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 #include "llvm/Transforms/InstCombine/InstCombiner.h"
19 using namespace llvm;
20 using namespace PatternMatch;
21
22 #define DEBUG_TYPE "instcombine"
23
canTryToConstantAddTwoShiftAmounts(Value * Sh0,Value * ShAmt0,Value * Sh1,Value * ShAmt1)24 bool canTryToConstantAddTwoShiftAmounts(Value *Sh0, Value *ShAmt0, Value *Sh1,
25 Value *ShAmt1) {
26 // We have two shift amounts from two different shifts. The types of those
27 // shift amounts may not match. If that's the case let's bailout now..
28 if (ShAmt0->getType() != ShAmt1->getType())
29 return false;
30
31 // As input, we have the following pattern:
32 // Sh0 (Sh1 X, Q), K
33 // We want to rewrite that as:
34 // Sh x, (Q+K) iff (Q+K) u< bitwidth(x)
35 // While we know that originally (Q+K) would not overflow
36 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
37 // shift amounts. so it may now overflow in smaller bitwidth.
38 // To ensure that does not happen, we need to ensure that the total maximal
39 // shift amount is still representable in that smaller bit width.
40 unsigned MaximalPossibleTotalShiftAmount =
41 (Sh0->getType()->getScalarSizeInBits() - 1) +
42 (Sh1->getType()->getScalarSizeInBits() - 1);
43 APInt MaximalRepresentableShiftAmount =
44 APInt::getAllOnesValue(ShAmt0->getType()->getScalarSizeInBits());
45 return MaximalRepresentableShiftAmount.uge(MaximalPossibleTotalShiftAmount);
46 }
47
48 // Given pattern:
49 // (x shiftopcode Q) shiftopcode K
50 // we should rewrite it as
51 // x shiftopcode (Q+K) iff (Q+K) u< bitwidth(x) and
52 //
53 // This is valid for any shift, but they must be identical, and we must be
54 // careful in case we have (zext(Q)+zext(K)) and look past extensions,
55 // (Q+K) must not overflow or else (Q+K) u< bitwidth(x) is bogus.
56 //
57 // AnalyzeForSignBitExtraction indicates that we will only analyze whether this
58 // pattern has any 2 right-shifts that sum to 1 less than original bit width.
reassociateShiftAmtsOfTwoSameDirectionShifts(BinaryOperator * Sh0,const SimplifyQuery & SQ,bool AnalyzeForSignBitExtraction)59 Value *InstCombinerImpl::reassociateShiftAmtsOfTwoSameDirectionShifts(
60 BinaryOperator *Sh0, const SimplifyQuery &SQ,
61 bool AnalyzeForSignBitExtraction) {
62 // Look for a shift of some instruction, ignore zext of shift amount if any.
63 Instruction *Sh0Op0;
64 Value *ShAmt0;
65 if (!match(Sh0,
66 m_Shift(m_Instruction(Sh0Op0), m_ZExtOrSelf(m_Value(ShAmt0)))))
67 return nullptr;
68
69 // If there is a truncation between the two shifts, we must make note of it
70 // and look through it. The truncation imposes additional constraints on the
71 // transform.
72 Instruction *Sh1;
73 Value *Trunc = nullptr;
74 match(Sh0Op0,
75 m_CombineOr(m_CombineAnd(m_Trunc(m_Instruction(Sh1)), m_Value(Trunc)),
76 m_Instruction(Sh1)));
77
78 // Inner shift: (x shiftopcode ShAmt1)
79 // Like with other shift, ignore zext of shift amount if any.
80 Value *X, *ShAmt1;
81 if (!match(Sh1, m_Shift(m_Value(X), m_ZExtOrSelf(m_Value(ShAmt1)))))
82 return nullptr;
83
84 // Verify that it would be safe to try to add those two shift amounts.
85 if (!canTryToConstantAddTwoShiftAmounts(Sh0, ShAmt0, Sh1, ShAmt1))
86 return nullptr;
87
88 // We are only looking for signbit extraction if we have two right shifts.
89 bool HadTwoRightShifts = match(Sh0, m_Shr(m_Value(), m_Value())) &&
90 match(Sh1, m_Shr(m_Value(), m_Value()));
91 // ... and if it's not two right-shifts, we know the answer already.
92 if (AnalyzeForSignBitExtraction && !HadTwoRightShifts)
93 return nullptr;
94
95 // The shift opcodes must be identical, unless we are just checking whether
96 // this pattern can be interpreted as a sign-bit-extraction.
97 Instruction::BinaryOps ShiftOpcode = Sh0->getOpcode();
98 bool IdenticalShOpcodes = Sh0->getOpcode() == Sh1->getOpcode();
99 if (!IdenticalShOpcodes && !AnalyzeForSignBitExtraction)
100 return nullptr;
101
102 // If we saw truncation, we'll need to produce extra instruction,
103 // and for that one of the operands of the shift must be one-use,
104 // unless of course we don't actually plan to produce any instructions here.
105 if (Trunc && !AnalyzeForSignBitExtraction &&
106 !match(Sh0, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
107 return nullptr;
108
109 // Can we fold (ShAmt0+ShAmt1) ?
110 auto *NewShAmt = dyn_cast_or_null<Constant>(
111 SimplifyAddInst(ShAmt0, ShAmt1, /*isNSW=*/false, /*isNUW=*/false,
112 SQ.getWithInstruction(Sh0)));
113 if (!NewShAmt)
114 return nullptr; // Did not simplify.
115 unsigned NewShAmtBitWidth = NewShAmt->getType()->getScalarSizeInBits();
116 unsigned XBitWidth = X->getType()->getScalarSizeInBits();
117 // Is the new shift amount smaller than the bit width of inner/new shift?
118 if (!match(NewShAmt, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
119 APInt(NewShAmtBitWidth, XBitWidth))))
120 return nullptr; // FIXME: could perform constant-folding.
121
122 // If there was a truncation, and we have a right-shift, we can only fold if
123 // we are left with the original sign bit. Likewise, if we were just checking
124 // that this is a sighbit extraction, this is the place to check it.
125 // FIXME: zero shift amount is also legal here, but we can't *easily* check
126 // more than one predicate so it's not really worth it.
127 if (HadTwoRightShifts && (Trunc || AnalyzeForSignBitExtraction)) {
128 // If it's not a sign bit extraction, then we're done.
129 if (!match(NewShAmt,
130 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
131 APInt(NewShAmtBitWidth, XBitWidth - 1))))
132 return nullptr;
133 // If it is, and that was the question, return the base value.
134 if (AnalyzeForSignBitExtraction)
135 return X;
136 }
137
138 assert(IdenticalShOpcodes && "Should not get here with different shifts.");
139
140 // All good, we can do this fold.
141 NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, X->getType());
142
143 BinaryOperator *NewShift = BinaryOperator::Create(ShiftOpcode, X, NewShAmt);
144
145 // The flags can only be propagated if there wasn't a trunc.
146 if (!Trunc) {
147 // If the pattern did not involve trunc, and both of the original shifts
148 // had the same flag set, preserve the flag.
149 if (ShiftOpcode == Instruction::BinaryOps::Shl) {
150 NewShift->setHasNoUnsignedWrap(Sh0->hasNoUnsignedWrap() &&
151 Sh1->hasNoUnsignedWrap());
152 NewShift->setHasNoSignedWrap(Sh0->hasNoSignedWrap() &&
153 Sh1->hasNoSignedWrap());
154 } else {
155 NewShift->setIsExact(Sh0->isExact() && Sh1->isExact());
156 }
157 }
158
159 Instruction *Ret = NewShift;
160 if (Trunc) {
161 Builder.Insert(NewShift);
162 Ret = CastInst::Create(Instruction::Trunc, NewShift, Sh0->getType());
163 }
164
165 return Ret;
166 }
167
168 // If we have some pattern that leaves only some low bits set, and then performs
169 // left-shift of those bits, if none of the bits that are left after the final
170 // shift are modified by the mask, we can omit the mask.
