1 //===- InstCombineShifts.cpp ----------------------------------------------===//
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
7 //===----------------------------------------------------------------------===//
9 // This file implements the visitShl, visitLShr, and visitAShr functions.
11 //===----------------------------------------------------------------------===//
13 #include "InstCombineInternal.h"
14 #include "llvm/Analysis/InstructionSimplify.h"
15 #include "llvm/IR/IntrinsicInst.h"
16 #include "llvm/IR/PatternMatch.h"
17 #include "llvm/Transforms/InstCombine/InstCombiner.h"
19 using namespace PatternMatch;
21 #define DEBUG_TYPE "instcombine"
23 bool canTryToConstantAddTwoShiftAmounts(Value *Sh0, Value *ShAmt0, Value *Sh1,
25 // We have two shift amounts from two different shifts. The types of those
26 // shift amounts may not match. If that's the case let's bailout now..
27 if (ShAmt0->getType() != ShAmt1->getType())
30 // As input, we have the following pattern:
32 // We want to rewrite that as:
33 // Sh x, (Q+K) iff (Q+K) u< bitwidth(x)
34 // While we know that originally (Q+K) would not overflow
35 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
36 // shift amounts. so it may now overflow in smaller bitwidth.
37 // To ensure that does not happen, we need to ensure that the total maximal
38 // shift amount is still representable in that smaller bit width.
39 unsigned MaximalPossibleTotalShiftAmount =
40 (Sh0->getType()->getScalarSizeInBits() - 1) +
41 (Sh1->getType()->getScalarSizeInBits() - 1);
42 APInt MaximalRepresentableShiftAmount =
43 APInt::getAllOnes(ShAmt0->getType()->getScalarSizeInBits());
44 return MaximalRepresentableShiftAmount.uge(MaximalPossibleTotalShiftAmount);
48 // (x shiftopcode Q) shiftopcode K
49 // we should rewrite it as
50 // x shiftopcode (Q+K) iff (Q+K) u< bitwidth(x) and
52 // This is valid for any shift, but they must be identical, and we must be
53 // careful in case we have (zext(Q)+zext(K)) and look past extensions,
54 // (Q+K) must not overflow or else (Q+K) u< bitwidth(x) is bogus.
56 // AnalyzeForSignBitExtraction indicates that we will only analyze whether this
57 // pattern has any 2 right-shifts that sum to 1 less than original bit width.
58 Value *InstCombinerImpl::reassociateShiftAmtsOfTwoSameDirectionShifts(
59 BinaryOperator *Sh0, const SimplifyQuery &SQ,
60 bool AnalyzeForSignBitExtraction) {
61 // Look for a shift of some instruction, ignore zext of shift amount if any.
65 m_Shift(m_Instruction(Sh0Op0), m_ZExtOrSelf(m_Value(ShAmt0)))))
68 // If there is a truncation between the two shifts, we must make note of it
69 // and look through it. The truncation imposes additional constraints on the
72 Value *Trunc = nullptr;
74 m_CombineOr(m_CombineAnd(m_Trunc(m_Instruction(Sh1)), m_Value(Trunc)),
77 // Inner shift: (x shiftopcode ShAmt1)
78 // Like with other shift, ignore zext of shift amount if any.
80 if (!match(Sh1, m_Shift(m_Value(X), m_ZExtOrSelf(m_Value(ShAmt1)))))
83 // Verify that it would be safe to try to add those two shift amounts.
84 if (!canTryToConstantAddTwoShiftAmounts(Sh0, ShAmt0, Sh1, ShAmt1))
87 // We are only looking for signbit extraction if we have two right shifts.
88 bool HadTwoRightShifts = match(Sh0, m_Shr(m_Value(), m_Value())) &&
89 match(Sh1, m_Shr(m_Value(), m_Value()));
90 // ... and if it's not two right-shifts, we know the answer already.
91 if (AnalyzeForSignBitExtraction && !HadTwoRightShifts)
94 // The shift opcodes must be identical, unless we are just checking whether
95 // this pattern can be interpreted as a sign-bit-extraction.
96 Instruction::BinaryOps ShiftOpcode = Sh0->getOpcode();
97 bool IdenticalShOpcodes = Sh0->getOpcode() == Sh1->getOpcode();
98 if (!IdenticalShOpcodes && !AnalyzeForSignBitExtraction)
101 // If we saw truncation, we'll need to produce extra instruction,
102 // and for that one of the operands of the shift must be one-use,
103 // unless of course we don't actually plan to produce any instructions here.
104 if (Trunc && !AnalyzeForSignBitExtraction &&
105 !match(Sh0, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
108 // Can we fold (ShAmt0+ShAmt1) ?
109 auto *NewShAmt = dyn_cast_or_null<Constant>(
110 simplifyAddInst(ShAmt0, ShAmt1, /*isNSW=*/false, /*isNUW=*/false,
111 SQ.getWithInstruction(Sh0)));
113 return nullptr; // Did not simplify.
114 unsigned NewShAmtBitWidth = NewShAmt->getType()->getScalarSizeInBits();
115 unsigned XBitWidth = X->getType()->getScalarSizeInBits();
116 // Is the new shift amount smaller than the bit width of inner/new shift?
117 if (!match(NewShAmt, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
118 APInt(NewShAmtBitWidth, XBitWidth))))
119 return nullptr; // FIXME: could perform constant-folding.
121 // If there was a truncation, and we have a right-shift, we can only fold if
122 // we are left with the original sign bit. Likewise, if we were just checking
123 // that this is a sighbit extraction, this is the place to check it.
124 // FIXME: zero shift amount is also legal here, but we can't *easily* check
125 // more than one predicate so it's not really worth it.
126 if (HadTwoRightShifts && (Trunc || AnalyzeForSignBitExtraction)) {
127 // If it's not a sign bit extraction, then we're done.
129 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
130 APInt(NewShAmtBitWidth, XBitWidth - 1))))
132 // If it is, and that was the question, return the base value.
133 if (AnalyzeForSignBitExtraction)
137 assert(IdenticalShOpcodes && "Should not get here with different shifts.");
139 // All good, we can do this fold.
140 NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, X->getType());
142 BinaryOperator *NewShift = BinaryOperator::Create(ShiftOpcode, X, NewShAmt);
144 // The flags can only be propagated if there wasn't a trunc.
146 // If the pattern did not involve trunc, and both of the original shifts
147 // had the same flag set, preserve the flag.
148 if (ShiftOpcode == Instruction::BinaryOps::Shl) {
149 NewShift->setHasNoUnsignedWrap(Sh0->hasNoUnsignedWrap() &&
150 Sh1->hasNoUnsignedWrap());
151 NewShift->setHasNoSignedWrap(Sh0->hasNoSignedWrap() &&
152 Sh1->hasNoSignedWrap());
154 NewShift->setIsExact(Sh0->isExact() && Sh1->isExact());
158 Instruction *Ret = NewShift;
160 Builder.Insert(NewShift);
161 Ret = CastInst::Create(Instruction::Trunc, NewShift, Sh0->getType());
167 // If we have some pattern that leaves only some low bits set, and then performs
168 // left-shift of those bits, if none of the bits that are left after the final
169 // shift are modified by the mask, we can omit the mask.
171 // There are many variants to this pattern:
172 // a) (x & ((1 << MaskShAmt) - 1)) << ShiftShAmt
173 // b) (x & (~(-1 << MaskShAmt))) << ShiftShAmt
174 // c) (x & (-1 l>> MaskShAmt)) << ShiftShAmt
175 // d) (x & ((-1 << MaskShAmt) l>> MaskShAmt)) << ShiftShAmt
176 // e) ((x << MaskShAmt) l>> MaskShAmt) << ShiftShAmt
177 // f) ((x << MaskShAmt) a>> MaskShAmt) << ShiftShAmt
178 // All these patterns can be simplified to just:
181 // a,b) (MaskShAmt+ShiftShAmt) u>= bitwidth(x)
182 // c,d,e,f) (ShiftShAmt-MaskShAmt) s>= 0 (i.e. ShiftShAmt u>= MaskShAmt)
184 dropRedundantMaskingOfLeftShiftInput(BinaryOperator *OuterShift,
185 const SimplifyQuery &Q,
186 InstCombiner::BuilderTy &Builder) {
187 assert(OuterShift->getOpcode() == Instruction::BinaryOps::Shl &&
188 "The input must be 'shl'!");
190 Value *Masked, *ShiftShAmt;
192 m_Shift(m_Value(Masked), m_ZExtOrSelf(m_Value(ShiftShAmt))));
194 // *If* there is a truncation between an outer shift and a possibly-mask,
195 // then said truncation *must* be one-use, else we can't perform the fold.
197 if (match(Masked, m_CombineAnd(m_Trunc(m_Value(Masked)), m_Value(Trunc))) &&
201 Type *NarrowestTy = OuterShift->getType();
202 Type *WidestTy = Masked->getType();
203 bool HadTrunc = WidestTy != NarrowestTy;
205 // The mask must be computed in a type twice as wide to ensure
206 // that no bits are lost if the sum-of-shifts is wider than the base type.
