InstCombineShifts.cpp revision 226633
1//===- InstCombineShifts.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 visitShl, visitLShr, and visitAShr functions. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/IntrinsicInst.h" 16#include "llvm/Analysis/ConstantFolding.h" 17#include "llvm/Analysis/InstructionSimplify.h" 18#include "llvm/Support/PatternMatch.h" 19using namespace llvm; 20using namespace PatternMatch; 21 22Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) { 23 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType()); 24 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 25 26 // See if we can fold away this shift. 27 if (SimplifyDemandedInstructionBits(I)) 28 return &I; 29 30 // Try to fold constant and into select arguments. 31 if (isa<Constant>(Op0)) 32 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 33 if (Instruction *R = FoldOpIntoSelect(I, SI)) 34 return R; 35 36 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1)) 37 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I)) 38 return Res; 39 40 // X shift (A srem B) -> X shift (A and B-1) iff B is a power of 2. 41 // Because shifts by negative values (which could occur if A were negative) 42 // are undefined. 43 Value *A; const APInt *B; 44 if (Op1->hasOneUse() && match(Op1, m_SRem(m_Value(A), m_Power2(B)))) { 45 // FIXME: Should this get moved into SimplifyDemandedBits by saying we don't 46 // demand the sign bit (and many others) here?? 47 Value *Rem = Builder->CreateAnd(A, ConstantInt::get(I.getType(), *B-1), 48 Op1->getName()); 49 I.setOperand(1, Rem); 50 return &I; 51 } 52 53 return 0; 54} 55 56/// CanEvaluateShifted - See if we can compute the specified value, but shifted 57/// logically to the left or right by some number of bits. This should return 58/// true if the expression can be computed for the same cost as the current 59/// expression tree. This is used to eliminate extraneous shifting from things 60/// like: 61/// %C = shl i128 %A, 64 62/// %D = shl i128 %B, 96 63/// %E = or i128 %C, %D 64/// %F = lshr i128 %E, 64 65/// where the client will ask if E can be computed shifted right by 64-bits. If 66/// this succeeds, the GetShiftedValue function will be called to produce the 67/// value. 68static bool CanEvaluateShifted(Value *V, unsigned NumBits, bool isLeftShift, 69 InstCombiner &IC) { 70 // We can always evaluate constants shifted. 71 if (isa<Constant>(V)) 72 return true; 73 74 Instruction *I = dyn_cast<Instruction>(V); 75 if (!I) return false; 76 77 // If this is the opposite shift, we can directly reuse the input of the shift 78 // if the needed bits are already zero in the input. This allows us to reuse 79 // the value which means that we don't care if the shift has multiple uses. 80 // TODO: Handle opposite shift by exact value. 81 ConstantInt *CI = 0; 82 if ((isLeftShift && match(I, m_LShr(m_Value(), m_ConstantInt(CI)))) || 83 (!isLeftShift && match(I, m_Shl(m_Value(), m_ConstantInt(CI))))) { 84 if (CI->getZExtValue() == NumBits) { 85 // TODO: Check that the input bits are already zero with MaskedValueIsZero 86#if 0 87 // If this is a truncate of a logical shr, we can truncate it to a smaller 88 // lshr iff we know that the bits we would otherwise be shifting in are 89 // already zeros. 90 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 91 uint32_t BitWidth = Ty->getScalarSizeInBits(); 92 if (MaskedValueIsZero(I->getOperand(0), 93 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && 94 CI->getLimitedValue(BitWidth) < BitWidth) { 95 return CanEvaluateTruncated(I->getOperand(0), Ty); 96 } 97#endif 98 99 } 100 } 101 102 // We can't mutate something that has multiple uses: doing so would 103 // require duplicating the instruction in general, which isn't profitable. 104 if (!I->hasOneUse()) return false; 105 106 switch (I->getOpcode()) { 107 default: return false; 108 case Instruction::And: 109 case Instruction::Or: 110 case Instruction::Xor: 111 // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted. 112 return CanEvaluateShifted(I->getOperand(0), NumBits, isLeftShift, IC) && 113 CanEvaluateShifted(I->getOperand(1), NumBits, isLeftShift, IC); 114 115 case Instruction::Shl: { 116 // We can often fold the shift into shifts-by-a-constant. 