171 //
172 // There are many variants to this pattern:
173 // a) (x & ((1 << MaskShAmt) - 1)) << ShiftShAmt
174 // b) (x & (~(-1 << MaskShAmt))) << ShiftShAmt
175 // c) (x & (-1 >> MaskShAmt)) << ShiftShAmt
176 // d) (x & ((-1 << MaskShAmt) >> MaskShAmt)) << ShiftShAmt
177 // e) ((x << MaskShAmt) l>> MaskShAmt) << ShiftShAmt
178 // f) ((x << MaskShAmt) a>> MaskShAmt) << ShiftShAmt
179 // All these patterns can be simplified to just:
180 // x << ShiftShAmt
181 // iff:
182 // a,b) (MaskShAmt+ShiftShAmt) u>= bitwidth(x)
183 // c,d,e,f) (ShiftShAmt-MaskShAmt) s>= 0 (i.e. ShiftShAmt u>= MaskShAmt)
184 static Instruction *
dropRedundantMaskingOfLeftShiftInput(BinaryOperator * OuterShift,const SimplifyQuery & Q,InstCombiner::BuilderTy & Builder)185 dropRedundantMaskingOfLeftShiftInput(BinaryOperator *OuterShift,
186 const SimplifyQuery &Q,
187 InstCombiner::BuilderTy &Builder) {
188 assert(OuterShift->getOpcode() == Instruction::BinaryOps::Shl &&
189 "The input must be 'shl'!");
190
191 Value *Masked, *ShiftShAmt;
192 match(OuterShift,
193 m_Shift(m_Value(Masked), m_ZExtOrSelf(m_Value(ShiftShAmt))));
194
195 // *If* there is a truncation between an outer shift and a possibly-mask,
196 // then said truncation *must* be one-use, else we can't perform the fold.
197 Value *Trunc;
198 if (match(Masked, m_CombineAnd(m_Trunc(m_Value(Masked)), m_Value(Trunc))) &&
199 !Trunc->hasOneUse())
200 return nullptr;
201
202 Type *NarrowestTy = OuterShift->getType();
203 Type *WidestTy = Masked->getType();
204 bool HadTrunc = WidestTy != NarrowestTy;
205
206 // The mask must be computed in a type twice as wide to ensure
207 // that no bits are lost if the sum-of-shifts is wider than the base type.
208 Type *ExtendedTy = WidestTy->getExtendedType();
209
210 Value *MaskShAmt;
211
212 // ((1 << MaskShAmt) - 1)
213 auto MaskA = m_Add(m_Shl(m_One(), m_Value(MaskShAmt)), m_AllOnes());
214 // (~(-1 << maskNbits))
215 auto MaskB = m_Xor(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_AllOnes());
216 // (-1 >> MaskShAmt)
217 auto MaskC = m_Shr(m_AllOnes(), m_Value(MaskShAmt));
218 // ((-1 << MaskShAmt) >> MaskShAmt)
219 auto MaskD =
220 m_Shr(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_Deferred(MaskShAmt));
221
222 Value *X;
223 Constant *NewMask;
224
225 if (match(Masked, m_c_And(m_CombineOr(MaskA, MaskB), m_Value(X)))) {
226 // Peek through an optional zext of the shift amount.
227 match(MaskShAmt, m_ZExtOrSelf(m_Value(MaskShAmt)));
228
229 // Verify that it would be safe to try to add those two shift amounts.
230 if (!canTryToConstantAddTwoShiftAmounts(OuterShift, ShiftShAmt, Masked,
231 MaskShAmt))
232 return nullptr;
233
234 // Can we simplify (MaskShAmt+ShiftShAmt) ?
235 auto *SumOfShAmts = dyn_cast_or_null<Constant>(SimplifyAddInst(
236 MaskShAmt, ShiftShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q));
237 if (!SumOfShAmts)
238 return nullptr; // Did not simplify.
239 // In this pattern SumOfShAmts correlates with the number of low bits
240 // that shall remain in the root value (OuterShift).
241
242 // An extend of an undef value becomes zero because the high bits are never
243 // completely unknown. Replace the the `undef` shift amounts with final
244 // shift bitwidth to ensure that the value remains undef when creating the
245 // subsequent shift op.
246 SumOfShAmts = Constant::replaceUndefsWith(
247 SumOfShAmts, ConstantInt::get(SumOfShAmts->getType()->getScalarType(),
248 ExtendedTy->getScalarSizeInBits()));
249 auto *ExtendedSumOfShAmts = ConstantExpr::getZExt(SumOfShAmts, ExtendedTy);
250 // And compute the mask as usual: ~(-1 << (SumOfShAmts))
251 auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(ExtendedTy);
252 auto *ExtendedInvertedMask =
253 ConstantExpr::getShl(ExtendedAllOnes, ExtendedSumOfShAmts);
254 NewMask = ConstantExpr::getNot(ExtendedInvertedMask);
255 } else if (match(Masked, m_c_And(m_CombineOr(MaskC, MaskD), m_Value(X))) ||
256 match(Masked, m_Shr(m_Shl(m_Value(X), m_Value(MaskShAmt)),
257 m_Deferred(MaskShAmt)))) {
258 // Peek through an optional zext of the shift amount.
259 match(MaskShAmt, m_ZExtOrSelf(m_Value(MaskShAmt)));
260
261 // Verify that it would be safe to try to add those two shift amounts.
262 if (!canTryToConstantAddTwoShiftAmounts(OuterShift, ShiftShAmt, Masked,
263 MaskShAmt))
264 return nullptr;
265
266 // Can we simplify (ShiftShAmt-MaskShAmt) ?
267 auto *ShAmtsDiff = dyn_cast_or_null<Constant>(SimplifySubInst(
268 ShiftShAmt, MaskShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q));
269 if (!ShAmtsDiff)
270 return nullptr; // Did not simplify.
271 // In this pattern ShAmtsDiff correlates with the number of high bits that
272 // shall be unset in the root value (OuterShift).
273
274 // An extend of an undef value becomes zero because the high bits are never
275 // completely unknown. Replace the the `undef` shift amounts with negated
276 // bitwidth of innermost shift to ensure that the value remains undef when
277 // creating the subsequent shift op.
278 unsigned WidestTyBitWidth = WidestTy->getScalarSizeInBits();
279 ShAmtsDiff = Constant::replaceUndefsWith(
280 ShAmtsDiff, ConstantInt::get(ShAmtsDiff->getType()->getScalarType(),
281 -WidestTyBitWidth));
282 auto *ExtendedNumHighBitsToClear = ConstantExpr::getZExt(
283 ConstantExpr::getSub(ConstantInt::get(ShAmtsDiff->getType(),
284 WidestTyBitWidth,
285 /*isSigned=*/false),
286 ShAmtsDiff),
287 ExtendedTy);
288 // And compute the mask as usual: (-1 l>> (NumHighBitsToClear))
289 auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(ExtendedTy);
290 NewMask =
291 ConstantExpr::getLShr(ExtendedAllOnes, ExtendedNumHighBitsToClear);
292 } else
293 return nullptr; // Don't know anything about this pattern.
294
295 NewMask = ConstantExpr::getTrunc(NewMask, NarrowestTy);
296
297 // Does this mask has any unset bits? If not then we can just not apply it.
298 bool NeedMask = !match(NewMask, m_AllOnes());
299
300 // If we need to apply a mask, there are several more restrictions we have.
301 if (NeedMask) {
302 // The old masking instruction must go away.
303 if (!Masked->hasOneUse())
304 return nullptr;
305 // The original "masking" instruction must not have been`ashr`.
306 if (match(Masked, m_AShr(m_Value(), m_Value())))
307 return nullptr;
308 }
309
310 // If we need to apply truncation, let's do it first, since we can.
311 // We have already ensured that the old truncation will go away.
312 if (HadTrunc)
313 X = Builder.CreateTrunc(X, NarrowestTy);
314
315 // No 'NUW'/'NSW'! We no longer know that we won't shift-out non-0 bits.
316 // We didn't change the Type of this outermost shift, so we can just do it.
