1 //===- InstCombineMulDivRem.cpp -------------------------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the visit functions for mul, fmul, sdiv, udiv, fdiv, 11 // srem, urem, frem. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "InstCombine.h" 16 #include "llvm/Analysis/InstructionSimplify.h" 17 #include "llvm/IR/IntrinsicInst.h" 18 #include "llvm/Support/PatternMatch.h" 19 using namespace llvm; 20 using namespace PatternMatch; 21 22 23 /// simplifyValueKnownNonZero - The specific integer value is used in a context 24 /// where it is known to be non-zero. If this allows us to simplify the 25 /// computation, do so and return the new operand, otherwise return null. 26 static Value *simplifyValueKnownNonZero(Value *V, InstCombiner &IC) { 27 // If V has multiple uses, then we would have to do more analysis to determine 28 // if this is safe. For example, the use could be in dynamically unreached 29 // code. 30 if (!V->hasOneUse()) return 0; 31 32 bool MadeChange = false; 33 34 // ((1 << A) >>u B) --> (1 << (A-B)) 35 // Because V cannot be zero, we know that B is less than A. 36 Value *A = 0, *B = 0, *PowerOf2 = 0; 37 if (match(V, m_LShr(m_OneUse(m_Shl(m_Value(PowerOf2), m_Value(A))), 38 m_Value(B))) && 39 // The "1" can be any value known to be a power of 2. 40 isKnownToBeAPowerOfTwo(PowerOf2)) { 41 A = IC.Builder->CreateSub(A, B); 42 return IC.Builder->CreateShl(PowerOf2, A); 43 } 44 45 // (PowerOfTwo >>u B) --> isExact since shifting out the result would make it 46 // inexact. Similarly for <<. 47 if (BinaryOperator *I = dyn_cast<BinaryOperator>(V)) 48 if (I->isLogicalShift() && isKnownToBeAPowerOfTwo(I->getOperand(0))) { 49 // We know that this is an exact/nuw shift and that the input is a 50 // non-zero context as well. 51 if (Value *V2 = simplifyValueKnownNonZero(I->getOperand(0), IC)) { 52 I->setOperand(0, V2); 53 MadeChange = true; 54 } 55 56 if (I->getOpcode() == Instruction::LShr && !I->isExact()) { 57 I->setIsExact(); 58 MadeChange = true; 59 } 60 61 if (I->getOpcode() == Instruction::Shl && !I->hasNoUnsignedWrap()) { 62 I->setHasNoUnsignedWrap(); 63 MadeChange = true; 64 } 65 } 66 67 // TODO: Lots more we could do here: 68 // If V is a phi node, we can call this on each of its operands. 69 // "select cond, X, 0" can simplify to "X". 70 71 return MadeChange ? V : 0; 72 } 73 74 75 /// MultiplyOverflows - True if the multiply can not be expressed in an int 76 /// this size. 77 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) { 78 uint32_t W = C1->getBitWidth(); 79 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue(); 80 if (sign) { 81 LHSExt = LHSExt.sext(W * 2); 82 RHSExt = RHSExt.sext(W * 2); 83 } else { 84 LHSExt = LHSExt.zext(W * 2); 85 RHSExt = RHSExt.zext(W * 2); 86 } 87 88 APInt MulExt = LHSExt * RHSExt; 89 90 if (!sign) 91 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W)); 92 93 APInt Min = APInt::getSignedMinValue(W).sext(W * 2); 94 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2); 95 return MulExt.slt(Min) || MulExt.sgt(Max); 96 } 97 98 Instruction *InstCombiner::visitMul(BinaryOperator &I) { 99 bool Changed = SimplifyAssociativeOrCommutative(I); 100 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 101 102 if (Value *V = SimplifyMulInst(Op0, Op1, TD)) 103 return ReplaceInstUsesWith(I, V); 104 105 if (Value *V = SimplifyUsingDistributiveLaws(I)) 106 return ReplaceInstUsesWith(I, V); 107 108 if (match(Op1, m_AllOnes())) // X * -1 == 0 - X 109 return BinaryOperator::CreateNeg(Op0, I.getName()); 110 111 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 112 113 // ((X << C1)*C2) == (X * (C2 << C1)) 114 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0)) 115 if (SI->getOpcode() == Instruction::Shl) 116 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1))) 117 return BinaryOperator::CreateMul(SI->getOperand(0), 118 ConstantExpr::getShl(CI, ShOp)); 119 120 const APInt &Val = CI->getValue(); 121 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C 122 Constant *NewCst = ConstantInt::get(Op0->getType(), Val.logBase2()); 123 BinaryOperator *Shl = BinaryOperator::CreateShl(Op0, NewCst); 124 if (I.hasNoSignedWrap()) Shl->setHasNoSignedWrap(); 125 if (I.hasNoUnsignedWrap()) Shl->setHasNoUnsignedWrap(); 126 return Shl; 127 } 128 129 // Canonicalize (X+C1)*CI -> X*CI+C1*CI. 130 { Value *X; ConstantInt *C1; 131 if (Op0->hasOneUse() && 132 match(Op0, m_Add(m_Value(X), m_ConstantInt(C1)))) { 133 Value *Add = Builder->CreateMul(X, CI); 134 return BinaryOperator::CreateAdd(Add, Builder->CreateMul(C1, CI)); 135 } 136 } 137 138 // (Y - X) * (-(2**n)) -> (X - Y) * (2**n), for positive nonzero n 139 // (Y + const) * (-(2**n)) -> (-constY) * (2**n), for positive nonzero n 140 // The "* (2**n)" thus becomes a potential shifting opportunity. 