207 Type *ExtendedTy = WidestTy->getExtendedType();
211 // ((1 << MaskShAmt) - 1)
212 auto MaskA = m_Add(m_Shl(m_One(), m_Value(MaskShAmt)), m_AllOnes());
213 // (~(-1 << maskNbits))
214 auto MaskB = m_Xor(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_AllOnes());
215 // (-1 l>> MaskShAmt)
216 auto MaskC = m_LShr(m_AllOnes(), m_Value(MaskShAmt));
217 // ((-1 << MaskShAmt) l>> MaskShAmt)
219 m_LShr(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_Deferred(MaskShAmt));
224 if (match(Masked, m_c_And(m_CombineOr(MaskA, MaskB), m_Value(X)))) {
225 // Peek through an optional zext of the shift amount.
226 match(MaskShAmt, m_ZExtOrSelf(m_Value(MaskShAmt)));
228 // Verify that it would be safe to try to add those two shift amounts.
229 if (!canTryToConstantAddTwoShiftAmounts(OuterShift, ShiftShAmt, Masked,
233 // Can we simplify (MaskShAmt+ShiftShAmt) ?
234 auto *SumOfShAmts = dyn_cast_or_null<Constant>(simplifyAddInst(
235 MaskShAmt, ShiftShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q));
237 return nullptr; // Did not simplify.
238 // In this pattern SumOfShAmts correlates with the number of low bits
239 // that shall remain in the root value (OuterShift).
241 // An extend of an undef value becomes zero because the high bits are never
242 // completely unknown. Replace the `undef` shift amounts with final
243 // shift bitwidth to ensure that the value remains undef when creating the
244 // subsequent shift op.
245 SumOfShAmts = Constant::replaceUndefsWith(
246 SumOfShAmts, ConstantInt::get(SumOfShAmts->getType()->getScalarType(),
247 ExtendedTy->getScalarSizeInBits()));
248 auto *ExtendedSumOfShAmts = ConstantExpr::getZExt(SumOfShAmts, ExtendedTy);
249 // And compute the mask as usual: ~(-1 << (SumOfShAmts))
250 auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(ExtendedTy);
251 auto *ExtendedInvertedMask =
252 ConstantExpr::getShl(ExtendedAllOnes, ExtendedSumOfShAmts);
253 NewMask = ConstantExpr::getNot(ExtendedInvertedMask);
254 } else if (match(Masked, m_c_And(m_CombineOr(MaskC, MaskD), m_Value(X))) ||
255 match(Masked, m_Shr(m_Shl(m_Value(X), m_Value(MaskShAmt)),
256 m_Deferred(MaskShAmt)))) {
257 // Peek through an optional zext of the shift amount.
258 match(MaskShAmt, m_ZExtOrSelf(m_Value(MaskShAmt)));
260 // Verify that it would be safe to try to add those two shift amounts.
261 if (!canTryToConstantAddTwoShiftAmounts(OuterShift, ShiftShAmt, Masked,
265 // Can we simplify (ShiftShAmt-MaskShAmt) ?
266 auto *ShAmtsDiff = dyn_cast_or_null<Constant>(simplifySubInst(
267 ShiftShAmt, MaskShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q));
269 return nullptr; // Did not simplify.
270 // In this pattern ShAmtsDiff correlates with the number of high bits that
271 // shall be unset in the root value (OuterShift).
273 // An extend of an undef value becomes zero because the high bits are never
274 // completely unknown. Replace the `undef` shift amounts with negated
275 // bitwidth of innermost shift to ensure that the value remains undef when
276 // creating the subsequent shift op.
277 unsigned WidestTyBitWidth = WidestTy->getScalarSizeInBits();
278 ShAmtsDiff = Constant::replaceUndefsWith(
279 ShAmtsDiff, ConstantInt::get(ShAmtsDiff->getType()->getScalarType(),
281 auto *ExtendedNumHighBitsToClear = ConstantExpr::getZExt(
282 ConstantExpr::getSub(ConstantInt::get(ShAmtsDiff->getType(),
287 // And compute the mask as usual: (-1 l>> (NumHighBitsToClear))
288 auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(ExtendedTy);
290 ConstantExpr::getLShr(ExtendedAllOnes, ExtendedNumHighBitsToClear);
292 return nullptr; // Don't know anything about this pattern.
294 NewMask = ConstantExpr::getTrunc(NewMask, NarrowestTy);
296 // Does this mask has any unset bits? If not then we can just not apply it.
297 bool NeedMask = !match(NewMask, m_AllOnes());
299 // If we need to apply a mask, there are several more restrictions we have.
301 // The old masking instruction must go away.
302 if (!Masked->hasOneUse())
304 // The original "masking" instruction must not have been`ashr`.
305 if (match(Masked, m_AShr(m_Value(), m_Value())))
309 // If we need to apply truncation, let's do it first, since we can.
310 // We have already ensured that the old truncation will go away.
312 X = Builder.CreateTrunc(X, NarrowestTy);
314 // No 'NUW'/'NSW'! We no longer know that we won't shift-out non-0 bits.
315 // We didn't change the Type of this outermost shift, so we can just do it.
316 auto *NewShift = BinaryOperator::Create(OuterShift->getOpcode(), X,
317 OuterShift->getOperand(1));
321 Builder.Insert(NewShift);
322 return BinaryOperator::Create(Instruction::And, NewShift, NewMask);
325 /// If we have a shift-by-constant of a bin op (bitwise logic op or add/sub w/
326 /// shl) that itself has a shift-by-constant operand with identical opcode, we
327 /// may be able to convert that into 2 independent shifts followed by the logic
328 /// op. This eliminates a use of an intermediate value (reduces dependency
330 static Instruction *foldShiftOfShiftedBinOp(BinaryOperator &I,
331 InstCombiner::BuilderTy &Builder) {
332 assert(I.isShift() && "Expected a shift as input");
333 auto *BinInst = dyn_cast<BinaryOperator>(I.getOperand(0));
335 (!BinInst->isBitwiseLogicOp() &&
336 BinInst->getOpcode() != Instruction::Add &&
337 BinInst->getOpcode() != Instruction::Sub) ||
338 !BinInst->hasOneUse())
342 if (!match(I.getOperand(1), m_Constant(C1)))
345 Instruction::BinaryOps ShiftOpcode = I.getOpcode();
346 // Transform for add/sub only works with shl.
347 if ((BinInst->getOpcode() == Instruction::Add ||
348 BinInst->getOpcode() == Instruction::Sub) &&
349 ShiftOpcode != Instruction::Shl)
352 Type *Ty = I.getType();
354 // Find a matching one-use shift by constant. The fold is not valid if the sum
355 // of the shift values equals or exceeds bitwidth.
356 // TODO: Remove the one-use check if the other logic operand (Y) is constant.
358 auto matchFirstShift = [&](Value *V) {
359 APInt Threshold(Ty->getScalarSizeInBits(), Ty->getScalarSizeInBits());
361 m_OneUse(m_BinOp(ShiftOpcode, m_Value(X), m_Constant(C0)))) &&
362 match(ConstantExpr::getAdd(C0, C1),
363 m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, Threshold));
366 // Logic ops and Add are commutative, so check each operand for a match. Sub
367 // is not so we cannot reoder if we match operand(1) and need to keep the
368 // operands in their original positions.
369 bool FirstShiftIsOp1 = false;
370 if (matchFirstShift(BinInst->getOperand(0)))
371 Y = BinInst->getOperand(1);
372 else if (matchFirstShift(BinInst->getOperand(1))) {
373 Y = BinInst->getOperand(0);
374 FirstShiftIsOp1 = BinInst->getOpcode() == Instruction::Sub;
378 // shift (binop (shift X, C0), Y), C1 -> binop (shift X, C0+C1), (shift Y, C1)
379 Constant *ShiftSumC = ConstantExpr::getAdd(C0, C1);
380 Value *NewShift1 = Builder.CreateBinOp(ShiftOpcode, X, ShiftSumC);
381 Value *NewShift2 = Builder.CreateBinOp(ShiftOpcode, Y, C1);
382 Value *Op1 = FirstShiftIsOp1 ? NewShift2 : NewShift1;
383 Value *Op2 = FirstShiftIsOp1 ? NewShift1 : NewShift2;
384 return BinaryOperator::Create(BinInst->getOpcode(), Op1, Op2);
387 Instruction *InstCombinerImpl::commonShiftTransforms(BinaryOperator &I) {
388 if (Instruction *Phi = foldBinopWithPhiOperands(I))
391 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
392 assert(Op0->getType() == Op1->getType());
393 Type *Ty = I.getType();
395 // If the shift amount is a one-use `sext`, we can demote it to `zext`.
397 if (match(Op1, m_OneUse(m_SExt(m_Value(Y))))) {
398 Value *NewExt = Builder.CreateZExt(Y, Ty, Op1->getName());
399 return BinaryOperator::Create(I.getOpcode(), Op0, NewExt);
402 // See if we can fold away this shift.
403 if (SimplifyDemandedInstructionBits(I))
406 // Try to fold constant and into select arguments.