117 CI = dyn_cast<ConstantInt>(I->getOperand(1)); 118 if (CI == 0) return false; 119 120 // We can always fold shl(c1)+shl(c2) -> shl(c1+c2). 121 if (isLeftShift) return true; 122 123 // We can always turn shl(c)+shr(c) -> and(c2). 124 if (CI->getValue() == NumBits) return true; 125 126 unsigned TypeWidth = I->getType()->getScalarSizeInBits(); 127 128 // We can turn shl(c1)+shr(c2) -> shl(c3)+and(c4), but it isn't 129 // profitable unless we know the and'd out bits are already zero. 130 if (CI->getZExtValue() > NumBits) { 131 unsigned LowBits = TypeWidth - CI->getZExtValue(); 132 if (MaskedValueIsZero(I->getOperand(0), 133 APInt::getLowBitsSet(TypeWidth, NumBits) << LowBits)) 134 return true; 135 } 136 137 return false; 138 } 139 case Instruction::LShr: { 140 // We can often fold the shift into shifts-by-a-constant. 141 CI = dyn_cast<ConstantInt>(I->getOperand(1)); 142 if (CI == 0) return false; 143 144 // We can always fold lshr(c1)+lshr(c2) -> lshr(c1+c2). 145 if (!isLeftShift) return true; 146 147 // We can always turn lshr(c)+shl(c) -> and(c2). 148 if (CI->getValue() == NumBits) return true; 149 150 unsigned TypeWidth = I->getType()->getScalarSizeInBits(); 151 152 // We can always turn lshr(c1)+shl(c2) -> lshr(c3)+and(c4), but it isn't 153 // profitable unless we know the and'd out bits are already zero. 154 if (CI->getZExtValue() > NumBits) { 155 unsigned LowBits = CI->getZExtValue() - NumBits; 156 if (MaskedValueIsZero(I->getOperand(0), 157 APInt::getLowBitsSet(TypeWidth, NumBits) << LowBits)) 158 return true; 159 } 160 161 return false; 162 } 163 case Instruction::Select: { 164 SelectInst *SI = cast<SelectInst>(I); 165 return CanEvaluateShifted(SI->getTrueValue(), NumBits, isLeftShift, IC) && 166 CanEvaluateShifted(SI->getFalseValue(), NumBits, isLeftShift, IC); 167 } 168 case Instruction::PHI: { 169 // We can change a phi if we can change all operands. Note that we never 170 // get into trouble with cyclic PHIs here because we only consider 171 // instructions with a single use. 172 PHINode *PN = cast<PHINode>(I); 173 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 174 if (!CanEvaluateShifted(PN->getIncomingValue(i), NumBits, isLeftShift,IC)) 175 return false; 176 return true; 177 } 178 } 179} 180 181/// GetShiftedValue - When CanEvaluateShifted returned true for an expression, 182/// this value inserts the new computation that produces the shifted value. 183static Value *GetShiftedValue(Value *V, unsigned NumBits, bool isLeftShift, 184 InstCombiner &IC) { 185 // We can always evaluate constants shifted. 186 if (Constant *C = dyn_cast<Constant>(V)) { 187 if (isLeftShift) 188 V = IC.Builder->CreateShl(C, NumBits); 189 else 190 V = IC.Builder->CreateLShr(C, NumBits); 191 // If we got a constantexpr back, try to simplify it with TD info. 192 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 193 V = ConstantFoldConstantExpression(CE, IC.getTargetData()); 194 return V; 195 } 196 197 Instruction *I = cast<Instruction>(V); 198 IC.Worklist.Add(I); 199 200 switch (I->getOpcode()) { 201 default: assert(0 && "Inconsistency with CanEvaluateShifted"); 202 case Instruction::And: 203 case Instruction::Or: 204 case Instruction::Xor: 205 // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted. 206 I->setOperand(0, GetShiftedValue(I->getOperand(0), NumBits,isLeftShift,IC)); 207 I->setOperand(1, GetShiftedValue(I->getOperand(1), NumBits,isLeftShift,IC)); 208 return I; 209 210 case Instruction::Shl: { 211 BinaryOperator *BO = cast<BinaryOperator>(I); 212 unsigned TypeWidth = BO->getType()->getScalarSizeInBits(); 213 214 // We only accept shifts-by-a-constant in CanEvaluateShifted. 215 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1)); 216 217 // We can always fold shl(c1)+shl(c2) -> shl(c1+c2). 