317 auto *NewShift = BinaryOperator::Create(OuterShift->getOpcode(), X,
318 OuterShift->getOperand(1));
319 if (!NeedMask)
320 return NewShift;
321
322 Builder.Insert(NewShift);
323 return BinaryOperator::Create(Instruction::And, NewShift, NewMask);
324 }
325
326 /// If we have a shift-by-constant of a bitwise logic op that itself has a
327 /// shift-by-constant operand with identical opcode, we may be able to convert
328 /// that into 2 independent shifts followed by the logic op. This eliminates a
329 /// a use of an intermediate value (reduces dependency chain).
foldShiftOfShiftedLogic(BinaryOperator & I,InstCombiner::BuilderTy & Builder)330 static Instruction *foldShiftOfShiftedLogic(BinaryOperator &I,
331 InstCombiner::BuilderTy &Builder) {
332 assert(I.isShift() && "Expected a shift as input");
333 auto *LogicInst = dyn_cast<BinaryOperator>(I.getOperand(0));
334 if (!LogicInst || !LogicInst->isBitwiseLogicOp() || !LogicInst->hasOneUse())
335 return nullptr;
336
337 Constant *C0, *C1;
338 if (!match(I.getOperand(1), m_Constant(C1)))
339 return nullptr;
340
341 Instruction::BinaryOps ShiftOpcode = I.getOpcode();
342 Type *Ty = I.getType();
343
344 // Find a matching one-use shift by constant. The fold is not valid if the sum
345 // of the shift values equals or exceeds bitwidth.
346 // TODO: Remove the one-use check if the other logic operand (Y) is constant.
347 Value *X, *Y;
348 auto matchFirstShift = [&](Value *V) {
349 BinaryOperator *BO;
350 APInt Threshold(Ty->getScalarSizeInBits(), Ty->getScalarSizeInBits());
351 return match(V, m_BinOp(BO)) && BO->getOpcode() == ShiftOpcode &&
352 match(V, m_OneUse(m_Shift(m_Value(X), m_Constant(C0)))) &&
353 match(ConstantExpr::getAdd(C0, C1),
354 m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, Threshold));
355 };
356
357 // Logic ops are commutative, so check each operand for a match.
358 if (matchFirstShift(LogicInst->getOperand(0)))
359 Y = LogicInst->getOperand(1);
360 else if (matchFirstShift(LogicInst->getOperand(1)))
361 Y = LogicInst->getOperand(0);
362 else
363 return nullptr;
364
365 // shift (logic (shift X, C0), Y), C1 -> logic (shift X, C0+C1), (shift Y, C1)
366 Constant *ShiftSumC = ConstantExpr::getAdd(C0, C1);
367 Value *NewShift1 = Builder.CreateBinOp(ShiftOpcode, X, ShiftSumC);
368 Value *NewShift2 = Builder.CreateBinOp(ShiftOpcode, Y, I.getOperand(1));
369 return BinaryOperator::Create(LogicInst->getOpcode(), NewShift1, NewShift2);
370 }
371
commonShiftTransforms(BinaryOperator & I)372 Instruction *InstCombinerImpl::commonShiftTransforms(BinaryOperator &I) {
373 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
374 assert(Op0->getType() == Op1->getType());
375
376 // If the shift amount is a one-use `sext`, we can demote it to `zext`.
377 Value *Y;
378 if (match(Op1, m_OneUse(m_SExt(m_Value(Y))))) {
379 Value *NewExt = Builder.CreateZExt(Y, I.getType(), Op1->getName());
380 return BinaryOperator::Create(I.getOpcode(), Op0, NewExt);
381 }
382
383 // See if we can fold away this shift.
384 if (SimplifyDemandedInstructionBits(I))
385 return &I;
386
387 // Try to fold constant and into select arguments.
388 if (isa<Constant>(Op0))
389 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
390 if (Instruction *R = FoldOpIntoSelect(I, SI))
391 return R;
392
393 if (Constant *CUI = dyn_cast<Constant>(Op1))
394 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
395 return Res;
396
397 if (auto *NewShift = cast_or_null<Instruction>(
398 reassociateShiftAmtsOfTwoSameDirectionShifts(&I, SQ)))
399 return NewShift;
400
401 // (C1 shift (A add C2)) -> (C1 shift C2) shift A)
402 // iff A and C2 are both positive.
403 Value *A;
404 Constant *C;
405 if (match(Op0, m_Constant()) && match(Op1, m_Add(m_Value(A), m_Constant(C))))
406 if (isKnownNonNegative(A, DL, 0, &AC, &I, &DT) &&
407 isKnownNonNegative(C, DL, 0, &AC, &I, &DT))
408 return BinaryOperator::Create(
409 I.getOpcode(), Builder.CreateBinOp(I.getOpcode(), Op0, C), A);
410
411 // X shift (A srem C) -> X shift (A and (C - 1)) iff C is a power of 2.
412 // Because shifts by negative values (which could occur if A were negative)
413 // are undefined.
414 if (Op1->hasOneUse() && match(Op1, m_SRem(m_Value(A), m_Constant(C))) &&
415 match(C, m_Power2())) {
416 // FIXME: Should this get moved into SimplifyDemandedBits by saying we don't
417 // demand the sign bit (and many others) here??
418 Constant *Mask = ConstantExpr::getSub(C, ConstantInt::get(I.getType(), 1));
419 Value *Rem = Builder.CreateAnd(A, Mask, Op1->getName());
420 return replaceOperand(I, 1, Rem);
421 }
422
423 if (Instruction *Logic = foldShiftOfShiftedLogic(I, Builder))
424 return Logic;
425
426 return nullptr;
427 }
428
429 /// Return true if we can simplify two logical (either left or right) shifts
430 /// that have constant shift amounts: OuterShift (InnerShift X, C1), C2.
canEvaluateShiftedShift(unsigned OuterShAmt,bool IsOuterShl,Instruction * InnerShift,InstCombinerImpl & IC,Instruction * CxtI)431 static bool canEvaluateShiftedShift(unsigned OuterShAmt, bool IsOuterShl,
432 Instruction *InnerShift,
433 InstCombinerImpl &IC, Instruction *CxtI) {
434 assert(InnerShift->isLogicalShift() && "Unexpected instruction type");
435
436 // We need constant scalar or constant splat shifts.
437 const APInt *InnerShiftConst;
438 if (!match(InnerShift->getOperand(1), m_APInt(InnerShiftConst)))
439 return false;
440
441 // Two logical shifts in the same direction:
442 // shl (shl X, C1), C2 --> shl X, C1 + C2
443 // lshr (lshr X, C1), C2 --> lshr X, C1 + C2
444 bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl;
445 if (IsInnerShl == IsOuterShl)
446 return true;
447
448 // Equal shift amounts in opposite directions become bitwise 'and':
449 // lshr (shl X, C), C --> and X, C'
450 // shl (lshr X, C), C --> and X, C'
451 if (*InnerShiftConst == OuterShAmt)
452 return true;
453
454 // If the 2nd shift is bigger than the 1st, we can fold:
455 // lshr (shl X, C1), C2 --> and (shl X, C1 - C2), C3
456 // shl (lshr X, C1), C2 --> and (lshr X, C1 - C2), C3
457 // but it isn't profitable unless we know the and'd out bits are already zero.
458 // Also, check that the inner shift is valid (less than the type width) or
459 // we'll crash trying to produce the bit mask for the 'and'.
460 unsigned TypeWidth = InnerShift->getType()->getScalarSizeInBits();
461 if (InnerShiftConst->ugt(OuterShAmt) && InnerShiftConst->ult(TypeWidth)) {
462 unsigned InnerShAmt = InnerShiftConst->getZExtValue();
463 unsigned MaskShift =
464 IsInnerShl ? TypeWidth - InnerShAmt : InnerShAmt - OuterShAmt;
465 APInt Mask = APInt::getLowBitsSet(TypeWidth, OuterShAmt) << MaskShift;
466 if (IC.MaskedValueIsZero(InnerShift->getOperand(0), Mask, 0, CxtI))
467 return true;
468 }
469
470 return false;
471 }
472
473 /// See if we can compute the specified value, but shifted logically to the left
474 /// or right by some number of bits. This should return true if the expression
475 /// can be computed for the same cost as the current expression tree. This is
476 /// used to eliminate extraneous shifting from things like:
477 /// %C = shl i128 %A, 64
478 /// %D = shl i128 %B, 96
479 /// %E = or i128 %C, %D
480 /// %F = lshr i128 %E, 64
481 /// where the client will ask if E can be computed shifted right by 64-bits. If
482 /// this succeeds, getShiftedValue() will be called to produce the value.
canEvaluateShifted(Value * V,unsigned NumBits,bool IsLeftShift,InstCombinerImpl & IC,Instruction * CxtI)483 static bool canEvaluateShifted(Value *V, unsigned NumBits, bool IsLeftShift,
484 InstCombinerImpl &IC, Instruction *CxtI) {
485 // We can always evaluate constants shifted.
