_Value(X)))) && match(D, m_Constant(C))) || 936 (match(D, m_OneUse(m_ZExt(m_Value(X)))) && match(N, m_Constant(C)))) { 937 // If the constant is the same in the smaller type, use the narrow version. 938 Constant *TruncC = ConstantExpr::getTrunc(C, X->getType()); 939 if (ConstantExpr::getZExt(TruncC, Ty) != C) 940 return nullptr; 941 942 // udiv (zext X), C --> zext (udiv X, C') 943 // urem (zext X), C --> zext (urem X, C') 944 // udiv C, (zext X) --> zext (udiv C', X) 945 // urem C, (zext X) --> zext (urem C', X) 946 Value *NarrowOp = isa<Constant>(D) ? Builder.CreateBinOp(Opcode, X, TruncC) 947 : Builder.CreateBinOp(Opcode, TruncC, X); 948 return new ZExtInst(NarrowOp, Ty); 949 } 950 951 return nullptr; 952 } 953 954 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) { 955 if (Value *V = SimplifyUDivInst(I.getOperand(0), I.getOperand(1), 956 SQ.getWithInstruction(&I))) 957 return replaceInstUsesWith(I, V); 958 959 if (Instruction *X = foldVectorBinop(I)) 960 return X; 961 962 // Handle the integer div common cases 963 if (Instruction *Common = commonIDivTransforms(I)) 964 return Common; 965 966 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 967 Value *X; 968 const APInt *C1, *C2; 969 if (match(Op0, m_LShr(m_Value(X), m_APInt(C1))) && match(Op1, m_APInt(C2))) { 970 // (X lshr C1) udiv C2 --> X udiv (C2 << C1) 971 bool Overflow; 972 APInt C2ShlC1 = C2->ushl_ov(*C1, Overflow); 973 if (!Overflow) { 974 bool IsExact = I.isExact() && match(Op0, m_Exact(m_Value())); 975 BinaryOperator *BO = BinaryOperator::CreateUDiv( 976 X, ConstantInt::get(X->getType(), C2ShlC1)); 977 if (IsExact) 978 BO->setIsExact(); 979 return BO; 980 } 981 } 982 983 // Op0 / C where C is large (negative) --> zext (Op0 >= C) 984 // TODO: Could use isKnownNegative() to handle non-constant values. 985 Type *Ty = I.getType(); 986 if (match(Op1, m_Negative())) { 987 Value *Cmp = Builder.CreateICmpUGE(Op0, Op1); 988 return CastInst::CreateZExtOrBitCast(Cmp, Ty); 989 } 990 // Op0 / (sext i1 X) --> zext (Op0 == -1) (if X is 0, the div is undefined) 991 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) { 992 Value *Cmp = Builder.CreateICmpEQ(Op0, ConstantInt::getAllOnesValue(Ty)); 993 return CastInst::CreateZExtOrBitCast(Cmp, Ty); 994 } 995 996 if (Instruction *NarrowDiv = narrowUDivURem(I, Builder)) 997 return NarrowDiv; 998 999 // If the udiv operands are non-overflowing multiplies with a common operand, 1000 // then eliminate the common factor: 1001 // (A * B) / (A * X) --> B / X (and commuted variants) 1002 // TODO: The code would be reduced if we had m_c_NUWMul pattern matching. 1003 // TODO: If -reassociation handled this generally, we could remove this. 1004 Value *A, *B; 1005 if (match(Op0, m_NUWMul(m_Value(A), m_Value(B)))) { 1006 if (match(Op1, m_NUWMul(m_Specific(A), m_Value(X))) || 1007 match(Op1, m_NUWMul(m_Value(X), m_Specific(A)))) 1008 return BinaryOperator::CreateUDiv(B, X); 1009 if (match(Op1, m_NUWMul(m_Specific(B), m_Value(X))) || 1010 match(Op1, m_NUWMul(m_Value(X), m_Specific(B)))) 1011 return BinaryOperator::CreateUDiv(A, X); 1012 } 1013 1014 // (LHS udiv (select (select (...)))) -> (LHS >> (select (select (...)))) 1015 SmallVector<UDivFoldAction, 6> UDivActions; 1016 if (visitUDivOperand(Op0, Op1, I, UDivActions)) 1017 for (unsigned i = 0, e = UDivActions.size(); i != e; ++i) { 1018 FoldUDivOperandCb Action = UDivActions[i].FoldAction; 1019 Value *ActionOp1 = UDivActions[i].OperandToFold; 1020 Instruction *Inst; 1021 if (Action) 1022 Inst = Action(Op0, ActionOp1, I, *this); 1023 else { 1024 // This action joins two actions together. The RHS of this action is 1025 // simply the last action we processed, we saved the LHS action index in 1026 // the joining action. 