LLVM 19.0.0git
InstCombineSimplifyDemanded.cpp
Go to the documentation of this file.
1//===- InstCombineSimplifyDemanded.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 contains logic for simplifying instructions based on information
10// about how they are used.
11//
12//===----------------------------------------------------------------------===//
13
14#include "InstCombineInternal.h"
21
22using namespace llvm;
23using namespace llvm::PatternMatch;
24
25#define DEBUG_TYPE "instcombine"
26
27static cl::opt<bool>
28 VerifyKnownBits("instcombine-verify-known-bits",
29 cl::desc("Verify that computeKnownBits() and "
30 "SimplifyDemandedBits() are consistent"),
31 cl::Hidden, cl::init(false));
32
33/// Check to see if the specified operand of the specified instruction is a
34/// constant integer. If so, check to see if there are any bits set in the
35/// constant that are not demanded. If so, shrink the constant and return true.
36static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
37 const APInt &Demanded) {
38 assert(I && "No instruction?");
39 assert(OpNo < I->getNumOperands() && "Operand index too large");
40
41 // The operand must be a constant integer or splat integer.
42 Value *Op = I->getOperand(OpNo);
43 const APInt *C;
44 if (!match(Op, m_APInt(C)))
45 return false;
46
47 // If there are no bits set that aren't demanded, nothing to do.
48 if (C->isSubsetOf(Demanded))
49 return false;
50
51 // This instruction is producing bits that are not demanded. Shrink the RHS.
52 I->setOperand(OpNo, ConstantInt::get(Op->getType(), *C & Demanded));
53
54 return true;
55}
56
57/// Returns the bitwidth of the given scalar or pointer type. For vector types,
58/// returns the element type's bitwidth.
59static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
60 if (unsigned BitWidth = Ty->getScalarSizeInBits())
61 return BitWidth;
62
63 return DL.getPointerTypeSizeInBits(Ty);
64}
65
66/// Inst is an integer instruction that SimplifyDemandedBits knows about. See if
67/// the instruction has any properties that allow us to simplify its operands.
69 KnownBits &Known) {
70 APInt DemandedMask(APInt::getAllOnes(Known.getBitWidth()));
71 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask, Known,
72 0, &Inst);
73 if (!V) return false;
74 if (V == &Inst) return true;
75 replaceInstUsesWith(Inst, V);
76 return true;
77}
78
79/// Inst is an integer instruction that SimplifyDemandedBits knows about. See if
80/// the instruction has any properties that allow us to simplify its operands.
82 KnownBits Known(getBitWidth(Inst.getType(), DL));
83 return SimplifyDemandedInstructionBits(Inst, Known);
84}
85
86/// This form of SimplifyDemandedBits simplifies the specified instruction
87/// operand if possible, updating it in place. It returns true if it made any
88/// change and false otherwise.
90 const APInt &DemandedMask,
91 KnownBits &Known, unsigned Depth) {
92 Use &U = I->getOperandUse(OpNo);
93 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask, Known,
94 Depth, I);
95 if (!NewVal) return false;
96 if (Instruction* OpInst = dyn_cast<Instruction>(U))
97 salvageDebugInfo(*OpInst);
98
99 replaceUse(U, NewVal);
100 return true;
101}
102
103/// This function attempts to replace V with a simpler value based on the
104/// demanded bits. When this function is called, it is known that only the bits
105/// set in DemandedMask of the result of V are ever used downstream.
106/// Consequently, depending on the mask and V, it may be possible to replace V
107/// with a constant or one of its operands. In such cases, this function does
108/// the replacement and returns true. In all other cases, it returns false after
109/// analyzing the expression and setting KnownOne and known to be one in the
110/// expression. Known.Zero contains all the bits that are known to be zero in
111/// the expression. These are provided to potentially allow the caller (which
112/// might recursively be SimplifyDemandedBits itself) to simplify the
113/// expression.
114/// Known.One and Known.Zero always follow the invariant that:
115/// Known.One & Known.Zero == 0.
116/// That is, a bit can't be both 1 and 0. The bits in Known.One and Known.Zero
117/// are accurate even for bits not in DemandedMask. Note
118/// also that the bitwidth of V, DemandedMask, Known.Zero and Known.One must all
119/// be the same.
120///
121/// This returns null if it did not change anything and it permits no
122/// simplification. This returns V itself if it did some simplification of V's
123/// operands based on the information about what bits are demanded. This returns
124/// some other non-null value if it found out that V is equal to another value
125/// in the context where the specified bits are demanded, but not for all users.
127 KnownBits &Known,
128 unsigned Depth,
129 Instruction *CxtI) {
130 assert(V != nullptr && "Null pointer of Value???");
131 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth");
132 uint32_t BitWidth = DemandedMask.getBitWidth();
133 Type *VTy = V->getType();
134 assert(
135 (!VTy->isIntOrIntVectorTy() || VTy->getScalarSizeInBits() == BitWidth) &&
136 Known.getBitWidth() == BitWidth &&
137 "Value *V, DemandedMask and Known must have same BitWidth");
138
139 if (isa<Constant>(V)) {
140 computeKnownBits(V, Known, Depth, CxtI);
141 return nullptr;
142 }
143
144 Known.resetAll();
145 if (DemandedMask.isZero()) // Not demanding any bits from V.
146 return UndefValue::get(VTy);
147
149 return nullptr;
150
151 Instruction *I = dyn_cast<Instruction>(V);
152 if (!I) {
153 computeKnownBits(V, Known, Depth, CxtI);
154 return nullptr; // Only analyze instructions.
155 }
156
157 // If there are multiple uses of this value and we aren't at the root, then
158 // we can't do any simplifications of the operands, because DemandedMask
159 // only reflects the bits demanded by *one* of the users.
160 if (Depth != 0 && !I->hasOneUse())
161 return SimplifyMultipleUseDemandedBits(I, DemandedMask, Known, Depth, CxtI);
162
163 KnownBits LHSKnown(BitWidth), RHSKnown(BitWidth);
164 // If this is the root being simplified, allow it to have multiple uses,
165 // just set the DemandedMask to all bits so that we can try to simplify the
166 // operands. This allows visitTruncInst (for example) to simplify the
167 // operand of a trunc without duplicating all the logic below.
168 if (Depth == 0 && !V->hasOneUse())
169 DemandedMask.setAllBits();
170
171 // Update flags after simplifying an operand based on the fact that some high
172 // order bits are not demanded.
173 auto disableWrapFlagsBasedOnUnusedHighBits = [](Instruction *I,
174 unsigned NLZ) {
175 if (NLZ > 0) {
176 // Disable the nsw and nuw flags here: We can no longer guarantee that
177 // we won't wrap after simplification. Removing the nsw/nuw flags is
178 // legal here because the top bit is not demanded.
179 I->setHasNoSignedWrap(false);
180 I->setHasNoUnsignedWrap(false);
181 }
182 return I;
183 };
184
185 // If the high-bits of an ADD/SUB/MUL are not demanded, then we do not care
186 // about the high bits of the operands.
187 auto simplifyOperandsBasedOnUnusedHighBits = [&](APInt &DemandedFromOps) {
188 unsigned NLZ = DemandedMask.countl_zero();
189 // Right fill the mask of bits for the operands to demand the most
190 // significant bit and all those below it.
191 DemandedFromOps = APInt::getLowBitsSet(BitWidth, BitWidth - NLZ);
192 if (ShrinkDemandedConstant(I, 0, DemandedFromOps) ||
193 SimplifyDemandedBits(I, 0, DemandedFromOps, LHSKnown, Depth + 1) ||
194 ShrinkDemandedConstant(I, 1, DemandedFromOps) ||
195 SimplifyDemandedBits(I, 1, DemandedFromOps, RHSKnown, Depth + 1)) {
196 disableWrapFlagsBasedOnUnusedHighBits(I, NLZ);
197 return true;
198 }
199 return false;
200 };
201
202 switch (I->getOpcode()) {
203 default:
204 computeKnownBits(I, Known, Depth, CxtI);
205 break;
206 case Instruction::And: {
207 // If either the LHS or the RHS are Zero, the result is zero.
208 if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) ||
209 SimplifyDemandedBits(I, 0, DemandedMask & ~RHSKnown.Zero, LHSKnown,
210 Depth + 1))
211 return I;
212 assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
213 assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
214
215 Known = analyzeKnownBitsFromAndXorOr(cast<Operator>(I), LHSKnown, RHSKnown,
217
218 // If the client is only demanding bits that we know, return the known
219 // constant.
220 if (DemandedMask.isSubsetOf(Known.Zero | Known.One))
221 return Constant::getIntegerValue(VTy, Known.One);
222
223 // If all of the demanded bits are known 1 on one side, return the other.
224 // These bits cannot contribute to the result of the 'and'.
225 if (DemandedMask.isSubsetOf(LHSKnown.Zero | RHSKnown.One))
226 return I->getOperand(0);
227 if (DemandedMask.isSubsetOf(RHSKnown.Zero | LHSKnown.One))
228 return I->getOperand(1);
229
230 // If the RHS is a constant, see if we can simplify it.
231 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnown.Zero))
232 return I;
233
234 break;
235 }
236 case Instruction::Or: {
237 // If either the LHS or the RHS are One, the result is One.
238 if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) ||
239 SimplifyDemandedBits(I, 0, DemandedMask & ~RHSKnown.One, LHSKnown,
240 Depth + 1)) {
241 // Disjoint flag may not longer hold.
242 I->dropPoisonGeneratingFlags();
243 return I;
244 }
245 assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
246 assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
247
248 Known = analyzeKnownBitsFromAndXorOr(cast<Operator>(I), LHSKnown, RHSKnown,
250
251 // If the client is only demanding bits that we know, return the known
252 // constant.
253 if (DemandedMask.isSubsetOf(Known.Zero | Known.One))
254 return Constant::getIntegerValue(VTy, Known.One);
255
256 // If all of the demanded bits are known zero on one side, return the other.
257 // These bits cannot contribute to the result of the 'or'.
258 if (DemandedMask.isSubsetOf(LHSKnown.One | RHSKnown.Zero))
259 return I->getOperand(0);
260 if (DemandedMask.isSubsetOf(RHSKnown.One | LHSKnown.Zero))
261 return I->getOperand(1);
262
263 // If the RHS is a constant, see if we can simplify it.
264 if (ShrinkDemandedConstant(I, 1, DemandedMask))
265 return I;
266
267 // Infer disjoint flag if no common bits are set.
268 if (!cast<PossiblyDisjointInst>(I)->isDisjoint()) {
269 WithCache<const Value *> LHSCache(I->getOperand(0), LHSKnown),
270 RHSCache(I->getOperand(1), RHSKnown);
271 if (haveNoCommonBitsSet(LHSCache, RHSCache, SQ.getWithInstruction(I))) {
272 cast<PossiblyDisjointInst>(I)->setIsDisjoint(true);
273 return I;
274 }
275 }
276
277 break;
278 }
279 case Instruction::Xor: {
280 if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) ||
281 SimplifyDemandedBits(I, 0, DemandedMask, LHSKnown, Depth + 1))
282 return I;
283 Value *LHS, *RHS;
284 if (DemandedMask == 1 &&
285 match(I->getOperand(0), m_Intrinsic<Intrinsic::ctpop>(m_Value(LHS))) &&
286 match(I->getOperand(1), m_Intrinsic<Intrinsic::ctpop>(m_Value(RHS)))) {
287 // (ctpop(X) ^ ctpop(Y)) & 1 --> ctpop(X^Y) & 1
290 auto *Xor = Builder.CreateXor(LHS, RHS);
291 return Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, Xor);
292 }
293
294 assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
295 assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
296
297 Known = analyzeKnownBitsFromAndXorOr(cast<Operator>(I), LHSKnown, RHSKnown,
299
300 // If the client is only demanding bits that we know, return the known
301 // constant.
302 if (DemandedMask.isSubsetOf(Known.Zero | Known.One))
303 return Constant::getIntegerValue(VTy, Known.One);
304
305 // If all of the demanded bits are known zero on one side, return the other.
306 // These bits cannot contribute to the result of the 'xor'.
307 if (DemandedMask.isSubsetOf(RHSKnown.Zero))
308 return I->getOperand(0);
309 if (DemandedMask.isSubsetOf(LHSKnown.Zero))
310 return I->getOperand(1);
311
312 // If all of the demanded bits are known to be zero on one side or the
313 // other, turn this into an *inclusive* or.
314 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
315 if (DemandedMask.isSubsetOf(RHSKnown.Zero | LHSKnown.Zero)) {
316 Instruction *Or =
317 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1));
318 if (DemandedMask.isAllOnes())
319 cast<PossiblyDisjointInst>(Or)->setIsDisjoint(true);
320 Or->takeName(I);
321 return InsertNewInstWith(Or, I->getIterator());
322 }
323
324 // If all of the demanded bits on one side are known, and all of the set
325 // bits on that side are also known to be set on the other side, turn this
326 // into an AND, as we know the bits will be cleared.
327 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
328 if (DemandedMask.isSubsetOf(RHSKnown.Zero|RHSKnown.One) &&
329 RHSKnown.One.isSubsetOf(LHSKnown.One)) {
331 ~RHSKnown.One & DemandedMask);
332 Instruction *And = BinaryOperator::CreateAnd(I->getOperand(0), AndC);
333 return InsertNewInstWith(And, I->getIterator());
334 }
335
336 // If the RHS is a constant, see if we can change it. Don't alter a -1
337 // constant because that's a canonical 'not' op, and that is better for
338 // combining, SCEV, and codegen.
339 const APInt *C;
340 if (match(I->getOperand(1), m_APInt(C)) && !C->isAllOnes()) {
341 if ((*C | ~DemandedMask).isAllOnes()) {
342 // Force bits to 1 to create a 'not' op.
343 I->setOperand(1, ConstantInt::getAllOnesValue(VTy));
344 return I;
345 }
346 // If we can't turn this into a 'not', try to shrink the constant.
347 if (ShrinkDemandedConstant(I, 1, DemandedMask))
348 return I;
349 }
350
351 // If our LHS is an 'and' and if it has one use, and if any of the bits we
352 // are flipping are known to be set, then the xor is just resetting those
353 // bits to zero. We can just knock out bits from the 'and' and the 'xor',
354 // simplifying both of them.