486 if (isa<Constant>(V))
487 return true;
488
489 Instruction *I = dyn_cast<Instruction>(V);
490 if (!I) return false;
491
492 // We can't mutate something that has multiple uses: doing so would
493 // require duplicating the instruction in general, which isn't profitable.
494 if (!I->hasOneUse()) return false;
495
496 switch (I->getOpcode()) {
497 default: return false;
498 case Instruction::And:
499 case Instruction::Or:
500 case Instruction::Xor:
501 // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
502 return canEvaluateShifted(I->getOperand(0), NumBits, IsLeftShift, IC, I) &&
503 canEvaluateShifted(I->getOperand(1), NumBits, IsLeftShift, IC, I);
504
505 case Instruction::Shl:
506 case Instruction::LShr:
507 return canEvaluateShiftedShift(NumBits, IsLeftShift, I, IC, CxtI);
508
509 case Instruction::Select: {
510 SelectInst *SI = cast<SelectInst>(I);
511 Value *TrueVal = SI->getTrueValue();
512 Value *FalseVal = SI->getFalseValue();
513 return canEvaluateShifted(TrueVal, NumBits, IsLeftShift, IC, SI) &&
514 canEvaluateShifted(FalseVal, NumBits, IsLeftShift, IC, SI);
515 }
516 case Instruction::PHI: {
517 // We can change a phi if we can change all operands. Note that we never
518 // get into trouble with cyclic PHIs here because we only consider
519 // instructions with a single use.
520 PHINode *PN = cast<PHINode>(I);
521 for (Value *IncValue : PN->incoming_values())
522 if (!canEvaluateShifted(IncValue, NumBits, IsLeftShift, IC, PN))
523 return false;
524 return true;
525 }
526 }
527 }
528
529 /// Fold OuterShift (InnerShift X, C1), C2.
530 /// See canEvaluateShiftedShift() for the constraints on these instructions.
foldShiftedShift(BinaryOperator * InnerShift,unsigned OuterShAmt,bool IsOuterShl,InstCombiner::BuilderTy & Builder)531 static Value *foldShiftedShift(BinaryOperator *InnerShift, unsigned OuterShAmt,
532 bool IsOuterShl,
533 InstCombiner::BuilderTy &Builder) {
534 bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl;
535 Type *ShType = InnerShift->getType();
536 unsigned TypeWidth = ShType->getScalarSizeInBits();
537
538 // We only accept shifts-by-a-constant in canEvaluateShifted().
539 const APInt *C1;
540 match(InnerShift->getOperand(1), m_APInt(C1));
541 unsigned InnerShAmt = C1->getZExtValue();
542
543 // Change the shift amount and clear the appropriate IR flags.
544 auto NewInnerShift = [&](unsigned ShAmt) {
545 InnerShift->setOperand(1, ConstantInt::get(ShType, ShAmt));
546 if (IsInnerShl) {
547 InnerShift->setHasNoUnsignedWrap(false);
548 InnerShift->setHasNoSignedWrap(false);
549 } else {
550 InnerShift->setIsExact(false);
551 }
552 return InnerShift;
553 };
554
555 // Two logical shifts in the same direction:
556 // shl (shl X, C1), C2 --> shl X, C1 + C2
557 // lshr (lshr X, C1), C2 --> lshr X, C1 + C2
558 if (IsInnerShl == IsOuterShl) {
559 // If this is an oversized composite shift, then unsigned shifts get 0.
560 if (InnerShAmt + OuterShAmt >= TypeWidth)
561 return Constant::getNullValue(ShType);
562
563 return NewInnerShift(InnerShAmt + OuterShAmt);
564 }
565
566 // Equal shift amounts in opposite directions become bitwise 'and':
567 // lshr (shl X, C), C --> and X, C'
568 // shl (lshr X, C), C --> and X, C'
569 if (InnerShAmt == OuterShAmt) {
570 APInt Mask = IsInnerShl
571 ? APInt::getLowBitsSet(TypeWidth, TypeWidth - OuterShAmt)
572 : APInt::getHighBitsSet(TypeWidth, TypeWidth - OuterShAmt);
573 Value *And = Builder.CreateAnd(InnerShift->getOperand(0),
574 ConstantInt::get(ShType, Mask));
575 if (auto *AndI = dyn_cast<Instruction>(And)) {
576 AndI->moveBefore(InnerShift);
577 AndI->takeName(InnerShift);
578 }
579 return And;
580 }
581
582 assert(InnerShAmt > OuterShAmt &&
583 "Unexpected opposite direction logical shift pair");
584
585 // In general, we would need an 'and' for this transform, but
586 // canEvaluateShiftedShift() guarantees that the masked-off bits are not used.
587 // lshr (shl X, C1), C2 --> shl X, C1 - C2
588 // shl (lshr X, C1), C2 --> lshr X, C1 - C2
589 return NewInnerShift(InnerShAmt - OuterShAmt);
590 }
591
592 /// When canEvaluateShifted() returns true for an expression, this function
593 /// inserts the new computation that produces the shifted value.
getShiftedValue(Value * V,unsigned NumBits,bool isLeftShift,InstCombinerImpl & IC,const DataLayout & DL)594 static Value *getShiftedValue(Value *V, unsigned NumBits, bool isLeftShift,
595 InstCombinerImpl &IC, const DataLayout &DL) {
596 // We can always evaluate constants shifted.
597 if (Constant *C = dyn_cast<Constant>(V)) {
598 if (isLeftShift)
599 return IC.Builder.CreateShl(C, NumBits);
600 else
601 return IC.Builder.CreateLShr(C, NumBits);
602 }
603
604 Instruction *I = cast<Instruction>(V);
605 IC.addToWorklist(I);
606
607 switch (I->getOpcode()) {
608 default: llvm_unreachable("Inconsistency with CanEvaluateShifted");
609 case Instruction::And:
610 case Instruction::Or:
611 case Instruction::Xor:
612 // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
613 I->setOperand(
614 0, getShiftedValue(I->getOperand(0), NumBits, isLeftShift, IC, DL));
615 I->setOperand(
616 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL));
617 return I;
618
619 case Instruction::Shl:
620 case Instruction::LShr:
621 return foldShiftedShift(cast<BinaryOperator>(I), NumBits, isLeftShift,
622 IC.Builder);
623
624 case Instruction::Select:
625 I->setOperand(
626 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL));
627 I->setOperand(
628 2, getShiftedValue(I->getOperand(2), NumBits, isLeftShift, IC, DL));
629 return I;
630 case Instruction::PHI: {
631 // We can change a phi if we can change all operands. Note that we never
632 // get into trouble with cyclic PHIs here because we only consider
633 // instructions with a single use.
634 PHINode *PN = cast<PHINode>(I);
635 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
636 PN->setIncomingValue(i, getShiftedValue(PN->getIncomingValue(i), NumBits,
637 isLeftShift, IC, DL));
638 return PN;
639 }
640 }
641 }
642
643 // If this is a bitwise operator or add with a constant RHS we might be able
644 // to pull it through a shift.
canShiftBinOpWithConstantRHS(BinaryOperator & Shift,BinaryOperator * BO)645 static bool canShiftBinOpWithConstantRHS(BinaryOperator &Shift,
646 BinaryOperator *BO) {
647 switch (BO->getOpcode()) {
648 default:
649 return false; // Do not perform transform!
650 case Instruction::Add:
651 return Shift.getOpcode() == Instruction::Shl;
652 case Instruction::Or:
653 case Instruction::And:
654 return true;
655 case Instruction::Xor:
656 // Do not change a 'not' of logical shift because that would create a normal
657 // 'xor'. The 'not' is likely better for analysis, SCEV, and codegen.
658 return !(Shift.isLogicalShift() && match(BO, m_Not(m_Value())));
659 }
660 }
661
FoldShiftByConstant(Value * Op0,Constant * Op1,BinaryOperator & I)662 Instruction *InstCombinerImpl::FoldShiftByConstant(Value *Op0, Constant *Op1,
663 BinaryOperator &I) {
664 bool isLeftShift = I.getOpcode() == Instruction::Shl;
665
666 const APInt *Op1C;
667 if (!match(Op1, m_APInt(Op1C)))
668 return nullptr;
669
670 // See if we can propagate this shift into the input, this covers the trivial
671 // cast of lshr(shl(x,c1),c2) as well as other more complex cases.
