1 //===- InstCombineShifts.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 visitShl, visitLShr, and visitAShr functions. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "InstCombineInternal.h" 14 #include "llvm/Analysis/ConstantFolding.h" 15 #include "llvm/Analysis/InstructionSimplify.h" 16 #include "llvm/IR/IntrinsicInst.h" 17 #include "llvm/IR/PatternMatch.h" 18 using namespace llvm; 19 using namespace PatternMatch; 20 21 #define DEBUG_TYPE "instcombine" 22 23 // Given pattern: 24 // (x shiftopcode Q) shiftopcode K 25 // we should rewrite it as 26 // x shiftopcode (Q+K) iff (Q+K) u< bitwidth(x) and 27 // 28 // This is valid for any shift, but they must be identical, and we must be 29 // careful in case we have (zext(Q)+zext(K)) and look past extensions, 30 // (Q+K) must not overflow or else (Q+K) u< bitwidth(x) is bogus. 31 // 32 // AnalyzeForSignBitExtraction indicates that we will only analyze whether this 33 // pattern has any 2 right-shifts that sum to 1 less than original bit width. 34 Value *InstCombiner::reassociateShiftAmtsOfTwoSameDirectionShifts( 35 BinaryOperator *Sh0, const SimplifyQuery &SQ, 36 bool AnalyzeForSignBitExtraction) { 37 // Look for a shift of some instruction, ignore zext of shift amount if any. 38 Instruction *Sh0Op0; 39 Value *ShAmt0; 40 if (!match(Sh0, 41 m_Shift(m_Instruction(Sh0Op0), m_ZExtOrSelf(m_Value(ShAmt0))))) 42 return nullptr; 43 44 // If there is a truncation between the two shifts, we must make note of it 45 // and look through it. The truncation imposes additional constraints on the 46 // transform. 47 Instruction *Sh1; 48 Value *Trunc = nullptr; 49 match(Sh0Op0, 50 m_CombineOr(m_CombineAnd(m_Trunc(m_Instruction(Sh1)), m_Value(Trunc)), 51 m_Instruction(Sh1))); 52 53 // Inner shift: (x shiftopcode ShAmt1) 54 // Like with other shift, ignore zext of shift amount if any. 55 Value *X, *ShAmt1; 56 if (!match(Sh1, m_Shift(m_Value(X), m_ZExtOrSelf(m_Value(ShAmt1))))) 57 return nullptr; 58 59 // We have two shift amounts from two different shifts. The types of those 60 // shift amounts may not match. If that's the case let's bailout now.. 61 if (ShAmt0->getType() != ShAmt1->getType()) 62 return nullptr; 63 64 // As input, we have the following pattern: 65 // Sh0 (Sh1 X, Q), K 66 // We want to rewrite that as: 67 // Sh x, (Q+K) iff (Q+K) u< bitwidth(x) 68 // While we know that originally (Q+K) would not overflow 69 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of 70 // shift amounts. so it may now overflow in smaller bitwidth. 71 // To ensure that does not happen, we need to ensure that the total maximal 72 // shift amount is still representable in that smaller bit width. 73 unsigned MaximalPossibleTotalShiftAmount = 74 (Sh0->getType()->getScalarSizeInBits() - 1) + 75 (Sh1->getType()->getScalarSizeInBits() - 1); 76 APInt MaximalRepresentableShiftAmount = 77 APInt::getAllOnesValue(ShAmt0->getType()->getScalarSizeInBits()); 78 if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount)) 79 return nullptr; 80 81 // We are only looking for signbit extraction if we have two right shifts. 82 bool HadTwoRightShifts = match(Sh0, m_Shr(m_Value(), m_Value())) && 83 match(Sh1, m_Shr(m_Value(), m_Value())); 84 // ... and if it's not two right-shifts, we know the answer already. 85 if (AnalyzeForSignBitExtraction && !HadTwoRightShifts) 86 return nullptr; 87 88 // The shift opcodes must be identical, unless we are just checking whether 89 // this pattern can be interpreted as a sign-bit-extraction. 90 Instruction::BinaryOps ShiftOpcode = Sh0->getOpcode(); 91 bool IdenticalShOpcodes = Sh0->getOpcode() == Sh1->getOpcode(); 92 if (!IdenticalShOpcodes && !AnalyzeForSignBitExtraction) 93 return nullptr; 94 95 // If we saw truncation, we'll need to produce extra instruction, 96 // and for that one of the operands of the shift must be one-use, 97 // unless of course we don't actually plan to produce any instructions here. 98 if (Trunc && !AnalyzeForSignBitExtraction && 99 !match(Sh0, m_c_BinOp(m_OneUse(m_Value()), m_Value()))) 100 return nullptr; 101 102 // Can we fold (ShAmt0+ShAmt1) ? 103 auto *NewShAmt = dyn_cast_or_null<Constant>( 104 SimplifyAddInst(ShAmt0, ShAmt1, /*isNSW=*/false, /*isNUW=*/false, 105 SQ.getWithInstruction(Sh0))); 106 if (!NewShAmt) 107 return nullptr; // Did not simplify. 108 unsigned NewShAmtBitWidth = NewShAmt->getType()->getScalarSizeInBits(); 109 unsigned XBitWidth = X->getType()->getScalarSizeInBits(); 110 // Is the new shift amount smaller than the bit width of inner/new shift? 111 if (!match(NewShAmt, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT, 112 APInt(NewShAmtBitWidth, XBitWidth)))) 113 return nullptr; // FIXME: could perform constant-folding. 114 115 // If there was a truncation, and we have a right-shift, we can only fold if 116 // we are left with the original sign bit. Likewise, if we were just checking 117 // that this is a sighbit extraction, this is the place to check it. 118 // FIXME: zero shift amount is also legal here, but we can't *easily* check 119 // more than one predicate so it's not really worth it. 120 if (HadTwoRightShifts && (Trunc || AnalyzeForSignBitExtraction)) { 121 // If it's not a sign bit extraction, then we're done. 122 if (!match(NewShAmt, 123 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ, 124 APInt(NewShAmtBitWidth, XBitWidth - 1)))) 125 return nullptr; 126 // If it is, and that was the question, return the base value. 127 if (AnalyzeForSignBitExtraction) 128 return X; 129 } 130 131 assert(IdenticalShOpcodes && "Should not get here with different shifts."); 132 133 // All good, we can do this fold. 134 NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, X->getType()); 135 136 BinaryOperator *NewShift = BinaryOperator::Create(ShiftOpcode, X, NewShAmt); 137 138 // The flags can only be propagated if there wasn't a trunc. 