355 if (Instruction *LHSInst = dyn_cast<Instruction>(I->getOperand(0))) {
356 ConstantInt *AndRHS, *XorRHS;
357 if (LHSInst->getOpcode() == Instruction::And && LHSInst->hasOneUse() &&
358 match(I->getOperand(1), m_ConstantInt(XorRHS)) &&
359 match(LHSInst->getOperand(1), m_ConstantInt(AndRHS)) &&
360 (LHSKnown.One & RHSKnown.One & DemandedMask) != 0) {
361 APInt NewMask = ~(LHSKnown.One & RHSKnown.One & DemandedMask);
362
363 Constant *AndC = ConstantInt::get(VTy, NewMask & AndRHS->getValue());
364 Instruction *NewAnd = BinaryOperator::CreateAnd(I->getOperand(0), AndC);
365 InsertNewInstWith(NewAnd, I->getIterator());
366
367 Constant *XorC = ConstantInt::get(VTy, NewMask & XorRHS->getValue());
368 Instruction *NewXor = BinaryOperator::CreateXor(NewAnd, XorC);
369 return InsertNewInstWith(NewXor, I->getIterator());
370 }
371 }
372 break;
373 }
374 case Instruction::Select: {
375 if (SimplifyDemandedBits(I, 2, DemandedMask, RHSKnown, Depth + 1) ||
376 SimplifyDemandedBits(I, 1, DemandedMask, LHSKnown, Depth + 1))
377 return I;
378 assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
379 assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
380
381 // If the operands are constants, see if we can simplify them.
382 // This is similar to ShrinkDemandedConstant, but for a select we want to
383 // try to keep the selected constants the same as icmp value constants, if
384 // we can. This helps not break apart (or helps put back together)
385 // canonical patterns like min and max.
386 auto CanonicalizeSelectConstant = [](Instruction *I, unsigned OpNo,
387 const APInt &DemandedMask) {
388 const APInt *SelC;
389 if (!match(I->getOperand(OpNo), m_APInt(SelC)))
390 return false;
391
392 // Get the constant out of the ICmp, if there is one.
393 // Only try this when exactly 1 operand is a constant (if both operands
394 // are constant, the icmp should eventually simplify). Otherwise, we may
395 // invert the transform that reduces set bits and infinite-loop.
396 Value *X;
397 const APInt *CmpC;
399 if (!match(I->getOperand(0), m_ICmp(Pred, m_Value(X), m_APInt(CmpC))) ||
400 isa<Constant>(X) || CmpC->getBitWidth() != SelC->getBitWidth())
401 return ShrinkDemandedConstant(I, OpNo, DemandedMask);
402
403 // If the constant is already the same as the ICmp, leave it as-is.
404 if (*CmpC == *SelC)
405 return false;
406 // If the constants are not already the same, but can be with the demand
407 // mask, use the constant value from the ICmp.
408 if ((*CmpC & DemandedMask) == (*SelC & DemandedMask)) {
409 I->setOperand(OpNo, ConstantInt::get(I->getType(), *CmpC));
410 return true;
411 }
412 return ShrinkDemandedConstant(I, OpNo, DemandedMask);
413 };
414 if (CanonicalizeSelectConstant(I, 1, DemandedMask) ||
415 CanonicalizeSelectConstant(I, 2, DemandedMask))
416 return I;
417
418 // Only known if known in both the LHS and RHS.
419 Known = LHSKnown.intersectWith(RHSKnown);
420 break;
421 }
422 case Instruction::Trunc: {
423 // If we do not demand the high bits of a right-shifted and truncated value,
424 // then we may be able to truncate it before the shift.
425 Value *X;
426 const APInt *C;
427 if (match(I->getOperand(0), m_OneUse(m_LShr(m_Value(X), m_APInt(C))))) {
428 // The shift amount must be valid (not poison) in the narrow type, and
429 // it must not be greater than the high bits demanded of the result.
430 if (C->ult(VTy->getScalarSizeInBits()) &&
431 C->ule(DemandedMask.countl_zero())) {
432 // trunc (lshr X, C) --> lshr (trunc X), C
435 Value *Trunc = Builder.CreateTrunc(X, VTy);
436 return Builder.CreateLShr(Trunc, C->getZExtValue());
437 }
438 }
439 }
440 [[fallthrough]];
441 case Instruction::ZExt: {
442 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
443
444 APInt InputDemandedMask = DemandedMask.zextOrTrunc(SrcBitWidth);
445 KnownBits InputKnown(SrcBitWidth);
446 if (SimplifyDemandedBits(I, 0, InputDemandedMask, InputKnown, Depth + 1)) {
447 // For zext nneg, we may have dropped the instruction which made the
448 // input non-negative.
449 I->dropPoisonGeneratingFlags();
450 return I;
451 }
452 assert(InputKnown.getBitWidth() == SrcBitWidth && "Src width changed?");
453 if (I->getOpcode() == Instruction::ZExt && I->hasNonNeg() &&
454 !InputKnown.isNegative())
455 InputKnown.makeNonNegative();
456 Known = InputKnown.zextOrTrunc(BitWidth);
457
458 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
459 break;
460 }
461 case Instruction::SExt: {
462 // Compute the bits in the result that are not present in the input.
463 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
464
465 APInt InputDemandedBits = DemandedMask.trunc(SrcBitWidth);
466
467 // If any of the sign extended bits are demanded, we know that the sign
468 // bit is demanded.
469 if (DemandedMask.getActiveBits() > SrcBitWidth)
470 InputDemandedBits.setBit(SrcBitWidth-1);
471
472 KnownBits InputKnown(SrcBitWidth);
473 if (SimplifyDemandedBits(I, 0, InputDemandedBits, InputKnown, Depth + 1))
474 return I;
475
476 // If the input sign bit is known zero, or if the NewBits are not demanded
477 // convert this into a zero extension.
478 if (InputKnown.isNonNegative() ||
479 DemandedMask.getActiveBits() <= SrcBitWidth) {
480 // Convert to ZExt cast.
481 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy);
482 NewCast->takeName(I);
483 return InsertNewInstWith(NewCast, I->getIterator());
484 }
485
486 // If the sign bit of the input is known set or clear, then we know the
487 // top bits of the result.
488 Known = InputKnown.sext(BitWidth);
489 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
490 break;
491 }
492 case Instruction::Add: {
493 if ((DemandedMask & 1) == 0) {
494 // If we do not need the low bit, try to convert bool math to logic:
495 // add iN (zext i1 X), (sext i1 Y) --> sext (~X & Y) to iN
496 Value *X, *Y;
498 m_OneUse(m_SExt(m_Value(Y))))) &&
499 X->getType()->isIntOrIntVectorTy(1) && X->getType() == Y->getType()) {
500 // Truth table for inputs and output signbits:
501 // X:0 | X:1
502 // ----------
503 // Y:0 | 0 | 0 |
504 // Y:1 | -1 | 0 |
505 // ----------
509 return Builder.CreateSExt(AndNot, VTy);
510 }
511
512 // add iN (sext i1 X), (sext i1 Y) --> sext (X | Y) to iN
513 // TODO: Relax the one-use checks because we are removing an instruction?
515 m_OneUse(m_SExt(m_Value(Y))))) &&
516 X->getType()->isIntOrIntVectorTy(1) && X->getType() == Y->getType()) {
517 // Truth table for inputs and output signbits:
518 // X:0 | X:1
519 // -----------
520 // Y:0 | -1 | -1 |
521 // Y:1 | -1 | 0 |
522 // -----------
525 Value *Or = Builder.CreateOr(X, Y);
526 return Builder.CreateSExt(Or, VTy);
527 }
528 }
529
530 // Right fill the mask of bits for the operands to demand the most
531 // significant bit and all those below it.
532 unsigned NLZ = DemandedMask.countl_zero();
533 APInt DemandedFromOps = APInt::getLowBitsSet(BitWidth, BitWidth - NLZ);
534 if (ShrinkDemandedConstant(I, 1, DemandedFromOps) ||
535 SimplifyDemandedBits(I, 1, DemandedFromOps, RHSKnown, Depth + 1))
536 return disableWrapFlagsBasedOnUnusedHighBits(I, NLZ);
537
538 // If low order bits are not demanded and known to be zero in one operand,
539 // then we don't need to demand them from the other operand, since they
540 // can't cause overflow into any bits that are demanded in the result.
541 unsigned NTZ = (~DemandedMask & RHSKnown.Zero).countr_one();
542 APInt DemandedFromLHS = DemandedFromOps;
543 DemandedFromLHS.clearLowBits(NTZ);
544 if (ShrinkDemandedConstant(I, 0, DemandedFromLHS) ||
545 SimplifyDemandedBits(I, 0, DemandedFromLHS, LHSKnown, Depth + 1))
546 return disableWrapFlagsBasedOnUnusedHighBits(I, NLZ);
547
548 // If we are known to be adding zeros to every bit below
549 // the highest demanded bit, we just return the other side.
550 if (DemandedFromOps.isSubsetOf(RHSKnown.Zero))
551 return I->getOperand(0);
552 if (DemandedFromOps.isSubsetOf(LHSKnown.Zero))
553 return I->getOperand(1);
554
555 // (add X, C) --> (xor X, C) IFF C is equal to the top bit of the DemandMask
556 {
557 const APInt *C;
558 if (match(I->getOperand(1), m_APInt(C)) &&
559 C->isOneBitSet(DemandedMask.getActiveBits() - 1)) {
562 return Builder.CreateXor(I->getOperand(0), ConstantInt::get(VTy, *C));
563 }
564 }
565
566 // Otherwise just compute the known bits of the result.
567 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
568 bool NUW = cast<OverflowingBinaryOperator>(I)->hasNoUnsignedWrap();
569 Known = KnownBits::computeForAddSub(true, NSW, NUW, LHSKnown, RHSKnown);
570 break;
571 }
572 case Instruction::Sub: {
573 // Right fill the mask of bits for the operands to demand the most
574 // significant bit and all those below it.
575 unsigned NLZ = DemandedMask.countl_zero();
576 APInt DemandedFromOps = APInt::getLowBitsSet(BitWidth, BitWidth - NLZ);
577 if (ShrinkDemandedConstant(I, 1, DemandedFromOps) ||
578 SimplifyDemandedBits(I, 1, DemandedFromOps, RHSKnown, Depth + 1))
579 return disableWrapFlagsBasedOnUnusedHighBits(I, NLZ);
580
581 // If low order bits are not demanded and are known to be zero in RHS,
582 // then we don't need to demand them from LHS, since they can't cause a
583 // borrow from any bits that are demanded in the result.
584 unsigned NTZ = (~DemandedMask & RHSKnown.Zero).countr_one();
585 APInt DemandedFromLHS = DemandedFromOps;
586 DemandedFromLHS.clearLowBits(NTZ);
587 if (ShrinkDemandedConstant(I, 0, DemandedFromLHS) ||
588 SimplifyDemandedBits(I, 0, DemandedFromLHS, LHSKnown, Depth + 1))
589 return disableWrapFlagsBasedOnUnusedHighBits(I, NLZ);
590
591 // If we are known to be subtracting zeros from every bit below
592 // the highest demanded bit, we just return the other side.
593 if (DemandedFromOps.isSubsetOf(RHSKnown.Zero))
594 return I->getOperand(0);
595 // We can't do this with the LHS for subtraction, unless we are only
596 // demanding the LSB.
597 if (DemandedFromOps.isOne() && DemandedFromOps.isSubsetOf(LHSKnown.Zero))
598 return I->getOperand(1);
599
600 // Otherwise just compute the known bits of the result.
601 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
602 bool NUW = cast<OverflowingBinaryOperator>(I)->hasNoUnsignedWrap();
603 Known = KnownBits::computeForAddSub(false, NSW, NUW, LHSKnown, RHSKnown);
604 break;
605 }
606 case Instruction::Mul: {
607 APInt DemandedFromOps;
608 if (simplifyOperandsBasedOnUnusedHighBits(DemandedFromOps))
609 return I;
610
611 if (DemandedMask.isPowerOf2()) {
612 // The LSB of X*Y is set only if (X & 1) == 1 and (Y & 1) == 1.
613 // If we demand exactly one bit N and we have "X * (C' << N)" where C' is
614 // odd (has LSB set), then the left-shifted low bit of X is the answer.
615 unsigned CTZ = DemandedMask.countr_zero();
616 const APInt *C;
617 if (match(I->getOperand(1), m_APInt(C)) && C->countr_zero() == CTZ) {
618 Constant *ShiftC = ConstantInt::get(VTy, CTZ);
619 Instruction *Shl = BinaryOperator::CreateShl(I->getOperand(0), ShiftC);
620 return InsertNewInstWith(Shl, I->getIterator());
621 }
622 }
623 // For a squared value "X * X", the bottom 2 bits are 0 and X[0] because:
624 // X * X is odd iff X is odd.
625 // 'Quadratic Reciprocity': X * X -> 0 for bit[1]
626 if (I->getOperand(0) == I->getOperand(1) && DemandedMask.ult(4)) {
627 Constant *One = ConstantInt::get(VTy, 1);
628 Instruction *And1 = BinaryOperator::CreateAnd(I->getOperand(0), One);
629 return InsertNewInstWith(And1, I->getIterator());
630 }
631
632 computeKnownBits(I, Known, Depth, CxtI);
633 break;
634 }
635 case Instruction::Shl: {
636 const APInt *SA;
637 if (match(I->getOperand(1), m_APInt(SA))) {
638 const APInt *ShrAmt;
639 if (match(I->getOperand(0), m_Shr(m_Value(), m_APInt(ShrAmt))))
640 if (Instruction *Shr = dyn_cast<Instruction>(I->getOperand(0)))
641 if (Value *R = simplifyShrShlDemandedBits(Shr, *ShrAmt, I, *SA,
642 DemandedMask, Known))
643 return R;
644
645 // Do not simplify if shl is part of funnel-shift pattern
646 if (I->hasOneUse()) {
647 auto *Inst = dyn_cast<Instruction>(I->user_back());
648 if (Inst && Inst->getOpcode() == BinaryOperator::Or) {
649 if (auto Opt = convertOrOfShiftsToFunnelShift(*Inst)) {
650 auto [IID, FShiftArgs] = *Opt;
651 if ((IID == Intrinsic::fshl || IID == Intrinsic::fshr) &&
652 FShiftArgs[0] == FShiftArgs[1])
653 return nullptr;
654 }
655 }
656 }
657
658 // We only want bits that already match the signbit then we don't
659 // need to shift.
