1 //===- InstCombineShifts.cpp ----------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visitShl, visitLShr, and visitAShr functions.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombineInternal.h"
15 #include "llvm/Analysis/ConstantFolding.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/IR/IntrinsicInst.h"
18 #include "llvm/IR/PatternMatch.h"
20 using namespace PatternMatch;
22 #define DEBUG_TYPE "instcombine"
24 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
25 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
26 assert(Op0->getType() == Op1->getType());
28 // See if we can fold away this shift.
29 if (SimplifyDemandedInstructionBits(I))
32 // Try to fold constant and into select arguments.
33 if (isa<Constant>(Op0))
34 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
35 if (Instruction *R = FoldOpIntoSelect(I, SI))
38 if (Constant *CUI = dyn_cast<Constant>(Op1))
39 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
42 // (C1 shift (A add C2)) -> (C1 shift C2) shift A)
43 // iff A and C2 are both positive.
46 if (match(Op0, m_Constant()) && match(Op1, m_Add(m_Value(A), m_Constant(C))))
47 if (isKnownNonNegative(A, DL, 0, &AC, &I, &DT) &&
48 isKnownNonNegative(C, DL, 0, &AC, &I, &DT))
49 return BinaryOperator::Create(
50 I.getOpcode(), Builder.CreateBinOp(I.getOpcode(), Op0, C), A);
52 // X shift (A srem B) -> X shift (A and B-1) iff B is a power of 2.
53 // Because shifts by negative values (which could occur if A were negative)
56 if (Op1->hasOneUse() && match(Op1, m_SRem(m_Value(A), m_Power2(B)))) {
57 // FIXME: Should this get moved into SimplifyDemandedBits by saying we don't
58 // demand the sign bit (and many others) here??
59 Value *Rem = Builder.CreateAnd(A, ConstantInt::get(I.getType(), *B - 1),
68 /// Return true if we can simplify two logical (either left or right) shifts
69 /// that have constant shift amounts: OuterShift (InnerShift X, C1), C2.
70 static bool canEvaluateShiftedShift(unsigned OuterShAmt, bool IsOuterShl,
71 Instruction *InnerShift, InstCombiner &IC,
73 assert(InnerShift->isLogicalShift() && "Unexpected instruction type");
75 // We need constant scalar or constant splat shifts.
76 const APInt *InnerShiftConst;
77 if (!match(InnerShift->getOperand(1), m_APInt(InnerShiftConst)))
80 // Two logical shifts in the same direction:
81 // shl (shl X, C1), C2 --> shl X, C1 + C2
82 // lshr (lshr X, C1), C2 --> lshr X, C1 + C2
83 bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl;
84 if (IsInnerShl == IsOuterShl)
87 // Equal shift amounts in opposite directions become bitwise 'and':
88 // lshr (shl X, C), C --> and X, C'
89 // shl (lshr X, C), C --> and X, C'
90 unsigned InnerShAmt = InnerShiftConst->getZExtValue();
91 if (InnerShAmt == OuterShAmt)
94 // If the 2nd shift is bigger than the 1st, we can fold:
95 // lshr (shl X, C1), C2 --> and (shl X, C1 - C2), C3
96 // shl (lshr X, C1), C2 --> and (lshr X, C1 - C2), C3
97 // but it isn't profitable unless we know the and'd out bits are already zero.
98 // Also, check that the inner shift is valid (less than the type width) or
99 // we'll crash trying to produce the bit mask for the 'and'.
100 unsigned TypeWidth = InnerShift->getType()->getScalarSizeInBits();
101 if (InnerShAmt > OuterShAmt && InnerShAmt < TypeWidth) {
103 IsInnerShl ? TypeWidth - InnerShAmt : InnerShAmt - OuterShAmt;
104 APInt Mask = APInt::getLowBitsSet(TypeWidth, OuterShAmt) << MaskShift;
105 if (IC.MaskedValueIsZero(InnerShift->getOperand(0), Mask, 0, CxtI))
112 /// See if we can compute the specified value, but shifted logically to the left
113 /// or right by some number of bits. This should return true if the expression
114 /// can be computed for the same cost as the current expression tree. This is
115 /// used to eliminate extraneous shifting from things like:
116 /// %C = shl i128 %A, 64
117 /// %D = shl i128 %B, 96
118 /// %E = or i128 %C, %D
119 /// %F = lshr i128 %E, 64
120 /// where the client will ask if E can be computed shifted right by 64-bits. If
121 /// this succeeds, getShiftedValue() will be called to produce the value.