139 if (!Trunc) { 140 // If the pattern did not involve trunc, and both of the original shifts 141 // had the same flag set, preserve the flag. 142 if (ShiftOpcode == Instruction::BinaryOps::Shl) { 143 NewShift->setHasNoUnsignedWrap(Sh0->hasNoUnsignedWrap() && 144 Sh1->hasNoUnsignedWrap()); 145 NewShift->setHasNoSignedWrap(Sh0->hasNoSignedWrap() && 146 Sh1->hasNoSignedWrap()); 147 } else { 148 NewShift->setIsExact(Sh0->isExact() && Sh1->isExact()); 149 } 150 } 151 152 Instruction *Ret = NewShift; 153 if (Trunc) { 154 Builder.Insert(NewShift); 155 Ret = CastInst::Create(Instruction::Trunc, NewShift, Sh0->getType()); 156 } 157 158 return Ret; 159 } 160 161 // If we have some pattern that leaves only some low bits set, and then performs 162 // left-shift of those bits, if none of the bits that are left after the final 163 // shift are modified by the mask, we can omit the mask. 164 // 165 // There are many variants to this pattern: 166 // a) (x & ((1 << MaskShAmt) - 1)) << ShiftShAmt 167 // b) (x & (~(-1 << MaskShAmt))) << ShiftShAmt 168 // c) (x & (-1 >> MaskShAmt)) << ShiftShAmt 169 // d) (x & ((-1 << MaskShAmt) >> MaskShAmt)) << ShiftShAmt 170 // e) ((x << MaskShAmt) l>> MaskShAmt) << ShiftShAmt 171 // f) ((x << MaskShAmt) a>> MaskShAmt) << ShiftShAmt 172 // All these patterns can be simplified to just: 173 // x << ShiftShAmt 174 // iff: 175 // a,b) (MaskShAmt+ShiftShAmt) u>= bitwidth(x) 176 // c,d,e,f) (ShiftShAmt-MaskShAmt) s>= 0 (i.e. ShiftShAmt u>= MaskShAmt) 177 static Instruction * 178 dropRedundantMaskingOfLeftShiftInput(BinaryOperator *OuterShift, 179 const SimplifyQuery &Q, 180 InstCombiner::BuilderTy &Builder) { 181 assert(OuterShift->getOpcode() == Instruction::BinaryOps::Shl && 182 "The input must be 'shl'!"); 183 184 Value *Masked, *ShiftShAmt; 185 match(OuterShift, 186 m_Shift(m_Value(Masked), m_ZExtOrSelf(m_Value(ShiftShAmt)))); 187 188 // *If* there is a truncation between an outer shift and a possibly-mask, 189 // then said truncation *must* be one-use, else we can't perform the fold. 190 Value *Trunc; 191 if (match(Masked, m_CombineAnd(m_Trunc(m_Value(Masked)), m_Value(Trunc))) && 192 !Trunc->hasOneUse()) 193 return nullptr; 194 195 Type *NarrowestTy = OuterShift->getType(); 196 Type *WidestTy = Masked->getType(); 197 bool HadTrunc = WidestTy != NarrowestTy; 198 199 // The mask must be computed in a type twice as wide to ensure 200 // that no bits are lost if the sum-of-shifts is wider than the base type. 201 Type *ExtendedTy = WidestTy->getExtendedType(); 202 203 Value *MaskShAmt; 204 205 // ((1 << MaskShAmt) - 1) 206 auto MaskA = m_Add(m_Shl(m_One(), m_Value(MaskShAmt)), m_AllOnes()); 207 // (~(-1 << maskNbits)) 208 auto MaskB = m_Xor(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_AllOnes()); 209 // (-1 >> MaskShAmt) 210 auto MaskC = m_Shr(m_AllOnes(), m_Value(MaskShAmt)); 211 // ((-1 << MaskShAmt) >> MaskShAmt) 212 auto MaskD = 213 m_Shr(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_Deferred(MaskShAmt)); 214 215 Value *X; 216 Constant *NewMask; 217 218 if (match(Masked, m_c_And(m_CombineOr(MaskA, MaskB), m_Value(X)))) { 219 // Peek through an optional zext of the shift amount. 220 match(MaskShAmt, m_ZExtOrSelf(m_Value(MaskShAmt))); 221 222 // We have two shift amounts from two different shifts. The types of those 223 // shift amounts may not match. If that's the case let's bailout now. 224 if (MaskShAmt->getType() != ShiftShAmt->getType()) 225 return nullptr; 226 227 // Can we simplify (MaskShAmt+ShiftShAmt) ? 228 auto *SumOfShAmts = dyn_cast_or_null<Constant>(SimplifyAddInst( 229 MaskShAmt, ShiftShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q)); 230 if (!SumOfShAmts) 231 return nullptr; // Did not simplify. 232 // In this pattern SumOfShAmts correlates with the number of low bits 233 // that shall remain in the root value (OuterShift). 234 235 // An extend of an undef value becomes zero because the high bits are never 236 // completely unknown. Replace the the `undef` shift amounts with final 237 // shift bitwidth to ensure that the value remains undef when creating the 238 // subsequent shift op. 239 SumOfShAmts = Constant::replaceUndefsWith( 240 SumOfShAmts, ConstantInt::get(SumOfShAmts->getType()->getScalarType(), 241 ExtendedTy->getScalarSizeInBits())); 242 auto *ExtendedSumOfShAmts = ConstantExpr::getZExt(SumOfShAmts, ExtendedTy); 243 // And compute the mask as usual: ~(-1 << (SumOfShAmts)) 244 auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(ExtendedTy); 245 auto *ExtendedInvertedMask = 246 ConstantExpr::getShl(ExtendedAllOnes, ExtendedSumOfShAmts); 247 NewMask = ConstantExpr::getNot(ExtendedInvertedMask); 248 } else if (match(Masked, m_c_And(m_CombineOr(MaskC, MaskD), m_Value(X))) || 249 match(Masked, m_Shr(m_Shl(m_Value(X), m_Value(MaskShAmt)), 250 m_Deferred(MaskShAmt)))) { 251 // Peek through an optional zext of the shift amount. 252 match(MaskShAmt, m_ZExtOrSelf(m_Value(MaskShAmt))); 253 254 // We have two shift amounts from two different shifts. The types of those 255 // shift amounts may not match. If that's the case let's bailout now. 256 if (MaskShAmt->getType() != ShiftShAmt->getType()) 257 return nullptr; 258 259 // Can we simplify (ShiftShAmt-MaskShAmt) ? 260 auto *ShAmtsDiff = dyn_cast_or_null<Constant>(SimplifySubInst( 261 ShiftShAmt, MaskShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q)); 262 if (!ShAmtsDiff) 263 return nullptr; // Did not simplify. 264 // In this pattern ShAmtsDiff correlates with the number of high bits that 265 // shall be unset in the root value (OuterShift). 266 267 // An extend of an undef value becomes zero because the high bits are never 268 // completely unknown. Replace the the `undef` shift amounts with negated 269 // bitwidth of innermost shift to ensure that the value remains undef when 270 // creating the subsequent shift op. 271 unsigned WidestTyBitWidth = WidestTy->getScalarSizeInBits(); 272 ShAmtsDiff = Constant::replaceUndefsWith( 273 ShAmtsDiff, ConstantInt::get(ShAmtsDiff->getType()->getScalarType(), 274 -WidestTyBitWidth)); 275 auto *ExtendedNumHighBitsToClear = ConstantExpr::getZExt( 276 ConstantExpr::getSub(ConstantInt::get(ShAmtsDiff->getType(), 277 WidestTyBitWidth, 278 /*isSigned=*/false), 279 ShAmtsDiff), 280 ExtendedTy); 281 // And compute the mask as usual: (-1 l>> (NumHighBitsToClear)) 282 auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(ExtendedTy); 283 NewMask = 284 ConstantExpr::getLShr(ExtendedAllOnes, ExtendedNumHighBitsToClear); 285 } else 286 return nullptr; // Don't know anything about this pattern. 287 288 NewMask = ConstantExpr::getTrunc(NewMask, NarrowestTy); 289 290 // Does this mask has any unset bits? If not then we can just not apply it. 291 bool NeedMask = !match(NewMask, m_AllOnes()); 292 293 // If we need to apply a mask, there are several more restrictions we have. 294 if (NeedMask) { 295 // The old masking instruction must go away. 296 if (!Masked->hasOneUse()) 297 return nullptr; 298 // The original "masking" instruction must not have been`ashr`. 