1 //===- InstCombineSimplifyDemanded.cpp ------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains logic for simplifying instructions based on information
11 // about how they are used.
13 //===----------------------------------------------------------------------===//
15 #include "InstCombineInternal.h"
16 #include "llvm/Analysis/ValueTracking.h"
17 #include "llvm/IR/IntrinsicInst.h"
18 #include "llvm/IR/PatternMatch.h"
19 #include "llvm/Support/KnownBits.h"
22 using namespace llvm::PatternMatch;
24 #define DEBUG_TYPE "instcombine"
26 /// Check to see if the specified operand of the specified instruction is a
27 /// constant integer. If so, check to see if there are any bits set in the
28 /// constant that are not demanded. If so, shrink the constant and return true.
29 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
30 const APInt &Demanded) {
31 assert(I && "No instruction?");
32 assert(OpNo < I->getNumOperands() && "Operand index too large");
34 // The operand must be a constant integer or splat integer.
35 Value *Op = I->getOperand(OpNo);
37 if (!match(Op, m_APInt(C)))
40 // If there are no bits set that aren't demanded, nothing to do.
41 if (C->isSubsetOf(Demanded))
44 // This instruction is producing bits that are not demanded. Shrink the RHS.
45 I->setOperand(OpNo, ConstantInt::get(Op->getType(), *C & Demanded));
52 /// Inst is an integer instruction that SimplifyDemandedBits knows about. See if
53 /// the instruction has any properties that allow us to simplify its operands.
54 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
55 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
56 KnownBits Known(BitWidth);
57 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
59 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask, Known,
62 if (V == &Inst) return true;
63 replaceInstUsesWith(Inst, V);
67 /// This form of SimplifyDemandedBits simplifies the specified instruction
68 /// operand if possible, updating it in place. It returns true if it made any
69 /// change and false otherwise.
70 bool InstCombiner::SimplifyDemandedBits(Instruction *I, unsigned OpNo,
71 const APInt &DemandedMask,
74 Use &U = I->getOperandUse(OpNo);
75 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask, Known,
77 if (!NewVal) return false;
83 /// This function attempts to replace V with a simpler value based on the
84 /// demanded bits. When this function is called, it is known that only the bits
85 /// set in DemandedMask of the result of V are ever used downstream.
86 /// Consequently, depending on the mask and V, it may be possible to replace V
87 /// with a constant or one of its operands. In such cases, this function does
88 /// the replacement and returns true. In all other cases, it returns false after
89 /// analyzing the expression and setting KnownOne and known to be one in the
90 /// expression. Known.Zero contains all the bits that are known to be zero in
91 /// the expression. These are provided to potentially allow the caller (which
92 /// might recursively be SimplifyDemandedBits itself) to simplify the
94 /// Known.One and Known.Zero always follow the invariant that:
95 /// Known.One & Known.Zero == 0.
96 /// That is, a bit can't be both 1 and 0. Note that the bits in Known.One and
97 /// Known.Zero may only be accurate for those bits set in DemandedMask. Note
98 /// also that the bitwidth of V, DemandedMask, Known.Zero and Known.One must all
101 /// This returns null if it did not change anything and it permits no
102 /// simplification. This returns V itself if it did some simplification of V's
103 /// operands based on the information about what bits are demanded. This returns
104 /// some other non-null value if it found out that V is equal to another value
105 /// in the context where the specified bits are demanded, but not for all users.
106 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
107 KnownBits &Known, unsigned Depth,
109 assert(V != nullptr && "Null pointer of Value???");
110 assert(Depth <= 6 && "Limit Search Depth");
111 uint32_t BitWidth = DemandedMask.getBitWidth();
112 Type *VTy = V->getType();
114 (!VTy->isIntOrIntVectorTy() || VTy->getScalarSizeInBits() == BitWidth) &&
115 Known.getBitWidth() == BitWidth &&
116 "Value *V, DemandedMask and Known must have same BitWidth");
118 if (isa<Constant>(V)) {
119 computeKnownBits(V, Known, Depth, CxtI);
124 if (DemandedMask.isNullValue()) // Not demanding any bits from V.
125 return UndefValue::get(VTy);
127 if (Depth == 6) // Limit search depth.
130 Instruction *I = dyn_cast<Instruction>(V);
132 computeKnownBits(V, Known, Depth, CxtI);
133 return nullptr; // Only analyze instructions.
136 // If there are multiple uses of this value and we aren't at the root, then
137 // we can't do any simplifications of the operands, because DemandedMask
138 // only reflects the bits demanded by *one* of the users.
139 if (Depth != 0 && !I->hasOneUse())
140 return SimplifyMultipleUseDemandedBits(I, DemandedMask, Known, Depth, CxtI);
142 KnownBits LHSKnown(BitWidth), RHSKnown(BitWidth);
144 // If this is the root being simplified, allow it to have multiple uses,
145 // just set the DemandedMask to all bits so that we can try to simplify the
146 // operands. This allows visitTruncInst (for example) to simplify the
147 // operand of a trunc without duplicating all the logic below.
148 if (Depth == 0 && !V->hasOneUse())
149 DemandedMask.setAllBits();
151 switch (I->getOpcode()) {
153 computeKnownBits(I, Known, Depth, CxtI);
155 case Instruction::And: {
156 // If either the LHS or the RHS are Zero, the result is zero.
157 if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) ||
158 SimplifyDemandedBits(I, 0, DemandedMask & ~RHSKnown.Zero, LHSKnown,
161 assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
162 assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
164 // Output known-0 are known to be clear if zero in either the LHS | RHS.
165 APInt IKnownZero = RHSKnown.Zero | LHSKnown.Zero;
166 // Output known-1 bits are only known if set in both the LHS & RHS.
167 APInt IKnownOne = RHSKnown.One & LHSKnown.One;
169 // If the client is only demanding bits that we know, return the known
171 if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
172 return Constant::getIntegerValue(VTy, IKnownOne);
174 // If all of the demanded bits are known 1 on one side, return the other.
175 // These bits cannot contribute to the result of the 'and'.
176 if (DemandedMask.isSubsetOf(LHSKnown.Zero | RHSKnown.One))
177 return I->getOperand(0);
178 if (DemandedMask.isSubsetOf(RHSKnown.Zero | LHSKnown.One))
179 return I->getOperand(1);
181 // If the RHS is a constant, see if we can simplify it.
182 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnown.Zero))
185 Known.Zero = std::move(IKnownZero);
186 Known.One = std::move(IKnownOne);
189 case Instruction::Or: {
190 // If either the LHS or the RHS are One, the result is One.
191 if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) ||
192 SimplifyDemandedBits(I, 0, DemandedMask & ~RHSKnown.One, LHSKnown,
195 assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
196 assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
198 // Output known-0 bits are only known if clear in both the LHS & RHS.