122 static bool canEvaluateShifted(Value *V, unsigned NumBits, bool IsLeftShift,
123 InstCombiner &IC, Instruction *CxtI) {
124 // We can always evaluate constants shifted.
125 if (isa<Constant>(V))
128 Instruction *I = dyn_cast<Instruction>(V);
129 if (!I) return false;
131 // If this is the opposite shift, we can directly reuse the input of the shift
132 // if the needed bits are already zero in the input. This allows us to reuse
133 // the value which means that we don't care if the shift has multiple uses.
134 // TODO: Handle opposite shift by exact value.
135 ConstantInt *CI = nullptr;
136 if ((IsLeftShift && match(I, m_LShr(m_Value(), m_ConstantInt(CI)))) ||
137 (!IsLeftShift && match(I, m_Shl(m_Value(), m_ConstantInt(CI))))) {
138 if (CI->getZExtValue() == NumBits) {
139 // TODO: Check that the input bits are already zero with MaskedValueIsZero
141 // If this is a truncate of a logical shr, we can truncate it to a smaller
142 // lshr iff we know that the bits we would otherwise be shifting in are
144 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
145 uint32_t BitWidth = Ty->getScalarSizeInBits();
146 if (MaskedValueIsZero(I->getOperand(0),
147 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
148 CI->getLimitedValue(BitWidth) < BitWidth) {
149 return CanEvaluateTruncated(I->getOperand(0), Ty);
156 // We can't mutate something that has multiple uses: doing so would
157 // require duplicating the instruction in general, which isn't profitable.
158 if (!I->hasOneUse()) return false;
160 switch (I->getOpcode()) {
161 default: return false;
162 case Instruction::And:
163 case Instruction::Or:
164 case Instruction::Xor:
165 // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
166 return canEvaluateShifted(I->getOperand(0), NumBits, IsLeftShift, IC, I) &&
167 canEvaluateShifted(I->getOperand(1), NumBits, IsLeftShift, IC, I);
169 case Instruction::Shl:
170 case Instruction::LShr:
171 return canEvaluateShiftedShift(NumBits, IsLeftShift, I, IC, CxtI);
173 case Instruction::Select: {
174 SelectInst *SI = cast<SelectInst>(I);
175 Value *TrueVal = SI->getTrueValue();
176 Value *FalseVal = SI->getFalseValue();
177 return canEvaluateShifted(TrueVal, NumBits, IsLeftShift, IC, SI) &&
178 canEvaluateShifted(FalseVal, NumBits, IsLeftShift, IC, SI);
180 case Instruction::PHI: {
181 // We can change a phi if we can change all operands. Note that we never
182 // get into trouble with cyclic PHIs here because we only consider
183 // instructions with a single use.
184 PHINode *PN = cast<PHINode>(I);
185 for (Value *IncValue : PN->incoming_values())
186 if (!canEvaluateShifted(IncValue, NumBits, IsLeftShift, IC, PN))
193 /// Fold OuterShift (InnerShift X, C1), C2.
194 /// See canEvaluateShiftedShift() for the constraints on these instructions.
195 static Value *foldShiftedShift(BinaryOperator *InnerShift, unsigned OuterShAmt,
197 InstCombiner::BuilderTy &Builder) {
198 bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl;
199 Type *ShType = InnerShift->getType();
200 unsigned TypeWidth = ShType->getScalarSizeInBits();
202 // We only accept shifts-by-a-constant in canEvaluateShifted().
204 match(InnerShift->getOperand(1), m_APInt(C1));
205 unsigned InnerShAmt = C1->getZExtValue();
207 // Change the shift amount and clear the appropriate IR flags.