299 if (match(Masked, m_AShr(m_Value(), m_Value()))) 300 return nullptr; 301 } 302 303 // If we need to apply truncation, let's do it first, since we can. 304 // We have already ensured that the old truncation will go away. 305 if (HadTrunc) 306 X = Builder.CreateTrunc(X, NarrowestTy); 307 308 // No 'NUW'/'NSW'! We no longer know that we won't shift-out non-0 bits. 309 // We didn't change the Type of this outermost shift, so we can just do it. 310 auto *NewShift = BinaryOperator::Create(OuterShift->getOpcode(), X, 311 OuterShift->getOperand(1)); 312 if (!NeedMask) 313 return NewShift; 314 315 Builder.Insert(NewShift); 316 return BinaryOperator::Create(Instruction::And, NewShift, NewMask); 317 } 318 319 /// If we have a shift-by-constant of a bitwise logic op that itself has a 320 /// shift-by-constant operand with identical opcode, we may be able to convert 321 /// that into 2 independent shifts followed by the logic op. This eliminates a 322 /// a use of an intermediate value (reduces dependency chain). 323 static Instruction *foldShiftOfShiftedLogic(BinaryOperator &I, 324 InstCombiner::BuilderTy &Builder) { 325 assert(I.isShift() && "Expected a shift as input"); 326 auto *LogicInst = dyn_cast<BinaryOperator>(I.getOperand(0)); 327 if (!LogicInst || !LogicInst->isBitwiseLogicOp() || !LogicInst->hasOneUse()) 328 return nullptr; 329 330 const APInt *C0, *C1; 331 if (!match(I.getOperand(1), m_APInt(C1))) 332 return nullptr; 333 334 Instruction::BinaryOps ShiftOpcode = I.getOpcode(); 335 Type *Ty = I.getType(); 336 337 // Find a matching one-use shift by constant. The fold is not valid if the sum 338 // of the shift values equals or exceeds bitwidth. 339 // TODO: Remove the one-use check if the other logic operand (Y) is constant. 340 Value *X, *Y; 341 auto matchFirstShift = [&](Value *V) { 342 return !isa<ConstantExpr>(V) && 343 match(V, m_OneUse(m_Shift(m_Value(X), m_APInt(C0)))) && 344 cast<BinaryOperator>(V)->getOpcode() == ShiftOpcode && 345 (*C0 + *C1).ult(Ty->getScalarSizeInBits()); 346 }; 347 348 // Logic ops are commutative, so check each operand for a match. 349 if (matchFirstShift(LogicInst->getOperand(0))) 350 Y = LogicInst->getOperand(1); 351 else if (matchFirstShift(LogicInst->getOperand(1))) 352 Y = LogicInst->getOperand(0); 353 else 354 return nullptr; 355 356 // shift (logic (shift X, C0), Y), C1 -> logic (shift X, C0+C1), (shift Y, C1) 357 Constant *ShiftSumC = ConstantInt::get(Ty, *C0 + *C1); 358 Value *NewShift1 = Builder.CreateBinOp(ShiftOpcode, X, ShiftSumC); 359 Value *NewShift2 = Builder.CreateBinOp(ShiftOpcode, Y, I.getOperand(1)); 360 return BinaryOperator::Create(LogicInst->getOpcode(), NewShift1, NewShift2); 361 } 362 363 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) { 364 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 365 assert(Op0->getType() == Op1->getType()); 366 367 // If the shift amount is a one-use `sext`, we can demote it to `zext`. 368 Value *Y; 369 if (match(Op1, m_OneUse(m_SExt(m_Value(Y))))) { 370 Value *NewExt = Builder.CreateZExt(Y, I.getType(), Op1->getName()); 371 return BinaryOperator::Create(I.getOpcode(), Op0, NewExt); 372 } 373 374 // See if we can fold away this shift. 375 if (SimplifyDemandedInstructionBits(I)) 376 return &I; 377 378 // Try to fold constant and into select arguments. 379 if (isa<Constant>(Op0)) 380 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) 381 if (Instruction *R = FoldOpIntoSelect(I, SI)) 382 return R; 383 384 if (Constant *CUI = dyn_cast<Constant>(Op1)) 385 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I)) 386 return Res; 387 388 if (auto *NewShift = cast_or_null<Instruction>( 389 reassociateShiftAmtsOfTwoSameDirectionShifts(&I, SQ))) 390 return NewShift; 391 392 // (C1 shift (A add C2)) -> (C1 shift C2) shift A) 393 // iff A and C2 are both positive. 394 Value *A; 395 Constant *C; 396 if (match(Op0, m_Constant()) && match(Op1, m_Add(m_Value(A), m_Constant(C)))) 397 if (isKnownNonNegative(A, DL, 0, &AC, &I, &DT) && 398 isKnownNonNegative(C, DL, 0, &AC, &I, &DT)) 399 return BinaryOperator::Create( 400 I.getOpcode(), Builder.CreateBinOp(I.getOpcode(), Op0, C), A); 401 402 // X shift (A srem B) -> X shift (A and B-1) iff B is a power of 2. 403 // Because shifts by negative values (which could occur if A were negative) 404 // are undefined. 405 const APInt *B; 406 if (Op1->hasOneUse() && match(Op1, m_SRem(m_Value(A), m_Power2(B)))) { 407 // FIXME: Should this get moved into SimplifyDemandedBits by saying we don't 408 // demand the sign bit (and many others) here?? 409 Value *Rem = Builder.CreateAnd(A, ConstantInt::get(I.getType(), *B - 1), 410 Op1->getName()); 411 return replaceOperand(I, 1, Rem); 412 } 413 414 if (Instruction *Logic = foldShiftOfShiftedLogic(I, Builder)) 415 return Logic; 416 417 return nullptr; 418 } 419 420 /// Return true if we can simplify two logical (either left or right) shifts 421 /// that have constant shift amounts: OuterShift (InnerShift X, C1), C2. 422 static bool canEvaluateShiftedShift(unsigned OuterShAmt, bool IsOuterShl, 423 Instruction *InnerShift, InstCombiner &IC, 424 Instruction *CxtI) { 425 assert(InnerShift->isLogicalShift() && "Unexpected instruction type"); 426 427 // We need constant scalar or constant splat shifts. 428 const APInt *InnerShiftConst; 429 if (!match(InnerShift->getOperand(1), m_APInt(InnerShiftConst))) 430 return false; 431 432 // Two logical shifts in the same direction: 433 // shl (shl X, C1), C2 --> shl X, C1 + C2 434 // lshr (lshr X, C1), C2 --> lshr X, C1 + C2 435 bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl; 436 if (IsInnerShl == IsOuterShl) 437 return true; 438 439 // Equal shift amounts in opposite directions become bitwise 'and': 440 // lshr (shl X, C), C --> and X, C' 441 // shl (lshr X, C), C --> and X, C' 442 if (*InnerShiftConst == OuterShAmt) 443 return true; 444 445 // If the 2nd shift is bigger than the 1st, we can fold: 446 // lshr (shl X, C1), C2 --> and (shl X, C1 - C2), C3 447 // shl (lshr X, C1), C2 --> and (lshr X, C1 - C2), C3 448 // but it isn't profitable unless we know the and'd out bits are already zero. 449 // Also, check that the inner shift is valid (less than the type width) or 450 // we'll crash trying to produce the bit mask for the 'and'. 451 unsigned TypeWidth = InnerShift->getType()->getScalarSizeInBits(); 452 if (InnerShiftConst->ugt(OuterShAmt) && InnerShiftConst->ult(TypeWidth)) { 453 unsigned InnerShAmt = InnerShiftConst->getZExtValue(); 454 unsigned MaskShift = 455 IsInnerShl ? TypeWidth - InnerShAmt : InnerShAmt - OuterShAmt; 456 APInt Mask = APInt::getLowBitsSet(TypeWidth, OuterShAmt) << MaskShift; 457 if (IC.MaskedValueIsZero(InnerShift->getOperand(0), Mask, 0, CxtI)) 458 return true; 459 } 460 461 return false; 462 } 463 464 /// See if we can compute the specified value, but shifted logically to the left 465 /// or right by some number of bits. This should return true if the expression 466 /// can be computed for the same cost as the current expression tree. This is 467 /// used to eliminate extraneous shifting from things like: 468 /// %C = shl i128 %A, 64 469 /// %D = shl i128 %B, 96 470 /// %E = or i128 %C, %D 471 /// %F = lshr i128 %E, 64 472 /// where the client will ask if E can be computed shifted right by 64-bits. If 473 /// this succeeds, getShiftedValue() will be called to produce the value. 