218 if (isLeftShift) { 219 // If this is oversized composite shift, then unsigned shifts get 0. 220 unsigned NewShAmt = NumBits+CI->getZExtValue(); 221 if (NewShAmt >= TypeWidth) 222 return Constant::getNullValue(I->getType()); 223 224 BO->setOperand(1, ConstantInt::get(BO->getType(), NewShAmt)); 225 BO->setHasNoUnsignedWrap(false); 226 BO->setHasNoSignedWrap(false); 227 return I; 228 } 229 230 // We turn shl(c)+lshr(c) -> and(c2) if the input doesn't already have 231 // zeros. 232 if (CI->getValue() == NumBits) { 233 APInt Mask(APInt::getLowBitsSet(TypeWidth, TypeWidth - NumBits)); 234 V = IC.Builder->CreateAnd(BO->getOperand(0), 235 ConstantInt::get(BO->getContext(), Mask)); 236 if (Instruction *VI = dyn_cast<Instruction>(V)) { 237 VI->moveBefore(BO); 238 VI->takeName(BO); 239 } 240 return V; 241 } 242 243 // We turn shl(c1)+shr(c2) -> shl(c3)+and(c4), but only when we know that 244 // the and won't be needed. 245 assert(CI->getZExtValue() > NumBits); 246 BO->setOperand(1, ConstantInt::get(BO->getType(), 247 CI->getZExtValue() - NumBits)); 248 BO->setHasNoUnsignedWrap(false); 249 BO->setHasNoSignedWrap(false); 250 return BO; 251 } 252 case Instruction::LShr: { 253 BinaryOperator *BO = cast<BinaryOperator>(I); 254 unsigned TypeWidth = BO->getType()->getScalarSizeInBits(); 255 // We only accept shifts-by-a-constant in CanEvaluateShifted. 256 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1)); 257 258 // We can always fold lshr(c1)+lshr(c2) -> lshr(c1+c2). 259 if (!isLeftShift) { 260 // If this is oversized composite shift, then unsigned shifts get 0. 261 unsigned NewShAmt = NumBits+CI->getZExtValue(); 262 if (NewShAmt >= TypeWidth) 263 return Constant::getNullValue(BO->getType()); 264 265 BO->setOperand(1, ConstantInt::get(BO->getType(), NewShAmt)); 266 BO->setIsExact(false); 267 return I; 268 } 269 270 // We turn lshr(c)+shl(c) -> and(c2) if the input doesn't already have 271 // zeros. 272 if (CI->getValue() == NumBits) { 273 APInt Mask(APInt::getHighBitsSet(TypeWidth, TypeWidth - NumBits)); 274 V = IC.Builder->CreateAnd(I->getOperand(0), 275 ConstantInt::get(BO->getContext(), Mask)); 276 if (Instruction *VI = dyn_cast<Instruction>(V)) { 277 VI->moveBefore(I); 278 VI->takeName(I); 279 } 280 return V; 281 } 282 283 // We turn lshr(c1)+shl(c2) -> lshr(c3)+and(c4), but only when we know that 284 // the and won't be needed. 285 assert(CI->getZExtValue() > NumBits); 286 BO->setOperand(1, ConstantInt::get(BO->getType(), 287 CI->getZExtValue() - NumBits)); 288 BO->setIsExact(false); 289 return BO; 290 } 291 292 case Instruction::Select: 293 I->setOperand(1, GetShiftedValue(I->getOperand(1), NumBits,isLeftShift,IC)); 294 I->setOperand(2, GetShiftedValue(I->getOperand(2), NumBits,isLeftShift,IC)); 295 return I; 296 case Instruction::PHI: { 297 // We can change a phi if we can change all operands. Note that we never 298 // get into trouble with cyclic PHIs here because we only consider 299 // instructions with a single use. 300 PHINode *PN = cast<PHINode>(I); 301 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 302 PN->setIncomingValue(i, GetShiftedValue(PN->getIncomingValue(i), 303 NumBits, isLeftShift, IC)); 304 return PN; 305 } 306 } 307} 308 309 310 311Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1, 312 BinaryOperator &I) { 313 bool isLeftShift = I.getOpcode() == Instruction::Shl; 314 315 316 // See if we can propagate this shift into the input, this covers the trivial 317 // cast of lshr(shl(x,c1),c2) as well as other more complex cases. 318 if (I.getOpcode() != Instruction::AShr && 319 CanEvaluateShifted(Op0, Op1->getZExtValue(), isLeftShift, *this)) { 320 DEBUG(dbgs() << "ICE: GetShiftedValue propagating shift through expression" 321 " to eliminate shift:\n IN: " << *Op0 << "\n SH: " << I <<"\n"); 322 323 return ReplaceInstUsesWith(I, 324 GetShiftedValue(Op0, Op1->getZExtValue(), isLeftShift, *this)); 325 } 326 327 328 // See if we can simplify any instructions used by the instruction whose sole 329 // purpose is to compute bits we don't care about. 