660 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth - 1);
661 if (DemandedMask.countr_zero() >= ShiftAmt) {
662 if (I->hasNoSignedWrap()) {
663 unsigned NumHiDemandedBits = BitWidth - DemandedMask.countr_zero();
664 unsigned SignBits =
665 ComputeNumSignBits(I->getOperand(0), Depth + 1, CxtI);
666 if (SignBits > ShiftAmt && SignBits - ShiftAmt >= NumHiDemandedBits)
667 return I->getOperand(0);
668 }
669
670 // If we can pre-shift a right-shifted constant to the left without
671 // losing any high bits and we don't demand the low bits, then eliminate
672 // the left-shift:
673 // (C >> X) << LeftShiftAmtC --> (C << LeftShiftAmtC) >> X
674 Value *X;
675 Constant *C;
676 if (match(I->getOperand(0), m_LShr(m_ImmConstant(C), m_Value(X)))) {
677 Constant *LeftShiftAmtC = ConstantInt::get(VTy, ShiftAmt);
678 Constant *NewC = ConstantFoldBinaryOpOperands(Instruction::Shl, C,
679 LeftShiftAmtC, DL);
680 if (ConstantFoldBinaryOpOperands(Instruction::LShr, NewC,
681 LeftShiftAmtC, DL) == C) {
682 Instruction *Lshr = BinaryOperator::CreateLShr(NewC, X);
683 return InsertNewInstWith(Lshr, I->getIterator());
684 }
685 }
686 }
687
688 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
689
690 // If the shift is NUW/NSW, then it does demand the high bits.
691 ShlOperator *IOp = cast<ShlOperator>(I);
692 if (IOp->hasNoSignedWrap())
693 DemandedMaskIn.setHighBits(ShiftAmt+1);
694 else if (IOp->hasNoUnsignedWrap())
695 DemandedMaskIn.setHighBits(ShiftAmt);
696
697 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1))
698 return I;
699 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
700
701 Known = KnownBits::shl(Known,
703 /* NUW */ IOp->hasNoUnsignedWrap(),
704 /* NSW */ IOp->hasNoSignedWrap());
705 } else {
706 // This is a variable shift, so we can't shift the demand mask by a known
707 // amount. But if we are not demanding high bits, then we are not
708 // demanding those bits from the pre-shifted operand either.
709 if (unsigned CTLZ = DemandedMask.countl_zero()) {
710 APInt DemandedFromOp(APInt::getLowBitsSet(BitWidth, BitWidth - CTLZ));
711 if (SimplifyDemandedBits(I, 0, DemandedFromOp, Known, Depth + 1)) {
712 // We can't guarantee that nsw/nuw hold after simplifying the operand.
713 I->dropPoisonGeneratingFlags();
714 return I;
715 }
716 }
717 computeKnownBits(I, Known, Depth, CxtI);
718 }
719 break;
720 }
721 case Instruction::LShr: {
722 const APInt *SA;
723 if (match(I->getOperand(1), m_APInt(SA))) {
724 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
725
726 // Do not simplify if lshr is part of funnel-shift pattern
727 if (I->hasOneUse()) {
728 auto *Inst = dyn_cast<Instruction>(I->user_back());
729 if (Inst && Inst->getOpcode() == BinaryOperator::Or) {
730 if (auto Opt = convertOrOfShiftsToFunnelShift(*Inst)) {
731 auto [IID, FShiftArgs] = *Opt;
732 if ((IID == Intrinsic::fshl || IID == Intrinsic::fshr) &&
733 FShiftArgs[0] == FShiftArgs[1])
734 return nullptr;
735 }
736 }
737 }
738
739 // If we are just demanding the shifted sign bit and below, then this can
740 // be treated as an ASHR in disguise.
741 if (DemandedMask.countl_zero() >= ShiftAmt) {
742 // If we only want bits that already match the signbit then we don't
743 // need to shift.
744 unsigned NumHiDemandedBits = BitWidth - DemandedMask.countr_zero();
745 unsigned SignBits =
746 ComputeNumSignBits(I->getOperand(0), Depth + 1, CxtI);
747 if (SignBits >= NumHiDemandedBits)
748 return I->getOperand(0);
749
750 // If we can pre-shift a left-shifted constant to the right without
751 // losing any low bits (we already know we don't demand the high bits),
752 // then eliminate the right-shift:
753 // (C << X) >> RightShiftAmtC --> (C >> RightShiftAmtC) << X
754 Value *X;
755 Constant *C;
756 if (match(I->getOperand(0), m_Shl(m_ImmConstant(C), m_Value(X)))) {
757 Constant *RightShiftAmtC = ConstantInt::get(VTy, ShiftAmt);
758 Constant *NewC = ConstantFoldBinaryOpOperands(Instruction::LShr, C,
759 RightShiftAmtC, DL);
760 if (ConstantFoldBinaryOpOperands(Instruction::Shl, NewC,
761 RightShiftAmtC, DL) == C) {
762 Instruction *Shl = BinaryOperator::CreateShl(NewC, X);
763 return InsertNewInstWith(Shl, I->getIterator());
764 }
765 }
766 }
767
768 // Unsigned shift right.
769 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
770 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1)) {
771 // exact flag may not longer hold.
772 I->dropPoisonGeneratingFlags();
773 return I;
774 }
775 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
776 Known.Zero.lshrInPlace(ShiftAmt);
777 Known.One.lshrInPlace(ShiftAmt);
778 if (ShiftAmt)
779 Known.Zero.setHighBits(ShiftAmt); // high bits known zero.
780 } else {
781 computeKnownBits(I, Known, Depth, CxtI);
782 }
783 break;
784 }
785 case Instruction::AShr: {
786 unsigned SignBits = ComputeNumSignBits(I->getOperand(0), Depth + 1, CxtI);
787
788 // If we only want bits that already match the signbit then we don't need
789 // to shift.
790 unsigned NumHiDemandedBits = BitWidth - DemandedMask.countr_zero();
791 if (SignBits >= NumHiDemandedBits)
792 return I->getOperand(0);
793
794 // If this is an arithmetic shift right and only the low-bit is set, we can
795 // always convert this into a logical shr, even if the shift amount is
796 // variable. The low bit of the shift cannot be an input sign bit unless
797 // the shift amount is >= the size of the datatype, which is undefined.
798 if (DemandedMask.isOne()) {
799 // Perform the logical shift right.
800 Instruction *NewVal = BinaryOperator::CreateLShr(
801 I->getOperand(0), I->getOperand(1), I->getName());
802 return InsertNewInstWith(NewVal, I->getIterator());
803 }
804
805 const APInt *SA;
806 if (match(I->getOperand(1), m_APInt(SA))) {
807 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
808
809 // Signed shift right.
810 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
811 // If any of the high bits are demanded, we should set the sign bit as
812 // demanded.
813 if (DemandedMask.countl_zero() <= ShiftAmt)
814 DemandedMaskIn.setSignBit();
815
816 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1)) {
817 // exact flag may not longer hold.
818 I->dropPoisonGeneratingFlags();
819 return I;
820 }
821
822 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
823 // Compute the new bits that are at the top now plus sign bits.
825 BitWidth, std::min(SignBits + ShiftAmt - 1, BitWidth)));
826 Known.Zero.lshrInPlace(ShiftAmt);
827 Known.One.lshrInPlace(ShiftAmt);
828
829 // If the input sign bit is known to be zero, or if none of the top bits
830 // are demanded, turn this into an unsigned shift right.
831 assert(BitWidth > ShiftAmt && "Shift amount not saturated?");
832 if (Known.Zero[BitWidth-ShiftAmt-1] ||
833 !DemandedMask.intersects(HighBits)) {
834 BinaryOperator *LShr = BinaryOperator::CreateLShr(I->getOperand(0),
835 I->getOperand(1));
836 LShr->setIsExact(cast<BinaryOperator>(I)->isExact());
837 LShr->takeName(I);
838 return InsertNewInstWith(LShr, I->getIterator());
839 } else if (Known.One[BitWidth-ShiftAmt-1]) { // New bits are known one.
840 Known.One |= HighBits;
841 // SignBits may be out-of-sync with Known.countMinSignBits(). Mask out
842 // high bits of Known.Zero to avoid conflicts.
843 Known.Zero &= ~HighBits;
844 }
845 } else {
846 computeKnownBits(I, Known, Depth, CxtI);
847 }
848 break;
849 }
850 case Instruction::UDiv: {
851 // UDiv doesn't demand low bits that are zero in the divisor.
852 const APInt *SA;
853 if (match(I->getOperand(1), m_APInt(SA))) {
854 // TODO: Take the demanded mask of the result into account.
855 unsigned RHSTrailingZeros = SA->countr_zero();
856 APInt DemandedMaskIn =
857 APInt::getHighBitsSet(BitWidth, BitWidth - RHSTrailingZeros);
858 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, LHSKnown, Depth + 1)) {
859 // We can't guarantee that "exact" is still true after changing the
860 // the dividend.
861 I->dropPoisonGeneratingFlags();
862 return I;
863 }
864
865 Known = KnownBits::udiv(LHSKnown, KnownBits::makeConstant(*SA),
866 cast<BinaryOperator>(I)->isExact());
867 } else {
868 computeKnownBits(I, Known, Depth, CxtI);
869 }
870 break;
871 }
872 case Instruction::SRem: {
873 const APInt *Rem;
874 if (match(I->getOperand(1), m_APInt(Rem))) {
875 // X % -1 demands all the bits because we don't want to introduce
876 // INT_MIN % -1 (== undef) by accident.
877 if (Rem->isAllOnes())
878 break;
879 APInt RA = Rem->abs();
880 if (RA.isPowerOf2()) {
881 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
882 return I->getOperand(0);
883
884 APInt LowBits = RA - 1;
885 APInt Mask2 = LowBits | APInt::getSignMask(BitWidth);
886 if (SimplifyDemandedBits(I, 0, Mask2, LHSKnown, Depth + 1))
887 return I;
888
889 // The low bits of LHS are unchanged by the srem.
890 Known.Zero = LHSKnown.Zero & LowBits;
891 Known.One = LHSKnown.One & LowBits;
892
893 // If LHS is non-negative or has all low bits zero, then the upper bits
894 // are all zero.
895 if (LHSKnown.isNonNegative() || LowBits.isSubsetOf(LHSKnown.Zero))
896 Known.Zero |= ~LowBits;
897
898 // If LHS is negative and not all low bits are zero, then the upper bits
899 // are all one.
900 if (LHSKnown.isNegative() && LowBits.intersects(LHSKnown.One))
901 Known.One |= ~LowBits;
902
903 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
904 break;
905 }
906 }
907
908 computeKnownBits(I, Known, Depth, CxtI);
909 break;
910 }
911 case Instruction::URem: {
913 if (SimplifyDemandedBits(I, 0, AllOnes, LHSKnown, Depth + 1) ||
914 SimplifyDemandedBits(I, 1, AllOnes, RHSKnown, Depth + 1))
915 return I;
916
917 Known = KnownBits::urem(LHSKnown, RHSKnown);
918 break;
919 }
920 case Instruction::Call: {
921 bool KnownBitsComputed = false;
922 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
923 switch (II->getIntrinsicID()) {
924 case Intrinsic::abs: {
925 if (DemandedMask == 1)
926 return II->getArgOperand(0);
927 break;
928 }
929 case Intrinsic::ctpop: {
930 // Checking if the number of clear bits is odd (parity)? If the type has
931 // an even number of bits, that's the same as checking if the number of
932 // set bits is odd, so we can eliminate the 'not' op.
933 Value *X;
934 if (DemandedMask == 1 && VTy->getScalarSizeInBits() % 2 == 0 &&
935 match(II->getArgOperand(0), m_Not(m_Value(X)))) {
937 II->getModule(), Intrinsic::ctpop, VTy);
938 return InsertNewInstWith(CallInst::Create(Ctpop, {X}), I->getIterator());
939 }
940 break;
941 }
942 case Intrinsic::bswap: {
943 // If the only bits demanded come from one byte of the bswap result,
944 // just shift the input byte into position to eliminate the bswap.
945 unsigned NLZ = DemandedMask.countl_zero();
946 unsigned NTZ = DemandedMask.countr_zero();
947
948 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
949 // we need all the bits down to bit 8. Likewise, round NLZ. If we
950 // have 14 leading zeros, round to 8.
951 NLZ = alignDown(NLZ, 8);
952 NTZ = alignDown(NTZ, 8);
953 // If we need exactly one byte, we can do this transformation.
954 if (BitWidth - NLZ - NTZ == 8) {
955 // Replace this with either a left or right shift to get the byte into
956 // the right place.
957 Instruction *NewVal;
958 if (NLZ > NTZ)
959 NewVal = BinaryOperator::CreateLShr(
960 II->getArgOperand(0), ConstantInt::get(VTy, NLZ - NTZ));
961 else
962 NewVal = BinaryOperator::CreateShl(
963 II->getArgOperand(0), ConstantInt::get(VTy, NTZ - NLZ));
964 NewVal->takeName(I);
965 return InsertNewInstWith(NewVal, I->getIterator());
966 }
967 break;
968 }
969 case Intrinsic::ptrmask: {
970 unsigned MaskWidth = I->getOperand(1)->getType()->getScalarSizeInBits();
971 RHSKnown = KnownBits(MaskWidth);
972 // If either the LHS or the RHS are Zero, the result is zero.
973 if (SimplifyDemandedBits(I, 0, DemandedMask, LHSKnown, Depth + 1) ||
975 I, 1, (DemandedMask & ~LHSKnown.Zero).zextOrTrunc(MaskWidth),
976 RHSKnown, Depth + 1))
977 return I;
978
979 // TODO: Should be 1-extend
980 RHSKnown = RHSKnown.anyextOrTrunc(BitWidth);
981 assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
982 assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
983
984 Known = LHSKnown & RHSKnown;
985 KnownBitsComputed = true;
986
987 // If the client is only demanding bits we know to be zero, return
988 // `llvm.ptrmask(p, 0)`. We can't return `null` here due to pointer
989 // provenance, but making the mask zero will be easily optimizable in
990 // the backend.