474 static bool canEvaluateShifted(Value *V, unsigned NumBits, bool IsLeftShift, 475 InstCombiner &IC, Instruction *CxtI) { 476 // We can always evaluate constants shifted. 477 if (isa<Constant>(V)) 478 return true; 479 480 Instruction *I = dyn_cast<Instruction>(V); 481 if (!I) return false; 482 483 // If this is the opposite shift, we can directly reuse the input of the shift 484 // if the needed bits are already zero in the input. This allows us to reuse 485 // the value which means that we don't care if the shift has multiple uses. 486 // TODO: Handle opposite shift by exact value. 487 ConstantInt *CI = nullptr; 488 if ((IsLeftShift && match(I, m_LShr(m_Value(), m_ConstantInt(CI)))) || 489 (!IsLeftShift && match(I, m_Shl(m_Value(), m_ConstantInt(CI))))) { 490 if (CI->getValue() == NumBits) { 491 // TODO: Check that the input bits are already zero with MaskedValueIsZero 492 #if 0 493 // If this is a truncate of a logical shr, we can truncate it to a smaller 494 // lshr iff we know that the bits we would otherwise be shifting in are 495 // already zeros. 496 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 497 uint32_t BitWidth = Ty->getScalarSizeInBits(); 498 if (MaskedValueIsZero(I->getOperand(0), 499 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && 500 CI->getLimitedValue(BitWidth) < BitWidth) { 501 return CanEvaluateTruncated(I->getOperand(0), Ty); 502 } 503 #endif 504 505 } 506 } 507 508 // We can't mutate something that has multiple uses: doing so would 509 // require duplicating the instruction in general, which isn't profitable. 510 if (!I->hasOneUse()) return false; 511 512 switch (I->getOpcode()) { 513 default: return false; 514 case Instruction::And: 515 case Instruction::Or: 516 case Instruction::Xor: 517 // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted. 518 return canEvaluateShifted(I->getOperand(0), NumBits, IsLeftShift, IC, I) && 519 canEvaluateShifted(I->getOperand(1), NumBits, IsLeftShift, IC, I); 520 521 case Instruction::Shl: 522 case Instruction::LShr: 523 return canEvaluateShiftedShift(NumBits, IsLeftShift, I, IC, CxtI); 524 525 case Instruction::Select: { 526 SelectInst *SI = cast<SelectInst>(I); 527 Value *TrueVal = SI->getTrueValue(); 528 Value *FalseVal = SI->getFalseValue(); 529 return canEvaluateShifted(TrueVal, NumBits, IsLeftShift, IC, SI) && 530 canEvaluateShifted(FalseVal, NumBits, IsLeftShift, IC, SI); 531 } 532 case Instruction::PHI: { 533 // We can change a phi if we can change all operands. Note that we never 534 // get into trouble with cyclic PHIs here because we only consider 535 // instructions with a single use. 536 PHINode *PN = cast<PHINode>(I); 537 for (Value *IncValue : PN->incoming_values()) 538 if (!canEvaluateShifted(IncValue, NumBits, IsLeftShift, IC, PN)) 539 return false; 540 return true; 541 } 542 } 543 } 544 545 /// Fold OuterShift (InnerShift X, C1), C2. 546 /// See canEvaluateShiftedShift() for the constraints on these instructions. 547 static Value *foldShiftedShift(BinaryOperator *InnerShift, unsigned OuterShAmt, 548 bool IsOuterShl, 549 InstCombiner::BuilderTy &Builder) { 550 bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl; 551 Type *ShType = InnerShift->getType(); 552 unsigned TypeWidth = ShType->getScalarSizeInBits(); 553 554 // We only accept shifts-by-a-constant in canEvaluateShifted(). 555 const APInt *C1; 556 match(InnerShift->getOperand(1), m_APInt(C1)); 557 unsigned InnerShAmt = C1->getZExtValue(); 558 559 // Change the shift amount and clear the appropriate IR flags. 560 auto NewInnerShift = [&](unsigned ShAmt) { 561 InnerShift->setOperand(1, ConstantInt::get(ShType, ShAmt)); 562 if (IsInnerShl) { 563 InnerShift->setHasNoUnsignedWrap(false); 564 InnerShift->setHasNoSignedWrap(false); 565 } else { 566 InnerShift->setIsExact(false); 567 } 568 return InnerShift; 569 }; 570 571 // Two logical shifts in the same direction: 572 // shl (shl X, C1), C2 --> shl X, C1 + C2 573 // lshr (lshr X, C1), C2 --> lshr X, C1 + C2 574 if (IsInnerShl == IsOuterShl) { 575 // If this is an oversized composite shift, then unsigned shifts get 0. 576 if (InnerShAmt + OuterShAmt >= TypeWidth) 577 return Constant::getNullValue(ShType); 578 579 return NewInnerShift(InnerShAmt + OuterShAmt); 580 } 581 582 // Equal shift amounts in opposite directions become bitwise 'and': 583 // lshr (shl X, C), C --> and X, C' 584 // shl (lshr X, C), C --> and X, C' 585 if (InnerShAmt == OuterShAmt) { 586 APInt Mask = IsInnerShl 587 ? APInt::getLowBitsSet(TypeWidth, TypeWidth - OuterShAmt) 588 : APInt::getHighBitsSet(TypeWidth, TypeWidth - OuterShAmt); 589 Value *And = Builder.CreateAnd(InnerShift->getOperand(0), 590 ConstantInt::get(ShType, Mask)); 591 if (auto *AndI = dyn_cast<Instruction>(And)) { 592 AndI->moveBefore(InnerShift); 593 AndI->takeName(InnerShift); 594 } 595 return And; 596 } 597 598 assert(InnerShAmt > OuterShAmt && 599 "Unexpected opposite direction logical shift pair"); 600 601 // In general, we would need an 'and' for this transform, but 602 // canEvaluateShiftedShift() guarantees that the masked-off bits are not used. 603 // lshr (shl X, C1), C2 --> shl X, C1 - C2 604 // shl (lshr X, C1), C2 --> lshr X, C1 - C2 605 return NewInnerShift(InnerShAmt - OuterShAmt); 606 } 607 608 /// When canEvaluateShifted() returns true for an expression, this function 609 /// inserts the new computation that produces the shifted value. 610 static Value *getShiftedValue(Value *V, unsigned NumBits, bool isLeftShift, 611 InstCombiner &IC, const DataLayout &DL) { 612 // We can always evaluate constants shifted. 