672 if (I.getOpcode() != Instruction::AShr &&
673 canEvaluateShifted(Op0, Op1C->getZExtValue(), isLeftShift, *this, &I)) {
674 LLVM_DEBUG(
675 dbgs() << "ICE: GetShiftedValue propagating shift through expression"
676 " to eliminate shift:\n IN: "
677 << *Op0 << "\n SH: " << I << "\n");
678
679 return replaceInstUsesWith(
680 I, getShiftedValue(Op0, Op1C->getZExtValue(), isLeftShift, *this, DL));
681 }
682
683 // See if we can simplify any instructions used by the instruction whose sole
684 // purpose is to compute bits we don't care about.
685 Type *Ty = I.getType();
686 unsigned TypeBits = Ty->getScalarSizeInBits();
687 assert(!Op1C->uge(TypeBits) &&
688 "Shift over the type width should have been removed already");
689
690 if (Instruction *FoldedShift = foldBinOpIntoSelectOrPhi(I))
691 return FoldedShift;
692
693 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
694 if (auto *TI = dyn_cast<TruncInst>(Op0)) {
695 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
696 // place. Don't try to do this transformation in this case. Also, we
697 // require that the input operand is a shift-by-constant so that we have
698 // confidence that the shifts will get folded together. We could do this
699 // xform in more cases, but it is unlikely to be profitable.
700 const APInt *TrShiftAmt;
701 if (I.isLogicalShift() &&
702 match(TI->getOperand(0), m_Shift(m_Value(), m_APInt(TrShiftAmt)))) {
703 auto *TrOp = cast<Instruction>(TI->getOperand(0));
704 Type *SrcTy = TrOp->getType();
705
706 // Okay, we'll do this xform. Make the shift of shift.
707 Constant *ShAmt = ConstantExpr::getZExt(Op1, SrcTy);
708 // (shift2 (shift1 & 0x00FF), c2)
709 Value *NSh = Builder.CreateBinOp(I.getOpcode(), TrOp, ShAmt, I.getName());
710
711 // For logical shifts, the truncation has the effect of making the high
712 // part of the register be zeros. Emulate this by inserting an AND to
713 // clear the top bits as needed. This 'and' will usually be zapped by
714 // other xforms later if dead.
715 unsigned SrcSize = SrcTy->getScalarSizeInBits();
716 Constant *MaskV =
717 ConstantInt::get(SrcTy, APInt::getLowBitsSet(SrcSize, TypeBits));
718
719 // The mask we constructed says what the trunc would do if occurring
720 // between the shifts. We want to know the effect *after* the second
721 // shift. We know that it is a logical shift by a constant, so adjust the
722 // mask as appropriate.
723 MaskV = ConstantExpr::get(I.getOpcode(), MaskV, ShAmt);
724 // shift1 & 0x00FF
725 Value *And = Builder.CreateAnd(NSh, MaskV, TI->getName());
726 // Return the value truncated to the interesting size.
727 return new TruncInst(And, Ty);
728 }
729 }
730
731 if (Op0->hasOneUse()) {
732 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
733 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
734 Value *V1;
735 const APInt *CC;
736 switch (Op0BO->getOpcode()) {
737 default: break;
738 case Instruction::Add:
739 case Instruction::And:
740 case Instruction::Or:
741 case Instruction::Xor: {
742 // These operators commute.
743 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
744 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
745 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
746 m_Specific(Op1)))) {
747 Value *YS = // (Y << C)
748 Builder.CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
749 // (X + (Y << C))
750 Value *X = Builder.CreateBinOp(Op0BO->getOpcode(), YS, V1,
751 Op0BO->getOperand(1)->getName());
752 unsigned Op1Val = Op1C->getLimitedValue(TypeBits);
753 APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val);
754 Constant *Mask = ConstantInt::get(Ty, Bits);
755 return BinaryOperator::CreateAnd(X, Mask);
756 }
757
758 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
759 Value *Op0BOOp1 = Op0BO->getOperand(1);
760 if (isLeftShift && Op0BOOp1->hasOneUse() &&
761 match(Op0BOOp1, m_And(m_OneUse(m_Shr(m_Value(V1), m_Specific(Op1))),
762 m_APInt(CC)))) {
763 Value *YS = // (Y << C)
764 Builder.CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
765 // X & (CC << C)
766 Value *XM = Builder.CreateAnd(
767 V1, ConstantExpr::getShl(ConstantInt::get(Ty, *CC), Op1),
768 V1->getName() + ".mask");
769 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
770 }
771 LLVM_FALLTHROUGH;
772 }
773
774 case Instruction::Sub: {
775 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
776 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
777 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
778 m_Specific(Op1)))) {
779 Value *YS = // (Y << C)
780 Builder.CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
781 // (X + (Y << C))
782 Value *X = Builder.CreateBinOp(Op0BO->getOpcode(), V1, YS,
783 Op0BO->getOperand(0)->getName());
784 unsigned Op1Val = Op1C->getLimitedValue(TypeBits);
785 APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val);
786 Constant *Mask = ConstantInt::get(Ty, Bits);
787 return BinaryOperator::CreateAnd(X, Mask);
788 }
789
790 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
791 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
792 match(Op0BO->getOperand(0),
793 m_And(m_OneUse(m_Shr(m_Value(V1), m_Specific(Op1))),
794 m_APInt(CC)))) {
795 Value *YS = // (Y << C)
796 Builder.CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
797 // X & (CC << C)
798 Value *XM = Builder.CreateAnd(
799 V1, ConstantExpr::getShl(ConstantInt::get(Ty, *CC), Op1),
800 V1->getName() + ".mask");
801 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
802 }
803
804 break;
805 }
806 }
807
808 // If the operand is a bitwise operator with a constant RHS, and the
809 // shift is the only use, we can pull it out of the shift.
810 const APInt *Op0C;
811 if (match(Op0BO->getOperand(1), m_APInt(Op0C))) {
812 if (canShiftBinOpWithConstantRHS(I, Op0BO)) {
813 Constant *NewRHS = ConstantExpr::get(I.getOpcode(),
814 cast<Constant>(Op0BO->getOperand(1)), Op1);
815
816 Value *NewShift =
817 Builder.CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
818 NewShift->takeName(Op0BO);
819
820 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
821 NewRHS);
822 }
823 }
824
825 // If the operand is a subtract with a constant LHS, and the shift
826 // is the only use, we can pull it out of the shift.
827 // This folds (shl (sub C1, X), C2) -> (sub (C1 << C2), (shl X, C2))
828 if (isLeftShift && Op0BO->getOpcode() == Instruction::Sub &&
829 match(Op0BO->getOperand(0), m_APInt(Op0C))) {
830 Constant *NewRHS = ConstantExpr::get(I.getOpcode(),
831 cast<Constant>(Op0BO->getOperand(0)), Op1);
832
833 Value *NewShift = Builder.CreateShl(Op0BO->getOperand(1), Op1);
834 NewShift->takeName(Op0BO);
835
836 return BinaryOperator::CreateSub(NewRHS, NewShift);
837 }
838 }
839
840 // If we have a select that conditionally executes some binary operator,
841 // see if we can pull it the select and operator through the shift.