991 if (DemandedMask.isSubsetOf(Known.Zero) &&
992 !match(I->getOperand(1), m_Zero()))
993 return replaceOperand(
994 *I, 1, Constant::getNullValue(I->getOperand(1)->getType()));
995
996 // Mask in demanded space does nothing.
997 // NOTE: We may have attributes associated with the return value of the
998 // llvm.ptrmask intrinsic that will be lost when we just return the
999 // operand. We should try to preserve them.
1000 if (DemandedMask.isSubsetOf(RHSKnown.One | LHSKnown.Zero))
1001 return I->getOperand(0);
1002
1003 // If the RHS is a constant, see if we can simplify it.
1005 I, 1, (DemandedMask & ~LHSKnown.Zero).zextOrTrunc(MaskWidth)))
1006 return I;
1007
1008 // Combine:
1009 // (ptrmask (getelementptr i8, ptr p, imm i), imm mask)
1010 // -> (ptrmask (getelementptr i8, ptr p, imm (i & mask)), imm mask)
1011 // where only the low bits known to be zero in the pointer are changed
1012 Value *InnerPtr;
1013 uint64_t GEPIndex;
1014 uint64_t PtrMaskImmediate;
1015 if (match(I, m_Intrinsic<Intrinsic::ptrmask>(
1016 m_PtrAdd(m_Value(InnerPtr), m_ConstantInt(GEPIndex)),
1017 m_ConstantInt(PtrMaskImmediate)))) {
1018
1019 LHSKnown = computeKnownBits(InnerPtr, Depth + 1, I);
1020 if (!LHSKnown.isZero()) {
1021 const unsigned trailingZeros = LHSKnown.countMinTrailingZeros();
1022 uint64_t PointerAlignBits = (uint64_t(1) << trailingZeros) - 1;
1023
1024 uint64_t HighBitsGEPIndex = GEPIndex & ~PointerAlignBits;
1025 uint64_t MaskedLowBitsGEPIndex =
1026 GEPIndex & PointerAlignBits & PtrMaskImmediate;
1027
1028 uint64_t MaskedGEPIndex = HighBitsGEPIndex | MaskedLowBitsGEPIndex;
1029
1030 if (MaskedGEPIndex != GEPIndex) {
1031 auto *GEP = cast<GetElementPtrInst>(II->getArgOperand(0));
1033 Type *GEPIndexType =
1034 DL.getIndexType(GEP->getPointerOperand()->getType());
1035 Value *MaskedGEP = Builder.CreateGEP(
1036 GEP->getSourceElementType(), InnerPtr,
1037 ConstantInt::get(GEPIndexType, MaskedGEPIndex),
1038 GEP->getName(), GEP->isInBounds());
1039
1040 replaceOperand(*I, 0, MaskedGEP);
1041 return I;
1042 }
1043 }
1044 }
1045
1046 break;
1047 }
1048
1049 case Intrinsic::fshr:
1050 case Intrinsic::fshl: {
1051 const APInt *SA;
1052 if (!match(I->getOperand(2), m_APInt(SA)))
1053 break;
1054
1055 // Normalize to funnel shift left. APInt shifts of BitWidth are well-
1056 // defined, so no need to special-case zero shifts here.
1057 uint64_t ShiftAmt = SA->urem(BitWidth);
1058 if (II->getIntrinsicID() == Intrinsic::fshr)
1059 ShiftAmt = BitWidth - ShiftAmt;
1060
1061 APInt DemandedMaskLHS(DemandedMask.lshr(ShiftAmt));
1062 APInt DemandedMaskRHS(DemandedMask.shl(BitWidth - ShiftAmt));
1063 if (I->getOperand(0) != I->getOperand(1)) {
1064 if (SimplifyDemandedBits(I, 0, DemandedMaskLHS, LHSKnown,
1065 Depth + 1) ||
1066 SimplifyDemandedBits(I, 1, DemandedMaskRHS, RHSKnown, Depth + 1))
1067 return I;
1068 } else { // fshl is a rotate
1069 // Avoid converting rotate into funnel shift.
1070 // Only simplify if one operand is constant.
1071 LHSKnown = computeKnownBits(I->getOperand(0), Depth + 1, I);
1072 if (DemandedMaskLHS.isSubsetOf(LHSKnown.Zero | LHSKnown.One) &&
1073 !match(I->getOperand(0), m_SpecificInt(LHSKnown.One))) {
1074 replaceOperand(*I, 0, Constant::getIntegerValue(VTy, LHSKnown.One));
1075 return I;
1076 }
1077
1078 RHSKnown = computeKnownBits(I->getOperand(1), Depth + 1, I);
1079 if (DemandedMaskRHS.isSubsetOf(RHSKnown.Zero | RHSKnown.One) &&
1080 !match(I->getOperand(1), m_SpecificInt(RHSKnown.One))) {
1081 replaceOperand(*I, 1, Constant::getIntegerValue(VTy, RHSKnown.One));
1082 return I;
1083 }
1084 }
1085
1086 Known.Zero = LHSKnown.Zero.shl(ShiftAmt) |
1087 RHSKnown.Zero.lshr(BitWidth - ShiftAmt);
1088 Known.One = LHSKnown.One.shl(ShiftAmt) |
1089 RHSKnown.One.lshr(BitWidth - ShiftAmt);
1090 KnownBitsComputed = true;
1091 break;
1092 }
1093 case Intrinsic::umax: {
1094 // UMax(A, C) == A if ...
1095 // The lowest non-zero bit of DemandMask is higher than the highest
1096 // non-zero bit of C.
1097 const APInt *C;
1098 unsigned CTZ = DemandedMask.countr_zero();
1099 if (match(II->getArgOperand(1), m_APInt(C)) &&
1100 CTZ >= C->getActiveBits())
1101 return II->getArgOperand(0);
1102 break;
1103 }
1104 case Intrinsic::umin: {
1105 // UMin(A, C) == A if ...
1106 // The lowest non-zero bit of DemandMask is higher than the highest
1107 // non-one bit of C.
1108 // This comes from using DeMorgans on the above umax example.
1109 const APInt *C;
1110 unsigned CTZ = DemandedMask.countr_zero();
1111 if (match(II->getArgOperand(1), m_APInt(C)) &&
1112 CTZ >= C->getBitWidth() - C->countl_one())
1113 return II->getArgOperand(0);
1114 break;
1115 }
1116 default: {
1117 // Handle target specific intrinsics
1118 std::optional<Value *> V = targetSimplifyDemandedUseBitsIntrinsic(
1119 *II, DemandedMask, Known, KnownBitsComputed);
1120 if (V)
1121 return *V;
1122 break;
1123 }
1124 }
1125 }
1126
1127 if (!KnownBitsComputed)
1128 computeKnownBits(V, Known, Depth, CxtI);
1129 break;
1130 }
1131 }
1132
1133 if (V->getType()->isPointerTy()) {
1134 Align Alignment = V->getPointerAlignment(DL);
1135 Known.Zero.setLowBits(Log2(Alignment));
1136 }
1137
1138 // If the client is only demanding bits that we know, return the known
1139 // constant. We can't directly simplify pointers as a constant because of
1140 // pointer provenance.
1141 // TODO: We could return `(inttoptr const)` for pointers.
1142 if (!V->getType()->isPointerTy() && DemandedMask.isSubsetOf(Known.Zero | Known.One))
1143 return Constant::getIntegerValue(VTy, Known.One);
1144
1145 if (VerifyKnownBits) {
1146 KnownBits ReferenceKnown = computeKnownBits(V, Depth, CxtI);
1147 if (Known != ReferenceKnown) {
1148 errs() << "Mismatched known bits for " << *V << " in "
1149 << I->getFunction()->getName() << "\n";
1150 errs() << "computeKnownBits(): " << ReferenceKnown << "\n";
1151 errs() << "SimplifyDemandedBits(): " << Known << "\n";
1152 std::abort();
1153 }
1154 }
1155
1156 return nullptr;
1157}
1158
1159/// Helper routine of SimplifyDemandedUseBits. It computes Known
1160/// bits. It also tries to handle simplifications that can be done based on
1161/// DemandedMask, but without modifying the Instruction.
1163 Instruction *I, const APInt &DemandedMask, KnownBits &Known, unsigned Depth,
1164 Instruction *CxtI) {
1165 unsigned BitWidth = DemandedMask.getBitWidth();
1166 Type *ITy = I->getType();
1167
1168 KnownBits LHSKnown(BitWidth);
1169 KnownBits RHSKnown(BitWidth);
1170
1171 // Despite the fact that we can't simplify this instruction in all User's
1172 // context, we can at least compute the known bits, and we can
1173 // do simplifications that apply to *just* the one user if we know that
1174 // this instruction has a simpler value in that context.
1175 switch (I->getOpcode()) {
1176 case Instruction::And: {
1177 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
1178 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1, CxtI);
1179 Known = analyzeKnownBitsFromAndXorOr(cast<Operator>(I), LHSKnown, RHSKnown,
1180 Depth, SQ.getWithInstruction(CxtI));
1182
1183 // If the client is only demanding bits that we know, return the known
1184 // constant.
1185 if (DemandedMask.isSubsetOf(Known.Zero | Known.One))
1186 return Constant::getIntegerValue(ITy, Known.One);
1187
1188 // If all of the demanded bits are known 1 on one side, return the other.
1189 // These bits cannot contribute to the result of the 'and' in this context.
1190 if (DemandedMask.isSubsetOf(LHSKnown.Zero | RHSKnown.One))
1191 return I->getOperand(0);
1192 if (DemandedMask.isSubsetOf(RHSKnown.Zero | LHSKnown.One))
1193 return I->getOperand(1);
1194
1195 break;
1196 }
1197 case Instruction::Or: {
1198 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
1199 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1, CxtI);
1200 Known = analyzeKnownBitsFromAndXorOr(cast<Operator>(I), LHSKnown, RHSKnown,
1201 Depth, SQ.getWithInstruction(CxtI));
1203
1204 // If the client is only demanding bits that we know, return the known
1205 // constant.
1206 if (DemandedMask.isSubsetOf(Known.Zero | Known.One))
1207 return Constant::getIntegerValue(ITy, Known.One);
1208
1209 // We can simplify (X|Y) -> X or Y in the user's context if we know that
1210 // only bits from X or Y are demanded.
1211 // If all of the demanded bits are known zero on one side, return the other.
1212 // These bits cannot contribute to the result of the 'or' in this context.
1213 if (DemandedMask.isSubsetOf(LHSKnown.One | RHSKnown.Zero))
1214 return I->getOperand(0);
1215 if (DemandedMask.isSubsetOf(RHSKnown.One | LHSKnown.Zero))
1216 return I->getOperand(1);
1217
1218 break;
1219 }
1220 case Instruction::Xor: {
1221 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
1222 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1, CxtI);
1223 Known = analyzeKnownBitsFromAndXorOr(cast<Operator>(I), LHSKnown, RHSKnown,
1224 Depth, SQ.getWithInstruction(CxtI));
1226
1227 // If the client is only demanding bits that we know, return the known
1228 // constant.
1229 if (DemandedMask.isSubsetOf(Known.Zero | Known.One))
1230 return Constant::getIntegerValue(ITy, Known.One);
1231
1232 // We can simplify (X^Y) -> X or Y in the user's context if we know that
1233 // only bits from X or Y are demanded.
1234 // If all of the demanded bits are known zero on one side, return the other.
1235 if (DemandedMask.isSubsetOf(RHSKnown.Zero))
1236 return I->getOperand(0);
1237 if (DemandedMask.isSubsetOf(LHSKnown.Zero))
1238 return I->getOperand(1);
1239
1240 break;
1241 }
1242 case Instruction::Add: {
1243 unsigned NLZ = DemandedMask.countl_zero();
1244 APInt DemandedFromOps = APInt::getLowBitsSet(BitWidth, BitWidth - NLZ);
1245
1246 // If an operand adds zeros to every bit below the highest demanded bit,
1247 // that operand doesn't change the result. Return the other side.
1248 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
1249 if (DemandedFromOps.isSubsetOf(RHSKnown.Zero))
1250 return I->getOperand(0);
1251
1252 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1, CxtI);
1253 if (DemandedFromOps.isSubsetOf(LHSKnown.Zero))
1254 return I->getOperand(1);
1255
1256 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1257 bool NUW = cast<OverflowingBinaryOperator>(I)->hasNoUnsignedWrap();
1258 Known =
1259 KnownBits::computeForAddSub(/*Add=*/true, NSW, NUW, LHSKnown, RHSKnown);
1261 break;
1262 }
1263 case Instruction::Sub: {
1264 unsigned NLZ = DemandedMask.countl_zero();
1265 APInt DemandedFromOps = APInt::getLowBitsSet(BitWidth, BitWidth - NLZ);
1266
1267 // If an operand subtracts zeros from every bit below the highest demanded
1268 // bit, that operand doesn't change the result. Return the other side.
1269 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
1270 if (DemandedFromOps.isSubsetOf(RHSKnown.Zero))
1271 return I->getOperand(0);
1272
1273 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1274 bool NUW = cast<OverflowingBinaryOperator>(I)->hasNoUnsignedWrap();
1275 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1, CxtI);
1276 Known = KnownBits::computeForAddSub(/*Add=*/false, NSW, NUW, LHSKnown,
1277 RHSKnown);
1279 break;
1280 }
1281 case Instruction::AShr: {
1282 // Compute the Known bits to simplify things downstream.
1283 computeKnownBits(I, Known, Depth, CxtI);
1284
1285 // If this user is only demanding bits that we know, return the known
1286 // constant.
1287 if (DemandedMask.isSubsetOf(Known.Zero | Known.One))
1288 return Constant::getIntegerValue(ITy, Known.One);
1289
1290 // If the right shift operand 0 is a result of a left shift by the same
1291 // amount, this is probably a zero/sign extension, which may be unnecessary,
1292 // if we do not demand any of the new sign bits. So, return the original
1293 // operand instead.