141 { 142 const APInt & Val = CI->getValue(); 143 const APInt &PosVal = Val.abs(); 144 if (Val.isNegative() && PosVal.isPowerOf2()) { 145 Value *X = 0, *Y = 0; 146 if (Op0->hasOneUse()) { 147 ConstantInt *C1; 148 Value *Sub = 0; 149 if (match(Op0, m_Sub(m_Value(Y), m_Value(X)))) 150 Sub = Builder->CreateSub(X, Y, "suba"); 151 else if (match(Op0, m_Add(m_Value(Y), m_ConstantInt(C1)))) 152 Sub = Builder->CreateSub(Builder->CreateNeg(C1), Y, "subc"); 153 if (Sub) 154 return 155 BinaryOperator::CreateMul(Sub, 156 ConstantInt::get(Y->getType(), PosVal)); 157 } 158 } 159 } 160 } 161 162 // Simplify mul instructions with a constant RHS. 163 if (isa<Constant>(Op1)) { 164 // Try to fold constant mul into select arguments. 165 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 166 if (Instruction *R = FoldOpIntoSelect(I, SI)) 167 return R; 168 169 if (isa<PHINode>(Op0)) 170 if (Instruction *NV = FoldOpIntoPhi(I)) 171 return NV; 172 } 173 174 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y 175 if (Value *Op1v = dyn_castNegVal(Op1)) 176 return BinaryOperator::CreateMul(Op0v, Op1v); 177 178 // (X / Y) * Y = X - (X % Y) 179 // (X / Y) * -Y = (X % Y) - X 180 { 181 Value *Op1C = Op1; 182 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0); 183 if (!BO || 184 (BO->getOpcode() != Instruction::UDiv && 185 BO->getOpcode() != Instruction::SDiv)) { 186 Op1C = Op0; 187 BO = dyn_cast<BinaryOperator>(Op1); 188 } 189 Value *Neg = dyn_castNegVal(Op1C); 190 if (BO && BO->hasOneUse() && 191 (BO->getOperand(1) == Op1C || BO->getOperand(1) == Neg) && 192 (BO->getOpcode() == Instruction::UDiv || 193 BO->getOpcode() == Instruction::SDiv)) { 194 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1); 195 196 // If the division is exact, X % Y is zero, so we end up with X or -X. 197 if (PossiblyExactOperator *SDiv = dyn_cast<PossiblyExactOperator>(BO)) 198 if (SDiv->isExact()) { 199 if (Op1BO == Op1C) 200 return ReplaceInstUsesWith(I, Op0BO); 201 return BinaryOperator::CreateNeg(Op0BO); 202 } 203 204 Value *Rem; 205 if (BO->getOpcode() == Instruction::UDiv) 206 Rem = Builder->CreateURem(Op0BO, Op1BO); 207 else 208 Rem = Builder->CreateSRem(Op0BO, Op1BO); 209 Rem->takeName(BO); 210 211 if (Op1BO == Op1C) 212 return BinaryOperator::CreateSub(Op0BO, Rem); 213 return BinaryOperator::CreateSub(Rem, Op0BO); 214 } 215 } 216 217 /// i1 mul -> i1 and. 218 if (I.getType()->isIntegerTy(1)) 219 return BinaryOperator::CreateAnd(Op0, Op1); 220 221 // X*(1 << Y) --> X << Y 222 // (1 << Y)*X --> X << Y 223 { 224 Value *Y; 225 if (match(Op0, m_Shl(m_One(), m_Value(Y)))) 226 return BinaryOperator::CreateShl(Op1, Y); 227 if (match(Op1, m_Shl(m_One(), m_Value(Y)))) 228 return BinaryOperator::CreateShl(Op0, Y); 229 } 230 231 // If one of the operands of the multiply is a cast from a boolean value, then 232 // we know the bool is either zero or one, so this is a 'masking' multiply. 233 // X * Y (where Y is 0 or 1) -> X & (0-Y) 234 if (!I.getType()->isVectorTy()) { 235 // -2 is "-1 << 1" so it is all bits set except the low one. 236 APInt Negative2(I.getType()->getPrimitiveSizeInBits(), (uint64_t)-2, true); 237 238 Value *BoolCast = 0, *OtherOp = 0; 239 if (MaskedValueIsZero(Op0, Negative2)) 240 BoolCast = Op0, OtherOp = Op1; 241 else if (MaskedValueIsZero(Op1, Negative2)) 242 BoolCast = Op1, OtherOp = Op0; 243 244 if (BoolCast) { 245 Value *V = Builder->CreateSub(Constant::getNullValue(I.getType()), 246 BoolCast); 247 return BinaryOperator::CreateAnd(V, OtherOp); 248 } 249 } 250 251 return Changed ? &I : 0; 252 } 253 254 // 255 // Detect pattern: 256 // 257 // log2(Y*0.5) 258 // 259 // And check for corresponding fast math flags 260 // 261 262 static void detectLog2OfHalf(Value *&Op, Value *&Y, IntrinsicInst *&Log2) { 263 264 if (!Op->hasOneUse()) 265 return; 266 267 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op); 268 if (!II) 269 return; 270 if (II->getIntrinsicID() != Intrinsic::log2 || !II->hasUnsafeAlgebra()) 271 return; 272 Log2 = II; 273 274 Value *OpLog2Of = II->getArgOperand(0); 275 if (!OpLog2Of->hasOneUse()) 276 return; 277 278 Instruction *I = dyn_cast<Instruction>(OpLog2Of); 279 if (!I) 280 return; 281 if (I->getOpcode() != Instruction::FMul || !I->hasUnsafeAlgebra()) 282 return; 283 284 ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(0)); 285 if (CFP && CFP->isExactlyValue(0.5)) { 286 Y = I->getOperand(1); 287 return; 288 } 289 CFP = dyn_cast<ConstantFP>(I->getOperand(1)); 290 if (CFP && CFP->isExactlyValue(0.5)) 291 Y = I->getOperand(0); 292 } 293 294 /// Helper function of InstCombiner::visitFMul(BinaryOperator(). It returns 295 /// true iff the given value is FMul or FDiv with one and only one operand 296 /// being a normal constant (i.e. not Zero/NaN/Infinity). 