407 if (isa<Constant>(Op0))
408 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
409 if (Instruction *R = FoldOpIntoSelect(I, SI))
412 if (Constant *CUI = dyn_cast<Constant>(Op1))
413 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
416 if (auto *NewShift = cast_or_null<Instruction>(
417 reassociateShiftAmtsOfTwoSameDirectionShifts(&I, SQ)))
420 // Pre-shift a constant shifted by a variable amount with constant offset:
421 // C shift (A add nuw C1) --> (C shift C1) shift A
424 if (match(Op0, m_Constant(C)) &&
425 match(Op1, m_NUWAdd(m_Value(A), m_Constant(C1)))) {
426 Value *NewC = Builder.CreateBinOp(I.getOpcode(), C, C1);
427 return BinaryOperator::Create(I.getOpcode(), NewC, A);
430 unsigned BitWidth = Ty->getScalarSizeInBits();
432 const APInt *AC, *AddC;
433 // Try to pre-shift a constant shifted by a variable amount added with a
435 // C << (X - AddC) --> (C >> AddC) << X
437 // C >> (X - AddC) --> (C << AddC) >> X
438 if (match(Op0, m_APInt(AC)) && match(Op1, m_Add(m_Value(A), m_APInt(AddC))) &&
439 AddC->isNegative() && (-*AddC).ult(BitWidth)) {
440 assert(!AC->isZero() && "Expected simplify of shifted zero");
441 unsigned PosOffset = (-*AddC).getZExtValue();
443 auto isSuitableForPreShift = [PosOffset, &I, AC]() {
444 switch (I.getOpcode()) {
447 case Instruction::Shl:
448 return (I.hasNoSignedWrap() || I.hasNoUnsignedWrap()) &&
449 AC->eq(AC->lshr(PosOffset).shl(PosOffset));
450 case Instruction::LShr:
451 return I.isExact() && AC->eq(AC->shl(PosOffset).lshr(PosOffset));
452 case Instruction::AShr:
453 return I.isExact() && AC->eq(AC->shl(PosOffset).ashr(PosOffset));
456 if (isSuitableForPreShift()) {
457 Constant *NewC = ConstantInt::get(Ty, I.getOpcode() == Instruction::Shl
458 ? AC->lshr(PosOffset)
459 : AC->shl(PosOffset));
460 BinaryOperator *NewShiftOp =
461 BinaryOperator::Create(I.getOpcode(), NewC, A);
462 if (I.getOpcode() == Instruction::Shl) {
463 NewShiftOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
465 NewShiftOp->setIsExact();
471 // X shift (A srem C) -> X shift (A and (C - 1)) iff C is a power of 2.
472 // Because shifts by negative values (which could occur if A were negative)
474 if (Op1->hasOneUse() && match(Op1, m_SRem(m_Value(A), m_Constant(C))) &&
475 match(C, m_Power2())) {
476 // FIXME: Should this get moved into SimplifyDemandedBits by saying we don't
477 // demand the sign bit (and many others) here??
478 Constant *Mask = ConstantExpr::getSub(C, ConstantInt::get(Ty, 1));
479 Value *Rem = Builder.CreateAnd(A, Mask, Op1->getName());
480 return replaceOperand(I, 1, Rem);
483 if (Instruction *Logic = foldShiftOfShiftedBinOp(I, Builder))
486 if (match(Op1, m_Or(m_Value(), m_SpecificInt(BitWidth - 1))))
487 return replaceOperand(I, 1, ConstantInt::get(Ty, BitWidth - 1));
492 /// Return true if we can simplify two logical (either left or right) shifts
493 /// that have constant shift amounts: OuterShift (InnerShift X, C1), C2.
494 static bool canEvaluateShiftedShift(unsigned OuterShAmt, bool IsOuterShl,
495 Instruction *InnerShift,
496 InstCombinerImpl &IC, Instruction *CxtI) {
497 assert(InnerShift->isLogicalShift() && "Unexpected instruction type");
499 // We need constant scalar or constant splat shifts.
500 const APInt *InnerShiftConst;
501 if (!match(InnerShift->getOperand(1), m_APInt(InnerShiftConst)))
504 // Two logical shifts in the same direction:
505 // shl (shl X, C1), C2 --> shl X, C1 + C2
506 // lshr (lshr X, C1), C2 --> lshr X, C1 + C2
507 bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl;
508 if (IsInnerShl == IsOuterShl)
511 // Equal shift amounts in opposite directions become bitwise 'and':
512 // lshr (shl X, C), C --> and X, C'
513 // shl (lshr X, C), C --> and X, C'
514 if (*InnerShiftConst == OuterShAmt)
517 // If the 2nd shift is bigger than the 1st, we can fold:
518 // lshr (shl X, C1), C2 --> and (shl X, C1 - C2), C3
519 // shl (lshr X, C1), C2 --> and (lshr X, C1 - C2), C3
520 // but it isn't profitable unless we know the and'd out bits are already zero.
521 // Also, check that the inner shift is valid (less than the type width) or
522 // we'll crash trying to produce the bit mask for the 'and'.
523 unsigned TypeWidth = InnerShift->getType()->getScalarSizeInBits();
524 if (InnerShiftConst->ugt(OuterShAmt) && InnerShiftConst->ult(TypeWidth)) {
525 unsigned InnerShAmt = InnerShiftConst->getZExtValue();
527 IsInnerShl ? TypeWidth - InnerShAmt : InnerShAmt - OuterShAmt;
528 APInt Mask = APInt::getLowBitsSet(TypeWidth, OuterShAmt) << MaskShift;
529 if (IC.MaskedValueIsZero(InnerShift->getOperand(0), Mask, 0, CxtI))
536 /// See if we can compute the specified value, but shifted logically to the left
537 /// or right by some number of bits. This should return true if the expression
538 /// can be computed for the same cost as the current expression tree. This is
539 /// used to eliminate extraneous shifting from things like:
540 /// %C = shl i128 %A, 64
541 /// %D = shl i128 %B, 96
542 /// %E = or i128 %C, %D
543 /// %F = lshr i128 %E, 64
544 /// where the client will ask if E can be computed shifted right by 64-bits. If
545 /// this succeeds, getShiftedValue() will be called to produce the value.
546 static bool canEvaluateShifted(Value *V, unsigned NumBits, bool IsLeftShift,
547 InstCombinerImpl &IC, Instruction *CxtI) {
548 // We can always evaluate constants shifted.
549 if (isa<Constant>(V))
552 Instruction *I = dyn_cast<Instruction>(V);
553 if (!I) return false;
555 // We can't mutate something that has multiple uses: doing so would
556 // require duplicating the instruction in general, which isn't profitable.
557 if (!I->hasOneUse()) return false;
559 switch (I->getOpcode()) {
560 default: return false;
561 case Instruction::And:
562 case Instruction::Or:
563 case Instruction::Xor:
564 // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
565 return canEvaluateShifted(I->getOperand(0), NumBits, IsLeftShift, IC, I) &&
566 canEvaluateShifted(I->getOperand(1), NumBits, IsLeftShift, IC, I);
568 case Instruction::Shl:
569 case Instruction::LShr:
570 return canEvaluateShiftedShift(NumBits, IsLeftShift, I, IC, CxtI);
572 case Instruction::Select: {
573 SelectInst *SI = cast<SelectInst>(I);
574 Value *TrueVal = SI->getTrueValue();
575 Value *FalseVal = SI->getFalseValue();
576 return canEvaluateShifted(TrueVal, NumBits, IsLeftShift, IC, SI) &&
577 canEvaluateShifted(FalseVal, NumBits, IsLeftShift, IC, SI);
579 case Instruction::PHI: {
580 // We can change a phi if we can change all operands. Note that we never
581 // get into trouble with cyclic PHIs here because we only consider
582 // instructions with a single use.
583 PHINode *PN = cast<PHINode>(I);
584 for (Value *IncValue : PN->incoming_values())
585 if (!canEvaluateShifted(IncValue, NumBits, IsLeftShift, IC, PN))
589 case Instruction::Mul: {
590 const APInt *MulConst;
591 // We can fold (shr (mul X, -(1 << C)), C) -> (and (neg X), C`)
592 return !IsLeftShift && match(I->getOperand(1), m_APInt(MulConst)) &&
593 MulConst->isNegatedPowerOf2() && MulConst->countr_zero() == NumBits;
598 /// Fold OuterShift (InnerShift X, C1), C2.
599 /// See canEvaluateShiftedShift() for the constraints on these instructions.
600 static Value *foldShiftedShift(BinaryOperator *InnerShift, unsigned OuterShAmt,
602 InstCombiner::BuilderTy &Builder) {
603 bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl;
604 Type *ShType = InnerShift->getType();
605 unsigned TypeWidth = ShType->getScalarSizeInBits();
607 // We only accept shifts-by-a-constant in canEvaluateShifted().
609 match(InnerShift->getOperand(1), m_APInt(C1));
610 unsigned InnerShAmt = C1->getZExtValue();
612 // Change the shift amount and clear the appropriate IR flags.