199 APInt IKnownZero = RHSKnown.Zero & LHSKnown.Zero;
200 // Output known-1 are known. to be set if s.et in either the LHS | RHS.
201 APInt IKnownOne = RHSKnown.One | LHSKnown.One;
203 // If the client is only demanding bits that we know, return the known
205 if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
206 return Constant::getIntegerValue(VTy, IKnownOne);
208 // If all of the demanded bits are known zero on one side, return the other.
209 // These bits cannot contribute to the result of the 'or'.
210 if (DemandedMask.isSubsetOf(LHSKnown.One | RHSKnown.Zero))
211 return I->getOperand(0);
212 if (DemandedMask.isSubsetOf(RHSKnown.One | LHSKnown.Zero))
213 return I->getOperand(1);
215 // If the RHS is a constant, see if we can simplify it.
216 if (ShrinkDemandedConstant(I, 1, DemandedMask))
219 Known.Zero = std::move(IKnownZero);
220 Known.One = std::move(IKnownOne);
223 case Instruction::Xor: {
224 if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) ||
225 SimplifyDemandedBits(I, 0, DemandedMask, LHSKnown, Depth + 1))
227 assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
228 assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
230 // Output known-0 bits are known if clear or set in both the LHS & RHS.
231 APInt IKnownZero = (RHSKnown.Zero & LHSKnown.Zero) |
232 (RHSKnown.One & LHSKnown.One);
233 // Output known-1 are known to be set if set in only one of the LHS, RHS.
234 APInt IKnownOne = (RHSKnown.Zero & LHSKnown.One) |
235 (RHSKnown.One & LHSKnown.Zero);
237 // If the client is only demanding bits that we know, return the known
239 if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
240 return Constant::getIntegerValue(VTy, IKnownOne);
242 // If all of the demanded bits are known zero on one side, return the other.
243 // These bits cannot contribute to the result of the 'xor'.
244 if (DemandedMask.isSubsetOf(RHSKnown.Zero))
245 return I->getOperand(0);
246 if (DemandedMask.isSubsetOf(LHSKnown.Zero))
247 return I->getOperand(1);
249 // If all of the demanded bits are known to be zero on one side or the
250 // other, turn this into an *inclusive* or.
251 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
252 if (DemandedMask.isSubsetOf(RHSKnown.Zero | LHSKnown.Zero)) {
254 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
256 return InsertNewInstWith(Or, *I);
259 // If all of the demanded bits on one side are known, and all of the set
260 // bits on that side are also known to be set on the other side, turn this
261 // into an AND, as we know the bits will be cleared.
262 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
263 if (DemandedMask.isSubsetOf(RHSKnown.Zero|RHSKnown.One) &&
264 RHSKnown.One.isSubsetOf(LHSKnown.One)) {
265 Constant *AndC = Constant::getIntegerValue(VTy,
266 ~RHSKnown.One & DemandedMask);
267 Instruction *And = BinaryOperator::CreateAnd(I->getOperand(0), AndC);
268 return InsertNewInstWith(And, *I);
271 // If the RHS is a constant, see if we can simplify it.
272 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
273 if (ShrinkDemandedConstant(I, 1, DemandedMask))
276 // If our LHS is an 'and' and if it has one use, and if any of the bits we
277 // are flipping are known to be set, then the xor is just resetting those
278 // bits to zero. We can just knock out bits from the 'and' and the 'xor',
279 // simplifying both of them.
280 if (Instruction *LHSInst = dyn_cast<Instruction>(I->getOperand(0)))
281 if (LHSInst->getOpcode() == Instruction::And && LHSInst->hasOneUse() &&
282 isa<ConstantInt>(I->getOperand(1)) &&
283 isa<ConstantInt>(LHSInst->getOperand(1)) &&
284 (LHSKnown.One & RHSKnown.One & DemandedMask) != 0) {
285 ConstantInt *AndRHS = cast<ConstantInt>(LHSInst->getOperand(1));
286 ConstantInt *XorRHS = cast<ConstantInt>(I->getOperand(1));
287 APInt NewMask = ~(LHSKnown.One & RHSKnown.One & DemandedMask);
290 ConstantInt::get(I->getType(), NewMask & AndRHS->getValue());
291 Instruction *NewAnd = BinaryOperator::CreateAnd(I->getOperand(0), AndC);
292 InsertNewInstWith(NewAnd, *I);
295 ConstantInt::get(I->getType(), NewMask & XorRHS->getValue());
296 Instruction *NewXor = BinaryOperator::CreateXor(NewAnd, XorC);
297 return InsertNewInstWith(NewXor, *I);
300 // Output known-0 bits are known if clear or set in both the LHS & RHS.
301 Known.Zero = std::move(IKnownZero);
302 // Output known-1 are known to be set if set in only one of the LHS, RHS.
303 Known.One = std::move(IKnownOne);
306 case Instruction::Select:
307 // If this is a select as part of a min/max pattern, don't simplify any
308 // further in case we break the structure.
310 if (matchSelectPattern(I, LHS, RHS).Flavor != SPF_UNKNOWN)
313 if (SimplifyDemandedBits(I, 2, DemandedMask, RHSKnown, Depth + 1) ||
314 SimplifyDemandedBits(I, 1, DemandedMask, LHSKnown, Depth + 1))
316 assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
317 assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
319 // If the operands are constants, see if we can simplify them.
320 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
321 ShrinkDemandedConstant(I, 2, DemandedMask))
324 // Only known if known in both the LHS and RHS.
325 Known.One = RHSKnown.One & LHSKnown.One;
326 Known.Zero = RHSKnown.Zero & LHSKnown.Zero;
328 case Instruction::ZExt:
329 case Instruction::Trunc: {
330 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
332 APInt InputDemandedMask = DemandedMask.zextOrTrunc(SrcBitWidth);
333 KnownBits InputKnown(SrcBitWidth);
334 if (SimplifyDemandedBits(I, 0, InputDemandedMask, InputKnown, Depth + 1))
336 Known = Known.zextOrTrunc(BitWidth);
337 // Any top bits are known to be zero.
338 if (BitWidth > SrcBitWidth)
339 Known.Zero.setBitsFrom(SrcBitWidth);
340 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
343 case Instruction::BitCast:
344 if (!I->getOperand(0)->getType()->isIntOrIntVectorTy())
345 return nullptr; // vector->int or fp->int?
347 if (VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
348 if (VectorType *SrcVTy =
349 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
350 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
351 // Don't touch a bitcast between vectors of different element counts.
354 // Don't touch a scalar-to-vector bitcast.
356 } else if (I->getOperand(0)->getType()->isVectorTy())
357 // Don't touch a vector-to-scalar bitcast.
360 if (SimplifyDemandedBits(I, 0, DemandedMask, Known, Depth + 1))
362 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
364 case Instruction::SExt: {
365 // Compute the bits in the result that are not present in the input.