208 auto NewInnerShift = [&](unsigned ShAmt) {
209 InnerShift->setOperand(1, ConstantInt::get(ShType, ShAmt));
211 InnerShift->setHasNoUnsignedWrap(false);
212 InnerShift->setHasNoSignedWrap(false);
214 InnerShift->setIsExact(false);
219 // Two logical shifts in the same direction:
220 // shl (shl X, C1), C2 --> shl X, C1 + C2
221 // lshr (lshr X, C1), C2 --> lshr X, C1 + C2
222 if (IsInnerShl == IsOuterShl) {
223 // If this is an oversized composite shift, then unsigned shifts get 0.
224 if (InnerShAmt + OuterShAmt >= TypeWidth)
225 return Constant::getNullValue(ShType);
227 return NewInnerShift(InnerShAmt + OuterShAmt);
230 // Equal shift amounts in opposite directions become bitwise 'and':
231 // lshr (shl X, C), C --> and X, C'
232 // shl (lshr X, C), C --> and X, C'
233 if (InnerShAmt == OuterShAmt) {
234 APInt Mask = IsInnerShl
235 ? APInt::getLowBitsSet(TypeWidth, TypeWidth - OuterShAmt)
236 : APInt::getHighBitsSet(TypeWidth, TypeWidth - OuterShAmt);
237 Value *And = Builder.CreateAnd(InnerShift->getOperand(0),
238 ConstantInt::get(ShType, Mask));
239 if (auto *AndI = dyn_cast<Instruction>(And)) {
240 AndI->moveBefore(InnerShift);
241 AndI->takeName(InnerShift);
246 assert(InnerShAmt > OuterShAmt &&
247 "Unexpected opposite direction logical shift pair");
249 // In general, we would need an 'and' for this transform, but
250 // canEvaluateShiftedShift() guarantees that the masked-off bits are not used.
251 // lshr (shl X, C1), C2 --> shl X, C1 - C2
252 // shl (lshr X, C1), C2 --> lshr X, C1 - C2
253 return NewInnerShift(InnerShAmt - OuterShAmt);
256 /// When canEvaluateShifted() returns true for an expression, this function
257 /// inserts the new computation that produces the shifted value.
258 static Value *getShiftedValue(Value *V, unsigned NumBits, bool isLeftShift,
259 InstCombiner &IC, const DataLayout &DL) {
260 // We can always evaluate constants shifted.
261 if (Constant *C = dyn_cast<Constant>(V)) {
263 V = IC.Builder.CreateShl(C, NumBits);
265 V = IC.Builder.CreateLShr(C, NumBits);
266 // If we got a constantexpr back, try to simplify it with TD info.
267 if (auto *C = dyn_cast<Constant>(V))
269 ConstantFoldConstant(C, DL, &IC.getTargetLibraryInfo()))
274 Instruction *I = cast<Instruction>(V);
277 switch (I->getOpcode()) {
278 default: llvm_unreachable("Inconsistency with CanEvaluateShifted");
279 case Instruction::And:
280 case Instruction::Or:
281 case Instruction::Xor:
282 // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
284 0, getShiftedValue(I->getOperand(0), NumBits, isLeftShift, IC, DL));
286 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL));
289 case Instruction::Shl:
290 case Instruction::LShr:
291 return foldShiftedShift(cast<BinaryOperator>(I), NumBits, isLeftShift,
294 case Instruction::Select:
296 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL));
298 2, getShiftedValue(I->getOperand(2), NumBits, isLeftShift, IC, DL));
300 case Instruction::PHI: {
301 // We can change a phi if we can change all operands. Note that we never
302 // get into trouble with cyclic PHIs here because we only consider
303 // instructions with a single use.
304 PHINode *PN = cast<PHINode>(I);
305 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
306 PN->setIncomingValue(i, getShiftedValue(PN->getIncomingValue(i), NumBits,
307 isLeftShift, IC, DL));
313 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, Constant *Op1,
315 bool isLeftShift = I.getOpcode() == Instruction::Shl;
318 if (!match(Op1, m_APInt(Op1C)))
321 // See if we can propagate this shift into the input, this covers the trivial
322 // cast of lshr(shl(x,c1),c2) as well as other more complex cases.