613 auto NewInnerShift = [&](unsigned ShAmt) {
614 InnerShift->setOperand(1, ConstantInt::get(ShType, ShAmt));
616 InnerShift->setHasNoUnsignedWrap(false);
617 InnerShift->setHasNoSignedWrap(false);
619 InnerShift->setIsExact(false);
624 // Two logical shifts in the same direction:
625 // shl (shl X, C1), C2 --> shl X, C1 + C2
626 // lshr (lshr X, C1), C2 --> lshr X, C1 + C2
627 if (IsInnerShl == IsOuterShl) {
628 // If this is an oversized composite shift, then unsigned shifts get 0.
629 if (InnerShAmt + OuterShAmt >= TypeWidth)
630 return Constant::getNullValue(ShType);
632 return NewInnerShift(InnerShAmt + OuterShAmt);
635 // Equal shift amounts in opposite directions become bitwise 'and':
636 // lshr (shl X, C), C --> and X, C'
637 // shl (lshr X, C), C --> and X, C'
638 if (InnerShAmt == OuterShAmt) {
639 APInt Mask = IsInnerShl
640 ? APInt::getLowBitsSet(TypeWidth, TypeWidth - OuterShAmt)
641 : APInt::getHighBitsSet(TypeWidth, TypeWidth - OuterShAmt);
642 Value *And = Builder.CreateAnd(InnerShift->getOperand(0),
643 ConstantInt::get(ShType, Mask));
644 if (auto *AndI = dyn_cast<Instruction>(And)) {
645 AndI->moveBefore(InnerShift);
646 AndI->takeName(InnerShift);
651 assert(InnerShAmt > OuterShAmt &&
652 "Unexpected opposite direction logical shift pair");
654 // In general, we would need an 'and' for this transform, but
655 // canEvaluateShiftedShift() guarantees that the masked-off bits are not used.
656 // lshr (shl X, C1), C2 --> shl X, C1 - C2
657 // shl (lshr X, C1), C2 --> lshr X, C1 - C2
658 return NewInnerShift(InnerShAmt - OuterShAmt);
661 /// When canEvaluateShifted() returns true for an expression, this function
662 /// inserts the new computation that produces the shifted value.
663 static Value *getShiftedValue(Value *V, unsigned NumBits, bool isLeftShift,
664 InstCombinerImpl &IC, const DataLayout &DL) {
665 // We can always evaluate constants shifted.
666 if (Constant *C = dyn_cast<Constant>(V)) {
668 return IC.Builder.CreateShl(C, NumBits);
670 return IC.Builder.CreateLShr(C, NumBits);
673 Instruction *I = cast<Instruction>(V);
676 switch (I->getOpcode()) {
677 default: llvm_unreachable("Inconsistency with CanEvaluateShifted");
678 case Instruction::And:
679 case Instruction::Or:
680 case Instruction::Xor:
681 // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
683 0, getShiftedValue(I->getOperand(0), NumBits, isLeftShift, IC, DL));
685 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL));
688 case Instruction::Shl:
689 case Instruction::LShr:
690 return foldShiftedShift(cast<BinaryOperator>(I), NumBits, isLeftShift,
693 case Instruction::Select:
695 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL));
697 2, getShiftedValue(I->getOperand(2), NumBits, isLeftShift, IC, DL));
699 case Instruction::PHI: {
700 // We can change a phi if we can change all operands. Note that we never
701 // get into trouble with cyclic PHIs here because we only consider
702 // instructions with a single use.
703 PHINode *PN = cast<PHINode>(I);
704 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
705 PN->setIncomingValue(i, getShiftedValue(PN->getIncomingValue(i), NumBits,
706 isLeftShift, IC, DL));
709 case Instruction::Mul: {
710 assert(!isLeftShift && "Unexpected shift direction!");
711 auto *Neg = BinaryOperator::CreateNeg(I->getOperand(0));
712 IC.InsertNewInstWith(Neg, *I);
713 unsigned TypeWidth = I->getType()->getScalarSizeInBits();
714 APInt Mask = APInt::getLowBitsSet(TypeWidth, TypeWidth - NumBits);
715 auto *And = BinaryOperator::CreateAnd(Neg,
716 ConstantInt::get(I->getType(), Mask));
718 return IC.InsertNewInstWith(And, *I);
723 // If this is a bitwise operator or add with a constant RHS we might be able
724 // to pull it through a shift.
725 static bool canShiftBinOpWithConstantRHS(BinaryOperator &Shift,
726 BinaryOperator *BO) {
727 switch (BO->getOpcode()) {
729 return false; // Do not perform transform!
730 case Instruction::Add:
731 return Shift.getOpcode() == Instruction::Shl;
732 case Instruction::Or:
733 case Instruction::And:
735 case Instruction::Xor:
736 // Do not change a 'not' of logical shift because that would create a normal
737 // 'xor'. The 'not' is likely better for analysis, SCEV, and codegen.
738 return !(Shift.isLogicalShift() && match(BO, m_Not(m_Value())));
742 Instruction *InstCombinerImpl::FoldShiftByConstant(Value *Op0, Constant *C1,
744 // (C2 << X) << C1 --> (C2 << C1) << X
745 // (C2 >> X) >> C1 --> (C2 >> C1) >> X
748 if (match(Op0, m_BinOp(I.getOpcode(), m_Constant(C2), m_Value(X))))
749 return BinaryOperator::Create(
750 I.getOpcode(), Builder.CreateBinOp(I.getOpcode(), C2, C1), X);
752 bool IsLeftShift = I.getOpcode() == Instruction::Shl;
753 Type *Ty = I.getType();
754 unsigned TypeBits = Ty->getScalarSizeInBits();
756 // (X / +DivC) >> (Width - 1) --> ext (X <= -DivC)
757 // (X / -DivC) >> (Width - 1) --> ext (X >= +DivC)
759 if (!IsLeftShift && match(C1, m_SpecificIntAllowUndef(TypeBits - 1)) &&
760 match(Op0, m_SDiv(m_Value(X), m_APInt(DivC))) && !DivC->isZero() &&
761 !DivC->isMinSignedValue()) {
762 Constant *NegDivC = ConstantInt::get(Ty, -(*DivC));
763 ICmpInst::Predicate Pred =
764 DivC->isNegative() ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_SLE;
765 Value *Cmp = Builder.CreateICmp(Pred, X, NegDivC);
766 auto ExtOpcode = (I.getOpcode() == Instruction::AShr) ? Instruction::SExt
768 return CastInst::Create(ExtOpcode, Cmp, Ty);
772 if (!match(C1, m_APInt(Op1C)))
775 assert(!Op1C->uge(TypeBits) &&
776 "Shift over the type width should have been removed already");
778 // See if we can propagate this shift into the input, this covers the trivial
779 // cast of lshr(shl(x,c1),c2) as well as other more complex cases.
780 if (I.getOpcode() != Instruction::AShr &&
781 canEvaluateShifted(Op0, Op1C->getZExtValue(), IsLeftShift, *this, &I)) {
783 dbgs() << "ICE: GetShiftedValue propagating shift through expression"
784 " to eliminate shift:\n IN: "
785 << *Op0 << "\n SH: " << I << "\n");
787 return replaceInstUsesWith(
788 I, getShiftedValue(Op0, Op1C->getZExtValue(), IsLeftShift, *this, DL));
791 if (Instruction *FoldedShift = foldBinOpIntoSelectOrPhi(I))
794 if (!Op0->hasOneUse())
797 if (auto *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
798 // If the operand is a bitwise operator with a constant RHS, and the
799 // shift is the only use, we can pull it out of the shift.
801 if (match(Op0BO->getOperand(1), m_APInt(Op0C))) {
802 if (canShiftBinOpWithConstantRHS(I, Op0BO)) {
804 Builder.CreateBinOp(I.getOpcode(), Op0BO->getOperand(1), C1);
807 Builder.CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), C1);
808 NewShift->takeName(Op0BO);
810 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift, NewRHS);
815 // If we have a select that conditionally executes some binary operator,
816 // see if we can pull it the select and operator through the shift.
818 // For example, turning:
819 // shl (select C, (add X, C1), X), C2
822 // select C, (add Y, C1 << C2), Y
826 if (match(Op0, m_Select(m_Value(Cond), m_OneUse(m_BinOp(TBO)),
827 m_Value(FalseVal)))) {
829 if (!isa<Constant>(FalseVal) && TBO->getOperand(0) == FalseVal &&
830 match(TBO->getOperand(1), m_APInt(C)) &&
831 canShiftBinOpWithConstantRHS(I, TBO)) {
833 Builder.CreateBinOp(I.getOpcode(), TBO->getOperand(1), C1);
835 Value *NewShift = Builder.CreateBinOp(I.getOpcode(), FalseVal, C1);
836 Value *NewOp = Builder.CreateBinOp(TBO->getOpcode(), NewShift, NewRHS);
837 return SelectInst::Create(Cond, NewOp, NewShift);
843 if (match(Op0, m_Select(m_Value(Cond), m_Value(TrueVal),
844 m_OneUse(m_BinOp(FBO))))) {
846 if (!isa<Constant>(TrueVal) && FBO->getOperand(0) == TrueVal &&
847 match(FBO->getOperand(1), m_APInt(C)) &&
848 canShiftBinOpWithConstantRHS(I, FBO)) {
850 Builder.CreateBinOp(I.getOpcode(), FBO->getOperand(1), C1);
852 Value *NewShift = Builder.CreateBinOp(I.getOpcode(), TrueVal, C1);
853 Value *NewOp = Builder.CreateBinOp(FBO->getOpcode(), NewShift, NewRHS);
854 return SelectInst::Create(Cond, NewShift, NewOp);
862 // (lshr (add (zext X), (zext Y)), K)
863 // -> (icmp ult (add X, Y), X)
865 // - The add's operands are zexts from a K-bits integer to a bigger type.