366 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
368 APInt InputDemandedBits = DemandedMask.trunc(SrcBitWidth);
370 // If any of the sign extended bits are demanded, we know that the sign
372 if (DemandedMask.getActiveBits() > SrcBitWidth)
373 InputDemandedBits.setBit(SrcBitWidth-1);
375 KnownBits InputKnown(SrcBitWidth);
376 if (SimplifyDemandedBits(I, 0, InputDemandedBits, InputKnown, Depth + 1))
379 // If the input sign bit is known zero, or if the NewBits are not demanded
380 // convert this into a zero extension.
381 if (InputKnown.isNonNegative() ||
382 DemandedMask.getActiveBits() <= SrcBitWidth) {
383 // Convert to ZExt cast.
384 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
385 return InsertNewInstWith(NewCast, *I);
388 // If the sign bit of the input is known set or clear, then we know the
389 // top bits of the result.
390 Known = InputKnown.sext(BitWidth);
391 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
394 case Instruction::Add:
395 case Instruction::Sub: {
396 /// If the high-bits of an ADD/SUB are not demanded, then we do not care
397 /// about the high bits of the operands.
398 unsigned NLZ = DemandedMask.countLeadingZeros();
399 // Right fill the mask of bits for this ADD/SUB to demand the most
400 // significant bit and all those below it.
401 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
402 if (ShrinkDemandedConstant(I, 0, DemandedFromOps) ||
403 SimplifyDemandedBits(I, 0, DemandedFromOps, LHSKnown, Depth + 1) ||
404 ShrinkDemandedConstant(I, 1, DemandedFromOps) ||
405 SimplifyDemandedBits(I, 1, DemandedFromOps, RHSKnown, Depth + 1)) {
407 // Disable the nsw and nuw flags here: We can no longer guarantee that
408 // we won't wrap after simplification. Removing the nsw/nuw flags is
409 // legal here because the top bit is not demanded.
410 BinaryOperator &BinOP = *cast<BinaryOperator>(I);
411 BinOP.setHasNoSignedWrap(false);
412 BinOP.setHasNoUnsignedWrap(false);
417 // If we are known to be adding/subtracting zeros to every bit below
418 // the highest demanded bit, we just return the other side.
419 if (DemandedFromOps.isSubsetOf(RHSKnown.Zero))
420 return I->getOperand(0);
421 // We can't do this with the LHS for subtraction, unless we are only
422 // demanding the LSB.
423 if ((I->getOpcode() == Instruction::Add ||
424 DemandedFromOps.isOneValue()) &&
425 DemandedFromOps.isSubsetOf(LHSKnown.Zero))
426 return I->getOperand(1);
428 // Otherwise just compute the known bits of the result.
429 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
430 Known = KnownBits::computeForAddSub(I->getOpcode() == Instruction::Add,
431 NSW, LHSKnown, RHSKnown);
434 case Instruction::Shl: {
436 if (match(I->getOperand(1), m_APInt(SA))) {
438 if (match(I->getOperand(0), m_Shr(m_Value(), m_APInt(ShrAmt))))
439 if (Instruction *Shr = dyn_cast<Instruction>(I->getOperand(0)))
440 if (Value *R = simplifyShrShlDemandedBits(Shr, *ShrAmt, I, *SA,
441 DemandedMask, Known))
444 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
445 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
447 // If the shift is NUW/NSW, then it does demand the high bits.
448 ShlOperator *IOp = cast<ShlOperator>(I);
449 if (IOp->hasNoSignedWrap())
450 DemandedMaskIn.setHighBits(ShiftAmt+1);
451 else if (IOp->hasNoUnsignedWrap())
452 DemandedMaskIn.setHighBits(ShiftAmt);
454 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1))
456 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
457 Known.Zero <<= ShiftAmt;
458 Known.One <<= ShiftAmt;
459 // low bits known zero.
461 Known.Zero.setLowBits(ShiftAmt);
465 case Instruction::LShr: {
467 if (match(I->getOperand(1), m_APInt(SA))) {
468 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
470 // Unsigned shift right.
471 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
473 // If the shift is exact, then it does demand the low bits (and knows that
475 if (cast<LShrOperator>(I)->isExact())
476 DemandedMaskIn.setLowBits(ShiftAmt);
478 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1))
480 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
481 Known.Zero.lshrInPlace(ShiftAmt);
482 Known.One.lshrInPlace(ShiftAmt);
484 Known.Zero.setHighBits(ShiftAmt); // high bits known zero.
488 case Instruction::AShr: {
489 // If this is an arithmetic shift right and only the low-bit is set, we can
490 // always convert this into a logical shr, even if the shift amount is
491 // variable. The low bit of the shift cannot be an input sign bit unless
492 // the shift amount is >= the size of the datatype, which is undefined.
493 if (DemandedMask.isOneValue()) {
494 // Perform the logical shift right.
495 Instruction *NewVal = BinaryOperator::CreateLShr(
496 I->getOperand(0), I->getOperand(1), I->getName());
497 return InsertNewInstWith(NewVal, *I);
500 // If the sign bit is the only bit demanded by this ashr, then there is no
501 // need to do it, the shift doesn't change the high bit.
502 if (DemandedMask.isSignMask())
503 return I->getOperand(0);
506 if (match(I->getOperand(1), m_APInt(SA))) {
507 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
509 // Signed shift right.
510 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
511 // If any of the high bits are demanded, we should set the sign bit as
513 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
514 DemandedMaskIn.setSignBit();
516 // If the shift is exact, then it does demand the low bits (and knows that
518 if (cast<AShrOperator>(I)->isExact())
519 DemandedMaskIn.setLowBits(ShiftAmt);
521 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1))
524 unsigned SignBits = ComputeNumSignBits(I->getOperand(0), Depth + 1, CxtI);
526 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
527 // Compute the new bits that are at the top now plus sign bits.
528 APInt HighBits(APInt::getHighBitsSet(
529 BitWidth, std::min(SignBits + ShiftAmt - 1, BitWidth)));
530 Known.Zero.lshrInPlace(ShiftAmt);
531 Known.One.lshrInPlace(ShiftAmt);
533 // If the input sign bit is known to be zero, or if none of the top bits
534 // are demanded, turn this into an unsigned shift right.
535 assert(BitWidth > ShiftAmt && "Shift amount not saturated?");
536 if (Known.Zero[BitWidth-ShiftAmt-1] ||
537 !DemandedMask.intersects(HighBits)) {
538 BinaryOperator *LShr = BinaryOperator::CreateLShr(I->getOperand(0),
540 LShr->setIsExact(cast<BinaryOperator>(I)->isExact());
541 return InsertNewInstWith(LShr, *I);
542 } else if (Known.One[BitWidth-ShiftAmt-1]) { // New bits are known one.