1294 const APInt *ShiftRC;
1295 const APInt *ShiftLC;
1296 Value *X;
1297 unsigned BitWidth = DemandedMask.getBitWidth();
1298 if (match(I,
1299 m_AShr(m_Shl(m_Value(X), m_APInt(ShiftLC)), m_APInt(ShiftRC))) &&
1300 ShiftLC == ShiftRC && ShiftLC->ult(BitWidth) &&
1301 DemandedMask.isSubsetOf(APInt::getLowBitsSet(
1302 BitWidth, BitWidth - ShiftRC->getZExtValue()))) {
1303 return X;
1304 }
1305
1306 break;
1307 }
1308 default:
1309 // Compute the Known bits to simplify things downstream.
1310 computeKnownBits(I, Known, Depth, CxtI);
1311
1312 // If this user is only demanding bits that we know, return the known
1313 // constant.
1314 if (DemandedMask.isSubsetOf(Known.Zero|Known.One))
1315 return Constant::getIntegerValue(ITy, Known.One);
1316
1317 break;
1318 }
1319
1320 return nullptr;
1321}
1322
1323/// Helper routine of SimplifyDemandedUseBits. It tries to simplify
1324/// "E1 = (X lsr C1) << C2", where the C1 and C2 are constant, into
1325/// "E2 = X << (C2 - C1)" or "E2 = X >> (C1 - C2)", depending on the sign
1326/// of "C2-C1".
1327///
1328/// Suppose E1 and E2 are generally different in bits S={bm, bm+1,
1329/// ..., bn}, without considering the specific value X is holding.
1330/// This transformation is legal iff one of following conditions is hold:
1331/// 1) All the bit in S are 0, in this case E1 == E2.
1332/// 2) We don't care those bits in S, per the input DemandedMask.
1333/// 3) Combination of 1) and 2). Some bits in S are 0, and we don't care the
1334/// rest bits.
1335///
1336/// Currently we only test condition 2).
1337///
1338/// As with SimplifyDemandedUseBits, it returns NULL if the simplification was
1339/// not successful.
1341 Instruction *Shr, const APInt &ShrOp1, Instruction *Shl,
1342 const APInt &ShlOp1, const APInt &DemandedMask, KnownBits &Known) {
1343 if (!ShlOp1 || !ShrOp1)
1344 return nullptr; // No-op.
1345
1346 Value *VarX = Shr->getOperand(0);
1347 Type *Ty = VarX->getType();
1348 unsigned BitWidth = Ty->getScalarSizeInBits();
1349 if (ShlOp1.uge(BitWidth) || ShrOp1.uge(BitWidth))
1350 return nullptr; // Undef.
1351
1352 unsigned ShlAmt = ShlOp1.getZExtValue();
1353 unsigned ShrAmt = ShrOp1.getZExtValue();
1354
1355 Known.One.clearAllBits();
1356 Known.Zero.setLowBits(ShlAmt - 1);
1357 Known.Zero &= DemandedMask;
1358
1359 APInt BitMask1(APInt::getAllOnes(BitWidth));
1360 APInt BitMask2(APInt::getAllOnes(BitWidth));
1361
1362 bool isLshr = (Shr->getOpcode() == Instruction::LShr);
1363 BitMask1 = isLshr ? (BitMask1.lshr(ShrAmt) << ShlAmt) :
1364 (BitMask1.ashr(ShrAmt) << ShlAmt);
1365
1366 if (ShrAmt <= ShlAmt) {
1367 BitMask2 <<= (ShlAmt - ShrAmt);
1368 } else {
1369 BitMask2 = isLshr ? BitMask2.lshr(ShrAmt - ShlAmt):
1370 BitMask2.ashr(ShrAmt - ShlAmt);
1371 }
1372
1373 // Check if condition-2 (see the comment to this function) is satified.
1374 if ((BitMask1 & DemandedMask) == (BitMask2 & DemandedMask)) {
1375 if (ShrAmt == ShlAmt)
1376 return VarX;
1377
1378 if (!Shr->hasOneUse())
1379 return nullptr;
1380
1381 BinaryOperator *New;
1382 if (ShrAmt < ShlAmt) {
1383 Constant *Amt = ConstantInt::get(VarX->getType(), ShlAmt - ShrAmt);
1384 New = BinaryOperator::CreateShl(VarX, Amt);
1385 BinaryOperator *Orig = cast<BinaryOperator>(Shl);
1386 New->setHasNoSignedWrap(Orig->hasNoSignedWrap());
1387 New->setHasNoUnsignedWrap(Orig->hasNoUnsignedWrap());
1388 } else {
1389 Constant *Amt = ConstantInt::get(VarX->getType(), ShrAmt - ShlAmt);
1390 New = isLshr ? BinaryOperator::CreateLShr(VarX, Amt) :
1391 BinaryOperator::CreateAShr(VarX, Amt);
1392 if (cast<BinaryOperator>(Shr)->isExact())
1393 New->setIsExact(true);
1394 }
1395
1396 return InsertNewInstWith(New, Shl->getIterator());
1397 }
1398
1399 return nullptr;
1400}
1401
1402/// The specified value produces a vector with any number of elements.
1403/// This method analyzes which elements of the operand are poison and
1404/// returns that information in PoisonElts.
1405///
1406/// DemandedElts contains the set of elements that are actually used by the
1407/// caller, and by default (AllowMultipleUsers equals false) the value is
1408/// simplified only if it has a single caller. If AllowMultipleUsers is set
1409/// to true, DemandedElts refers to the union of sets of elements that are
1410/// used by all callers.
1411///
1412/// If the information about demanded elements can be used to simplify the
1413/// operation, the operation is simplified, then the resultant value is
1414/// returned. This returns null if no change was made.
1416 APInt DemandedElts,
1417 APInt &PoisonElts,
1418 unsigned Depth,
1419 bool AllowMultipleUsers) {
1420 // Cannot analyze scalable type. The number of vector elements is not a
1421 // compile-time constant.
1422 if (isa<ScalableVectorType>(V->getType()))
1423 return nullptr;
1424
1425 unsigned VWidth = cast<FixedVectorType>(V->getType())->getNumElements();
1426 APInt EltMask(APInt::getAllOnes(VWidth));
1427 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1428
1429 if (match(V, m_Poison())) {
1430 // If the entire vector is poison, just return this info.
1431 PoisonElts = EltMask;
1432 return nullptr;
1433 }
1434
1435 if (DemandedElts.isZero()) { // If nothing is demanded, provide poison.
1436 PoisonElts = EltMask;
1437 return PoisonValue::get(V->getType());
1438 }
1439
1440 PoisonElts = 0;
1441
1442 if (auto *C = dyn_cast<Constant>(V)) {
1443 // Check if this is identity. If so, return 0 since we are not simplifying
1444 // anything.
1445 if (DemandedElts.isAllOnes())
1446 return nullptr;
1447
1448 Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1449 Constant *Poison = PoisonValue::get(EltTy);
1451 for (unsigned i = 0; i != VWidth; ++i) {
1452 if (!DemandedElts[i]) { // If not demanded, set to poison.
1453 Elts.push_back(Poison);
1454 PoisonElts.setBit(i);
1455 continue;
1456 }
1457
1458 Constant *Elt = C->getAggregateElement(i);
1459 if (!Elt) return nullptr;
1460
1461 Elts.push_back(Elt);
1462 if (isa<PoisonValue>(Elt)) // Already poison.
1463 PoisonElts.setBit(i);
1464 }
1465
1466 // If we changed the constant, return it.
1467 Constant *NewCV = ConstantVector::get(Elts);
1468 return NewCV != C ? NewCV : nullptr;
1469 }
1470
1471 // Limit search depth.
1472 if (Depth == 10)
1473 return nullptr;
1474
1475 if (!AllowMultipleUsers) {
1476 // If multiple users are using the root value, proceed with
1477 // simplification conservatively assuming that all elements
1478 // are needed.
1479 if (!V->hasOneUse()) {
1480 // Quit if we find multiple users of a non-root value though.
1481 // They'll be handled when it's their turn to be visited by
1482 // the main instcombine process.
1483 if (Depth != 0)
1484 // TODO: Just compute the PoisonElts information recursively.
1485 return nullptr;
1486
1487 // Conservatively assume that all elements are needed.
1488 DemandedElts = EltMask;
1489 }
1490 }
1491
1492 Instruction *I = dyn_cast<Instruction>(V);
1493 if (!I) return nullptr; // Only analyze instructions.
1494
1495 bool MadeChange = false;
1496 auto simplifyAndSetOp = [&](Instruction *Inst, unsigned OpNum,
1497 APInt Demanded, APInt &Undef) {
1498 auto *II = dyn_cast<IntrinsicInst>(Inst);
1499 Value *Op = II ? II->getArgOperand(OpNum) : Inst->getOperand(OpNum);
1500 if (Value *V = SimplifyDemandedVectorElts(Op, Demanded, Undef, Depth + 1)) {
1501 replaceOperand(*Inst, OpNum, V);
1502 MadeChange = true;
1503 }
1504 };
1505
1506 APInt PoisonElts2(VWidth, 0);
1507 APInt PoisonElts3(VWidth, 0);
1508 switch (I->getOpcode()) {
1509 default: break;
1510
1511 case Instruction::GetElementPtr: {
1512 // The LangRef requires that struct geps have all constant indices. As
1513 // such, we can't convert any operand to partial undef.
1514 auto mayIndexStructType = [](GetElementPtrInst &GEP) {
1515 for (auto I = gep_type_begin(GEP), E = gep_type_end(GEP);
1516 I != E; I++)
1517 if (I.isStruct())
1518 return true;
1519 return false;
1520 };
1521 if (mayIndexStructType(cast<GetElementPtrInst>(*I)))
1522 break;
1523
1524 // Conservatively track the demanded elements back through any vector
1525 // operands we may have. We know there must be at least one, or we
1526 // wouldn't have a vector result to get here. Note that we intentionally
1527 // merge the undef bits here since gepping with either an poison base or
1528 // index results in poison.
1529 for (unsigned i = 0; i < I->getNumOperands(); i++) {
1530 if (i == 0 ? match(I->getOperand(i), m_Undef())
1531 : match(I->getOperand(i), m_Poison())) {
1532 // If the entire vector is undefined, just return this info.
1533 PoisonElts = EltMask;
1534 return nullptr;
1535 }
1536 if (I->getOperand(i)->getType()->isVectorTy()) {
1537 APInt PoisonEltsOp(VWidth, 0);
1538 simplifyAndSetOp(I, i, DemandedElts, PoisonEltsOp);
1539 // gep(x, undef) is not undef, so skip considering idx ops here
1540 // Note that we could propagate poison, but we can't distinguish between
1541 // undef & poison bits ATM
1542 if (i == 0)
1543 PoisonElts |= PoisonEltsOp;
1544 }
1545 }
1546
1547 break;
1548 }
1549 case Instruction::InsertElement: {
1550 // If this is a variable index, we don't know which element it overwrites.
1551 // demand exactly the same input as we produce.
1552 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1553 if (!Idx) {
1554 // Note that we can't propagate undef elt info, because we don't know
1555 // which elt is getting updated.
1556 simplifyAndSetOp(I, 0, DemandedElts, PoisonElts2);
1557 break;
1558 }
1559
1560 // The element inserted overwrites whatever was there, so the input demanded
1561 // set is simpler than the output set.
1562 unsigned IdxNo = Idx->getZExtValue();
1563 APInt PreInsertDemandedElts = DemandedElts;
1564 if (IdxNo < VWidth)
1565 PreInsertDemandedElts.clearBit(IdxNo);
1566
1567 // If we only demand the element that is being inserted and that element
1568 // was extracted from the same index in another vector with the same type,
1569 // replace this insert with that other vector.
1570 // Note: This is attempted before the call to simplifyAndSetOp because that
1571 // may change PoisonElts to a value that does not match with Vec.
1572 Value *Vec;
1573 if (PreInsertDemandedElts == 0 &&
1574 match(I->getOperand(1),
1575 m_ExtractElt(m_Value(Vec), m_SpecificInt(IdxNo))) &&
1576 Vec->getType() == I->getType()) {
1577 return Vec;
1578 }
1579
1580 simplifyAndSetOp(I, 0, PreInsertDemandedElts, PoisonElts);
1581
1582 // If this is inserting an element that isn't demanded, remove this
1583 // insertelement.
1584 if (IdxNo >= VWidth || !DemandedElts[IdxNo]) {
1585 Worklist.push(I);
1586 return I->getOperand(0);
1587 }
1588
1589 // The inserted element is defined.
1590 PoisonElts.clearBit(IdxNo);
1591 break;
1592 }
1593 case Instruction::ShuffleVector: {
1594 auto *Shuffle = cast<ShuffleVectorInst>(I);
1595 assert(Shuffle->getOperand(0)->getType() ==
1596 Shuffle->getOperand(1)->getType() &&
1597 "Expected shuffle operands to have same type");
1598 unsigned OpWidth = cast<FixedVectorType>(Shuffle->getOperand(0)->getType())
1599 ->getNumElements();
1600 // Handle trivial case of a splat. Only check the first element of LHS
1601 // operand.
1602 if (all_of(Shuffle->getShuffleMask(), [](int Elt) { return Elt == 0; }) &&
1603 DemandedElts.isAllOnes()) {
1604 if (!isa<PoisonValue>(I->getOperand(1))) {
1605 I->setOperand(1, PoisonValue::get(I->getOperand(1)->getType()));
1606 MadeChange = true;
1607 }
1608 APInt LeftDemanded(OpWidth, 1);
1609 APInt LHSPoisonElts(OpWidth, 0);
1610 simplifyAndSetOp(I, 0, LeftDemanded, LHSPoisonElts);
1611 if (LHSPoisonElts[0])
1612 PoisonElts = EltMask;
1613 else
1614 PoisonElts.clearAllBits();
1615 break;
1616 }
1617
1618 APInt LeftDemanded(OpWidth, 0), RightDemanded(OpWidth, 0);
1619 for (unsigned i = 0; i < VWidth; i++) {
1620 if (DemandedElts[i]) {
1621 unsigned MaskVal = Shuffle->getMaskValue(i);
1622 if (MaskVal != -1u) {
1623 assert(MaskVal < OpWidth * 2 &&
1624 "shufflevector mask index out of range!");
1625 if (MaskVal < OpWidth)
1626 LeftDemanded.setBit(MaskVal);
1627 else
1628 RightDemanded.setBit(MaskVal - OpWidth);
1629 }
1630 }
1631 }
1632
1633 APInt LHSPoisonElts(OpWidth, 0);
1634 simplifyAndSetOp(I, 0, LeftDemanded, LHSPoisonElts);
1635
1636 APInt RHSPoisonElts(OpWidth, 0);
1637 simplifyAndSetOp(I, 1, RightDemanded, RHSPoisonElts);
1638
1639 // If this shuffle does not change the vector length and the elements
1640 // demanded by this shuffle are an identity mask, then this shuffle is
1641 // unnecessary.