866 // - The add is only used by the shr, or by iK (or narrower) truncates.
867 // - The lshr type has more than 2 bits (other types are boolean math).
870 // - The resulting add cannot have nuw/nsw, else on overflow we get a
871 // poison value and the transform isn't legal anymore.
872 Instruction *InstCombinerImpl::foldLShrOverflowBit(BinaryOperator &I) {
873 assert(I.getOpcode() == Instruction::LShr);
875 Value *Add = I.getOperand(0);
876 Value *ShiftAmt = I.getOperand(1);
877 Type *Ty = I.getType();
879 if (Ty->getScalarSizeInBits() < 3)
882 const APInt *ShAmtAPInt = nullptr;
883 Value *X = nullptr, *Y = nullptr;
884 if (!match(ShiftAmt, m_APInt(ShAmtAPInt)) ||
886 m_Add(m_OneUse(m_ZExt(m_Value(X))), m_OneUse(m_ZExt(m_Value(Y))))))
889 const unsigned ShAmt = ShAmtAPInt->getZExtValue();
893 // X/Y are zexts from `ShAmt`-sized ints.
894 if (X->getType()->getScalarSizeInBits() != ShAmt ||
895 Y->getType()->getScalarSizeInBits() != ShAmt)
898 // Make sure that `Add` is only used by `I` and `ShAmt`-truncates.
899 if (!Add->hasOneUse()) {
900 for (User *U : Add->users()) {
904 TruncInst *Trunc = dyn_cast<TruncInst>(U);
905 if (!Trunc || Trunc->getType()->getScalarSizeInBits() > ShAmt)
910 // Insert at Add so that the newly created `NarrowAdd` will dominate it's
911 // users (i.e. `Add`'s users).
912 Instruction *AddInst = cast<Instruction>(Add);
913 Builder.SetInsertPoint(AddInst);
915 Value *NarrowAdd = Builder.CreateAdd(X, Y, "add.narrowed");
917 Builder.CreateICmpULT(NarrowAdd, X, "add.narrowed.overflow");
919 // Replace the uses of the original add with a zext of the
920 // NarrowAdd's result. Note that all users at this stage are known to
921 // be ShAmt-sized truncs, or the lshr itself.
922 if (!Add->hasOneUse()) {
923 replaceInstUsesWith(*AddInst, Builder.CreateZExt(NarrowAdd, Ty));
924 eraseInstFromFunction(*AddInst);
927 // Replace the LShr with a zext of the overflow check.
928 return new ZExtInst(Overflow, Ty);
931 Instruction *InstCombinerImpl::visitShl(BinaryOperator &I) {
932 const SimplifyQuery Q = SQ.getWithInstruction(&I);
934 if (Value *V = simplifyShlInst(I.getOperand(0), I.getOperand(1),
935 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(), Q))
936 return replaceInstUsesWith(I, V);
938 if (Instruction *X = foldVectorBinop(I))
941 if (Instruction *V = commonShiftTransforms(I))
944 if (Instruction *V = dropRedundantMaskingOfLeftShiftInput(&I, Q, Builder))
947 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
948 Type *Ty = I.getType();
949 unsigned BitWidth = Ty->getScalarSizeInBits();
952 if (match(Op1, m_APInt(C))) {
953 unsigned ShAmtC = C->getZExtValue();
955 // shl (zext X), C --> zext (shl X, C)
956 // This is only valid if X would have zeros shifted out.
958 if (match(Op0, m_OneUse(m_ZExt(m_Value(X))))) {
959 unsigned SrcWidth = X->getType()->getScalarSizeInBits();
960 if (ShAmtC < SrcWidth &&
961 MaskedValueIsZero(X, APInt::getHighBitsSet(SrcWidth, ShAmtC), 0, &I))
962 return new ZExtInst(Builder.CreateShl(X, ShAmtC), Ty);
965 // (X >> C) << C --> X & (-1 << C)
966 if (match(Op0, m_Shr(m_Value(X), m_Specific(Op1)))) {
967 APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmtC));
968 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask));
972 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_APInt(C1)))) &&
974 unsigned ShrAmt = C1->getZExtValue();
975 if (ShrAmt < ShAmtC) {
976 // If C1 < C: (X >>?,exact C1) << C --> X << (C - C1)
977 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmtC - ShrAmt);
978 auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
979 NewShl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
980 NewShl->setHasNoSignedWrap(I.hasNoSignedWrap());
983 if (ShrAmt > ShAmtC) {
984 // If C1 > C: (X >>?exact C1) << C --> X >>?exact (C1 - C)
985 Constant *ShiftDiff = ConstantInt::get(Ty, ShrAmt - ShAmtC);
986 auto *NewShr = BinaryOperator::Create(
987 cast<BinaryOperator>(Op0)->getOpcode(), X, ShiftDiff);
988 NewShr->setIsExact(true);
993 if (match(Op0, m_OneUse(m_Shr(m_Value(X), m_APInt(C1)))) &&
995 unsigned ShrAmt = C1->getZExtValue();
996 if (ShrAmt < ShAmtC) {
997 // If C1 < C: (X >>? C1) << C --> (X << (C - C1)) & (-1 << C)
998 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmtC - ShrAmt);
999 auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
1000 NewShl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1001 NewShl->setHasNoSignedWrap(I.hasNoSignedWrap());
1002 Builder.Insert(NewShl);
1003 APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmtC));
1004 return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask));
1006 if (ShrAmt > ShAmtC) {
1007 // If C1 > C: (X >>? C1) << C --> (X >>? (C1 - C)) & (-1 << C)
1008 Constant *ShiftDiff = ConstantInt::get(Ty, ShrAmt - ShAmtC);
1009 auto *OldShr = cast<BinaryOperator>(Op0);
1011 BinaryOperator::Create(OldShr->getOpcode(), X, ShiftDiff);
1012 NewShr->setIsExact(OldShr->isExact());
1013 Builder.Insert(NewShr);
1014 APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmtC));
1015 return BinaryOperator::CreateAnd(NewShr, ConstantInt::get(Ty, Mask));
1019 // Similar to above, but look through an intermediate trunc instruction.
1020 BinaryOperator *Shr;
1021 if (match(Op0, m_OneUse(m_Trunc(m_OneUse(m_BinOp(Shr))))) &&
1022 match(Shr, m_Shr(m_Value(X), m_APInt(C1)))) {
1023 // The larger shift direction survives through the transform.
1024 unsigned ShrAmtC = C1->getZExtValue();
1025 unsigned ShDiff = ShrAmtC > ShAmtC ? ShrAmtC - ShAmtC : ShAmtC - ShrAmtC;
1026 Constant *ShiftDiffC = ConstantInt::get(X->getType(), ShDiff);
1027 auto ShiftOpc = ShrAmtC > ShAmtC ? Shr->getOpcode() : Instruction::Shl;
1030 // (trunc (X >> C1)) << C --> (trunc (X >> (C1 - C))) && (-1 << C)
1032 // (trunc (X >> C1)) << C --> (trunc (X << (C - C1))) && (-1 << C)
1033 Value *NewShift = Builder.CreateBinOp(ShiftOpc, X, ShiftDiffC, "sh.diff");
1034 Value *Trunc = Builder.CreateTrunc(NewShift, Ty, "tr.sh.diff");
1035 APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmtC));
1036 return BinaryOperator::CreateAnd(Trunc, ConstantInt::get(Ty, Mask));
1039 if (match(Op0, m_Shl(m_Value(X), m_APInt(C1))) && C1->ult(BitWidth)) {
1040 unsigned AmtSum = ShAmtC + C1->getZExtValue();
1041 // Oversized shifts are simplified to zero in InstSimplify.
1042 if (AmtSum < BitWidth)
1043 // (X << C1) << C2 --> X << (C1 + C2)
1044 return BinaryOperator::CreateShl(X, ConstantInt::get(Ty, AmtSum));
1047 // If we have an opposite shift by the same amount, we may be able to
1048 // reorder binops and shifts to eliminate math/logic.
1049 auto isSuitableBinOpcode = [](Instruction::BinaryOps BinOpcode) {
1050 switch (BinOpcode) {
1053 case Instruction::Add:
1054 case Instruction::And:
1055 case Instruction::Or:
1056 case Instruction::Xor:
1057 case Instruction::Sub:
1058 // NOTE: Sub is not commutable and the tranforms below may not be valid
1059 // when the shift-right is operand 1 (RHS) of the sub.