543 Known.One |= HighBits;
548 case Instruction::SRem:
549 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
550 // X % -1 demands all the bits because we don't want to introduce
551 // INT_MIN % -1 (== undef) by accident.
552 if (Rem->isMinusOne())
554 APInt RA = Rem->getValue().abs();
555 if (RA.isPowerOf2()) {
556 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
557 return I->getOperand(0);
559 APInt LowBits = RA - 1;
560 APInt Mask2 = LowBits | APInt::getSignMask(BitWidth);
561 if (SimplifyDemandedBits(I, 0, Mask2, LHSKnown, Depth + 1))
564 // The low bits of LHS are unchanged by the srem.
565 Known.Zero = LHSKnown.Zero & LowBits;
566 Known.One = LHSKnown.One & LowBits;
568 // If LHS is non-negative or has all low bits zero, then the upper bits
570 if (LHSKnown.isNonNegative() || LowBits.isSubsetOf(LHSKnown.Zero))
571 Known.Zero |= ~LowBits;
573 // If LHS is negative and not all low bits are zero, then the upper bits
575 if (LHSKnown.isNegative() && LowBits.intersects(LHSKnown.One))
576 Known.One |= ~LowBits;
578 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
583 // The sign bit is the LHS's sign bit, except when the result of the
584 // remainder is zero.
585 if (DemandedMask.isSignBitSet()) {
586 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1, CxtI);
587 // If it's known zero, our sign bit is also zero.
588 if (LHSKnown.isNonNegative())
589 Known.makeNonNegative();
592 case Instruction::URem: {
593 KnownBits Known2(BitWidth);
594 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
595 if (SimplifyDemandedBits(I, 0, AllOnes, Known2, Depth + 1) ||
596 SimplifyDemandedBits(I, 1, AllOnes, Known2, Depth + 1))
599 unsigned Leaders = Known2.countMinLeadingZeros();
600 Known.Zero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
603 case Instruction::Call:
604 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
605 switch (II->getIntrinsicID()) {
607 case Intrinsic::bswap: {
608 // If the only bits demanded come from one byte of the bswap result,
609 // just shift the input byte into position to eliminate the bswap.
610 unsigned NLZ = DemandedMask.countLeadingZeros();
611 unsigned NTZ = DemandedMask.countTrailingZeros();
613 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
614 // we need all the bits down to bit 8. Likewise, round NLZ. If we
615 // have 14 leading zeros, round to 8.
618 // If we need exactly one byte, we can do this transformation.
619 if (BitWidth-NLZ-NTZ == 8) {
620 unsigned ResultBit = NTZ;
621 unsigned InputBit = BitWidth-NTZ-8;
623 // Replace this with either a left or right shift to get the byte into
626 if (InputBit > ResultBit)
627 NewVal = BinaryOperator::CreateLShr(II->getArgOperand(0),
628 ConstantInt::get(I->getType(), InputBit-ResultBit));
630 NewVal = BinaryOperator::CreateShl(II->getArgOperand(0),
631 ConstantInt::get(I->getType(), ResultBit-InputBit));
633 return InsertNewInstWith(NewVal, *I);
636 // TODO: Could compute known zero/one bits based on the input.
639 case Intrinsic::x86_mmx_pmovmskb:
640 case Intrinsic::x86_sse_movmsk_ps:
641 case Intrinsic::x86_sse2_movmsk_pd:
642 case Intrinsic::x86_sse2_pmovmskb_128:
643 case Intrinsic::x86_avx_movmsk_ps_256:
644 case Intrinsic::x86_avx_movmsk_pd_256:
645 case Intrinsic::x86_avx2_pmovmskb: {
646 // MOVMSK copies the vector elements' sign bits to the low bits
647 // and zeros the high bits.
649 if (II->getIntrinsicID() == Intrinsic::x86_mmx_pmovmskb) {
650 ArgWidth = 8; // Arg is x86_mmx, but treated as <8 x i8>.
652 auto Arg = II->getArgOperand(0);
653 auto ArgType = cast<VectorType>(Arg->getType());
654 ArgWidth = ArgType->getNumElements();
657 // If we don't need any of low bits then return zero,
658 // we know that DemandedMask is non-zero already.
659 APInt DemandedElts = DemandedMask.zextOrTrunc(ArgWidth);
660 if (DemandedElts.isNullValue())
661 return ConstantInt::getNullValue(VTy);
663 // We know that the upper bits are set to zero.
664 Known.Zero.setBitsFrom(ArgWidth);
667 case Intrinsic::x86_sse42_crc32_64_64:
668 Known.Zero.setBitsFrom(32);
672 computeKnownBits(V, Known, Depth, CxtI);
676 // If the client is only demanding bits that we know, return the known
678 if (DemandedMask.isSubsetOf(Known.Zero|Known.One))
679 return Constant::getIntegerValue(VTy, Known.One);
683 /// Helper routine of SimplifyDemandedUseBits. It computes Known
684 /// bits. It also tries to handle simplifications that can be done based on
685 /// DemandedMask, but without modifying the Instruction.
686 Value *InstCombiner::SimplifyMultipleUseDemandedBits(Instruction *I,
687 const APInt &DemandedMask,
691 unsigned BitWidth = DemandedMask.getBitWidth();
692 Type *ITy = I->getType();
694 KnownBits LHSKnown(BitWidth);
695 KnownBits RHSKnown(BitWidth);
697 // Despite the fact that we can't simplify this instruction in all User's
698 // context, we can at least compute the known bits, and we can
699 // do simplifications that apply to *just* the one user if we know that
700 // this instruction has a simpler value in that context.
701 switch (I->getOpcode()) {
702 case Instruction::And: {
703 // If either the LHS or the RHS are Zero, the result is zero.
704 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
705 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1,
708 // Output known-0 are known to be clear if zero in either the LHS | RHS.
709 APInt IKnownZero = RHSKnown.Zero | LHSKnown.Zero;
710 // Output known-1 bits are only known if set in both the LHS & RHS.
711 APInt IKnownOne = RHSKnown.One & LHSKnown.One;
713 // If the client is only demanding bits that we know, return the known
715 if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
716 return Constant::getIntegerValue(ITy, IKnownOne);
718 // If all of the demanded bits are known 1 on one side, return the other.
719 // These bits cannot contribute to the result of the 'and' in this
721 if (DemandedMask.isSubsetOf(LHSKnown.Zero | RHSKnown.One))
722 return I->getOperand(0);
723 if (DemandedMask.isSubsetOf(RHSKnown.Zero | LHSKnown.One))
724 return I->getOperand(1);
726 Known.Zero = std::move(IKnownZero);
727 Known.One = std::move(IKnownOne);
730 case Instruction::Or: {
731 // We can simplify (X|Y) -> X or Y in the user's context if we know that
732 // only bits from X or Y are demanded.