297 static bool isFMulOrFDivWithConstant(Value *V) { 298 Instruction *I = dyn_cast<Instruction>(V); 299 if (!I || (I->getOpcode() != Instruction::FMul && 300 I->getOpcode() != Instruction::FDiv)) 301 return false; 302 303 ConstantFP *C0 = dyn_cast<ConstantFP>(I->getOperand(0)); 304 ConstantFP *C1 = dyn_cast<ConstantFP>(I->getOperand(1)); 305 306 if (C0 && C1) 307 return false; 308 309 return (C0 && C0->getValueAPF().isNormal()) || 310 (C1 && C1->getValueAPF().isNormal()); 311 } 312 313 static bool isNormalFp(const ConstantFP *C) { 314 const APFloat &Flt = C->getValueAPF(); 315 return Flt.isNormal() && !Flt.isDenormal(); 316 } 317 318 /// foldFMulConst() is a helper routine of InstCombiner::visitFMul(). 319 /// The input \p FMulOrDiv is a FMul/FDiv with one and only one operand 320 /// being a constant (i.e. isFMulOrFDivWithConstant(FMulOrDiv) == true). 321 /// This function is to simplify "FMulOrDiv * C" and returns the 322 /// resulting expression. Note that this function could return NULL in 323 /// case the constants cannot be folded into a normal floating-point. 324 /// 325 Value *InstCombiner::foldFMulConst(Instruction *FMulOrDiv, ConstantFP *C, 326 Instruction *InsertBefore) { 327 assert(isFMulOrFDivWithConstant(FMulOrDiv) && "V is invalid"); 328 329 Value *Opnd0 = FMulOrDiv->getOperand(0); 330 Value *Opnd1 = FMulOrDiv->getOperand(1); 331 332 ConstantFP *C0 = dyn_cast<ConstantFP>(Opnd0); 333 ConstantFP *C1 = dyn_cast<ConstantFP>(Opnd1); 334 335 BinaryOperator *R = 0; 336 337 // (X * C0) * C => X * (C0*C) 338 if (FMulOrDiv->getOpcode() == Instruction::FMul) { 339 Constant *F = ConstantExpr::getFMul(C1 ? C1 : C0, C); 340 if (isNormalFp(cast<ConstantFP>(F))) 341 R = BinaryOperator::CreateFMul(C1 ? Opnd0 : Opnd1, F); 342 } else { 343 if (C0) { 344 // (C0 / X) * C => (C0 * C) / X 345 ConstantFP *F = cast<ConstantFP>(ConstantExpr::getFMul(C0, C)); 346 if (isNormalFp(F)) 347 R = BinaryOperator::CreateFDiv(F, Opnd1); 348 } else { 349 // (X / C1) * C => X * (C/C1) if C/C1 is not a denormal 350 ConstantFP *F = cast<ConstantFP>(ConstantExpr::getFDiv(C, C1)); 351 if (isNormalFp(F)) { 352 R = BinaryOperator::CreateFMul(Opnd0, F); 353 } else { 354 // (X / C1) * C => X / (C1/C) 355 Constant *F = ConstantExpr::getFDiv(C1, C); 356 if (isNormalFp(cast<ConstantFP>(F))) 357 R = BinaryOperator::CreateFDiv(Opnd0, F); 358 } 359 } 360 } 361 362 if (R) { 363 R->setHasUnsafeAlgebra(true); 364 InsertNewInstWith(R, *InsertBefore); 365 } 366 367 return R; 368 } 369 370 Instruction *InstCombiner::visitFMul(BinaryOperator &I) { 371 bool Changed = SimplifyAssociativeOrCommutative(I); 372 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 373 374 if (isa<Constant>(Op0)) 375 std::swap(Op0, Op1); 376 377 if (Value *V = SimplifyFMulInst(Op0, Op1, I.getFastMathFlags(), TD)) 378 return ReplaceInstUsesWith(I, V); 379 380 bool AllowReassociate = I.hasUnsafeAlgebra(); 381 382 // Simplify mul instructions with a constant RHS. 383 if (isa<Constant>(Op1)) { 384 // Try to fold constant mul into select arguments. 385 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 386 if (Instruction *R = FoldOpIntoSelect(I, SI)) 387 return R; 388 389 if (isa<PHINode>(Op0)) 390 if (Instruction *NV = FoldOpIntoPhi(I)) 391 return NV; 392 393 ConstantFP *C = dyn_cast<ConstantFP>(Op1); 394 if (C && AllowReassociate && C->getValueAPF().isNormal()) { 395 // Let MDC denote an expression in one of these forms: 396 // X * C, C/X, X/C, where C is a constant. 397 // 398 // Try to simplify "MDC * Constant" 399 if (isFMulOrFDivWithConstant(Op0)) { 400 Value *V = foldFMulConst(cast<Instruction>(Op0), C, &I); 401 if (V) 402 return ReplaceInstUsesWith(I, V); 403 } 404 405 // (MDC +/- C1) * C => (MDC * C) +/- (C1 * C) 406 Instruction *FAddSub = dyn_cast<Instruction>(Op0); 407 if (FAddSub && 408 (FAddSub->getOpcode() == Instruction::FAdd || 409 FAddSub->getOpcode() == Instruction::FSub)) { 410 Value *Opnd0 = FAddSub->getOperand(0); 411 Value *Opnd1 = FAddSub->getOperand(1); 412 ConstantFP *C0 = dyn_cast<ConstantFP>(Opnd0); 413 ConstantFP *C1 = dyn_cast<ConstantFP>(Opnd1); 414 bool Swap = false; 415 if (C0) { 416 std::swap(C0, C1); 417 std::swap(Opnd0, Opnd1); 418 Swap = true; 419 } 420 421 if (C1 && C1->getValueAPF().isNormal() && 422 isFMulOrFDivWithConstant(Opnd0)) { 423 Value *M1 = ConstantExpr::getFMul(C1, C); 424 Value *M0 = isNormalFp(cast<ConstantFP>(M1)) ? 