613 if (Constant *C = dyn_cast<Constant>(V)) { 614 if (isLeftShift) 615 return IC.Builder.CreateShl(C, NumBits); 616 else 617 return IC.Builder.CreateLShr(C, NumBits); 618 } 619 620 Instruction *I = cast<Instruction>(V); 621 IC.Worklist.push(I); 622 623 switch (I->getOpcode()) { 624 default: llvm_unreachable("Inconsistency with CanEvaluateShifted"); 625 case Instruction::And: 626 case Instruction::Or: 627 case Instruction::Xor: 628 // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted. 629 I->setOperand( 630 0, getShiftedValue(I->getOperand(0), NumBits, isLeftShift, IC, DL)); 631 I->setOperand( 632 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL)); 633 return I; 634 635 case Instruction::Shl: 636 case Instruction::LShr: 637 return foldShiftedShift(cast<BinaryOperator>(I), NumBits, isLeftShift, 638 IC.Builder); 639 640 case Instruction::Select: 641 I->setOperand( 642 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL)); 643 I->setOperand( 644 2, getShiftedValue(I->getOperand(2), NumBits, isLeftShift, IC, DL)); 645 return I; 646 case Instruction::PHI: { 647 // We can change a phi if we can change all operands. Note that we never 648 // get into trouble with cyclic PHIs here because we only consider 649 // instructions with a single use. 650 PHINode *PN = cast<PHINode>(I); 651 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 652 PN->setIncomingValue(i, getShiftedValue(PN->getIncomingValue(i), NumBits, 653 isLeftShift, IC, DL)); 654 return PN; 655 } 656 } 657 } 658 659 // If this is a bitwise operator or add with a constant RHS we might be able 660 // to pull it through a shift. 661 static bool canShiftBinOpWithConstantRHS(BinaryOperator &Shift, 662 BinaryOperator *BO) { 663 switch (BO->getOpcode()) { 664 default: 665 return false; // Do not perform transform! 666 case Instruction::Add: 667 return Shift.getOpcode() == Instruction::Shl; 668 case Instruction::Or: 669 case Instruction::Xor: 670 case Instruction::And: 671 return true; 672 } 673 } 674 675 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, Constant *Op1, 676 BinaryOperator &I) { 677 bool isLeftShift = I.getOpcode() == Instruction::Shl; 678 679 const APInt *Op1C; 680 if (!match(Op1, m_APInt(Op1C))) 681 return nullptr; 682 683 // See if we can propagate this shift into the input, this covers the trivial 684 // cast of lshr(shl(x,c1),c2) as well as other more complex cases. 685 if (I.getOpcode() != Instruction::AShr && 686 canEvaluateShifted(Op0, Op1C->getZExtValue(), isLeftShift, *this, &I)) { 687 LLVM_DEBUG( 688 dbgs() << "ICE: GetShiftedValue propagating shift through expression" 689 " to eliminate shift:\n IN: " 690 << *Op0 << "\n SH: " << I << "\n"); 691 692 return replaceInstUsesWith( 693 I, getShiftedValue(Op0, Op1C->getZExtValue(), isLeftShift, *this, DL)); 694 } 695 696 // See if we can simplify any instructions used by the instruction whose sole 697 // purpose is to compute bits we don't care about. 698 unsigned TypeBits = Op0->getType()->getScalarSizeInBits(); 699 700 assert(!Op1C->uge(TypeBits) && 701 "Shift over the type width should have been removed already"); 702 703 if (Instruction *FoldedShift = foldBinOpIntoSelectOrPhi(I)) 704 return FoldedShift; 705 706 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2)) 707 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) { 708 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0)); 709 // If 'shift2' is an ashr, we would have to get the sign bit into a funny 710 // place. Don't try to do this transformation in this case. Also, we 711 // require that the input operand is a shift-by-constant so that we have 712 // confidence that the shifts will get folded together. We could do this 713 // xform in more cases, but it is unlikely to be profitable. 714 if (TrOp && I.isLogicalShift() && TrOp->isShift() && 715 isa<ConstantInt>(TrOp->getOperand(1))) { 716 // Okay, we'll do this xform. Make the shift of shift. 717 Constant *ShAmt = 718 ConstantExpr::getZExt(cast<Constant>(Op1), TrOp->getType()); 719 // (shift2 (shift1 & 0x00FF), c2) 720 Value *NSh = Builder.CreateBinOp(I.getOpcode(), TrOp, ShAmt, I.getName()); 721 722 // For logical shifts, the truncation has the effect of making the high 723 // part of the register be zeros. Emulate this by inserting an AND to 724 // clear the top bits as needed. This 'and' will usually be zapped by 725 // other xforms later if dead. 726 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits(); 727 unsigned DstSize = TI->getType()->getScalarSizeInBits(); 728 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize)); 729 730 // The mask we constructed says what the trunc would do if occurring 731 // between the shifts. We want to know the effect *after* the second 732 // shift. We know that it is a logical shift by a constant, so adjust the 733 // mask as appropriate. 734 if (I.getOpcode() == Instruction::Shl) 735 MaskV <<= Op1C->getZExtValue(); 736 else { 737 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift"); 738 MaskV.lshrInPlace(Op1C->getZExtValue()); 739 } 740 741 // shift1 & 0x00FF 742 Value *And = Builder.CreateAnd(NSh, 743 ConstantInt::get(I.getContext(), MaskV), 744 TI->getName()); 745 746 // Return the value truncated to the interesting size. 747 return new TruncInst(And, I.getType()); 748 } 749 } 750 751 if (Op0->hasOneUse()) { 752 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) { 753 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C) 754 Value *V1, *V2; 755 ConstantInt *CC; 756 switch (Op0BO->getOpcode()) { 757 default: break; 758 case Instruction::Add: 759 case Instruction::And: 760 case Instruction::Or: 761 case Instruction::Xor: { 762 // These operators commute. 763 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C) 764 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() && 765 match(Op0BO->getOperand(1), m_Shr(m_Value(V1), 766 m_Specific(Op1)))) { 767 Value *YS = // (Y << C) 768 Builder.CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName()); 769 // (X + (Y << C)) 770 Value *X = Builder.CreateBinOp(Op0BO->getOpcode(), YS, V1, 771 Op0BO->getOperand(1)->getName()); 772 unsigned Op1Val = Op1C->getLimitedValue(TypeBits); 773 774 APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val); 775 Constant *Mask = ConstantInt::get(I.getContext(), Bits); 776 if (VectorType *VT = dyn_cast<VectorType>(X->getType())) 777 Mask = ConstantVector::getSplat(VT->getElementCount(), Mask); 778 return BinaryOperator::CreateAnd(X, Mask); 779 } 780 781 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C)) 782 Value *Op0BOOp1 = Op0BO->getOperand(1); 783 if (isLeftShift && Op0BOOp1->hasOneUse() && 784 match(Op0BOOp1, 785 m_And(m_OneUse(m_Shr(m_Value(V1), m_Specific(Op1))), 786 m_ConstantInt(CC)))) { 787 Value *YS = // (Y << C) 788 Builder.CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName()); 789 // X & (CC << C) 790 Value *XM = Builder.CreateAnd(V1, ConstantExpr::getShl(CC, Op1), 791 V1->getName()+".mask"); 792 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM); 793 } 794 LLVM_FALLTHROUGH; 795 } 796 797 case Instruction::Sub: { 798 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C) 799 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() && 800 match(Op0BO->getOperand(0), m_Shr(m_Value(V1), 801 m_Specific(Op1)))) { 802 