734 // If either the LHS or the RHS are One, the result is One.
735 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
736 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1,
739 // Output known-0 bits are only known if clear in both the LHS & RHS.
740 APInt IKnownZero = RHSKnown.Zero & LHSKnown.Zero;
741 // Output known-1 are known to be set if set in either the LHS | RHS.
742 APInt IKnownOne = RHSKnown.One | LHSKnown.One;
744 // If the client is only demanding bits that we know, return the known
746 if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
747 return Constant::getIntegerValue(ITy, IKnownOne);
749 // If all of the demanded bits are known zero on one side, return the
750 // other. These bits cannot contribute to the result of the 'or' in this
752 if (DemandedMask.isSubsetOf(LHSKnown.One | RHSKnown.Zero))
753 return I->getOperand(0);
754 if (DemandedMask.isSubsetOf(RHSKnown.One | LHSKnown.Zero))
755 return I->getOperand(1);
757 Known.Zero = std::move(IKnownZero);
758 Known.One = std::move(IKnownOne);
761 case Instruction::Xor: {
762 // We can simplify (X^Y) -> X or Y in the user's context if we know that
763 // only bits from X or Y are demanded.
765 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
766 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1,
769 // Output known-0 bits are known if clear or set in both the LHS & RHS.
770 APInt IKnownZero = (RHSKnown.Zero & LHSKnown.Zero) |
771 (RHSKnown.One & LHSKnown.One);
772 // Output known-1 are known to be set if set in only one of the LHS, RHS.
773 APInt IKnownOne = (RHSKnown.Zero & LHSKnown.One) |
774 (RHSKnown.One & LHSKnown.Zero);
776 // If the client is only demanding bits that we know, return the known
778 if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
779 return Constant::getIntegerValue(ITy, IKnownOne);
781 // If all of the demanded bits are known zero on one side, return the
783 if (DemandedMask.isSubsetOf(RHSKnown.Zero))
784 return I->getOperand(0);
785 if (DemandedMask.isSubsetOf(LHSKnown.Zero))
786 return I->getOperand(1);
788 // Output known-0 bits are known if clear or set in both the LHS & RHS.
789 Known.Zero = std::move(IKnownZero);
790 // Output known-1 are known to be set if set in only one of the LHS, RHS.
791 Known.One = std::move(IKnownOne);
795 // Compute the Known bits to simplify things downstream.
796 computeKnownBits(I, Known, Depth, CxtI);
798 // If this user is only demanding bits that we know, return the known
800 if (DemandedMask.isSubsetOf(Known.Zero|Known.One))
801 return Constant::getIntegerValue(ITy, Known.One);
810 /// Helper routine of SimplifyDemandedUseBits. It tries to simplify
811 /// "E1 = (X lsr C1) << C2", where the C1 and C2 are constant, into
812 /// "E2 = X << (C2 - C1)" or "E2 = X >> (C1 - C2)", depending on the sign
815 /// Suppose E1 and E2 are generally different in bits S={bm, bm+1,
816 /// ..., bn}, without considering the specific value X is holding.
817 /// This transformation is legal iff one of following conditions is hold:
818 /// 1) All the bit in S are 0, in this case E1 == E2.
819 /// 2) We don't care those bits in S, per the input DemandedMask.
820 /// 3) Combination of 1) and 2). Some bits in S are 0, and we don't care the
823 /// Currently we only test condition 2).
825 /// As with SimplifyDemandedUseBits, it returns NULL if the simplification was
828 InstCombiner::simplifyShrShlDemandedBits(Instruction *Shr, const APInt &ShrOp1,
829 Instruction *Shl, const APInt &ShlOp1,
830 const APInt &DemandedMask,
832 if (!ShlOp1 || !ShrOp1)
833 return nullptr; // No-op.
835 Value *VarX = Shr->getOperand(0);
836 Type *Ty = VarX->getType();
837 unsigned BitWidth = Ty->getScalarSizeInBits();
838 if (ShlOp1.uge(BitWidth) || ShrOp1.uge(BitWidth))
839 return nullptr; // Undef.
841 unsigned ShlAmt = ShlOp1.getZExtValue();
842 unsigned ShrAmt = ShrOp1.getZExtValue();
844 Known.One.clearAllBits();
845 Known.Zero.setLowBits(ShlAmt - 1);
846 Known.Zero &= DemandedMask;
848 APInt BitMask1(APInt::getAllOnesValue(BitWidth));
849 APInt BitMask2(APInt::getAllOnesValue(BitWidth));
851 bool isLshr = (Shr->getOpcode() == Instruction::LShr);
852 BitMask1 = isLshr ? (BitMask1.lshr(ShrAmt) << ShlAmt) :
853 (BitMask1.ashr(ShrAmt) << ShlAmt);
855 if (ShrAmt <= ShlAmt) {
856 BitMask2 <<= (ShlAmt - ShrAmt);
858 BitMask2 = isLshr ? BitMask2.lshr(ShrAmt - ShlAmt):
859 BitMask2.ashr(ShrAmt - ShlAmt);
862 // Check if condition-2 (see the comment to this function) is satified.
863 if ((BitMask1 & DemandedMask) == (BitMask2 & DemandedMask)) {
864 if (ShrAmt == ShlAmt)
867 if (!Shr->hasOneUse())
871 if (ShrAmt < ShlAmt) {
872 Constant *Amt = ConstantInt::get(VarX->getType(), ShlAmt - ShrAmt);
873 New = BinaryOperator::CreateShl(VarX, Amt);
874 BinaryOperator *Orig = cast<BinaryOperator>(Shl);
875 New->setHasNoSignedWrap(Orig->hasNoSignedWrap());
876 New->setHasNoUnsignedWrap(Orig->hasNoUnsignedWrap());
878 Constant *Amt = ConstantInt::get(VarX->getType(), ShrAmt - ShlAmt);
879 New = isLshr ? BinaryOperator::CreateLShr(VarX, Amt) :
880 BinaryOperator::CreateAShr(VarX, Amt);
881 if (cast<BinaryOperator>(Shr)->isExact())
882 New->setIsExact(true);
885 return InsertNewInstWith(New, *Shl);
891 /// The specified value produces a vector with any number of elements.
892 /// DemandedElts contains the set of elements that are actually used by the
893 /// caller. This method analyzes which elements of the operand are undef and
894 /// returns that information in UndefElts.
896 /// If the information about demanded elements can be used to simplify the
897 /// operation, the operation is simplified, then the resultant value is
898 /// returned. This returns null if no change was made.
899 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
902 unsigned VWidth = V->getType()->getVectorNumElements();
903 APInt EltMask(APInt::getAllOnesValue(VWidth));
904 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
906 if (isa<UndefValue>(V)) {
907 // If the entire vector is undefined, just return this info.