425 foldFMulConst(cast<Instruction>(Opnd0), C, &I) : 426 0; 427 if (M0 && M1) { 428 if (Swap && FAddSub->getOpcode() == Instruction::FSub) 429 std::swap(M0, M1); 430 431 Value *R = (FAddSub->getOpcode() == Instruction::FAdd) ? 432 BinaryOperator::CreateFAdd(M0, M1) : 433 BinaryOperator::CreateFSub(M0, M1); 434 Instruction *RI = cast<Instruction>(R); 435 RI->copyFastMathFlags(&I); 436 return RI; 437 } 438 } 439 } 440 } 441 } 442 443 444 // Under unsafe algebra do: 445 // X * log2(0.5*Y) = X*log2(Y) - X 446 if (I.hasUnsafeAlgebra()) { 447 Value *OpX = NULL; 448 Value *OpY = NULL; 449 IntrinsicInst *Log2; 450 detectLog2OfHalf(Op0, OpY, Log2); 451 if (OpY) { 452 OpX = Op1; 453 } else { 454 detectLog2OfHalf(Op1, OpY, Log2); 455 if (OpY) { 456 OpX = Op0; 457 } 458 } 459 // if pattern detected emit alternate sequence 460 if (OpX && OpY) { 461 Log2->setArgOperand(0, OpY); 462 Value *FMulVal = Builder->CreateFMul(OpX, Log2); 463 Instruction *FMul = cast<Instruction>(FMulVal); 464 FMul->copyFastMathFlags(Log2); 465 Instruction *FSub = BinaryOperator::CreateFSub(FMulVal, OpX); 466 FSub->copyFastMathFlags(Log2); 467 return FSub; 468 } 469 } 470 471 // Handle symmetric situation in a 2-iteration loop 472 Value *Opnd0 = Op0; 473 Value *Opnd1 = Op1; 474 for (int i = 0; i < 2; i++) { 475 bool IgnoreZeroSign = I.hasNoSignedZeros(); 476 if (BinaryOperator::isFNeg(Opnd0, IgnoreZeroSign)) { 477 Value *N0 = dyn_castFNegVal(Opnd0, IgnoreZeroSign); 478 Value *N1 = dyn_castFNegVal(Opnd1, IgnoreZeroSign); 479 480 // -X * -Y => X*Y 481 if (N1) 482 return BinaryOperator::CreateFMul(N0, N1); 483 484 if (Opnd0->hasOneUse()) { 485 // -X * Y => -(X*Y) (Promote negation as high as possible) 486 Value *T = Builder->CreateFMul(N0, Opnd1); 487 cast<Instruction>(T)->setDebugLoc(I.getDebugLoc()); 488 Instruction *Neg = BinaryOperator::CreateFNeg(T); 489 if (I.getFastMathFlags().any()) { 490 cast<Instruction>(T)->copyFastMathFlags(&I); 491 Neg->copyFastMathFlags(&I); 492 } 493 return Neg; 494 } 495 } 496 497 // (X*Y) * X => (X*X) * Y where Y != X 498 // The purpose is two-fold: 499 // 1) to form a power expression (of X). 500 // 2) potentially shorten the critical path: After transformation, the 501 // latency of the instruction Y is amortized by the expression of X*X, 502 // and therefore Y is in a "less critical" position compared to what it 503 // was before the transformation. 504 // 505 if (AllowReassociate) { 506 Value *Opnd0_0, *Opnd0_1; 507 if (Opnd0->hasOneUse() && 508 match(Opnd0, m_FMul(m_Value(Opnd0_0), m_Value(Opnd0_1)))) { 509 Value *Y = 0; 510 if (Opnd0_0 == Opnd1 && Opnd0_1 != Opnd1) 511 Y = Opnd0_1; 512 else if (Opnd0_1 == Opnd1 && Opnd0_0 != Opnd1) 513 Y = Opnd0_0; 514 515 if (Y) { 516 Instruction *T = cast<Instruction>(Builder->CreateFMul(Opnd1, Opnd1)); 517 T->copyFastMathFlags(&I); 518 T->setDebugLoc(I.getDebugLoc()); 519 520 Instruction *R = BinaryOperator::CreateFMul(T, Y); 521 R->copyFastMathFlags(&I); 522 return R; 523 } 524 } 525 } 526 527 if (!isa<Constant>(Op1)) 528 std::swap(Opnd0, Opnd1); 529 else 530 break; 531 } 532 533 return Changed ? &I : 0; 534 } 535 536 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select 537 /// instruction. 538 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) { 539 SelectInst *SI = cast<SelectInst>(I.getOperand(1)); 540 541 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y 542 int NonNullOperand = -1; 543 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1))) 544 if (ST->isNullValue()) 545 NonNullOperand = 2; 546 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y 547 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2))) 548 if (ST->isNullValue()) 549 NonNullOperand = 1; 550 551 if (NonNullOperand == -1) 552 return false; 553 554 Value *SelectCond = SI->getOperand(0); 555 556 // Change the div/rem to use 'Y' instead of the select. 557 I.setOperand(1, SI->getOperand(NonNullOperand)); 558 559 // Okay, we know we replace the operand of the div/rem with 'Y' with no 560 // problem. However, the select, or the condition of the select may have 561 // multiple uses. Based on our knowledge that the operand must be non-zero, 562 // propagate the known value for the select into other uses of it, and 563 // propagate a known value of the condition into its other users. 