330 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits(); 331 332 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate 333 // a signed shift. 334 // 335 if (Op1->uge(TypeBits)) { 336 if (I.getOpcode() != Instruction::AShr) 337 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType())); 338 // ashr i32 X, 32 --> ashr i32 X, 31 339 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1)); 340 return &I; 341 } 342 343 // ((X*C1) << C2) == (X * (C1 << C2)) 344 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) 345 if (BO->getOpcode() == Instruction::Mul && isLeftShift) 346 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1))) 347 return BinaryOperator::CreateMul(BO->getOperand(0), 348 ConstantExpr::getShl(BOOp, Op1)); 349 350 // Try to fold constant and into select arguments. 351 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 352 if (Instruction *R = FoldOpIntoSelect(I, SI)) 353 return R; 354 if (isa<PHINode>(Op0)) 355 if (Instruction *NV = FoldOpIntoPhi(I)) 356 return NV; 357 358 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2)) 359 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) { 360 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0)); 361 // If 'shift2' is an ashr, we would have to get the sign bit into a funny 362 // place. Don't try to do this transformation in this case. Also, we 363 // require that the input operand is a shift-by-constant so that we have 364 // confidence that the shifts will get folded together. We could do this 365 // xform in more cases, but it is unlikely to be profitable. 366 if (TrOp && I.isLogicalShift() && TrOp->isShift() && 367 isa<ConstantInt>(TrOp->getOperand(1))) { 368 // Okay, we'll do this xform. Make the shift of shift. 369 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType()); 370 // (shift2 (shift1 & 0x00FF), c2) 371 Value *NSh = Builder->CreateBinOp(I.getOpcode(), TrOp, ShAmt,I.getName()); 372 373 // For logical shifts, the truncation has the effect of making the high 374 // part of the register be zeros. Emulate this by inserting an AND to 375 // clear the top bits as needed. This 'and' will usually be zapped by 376 // other xforms later if dead. 377 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits(); 378 unsigned DstSize = TI->getType()->getScalarSizeInBits(); 379 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize)); 380 381 // The mask we constructed says what the trunc would do if occurring 382 // between the shifts. We want to know the effect *after* the second 383 // shift. We know that it is a logical shift by a constant, so adjust the 384 // mask as appropriate. 385 if (I.getOpcode() == Instruction::Shl) 386 MaskV <<= Op1->getZExtValue(); 387 else { 388 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift"); 389 MaskV = MaskV.lshr(Op1->getZExtValue()); 390 } 391 392 // shift1 & 0x00FF 393 Value *And = Builder->CreateAnd(NSh, 394 ConstantInt::get(I.getContext(), MaskV), 395 TI->getName()); 396 397 // Return the value truncated to the interesting size. 398 return new TruncInst(And, I.getType()); 399 } 400 } 401 402 if (Op0->hasOneUse()) { 403 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) { 404 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C) 405 Value *V1, *V2; 406 ConstantInt *CC; 407 switch (Op0BO->getOpcode()) { 408 default: break; 409 case Instruction::Add: 410 case Instruction::And: 411 case Instruction::Or: 412 case Instruction::Xor: { 413 // These operators commute. 