912 if (DemandedElts.isNullValue()) { // If nothing is demanded, provide undef.
914 return UndefValue::get(V->getType());
919 // Handle ConstantAggregateZero, ConstantVector, ConstantDataSequential.
920 if (Constant *C = dyn_cast<Constant>(V)) {
921 // Check if this is identity. If so, return 0 since we are not simplifying
923 if (DemandedElts.isAllOnesValue())
926 Type *EltTy = cast<VectorType>(V->getType())->getElementType();
927 Constant *Undef = UndefValue::get(EltTy);
929 SmallVector<Constant*, 16> Elts;
930 for (unsigned i = 0; i != VWidth; ++i) {
931 if (!DemandedElts[i]) { // If not demanded, set to undef.
932 Elts.push_back(Undef);
937 Constant *Elt = C->getAggregateElement(i);
938 if (!Elt) return nullptr;
940 if (isa<UndefValue>(Elt)) { // Already undef.
941 Elts.push_back(Undef);
943 } else { // Otherwise, defined.
948 // If we changed the constant, return it.
949 Constant *NewCV = ConstantVector::get(Elts);
950 return NewCV != C ? NewCV : nullptr;
953 // Limit search depth.
957 // If multiple users are using the root value, proceed with
958 // simplification conservatively assuming that all elements
960 if (!V->hasOneUse()) {
961 // Quit if we find multiple users of a non-root value though.
962 // They'll be handled when it's their turn to be visited by
963 // the main instcombine process.
965 // TODO: Just compute the UndefElts information recursively.
968 // Conservatively assume that all elements are needed.
969 DemandedElts = EltMask;
972 Instruction *I = dyn_cast<Instruction>(V);
973 if (!I) return nullptr; // Only analyze instructions.
975 bool MadeChange = false;
976 APInt UndefElts2(VWidth, 0);
977 APInt UndefElts3(VWidth, 0);
979 switch (I->getOpcode()) {
982 case Instruction::InsertElement: {
983 // If this is a variable index, we don't know which element it overwrites.
984 // demand exactly the same input as we produce.
985 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
987 // Note that we can't propagate undef elt info, because we don't know
988 // which elt is getting updated.
989 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
990 UndefElts2, Depth + 1);
991 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
995 // The element inserted overwrites whatever was there, so the input demanded
996 // set is simpler than the output set.
997 unsigned IdxNo = Idx->getZExtValue();
998 APInt PreInsertDemandedElts = DemandedElts;
1000 PreInsertDemandedElts.clearBit(IdxNo);
1001 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), PreInsertDemandedElts,
1002 UndefElts, Depth + 1);
1003 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1005 // If this is inserting an element that isn't demanded, remove this
1007 if (IdxNo >= VWidth || !DemandedElts[IdxNo]) {
1009 return I->getOperand(0);
1012 // The inserted element is defined.
1013 UndefElts.clearBit(IdxNo);
1016 case Instruction::ShuffleVector: {
1017 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1018 unsigned LHSVWidth =
1019 Shuffle->getOperand(0)->getType()->getVectorNumElements();
1020 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1021 for (unsigned i = 0; i < VWidth; i++) {
1022 if (DemandedElts[i]) {
1023 unsigned MaskVal = Shuffle->getMaskValue(i);
1024 if (MaskVal != -1u) {
1025 assert(MaskVal < LHSVWidth * 2 &&
1026 "shufflevector mask index out of range!");
1027 if (MaskVal < LHSVWidth)
1028 LeftDemanded.setBit(MaskVal);
1030 RightDemanded.setBit(MaskVal - LHSVWidth);
1035 APInt LHSUndefElts(LHSVWidth, 0);
1036 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1037 LHSUndefElts, Depth + 1);
1038 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1040 APInt RHSUndefElts(LHSVWidth, 0);
1041 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1042 RHSUndefElts, Depth + 1);
1043 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1045 bool NewUndefElts = false;
1046 unsigned LHSIdx = -1u, LHSValIdx = -1u;
1047 unsigned RHSIdx = -1u, RHSValIdx = -1u;
1048 bool LHSUniform = true;
1049 bool RHSUniform = true;
1050 for (unsigned i = 0; i < VWidth; i++) {
1051 unsigned MaskVal = Shuffle->getMaskValue(i);
1052 if (MaskVal == -1u) {
1053 UndefElts.setBit(i);
1054 } else if (!DemandedElts[i]) {
1055 NewUndefElts = true;
1056 UndefElts.setBit(i);
1057 } else if (MaskVal < LHSVWidth) {
1058 if (LHSUndefElts[MaskVal]) {
1059 NewUndefElts = true;
1060 UndefElts.setBit(i);
1062 LHSIdx = LHSIdx == -1u ? i : LHSVWidth;
1063 LHSValIdx = LHSValIdx == -1u ? MaskVal : LHSVWidth;
1064 LHSUniform = LHSUniform && (MaskVal == i);
1067 if (RHSUndefElts[MaskVal - LHSVWidth]) {
1068 NewUndefElts = true;
1069 UndefElts.setBit(i);
1071 RHSIdx = RHSIdx == -1u ? i : LHSVWidth;
1072 RHSValIdx = RHSValIdx == -1u ? MaskVal - LHSVWidth : LHSVWidth;
1073 RHSUniform = RHSUniform && (MaskVal - LHSVWidth == i);
1078 // Try to transform shuffle with constant vector and single element from
1079 // this constant vector to single insertelement instruction.
1080 // shufflevector V, C, <v1, v2, .., ci, .., vm> ->
1081 // insertelement V, C[ci], ci-n
1082 if (LHSVWidth == Shuffle->getType()->getNumElements()) {
1083 Value *Op = nullptr;
1084 Constant *Value = nullptr;
1087 // Find constant vector with the single element in shuffle (LHS or RHS).
1088 if (LHSIdx < LHSVWidth && RHSUniform) {
1089 if (auto *CV = dyn_cast<ConstantVector>(Shuffle->getOperand(0))) {
1090 Op = Shuffle->getOperand(1);
1091 Value = CV->getOperand(LHSValIdx);
1095 if (RHSIdx < LHSVWidth && LHSUniform) {
1096 if (auto *CV = dyn_cast<ConstantVector>(Shuffle->getOperand(1))) {
1097 Op = Shuffle->getOperand(0);
1098 Value = CV->getOperand(RHSValIdx);
1102 // Found constant vector with single element - convert to insertelement.
1104 Instruction *New = InsertElementInst::Create(
1105 Op, Value, ConstantInt::get(Type::getInt32Ty(I->getContext()), Idx),
1106 Shuffle->getName());
1107 InsertNewInstWith(New, *Shuffle);
1112 // Add additional discovered undefs.