564 565 // If the select and condition only have a single use, don't bother with this, 566 // early exit. 567 if (SI->use_empty() && SelectCond->hasOneUse()) 568 return true; 569 570 // Scan the current block backward, looking for other uses of SI. 571 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin(); 572 573 while (BBI != BBFront) { 574 --BBI; 575 // If we found a call to a function, we can't assume it will return, so 576 // information from below it cannot be propagated above it. 577 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI)) 578 break; 579 580 // Replace uses of the select or its condition with the known values. 581 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end(); 582 I != E; ++I) { 583 if (*I == SI) { 584 *I = SI->getOperand(NonNullOperand); 585 Worklist.Add(BBI); 586 } else if (*I == SelectCond) { 587 *I = NonNullOperand == 1 ? ConstantInt::getTrue(BBI->getContext()) : 588 ConstantInt::getFalse(BBI->getContext()); 589 Worklist.Add(BBI); 590 } 591 } 592 593 // If we past the instruction, quit looking for it. 594 if (&*BBI == SI) 595 SI = 0; 596 if (&*BBI == SelectCond) 597 SelectCond = 0; 598 599 // If we ran out of things to eliminate, break out of the loop. 600 if (SelectCond == 0 && SI == 0) 601 break; 602 603 } 604 return true; 605 } 606 607 608 /// This function implements the transforms common to both integer division 609 /// instructions (udiv and sdiv). It is called by the visitors to those integer 610 /// division instructions. 611 /// @brief Common integer divide transforms 612 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) { 613 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 614 615 // The RHS is known non-zero. 616 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this)) { 617 I.setOperand(1, V); 618 return &I; 619 } 620 621 // Handle cases involving: [su]div X, (select Cond, Y, Z) 622 // This does not apply for fdiv. 623 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I)) 624 return &I; 625 626 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) { 627 // (X / C1) / C2 -> X / (C1*C2) 628 if (Instruction *LHS = dyn_cast<Instruction>(Op0)) 629 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode()) 630 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) { 631 if (MultiplyOverflows(RHS, LHSRHS, 632 I.getOpcode()==Instruction::SDiv)) 633 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType())); 634 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0), 635 ConstantExpr::getMul(RHS, LHSRHS)); 636 } 637 638 if (!RHS->isZero()) { // avoid X udiv 0 639 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 640 if (Instruction *R = FoldOpIntoSelect(I, SI)) 641 return R; 642 if (isa<PHINode>(Op0)) 643 if (Instruction *NV = FoldOpIntoPhi(I)) 644 return NV; 645 } 646 } 647 648 // See if we can fold away this div instruction. 649 if (SimplifyDemandedInstructionBits(I)) 650 return &I; 651 652 // (X - (X rem Y)) / Y -> X / Y; usually originates as ((X / Y) * Y) / Y 653 Value *X = 0, *Z = 0; 654 if (match(Op0, m_Sub(m_Value(X), m_Value(Z)))) { // (X - Z) / Y; Y = Op1 655 bool isSigned = I.getOpcode() == Instruction::SDiv; 656 if ((isSigned && match(Z, m_SRem(m_Specific(X), m_Specific(Op1)))) || 657 (!isSigned && match(Z, m_URem(m_Specific(X), m_Specific(Op1))))) 658 return BinaryOperator::Create(I.getOpcode(), X, Op1); 659 } 660 661 return 0; 662 } 663 664 /// dyn_castZExtVal - Checks if V is a zext or constant that can 665 /// be truncated to Ty without losing bits. 666 static Value *dyn_castZExtVal(Value *V, Type *Ty) { 667 if (ZExtInst *Z = dyn_cast<ZExtInst>(V)) { 668 if (Z->getSrcTy() == Ty) 669 return Z->getOperand(0); 670 } else if (ConstantInt *C = dyn_cast<ConstantInt>(V)) { 671 if (C->getValue().getActiveBits() <= cast<IntegerType>(Ty)->getBitWidth()) 672 return ConstantExpr::getTrunc(C, Ty); 673 } 674 return 0; 675 } 676 677 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) { 678 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 679 680 if (Value *V = SimplifyUDivInst(Op0, Op1, TD)) 681 return ReplaceInstUsesWith(I, V); 682 683 // Handle the integer div common cases 684 if (Instruction *Common = commonIDivTransforms(I)) 685 return Common; 686 687 { 688 // X udiv 2^C -> X >> C 689 // Check to see if this is an unsigned division with an exact power of 2, 690 // if so, convert to a right shift. 