Value *YS = // (Y << C) 803 Builder.CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName()); 804 // (X + (Y << C)) 805 Value *X = Builder.CreateBinOp(Op0BO->getOpcode(), V1, YS, 806 Op0BO->getOperand(0)->getName()); 807 unsigned Op1Val = Op1C->getLimitedValue(TypeBits); 808 809 APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val); 810 Constant *Mask = ConstantInt::get(I.getContext(), Bits); 811 if (VectorType *VT = dyn_cast<VectorType>(X->getType())) 812 Mask = ConstantVector::getSplat(VT->getElementCount(), Mask); 813 return BinaryOperator::CreateAnd(X, Mask); 814 } 815 816 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C) 817 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() && 818 match(Op0BO->getOperand(0), 819 m_And(m_OneUse(m_Shr(m_Value(V1), m_Value(V2))), 820 m_ConstantInt(CC))) && V2 == Op1) { 821 Value *YS = // (Y << C) 822 Builder.CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName()); 823 // X & (CC << C) 824 Value *XM = Builder.CreateAnd(V1, ConstantExpr::getShl(CC, Op1), 825 V1->getName()+".mask"); 826 827 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS); 828 } 829 830 break; 831 } 832 } 833 834 835 // If the operand is a bitwise operator with a constant RHS, and the 836 // shift is the only use, we can pull it out of the shift. 837 const APInt *Op0C; 838 if (match(Op0BO->getOperand(1), m_APInt(Op0C))) { 839 if (canShiftBinOpWithConstantRHS(I, Op0BO)) { 840 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), 841 cast<Constant>(Op0BO->getOperand(1)), Op1); 842 843 Value *NewShift = 844 Builder.CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1); 845 NewShift->takeName(Op0BO); 846 847 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift, 848 NewRHS); 849 } 850 } 851 852 // If the operand is a subtract with a constant LHS, and the shift 853 // is the only use, we can pull it out of the shift. 854 // This folds (shl (sub C1, X), C2) -> (sub (C1 << C2), (shl X, C2)) 855 if (isLeftShift && Op0BO->getOpcode() == Instruction::Sub && 856 match(Op0BO->getOperand(0), m_APInt(Op0C))) { 857 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), 858 cast<Constant>(Op0BO->getOperand(0)), Op1); 859 860 Value *NewShift = Builder.CreateShl(Op0BO->getOperand(1), Op1); 861 NewShift->takeName(Op0BO); 862 863 return BinaryOperator::CreateSub(NewRHS, NewShift); 864 } 865 } 866 867 // If we have a select that conditionally executes some binary operator, 868 // see if we can pull it the select and operator through the shift. 869 // 870 // For example, turning: 871 // shl (select C, (add X, C1), X), C2 872 // Into: 873 // Y = shl X, C2 874 // select C, (add Y, C1 << C2), Y 875 Value *Cond; 876 BinaryOperator *TBO; 877 Value *FalseVal; 878 if (match(Op0, m_Select(m_Value(Cond), m_OneUse(m_BinOp(TBO)), 879 m_Value(FalseVal)))) { 880 const APInt *C; 881 if (!isa<Constant>(FalseVal) && TBO->getOperand(0) == FalseVal && 882 match(TBO->getOperand(1), m_APInt(C)) && 883 canShiftBinOpWithConstantRHS(I, TBO)) { 884 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), 885 cast<Constant>(TBO->getOperand(1)), Op1); 886 887 Value *NewShift = 888 Builder.CreateBinOp(I.getOpcode(), FalseVal, Op1); 889 Value *NewOp = Builder.CreateBinOp(TBO->getOpcode(), NewShift, 890 NewRHS); 891 return SelectInst::Create(Cond, NewOp, NewShift); 892 } 893 } 894 895 BinaryOperator *FBO; 896 Value *TrueVal; 897 if (match(Op0, m_Select(m_Value(Cond), m_Value(TrueVal), 898 m_OneUse(m_BinOp(FBO))))) { 899 const APInt *C; 900 if (!isa<Constant>(TrueVal) && FBO->getOperand(0) == TrueVal && 901 match(FBO->getOperand(1), m_APInt(C)) && 902 canShiftBinOpWithConstantRHS(I, FBO)) { 903 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), 904 cast<Constant>(FBO->getOperand(1)), Op1); 905 906 Value *NewShift = 907 Builder.CreateBinOp(I.getOpcode(), TrueVal, Op1); 908 Value *NewOp = Builder.CreateBinOp(FBO->getOpcode(), NewShift, 909 NewRHS); 910 return SelectInst::Create(Cond, NewShift, NewOp); 911 } 912 } 913 } 914 915 return nullptr; 916 } 917 918 Instruction *InstCombiner::visitShl(BinaryOperator &I) { 919 const SimplifyQuery Q = SQ.getWithInstruction(&I); 920 921 if (Value *V = SimplifyShlInst(I.getOperand(0), I.getOperand(1), 922 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(), Q)) 923 return replaceInstUsesWith(I, V); 924 925 if (Instruction *X = foldVectorBinop(I)) 926 return X; 927 928 if (Instruction *V = commonShiftTransforms(I)) 929 return V; 930 931 if (Instruction *V = dropRedundantMaskingOfLeftShiftInput(&I, Q, Builder)) 932 return V; 933 934 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 935 Type *Ty = I.getType(); 936 unsigned BitWidth = Ty->getScalarSizeInBits(); 937 938 const APInt *ShAmtAPInt; 939 if (match(Op1, m_APInt(ShAmtAPInt))) { 940 unsigned ShAmt = ShAmtAPInt->getZExtValue(); 941 942 // shl (zext X), ShAmt --> zext (shl X, ShAmt) 943 // This is only valid if X would have zeros shifted out. 944 Value *X; 945 if (match(Op0, m_OneUse(m_ZExt(m_Value(X))))) { 946 unsigned SrcWidth = X->getType()->getScalarSizeInBits(); 947 if (ShAmt < SrcWidth && 948 MaskedValueIsZero(X, APInt::getHighBitsSet(SrcWidth, ShAmt), 0, &I)) 949 return new ZExtInst(Builder.CreateShl(X, ShAmt), Ty); 950 } 951 952 // (X >> C) << C --> X & (-1 << C) 953 if (match(Op0, m_Shr(m_Value(X), m_Specific(Op1)))) { 954 APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)); 955 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask)); 956 } 957 958 // FIXME: we do not yet transform non-exact shr's. The backend (DAGCombine) 959 // needs a few fixes for the rotate pattern recognition first. 960 const APInt *ShOp1; 961 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_APInt(ShOp1))))) { 962 unsigned ShrAmt = ShOp1->getZExtValue(); 963 if (ShrAmt < ShAmt) { 964 // If C1 < C2: (X >>?,exact C1) << C2 --> X << (C2 - C1) 965 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShrAmt); 966 auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff); 967 NewShl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); 968 NewShl->setHasNoSignedWrap(I.hasNoSignedWrap()); 969 return NewShl; 970 } 971 if (ShrAmt > ShAmt) { 972 // If C1 > C2: (X >>?exact C1) << C2 --> X >>?exact (C1 - C2) 973 Constant *ShiftDiff = ConstantInt::get(Ty, ShrAmt - ShAmt); 974 auto *NewShr = BinaryOperator::Create( 975 cast<BinaryOperator>(Op0)->getOpcode(), X, ShiftDiff); 976 NewShr->setIsExact(true); 977 return NewShr; 978 } 979 } 980 981 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShOp1)))) { 982 unsigned AmtSum = ShAmt + ShOp1->getZExtValue(); 983 // Oversized shifts are simplified to zero in InstSimplify. 