323 if (I.getOpcode() != Instruction::AShr &&
324 canEvaluateShifted(Op0, Op1C->getZExtValue(), isLeftShift, *this, &I)) {
325 DEBUG(dbgs() << "ICE: GetShiftedValue propagating shift through expression"
326 " to eliminate shift:\n IN: " << *Op0 << "\n SH: " << I <<"\n");
328 return replaceInstUsesWith(
329 I, getShiftedValue(Op0, Op1C->getZExtValue(), isLeftShift, *this, DL));
332 // See if we can simplify any instructions used by the instruction whose sole
333 // purpose is to compute bits we don't care about.
334 unsigned TypeBits = Op0->getType()->getScalarSizeInBits();
336 assert(!Op1C->uge(TypeBits) &&
337 "Shift over the type width should have been removed already");
339 if (Instruction *FoldedShift = foldOpWithConstantIntoOperand(I))
342 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
343 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
344 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
345 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
346 // place. Don't try to do this transformation in this case. Also, we
347 // require that the input operand is a shift-by-constant so that we have
348 // confidence that the shifts will get folded together. We could do this
349 // xform in more cases, but it is unlikely to be profitable.
350 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
351 isa<ConstantInt>(TrOp->getOperand(1))) {
352 // Okay, we'll do this xform. Make the shift of shift.
354 ConstantExpr::getZExt(cast<Constant>(Op1), TrOp->getType());
355 // (shift2 (shift1 & 0x00FF), c2)
356 Value *NSh = Builder.CreateBinOp(I.getOpcode(), TrOp, ShAmt, I.getName());
358 // For logical shifts, the truncation has the effect of making the high
359 // part of the register be zeros. Emulate this by inserting an AND to
360 // clear the top bits as needed. This 'and' will usually be zapped by
361 // other xforms later if dead.
362 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
363 unsigned DstSize = TI->getType()->getScalarSizeInBits();
364 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
366 // The mask we constructed says what the trunc would do if occurring
367 // between the shifts. We want to know the effect *after* the second
368 // shift. We know that it is a logical shift by a constant, so adjust the
369 // mask as appropriate.
370 if (I.getOpcode() == Instruction::Shl)
371 MaskV <<= Op1C->getZExtValue();
373 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
374 MaskV.lshrInPlace(Op1C->getZExtValue());
378 Value *And = Builder.CreateAnd(NSh,
379 ConstantInt::get(I.getContext(), MaskV),
382 // Return the value truncated to the interesting size.
383 return new TruncInst(And, I.getType());
387 if (Op0->hasOneUse()) {
388 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
389 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
392 switch (Op0BO->getOpcode()) {
394 case Instruction::Add:
395 case Instruction::And:
396 case Instruction::Or:
397 case Instruction::Xor: {
398 // These operators commute.
399 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
400 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
401 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
403 Value *YS = // (Y << C)
404 Builder.CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
406 Value *X = Builder.CreateBinOp(Op0BO->getOpcode(), YS, V1,
407 Op0BO->getOperand(1)->getName());
408 unsigned Op1Val = Op1C->getLimitedValue(TypeBits);
410 APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val);
411 Constant *Mask = ConstantInt::get(I.getContext(), Bits);
412 if (VectorType *VT = dyn_cast<VectorType>(X->getType()))
413 Mask = ConstantVector::getSplat(VT->getNumElements(), Mask);
414 return BinaryOperator::CreateAnd(X, Mask);
417 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
418 Value *Op0BOOp1 = Op0BO->getOperand(1);
419 if (isLeftShift && Op0BOOp1->hasOneUse() &&
421 m_And(m_OneUse(m_Shr(m_Value(V1), m_Specific(Op1))),
422 m_ConstantInt(CC)))) {
423 Value *YS = // (Y << C)
424 Builder.CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
426 Value *XM = Builder.CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
427 V1->getName()+".mask");
428 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
433 case Instruction::Sub: {
434 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
435 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
436 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
438 Value *YS = // (Y << C)
439 Builder.CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
441 Value *X = Builder.CreateBinOp(Op0BO->getOpcode(), V1, YS,
442 Op0BO->getOperand(0)->getName());
443 unsigned Op1Val = Op1C->getLimitedValue(TypeBits);
445 APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val);
446 Constant *Mask = ConstantInt::get(I.getContext(), Bits);
447 if (VectorType *VT = dyn_cast<VectorType>(X->getType()))
448 Mask = ConstantVector::getSplat(VT->getNumElements(), Mask);
449 return BinaryOperator::CreateAnd(X, Mask);
452 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
453 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
454 match(Op0BO->getOperand(0),
455 m_And(m_OneUse(m_Shr(m_Value(V1), m_Value(V2))),
456 m_ConstantInt(CC))) && V2 == Op1) {
457 Value *YS = // (Y << C)
458 Builder.CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
460 Value *XM = Builder.CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
461 V1->getName()+".mask");
463 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
471 // If the operand is a bitwise operator with a constant RHS, and the
472 // shift is the only use, we can pull it out of the shift.