1063 BinaryOperator *Op0BO;
1064 if (match(Op0, m_OneUse(m_BinOp(Op0BO))) &&
1065 isSuitableBinOpcode(Op0BO->getOpcode())) {
1066 // Commute so shift-right is on LHS of the binop.
1067 // (Y bop (X >> C)) << C -> ((X >> C) bop Y) << C
1068 // (Y bop ((X >> C) & CC)) << C -> (((X >> C) & CC) bop Y) << C
1069 Value *Shr = Op0BO->getOperand(0);
1070 Value *Y = Op0BO->getOperand(1);
1073 if (Op0BO->isCommutative() && Y->hasOneUse() &&
1074 (match(Y, m_Shr(m_Value(), m_Specific(Op1))) ||
1075 match(Y, m_And(m_OneUse(m_Shr(m_Value(), m_Specific(Op1))),
1079 // ((X >> C) bop Y) << C -> (X bop (Y << C)) & (~0 << C)
1080 if (match(Shr, m_OneUse(m_Shr(m_Value(X), m_Specific(Op1))))) {
1082 Value *YS = Builder.CreateShl(Y, Op1, Op0BO->getName());
1085 Builder.CreateBinOp(Op0BO->getOpcode(), X, YS, Shr->getName());
1086 unsigned Op1Val = C->getLimitedValue(BitWidth);
1087 APInt Bits = APInt::getHighBitsSet(BitWidth, BitWidth - Op1Val);
1088 Constant *Mask = ConstantInt::get(Ty, Bits);
1089 return BinaryOperator::CreateAnd(B, Mask);
1092 // (((X >> C) & CC) bop Y) << C -> (X & (CC << C)) bop (Y << C)
1094 m_OneUse(m_And(m_OneUse(m_Shr(m_Value(X), m_Specific(Op1))),
1097 Value *YS = Builder.CreateShl(Y, Op1, Op0BO->getName());
1099 Value *M = Builder.CreateAnd(X, ConstantInt::get(Ty, CC->shl(*C)),
1100 X->getName() + ".mask");
1101 return BinaryOperator::Create(Op0BO->getOpcode(), M, YS);
1105 // (C1 - X) << C --> (C1 << C) - (X << C)
1106 if (match(Op0, m_OneUse(m_Sub(m_APInt(C1), m_Value(X))))) {
1107 Constant *NewLHS = ConstantInt::get(Ty, C1->shl(*C));
1108 Value *NewShift = Builder.CreateShl(X, Op1);
1109 return BinaryOperator::CreateSub(NewLHS, NewShift);
1112 // If the shifted-out value is known-zero, then this is a NUW shift.
1113 if (!I.hasNoUnsignedWrap() &&
1114 MaskedValueIsZero(Op0, APInt::getHighBitsSet(BitWidth, ShAmtC), 0,
1116 I.setHasNoUnsignedWrap();
1120 // If the shifted-out value is all signbits, then this is a NSW shift.
1121 if (!I.hasNoSignedWrap() && ComputeNumSignBits(Op0, 0, &I) > ShAmtC) {
1122 I.setHasNoSignedWrap();
1127 // Transform (x >> y) << y to x & (-1 << y)
1128 // Valid for any type of right-shift.
1130 if (match(Op0, m_OneUse(m_Shr(m_Value(X), m_Specific(Op1))))) {
1131 Constant *AllOnes = ConstantInt::getAllOnesValue(Ty);
1132 Value *Mask = Builder.CreateShl(AllOnes, Op1);
1133 return BinaryOperator::CreateAnd(Mask, X);
1137 if (match(Op1, m_Constant(C1))) {
1140 // (X * C2) << C1 --> X * (C2 << C1)
1141 if (match(Op0, m_Mul(m_Value(X), m_Constant(C2))))
1142 return BinaryOperator::CreateMul(X, ConstantExpr::getShl(C2, C1));
1144 // shl (zext i1 X), C1 --> select (X, 1 << C1, 0)
1145 if (match(Op0, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
1146 auto *NewC = ConstantExpr::getShl(ConstantInt::get(Ty, 1), C1);
1147 return SelectInst::Create(X, NewC, ConstantInt::getNullValue(Ty));
1151 if (match(Op0, m_One())) {
1152 // (1 << (C - x)) -> ((1 << C) >> x) if C is bitwidth - 1
1153 if (match(Op1, m_Sub(m_SpecificInt(BitWidth - 1), m_Value(X))))
1154 return BinaryOperator::CreateLShr(
1155 ConstantInt::get(Ty, APInt::getSignMask(BitWidth)), X);
1157 // Canonicalize "extract lowest set bit" using cttz to and-with-negate:
1158 // 1 << (cttz X) --> -X & X
1160 m_OneUse(m_Intrinsic<Intrinsic::cttz>(m_Value(X), m_Value())))) {
1161 Value *NegX = Builder.CreateNeg(X, "neg");
1162 return BinaryOperator::CreateAnd(NegX, X);
1165 // The only way to shift out the 1 is with an over-shift, so that would
1166 // be poison with or without "nuw". Undef is excluded because (undef << X)
1167 // is not undef (it is zero).
1168 Constant *ConstantOne = cast<Constant>(Op0);
1169 if (!I.hasNoUnsignedWrap() && !ConstantOne->containsUndefElement()) {
1170 I.setHasNoUnsignedWrap();
1178 Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) {
1179 if (Value *V = simplifyLShrInst(I.getOperand(0), I.getOperand(1), I.isExact(),
1180 SQ.getWithInstruction(&I)))
1181 return replaceInstUsesWith(I, V);
1183 if (Instruction *X = foldVectorBinop(I))
1186 if (Instruction *R = commonShiftTransforms(I))
1189 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1190 Type *Ty = I.getType();
1193 unsigned BitWidth = Ty->getScalarSizeInBits();
1195 // (iN (~X) u>> (N - 1)) --> zext (X > -1)
1196 if (match(Op0, m_OneUse(m_Not(m_Value(X)))) &&
1197 match(Op1, m_SpecificIntAllowUndef(BitWidth - 1)))
1198 return new ZExtInst(Builder.CreateIsNotNeg(X, "isnotneg"), Ty);
1200 if (match(Op1, m_APInt(C))) {
1201 unsigned ShAmtC = C->getZExtValue();
1202 auto *II = dyn_cast<IntrinsicInst>(Op0);
1203 if (II && isPowerOf2_32(BitWidth) && Log2_32(BitWidth) == ShAmtC &&
1204 (II->getIntrinsicID() == Intrinsic::ctlz ||
1205 II->getIntrinsicID() == Intrinsic::cttz ||
1206 II->getIntrinsicID() == Intrinsic::ctpop)) {
1207 // ctlz.i32(x)>>5 --> zext(x == 0)
1208 // cttz.i32(x)>>5 --> zext(x == 0)
1209 // ctpop.i32(x)>>5 --> zext(x == -1)
1210 bool IsPop = II->getIntrinsicID() == Intrinsic::ctpop;
1211 Constant *RHS = ConstantInt::getSigned(Ty, IsPop ? -1 : 0);
1212 Value *Cmp = Builder.CreateICmpEQ(II->getArgOperand(0), RHS);
1213 return new ZExtInst(Cmp, Ty);
1218 if (match(Op0, m_Shl(m_Value(X), m_APInt(C1))) && C1->ult(BitWidth)) {
1219 if (C1->ult(ShAmtC)) {
1220 unsigned ShlAmtC = C1->getZExtValue();
1221 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmtC - ShlAmtC);
1222 if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) {
1223 // (X <<nuw C1) >>u C --> X >>u (C - C1)
1224 auto *NewLShr = BinaryOperator::CreateLShr(X, ShiftDiff);
1225 NewLShr->setIsExact(I.isExact());
1228 if (Op0->hasOneUse()) {
1229 // (X << C1) >>u C --> (X >>u (C - C1)) & (-1 >> C)
1230 Value *NewLShr = Builder.CreateLShr(X, ShiftDiff, "", I.isExact());
1231 APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmtC));
1232 return BinaryOperator::CreateAnd(NewLShr, ConstantInt::get(Ty, Mask));
1234 } else if (C1->ugt(ShAmtC)) {
1235 unsigned ShlAmtC = C1->getZExtValue();
1236 Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmtC - ShAmtC);
1237 if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) {
1238 // (X <<nuw C1) >>u C --> X <<nuw (C1 - C)
1239 auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
1240 NewShl->setHasNoUnsignedWrap(true);
1243 if (Op0->hasOneUse()) {
1244 // (X << C1) >>u C --> X << (C1 - C) & (-1 >> C)
1245 Value *NewShl = Builder.CreateShl(X, ShiftDiff);
1246 APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmtC));
1247 return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask));
1250 assert(*C1 == ShAmtC);
1251 // (X << C) >>u C --> X & (-1 >>u C)
1252 APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmtC));
1253 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask));
1257 // ((X << C) + Y) >>u C --> (X + (Y >>u C)) & (-1 >>u C)
1258 // TODO: Consolidate with the more general transform that starts from shl
1259 // (the shifts are in the opposite order).