984 if (AmtSum < BitWidth) 985 // (X << C1) << C2 --> X << (C1 + C2) 986 return BinaryOperator::CreateShl(X, ConstantInt::get(Ty, AmtSum)); 987 } 988 989 // If the shifted-out value is known-zero, then this is a NUW shift. 990 if (!I.hasNoUnsignedWrap() && 991 MaskedValueIsZero(Op0, APInt::getHighBitsSet(BitWidth, ShAmt), 0, &I)) { 992 I.setHasNoUnsignedWrap(); 993 return &I; 994 } 995 996 // If the shifted-out value is all signbits, then this is a NSW shift. 997 if (!I.hasNoSignedWrap() && ComputeNumSignBits(Op0, 0, &I) > ShAmt) { 998 I.setHasNoSignedWrap(); 999 return &I; 1000 } 1001 } 1002 1003 // Transform (x >> y) << y to x & (-1 << y) 1004 // Valid for any type of right-shift. 1005 Value *X; 1006 if (match(Op0, m_OneUse(m_Shr(m_Value(X), m_Specific(Op1))))) { 1007 Constant *AllOnes = ConstantInt::getAllOnesValue(Ty); 1008 Value *Mask = Builder.CreateShl(AllOnes, Op1); 1009 return BinaryOperator::CreateAnd(Mask, X); 1010 } 1011 1012 Constant *C1; 1013 if (match(Op1, m_Constant(C1))) { 1014 Constant *C2; 1015 Value *X; 1016 // (C2 << X) << C1 --> (C2 << C1) << X 1017 if (match(Op0, m_OneUse(m_Shl(m_Constant(C2), m_Value(X))))) 1018 return BinaryOperator::CreateShl(ConstantExpr::getShl(C2, C1), X); 1019 1020 // (X * C2) << C1 --> X * (C2 << C1) 1021 if (match(Op0, m_Mul(m_Value(X), m_Constant(C2)))) 1022 return BinaryOperator::CreateMul(X, ConstantExpr::getShl(C2, C1)); 1023 1024 // shl (zext i1 X), C1 --> select (X, 1 << C1, 0) 1025 if (match(Op0, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) { 1026 auto *NewC = ConstantExpr::getShl(ConstantInt::get(Ty, 1), C1); 1027 return SelectInst::Create(X, NewC, ConstantInt::getNullValue(Ty)); 1028 } 1029 } 1030 1031 // (1 << (C - x)) -> ((1 << C) >> x) if C is bitwidth - 1 1032 if (match(Op0, m_One()) && 1033 match(Op1, m_Sub(m_SpecificInt(BitWidth - 1), m_Value(X)))) 1034 return BinaryOperator::CreateLShr( 1035 ConstantInt::get(Ty, APInt::getSignMask(BitWidth)), X); 1036 1037 return nullptr; 1038 } 1039 1040 Instruction *InstCombiner::visitLShr(BinaryOperator &I) { 1041 if (Value *V = SimplifyLShrInst(I.getOperand(0), I.getOperand(1), I.isExact(), 1042 SQ.getWithInstruction(&I))) 1043 return replaceInstUsesWith(I, V); 1044 1045 if (Instruction *X = foldVectorBinop(I)) 1046 return X; 1047 1048 if (Instruction *R = commonShiftTransforms(I)) 1049 return R; 1050 1051 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1052 Type *Ty = I.getType(); 1053 const APInt *ShAmtAPInt; 1054 if (match(Op1, m_APInt(ShAmtAPInt))) { 1055 unsigned ShAmt = ShAmtAPInt->getZExtValue(); 1056 unsigned BitWidth = Ty->getScalarSizeInBits(); 1057 auto *II = dyn_cast<IntrinsicInst>(Op0); 1058 if (II && isPowerOf2_32(BitWidth) && Log2_32(BitWidth) == ShAmt && 1059 (II->getIntrinsicID() == Intrinsic::ctlz || 1060 II->getIntrinsicID() == Intrinsic::cttz || 1061 II->getIntrinsicID() == Intrinsic::ctpop)) { 1062 // ctlz.i32(x)>>5 --> zext(x == 0) 1063 // cttz.i32(x)>>5 --> zext(x == 0) 1064 // ctpop.i32(x)>>5 --> zext(x == -1) 1065 bool IsPop = II->getIntrinsicID() == Intrinsic::ctpop; 1066 Constant *RHS = ConstantInt::getSigned(Ty, IsPop ? -1 : 0); 1067 Value *Cmp = Builder.CreateICmpEQ(II->getArgOperand(0), RHS); 1068 return new ZExtInst(Cmp, Ty); 1069 } 1070 1071 Value *X; 1072 const APInt *ShOp1; 1073 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShOp1))) && ShOp1->ult(BitWidth)) { 1074 if (ShOp1->ult(ShAmt)) { 1075 unsigned ShlAmt = ShOp1->getZExtValue(); 1076 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt); 1077 if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) { 1078 // (X <<nuw C1) >>u C2 --> X >>u (C2 - C1) 1079 auto *NewLShr = BinaryOperator::CreateLShr(X, ShiftDiff); 1080 NewLShr->setIsExact(I.isExact()); 1081 return NewLShr; 1082 } 1083 // (X << C1) >>u C2 --> (X >>u (C2 - C1)) & (-1 >> C2) 1084 Value *NewLShr = Builder.CreateLShr(X, ShiftDiff, "", I.isExact()); 1085 APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt)); 1086 return BinaryOperator::CreateAnd(NewLShr, ConstantInt::get(Ty, Mask)); 1087 } 1088 if (ShOp1->ugt(ShAmt)) { 1089 unsigned ShlAmt = ShOp1->getZExtValue(); 1090 Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt); 1091 if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) { 1092 // (X <<nuw C1) >>u C2 --> X <<nuw (C1 - C2) 1093 auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff); 1094 NewShl->setHasNoUnsignedWrap(true); 1095 return NewShl; 1096 } 1097 // (X << C1) >>u C2 --> X << (C1 - C2) & (-1 >> C2) 1098 Value *NewShl = Builder.CreateShl(X, ShiftDiff); 1099 APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt)); 1100 return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask)); 1101 } 1102 assert(*ShOp1 == ShAmt); 1103 // (X << C) >>u C --> X & (-1 >>u C) 1104 APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt)); 1105 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask)); 1106 } 1107 1108 if (match(Op0, m_OneUse(m_ZExt(m_Value(X)))) && 1109 (!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType()))) { 1110 assert(ShAmt < X->getType()->getScalarSizeInBits() && 1111 "Big shift not simplified to zero?"); 1112 // lshr (zext iM X to iN), C --> zext (lshr X, C) to iN 1113 Value *NewLShr = Builder.CreateLShr(X, ShAmt); 1114 return new ZExtInst(NewLShr, Ty); 1115 } 1116 1117 if (match(Op0, m_SExt(m_Value(X))) && 1118 (!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType()))) { 1119 // Are we moving the sign bit to the low bit and widening with high zeros? 1120 unsigned SrcTyBitWidth = X->getType()->getScalarSizeInBits(); 1121 if (ShAmt == BitWidth - 1) { 1122 // lshr (sext i1 X to iN), N-1 --> zext X to iN 1123 if (SrcTyBitWidth == 1) 1124 return new ZExtInst(X, Ty); 1125 1126 // lshr (sext iM X to iN), N-1 --> zext (lshr X, M-1) to iN 1127 if (Op0->hasOneUse()) { 1128 Value *NewLShr = Builder.CreateLShr(X, SrcTyBitWidth - 1); 1129 return new ZExtInst(NewLShr, Ty); 1130 } 1131 } 1132 1133 // lshr (sext iM X to iN), N-M --> zext (ashr X, min(N-M, M-1)) to iN 1134 if (ShAmt == BitWidth - SrcTyBitWidth && Op0->hasOneUse()) { 1135 // The new shift amount can't be more than the narrow source type. 1136 unsigned NewShAmt = std::min(ShAmt, SrcTyBitWidth - 1); 1137 Value *AShr = Builder.CreateAShr(X, NewShAmt); 1138 return new ZExtInst(AShr, Ty); 1139 } 1140 } 1141 1142 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShOp1)))) { 1143 unsigned AmtSum = ShAmt + ShOp1->getZExtValue(); 1144 // Oversized shifts are simplified to zero in InstSimplify. 1145 if (AmtSum < BitWidth) 1146 // (X >>u C1) >>u C2 --> X >>u (C1 + C2) 1147 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum)); 1148 } 1149 1150 // If the shifted-out value is known-zero, then this is an exact shift. 