473 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
474 bool isValid = true; // Valid only for And, Or, Xor
475 bool highBitSet = false; // Transform if high bit of constant set?
477 switch (Op0BO->getOpcode()) {
478 default: isValid = false; break; // Do not perform transform!
479 case Instruction::Add:
480 isValid = isLeftShift;
482 case Instruction::Or:
483 case Instruction::Xor:
486 case Instruction::And:
491 // If this is a signed shift right, and the high bit is modified
492 // by the logical operation, do not perform the transformation.
493 // The highBitSet boolean indicates the value of the high bit of
494 // the constant which would cause it to be modified for this
497 if (isValid && I.getOpcode() == Instruction::AShr)
498 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
501 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
504 Builder.CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
505 NewShift->takeName(Op0BO);
507 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
517 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
518 if (Value *V = SimplifyVectorOp(I))
519 return replaceInstUsesWith(I, V);
521 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
523 SimplifyShlInst(Op0, Op1, I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
524 SQ.getWithInstruction(&I)))
525 return replaceInstUsesWith(I, V);
527 if (Instruction *V = commonShiftTransforms(I))
530 const APInt *ShAmtAPInt;
531 if (match(Op1, m_APInt(ShAmtAPInt))) {
532 unsigned ShAmt = ShAmtAPInt->getZExtValue();
533 unsigned BitWidth = I.getType()->getScalarSizeInBits();
534 Type *Ty = I.getType();
536 // shl (zext X), ShAmt --> zext (shl X, ShAmt)
537 // This is only valid if X would have zeros shifted out.
539 if (match(Op0, m_ZExt(m_Value(X)))) {
540 unsigned SrcWidth = X->getType()->getScalarSizeInBits();
541 if (ShAmt < SrcWidth &&
542 MaskedValueIsZero(X, APInt::getHighBitsSet(SrcWidth, ShAmt), 0, &I))
543 return new ZExtInst(Builder.CreateShl(X, ShAmt), Ty);
546 // (X >>u C) << C --> X & (-1 << C)
547 if (match(Op0, m_LShr(m_Value(X), m_Specific(Op1)))) {
548 APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt));
549 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask));
552 // Be careful about hiding shl instructions behind bit masks. They are used
553 // to represent multiplies by a constant, and it is important that simple
554 // arithmetic expressions are still recognizable by scalar evolution.
555 // The inexact versions are deferred to DAGCombine, so we don't hide shl
556 // behind a bit mask.
558 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_APInt(ShOp1))))) {
559 unsigned ShrAmt = ShOp1->getZExtValue();
560 if (ShrAmt < ShAmt) {
561 // If C1 < C2: (X >>?,exact C1) << C2 --> X << (C2 - C1)
562 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShrAmt);
563 auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
564 NewShl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
565 NewShl->setHasNoSignedWrap(I.hasNoSignedWrap());
568 if (ShrAmt > ShAmt) {
569 // If C1 > C2: (X >>?exact C1) << C2 --> X >>?exact (C1 - C2)
570 Constant *ShiftDiff = ConstantInt::get(Ty, ShrAmt - ShAmt);
571 auto *NewShr = BinaryOperator::Create(
572 cast<BinaryOperator>(Op0)->getOpcode(), X, ShiftDiff);
573 NewShr->setIsExact(true);
578 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShOp1)))) {
579 unsigned AmtSum = ShAmt + ShOp1->getZExtValue();
580 // Oversized shifts are simplified to zero in InstSimplify.
581 if (AmtSum < BitWidth)
582 // (X << C1) << C2 --> X << (C1 + C2)
583 return BinaryOperator::CreateShl(X, ConstantInt::get(Ty, AmtSum));
586 // If the shifted-out value is known-zero, then this is a NUW shift.