1027 size_t SelectRHSIdx = i - 1; 1028 Value *SelectRHS = UDivActions[SelectRHSIdx].FoldResult; 1029 size_t SelectLHSIdx = UDivActions[i].SelectLHSIdx; 1030 Value *SelectLHS = UDivActions[SelectLHSIdx].FoldResult; 1031 Inst = SelectInst::Create(cast<SelectInst>(ActionOp1)->getCondition(), 1032 SelectLHS, SelectRHS); 1033 } 1034 1035 // If this is the last action to process, return it to the InstCombiner. 1036 // Otherwise, we insert it before the UDiv and record it so that we may 1037 // use it as part of a joining action (i.e., a SelectInst). 1038 if (e - i != 1) { 1039 Inst->insertBefore(&I); 1040 UDivActions[i].FoldResult = Inst; 1041 } else 1042 return Inst; 1043 } 1044 1045 return nullptr; 1046 } 1047 1048 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) { 1049 if (Value *V = SimplifySDivInst(I.getOperand(0), I.getOperand(1), 1050 SQ.getWithInstruction(&I))) 1051 return replaceInstUsesWith(I, V); 1052 1053 if (Instruction *X = foldVectorBinop(I)) 1054 return X; 1055 1056 // Handle the integer div common cases 1057 if (Instruction *Common = commonIDivTransforms(I)) 1058 return Common; 1059 1060 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1061 Value *X; 1062 // sdiv Op0, -1 --> -Op0 1063 // sdiv Op0, (sext i1 X) --> -Op0 (because if X is 0, the op is undefined) 1064 if (match(Op1, m_AllOnes()) || 1065 (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))) 1066 return BinaryOperator::CreateNeg(Op0); 1067 1068 // X / INT_MIN --> X == INT_MIN 1069 if (match(Op1, m_SignMask())) 1070 return new ZExtInst(Builder.CreateICmpEQ(Op0, Op1), I.getType()); 1071 1072 const APInt *Op1C; 1073 if (match(Op1, m_APInt(Op1C))) { 1074 // sdiv exact X, C --> ashr exact X, log2(C) 1075 if (I.isExact() && Op1C->isNonNegative() && Op1C->isPowerOf2()) { 1076 Value *ShAmt = ConstantInt::get(Op1->getType(), Op1C->exactLogBase2()); 1077 return BinaryOperator::CreateExactAShr(Op0, ShAmt, I.getName()); 1078 } 1079 1080 // If the dividend is sign-extended and the constant divisor is small enough 1081 // to fit in the source type, shrink the division to the narrower type: 1082 // (sext X) sdiv C --> sext (X sdiv C) 1083 Value *Op0Src; 1084 if (match(Op0, m_OneUse(m_SExt(m_Value(Op0Src)))) && 1085 Op0Src->getType()->getScalarSizeInBits() >= Op1C->getMinSignedBits()) { 1086 1087 // In the general case, we need to make sure that the dividend is not the 1088 // minimum signed value because dividing that by -1 is UB. But here, we 1089 // know that the -1 divisor case is already handled above. 1090 1091 Constant *NarrowDivisor = 1092 ConstantExpr::getTrunc(cast<Constant>(Op1), Op0Src->getType()); 1093 Value *NarrowOp = Builder.CreateSDiv(Op0Src, NarrowDivisor); 1094 return new SExtInst(NarrowOp, Op0->getType()); 1095 } 1096 1097 // -X / C --> X / -C (if the negation doesn't overflow). 