842 //
843 // For example, turning:
844 // shl (select C, (add X, C1), X), C2
845 // Into:
846 // Y = shl X, C2
847 // select C, (add Y, C1 << C2), Y
848 Value *Cond;
849 BinaryOperator *TBO;
850 Value *FalseVal;
851 if (match(Op0, m_Select(m_Value(Cond), m_OneUse(m_BinOp(TBO)),
852 m_Value(FalseVal)))) {
853 const APInt *C;
854 if (!isa<Constant>(FalseVal) && TBO->getOperand(0) == FalseVal &&
855 match(TBO->getOperand(1), m_APInt(C)) &&
856 canShiftBinOpWithConstantRHS(I, TBO)) {
857 Constant *NewRHS = ConstantExpr::get(I.getOpcode(),
858 cast<Constant>(TBO->getOperand(1)), Op1);
859
860 Value *NewShift =
861 Builder.CreateBinOp(I.getOpcode(), FalseVal, Op1);
862 Value *NewOp = Builder.CreateBinOp(TBO->getOpcode(), NewShift,
863 NewRHS);
864 return SelectInst::Create(Cond, NewOp, NewShift);
865 }
866 }
867
868 BinaryOperator *FBO;
869 Value *TrueVal;
870 if (match(Op0, m_Select(m_Value(Cond), m_Value(TrueVal),
871 m_OneUse(m_BinOp(FBO))))) {
872 const APInt *C;
873 if (!isa<Constant>(TrueVal) && FBO->getOperand(0) == TrueVal &&
874 match(FBO->getOperand(1), m_APInt(C)) &&
875 canShiftBinOpWithConstantRHS(I, FBO)) {
876 Constant *NewRHS = ConstantExpr::get(I.getOpcode(),
877 cast<Constant>(FBO->getOperand(1)), Op1);
878
879 Value *NewShift =
880 Builder.CreateBinOp(I.getOpcode(), TrueVal, Op1);
881 Value *NewOp = Builder.CreateBinOp(FBO->getOpcode(), NewShift,
882 NewRHS);
883 return SelectInst::Create(Cond, NewShift, NewOp);
884 }
885 }
886 }
887
888 return nullptr;
889 }
890
visitShl(BinaryOperator & I)891 Instruction *InstCombinerImpl::visitShl(BinaryOperator &I) {
892 const SimplifyQuery Q = SQ.getWithInstruction(&I);
893
894 if (Value *V = SimplifyShlInst(I.getOperand(0), I.getOperand(1),
895 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(), Q))
896 return replaceInstUsesWith(I, V);
897
898 if (Instruction *X = foldVectorBinop(I))
899 return X;
900
901 if (Instruction *V = commonShiftTransforms(I))
902 return V;
903
904 if (Instruction *V = dropRedundantMaskingOfLeftShiftInput(&I, Q, Builder))
905 return V;
906
907 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
908 Type *Ty = I.getType();
909 unsigned BitWidth = Ty->getScalarSizeInBits();
910
911 const APInt *ShAmtAPInt;
912 if (match(Op1, m_APInt(ShAmtAPInt))) {
913 unsigned ShAmt = ShAmtAPInt->getZExtValue();
914
915 // shl (zext X), ShAmt --> zext (shl X, ShAmt)
916 // This is only valid if X would have zeros shifted out.
917 Value *X;
918 if (match(Op0, m_OneUse(m_ZExt(m_Value(X))))) {
919 unsigned SrcWidth = X->getType()->getScalarSizeInBits();
920 if (ShAmt < SrcWidth &&
921 MaskedValueIsZero(X, APInt::getHighBitsSet(SrcWidth, ShAmt), 0, &I))
922 return new ZExtInst(Builder.CreateShl(X, ShAmt), Ty);
923 }
924
925 // (X >> C) << C --> X & (-1 << C)
926 if (match(Op0, m_Shr(m_Value(X), m_Specific(Op1)))) {
927 APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt));
928 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask));
929 }
930
931 const APInt *ShOp1;
932 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_APInt(ShOp1)))) &&
933 ShOp1->ult(BitWidth)) {
934 unsigned ShrAmt = ShOp1->getZExtValue();
935 if (ShrAmt < ShAmt) {
936 // If C1 < C2: (X >>?,exact C1) << C2 --> X << (C2 - C1)
937 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShrAmt);
938 auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
939 NewShl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
940 NewShl->setHasNoSignedWrap(I.hasNoSignedWrap());
941 return NewShl;
942 }
943 if (ShrAmt > ShAmt) {
944 // If C1 > C2: (X >>?exact C1) << C2 --> X >>?exact (C1 - C2)
945 Constant *ShiftDiff = ConstantInt::get(Ty, ShrAmt - ShAmt);
946 auto *NewShr = BinaryOperator::Create(
947 cast<BinaryOperator>(Op0)->getOpcode(), X, ShiftDiff);
948 NewShr->setIsExact(true);
949 return NewShr;
950 }
951 }
952
953 if (match(Op0, m_OneUse(m_Shr(m_Value(X), m_APInt(ShOp1)))) &&
954 ShOp1->ult(BitWidth)) {
955 unsigned ShrAmt = ShOp1->getZExtValue();
956 if (ShrAmt < ShAmt) {
957 // If C1 < C2: (X >>? C1) << C2 --> X << (C2 - C1) & (-1 << C2)
958 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShrAmt);
959 auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
960 NewShl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
961 NewShl->setHasNoSignedWrap(I.hasNoSignedWrap());
962 Builder.Insert(NewShl);
963 APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt));
964 return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask));
965 }
966 if (ShrAmt > ShAmt) {
967 // If C1 > C2: (X >>? C1) << C2 --> X >>? (C1 - C2) & (-1 << C2)
968 Constant *ShiftDiff = ConstantInt::get(Ty, ShrAmt - ShAmt);
969 auto *OldShr = cast<BinaryOperator>(Op0);
970 auto *NewShr =
971 BinaryOperator::Create(OldShr->getOpcode(), X, ShiftDiff);
972 NewShr->setIsExact(OldShr->isExact());
973 Builder.Insert(NewShr);
974 APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt));
975 return BinaryOperator::CreateAnd(NewShr, ConstantInt::get(Ty, Mask));
976 }
977 }
978
979 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShOp1))) && ShOp1->ult(BitWidth)) {
980 unsigned AmtSum = ShAmt + ShOp1->getZExtValue();
981 // Oversized shifts are simplified to zero in InstSimplify.
982 if (AmtSum < BitWidth)
983 // (X << C1) << C2 --> X << (C1 + C2)
984 return BinaryOperator::CreateShl(X, ConstantInt::get(Ty, AmtSum));
985 }
986
987 // If the shifted-out value is known-zero, then this is a NUW shift.
988 if (!I.hasNoUnsignedWrap() &&
989 MaskedValueIsZero(Op0, APInt::getHighBitsSet(BitWidth, ShAmt), 0, &I)) {
990 I.setHasNoUnsignedWrap();
991 return &I;
992 }
993
994 // If the shifted-out value is all signbits, then this is a NSW shift.
995 if (!I.hasNoSignedWrap() && ComputeNumSignBits(Op0, 0, &I) > ShAmt) {
996 I.setHasNoSignedWrap();
997 return &I;
998 }
999 }
1000
1001 // Transform (x >> y) << y to x & (-1 << y)
1002 // Valid for any type of right-shift.