691 const APInt *C; 692 if (match(Op1, m_Power2(C))) { 693 BinaryOperator *LShr = 694 BinaryOperator::CreateLShr(Op0, 695 ConstantInt::get(Op0->getType(), 696 C->logBase2())); 697 if (I.isExact()) LShr->setIsExact(); 698 return LShr; 699 } 700 } 701 702 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) { 703 // X udiv C, where C >= signbit 704 if (C->getValue().isNegative()) { 705 Value *IC = Builder->CreateICmpULT(Op0, C); 706 return SelectInst::Create(IC, Constant::getNullValue(I.getType()), 707 ConstantInt::get(I.getType(), 1)); 708 } 709 } 710 711 // (x lshr C1) udiv C2 --> x udiv (C2 << C1) 712 if (ConstantInt *C2 = dyn_cast<ConstantInt>(Op1)) { 713 Value *X; 714 ConstantInt *C1; 715 if (match(Op0, m_LShr(m_Value(X), m_ConstantInt(C1)))) { 716 APInt NC = C2->getValue().shl(C1->getLimitedValue(C1->getBitWidth()-1)); 717 return BinaryOperator::CreateUDiv(X, Builder->getInt(NC)); 718 } 719 } 720 721 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2) 722 { const APInt *CI; Value *N; 723 if (match(Op1, m_Shl(m_Power2(CI), m_Value(N))) || 724 match(Op1, m_ZExt(m_Shl(m_Power2(CI), m_Value(N))))) { 725 if (*CI != 1) 726 N = Builder->CreateAdd(N, 727 ConstantInt::get(N->getType(), CI->logBase2())); 728 if (ZExtInst *Z = dyn_cast<ZExtInst>(Op1)) 729 N = Builder->CreateZExt(N, Z->getDestTy()); 730 if (I.isExact()) 731 return BinaryOperator::CreateExactLShr(Op0, N); 732 return BinaryOperator::CreateLShr(Op0, N); 733 } 734 } 735 736 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2) 737 // where C1&C2 are powers of two. 738 { Value *Cond; const APInt *C1, *C2; 739 if (match(Op1, m_Select(m_Value(Cond), m_Power2(C1), m_Power2(C2)))) { 740 // Construct the "on true" case of the select 741 Value *TSI = Builder->CreateLShr(Op0, C1->logBase2(), Op1->getName()+".t", 742 I.isExact()); 743 744 // Construct the "on false" case of the select 745 Value *FSI = Builder->CreateLShr(Op0, C2->logBase2(), Op1->getName()+".f", 746 I.isExact()); 747 748 // construct the select instruction and return it. 749 return SelectInst::Create(Cond, TSI, FSI); 750 } 751 } 752 753 // (zext A) udiv (zext B) --> zext (A udiv B) 754 if (ZExtInst *ZOp0 = dyn_cast<ZExtInst>(Op0)) 755 if (Value *ZOp1 = dyn_castZExtVal(Op1, ZOp0->getSrcTy())) 756 return new ZExtInst(Builder->CreateUDiv(ZOp0->getOperand(0), ZOp1, "div", 757 I.isExact()), 758 I.getType()); 759 760 return 0; 761 } 762 763 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) { 764 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 765 766 if (Value *V = SimplifySDivInst(Op0, Op1, TD)) 767 return ReplaceInstUsesWith(I, V); 768 769 // Handle the integer div common cases 770 if (Instruction *Common = commonIDivTransforms(I)) 771 return Common; 772 773 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) { 774 // sdiv X, -1 == -X 775 if (RHS->isAllOnesValue()) 776 return BinaryOperator::CreateNeg(Op0); 777 778 // sdiv X, C --> ashr exact X, log2(C) 779 if (I.isExact() && RHS->getValue().isNonNegative() && 780 RHS->getValue().isPowerOf2()) { 781 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(), 782 RHS->getValue().exactLogBase2()); 783 return BinaryOperator::CreateExactAShr(Op0, ShAmt, I.getName()); 784 } 785 786 // -X/C --> X/-C provided the negation doesn't overflow. 787 if (SubOperator *Sub = dyn_cast<SubOperator>(Op0)) 788 if (match(Sub->getOperand(0), m_Zero()) && Sub->hasNoSignedWrap()) 789 return BinaryOperator::CreateSDiv(Sub->getOperand(1), 790 ConstantExpr::getNeg(RHS)); 791 } 792 793 // If the sign bits of both operands are zero (i.e. we can prove they are 794 // unsigned inputs), turn this into a udiv. 795 if (I.getType()->isIntegerTy()) { 796 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())); 797 if (MaskedValueIsZero(Op0, Mask)) { 798 if (MaskedValueIsZero(Op1, Mask)) { 799 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set 800 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName()); 801 } 802 803 if (match(Op1, m_Shl(m_Power2(), m_Value()))) { 804 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y) 805 // Safe because the only negative value (1 << Y) can take on is 806 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have 807 // the sign bit set. 808 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName()); 809 } 810 } 811 } 812 813 return 0; 814 } 815 816 /// CvtFDivConstToReciprocal tries to convert X/C into X*1/C if C not a special 817 /// FP value and: 818 /// 1) 1/C is exact, or 819 /// 2) reciprocal is allowed. 820 /// If the convertion was successful, the simplified expression "X * 1/C" is 821 /// returned; otherwise, NULL is returned. 