1098 // TODO: This could be enhanced to handle arbitrary vector constants by 1099 // checking if all elements are not the min-signed-val. 1100 if (!Op1C->isMinSignedValue() && 1101 match(Op0, m_NSWSub(m_Zero(), m_Value(X)))) { 1102 Constant *NegC = ConstantInt::get(I.getType(), -(*Op1C)); 1103 Instruction *BO = BinaryOperator::CreateSDiv(X, NegC); 1104 BO->setIsExact(I.isExact()); 1105 return BO; 1106 } 1107 } 1108 1109 // -X / Y --> -(X / Y) 1110 Value *Y; 1111 if (match(&I, m_SDiv(m_OneUse(m_NSWSub(m_Zero(), m_Value(X))), m_Value(Y)))) 1112 return BinaryOperator::CreateNSWNeg( 1113 Builder.CreateSDiv(X, Y, I.getName(), I.isExact())); 1114 1115 // If the sign bits of both operands are zero (i.e. we can prove they are 1116 // unsigned inputs), turn this into a udiv. 1117 APInt Mask(APInt::getSignMask(I.getType()->getScalarSizeInBits())); 1118 if (MaskedValueIsZero(Op0, Mask, 0, &I)) { 1119 if (MaskedValueIsZero(Op1, Mask, 0, &I)) { 1120 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set 1121 auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName()); 1122 BO->setIsExact(I.isExact()); 1123 return BO; 1124 } 1125 1126 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, &I)) { 1127 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y) 1128 // Safe because the only negative value (1 << Y) can take on is 1129 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have 1130 // the sign bit set. 1131 auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName()); 1132 BO->setIsExact(I.isExact()); 1133 return BO; 1134 } 1135 } 1136 1137 return nullptr; 1138 } 1139 1140 /// Remove negation and try to convert division into multiplication. 1141 static Instruction *foldFDivConstantDivisor(BinaryOperator &I) { 1142 Constant *C; 1143 if (!match(I.getOperand(1), m_Constant(C))) 1144 return nullptr; 1145 1146 // -X / C --> X / -C 1147 Value *X; 1148 if (match(I.getOperand(0), m_FNeg(m_Value(X)))) 1149 return BinaryOperator::CreateFDivFMF(X, ConstantExpr::getFNeg(C), &I); 1150 1151 // If the constant divisor has an exact inverse, this is always safe. If not, 1152 // then we can still create a reciprocal if fast-math-flags allow it and the 1153 // constant is a regular number (not zero, infinite, or denormal). 1154 if (!(C->hasExactInverseFP() || (I.hasAllowReciprocal() && C->isNormalFP()))) 1155 return nullptr; 1156 1157 // Disallow denormal constants because we don't know what would happen 1158 // on all targets. 1159 // TODO: Use Intrinsic::canonicalize or let function attributes tell us that 1160 // denorms are flushed? 1161 auto *RecipC = ConstantExpr::getFDiv(ConstantFP::get(I.getType(), 1.0), C); 1162 if (!RecipC->isNormalFP()) 1163 return nullptr; 1164 1165 // X / C --> X * (1 / C) 1166 return BinaryOperator::CreateFMulFMF(I.getOperand(0), RecipC, &I); 1167 } 1168 1169 /// Remove negation and try to reassociate constant math. 1170 static Instruction *foldFDivConstantDividend(BinaryOperator &I) { 1171 Constant *C; 1172 if (!match(I.getOperand(0), m_Constant(C))) 1173 return nullptr; 1174 1175 // C / -X --> -C / X 1176 Value *X; 1177 if (match(I.getOperand(1), m_FNeg(m_Value(X)))) 1178 return BinaryOperator::CreateFDivFMF(ConstantExpr::getFNeg(C), X, &I); 1179 1180 if (!I.hasAllowReassoc() || !I.hasAllowReciprocal()) 1181 return nullptr; 1182 1183 // Try to reassociate C / X expressions where X includes another constant. 1184 Constant *C2, *NewC = nullptr; 1185 if (match(I.getOperand(1), m_FMul(m_Value(X), m_Constant(C2)))) { 1186 // C / (X * C2) --> (C / C2) / X 1187 NewC = ConstantExpr::getFDiv(C, C2); 1188 } else if (match(I.getOperand(1), m_FDiv(m_Value(X), m_Constant(C2)))) { 1189 // C / (X / C2) --> (C * C2) / X 1190 NewC = ConstantExpr::getFMul(C, C2); 1191 } 1192 // Disallow denormal constants because we don't know what would happen 1193 // on all targets. 1194 // TODO: Use Intrinsic::canonicalize or let function attributes tell us that 1195 // denorms are flushed? 