1262 m_OneUse(m_c_Add(m_OneUse(m_Shl(m_Value(X), m_Specific(Op1))),
1264 Value *NewLshr = Builder.CreateLShr(Y, Op1);
1265 Value *NewAdd = Builder.CreateAdd(NewLshr, X);
1266 unsigned Op1Val = C->getLimitedValue(BitWidth);
1267 APInt Bits = APInt::getLowBitsSet(BitWidth, BitWidth - Op1Val);
1268 Constant *Mask = ConstantInt::get(Ty, Bits);
1269 return BinaryOperator::CreateAnd(NewAdd, Mask);
1272 if (match(Op0, m_OneUse(m_ZExt(m_Value(X)))) &&
1273 (!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType()))) {
1274 assert(ShAmtC < X->getType()->getScalarSizeInBits() &&
1275 "Big shift not simplified to zero?");
1276 // lshr (zext iM X to iN), C --> zext (lshr X, C) to iN
1277 Value *NewLShr = Builder.CreateLShr(X, ShAmtC);
1278 return new ZExtInst(NewLShr, Ty);
1281 if (match(Op0, m_SExt(m_Value(X)))) {
1282 unsigned SrcTyBitWidth = X->getType()->getScalarSizeInBits();
1283 // lshr (sext i1 X to iN), C --> select (X, -1 >> C, 0)
1284 if (SrcTyBitWidth == 1) {
1285 auto *NewC = ConstantInt::get(
1286 Ty, APInt::getLowBitsSet(BitWidth, BitWidth - ShAmtC));
1287 return SelectInst::Create(X, NewC, ConstantInt::getNullValue(Ty));
1290 if ((!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType())) &&
1292 // Are we moving the sign bit to the low bit and widening with high
1293 // zeros? lshr (sext iM X to iN), N-1 --> zext (lshr X, M-1) to iN
1294 if (ShAmtC == BitWidth - 1) {
1295 Value *NewLShr = Builder.CreateLShr(X, SrcTyBitWidth - 1);
1296 return new ZExtInst(NewLShr, Ty);
1299 // lshr (sext iM X to iN), N-M --> zext (ashr X, min(N-M, M-1)) to iN
1300 if (ShAmtC == BitWidth - SrcTyBitWidth) {
1301 // The new shift amount can't be more than the narrow source type.
1302 unsigned NewShAmt = std::min(ShAmtC, SrcTyBitWidth - 1);
1303 Value *AShr = Builder.CreateAShr(X, NewShAmt);
1304 return new ZExtInst(AShr, Ty);
1309 if (ShAmtC == BitWidth - 1) {
1310 // lshr i32 or(X,-X), 31 --> zext (X != 0)
1311 if (match(Op0, m_OneUse(m_c_Or(m_Neg(m_Value(X)), m_Deferred(X)))))
1312 return new ZExtInst(Builder.CreateIsNotNull(X), Ty);
1314 // lshr i32 (X -nsw Y), 31 --> zext (X < Y)
1315 if (match(Op0, m_OneUse(m_NSWSub(m_Value(X), m_Value(Y)))))
1316 return new ZExtInst(Builder.CreateICmpSLT(X, Y), Ty);
1318 // Check if a number is negative and odd:
1319 // lshr i32 (srem X, 2), 31 --> and (X >> 31), X
1320 if (match(Op0, m_OneUse(m_SRem(m_Value(X), m_SpecificInt(2))))) {
1321 Value *Signbit = Builder.CreateLShr(X, ShAmtC);
1322 return BinaryOperator::CreateAnd(Signbit, X);
1326 // (X >>u C1) >>u C --> X >>u (C1 + C)
1327 if (match(Op0, m_LShr(m_Value(X), m_APInt(C1)))) {
1328 // Oversized shifts are simplified to zero in InstSimplify.
1329 unsigned AmtSum = ShAmtC + C1->getZExtValue();
1330 if (AmtSum < BitWidth)
1331 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
1334 Instruction *TruncSrc;
1335 if (match(Op0, m_OneUse(m_Trunc(m_Instruction(TruncSrc)))) &&
1336 match(TruncSrc, m_LShr(m_Value(X), m_APInt(C1)))) {
1337 unsigned SrcWidth = X->getType()->getScalarSizeInBits();
1338 unsigned AmtSum = ShAmtC + C1->getZExtValue();
1340 // If the combined shift fits in the source width:
1341 // (trunc (X >>u C1)) >>u C --> and (trunc (X >>u (C1 + C)), MaskC
1343 // If the first shift covers the number of bits truncated, then the
1344 // mask instruction is eliminated (and so the use check is relaxed).
1345 if (AmtSum < SrcWidth &&
1346 (TruncSrc->hasOneUse() || C1->uge(SrcWidth - BitWidth))) {
1347 Value *SumShift = Builder.CreateLShr(X, AmtSum, "sum.shift");
1348 Value *Trunc = Builder.CreateTrunc(SumShift, Ty, I.getName());
1350 // If the first shift does not cover the number of bits truncated, then
1351 // we require a mask to get rid of high bits in the result.
1352 APInt MaskC = APInt::getAllOnes(BitWidth).lshr(ShAmtC);
1353 return BinaryOperator::CreateAnd(Trunc, ConstantInt::get(Ty, MaskC));
1358 if (match(Op0, m_NUWMul(m_Value(X), m_APInt(MulC)))) {
1359 // Look for a "splat" mul pattern - it replicates bits across each half of
1360 // a value, so a right shift is just a mask of the low bits:
1361 // lshr i[2N] (mul nuw X, (2^N)+1), N --> and iN X, (2^N)-1
1362 // TODO: Generalize to allow more than just half-width shifts?
1363 if (BitWidth > 2 && ShAmtC * 2 == BitWidth && (*MulC - 1).isPowerOf2() &&
1364 MulC->logBase2() == ShAmtC)
1365 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, *MulC - 2));
1367 // The one-use check is not strictly necessary, but codegen may not be
1368 // able to invert the transform and perf may suffer with an extra mul
1370 if (Op0->hasOneUse()) {
1371 APInt NewMulC = MulC->lshr(ShAmtC);
1372 // if c is divisible by (1 << ShAmtC):
1373 // lshr (mul nuw x, MulC), ShAmtC -> mul nuw x, (MulC >> ShAmtC)
1374 if (MulC->eq(NewMulC.shl(ShAmtC))) {
1376 BinaryOperator::CreateNUWMul(X, ConstantInt::get(Ty, NewMulC));
1377 BinaryOperator *OrigMul = cast<BinaryOperator>(Op0);
1378 NewMul->setHasNoSignedWrap(OrigMul->hasNoSignedWrap());
1384 // Try to narrow bswap.
1385 // In the case where the shift amount equals the bitwidth difference, the
1386 // shift is eliminated.
1387 if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::bswap>(
1388 m_OneUse(m_ZExt(m_Value(X))))))) {
1389 unsigned SrcWidth = X->getType()->getScalarSizeInBits();
1390 unsigned WidthDiff = BitWidth - SrcWidth;
1391 if (SrcWidth % 16 == 0) {
1392 Value *NarrowSwap = Builder.CreateUnaryIntrinsic(Intrinsic::bswap, X);
1393 if (ShAmtC >= WidthDiff) {
1394 // (bswap (zext X)) >> C --> zext (bswap X >> C')
1395 Value *NewShift = Builder.CreateLShr(NarrowSwap, ShAmtC - WidthDiff);
1396 return new ZExtInst(NewShift, Ty);
1398 // (bswap (zext X)) >> C --> (zext (bswap X)) << C'
1399 Value *NewZExt = Builder.CreateZExt(NarrowSwap, Ty);
1400 Constant *ShiftDiff = ConstantInt::get(Ty, WidthDiff - ShAmtC);
1401 return BinaryOperator::CreateShl(NewZExt, ShiftDiff);
1406 // Reduce add-carry of bools to logic:
1407 // ((zext BoolX) + (zext BoolY)) >> 1 --> zext (BoolX && BoolY)
1408 Value *BoolX, *BoolY;
1409 if (ShAmtC == 1 && match(Op0, m_Add(m_Value(X), m_Value(Y))) &&
1410 match(X, m_ZExt(m_Value(BoolX))) && match(Y, m_ZExt(m_Value(BoolY))) &&
1411 BoolX->getType()->isIntOrIntVectorTy(1) &&
1412 BoolY->getType()->isIntOrIntVectorTy(1) &&
1413 (X->hasOneUse() || Y->hasOneUse() || Op0->hasOneUse())) {
1414 Value *And = Builder.CreateAnd(BoolX, BoolY);
1415 return new ZExtInst(And, Ty);
1418 // If the shifted-out value is known-zero, then this is an exact shift.