414 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C) 415 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() && 416 match(Op0BO->getOperand(1), m_Shr(m_Value(V1), 417 m_Specific(Op1)))) { 418 Value *YS = // (Y << C) 419 Builder->CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName()); 420 // (X + (Y << C)) 421 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), YS, V1, 422 Op0BO->getOperand(1)->getName()); 423 uint32_t Op1Val = Op1->getLimitedValue(TypeBits); 424 return BinaryOperator::CreateAnd(X, ConstantInt::get(I.getContext(), 425 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val))); 426 } 427 428 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C)) 429 Value *Op0BOOp1 = Op0BO->getOperand(1); 430 if (isLeftShift && Op0BOOp1->hasOneUse() && 431 match(Op0BOOp1, 432 m_And(m_Shr(m_Value(V1), m_Specific(Op1)), 433 m_ConstantInt(CC))) && 434 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) { 435 Value *YS = // (Y << C) 436 Builder->CreateShl(Op0BO->getOperand(0), Op1, 437 Op0BO->getName()); 438 // X & (CC << C) 439 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1), 440 V1->getName()+".mask"); 441 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM); 442 } 443 } 444 445 // FALL THROUGH. 446 case Instruction::Sub: { 447 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C) 448 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() && 449 match(Op0BO->getOperand(0), m_Shr(m_Value(V1), 450 m_Specific(Op1)))) { 451 Value *YS = // (Y << C) 452 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName()); 453 // (X + (Y << C)) 454 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), V1, YS, 455 Op0BO->getOperand(0)->getName()); 456 uint32_t Op1Val = Op1->getLimitedValue(TypeBits); 457 return BinaryOperator::CreateAnd(X, ConstantInt::get(I.getContext(), 458 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val))); 459 } 460 461 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C) 462 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() && 463 match(Op0BO->getOperand(0), 464 m_And(m_Shr(m_Value(V1), m_Value(V2)), 465 m_ConstantInt(CC))) && V2 == Op1 && 466 cast<BinaryOperator>(Op0BO->getOperand(0)) 467 ->getOperand(0)->hasOneUse()) { 468 Value *YS = // (Y << C) 469 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName()); 470 // X & (CC << C) 471 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1), 472 V1->getName()+".mask"); 473 474 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS); 475 } 476 477 break; 478 } 479 } 480 481 482 // If the operand is an bitwise operator with a constant RHS, and the 483 // shift is the only use, we can pull it out of the shift. 484 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) { 485 bool isValid = true; // Valid only for And, Or, Xor 486 bool highBitSet = false; // Transform if high bit of constant set? 487 488 switch (Op0BO->getOpcode()) { 489 default: isValid = false; break; // Do not perform transform! 490 case Instruction::Add: 491 isValid = isLeftShift; 492 break; 493 case Instruction::Or: 494 case Instruction::Xor: 495 highBitSet = false; 496 break; 497 case Instruction::And: 498 highBitSet = true; 499 break; 500 } 501 502 // If this is a signed shift right, and the high bit is modified 503 // by the logical operation, do not perform the transformation. 504 // The highBitSet boolean indicates the value of the high bit of 505 // the constant which would cause it to be modified for this 506 // operation. 507 // 508 if (isValid && I.getOpcode() == Instruction::AShr) 509 isValid = Op0C->getValue()[TypeBits-1] == highBitSet; 510 511 if (isValid) { 512 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1); 513 514 Value *NewShift = 515 Builder->CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1); 516 NewShift->takeName(Op0BO); 517 518 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift, 519 NewRHS); 520 } 521 } 522 } 523 } 524 525 // Find out if this is a shift of a shift by a constant. 526 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0); 527 if (ShiftOp && !ShiftOp->isShift()) 528 ShiftOp = 0; 529 530 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) { 531 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1)); 532 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits); 533 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits); 534 assert(ShiftAmt2 != 0 && "Should have been simplified earlier"); 535 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future. 