1196 if (!NewC || !NewC->isNormalFP()) 1197 return nullptr; 1198 1199 return BinaryOperator::CreateFDivFMF(NewC, X, &I); 1200 } 1201 1202 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) { 1203 if (Value *V = SimplifyFDivInst(I.getOperand(0), I.getOperand(1), 1204 I.getFastMathFlags(), 1205 SQ.getWithInstruction(&I))) 1206 return replaceInstUsesWith(I, V); 1207 1208 if (Instruction *X = foldVectorBinop(I)) 1209 return X; 1210 1211 if (Instruction *R = foldFDivConstantDivisor(I)) 1212 return R; 1213 1214 if (Instruction *R = foldFDivConstantDividend(I)) 1215 return R; 1216 1217 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1218 if (isa<Constant>(Op0)) 1219 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 1220 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1221 return R; 1222 1223 if (isa<Constant>(Op1)) 1224 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 1225 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1226 return R; 1227 1228 if (I.hasAllowReassoc() && I.hasAllowReciprocal()) { 1229 Value *X, *Y; 1230 if (match(Op0, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))) && 1231 (!isa<Constant>(Y) || !isa<Constant>(Op1))) { 1232 // (X / Y) / Z => X / (Y * Z) 1233 Value *YZ = Builder.CreateFMulFMF(Y, Op1, &I); 1234 return BinaryOperator::CreateFDivFMF(X, YZ, &I); 1235 } 1236 if (match(Op1, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))) && 1237 (!isa<Constant>(Y) || !isa<Constant>(Op0))) { 1238 // Z / (X / Y) => (Y * Z) / X 1239 Value *YZ = Builder.CreateFMulFMF(Y, Op0, &I); 1240 return BinaryOperator::CreateFDivFMF(YZ, X, &I); 1241 } 1242 // Z / (1.0 / Y) => (Y * Z) 1243 // 1244 // This is a special case of Z / (X / Y) => (Y * Z) / X, with X = 1.0. The 1245 // m_OneUse check is avoided because even in the case of the multiple uses 1246 // for 1.0/Y, the number of instructions remain the same and a division is 1247 // replaced by a multiplication. 1248 if (match(Op1, m_FDiv(m_SpecificFP(1.0), m_Value(Y)))) 1249 return BinaryOperator::CreateFMulFMF(Y, Op0, &I); 1250 } 1251 1252 if (I.hasAllowReassoc() && Op0->hasOneUse() && Op1->hasOneUse()) { 1253 // sin(X) / cos(X) -> tan(X) 1254 // cos(X) / sin(X) -> 1/tan(X) (cotangent) 1255 Value *X; 1256 bool IsTan = match(Op0, m_Intrinsic<Intrinsic::sin>(m_Value(X))) && 1257 match(Op1, m_Intrinsic<Intrinsic::cos>(m_Specific(X))); 1258 bool IsCot = 1259 !IsTan && match(Op0, m_Intrinsic<Intrinsic::cos>(m_Value(X))) && 1260 match(Op1, m_Intrinsic<Intrinsic::sin>(m_Specific(X))); 1261 1262 if ((IsTan || IsCot) && 1263 hasFloatFn(&TLI, I.getType(), LibFunc_tan, LibFunc_tanf, LibFunc_tanl)) { 1264 IRBuilder<> B(&I); 1265 IRBuilder<>::FastMathFlagGuard FMFGuard(B); 1266 B.setFastMathFlags(I.getFastMathFlags()); 1267 AttributeList Attrs = 1268 cast<CallBase>(Op0)->getCalledFunction()->getAttributes(); 1269 Value *Res = emitUnaryFloatFnCall(X, &TLI, LibFunc_tan, LibFunc_tanf, 1270 LibFunc_tanl, B, Attrs); 1271 if (IsCot) 1272 Res = B.CreateFDiv(ConstantFP::get(I.getType(), 1.0), Res); 1273 return replaceInstUsesWith(I, Res); 1274 } 1275 } 1276 1277 // -X / -Y -> X / Y 1278 Value *X, *Y; 1279 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y)))) { 1280 I.setOperand(0, X); 1281 I.setOperand(1, Y); 1282 return &I; 1283 } 1284 1285 // X / (X * Y) --> 1.0 / Y 1286 // Reassociate to (X / X -> 1.0) is legal when NaNs are not allowed. 