1003 Value *X;
1004 if (match(Op0, m_OneUse(m_Shr(m_Value(X), m_Specific(Op1))))) {
1005 Constant *AllOnes = ConstantInt::getAllOnesValue(Ty);
1006 Value *Mask = Builder.CreateShl(AllOnes, Op1);
1007 return BinaryOperator::CreateAnd(Mask, X);
1008 }
1009
1010 Constant *C1;
1011 if (match(Op1, m_Constant(C1))) {
1012 Constant *C2;
1013 Value *X;
1014 // (C2 << X) << C1 --> (C2 << C1) << X
1015 if (match(Op0, m_OneUse(m_Shl(m_Constant(C2), m_Value(X)))))
1016 return BinaryOperator::CreateShl(ConstantExpr::getShl(C2, C1), X);
1017
1018 // (X * C2) << C1 --> X * (C2 << C1)
1019 if (match(Op0, m_Mul(m_Value(X), m_Constant(C2))))
1020 return BinaryOperator::CreateMul(X, ConstantExpr::getShl(C2, C1));
1021
1022 // shl (zext i1 X), C1 --> select (X, 1 << C1, 0)
1023 if (match(Op0, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
1024 auto *NewC = ConstantExpr::getShl(ConstantInt::get(Ty, 1), C1);
1025 return SelectInst::Create(X, NewC, ConstantInt::getNullValue(Ty));
1026 }
1027 }
1028
1029 // (1 << (C - x)) -> ((1 << C) >> x) if C is bitwidth - 1
1030 if (match(Op0, m_One()) &&
1031 match(Op1, m_Sub(m_SpecificInt(BitWidth - 1), m_Value(X))))
1032 return BinaryOperator::CreateLShr(
1033 ConstantInt::get(Ty, APInt::getSignMask(BitWidth)), X);
1034
1035 return nullptr;
1036 }
1037
visitLShr(BinaryOperator & I)1038 Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) {
1039 if (Value *V = SimplifyLShrInst(I.getOperand(0), I.getOperand(1), I.isExact(),
1040 SQ.getWithInstruction(&I)))
1041 return replaceInstUsesWith(I, V);
1042
1043 if (Instruction *X = foldVectorBinop(I))
1044 return X;
1045
1046 if (Instruction *R = commonShiftTransforms(I))
1047 return R;
1048
1049 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1050 Type *Ty = I.getType();
1051 const APInt *ShAmtAPInt;
1052 if (match(Op1, m_APInt(ShAmtAPInt))) {
1053 unsigned ShAmt = ShAmtAPInt->getZExtValue();
1054 unsigned BitWidth = Ty->getScalarSizeInBits();
1055 auto *II = dyn_cast<IntrinsicInst>(Op0);
1056 if (II && isPowerOf2_32(BitWidth) && Log2_32(BitWidth) == ShAmt &&
1057 (II->getIntrinsicID() == Intrinsic::ctlz ||
1058 II->getIntrinsicID() == Intrinsic::cttz ||
1059 II->getIntrinsicID() == Intrinsic::ctpop)) {
1060 // ctlz.i32(x)>>5 --> zext(x == 0)
1061 // cttz.i32(x)>>5 --> zext(x == 0)
1062 // ctpop.i32(x)>>5 --> zext(x == -1)
1063 bool IsPop = II->getIntrinsicID() == Intrinsic::ctpop;
1064 Constant *RHS = ConstantInt::getSigned(Ty, IsPop ? -1 : 0);
1065 Value *Cmp = Builder.CreateICmpEQ(II->getArgOperand(0), RHS);
1066 return new ZExtInst(Cmp, Ty);
1067 }
1068
1069 Value *X;
1070 const APInt *ShOp1;
1071 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShOp1))) && ShOp1->ult(BitWidth)) {
1072 if (ShOp1->ult(ShAmt)) {
1073 unsigned ShlAmt = ShOp1->getZExtValue();
1074 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt);
1075 if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) {
1076 // (X <<nuw C1) >>u C2 --> X >>u (C2 - C1)
1077 auto *NewLShr = BinaryOperator::CreateLShr(X, ShiftDiff);
1078 NewLShr->setIsExact(I.isExact());
1079 return NewLShr;
1080 }
1081 // (X << C1) >>u C2 --> (X >>u (C2 - C1)) & (-1 >> C2)
1082 Value *NewLShr = Builder.CreateLShr(X, ShiftDiff, "", I.isExact());
1083 APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt));
1084 return BinaryOperator::CreateAnd(NewLShr, ConstantInt::get(Ty, Mask));
1085 }
1086 if (ShOp1->ugt(ShAmt)) {
1087 unsigned ShlAmt = ShOp1->getZExtValue();
1088 Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt);
1089 if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) {
1090 // (X <<nuw C1) >>u C2 --> X <<nuw (C1 - C2)
1091 auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
1092 NewShl->setHasNoUnsignedWrap(true);
1093 return NewShl;
1094 }
1095 // (X << C1) >>u C2 --> X << (C1 - C2) & (-1 >> C2)
1096 Value *NewShl = Builder.CreateShl(X, ShiftDiff);
1097 APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt));
1098 return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask));
1099 }
1100 assert(*ShOp1 == ShAmt);
1101 // (X << C) >>u C --> X & (-1 >>u C)
1102 APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt));
1103 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask));
1104 }
1105
1106 if (match(Op0, m_OneUse(m_ZExt(m_Value(X)))) &&
1107 (!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType()))) {
1108 assert(ShAmt < X->getType()->getScalarSizeInBits() &&
1109 "Big shift not simplified to zero?");
1110 // lshr (zext iM X to iN), C --> zext (lshr X, C) to iN
1111 Value *NewLShr = Builder.CreateLShr(X, ShAmt);
1112 return new ZExtInst(NewLShr, Ty);
1113 }
1114
1115 if (match(Op0, m_SExt(m_Value(X))) &&
1116 (!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType()))) {
1117 // Are we moving the sign bit to the low bit and widening with high zeros?
1118 unsigned SrcTyBitWidth = X->getType()->getScalarSizeInBits();
1119 if (ShAmt == BitWidth - 1) {
1120 // lshr (sext i1 X to iN), N-1 --> zext X to iN
1121 if (SrcTyBitWidth == 1)
1122 return new ZExtInst(X, Ty);
1123
1124 // lshr (sext iM X to iN), N-1 --> zext (lshr X, M-1) to iN
1125 if (Op0->hasOneUse()) {
1126 Value *NewLShr = Builder.CreateLShr(X, SrcTyBitWidth - 1);
1127 return new ZExtInst(NewLShr, Ty);
1128 }
1129 }
1130
1131 // lshr (sext iM X to iN), N-M --> zext (ashr X, min(N-M, M-1)) to iN
1132 if (ShAmt == BitWidth - SrcTyBitWidth && Op0->hasOneUse()) {
1133 // The new shift amount can't be more than the narrow source type.
1134 unsigned NewShAmt = std::min(ShAmt, SrcTyBitWidth - 1);
1135 Value *AShr = Builder.CreateAShr(X, NewShAmt);
1136 return new ZExtInst(AShr, Ty);
1137 }
1138 }
1139
1140 Value *Y;
1141 if (ShAmt == BitWidth - 1) {
1142 // lshr i32 or(X,-X), 31 --> zext (X != 0)
1143 if (match(Op0, m_OneUse(m_c_Or(m_Neg(m_Value(X)), m_Deferred(X)))))
1144 return new ZExtInst(Builder.CreateIsNotNull(X), Ty);
1145
1146 // lshr i32 (X -nsw Y), 31 --> zext (X < Y)
1147 if (match(Op0, m_OneUse(m_NSWSub(m_Value(X), m_Value(Y)))))
1148 return new ZExtInst(Builder.CreateICmpSLT(X, Y), Ty);
1149
1150 // Check if a number is negative and odd:
1151 // lshr i32 (srem X, 2), 31 --> and (X >> 31), X
1152 if (match(Op0, m_OneUse(m_SRem(m_Value(X), m_SpecificInt(2))))) {
1153 Value *Signbit = Builder.CreateLShr(X, ShAmt);
1154 return BinaryOperator::CreateAnd(Signbit, X);
1155 }
1156 }
1157
1158 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShOp1)))) {
1159 unsigned AmtSum = ShAmt + ShOp1->getZExtValue();
1160 // Oversized shifts are simplified to zero in InstSimplify.
1161 if (AmtSum < BitWidth)
1162 // (X >>u C1) >>u C2 --> X >>u (C1 + C2)
1163 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
1164 }
1165
1166 // Look for a "splat" mul pattern - it replicates bits across each half of
1167 // a value, so a right shift is just a mask of the low bits:
1168 // lshr i32 (mul nuw X, Pow2+1), 16 --> and X, Pow2-1
1169 // TODO: Generalize to allow more than just half-width shifts?
1170 const APInt *MulC;
1171 if (match(Op0, m_NUWMul(m_Value(X), m_APInt(MulC))) &&
1172 ShAmt * 2 == BitWidth && (*MulC - 1).isPowerOf2() &&
1173 MulC->logBase2() == ShAmt)
1174 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, *MulC - 2));
1175
1176 // If the shifted-out value is known-zero, then this is an exact shift.
1177 if (!I.isExact() &&
1178 MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmt), 0, &I)) {
1179 I.setIsExact();
1180 return &I;
1181 }
1182 }
1183
1184 // Transform (x << y) >> y to x & (-1 >> y)
1185 Value *X;
1186 if (match(Op0, m_OneUse(m_Shl(m_Value(X), m_Specific(Op1))))) {
1187 Constant *AllOnes = ConstantInt::getAllOnesValue(Ty);
1188 Value *Mask = Builder.CreateLShr(AllOnes, Op1);
1189 return BinaryOperator::CreateAnd(Mask, X);
1190 }
1191
1192 return nullptr;
1193 }
1194
1195 Instruction *
foldVariableSignZeroExtensionOfVariableHighBitExtract(BinaryOperator & OldAShr)1196 InstCombinerImpl::foldVariableSignZeroExtensionOfVariableHighBitExtract(
1197 BinaryOperator &OldAShr) {
1198 assert(OldAShr.getOpcode() == Instruction::AShr &&
1199 "Must be called with arithmetic right-shift instruction only.");
1200
1201 // Check that constant C is a splat of the element-wise bitwidth of V.