536 Value *X = ShiftOp->getOperand(0); 537 538 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift. 539 540 IntegerType *Ty = cast<IntegerType>(I.getType()); 541 542 // Check for (X << c1) << c2 and (X >> c1) >> c2 543 if (I.getOpcode() == ShiftOp->getOpcode()) { 544 // If this is oversized composite shift, then unsigned shifts get 0, ashr 545 // saturates. 546 if (AmtSum >= TypeBits) { 547 if (I.getOpcode() != Instruction::AShr) 548 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType())); 549 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr. 550 } 551 552 return BinaryOperator::Create(I.getOpcode(), X, 553 ConstantInt::get(Ty, AmtSum)); 554 } 555 556 if (ShiftAmt1 == ShiftAmt2) { 557 // If we have ((X >>? C) << C), turn this into X & (-1 << C). 558 if (I.getOpcode() == Instruction::Shl && 559 ShiftOp->getOpcode() != Instruction::Shl) { 560 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1)); 561 return BinaryOperator::CreateAnd(X, 562 ConstantInt::get(I.getContext(),Mask)); 563 } 564 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C). 565 if (I.getOpcode() == Instruction::LShr && 566 ShiftOp->getOpcode() == Instruction::Shl) { 567 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1)); 568 return BinaryOperator::CreateAnd(X, 569 ConstantInt::get(I.getContext(), Mask)); 570 } 571 } else if (ShiftAmt1 < ShiftAmt2) { 572 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1; 573 574 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2) 575 if (I.getOpcode() == Instruction::Shl && 576 ShiftOp->getOpcode() != Instruction::Shl) { 577 assert(ShiftOp->getOpcode() == Instruction::LShr || 578 ShiftOp->getOpcode() == Instruction::AShr); 579 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff)); 580 581 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2)); 582 return BinaryOperator::CreateAnd(Shift, 583 ConstantInt::get(I.getContext(),Mask)); 584 } 585 586 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2) 587 if (I.getOpcode() == Instruction::LShr && 588 ShiftOp->getOpcode() == Instruction::Shl) { 589 assert(ShiftOp->getOpcode() == Instruction::Shl); 590 Value *Shift = Builder->CreateLShr(X, ConstantInt::get(Ty, ShiftDiff)); 591 592 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2)); 593 return BinaryOperator::CreateAnd(Shift, 594 ConstantInt::get(I.getContext(),Mask)); 595 } 596 597 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in. 598 } else { 599 assert(ShiftAmt2 < ShiftAmt1); 600 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2; 601 602 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2) 603 if (I.getOpcode() == Instruction::Shl && 604 ShiftOp->getOpcode() != Instruction::Shl) { 605 Value *Shift = Builder->CreateBinOp(ShiftOp->getOpcode(), X, 606 ConstantInt::get(Ty, ShiftDiff)); 607 608 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2)); 609 return BinaryOperator::CreateAnd(Shift, 610 ConstantInt::get(I.getContext(),Mask)); 611 } 612 613 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2) 614 if (I.getOpcode() == Instruction::LShr && 615 ShiftOp->getOpcode() == Instruction::Shl) { 616 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff)); 617 618 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2)); 619 return BinaryOperator::CreateAnd(Shift, 620 ConstantInt::get(I.getContext(),Mask)); 621 } 622 623 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in. 624 } 625 } 626 return 0; 627} 628 629Instruction *InstCombiner::visitShl(BinaryOperator &I) { 630 if (Value *V = SimplifyShlInst(I.getOperand(0), I.getOperand(1), 631 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(), 632 TD)) 633 return ReplaceInstUsesWith(I, V); 634 635 if (Instruction *V = commonShiftTransforms(I)) 636 return V; 637 638 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(I.getOperand(1))) { 639 unsigned ShAmt = Op1C->getZExtValue(); 640 641 // If the shifted-out value is known-zero, then this is a NUW shift. 