1642 //
1643 // We are assuming canonical form for the mask, so the source vector is
1644 // operand 0 and operand 1 is not used.
1645 //
1646 // Note that if an element is demanded and this shuffle mask is undefined
1647 // for that element, then the shuffle is not considered an identity
1648 // operation. The shuffle prevents poison from the operand vector from
1649 // leaking to the result by replacing poison with an undefined value.
1650 if (VWidth == OpWidth) {
1651 bool IsIdentityShuffle = true;
1652 for (unsigned i = 0; i < VWidth; i++) {
1653 unsigned MaskVal = Shuffle->getMaskValue(i);
1654 if (DemandedElts[i] && i != MaskVal) {
1655 IsIdentityShuffle = false;
1656 break;
1657 }
1658 }
1659 if (IsIdentityShuffle)
1660 return Shuffle->getOperand(0);
1661 }
1662
1663 bool NewPoisonElts = false;
1664 unsigned LHSIdx = -1u, LHSValIdx = -1u;
1665 unsigned RHSIdx = -1u, RHSValIdx = -1u;
1666 bool LHSUniform = true;
1667 bool RHSUniform = true;
1668 for (unsigned i = 0; i < VWidth; i++) {
1669 unsigned MaskVal = Shuffle->getMaskValue(i);
1670 if (MaskVal == -1u) {
1671 PoisonElts.setBit(i);
1672 } else if (!DemandedElts[i]) {
1673 NewPoisonElts = true;
1674 PoisonElts.setBit(i);
1675 } else if (MaskVal < OpWidth) {
1676 if (LHSPoisonElts[MaskVal]) {
1677 NewPoisonElts = true;
1678 PoisonElts.setBit(i);
1679 } else {
1680 LHSIdx = LHSIdx == -1u ? i : OpWidth;
1681 LHSValIdx = LHSValIdx == -1u ? MaskVal : OpWidth;
1682 LHSUniform = LHSUniform && (MaskVal == i);
1683 }
1684 } else {
1685 if (RHSPoisonElts[MaskVal - OpWidth]) {
1686 NewPoisonElts = true;
1687 PoisonElts.setBit(i);
1688 } else {
1689 RHSIdx = RHSIdx == -1u ? i : OpWidth;
1690 RHSValIdx = RHSValIdx == -1u ? MaskVal - OpWidth : OpWidth;
1691 RHSUniform = RHSUniform && (MaskVal - OpWidth == i);
1692 }
1693 }
1694 }
1695
1696 // Try to transform shuffle with constant vector and single element from
1697 // this constant vector to single insertelement instruction.
1698 // shufflevector V, C, <v1, v2, .., ci, .., vm> ->
1699 // insertelement V, C[ci], ci-n
1700 if (OpWidth ==
1701 cast<FixedVectorType>(Shuffle->getType())->getNumElements()) {
1702 Value *Op = nullptr;
1703 Constant *Value = nullptr;
1704 unsigned Idx = -1u;
1705
1706 // Find constant vector with the single element in shuffle (LHS or RHS).
1707 if (LHSIdx < OpWidth && RHSUniform) {
1708 if (auto *CV = dyn_cast<ConstantVector>(Shuffle->getOperand(0))) {
1709 Op = Shuffle->getOperand(1);
1710 Value = CV->getOperand(LHSValIdx);
1711 Idx = LHSIdx;
1712 }
1713 }
1714 if (RHSIdx < OpWidth && LHSUniform) {
1715 if (auto *CV = dyn_cast<ConstantVector>(Shuffle->getOperand(1))) {
1716 Op = Shuffle->getOperand(0);
1717 Value = CV->getOperand(RHSValIdx);
1718 Idx = RHSIdx;
1719 }
1720 }
1721 // Found constant vector with single element - convert to insertelement.
1722 if (Op && Value) {
1724 Op, Value, ConstantInt::get(Type::getInt64Ty(I->getContext()), Idx),
1725 Shuffle->getName());
1726 InsertNewInstWith(New, Shuffle->getIterator());
1727 return New;
1728 }
1729 }
1730 if (NewPoisonElts) {
1731 // Add additional discovered undefs.
1733 for (unsigned i = 0; i < VWidth; ++i) {
1734 if (PoisonElts[i])
1736 else
1737 Elts.push_back(Shuffle->getMaskValue(i));
1738 }
1739 Shuffle->setShuffleMask(Elts);
1740 MadeChange = true;
1741 }
1742 break;
1743 }
1744 case Instruction::Select: {
1745 // If this is a vector select, try to transform the select condition based
1746 // on the current demanded elements.
1747 SelectInst *Sel = cast<SelectInst>(I);
1748 if (Sel->getCondition()->getType()->isVectorTy()) {
1749 // TODO: We are not doing anything with PoisonElts based on this call.
1750 // It is overwritten below based on the other select operands. If an
1751 // element of the select condition is known undef, then we are free to
1752 // choose the output value from either arm of the select. If we know that
1753 // one of those values is undef, then the output can be undef.
1754 simplifyAndSetOp(I, 0, DemandedElts, PoisonElts);
1755 }
1756
1757 // Next, see if we can transform the arms of the select.
1758 APInt DemandedLHS(DemandedElts), DemandedRHS(DemandedElts);
1759 if (auto *CV = dyn_cast<ConstantVector>(Sel->getCondition())) {
1760 for (unsigned i = 0; i < VWidth; i++) {
1761 // isNullValue() always returns false when called on a ConstantExpr.
1762 // Skip constant expressions to avoid propagating incorrect information.
1763 Constant *CElt = CV->getAggregateElement(i);
1764 if (isa<ConstantExpr>(CElt))
1765 continue;
1766 // TODO: If a select condition element is undef, we can demand from
1767 // either side. If one side is known undef, choosing that side would
1768 // propagate undef.
1769 if (CElt->isNullValue())
1770 DemandedLHS.clearBit(i);
1771 else
1772 DemandedRHS.clearBit(i);
1773 }
1774 }
1775
1776 simplifyAndSetOp(I, 1, DemandedLHS, PoisonElts2);
1777 simplifyAndSetOp(I, 2, DemandedRHS, PoisonElts3);
1778
1779 // Output elements are undefined if the element from each arm is undefined.
1780 // TODO: This can be improved. See comment in select condition handling.
1781 PoisonElts = PoisonElts2 & PoisonElts3;
1782 break;
1783 }
1784 case Instruction::BitCast: {
1785 // Vector->vector casts only.
1786 VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1787 if (!VTy) break;
1788 unsigned InVWidth = cast<FixedVectorType>(VTy)->getNumElements();
1789 APInt InputDemandedElts(InVWidth, 0);
1790 PoisonElts2 = APInt(InVWidth, 0);
1791 unsigned Ratio;
1792
1793 if (VWidth == InVWidth) {
1794 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1795 // elements as are demanded of us.
1796 Ratio = 1;
1797 InputDemandedElts = DemandedElts;
1798 } else if ((VWidth % InVWidth) == 0) {
1799 // If the number of elements in the output is a multiple of the number of
1800 // elements in the input then an input element is live if any of the
1801 // corresponding output elements are live.
1802 Ratio = VWidth / InVWidth;
1803 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1804 if (DemandedElts[OutIdx])
1805 InputDemandedElts.setBit(OutIdx / Ratio);
1806 } else if ((InVWidth % VWidth) == 0) {
1807 // If the number of elements in the input is a multiple of the number of
1808 // elements in the output then an input element is live if the
1809 // corresponding output element is live.
1810 Ratio = InVWidth / VWidth;
1811 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1812 if (DemandedElts[InIdx / Ratio])
1813 InputDemandedElts.setBit(InIdx);
1814 } else {
1815 // Unsupported so far.
1816 break;
1817 }
1818
1819 simplifyAndSetOp(I, 0, InputDemandedElts, PoisonElts2);
1820
1821 if (VWidth == InVWidth) {
1822 PoisonElts = PoisonElts2;
1823 } else if ((VWidth % InVWidth) == 0) {
1824 // If the number of elements in the output is a multiple of the number of
1825 // elements in the input then an output element is undef if the
1826 // corresponding input element is undef.
1827 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1828 if (PoisonElts2[OutIdx / Ratio])
1829 PoisonElts.setBit(OutIdx);
1830 } else if ((InVWidth % VWidth) == 0) {
1831 // If the number of elements in the input is a multiple of the number of
1832 // elements in the output then an output element is undef if all of the
1833 // corresponding input elements are undef.
1834 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1835 APInt SubUndef = PoisonElts2.lshr(OutIdx * Ratio).zextOrTrunc(Ratio);
1836 if (SubUndef.popcount() == Ratio)
1837 PoisonElts.setBit(OutIdx);
1838 }
1839 } else {
1840 llvm_unreachable("Unimp");
1841 }
1842 break;
1843 }
1844 case Instruction::FPTrunc:
1845 case Instruction::FPExt:
1846 simplifyAndSetOp(I, 0, DemandedElts, PoisonElts);
1847 break;
1848
1849 case Instruction::Call: {
1850 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1851 if (!II) break;
1852 switch (II->getIntrinsicID()) {
1853 case Intrinsic::masked_gather: // fallthrough
1854 case Intrinsic::masked_load: {
1855 // Subtlety: If we load from a pointer, the pointer must be valid
1856 // regardless of whether the element is demanded. Doing otherwise risks
1857 // segfaults which didn't exist in the original program.
1858 APInt DemandedPtrs(APInt::getAllOnes(VWidth)),
1859 DemandedPassThrough(DemandedElts);
1860 if (auto *CV = dyn_cast<ConstantVector>(II->getOperand(2)))
1861 for (unsigned i = 0; i < VWidth; i++) {
1862 Constant *CElt = CV->getAggregateElement(i);
1863 if (CElt->isNullValue())
1864 DemandedPtrs.clearBit(i);
1865 else if (CElt->isAllOnesValue())
1866 DemandedPassThrough.clearBit(i);
1867 }
1868 if (II->getIntrinsicID() == Intrinsic::masked_gather)
1869 simplifyAndSetOp(II, 0, DemandedPtrs, PoisonElts2);
1870 simplifyAndSetOp(II, 3, DemandedPassThrough, PoisonElts3);
1871
1872 // Output elements are undefined if the element from both sources are.
1873 // TODO: can strengthen via mask as well.
1874 PoisonElts = PoisonElts2 & PoisonElts3;
1875 break;
1876 }
1877 default: {
1878 // Handle target specific intrinsics
1879 std::optional<Value *> V = targetSimplifyDemandedVectorEltsIntrinsic(
1880 *II, DemandedElts, PoisonElts, PoisonElts2, PoisonElts3,
1881 simplifyAndSetOp);
1882 if (V)
1883 return *V;
1884 break;
1885 }
1886 } // switch on IntrinsicID
1887 break;
1888 } // case Call
1889 } // switch on Opcode
1890
1891 // TODO: We bail completely on integer div/rem and shifts because they have
1892 // UB/poison potential, but that should be refined.
1893 BinaryOperator *BO;
1894 if (match(I, m_BinOp(BO)) && !BO->isIntDivRem() && !BO->isShift()) {
1895 Value *X = BO->getOperand(0);
1896 Value *Y = BO->getOperand(1);
1897
1898 // Look for an equivalent binop except that one operand has been shuffled.
1899 // If the demand for this binop only includes elements that are the same as
1900 // the other binop, then we may be able to replace this binop with a use of
1901 // the earlier one.
1902 //
1903 // Example:
1904 // %other_bo = bo (shuf X, {0}), Y
1905 // %this_extracted_bo = extelt (bo X, Y), 0
1906 // -->
1907 // %other_bo = bo (shuf X, {0}), Y
1908 // %this_extracted_bo = extelt %other_bo, 0
1909 //
1910 // TODO: Handle demand of an arbitrary single element or more than one
1911 // element instead of just element 0.
1912 // TODO: Unlike general demanded elements transforms, this should be safe
1913 // for any (div/rem/shift) opcode too.
1914 if (DemandedElts == 1 && !X->hasOneUse() && !Y->hasOneUse() &&
1915 BO->hasOneUse() ) {
1916
1917 auto findShufBO = [&](bool MatchShufAsOp0) -> User * {
1918 // Try to use shuffle-of-operand in place of an operand:
1919 // bo X, Y --> bo (shuf X), Y
1920 // bo X, Y --> bo X, (shuf Y)
1921 BinaryOperator::BinaryOps Opcode = BO->getOpcode();
1922 Value *ShufOp = MatchShufAsOp0 ? X : Y;
1923 Value *OtherOp = MatchShufAsOp0 ? Y : X;
1924 for (User *U : OtherOp->users()) {
1925 ArrayRef<int> Mask;
1926 auto Shuf = m_Shuffle(m_Specific(ShufOp), m_Value(), m_Mask(Mask));
1927 if (BO->isCommutative()
1928 ? match(U, m_c_BinOp(Opcode, Shuf, m_Specific(OtherOp)))
1929 : MatchShufAsOp0
1930 ? match(U, m_BinOp(Opcode, Shuf, m_Specific(OtherOp)))
1931 : match(U, m_BinOp(Opcode, m_Specific(OtherOp), Shuf)))
1932 if (match(Mask, m_ZeroMask()) && Mask[0] != PoisonMaskElem)
1933 if (DT.dominates(U, I))
1934 return U;
1935 }
1936 return nullptr;
1937 };
1938
1939 if (User *ShufBO = findShufBO(/* MatchShufAsOp0 */ true))
1940 return ShufBO;
1941 if (User *ShufBO = findShufBO(/* MatchShufAsOp0 */ false))
1942 return ShufBO;
1943 }
1944
1945 simplifyAndSetOp(I, 0, DemandedElts, PoisonElts);
1946 simplifyAndSetOp(I, 1, DemandedElts, PoisonElts2);
1947
1948 // Output elements are undefined if both are undefined. Consider things
1949 // like undef & 0. The result is known zero, not undef.