1151 if (!I.isExact() && 1152 MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmt), 0, &I)) { 1153 I.setIsExact(); 1154 return &I; 1155 } 1156 } 1157 1158 // Transform (x << y) >> y to x & (-1 >> y) 1159 Value *X; 1160 if (match(Op0, m_OneUse(m_Shl(m_Value(X), m_Specific(Op1))))) { 1161 Constant *AllOnes = ConstantInt::getAllOnesValue(Ty); 1162 Value *Mask = Builder.CreateLShr(AllOnes, Op1); 1163 return BinaryOperator::CreateAnd(Mask, X); 1164 } 1165 1166 return nullptr; 1167 } 1168 1169 Instruction * 1170 InstCombiner::foldVariableSignZeroExtensionOfVariableHighBitExtract( 1171 BinaryOperator &OldAShr) { 1172 assert(OldAShr.getOpcode() == Instruction::AShr && 1173 "Must be called with arithmetic right-shift instruction only."); 1174 1175 // Check that constant C is a splat of the element-wise bitwidth of V. 1176 auto BitWidthSplat = [](Constant *C, Value *V) { 1177 return match( 1178 C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ, 1179 APInt(C->getType()->getScalarSizeInBits(), 1180 V->getType()->getScalarSizeInBits()))); 1181 }; 1182 1183 // It should look like variable-length sign-extension on the outside: 1184 // (Val << (bitwidth(Val)-Nbits)) a>> (bitwidth(Val)-Nbits) 1185 Value *NBits; 1186 Instruction *MaybeTrunc; 1187 Constant *C1, *C2; 1188 if (!match(&OldAShr, 1189 m_AShr(m_Shl(m_Instruction(MaybeTrunc), 1190 m_ZExtOrSelf(m_Sub(m_Constant(C1), 1191 m_ZExtOrSelf(m_Value(NBits))))), 1192 m_ZExtOrSelf(m_Sub(m_Constant(C2), 1193 m_ZExtOrSelf(m_Deferred(NBits)))))) || 1194 !BitWidthSplat(C1, &OldAShr) || !BitWidthSplat(C2, &OldAShr)) 1195 return nullptr; 1196 1197 // There may or may not be a truncation after outer two shifts. 1198 Instruction *HighBitExtract; 1199 match(MaybeTrunc, m_TruncOrSelf(m_Instruction(HighBitExtract))); 1200 bool HadTrunc = MaybeTrunc != HighBitExtract; 1201 1202 // And finally, the innermost part of the pattern must be a right-shift. 1203 Value *X, *NumLowBitsToSkip; 1204 if (!match(HighBitExtract, m_Shr(m_Value(X), m_Value(NumLowBitsToSkip)))) 1205 return nullptr; 1206 1207 // Said right-shift must extract high NBits bits - C0 must be it's bitwidth. 1208 Constant *C0; 1209 if (!match(NumLowBitsToSkip, 1210 m_ZExtOrSelf( 1211 m_Sub(m_Constant(C0), m_ZExtOrSelf(m_Specific(NBits))))) || 1212 !BitWidthSplat(C0, HighBitExtract)) 1213 return nullptr; 1214 1215 // Since the NBits is identical for all shifts, if the outermost and 1216 // innermost shifts are identical, then outermost shifts are redundant. 1217 // If we had truncation, do keep it though. 1218 if (HighBitExtract->getOpcode() == OldAShr.getOpcode()) 1219 return replaceInstUsesWith(OldAShr, MaybeTrunc); 1220 1221 // Else, if there was a truncation, then we need to ensure that one 1222 // instruction will go away. 1223 if (HadTrunc && !match(&OldAShr, m_c_BinOp(m_OneUse(m_Value()), m_Value()))) 1224 return nullptr; 1225 1226 // Finally, bypass two innermost shifts, and perform the outermost shift on 1227 // the operands of the innermost shift. 1228 Instruction *NewAShr = 1229 BinaryOperator::Create(OldAShr.getOpcode(), X, NumLowBitsToSkip); 1230 NewAShr->copyIRFlags(HighBitExtract); // We can preserve 'exact'-ness. 1231 if (!HadTrunc) 1232 return NewAShr; 1233 1234 Builder.Insert(NewAShr); 1235 return TruncInst::CreateTruncOrBitCast(NewAShr, OldAShr.getType()); 1236 } 1237 1238 Instruction *InstCombiner::visitAShr(BinaryOperator &I) { 1239 if (Value *V = SimplifyAShrInst(I.getOperand(0), I.getOperand(1), I.isExact(), 1240 SQ.getWithInstruction(&I))) 1241 return replaceInstUsesWith(I, V); 1242 1243 if (Instruction *X = foldVectorBinop(I)) 1244 return X; 1245 1246 if (Instruction *R = commonShiftTransforms(I)) 1247 return R; 1248 1249 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1250 Type *Ty = I.getType(); 1251 unsigned BitWidth = Ty->getScalarSizeInBits(); 1252 const APInt *ShAmtAPInt; 1253 if (match(Op1, m_APInt(ShAmtAPInt)) && ShAmtAPInt->ult(BitWidth)) { 1254 unsigned ShAmt = ShAmtAPInt->getZExtValue(); 1255 1256 // If the shift amount equals the difference in width of the destination 1257 // and source scalar types: 1258 // ashr (shl (zext X), C), C --> sext X 1259 Value *X; 1260 if (match(Op0, m_Shl(m_ZExt(m_Value(X)), m_Specific(Op1))) && 1261 ShAmt == BitWidth - X->getType()->getScalarSizeInBits()) 1262 return new SExtInst(X, Ty); 1263 1264 // We can't handle (X << C1) >>s C2. It shifts arbitrary bits in. However, 1265 // we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits. 1266 const APInt *ShOp1; 1267 if (match(Op0, m_NSWShl(m_Value(X), m_APInt(ShOp1))) && 1268 ShOp1->ult(BitWidth)) { 1269 unsigned ShlAmt = ShOp1->getZExtValue(); 1270 if (ShlAmt < ShAmt) { 1271 // (X <<nsw C1) >>s C2 --> X >>s (C2 - C1) 1272 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt); 1273 auto *NewAShr = BinaryOperator::CreateAShr(X, ShiftDiff); 1274 NewAShr->setIsExact(I.isExact()); 1275 return NewAShr; 1276 } 1277 if (ShlAmt > ShAmt) { 1278 // (X <<nsw C1) >>s C2 --> X <<nsw (C1 - C2) 1279 Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt); 1280 auto *NewShl = BinaryOperator::Create(Instruction::Shl, X, ShiftDiff); 1281 NewShl->setHasNoSignedWrap(true); 1282 return NewShl; 1283 } 1284 } 1285 1286 if (match(Op0, m_AShr(m_Value(X), m_APInt(ShOp1))) && 1287 ShOp1->ult(BitWidth)) { 1288 unsigned AmtSum = ShAmt + ShOp1->getZExtValue(); 1289 // Oversized arithmetic shifts replicate the sign bit. 1290 AmtSum = std::min(AmtSum, BitWidth - 1); 1291 // (X >>s C1) >>s C2 --> X >>s (C1 + C2) 1292 return BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum)); 1293 } 1294 1295 if (match(Op0, m_OneUse(m_SExt(m_Value(X)))) && 1296 (Ty->isVectorTy() || shouldChangeType(Ty, X->getType()))) { 1297 // ashr (sext X), C --> sext (ashr X, C') 1298 Type *SrcTy = X->getType(); 1299 ShAmt = std::min(ShAmt, SrcTy->getScalarSizeInBits() - 1); 1300 Value *NewSh = Builder.CreateAShr(X, ConstantInt::get(SrcTy, ShAmt)); 1301 return new SExtInst(NewSh, Ty); 1302 } 1303 1304 // If the shifted-out value is known-zero, then this is an exact shift. 1305 if (!I.isExact() && 1306 MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmt), 0, &I)) { 1307 I.setIsExact(); 1308 return &I; 1309 } 1310 } 1311 1312 if (Instruction *R = foldVariableSignZeroExtensionOfVariableHighBitExtract(I)) 1313 return R; 1314 1315 // See if we can turn a signed shr into an unsigned shr. 1316 if (MaskedValueIsZero(Op0, APInt::getSignMask(BitWidth), 0, &I)) 1317 return BinaryOperator::CreateLShr(Op0, Op1); 1318 1319 return nullptr; 1320 } 1321