642 if (!I.hasNoUnsignedWrap() && 643 MaskedValueIsZero(I.getOperand(0), 644 APInt::getHighBitsSet(Op1C->getBitWidth(), ShAmt))) { 645 I.setHasNoUnsignedWrap(); 646 return &I; 647 } 648 649 // If the shifted out value is all signbits, this is a NSW shift. 650 if (!I.hasNoSignedWrap() && 651 ComputeNumSignBits(I.getOperand(0)) > ShAmt) { 652 I.setHasNoSignedWrap(); 653 return &I; 654 } 655 } 656 657 // (C1 << A) << C2 -> (C1 << C2) << A 658 Constant *C1, *C2; 659 Value *A; 660 if (match(I.getOperand(0), m_OneUse(m_Shl(m_Constant(C1), m_Value(A)))) && 661 match(I.getOperand(1), m_Constant(C2))) 662 return BinaryOperator::CreateShl(ConstantExpr::getShl(C1, C2), A); 663 664 return 0; 665} 666 667Instruction *InstCombiner::visitLShr(BinaryOperator &I) { 668 if (Value *V = SimplifyLShrInst(I.getOperand(0), I.getOperand(1), 669 I.isExact(), TD)) 670 return ReplaceInstUsesWith(I, V); 671 672 if (Instruction *R = commonShiftTransforms(I)) 673 return R; 674 675 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 676 677 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 678 unsigned ShAmt = Op1C->getZExtValue(); 679 680 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) { 681 unsigned BitWidth = Op0->getType()->getScalarSizeInBits(); 682 // ctlz.i32(x)>>5 --> zext(x == 0) 683 // cttz.i32(x)>>5 --> zext(x == 0) 684 // ctpop.i32(x)>>5 --> zext(x == -1) 685 if ((II->getIntrinsicID() == Intrinsic::ctlz || 686 II->getIntrinsicID() == Intrinsic::cttz || 687 II->getIntrinsicID() == Intrinsic::ctpop) && 688 isPowerOf2_32(BitWidth) && Log2_32(BitWidth) == ShAmt) { 689 bool isCtPop = II->getIntrinsicID() == Intrinsic::ctpop; 690 Constant *RHS = ConstantInt::getSigned(Op0->getType(), isCtPop ? -1:0); 691 Value *Cmp = Builder->CreateICmpEQ(II->getArgOperand(0), RHS); 692 return new ZExtInst(Cmp, II->getType()); 693 } 694 } 695 696 // If the shifted-out value is known-zero, then this is an exact shift. 697 if (!I.isExact() && 698 MaskedValueIsZero(Op0,APInt::getLowBitsSet(Op1C->getBitWidth(),ShAmt))){ 699 I.setIsExact(); 700 return &I; 701 } 702 } 703 704 return 0; 705} 706 707Instruction *InstCombiner::visitAShr(BinaryOperator &I) { 708 if (Value *V = SimplifyAShrInst(I.getOperand(0), I.getOperand(1), 709 I.isExact(), TD)) 710 return ReplaceInstUsesWith(I, V); 711 712 if (Instruction *R = commonShiftTransforms(I)) 713 return R; 714 715 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 716 717 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 718 unsigned ShAmt = Op1C->getZExtValue(); 719 720 // If the input is a SHL by the same constant (ashr (shl X, C), C), then we 721 // have a sign-extend idiom. 722 Value *X; 723 if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1)))) { 724 // If the left shift is just shifting out partial signbits, delete the 725 // extension. 726 if (cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap()) 727 return ReplaceInstUsesWith(I, X); 728 729 // If the input is an extension from the shifted amount value, e.g. 730 // %x = zext i8 %A to i32 731 // %y = shl i32 %x, 24 732 // %z = ashr %y, 24 733 // then turn this into "z = sext i8 A to i32". 734 if (ZExtInst *ZI = dyn_cast<ZExtInst>(X)) { 735 uint32_t SrcBits = ZI->getOperand(0)->getType()->getScalarSizeInBits(); 736 uint32_t DestBits = ZI->getType()->getScalarSizeInBits(); 737 if (Op1C->getZExtValue() == DestBits-SrcBits) 738 return new SExtInst(ZI->getOperand(0), ZI->getType()); 739 } 740 } 741 742 // If the shifted-out value is known-zero, then this is an exact shift. 743 if (!I.isExact() && 744 MaskedValueIsZero(Op0,APInt::getLowBitsSet(Op1C->getBitWidth(),ShAmt))){ 745 I.setIsExact(); 746 return &I; 747 } 748 } 749 750 // See if we can turn a signed shr into an unsigned shr. 751 if (MaskedValueIsZero(Op0, 752 APInt::getSignBit(I.getType()->getScalarSizeInBits()))) 753 return BinaryOperator::CreateLShr(Op0, Op1); 754 755 // Arithmetic shifting an all-sign-bit value is a no-op. 756 unsigned NumSignBits = ComputeNumSignBits(Op0); 757 if (NumSignBits == Op0->getType()->getScalarSizeInBits()) 758 return ReplaceInstUsesWith(I, Op0); 759 760 return 0; 761} 762 763