1113 SmallVector<Constant*, 16> Elts;
1114 for (unsigned i = 0; i < VWidth; ++i) {
1116 Elts.push_back(UndefValue::get(Type::getInt32Ty(I->getContext())));
1118 Elts.push_back(ConstantInt::get(Type::getInt32Ty(I->getContext()),
1119 Shuffle->getMaskValue(i)));
1121 I->setOperand(2, ConstantVector::get(Elts));
1126 case Instruction::Select: {
1127 APInt LeftDemanded(DemandedElts), RightDemanded(DemandedElts);
1128 if (ConstantVector* CV = dyn_cast<ConstantVector>(I->getOperand(0))) {
1129 for (unsigned i = 0; i < VWidth; i++) {
1130 Constant *CElt = CV->getAggregateElement(i);
1131 // Method isNullValue always returns false when called on a
1132 // ConstantExpr. If CElt is a ConstantExpr then skip it in order to
1133 // to avoid propagating incorrect information.
1134 if (isa<ConstantExpr>(CElt))
1136 if (CElt->isNullValue())
1137 LeftDemanded.clearBit(i);
1139 RightDemanded.clearBit(i);
1143 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), LeftDemanded, UndefElts,
1145 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1147 TmpV = SimplifyDemandedVectorElts(I->getOperand(2), RightDemanded,
1148 UndefElts2, Depth + 1);
1149 if (TmpV) { I->setOperand(2, TmpV); MadeChange = true; }
1151 // Output elements are undefined if both are undefined.
1152 UndefElts &= UndefElts2;
1155 case Instruction::BitCast: {
1156 // Vector->vector casts only.
1157 VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1159 unsigned InVWidth = VTy->getNumElements();
1160 APInt InputDemandedElts(InVWidth, 0);
1161 UndefElts2 = APInt(InVWidth, 0);
1164 if (VWidth == InVWidth) {
1165 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1166 // elements as are demanded of us.
1168 InputDemandedElts = DemandedElts;
1169 } else if ((VWidth % InVWidth) == 0) {
1170 // If the number of elements in the output is a multiple of the number of
1171 // elements in the input then an input element is live if any of the
1172 // corresponding output elements are live.
1173 Ratio = VWidth / InVWidth;
1174 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1175 if (DemandedElts[OutIdx])
1176 InputDemandedElts.setBit(OutIdx / Ratio);
1177 } else if ((InVWidth % VWidth) == 0) {
1178 // If the number of elements in the input is a multiple of the number of
1179 // elements in the output then an input element is live if the
1180 // corresponding output element is live.
1181 Ratio = InVWidth / VWidth;
1182 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1183 if (DemandedElts[InIdx / Ratio])
1184 InputDemandedElts.setBit(InIdx);
1186 // Unsupported so far.
1190 // div/rem demand all inputs, because they don't want divide by zero.
1191 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1192 UndefElts2, Depth + 1);
1194 I->setOperand(0, TmpV);
1198 if (VWidth == InVWidth) {
1199 UndefElts = UndefElts2;
1200 } else if ((VWidth % InVWidth) == 0) {
1201 // If the number of elements in the output is a multiple of the number of
1202 // elements in the input then an output element is undef if the
1203 // corresponding input element is undef.
1204 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1205 if (UndefElts2[OutIdx / Ratio])
1206 UndefElts.setBit(OutIdx);
1207 } else if ((InVWidth % VWidth) == 0) {
1208 // If the number of elements in the input is a multiple of the number of
1209 // elements in the output then an output element is undef if all of the
1210 // corresponding input elements are undef.
1211 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1212 APInt SubUndef = UndefElts2.lshr(OutIdx * Ratio).zextOrTrunc(Ratio);
1213 if (SubUndef.countPopulation() == Ratio)
1214 UndefElts.setBit(OutIdx);
1217 llvm_unreachable("Unimp");
1221 case Instruction::And:
1222 case Instruction::Or:
1223 case Instruction::Xor:
1224 case Instruction::Add:
1225 case Instruction::Sub:
1226 case Instruction::Mul:
1227 // div/rem demand all inputs, because they don't want divide by zero.
1228 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts, UndefElts,
1230 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1231 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1232 UndefElts2, Depth + 1);
1233 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1235 // Output elements are undefined if both are undefined. Consider things
1236 // like undef&0. The result is known zero, not undef.
1237 UndefElts &= UndefElts2;
1239 case Instruction::FPTrunc:
1240 case Instruction::FPExt:
1241 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts, UndefElts,
1243 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1246 case Instruction::Call: {
1247 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1249 switch (II->getIntrinsicID()) {
1252 case Intrinsic::x86_xop_vfrcz_ss:
1253 case Intrinsic::x86_xop_vfrcz_sd:
1254 // The instructions for these intrinsics are speced to zero upper bits not
1255 // pass them through like other scalar intrinsics. So we shouldn't just
1256 // use Arg0 if DemandedElts[0] is clear like we do for other intrinsics.
1257 // Instead we should return a zero vector.
1258 if (!DemandedElts[0]) {
1260 return ConstantAggregateZero::get(II->getType());
1263 // Only the lower element is used.
1265 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts,
1266 UndefElts, Depth + 1);
1267 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1269 // Only the lower element is undefined. The high elements are zero.
1270 UndefElts = UndefElts[0];
1273 // Unary scalar-as-vector operations that work column-wise.
1274 case Intrinsic::x86_sse_rcp_ss:
1275 case Intrinsic::x86_sse_rsqrt_ss:
1276 case Intrinsic::x86_sse_sqrt_ss:
1277 case Intrinsic::x86_sse2_sqrt_sd:
1278 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts,
1279 UndefElts, Depth + 1);
1280 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1282 // If lowest element of a scalar op isn't used then use Arg0.
1283 if (!DemandedElts[0]) {
1285 return II->getArgOperand(0);
1287 // TODO: If only low elt lower SQRT to FSQRT (with rounding/exceptions
1291 // Binary scalar-as-vector operations that work column-wise. The high
1292 // elements come from operand 0. The low element is a function of both
1294 case Intrinsic::x86_sse_min_ss:
1295 case Intrinsic::x86_sse_max_ss:
1296 case Intrinsic::x86_sse_cmp_ss:
1297 case Intrinsic::x86_sse2_min_sd:
1298 case Intrinsic::x86_sse2_max_sd:
1299 case Intrinsic::x86_sse2_cmp_sd: {
1300 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts,
1301 UndefElts, Depth + 1);
1302 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1304 // If lowest element of a scalar op isn't used then use Arg0.
1305 if (!DemandedElts[0]) {
1307 return II->getArgOperand(0);
1310 // Only lower element is used for operand 1.
1312 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(1), DemandedElts,
1313 UndefElts2, Depth + 1);
1314 if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1316 // Lower element is undefined if both lower elements are undefined.
1317 // Consider things like undef&0. The result is known zero, not undef.