1950 PoisonElts &= PoisonElts2;
1951 }
1952
1953 // If we've proven all of the lanes poison, return a poison value.
1954 // TODO: Intersect w/demanded lanes
1955 if (PoisonElts.isAllOnes())
1956 return PoisonValue::get(I->getType());
1957
1958 return MadeChange ? I : nullptr;
1959}
1960
1961/// For floating-point classes that resolve to a single bit pattern, return that
1962/// value.
1964 switch (Mask) {
1965 case fcPosZero:
1966 return ConstantFP::getZero(Ty);
1967 case fcNegZero:
1968 return ConstantFP::getZero(Ty, true);
1969 case fcPosInf:
1970 return ConstantFP::getInfinity(Ty);
1971 case fcNegInf:
1972 return ConstantFP::getInfinity(Ty, true);
1973 case fcNone:
1974 return PoisonValue::get(Ty);
1975 default:
1976 return nullptr;
1977 }
1978}
1979
1981 Value *V, const FPClassTest DemandedMask, KnownFPClass &Known,
1982 unsigned Depth, Instruction *CxtI) {
1983 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth");
1984 Type *VTy = V->getType();
1985
1986 assert(Known == KnownFPClass() && "expected uninitialized state");
1987
1988 if (DemandedMask == fcNone)
1989 return isa<UndefValue>(V) ? nullptr : PoisonValue::get(VTy);
1990
1992 return nullptr;
1993
1994 Instruction *I = dyn_cast<Instruction>(V);
1995 if (!I) {
1996 // Handle constants and arguments
1997 Known = computeKnownFPClass(V, fcAllFlags, CxtI, Depth + 1);
1998 Value *FoldedToConst =
1999 getFPClassConstant(VTy, DemandedMask & Known.KnownFPClasses);
2000 return FoldedToConst == V ? nullptr : FoldedToConst;
2001 }
2002
2003 if (!I->hasOneUse())
2004 return nullptr;
2005
2006 // TODO: Should account for nofpclass/FastMathFlags on current instruction
2007 switch (I->getOpcode()) {
2008 case Instruction::FNeg: {
2009 if (SimplifyDemandedFPClass(I, 0, llvm::fneg(DemandedMask), Known,
2010 Depth + 1))
2011 return I;
2012 Known.fneg();
2013 break;
2014 }
2015 case Instruction::Call: {
2016 CallInst *CI = cast<CallInst>(I);
2017 switch (CI->getIntrinsicID()) {
2018 case Intrinsic::fabs:
2019 if (SimplifyDemandedFPClass(I, 0, llvm::inverse_fabs(DemandedMask), Known,
2020 Depth + 1))
2021 return I;
2022 Known.fabs();
2023 break;
2024 case Intrinsic::arithmetic_fence:
2025 if (SimplifyDemandedFPClass(I, 0, DemandedMask, Known, Depth + 1))
2026 return I;
2027 break;
2028 case Intrinsic::copysign: {
2029 // Flip on more potentially demanded classes
2030 const FPClassTest DemandedMaskAnySign = llvm::unknown_sign(DemandedMask);
2031 if (SimplifyDemandedFPClass(I, 0, DemandedMaskAnySign, Known, Depth + 1))
2032 return I;
2033
2034 if ((DemandedMask & fcPositive) == fcNone) {
2035 // Roundabout way of replacing with fneg(fabs)
2036 I->setOperand(1, ConstantFP::get(VTy, -1.0));
2037 return I;
2038 }
2039
2040 if ((DemandedMask & fcNegative) == fcNone) {
2041 // Roundabout way of replacing with fabs
2042 I->setOperand(1, ConstantFP::getZero(VTy));
2043 return I;
2044 }
2045
2046 KnownFPClass KnownSign =
2047 computeKnownFPClass(I->getOperand(1), fcAllFlags, CxtI, Depth + 1);
2048 Known.copysign(KnownSign);
2049 break;
2050 }
2051 default:
2052 Known = computeKnownFPClass(I, ~DemandedMask, CxtI, Depth + 1);
2053 break;
2054 }
2055
2056 break;
2057 }
2058 case Instruction::Select: {
2059 KnownFPClass KnownLHS, KnownRHS;
2060 if (SimplifyDemandedFPClass(I, 2, DemandedMask, KnownRHS, Depth + 1) ||
2061 SimplifyDemandedFPClass(I, 1, DemandedMask, KnownLHS, Depth + 1))
2062 return I;
2063
2064 if (KnownLHS.isKnownNever(DemandedMask))
2065 return I->getOperand(2);
2066 if (KnownRHS.isKnownNever(DemandedMask))
2067 return I->getOperand(1);
2068
2069 // TODO: Recognize clamping patterns
2070 Known = KnownLHS | KnownRHS;
2071 break;
2072 }
2073 default:
2074 Known = computeKnownFPClass(I, ~DemandedMask, CxtI, Depth + 1);
2075 break;
2076 }
2077
2078 return getFPClassConstant(VTy, DemandedMask & Known.KnownFPClasses);
2079}
2080
2082 FPClassTest DemandedMask,
2083 KnownFPClass &Known,
2084 unsigned Depth) {
2085 Use &U = I->getOperandUse(OpNo);
2086 Value *NewVal =
2087 SimplifyDemandedUseFPClass(U.get(), DemandedMask, Known, Depth, I);
2088 if (!NewVal)
2089 return false;
2090 if (Instruction *OpInst = dyn_cast<Instruction>(U))
2091 salvageDebugInfo(*OpInst);
2092
2093 replaceUse(U, NewVal);
2094 return true;
2095}
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
Returns the sub type a function will return at a given Idx Should correspond to the result type of an ExtractValue instruction executed with just that one unsigned Idx
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
Hexagon Common GEP
This file provides internal interfaces used to implement the InstCombine.
static Constant * getFPClassConstant(Type *Ty, FPClassTest Mask)
For floating-point classes that resolve to a single bit pattern, return that value.
static cl::opt< bool > VerifyKnownBits("instcombine-verify-known-bits", cl::desc("Verify that computeKnownBits() and " "SimplifyDemandedBits() are consistent"), cl::Hidden, cl::init(false))
static unsigned getBitWidth(Type *Ty, const DataLayout &DL)
Returns the bitwidth of the given scalar or pointer type.
static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo, const APInt &Demanded)
Check to see if the specified operand of the specified instruction is a constant integer.
This file provides the interface for the instcombine pass implementation.
#define I(x, y, z)
Definition: MD5.cpp:58
static GCMetadataPrinterRegistry::Add< OcamlGCMetadataPrinter > Y("ocaml", "ocaml 3.10-compatible collector")
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
SI optimize exec mask operations pre RA
static unsigned getBitWidth(Type *Ty, const DataLayout &DL)
Returns the bitwidth of the given scalar or pointer type.
Value * RHS
Value * LHS
Class for arbitrary precision integers.
Definition: APInt.h:76
static APInt getAllOnes(unsigned numBits)
Return an APInt of a specified width with all bits set.
Definition: APInt.h:212
void clearBit(unsigned BitPosition)
Set a given bit to 0.
Definition: APInt.h:1385
static APInt getSignMask(unsigned BitWidth)
Get the SignMask for a specific bit width.
Definition: APInt.h:207
uint64_t getZExtValue() const
Get zero extended value.
Definition: APInt.h:1491
void setHighBits(unsigned hiBits)
Set the top hiBits bits.
Definition: APInt.h:1370
unsigned popcount() const
Count the number of bits set.
Definition: APInt.h:1620
APInt zextOrTrunc(unsigned width) const
Zero extend or truncate to width.
Definition: APInt.cpp:1002
unsigned getActiveBits() const
Compute the number of active bits in the value.
Definition: APInt.h:1463
APInt trunc(unsigned width) const
Truncate to new width.
Definition: APInt.cpp:906
void setBit(unsigned BitPosition)
Set the given bit to 1 whose position is given as "bitPosition".
Definition: APInt.h:1308
APInt abs() const
Get the absolute value.
Definition: APInt.h:1737
bool isAllOnes() const
Determine if all bits are set. This is true for zero-width values.
Definition: APInt.h:349
bool isZero() const
Determine if this value is zero, i.e. all bits are clear.
Definition: APInt.h:358
APInt urem(const APInt &RHS) const
Unsigned remainder operation.
Definition: APInt.cpp:1636
void setSignBit()
Set the sign bit to 1.
Definition: APInt.h:1318
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1439
bool ult(const APInt &RHS) const
Unsigned less than comparison.
Definition: APInt.h:1089
bool intersects(const APInt &RHS) const
This operation tests if there are any pairs of corresponding bits between this APInt and RHS that are...
Definition: APInt.h:1227
void clearAllBits()
Set every bit to 0.
Definition: APInt.h:1375
unsigned countr_zero() const
Count the number of trailing zero bits.
Definition: APInt.h:1589
unsigned countl_zero() const
The APInt version of std::countl_zero.
Definition: APInt.h:1548
void clearLowBits(unsigned loBits)
Set bottom loBits bits to 0.
Definition: APInt.h:1395
uint64_t getLimitedValue(uint64_t Limit=UINT64_MAX) const
If this value is smaller than the specified limit, return it, otherwise return the limit value.
Definition: APInt.h:453
APInt ashr(unsigned ShiftAmt) const
Arithmetic right-shift function.
Definition: APInt.h:805
void setAllBits()
Set every bit to 1.
Definition: APInt.h:1297
APInt shl(unsigned shiftAmt) const
Left-shift function.
Definition: APInt.h:851
bool isSubsetOf(const APInt &RHS) const
This operation checks that all bits set in this APInt are also set in RHS.
Definition: APInt.h:1235
bool isPowerOf2() const
Check if this APInt's value is a power of two greater than zero.
Definition: APInt.h:418
static APInt getLowBitsSet(unsigned numBits, unsigned loBitsSet)
Constructs an APInt value that has the bottom loBitsSet bits set.
Definition: APInt.h:284
static APInt getHighBitsSet(unsigned numBits, unsigned hiBitsSet)
Constructs an APInt value that has the top hiBitsSet bits set.
Definition: APInt.h:274
void setLowBits(unsigned loBits)
Set the bottom loBits bits.
Definition: APInt.h:1367
bool isOne() const
Determine if this is a value of 1.
Definition: APInt.h:367
void lshrInPlace(unsigned ShiftAmt)
Logical right-shift this APInt by ShiftAmt in place.
Definition: APInt.h:836
APInt lshr(unsigned shiftAmt) const
Logical right-shift function.
Definition: APInt.h:829
bool uge(const APInt &RHS) const
Unsigned greater or equal comparison.
Definition: APInt.h:1199
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory),...
Definition: ArrayRef.h:41
BinaryOps getOpcode() const
Definition: InstrTypes.h:486
Intrinsic::ID getIntrinsicID() const
Returns the intrinsic ID of the intrinsic called or Intrinsic::not_intrinsic if the called function i...
This class represents a function call, abstracting a target machine's calling convention.
static CallInst * Create(FunctionType *Ty, Value *F, const Twine &NameStr, BasicBlock::iterator InsertBefore)
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:574
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:966
static Constant * getInfinity(Type *Ty, bool Negative=false)
Definition: Constants.cpp:1083
static Constant * getZero(Type *Ty, bool Negative=false)
Definition: Constants.cpp:1037
This is the shared class of boolean and integer constants.
Definition: Constants.h:80
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:145
static Constant * get(ArrayRef< Constant * > V)
Definition: Constants.cpp:1398
This is an important base class in LLVM.
Definition: Constant.h:41
static Constant * getIntegerValue(Type *Ty, const APInt &V)
Return the value for an integer or pointer constant, or a vector thereof, with the given scalar value...
Definition: Constants.cpp:400
static Constant * getAllOnesValue(Type *Ty)
Definition: Constants.cpp:417
bool isAllOnesValue() const
Return true if this is the value that would be returned by getAllOnesValue.
Definition: Constants.cpp:107
static Constant * getNullValue(Type *Ty)
Constructor to create a '0' constant of arbitrary type.
Definition: Constants.cpp:370
Constant * getAggregateElement(unsigned Elt) const
For aggregates (struct/array/vector) return the constant that corresponds to the specified element if...
Definition: Constants.cpp:432
bool isNullValue() const
Return true if this is the value that would be returned by getNullValue.
Definition: Constants.cpp:90
This class represents an Operation in the Expression.
A parsed version of the target data layout string in and methods for querying it.
Definition: DataLayout.h:110
IntegerType * getIndexType(LLVMContext &C, unsigned AddressSpace) const
Returns the type of a GEP index in AddressSpace.
Definition: DataLayout.cpp:905
bool dominates(const BasicBlock *BB, const Use &U) const
Return true if the (end of the) basic block BB dominates the use U.
Definition: Dominators.cpp:122
an instruction for type-safe pointer arithmetic to access elements of arrays and structs
Definition: Instructions.h:973
CallInst * CreateUnaryIntrinsic(Intrinsic::ID ID, Value *V, Instruction *FMFSource=nullptr, const Twine &Name="")
Create a call to intrinsic ID with 1 operand which is mangled on its type.
Definition: IRBuilder.cpp:913
Value * CreateTrunc(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2007
Value * CreateSExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2023
Value * CreateLShr(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Definition: IRBuilder.h:1437
Value * CreateNot(Value *V, const Twine &Name="")
Definition: IRBuilder.h:1749
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1475
Value * CreateOr(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1497
void SetInsertPoint(BasicBlock *TheBB)
This specifies that created instructions should be appended to the end of the specified block.
Definition: IRBuilder.h:180
Value * CreateXor(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1519
Value * CreateGEP(Type *Ty, Value *Ptr, ArrayRef< Value * > IdxList, const Twine &Name="", bool IsInBounds=false)
Definition: IRBuilder.h:1866
static InsertElementInst * Create(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr, BasicBlock::iterator InsertBefore)
KnownFPClass computeKnownFPClass(Value *Val, FastMathFlags FMF, FPClassTest Interested=fcAllFlags, const Instruction *CtxI=nullptr, unsigned Depth=0) const
bool SimplifyDemandedBits(Instruction *I, unsigned Op, const APInt &DemandedMask, KnownBits &Known, unsigned Depth=0) override
This form of SimplifyDemandedBits simplifies the specified instruction operand if possible,...