587 if (!I.hasNoUnsignedWrap() &&
588 MaskedValueIsZero(Op0, APInt::getHighBitsSet(BitWidth, ShAmt), 0, &I)) {
589 I.setHasNoUnsignedWrap();
593 // If the shifted-out value is all signbits, then this is a NSW shift.
594 if (!I.hasNoSignedWrap() && ComputeNumSignBits(Op0, 0, &I) > ShAmt) {
595 I.setHasNoSignedWrap();
601 if (match(Op1, m_Constant(C1))) {
604 // (C2 << X) << C1 --> (C2 << C1) << X
605 if (match(Op0, m_OneUse(m_Shl(m_Constant(C2), m_Value(X)))))
606 return BinaryOperator::CreateShl(ConstantExpr::getShl(C2, C1), X);
608 // (X * C2) << C1 --> X * (C2 << C1)
609 if (match(Op0, m_Mul(m_Value(X), m_Constant(C2))))
610 return BinaryOperator::CreateMul(X, ConstantExpr::getShl(C2, C1));
616 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
617 if (Value *V = SimplifyVectorOp(I))
618 return replaceInstUsesWith(I, V);
620 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
622 SimplifyLShrInst(Op0, Op1, I.isExact(), SQ.getWithInstruction(&I)))
623 return replaceInstUsesWith(I, V);
625 if (Instruction *R = commonShiftTransforms(I))
628 Type *Ty = I.getType();
629 const APInt *ShAmtAPInt;
630 if (match(Op1, m_APInt(ShAmtAPInt))) {
631 unsigned ShAmt = ShAmtAPInt->getZExtValue();
632 unsigned BitWidth = Ty->getScalarSizeInBits();
633 auto *II = dyn_cast<IntrinsicInst>(Op0);
634 if (II && isPowerOf2_32(BitWidth) && Log2_32(BitWidth) == ShAmt &&
635 (II->getIntrinsicID() == Intrinsic::ctlz ||
636 II->getIntrinsicID() == Intrinsic::cttz ||
637 II->getIntrinsicID() == Intrinsic::ctpop)) {
638 // ctlz.i32(x)>>5 --> zext(x == 0)
639 // cttz.i32(x)>>5 --> zext(x == 0)
640 // ctpop.i32(x)>>5 --> zext(x == -1)
641 bool IsPop = II->getIntrinsicID() == Intrinsic::ctpop;
642 Constant *RHS = ConstantInt::getSigned(Ty, IsPop ? -1 : 0);
643 Value *Cmp = Builder.CreateICmpEQ(II->getArgOperand(0), RHS);
644 return new ZExtInst(Cmp, Ty);
649 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShOp1)))) {
650 unsigned ShlAmt = ShOp1->getZExtValue();
651 if (ShlAmt < ShAmt) {
652 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt);
653 if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) {
654 // (X <<nuw C1) >>u C2 --> X >>u (C2 - C1)
655 auto *NewLShr = BinaryOperator::CreateLShr(X, ShiftDiff);
656 NewLShr->setIsExact(I.isExact());
659 // (X << C1) >>u C2 --> (X >>u (C2 - C1)) & (-1 >> C2)
660 Value *NewLShr = Builder.CreateLShr(X, ShiftDiff, "", I.isExact());
661 APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt));
662 return BinaryOperator::CreateAnd(NewLShr, ConstantInt::get(Ty, Mask));
664 if (ShlAmt > ShAmt) {
665 Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt);
666 if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) {
667 // (X <<nuw C1) >>u C2 --> X <<nuw (C1 - C2)
668 auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
669 NewShl->setHasNoUnsignedWrap(true);
672 // (X << C1) >>u C2 --> X << (C1 - C2) & (-1 >> C2)
673 Value *NewShl = Builder.CreateShl(X, ShiftDiff);
674 APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt));
675 return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask));
677 assert(ShlAmt == ShAmt);
678 // (X << C) >>u C --> X & (-1 >>u C)
679 APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt));
680 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask));
683 if (match(Op0, m_SExt(m_Value(X))) &&
684 (!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType()))) {
685 // Are we moving the sign bit to the low bit and widening with high zeros?