1202 auto BitWidthSplat = [](Constant *C, Value *V) {
1203 return match(
1204 C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1205 APInt(C->getType()->getScalarSizeInBits(),
1206 V->getType()->getScalarSizeInBits())));
1207 };
1208
1209 // It should look like variable-length sign-extension on the outside:
1210 // (Val << (bitwidth(Val)-Nbits)) a>> (bitwidth(Val)-Nbits)
1211 Value *NBits;
1212 Instruction *MaybeTrunc;
1213 Constant *C1, *C2;
1214 if (!match(&OldAShr,
1215 m_AShr(m_Shl(m_Instruction(MaybeTrunc),
1216 m_ZExtOrSelf(m_Sub(m_Constant(C1),
1217 m_ZExtOrSelf(m_Value(NBits))))),
1218 m_ZExtOrSelf(m_Sub(m_Constant(C2),
1219 m_ZExtOrSelf(m_Deferred(NBits)))))) ||
1220 !BitWidthSplat(C1, &OldAShr) || !BitWidthSplat(C2, &OldAShr))
1221 return nullptr;
1222
1223 // There may or may not be a truncation after outer two shifts.
1224 Instruction *HighBitExtract;
1225 match(MaybeTrunc, m_TruncOrSelf(m_Instruction(HighBitExtract)));
1226 bool HadTrunc = MaybeTrunc != HighBitExtract;
1227
1228 // And finally, the innermost part of the pattern must be a right-shift.
1229 Value *X, *NumLowBitsToSkip;
1230 if (!match(HighBitExtract, m_Shr(m_Value(X), m_Value(NumLowBitsToSkip))))
1231 return nullptr;
1232
1233 // Said right-shift must extract high NBits bits - C0 must be it's bitwidth.
1234 Constant *C0;
1235 if (!match(NumLowBitsToSkip,
1236 m_ZExtOrSelf(
1237 m_Sub(m_Constant(C0), m_ZExtOrSelf(m_Specific(NBits))))) ||
1238 !BitWidthSplat(C0, HighBitExtract))
1239 return nullptr;
1240
1241 // Since the NBits is identical for all shifts, if the outermost and
1242 // innermost shifts are identical, then outermost shifts are redundant.
1243 // If we had truncation, do keep it though.
1244 if (HighBitExtract->getOpcode() == OldAShr.getOpcode())
1245 return replaceInstUsesWith(OldAShr, MaybeTrunc);
1246
1247 // Else, if there was a truncation, then we need to ensure that one
1248 // instruction will go away.
1249 if (HadTrunc && !match(&OldAShr, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
1250 return nullptr;
1251
1252 // Finally, bypass two innermost shifts, and perform the outermost shift on
1253 // the operands of the innermost shift.
1254 Instruction *NewAShr =
1255 BinaryOperator::Create(OldAShr.getOpcode(), X, NumLowBitsToSkip);
1256 NewAShr->copyIRFlags(HighBitExtract); // We can preserve 'exact'-ness.
1257 if (!HadTrunc)
1258 return NewAShr;
1259
1260 Builder.Insert(NewAShr);
1261 return TruncInst::CreateTruncOrBitCast(NewAShr, OldAShr.getType());
1262 }
1263
visitAShr(BinaryOperator & I)1264 Instruction *InstCombinerImpl::visitAShr(BinaryOperator &I) {
1265 if (Value *V = SimplifyAShrInst(I.getOperand(0), I.getOperand(1), I.isExact(),
1266 SQ.getWithInstruction(&I)))
1267 return replaceInstUsesWith(I, V);
1268
1269 if (Instruction *X = foldVectorBinop(I))
1270 return X;
1271
1272 if (Instruction *R = commonShiftTransforms(I))
1273 return R;
1274
1275 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1276 Type *Ty = I.getType();
1277 unsigned BitWidth = Ty->getScalarSizeInBits();
1278 const APInt *ShAmtAPInt;
1279 if (match(Op1, m_APInt(ShAmtAPInt)) && ShAmtAPInt->ult(BitWidth)) {
1280 unsigned ShAmt = ShAmtAPInt->getZExtValue();
1281
1282 // If the shift amount equals the difference in width of the destination
1283 // and source scalar types:
1284 // ashr (shl (zext X), C), C --> sext X
1285 Value *X;
1286 if (match(Op0, m_Shl(m_ZExt(m_Value(X)), m_Specific(Op1))) &&
1287 ShAmt == BitWidth - X->getType()->getScalarSizeInBits())
1288 return new SExtInst(X, Ty);
1289
1290 // We can't handle (X << C1) >>s C2. It shifts arbitrary bits in. However,
1291 // we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits.
1292 const APInt *ShOp1;
1293 if (match(Op0, m_NSWShl(m_Value(X), m_APInt(ShOp1))) &&
1294 ShOp1->ult(BitWidth)) {
1295 unsigned ShlAmt = ShOp1->getZExtValue();
1296 if (ShlAmt < ShAmt) {
1297 // (X <<nsw C1) >>s C2 --> X >>s (C2 - C1)
1298 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt);
1299 auto *NewAShr = BinaryOperator::CreateAShr(X, ShiftDiff);
1300 NewAShr->setIsExact(I.isExact());
1301 return NewAShr;
1302 }
1303 if (ShlAmt > ShAmt) {
1304 // (X <<nsw C1) >>s C2 --> X <<nsw (C1 - C2)
1305 Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt);
1306 auto *NewShl = BinaryOperator::Create(Instruction::Shl, X, ShiftDiff);
1307 NewShl->setHasNoSignedWrap(true);
1308 return NewShl;
1309 }
1310 }
1311
1312 if (match(Op0, m_AShr(m_Value(X), m_APInt(ShOp1))) &&
1313 ShOp1->ult(BitWidth)) {
1314 unsigned AmtSum = ShAmt + ShOp1->getZExtValue();
1315 // Oversized arithmetic shifts replicate the sign bit.
1316 AmtSum = std::min(AmtSum, BitWidth - 1);
1317 // (X >>s C1) >>s C2 --> X >>s (C1 + C2)
1318 return BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
1319 }
1320
1321 if (match(Op0, m_OneUse(m_SExt(m_Value(X)))) &&
1322 (Ty->isVectorTy() || shouldChangeType(Ty, X->getType()))) {
1323 // ashr (sext X), C --> sext (ashr X, C')
1324 Type *SrcTy = X->getType();
1325 ShAmt = std::min(ShAmt, SrcTy->getScalarSizeInBits() - 1);
1326 Value *NewSh = Builder.CreateAShr(X, ConstantInt::get(SrcTy, ShAmt));
1327 return new SExtInst(NewSh, Ty);
1328 }
1329
1330 if (ShAmt == BitWidth - 1) {
1331 // ashr i32 or(X,-X), 31 --> sext (X != 0)
1332 if (match(Op0, m_OneUse(m_c_Or(m_Neg(m_Value(X)), m_Deferred(X)))))
1333 return new SExtInst(Builder.CreateIsNotNull(X), Ty);
1334
1335 // ashr i32 (X -nsw Y), 31 --> sext (X < Y)
1336 Value *Y;
1337 if (match(Op0, m_OneUse(m_NSWSub(m_Value(X), m_Value(Y)))))
1338 return new SExtInst(Builder.CreateICmpSLT(X, Y), Ty);
1339 }
1340
1341 // If the shifted-out value is known-zero, then this is an exact shift.
1342 if (!I.isExact() &&
1343 MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmt), 0, &I)) {
1344 I.setIsExact();
1345 return &I;
1346 }
1347 }
1348
1349 if (Instruction *R = foldVariableSignZeroExtensionOfVariableHighBitExtract(I))
1350 return R;
1351
1352 // See if we can turn a signed shr into an unsigned shr.
1353 if (MaskedValueIsZero(Op0, APInt::getSignMask(BitWidth), 0, &I))
1354 return BinaryOperator::CreateLShr(Op0, Op1);
1355
1356 // ashr (xor %x, -1), %y --> xor (ashr %x, %y), -1
1357 Value *X;
1358 if (match(Op0, m_OneUse(m_Not(m_Value(X))))) {
1359 // Note that we must drop 'exact'-ness of the shift!
1360 // Note that we can't keep undef's in -1 vector constant!
1361 auto *NewAShr = Builder.CreateAShr(X, Op1, Op0->getName() + ".not");
1362 return BinaryOperator::CreateNot(NewAShr);
1363 }
1364
1365 return nullptr;
1366 }
1367