822 /// 823 static Instruction *CvtFDivConstToReciprocal(Value *Dividend, 824 ConstantFP *Divisor, 825 bool AllowReciprocal) { 826 const APFloat &FpVal = Divisor->getValueAPF(); 827 APFloat Reciprocal(FpVal.getSemantics()); 828 bool Cvt = FpVal.getExactInverse(&Reciprocal); 829 830 if (!Cvt && AllowReciprocal && FpVal.isNormal()) { 831 Reciprocal = APFloat(FpVal.getSemantics(), 1.0f); 832 (void)Reciprocal.divide(FpVal, APFloat::rmNearestTiesToEven); 833 Cvt = !Reciprocal.isDenormal(); 834 } 835 836 if (!Cvt) 837 return 0; 838 839 ConstantFP *R; 840 R = ConstantFP::get(Dividend->getType()->getContext(), Reciprocal); 841 return BinaryOperator::CreateFMul(Dividend, R); 842 } 843 844 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) { 845 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 846 847 if (Value *V = SimplifyFDivInst(Op0, Op1, TD)) 848 return ReplaceInstUsesWith(I, V); 849 850 bool AllowReassociate = I.hasUnsafeAlgebra(); 851 bool AllowReciprocal = I.hasAllowReciprocal(); 852 853 if (ConstantFP *Op1C = dyn_cast<ConstantFP>(Op1)) { 854 if (AllowReassociate) { 855 ConstantFP *C1 = 0; 856 ConstantFP *C2 = Op1C; 857 Value *X; 858 Instruction *Res = 0; 859 860 if (match(Op0, m_FMul(m_Value(X), m_ConstantFP(C1)))) { 861 // (X*C1)/C2 => X * (C1/C2) 862 // 863 Constant *C = ConstantExpr::getFDiv(C1, C2); 864 const APFloat &F = cast<ConstantFP>(C)->getValueAPF(); 865 if (F.isNormal() && !F.isDenormal()) 866 Res = BinaryOperator::CreateFMul(X, C); 867 } else if (match(Op0, m_FDiv(m_Value(X), m_ConstantFP(C1)))) { 868 // (X/C1)/C2 => X /(C2*C1) [=> X * 1/(C2*C1) if reciprocal is allowed] 869 // 870 Constant *C = ConstantExpr::getFMul(C1, C2); 871 const APFloat &F = cast<ConstantFP>(C)->getValueAPF(); 872 if (F.isNormal() && !F.isDenormal()) { 873 Res = CvtFDivConstToReciprocal(X, cast<ConstantFP>(C), 874 AllowReciprocal); 875 if (!Res) 876 Res = BinaryOperator::CreateFDiv(X, C); 877 } 878 } 879 880 if (Res) { 881 Res->setFastMathFlags(I.getFastMathFlags()); 882 return Res; 883 } 884 } 885 886 // X / C => X * 1/C 887 if (Instruction *T = CvtFDivConstToReciprocal(Op0, Op1C, AllowReciprocal)) 888 return T; 889 890 return 0; 891 } 892 893 if (AllowReassociate && isa<ConstantFP>(Op0)) { 894 ConstantFP *C1 = cast<ConstantFP>(Op0), *C2; 895 Constant *Fold = 0; 896 Value *X; 897 bool CreateDiv = true; 898 899 // C1 / (X*C2) => (C1/C2) / X 900 if (match(Op1, m_FMul(m_Value(X), m_ConstantFP(C2)))) 901 Fold = ConstantExpr::getFDiv(C1, C2); 902 else if (match(Op1, m_FDiv(m_Value(X), m_ConstantFP(C2)))) { 903 // C1 / (X/C2) => (C1*C2) / X 904 Fold = ConstantExpr::getFMul(C1, C2); 905 } else if (match(Op1, m_FDiv(m_ConstantFP(C2), m_Value(X)))) { 906 // C1 / (C2/X) => (C1/C2) * X 907 Fold = ConstantExpr::getFDiv(C1, C2); 908 CreateDiv = false; 909 } 910 911 if (Fold) { 912 const APFloat &FoldC = cast<ConstantFP>(Fold)->getValueAPF(); 913 if (FoldC.isNormal() && !FoldC.isDenormal()) { 914 Instruction *R = CreateDiv ? 915 BinaryOperator::CreateFDiv(Fold, X) : 916 BinaryOperator::CreateFMul(X, Fold); 917 R->setFastMathFlags(I.getFastMathFlags()); 918 return R; 919 } 920 } 921 return 0; 922 } 923 924 if (AllowReassociate) { 925 Value *X, *Y; 926 Value *NewInst = 0; 927 Instruction *SimpR = 0; 928 929 if (Op0->hasOneUse() && match(Op0, m_FDiv(m_Value(X), m_Value(Y)))) { 930 // (X/Y) / Z => X / (Y*Z) 931 // 932 if (!isa<ConstantFP>(Y) || !isa<ConstantFP>(Op1)) { 933 NewInst = Builder->CreateFMul(Y, Op1); 934 SimpR = BinaryOperator::CreateFDiv(X, NewInst); 935 } 936 } else if (Op1->hasOneUse() && match(Op1, m_FDiv(m_Value(X), m_Value(Y)))) { 937 // Z / (X/Y) => Z*Y / X 938 // 939 if (!isa<ConstantFP>(Y) || !isa<ConstantFP>(Op0)) { 940 NewInst = Builder->CreateFMul(Op0, Y); 941 SimpR = BinaryOperator::CreateFDiv(NewInst, X); 942 } 943 } 944 945 if (NewInst) { 946 if (Instruction *T = dyn_cast<Instruction>(NewInst)) 947 T->setDebugLoc(I.getDebugLoc()); 948 SimpR->setFastMathFlags(I.getFastMathFlags()); 949 return SimpR; 950 } 951 } 952 953 return 0; 954 } 955 956 /// This function implements the transforms common to both integer remainder 957 /// instructions (urem and srem). It is called by the visitors to those integer 958 /// remainder instructions. 959 /// @brief Common integer remainder transforms 960 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) { 961 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 962 963 // The RHS is known non-zero. 964 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this)) { 965 I.setOperand(1, V); 966 return &I; 967 } 968 969 // Handle cases involving: rem X, (select Cond, Y, Z) 970 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I)) 971 return &I; 972 973 if (isa<ConstantInt>(Op1)) { 974 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) { 975 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) { 976 if (Instruction *R = FoldOpIntoSelect(I, SI)) 977 return R; 978 } else if (isa<PHINode>(Op0I)) { 979 if (Instruction *NV = FoldOpIntoPhi(I)) 980 return NV; 981 } 982 983 // See if we can fold away this rem instruction. 