686 unsigned SrcTyBitWidth = X->getType()->getScalarSizeInBits();
687 if (ShAmt == BitWidth - 1) {
688 // lshr (sext i1 X to iN), N-1 --> zext X to iN
689 if (SrcTyBitWidth == 1)
690 return new ZExtInst(X, Ty);
692 // lshr (sext iM X to iN), N-1 --> zext (lshr X, M-1) to iN
693 if (Op0->hasOneUse()) {
694 Value *NewLShr = Builder.CreateLShr(X, SrcTyBitWidth - 1);
695 return new ZExtInst(NewLShr, Ty);
699 // lshr (sext iM X to iN), N-M --> zext (ashr X, min(N-M, M-1)) to iN
700 if (ShAmt == BitWidth - SrcTyBitWidth && Op0->hasOneUse()) {
701 // The new shift amount can't be more than the narrow source type.
702 unsigned NewShAmt = std::min(ShAmt, SrcTyBitWidth - 1);
703 Value *AShr = Builder.CreateAShr(X, NewShAmt);
704 return new ZExtInst(AShr, Ty);
708 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShOp1)))) {
709 unsigned AmtSum = ShAmt + ShOp1->getZExtValue();
710 // Oversized shifts are simplified to zero in InstSimplify.
711 if (AmtSum < BitWidth)
712 // (X >>u C1) >>u C2 --> X >>u (C1 + C2)
713 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
716 // If the shifted-out value is known-zero, then this is an exact shift.
718 MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmt), 0, &I)) {
726 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
727 if (Value *V = SimplifyVectorOp(I))
728 return replaceInstUsesWith(I, V);
730 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
732 SimplifyAShrInst(Op0, Op1, I.isExact(), SQ.getWithInstruction(&I)))
733 return replaceInstUsesWith(I, V);
735 if (Instruction *R = commonShiftTransforms(I))
738 Type *Ty = I.getType();
739 unsigned BitWidth = Ty->getScalarSizeInBits();
740 const APInt *ShAmtAPInt;
741 if (match(Op1, m_APInt(ShAmtAPInt))) {
742 unsigned ShAmt = ShAmtAPInt->getZExtValue();
744 // If the shift amount equals the difference in width of the destination
745 // and source scalar types:
746 // ashr (shl (zext X), C), C --> sext X
748 if (match(Op0, m_Shl(m_ZExt(m_Value(X)), m_Specific(Op1))) &&
749 ShAmt == BitWidth - X->getType()->getScalarSizeInBits())
750 return new SExtInst(X, Ty);
752 // We can't handle (X << C1) >>s C2. It shifts arbitrary bits in. However,
753 // we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits.
755 if (match(Op0, m_NSWShl(m_Value(X), m_APInt(ShOp1)))) {
756 unsigned ShlAmt = ShOp1->getZExtValue();
757 if (ShlAmt < ShAmt) {
758 // (X <<nsw C1) >>s C2 --> X >>s (C2 - C1)
759 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt);
760 auto *NewAShr = BinaryOperator::CreateAShr(X, ShiftDiff);
761 NewAShr->setIsExact(I.isExact());
764 if (ShlAmt > ShAmt) {
765 // (X <<nsw C1) >>s C2 --> X <<nsw (C1 - C2)
766 Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt);
767 auto *NewShl = BinaryOperator::Create(Instruction::Shl, X, ShiftDiff);
768 NewShl->setHasNoSignedWrap(true);
773 if (match(Op0, m_AShr(m_Value(X), m_APInt(ShOp1)))) {
774 unsigned AmtSum = ShAmt + ShOp1->getZExtValue();
775 // Oversized arithmetic shifts replicate the sign bit.
776 AmtSum = std::min(AmtSum, BitWidth - 1);
777 // (X >>s C1) >>s C2 --> X >>s (C1 + C2)
778 return BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
781 // If the shifted-out value is known-zero, then this is an exact shift.
783 MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmt), 0, &I)) {
789 // See if we can turn a signed shr into an unsigned shr.
790 if (MaskedValueIsZero(Op0, APInt::getSignMask(BitWidth), 0, &I))
791 return BinaryOperator::CreateLShr(Op0, Op1);