Value * SimplifyDemandedVectorElts(Value *V, APInt DemandedElts, APInt &PoisonElts, unsigned Depth=0, bool AllowMultipleUsers=false) override
The specified value produces a vector with any number of elements.
Value * SimplifyDemandedUseBits(Value *V, APInt DemandedMask, KnownBits &Known, unsigned Depth, Instruction *CxtI)
Attempts to replace V with a simpler value based on the demanded bits.
std::optional< std::pair< Intrinsic::ID, SmallVector< Value *, 3 > > > convertOrOfShiftsToFunnelShift(Instruction &Or)
Value * SimplifyMultipleUseDemandedBits(Instruction *I, const APInt &DemandedMask, KnownBits &Known, unsigned Depth, Instruction *CxtI)
Helper routine of SimplifyDemandedUseBits.
Value * simplifyShrShlDemandedBits(Instruction *Shr, const APInt &ShrOp1, Instruction *Shl, const APInt &ShlOp1, const APInt &DemandedMask, KnownBits &Known)
Helper routine of SimplifyDemandedUseBits.
Value * SimplifyDemandedUseFPClass(Value *V, FPClassTest DemandedMask, KnownFPClass &Known, unsigned Depth, Instruction *CxtI)
Attempts to replace V with a simpler value based on the demanded floating-point classes.
bool SimplifyDemandedFPClass(Instruction *I, unsigned Op, FPClassTest DemandedMask, KnownFPClass &Known, unsigned Depth=0)
bool SimplifyDemandedInstructionBits(Instruction &Inst)
Tries to simplify operands to an integer instruction based on its demanded bits.
SimplifyQuery SQ
Definition: InstCombiner.h:76
Instruction * replaceInstUsesWith(Instruction &I, Value *V)
A combiner-aware RAUW-like routine.
Definition: InstCombiner.h:385
void replaceUse(Use &U, Value *NewValue)
Replace use and add the previously used value to the worklist.
Definition: InstCombiner.h:417
InstructionWorklist & Worklist
A worklist of the instructions that need to be simplified.
Definition: InstCombiner.h:64
Instruction * InsertNewInstWith(Instruction *New, BasicBlock::iterator Old)
Same as InsertNewInstBefore, but also sets the debug loc.
Definition: InstCombiner.h:374
const DataLayout & DL
Definition: InstCombiner.h:75
unsigned ComputeNumSignBits(const Value *Op, unsigned Depth=0, const Instruction *CxtI=nullptr) const
Definition: InstCombiner.h:451
std::optional< Value * > targetSimplifyDemandedVectorEltsIntrinsic(IntrinsicInst &II, APInt DemandedElts, APInt &UndefElts, APInt &UndefElts2, APInt &UndefElts3, std::function< void(Instruction *, unsigned, APInt, APInt &)> SimplifyAndSetOp)
Instruction * replaceOperand(Instruction &I, unsigned OpNum, Value *V)
Replace operand of instruction and add old operand to the worklist.
Definition: InstCombiner.h:409
DominatorTree & DT
Definition: InstCombiner.h:74
std::optional< Value * > targetSimplifyDemandedUseBitsIntrinsic(IntrinsicInst &II, APInt DemandedMask, KnownBits &Known, bool &KnownBitsComputed)
void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth, const Instruction *CxtI) const
Definition: InstCombiner.h:430
BuilderTy & Builder
Definition: InstCombiner.h:60
void push(Instruction *I)
Push the instruction onto the worklist stack.
bool hasNoUnsignedWrap() const LLVM_READONLY
Determine whether the no unsigned wrap flag is set.
bool hasNoSignedWrap() const LLVM_READONLY
Determine whether the no signed wrap flag is set.
bool isCommutative() const LLVM_READONLY
Return true if the instruction is commutative:
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:252
void setIsExact(bool b=true)
Set or clear the exact flag on this instruction, which must be an operator which supports this flag.
bool isShift() const
Definition: Instruction.h:259
bool isIntDivRem() const
Definition: Instruction.h:258
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:47
Intrinsic::ID getIntrinsicID() const
Return the intrinsic ID of this intrinsic.
Definition: IntrinsicInst.h:54
bool hasNoSignedWrap() const
Test whether this operation is known to never undergo signed overflow, aka the nsw property.
Definition: Operator.h:108
bool hasNoUnsignedWrap() const
Test whether this operation is known to never undergo unsigned overflow, aka the nuw property.
Definition: Operator.h:102
static PoisonValue * get(Type *T)
Static factory methods - Return an 'poison' object of the specified type.
Definition: Constants.cpp:1827
This class represents the LLVM 'select' instruction.
const Value * getCondition() const
void push_back(const T &Elt)
Definition: SmallVector.h:426
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
Definition: SmallVector.h:1209
The instances of the Type class are immutable: once they are created, they are never changed.
Definition: Type.h:45
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:265
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:234
unsigned getScalarSizeInBits() const LLVM_READONLY
If this is a vector type, return the getPrimitiveSizeInBits value for the element type.
static IntegerType * getInt64Ty(LLVMContext &C)
static UndefValue * get(Type *T)
Static factory methods - Return an 'undef' object of the specified type.
Definition: Constants.cpp:1808
A Use represents the edge between a Value definition and its users.
Definition: Use.h:43
Value * getOperand(unsigned i) const
Definition: User.h:169
LLVM Value Representation.
Definition: Value.h:74
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:255
bool hasOneUse() const
Return true if there is exactly one use of this value.
Definition: Value.h:434
iterator_range< user_iterator > users()
Definition: Value.h:421
StringRef getName() const
Return a constant reference to the value's name.
Definition: Value.cpp:309
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:383
Base class of all SIMD vector types.
Definition: DerivedTypes.h:403
This class represents zero extension of integer types.
self_iterator getIterator()
Definition: ilist_node.h:109
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
@ C
The default llvm calling convention, compatible with C.
Definition: CallingConv.h:34
Function * getDeclaration(Module *M, ID id, ArrayRef< Type * > Tys=std::nullopt)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
Definition: Function.cpp:1469
class_match< PoisonValue > m_Poison()
Match an arbitrary poison constant.
Definition: PatternMatch.h:155
PtrAdd_match< PointerOpTy, OffsetOpTy > m_PtrAdd(const PointerOpTy &PointerOp, const OffsetOpTy &OffsetOp)
Matches GEP with i8 source element type.
BinaryOp_match< LHS, RHS, Instruction::Add > m_Add(const LHS &L, const RHS &R)
class_match< BinaryOperator > m_BinOp()
Match an arbitrary binary operation and ignore it.
Definition: PatternMatch.h:100
BinaryOp_match< LHS, RHS, Instruction::AShr > m_AShr(const LHS &L, const RHS &R)
specific_intval< false > m_SpecificInt(const APInt &V)
Match a specific integer value or vector with all elements equal to the value.
Definition: PatternMatch.h:918
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
specificval_ty m_Specific(const Value *V)
Match if we have a specific specified value.
Definition: PatternMatch.h:821
BinOpPred_match< LHS, RHS, is_right_shift_op > m_Shr(const LHS &L, const RHS &R)
Matches logical shift operations.
TwoOps_match< Val_t, Idx_t, Instruction::ExtractElement > m_ExtractElt(const Val_t &Val, const Idx_t &Idx)
Matches ExtractElementInst.
class_match< ConstantInt > m_ConstantInt()
Match an arbitrary ConstantInt and ignore it.
Definition: PatternMatch.h:163
CmpClass_match< LHS, RHS, ICmpInst, ICmpInst::Predicate > m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R)
OneUse_match< T > m_OneUse(const T &SubPattern)
Definition: PatternMatch.h:67
TwoOps_match< V1_t, V2_t, Instruction::ShuffleVector > m_Shuffle(const V1_t &v1, const V2_t &v2)
Matches ShuffleVectorInst independently of mask value.
match_combine_and< class_match< Constant >, match_unless< constantexpr_match > > m_ImmConstant()
Match an arbitrary immediate Constant and ignore it.
Definition: PatternMatch.h:800
CastInst_match< OpTy, ZExtInst > m_ZExt(const OpTy &Op)
Matches ZExt.
BinaryOp_match< LHS, RHS, Instruction::Add, true > m_c_Add(const LHS &L, const RHS &R)
Matches a Add with LHS and RHS in either order.
apint_match m_APInt(const APInt *&Res)
Match a ConstantInt or splatted ConstantVector, binding the specified pointer to the contained APInt.
Definition: PatternMatch.h:294
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:92
AnyBinaryOp_match< LHS, RHS, true > m_c_BinOp(const LHS &L, const RHS &R)
Matches a BinaryOperator with LHS and RHS in either order.
BinaryOp_match< LHS, RHS, Instruction::LShr > m_LShr(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::Shl > m_Shl(const LHS &L, const RHS &R)
auto m_Undef()
Match an arbitrary undef constant.
Definition: PatternMatch.h:152
BinaryOp_match< cst_pred_ty< is_all_ones >, ValTy, Instruction::Xor, true > m_Not(const ValTy &V)
Matches a 'Not' as 'xor V, -1' or 'xor -1, V'.
CastInst_match< OpTy, SExtInst > m_SExt(const OpTy &Op)
Matches SExt.
is_zero m_Zero()
Match any null constant or a vector with all elements equal to 0.
Definition: PatternMatch.h:561
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:450
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
bool haveNoCommonBitsSet(const WithCache< const Value * > &LHSCache, const WithCache< const Value * > &RHSCache, const SimplifyQuery &SQ)
Return true if LHS and RHS have no common bits set.
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly.
Definition: STLExtras.h:1722
int countr_one(T Value)
Count the number of ones from the least significant bit to the first zero bit.
Definition: bit.h:307
void salvageDebugInfo(const MachineRegisterInfo &MRI, MachineInstr &MI)
Assuming the instruction MI is going to be deleted, attempt to salvage debug users of MI by writing t...
Definition: Utils.cpp:1650
gep_type_iterator gep_type_end(const User *GEP)
void computeKnownBitsFromContext(const Value *V, KnownBits &Known, unsigned Depth, const SimplifyQuery &Q)
Merge bits known from context-dependent facts into Known.
KnownBits analyzeKnownBitsFromAndXorOr(const Operator *I, const KnownBits &KnownLHS, const KnownBits &KnownRHS, unsigned Depth, const SimplifyQuery &SQ)
Using KnownBits LHS/RHS produce the known bits for logic op (and/xor/or).
constexpr unsigned MaxAnalysisRecursionDepth
Definition: ValueTracking.h:47
FPClassTest fneg(FPClassTest Mask)
Return the test mask which returns true if the value's sign bit is flipped.
FPClassTest
Floating-point class tests, supported by 'is_fpclass' intrinsic.
FPClassTest inverse_fabs(FPClassTest Mask)
Return the test mask which returns true after fabs is applied to the value.
Constant * ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS, Constant *RHS, const DataLayout &DL)
Attempt to constant fold a binary operation with the specified operands.
constexpr int PoisonMaskElem
raw_fd_ostream & errs()
This returns a reference to a raw_ostream for standard error.
@ Or
Bitwise or logical OR of integers.
@ Xor
Bitwise or logical XOR of integers.
@ And
Bitwise or logical AND of integers.
FPClassTest unknown_sign(FPClassTest Mask)
Return the test mask which returns true if the value could have the same set of classes,...
constexpr unsigned BitWidth
Definition: BitmaskEnum.h:191
gep_type_iterator gep_type_begin(const User *GEP)
unsigned Log2(Align A)
Returns the log2 of the alignment.
Definition: Alignment.h:208
uint64_t alignDown(uint64_t Value, uint64_t Align, uint64_t Skew=0)
Returns the largest uint64_t less than or equal to Value and is Skew mod Align.
Definition: MathExtras.h:428
This struct is a compact representation of a valid (non-zero power of two) alignment.
Definition: Alignment.h:39
static KnownBits makeConstant(const APInt &C)
Create known bits from a known constant.
Definition: KnownBits.h:297
KnownBits anyextOrTrunc(unsigned BitWidth) const
Return known bits for an "any" extension or truncation of the value we're tracking.
Definition: KnownBits.h:182
bool isNonNegative() const
Returns true if this value is known to be non-negative.
Definition: KnownBits.h:104
void makeNonNegative()
Make this value non-negative.
Definition: KnownBits.h:120
static KnownBits urem(const KnownBits &LHS, const KnownBits &RHS)
Compute known bits for urem(LHS, RHS).
Definition: KnownBits.cpp:1018
bool hasConflict() const
Returns true if there is conflicting information.
Definition: KnownBits.h:47
unsigned getBitWidth() const
Get the bit width of this value.
Definition: KnownBits.h:40
void resetAll()
Resets the known state of all bits.
Definition: KnownBits.h:71
KnownBits intersectWith(const KnownBits &RHS) const
Returns KnownBits information that is known to be true for both this and RHS.
Definition: KnownBits.h:307
KnownBits sext(unsigned BitWidth) const
Return known bits for a sign extension of the value we're tracking.
Definition: KnownBits.h:176
KnownBits zextOrTrunc(unsigned BitWidth) const
Return known bits for a zero extension or truncation of the value we're tracking.
Definition: KnownBits.h:192
static KnownBits udiv(const KnownBits &LHS, const KnownBits &RHS, bool Exact=false)
Compute known bits for udiv(LHS, RHS).
Definition: KnownBits.cpp:976
static KnownBits computeForAddSub(bool Add, bool NSW, bool NUW, const KnownBits &LHS, const KnownBits &RHS)
Compute known bits resulting from adding LHS and RHS.
Definition: KnownBits.cpp:57
bool isNegative() const
Returns true if this value is known to be negative.
Definition: KnownBits.h:101
static KnownBits shl(const KnownBits &LHS, const KnownBits &RHS, bool NUW=false, bool NSW=false, bool ShAmtNonZero=false)
Compute known bits for shl(LHS, RHS).
Definition: KnownBits.cpp:279
FPClassTest KnownFPClasses
Floating-point classes the value could be one of.
void copysign(const KnownFPClass &Sign)
bool isKnownNever(FPClassTest Mask) const
Return true if it's known this can never be one of the mask entries.
SimplifyQuery getWithInstruction(const Instruction *I) const
Definition: SimplifyQuery.h:96