1287 // We can ignore the possibility that X is infinity because INF/INF is NaN. 1288 if (I.hasNoNaNs() && I.hasAllowReassoc() && 1289 match(Op1, m_c_FMul(m_Specific(Op0), m_Value(Y)))) { 1290 I.setOperand(0, ConstantFP::get(I.getType(), 1.0)); 1291 I.setOperand(1, Y); 1292 return &I; 1293 } 1294 1295 // X / fabs(X) -> copysign(1.0, X) 1296 // fabs(X) / X -> copysign(1.0, X) 1297 if (I.hasNoNaNs() && I.hasNoInfs() && 1298 (match(&I, 1299 m_FDiv(m_Value(X), m_Intrinsic<Intrinsic::fabs>(m_Deferred(X)))) || 1300 match(&I, m_FDiv(m_Intrinsic<Intrinsic::fabs>(m_Value(X)), 1301 m_Deferred(X))))) { 1302 Value *V = Builder.CreateBinaryIntrinsic( 1303 Intrinsic::copysign, ConstantFP::get(I.getType(), 1.0), X, &I); 1304 return replaceInstUsesWith(I, V); 1305 } 1306 return nullptr; 1307 } 1308 1309 /// This function implements the transforms common to both integer remainder 1310 /// instructions (urem and srem). It is called by the visitors to those integer 1311 /// remainder instructions. 1312 /// Common integer remainder transforms 1313 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) { 1314 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1315 1316 // The RHS is known non-zero. 1317 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I)) { 1318 I.setOperand(1, V); 1319 return &I; 1320 } 1321 1322 // Handle cases involving: rem X, (select Cond, Y, Z) 1323 if (simplifyDivRemOfSelectWithZeroOp(I)) 1324 return &I; 1325 1326 if (isa<Constant>(Op1)) { 1327 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) { 1328 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) { 1329 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1330 return R; 1331 } else if (auto *PN = dyn_cast<PHINode>(Op0I)) { 1332 const APInt *Op1Int; 1333 if (match(Op1, m_APInt(Op1Int)) && !Op1Int->isMinValue() && 1334 (I.getOpcode() == Instruction::URem || 1335 !Op1Int->isMinSignedValue())) { 1336 // foldOpIntoPhi will speculate instructions to the end of the PHI's 1337 // predecessor blocks, so do this only if we know the srem or urem 1338 // will not fault. 1339 if (Instruction *NV = foldOpIntoPhi(I, PN)) 1340 return NV; 1341 } 1342 } 1343 1344 // See if we can fold away this rem instruction. 1345 if (SimplifyDemandedInstructionBits(I)) 1346 return &I; 1347 } 1348 } 1349 1350 return nullptr; 1351 } 1352 1353 Instruction *InstCombiner::visitURem(BinaryOperator &I) { 1354 if (Value *V = SimplifyURemInst(I.getOperand(0), I.getOperand(1), 1355 SQ.getWithInstruction(&I))) 1356 return replaceInstUsesWith(I, V); 1357 1358 if (Instruction *X = foldVectorBinop(I)) 1359 return X; 1360 1361 if (Instruction *common = commonIRemTransforms(I)) 1362 return common; 1363 1364 if (Instruction *NarrowRem = narrowUDivURem(I, Builder)) 1365 return NarrowRem; 1366 1367 // X urem Y -> X and Y-1, where Y is a power of 2, 1368 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1369 Type *Ty = I.getType(); 1370 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, &I)) { 1371 // This may increase instruction count, we don't enforce that Y is a 1372 // constant. 1373 Constant *N1 = Constant::getAllOnesValue(Ty); 1374 Value *Add = Builder.CreateAdd(Op1, N1); 1375 return BinaryOperator::CreateAnd(Op0, Add); 1376 } 1377 1378 // 1 urem X -> zext(X != 1) 1379 if (match(Op0, m_One())) { 1380 Value *Cmp = Builder.CreateICmpNE(Op1, ConstantInt::get(Ty, 1)); 1381 return CastInst::CreateZExtOrBitCast(Cmp, Ty); 1382 } 1383 1384 // X urem C -> X < C ? X : X - C, where C >= signbit. 