1420 MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmtC), 0, &I)) {
1426 // Transform (x << y) >> y to x & (-1 >> y)
1427 if (match(Op0, m_OneUse(m_Shl(m_Value(X), m_Specific(Op1))))) {
1428 Constant *AllOnes = ConstantInt::getAllOnesValue(Ty);
1429 Value *Mask = Builder.CreateLShr(AllOnes, Op1);
1430 return BinaryOperator::CreateAnd(Mask, X);
1433 if (Instruction *Overflow = foldLShrOverflowBit(I))
1440 InstCombinerImpl::foldVariableSignZeroExtensionOfVariableHighBitExtract(
1441 BinaryOperator &OldAShr) {
1442 assert(OldAShr.getOpcode() == Instruction::AShr &&
1443 "Must be called with arithmetic right-shift instruction only.");
1445 // Check that constant C is a splat of the element-wise bitwidth of V.
1446 auto BitWidthSplat = [](Constant *C, Value *V) {
1448 C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1449 APInt(C->getType()->getScalarSizeInBits(),
1450 V->getType()->getScalarSizeInBits())));
1453 // It should look like variable-length sign-extension on the outside:
1454 // (Val << (bitwidth(Val)-Nbits)) a>> (bitwidth(Val)-Nbits)
1456 Instruction *MaybeTrunc;
1458 if (!match(&OldAShr,
1459 m_AShr(m_Shl(m_Instruction(MaybeTrunc),
1460 m_ZExtOrSelf(m_Sub(m_Constant(C1),
1461 m_ZExtOrSelf(m_Value(NBits))))),
1462 m_ZExtOrSelf(m_Sub(m_Constant(C2),
1463 m_ZExtOrSelf(m_Deferred(NBits)))))) ||
1464 !BitWidthSplat(C1, &OldAShr) || !BitWidthSplat(C2, &OldAShr))
1467 // There may or may not be a truncation after outer two shifts.
1468 Instruction *HighBitExtract;
1469 match(MaybeTrunc, m_TruncOrSelf(m_Instruction(HighBitExtract)));
1470 bool HadTrunc = MaybeTrunc != HighBitExtract;
1472 // And finally, the innermost part of the pattern must be a right-shift.
1473 Value *X, *NumLowBitsToSkip;
1474 if (!match(HighBitExtract, m_Shr(m_Value(X), m_Value(NumLowBitsToSkip))))
1477 // Said right-shift must extract high NBits bits - C0 must be it's bitwidth.
1479 if (!match(NumLowBitsToSkip,
1481 m_Sub(m_Constant(C0), m_ZExtOrSelf(m_Specific(NBits))))) ||
1482 !BitWidthSplat(C0, HighBitExtract))
1485 // Since the NBits is identical for all shifts, if the outermost and
1486 // innermost shifts are identical, then outermost shifts are redundant.
1487 // If we had truncation, do keep it though.
1488 if (HighBitExtract->getOpcode() == OldAShr.getOpcode())
1489 return replaceInstUsesWith(OldAShr, MaybeTrunc);
1491 // Else, if there was a truncation, then we need to ensure that one
1492 // instruction will go away.
1493 if (HadTrunc && !match(&OldAShr, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
1496 // Finally, bypass two innermost shifts, and perform the outermost shift on
1497 // the operands of the innermost shift.
1498 Instruction *NewAShr =
1499 BinaryOperator::Create(OldAShr.getOpcode(), X, NumLowBitsToSkip);
1500 NewAShr->copyIRFlags(HighBitExtract); // We can preserve 'exact'-ness.
1504 Builder.Insert(NewAShr);
1505 return TruncInst::CreateTruncOrBitCast(NewAShr, OldAShr.getType());
1508 Instruction *InstCombinerImpl::visitAShr(BinaryOperator &I) {
1509 if (Value *V = simplifyAShrInst(I.getOperand(0), I.getOperand(1), I.isExact(),
1510 SQ.getWithInstruction(&I)))
1511 return replaceInstUsesWith(I, V);
1513 if (Instruction *X = foldVectorBinop(I))
1516 if (Instruction *R = commonShiftTransforms(I))
1519 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1520 Type *Ty = I.getType();
1521 unsigned BitWidth = Ty->getScalarSizeInBits();
1522 const APInt *ShAmtAPInt;
1523 if (match(Op1, m_APInt(ShAmtAPInt)) && ShAmtAPInt->ult(BitWidth)) {
1524 unsigned ShAmt = ShAmtAPInt->getZExtValue();
1526 // If the shift amount equals the difference in width of the destination
1527 // and source scalar types:
1528 // ashr (shl (zext X), C), C --> sext X
1530 if (match(Op0, m_Shl(m_ZExt(m_Value(X)), m_Specific(Op1))) &&
1531 ShAmt == BitWidth - X->getType()->getScalarSizeInBits())
1532 return new SExtInst(X, Ty);
1534 // We can't handle (X << C1) >>s C2. It shifts arbitrary bits in. However,
1535 // we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits.
1537 if (match(Op0, m_NSWShl(m_Value(X), m_APInt(ShOp1))) &&
1538 ShOp1->ult(BitWidth)) {
1539 unsigned ShlAmt = ShOp1->getZExtValue();
1540 if (ShlAmt < ShAmt) {
1541 // (X <<nsw C1) >>s C2 --> X >>s (C2 - C1)
1542 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt);
1543 auto *NewAShr = BinaryOperator::CreateAShr(X, ShiftDiff);
1544 NewAShr->setIsExact(I.isExact());
1547 if (ShlAmt > ShAmt) {
1548 // (X <<nsw C1) >>s C2 --> X <<nsw (C1 - C2)
1549 Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt);
1550 auto *NewShl = BinaryOperator::Create(Instruction::Shl, X, ShiftDiff);
1551 NewShl->setHasNoSignedWrap(true);
1556 if (match(Op0, m_AShr(m_Value(X), m_APInt(ShOp1))) &&
1557 ShOp1->ult(BitWidth)) {
1558 unsigned AmtSum = ShAmt + ShOp1->getZExtValue();
1559 // Oversized arithmetic shifts replicate the sign bit.
1560 AmtSum = std::min(AmtSum, BitWidth - 1);
1561 // (X >>s C1) >>s C2 --> X >>s (C1 + C2)
1562 return BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
1565 if (match(Op0, m_OneUse(m_SExt(m_Value(X)))) &&
1566 (Ty->isVectorTy() || shouldChangeType(Ty, X->getType()))) {
1567 // ashr (sext X), C --> sext (ashr X, C')
1568 Type *SrcTy = X->getType();
1569 ShAmt = std::min(ShAmt, SrcTy->getScalarSizeInBits() - 1);
1570 Value *NewSh = Builder.CreateAShr(X, ConstantInt::get(SrcTy, ShAmt));
1571 return new SExtInst(NewSh, Ty);
1574 if (ShAmt == BitWidth - 1) {
1575 // ashr i32 or(X,-X), 31 --> sext (X != 0)
1576 if (match(Op0, m_OneUse(m_c_Or(m_Neg(m_Value(X)), m_Deferred(X)))))
1577 return new SExtInst(Builder.CreateIsNotNull(X), Ty);
1579 // ashr i32 (X -nsw Y), 31 --> sext (X < Y)
1581 if (match(Op0, m_OneUse(m_NSWSub(m_Value(X), m_Value(Y)))))
1582 return new SExtInst(Builder.CreateICmpSLT(X, Y), Ty);
1585 // If the shifted-out value is known-zero, then this is an exact shift.
1587 MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmt), 0, &I)) {
1593 // Prefer `-(x & 1)` over `(x << (bitwidth(x)-1)) a>> (bitwidth(x)-1)`
1594 // as the pattern to splat the lowest bit.
1595 // FIXME: iff X is already masked, we don't need the one-use check.
1597 if (match(Op1, m_SpecificIntAllowUndef(BitWidth - 1)) &&
1598 match(Op0, m_OneUse(m_Shl(m_Value(X),
1599 m_SpecificIntAllowUndef(BitWidth - 1))))) {
1600 Constant *Mask = ConstantInt::get(Ty, 1);
1601 // Retain the knowledge about the ignored lanes.
1602 Mask = Constant::mergeUndefsWith(
1603 Constant::mergeUndefsWith(Mask, cast<Constant>(Op1)),
1604 cast<Constant>(cast<Instruction>(Op0)->getOperand(1)));
1605 X = Builder.CreateAnd(X, Mask);
1606 return BinaryOperator::CreateNeg(X);
1609 if (Instruction *R = foldVariableSignZeroExtensionOfVariableHighBitExtract(I))
1612 // See if we can turn a signed shr into an unsigned shr.
1613 if (MaskedValueIsZero(Op0, APInt::getSignMask(BitWidth), 0, &I)) {
1614 Instruction *Lshr = BinaryOperator::CreateLShr(Op0, Op1);
1615 Lshr->setIsExact(I.isExact());
1619 // ashr (xor %x, -1), %y --> xor (ashr %x, %y), -1
1620 if (match(Op0, m_OneUse(m_Not(m_Value(X))))) {
1621 // Note that we must drop 'exact'-ness of the shift!
1622 // Note that we can't keep undef's in -1 vector constant!
1623 auto *NewAShr = Builder.CreateAShr(X, Op1, Op0->getName() + ".not");
1624 return BinaryOperator::CreateNot(NewAShr);