984 if (SimplifyDemandedInstructionBits(I)) 985 return &I; 986 } 987 } 988 989 return 0; 990 } 991 992 Instruction *InstCombiner::visitURem(BinaryOperator &I) { 993 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 994 995 if (Value *V = SimplifyURemInst(Op0, Op1, TD)) 996 return ReplaceInstUsesWith(I, V); 997 998 if (Instruction *common = commonIRemTransforms(I)) 999 return common; 1000 1001 // X urem C^2 -> X and C-1 1002 { const APInt *C; 1003 if (match(Op1, m_Power2(C))) 1004 return BinaryOperator::CreateAnd(Op0, 1005 ConstantInt::get(I.getType(), *C-1)); 1006 } 1007 1008 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1) 1009 if (match(Op1, m_Shl(m_Power2(), m_Value()))) { 1010 Constant *N1 = Constant::getAllOnesValue(I.getType()); 1011 Value *Add = Builder->CreateAdd(Op1, N1); 1012 return BinaryOperator::CreateAnd(Op0, Add); 1013 } 1014 1015 // urem X, (select Cond, 2^C1, 2^C2) --> 1016 // select Cond, (and X, C1-1), (and X, C2-1) 1017 // when C1&C2 are powers of two. 1018 { Value *Cond; const APInt *C1, *C2; 1019 if (match(Op1, m_Select(m_Value(Cond), m_Power2(C1), m_Power2(C2)))) { 1020 Value *TrueAnd = Builder->CreateAnd(Op0, *C1-1, Op1->getName()+".t"); 1021 Value *FalseAnd = Builder->CreateAnd(Op0, *C2-1, Op1->getName()+".f"); 1022 return SelectInst::Create(Cond, TrueAnd, FalseAnd); 1023 } 1024 } 1025 1026 // (zext A) urem (zext B) --> zext (A urem B) 1027 if (ZExtInst *ZOp0 = dyn_cast<ZExtInst>(Op0)) 1028 if (Value *ZOp1 = dyn_castZExtVal(Op1, ZOp0->getSrcTy())) 1029 return new ZExtInst(Builder->CreateURem(ZOp0->getOperand(0), ZOp1), 1030 I.getType()); 1031 1032 return 0; 1033 } 1034 1035 Instruction *InstCombiner::visitSRem(BinaryOperator &I) { 1036 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1037 1038 if (Value *V = SimplifySRemInst(Op0, Op1, TD)) 1039 return ReplaceInstUsesWith(I, V); 1040 1041 // Handle the integer rem common cases 1042 if (Instruction *Common = commonIRemTransforms(I)) 1043 return Common; 1044 1045 if (Value *RHSNeg = dyn_castNegVal(Op1)) 1046 if (!isa<Constant>(RHSNeg) || 1047 (isa<ConstantInt>(RHSNeg) && 1048 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) { 1049 // X % -Y -> X % Y 1050 Worklist.AddValue(I.getOperand(1)); 1051 I.setOperand(1, RHSNeg); 1052 return &I; 1053 } 1054 1055 // If the sign bits of both operands are zero (i.e. we can prove they are 1056 // unsigned inputs), turn this into a urem. 1057 if (I.getType()->isIntegerTy()) { 1058 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())); 1059 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) { 1060 // X srem Y -> X urem Y, iff X and Y don't have sign bit set 1061 return BinaryOperator::CreateURem(Op0, Op1, I.getName()); 1062 } 1063 } 1064 1065 // If it's a constant vector, flip any negative values positive. 1066 if (isa<ConstantVector>(Op1) || isa<ConstantDataVector>(Op1)) { 1067 Constant *C = cast<Constant>(Op1); 1068 unsigned VWidth = C->getType()->getVectorNumElements(); 1069 1070 bool hasNegative = false; 1071 bool hasMissing = false; 1072 for (unsigned i = 0; i != VWidth; ++i) { 1073 Constant *Elt = C->getAggregateElement(i); 1074 if (Elt == 0) { 1075 hasMissing = true; 1076 break; 1077 } 1078 1079 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elt)) 1080 if (RHS->isNegative()) 1081 hasNegative = true; 1082 } 1083 1084 if (hasNegative && !hasMissing) { 1085 SmallVector<Constant *, 16> Elts(VWidth); 1086 for (unsigned i = 0; i != VWidth; ++i) { 1087 Elts[i] = C->getAggregateElement(i); // Handle undef, etc. 1088 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elts[i])) { 1089 if (RHS->isNegative()) 1090 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS)); 1091 } 1092 } 1093 1094 Constant *NewRHSV = ConstantVector::get(Elts); 1095 if (NewRHSV != C) { // Don't loop on -MININT 1096 Worklist.AddValue(I.getOperand(1)); 1097 I.setOperand(1, NewRHSV); 1098 return &I; 1099 } 1100 } 1101 } 1102 1103 return 0; 1104 } 1105 1106 Instruction *InstCombiner::visitFRem(BinaryOperator &I) { 1107 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1108 1109 if (Value *V = SimplifyFRemInst(Op0, Op1, TD)) 1110 return ReplaceInstUsesWith(I, V); 1111 1112 // Handle cases involving: rem X, (select Cond, Y, Z) 1113 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I)) 1114 return &I; 1115 1116 return 0; 1117 } 1118