1385 if (match(Op1, m_Negative())) { 1386 Value *Cmp = Builder.CreateICmpULT(Op0, Op1); 1387 Value *Sub = Builder.CreateSub(Op0, Op1); 1388 return SelectInst::Create(Cmp, Op0, Sub); 1389 } 1390 1391 // If the divisor is a sext of a boolean, then the divisor must be max 1392 // unsigned value (-1). Therefore, the remainder is Op0 unless Op0 is also 1393 // max unsigned value. In that case, the remainder is 0: 1394 // urem Op0, (sext i1 X) --> (Op0 == -1) ? 0 : Op0 1395 Value *X; 1396 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) { 1397 Value *Cmp = Builder.CreateICmpEQ(Op0, ConstantInt::getAllOnesValue(Ty)); 1398 return SelectInst::Create(Cmp, ConstantInt::getNullValue(Ty), Op0); 1399 } 1400 1401 return nullptr; 1402 } 1403 1404 Instruction *InstCombiner::visitSRem(BinaryOperator &I) { 1405 if (Value *V = SimplifySRemInst(I.getOperand(0), I.getOperand(1), 1406 SQ.getWithInstruction(&I))) 1407 return replaceInstUsesWith(I, V); 1408 1409 if (Instruction *X = foldVectorBinop(I)) 1410 return X; 1411 1412 // Handle the integer rem common cases 1413 if (Instruction *Common = commonIRemTransforms(I)) 1414 return Common; 1415 1416 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1417 { 1418 const APInt *Y; 1419 // X % -Y -> X % Y 1420 if (match(Op1, m_Negative(Y)) && !Y->isMinSignedValue()) { 1421 Worklist.AddValue(I.getOperand(1)); 1422 I.setOperand(1, ConstantInt::get(I.getType(), -*Y)); 1423 return &I; 1424 } 1425 } 1426 1427 // -X srem Y --> -(X srem Y) 1428 Value *X, *Y; 1429 if (match(&I, m_SRem(m_OneUse(m_NSWSub(m_Zero(), m_Value(X))), m_Value(Y)))) 1430 return BinaryOperator::CreateNSWNeg(Builder.CreateSRem(X, Y)); 1431 1432 // If the sign bits of both operands are zero (i.e. we can prove they are 1433 // unsigned inputs), turn this into a urem. 1434 APInt Mask(APInt::getSignMask(I.getType()->getScalarSizeInBits())); 1435 if (MaskedValueIsZero(Op1, Mask, 0, &I) && 1436 MaskedValueIsZero(Op0, Mask, 0, &I)) { 1437 // X srem Y -> X urem Y, iff X and Y don't have sign bit set 1438 return BinaryOperator::CreateURem(Op0, Op1, I.getName()); 1439 } 1440 1441 // If it's a constant vector, flip any negative values positive. 1442 if (isa<ConstantVector>(Op1) || isa<ConstantDataVector>(Op1)) { 1443 Constant *C = cast<Constant>(Op1); 1444 unsigned VWidth = C->getType()->getVectorNumElements(); 1445 1446 bool hasNegative = false; 1447 bool hasMissing = false; 1448 for (unsigned i = 0; i != VWidth; ++i) { 1449 Constant *Elt = C->getAggregateElement(i); 1450 if (!Elt) { 1451 hasMissing = true; 1452 break; 1453 } 1454 1455 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elt)) 1456 if (RHS->isNegative()) 1457 hasNegative = true; 1458 } 1459 1460 if (hasNegative && !hasMissing) { 1461 SmallVector<Constant *, 16> Elts(VWidth); 1462 for (unsigned i = 0; i != VWidth; ++i) { 1463 Elts[i] = C->getAggregateElement(i); // Handle undef, etc. 1464 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elts[i])) { 1465 if (RHS->isNegative()) 1466 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS)); 1467 } 1468 } 1469 1470 Constant *NewRHSV = ConstantVector::get(Elts); 1471 if (NewRHSV != C) { // Don't loop on -MININT 1472 Worklist.AddValue(I.getOperand(1)); 1473 I.setOperand(1, NewRHSV); 1474 return &I; 1475 } 1476 } 1477 } 1478 1479 return nullptr; 1480 } 1481 1482 Instruction *InstCombiner::visitFRem(BinaryOperator &I) { 1483 if (Value *V = SimplifyFRemInst(I.getOperand(0), I.getOperand(1), 1484 I.getFastMathFlags(), 1485 SQ.getWithInstruction(&I))) 1486 return replaceInstUsesWith(I, V); 1487 1488 if (Instruction *X = foldVectorBinop(I)) 1489 return X; 1490 1491 return nullptr; 1492 } 1493