1319 UndefElts.clearBit(0);
1324 // Binary scalar-as-vector operations that work column-wise. The high
1325 // elements come from operand 0 and the low element comes from operand 1.
1326 case Intrinsic::x86_sse41_round_ss:
1327 case Intrinsic::x86_sse41_round_sd: {
1328 // Don't use the low element of operand 0.
1329 APInt DemandedElts2 = DemandedElts;
1330 DemandedElts2.clearBit(0);
1331 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts2,
1332 UndefElts, Depth + 1);
1333 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1335 // If lowest element of a scalar op isn't used then use Arg0.
1336 if (!DemandedElts[0]) {
1338 return II->getArgOperand(0);
1341 // Only lower element is used for operand 1.
1343 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(1), DemandedElts,
1344 UndefElts2, Depth + 1);
1345 if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1347 // Take the high undef elements from operand 0 and take the lower element
1349 UndefElts.clearBit(0);
1350 UndefElts |= UndefElts2[0];
1354 // Three input scalar-as-vector operations that work column-wise. The high
1355 // elements come from operand 0 and the low element is a function of all
1357 case Intrinsic::x86_avx512_mask_add_ss_round:
1358 case Intrinsic::x86_avx512_mask_div_ss_round:
1359 case Intrinsic::x86_avx512_mask_mul_ss_round:
1360 case Intrinsic::x86_avx512_mask_sub_ss_round:
1361 case Intrinsic::x86_avx512_mask_max_ss_round:
1362 case Intrinsic::x86_avx512_mask_min_ss_round:
1363 case Intrinsic::x86_avx512_mask_add_sd_round:
1364 case Intrinsic::x86_avx512_mask_div_sd_round:
1365 case Intrinsic::x86_avx512_mask_mul_sd_round:
1366 case Intrinsic::x86_avx512_mask_sub_sd_round:
1367 case Intrinsic::x86_avx512_mask_max_sd_round:
1368 case Intrinsic::x86_avx512_mask_min_sd_round:
1369 case Intrinsic::x86_fma_vfmadd_ss:
1370 case Intrinsic::x86_fma_vfmsub_ss:
1371 case Intrinsic::x86_fma_vfnmadd_ss:
1372 case Intrinsic::x86_fma_vfnmsub_ss:
1373 case Intrinsic::x86_fma_vfmadd_sd:
1374 case Intrinsic::x86_fma_vfmsub_sd:
1375 case Intrinsic::x86_fma_vfnmadd_sd:
1376 case Intrinsic::x86_fma_vfnmsub_sd:
1377 case Intrinsic::x86_avx512_mask_vfmadd_ss:
1378 case Intrinsic::x86_avx512_mask_vfmadd_sd:
1379 case Intrinsic::x86_avx512_maskz_vfmadd_ss:
1380 case Intrinsic::x86_avx512_maskz_vfmadd_sd:
1381 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts,
1382 UndefElts, Depth + 1);
1383 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1385 // If lowest element of a scalar op isn't used then use Arg0.
1386 if (!DemandedElts[0]) {
1388 return II->getArgOperand(0);
1391 // Only lower element is used for operand 1 and 2.
1393 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(1), DemandedElts,
1394 UndefElts2, Depth + 1);
1395 if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1396 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(2), DemandedElts,
1397 UndefElts3, Depth + 1);
1398 if (TmpV) { II->setArgOperand(2, TmpV); MadeChange = true; }
1400 // Lower element is undefined if all three lower elements are undefined.
1401 // Consider things like undef&0. The result is known zero, not undef.
1402 if (!UndefElts2[0] || !UndefElts3[0])
1403 UndefElts.clearBit(0);
1407 case Intrinsic::x86_avx512_mask3_vfmadd_ss:
1408 case Intrinsic::x86_avx512_mask3_vfmadd_sd:
1409 case Intrinsic::x86_avx512_mask3_vfmsub_ss:
1410 case Intrinsic::x86_avx512_mask3_vfmsub_sd:
1411 case Intrinsic::x86_avx512_mask3_vfnmsub_ss:
1412 case Intrinsic::x86_avx512_mask3_vfnmsub_sd:
1413 // These intrinsics get the passthru bits from operand 2.
1414 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(2), DemandedElts,
1415 UndefElts, Depth + 1);
1416 if (TmpV) { II->setArgOperand(2, TmpV); MadeChange = true; }
1418 // If lowest element of a scalar op isn't used then use Arg2.
1419 if (!DemandedElts[0]) {
1421 return II->getArgOperand(2);
1424 // Only lower element is used for operand 0 and 1.
1426 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts,
1427 UndefElts2, Depth + 1);
1428 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1429 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(1), DemandedElts,
1430 UndefElts3, Depth + 1);
1431 if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1433 // Lower element is undefined if all three lower elements are undefined.
1434 // Consider things like undef&0. The result is known zero, not undef.
1435 if (!UndefElts2[0] || !UndefElts3[0])
1436 UndefElts.clearBit(0);
1440 case Intrinsic::x86_sse2_pmulu_dq:
1441 case Intrinsic::x86_sse41_pmuldq:
1442 case Intrinsic::x86_avx2_pmul_dq:
1443 case Intrinsic::x86_avx2_pmulu_dq:
1444 case Intrinsic::x86_avx512_pmul_dq_512:
1445 case Intrinsic::x86_avx512_pmulu_dq_512: {
1446 Value *Op0 = II->getArgOperand(0);
1447 Value *Op1 = II->getArgOperand(1);
1448 unsigned InnerVWidth = Op0->getType()->getVectorNumElements();
1449 assert((VWidth * 2) == InnerVWidth && "Unexpected input size");
1451 APInt InnerDemandedElts(InnerVWidth, 0);
1452 for (unsigned i = 0; i != VWidth; ++i)
1453 if (DemandedElts[i])
1454 InnerDemandedElts.setBit(i * 2);
1456 UndefElts2 = APInt(InnerVWidth, 0);
1457 TmpV = SimplifyDemandedVectorElts(Op0, InnerDemandedElts, UndefElts2,
1459 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1461 UndefElts3 = APInt(InnerVWidth, 0);
1462 TmpV = SimplifyDemandedVectorElts(Op1, InnerDemandedElts, UndefElts3,
1464 if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1469 case Intrinsic::x86_sse2_packssdw_128:
1470 case Intrinsic::x86_sse2_packsswb_128:
1471 case Intrinsic::x86_sse2_packuswb_128:
1472 case Intrinsic::x86_sse41_packusdw:
1473 case Intrinsic::x86_avx2_packssdw:
1474 case Intrinsic::x86_avx2_packsswb:
1475 case Intrinsic::x86_avx2_packusdw:
1476 case Intrinsic::x86_avx2_packuswb:
1477 case Intrinsic::x86_avx512_packssdw_512:
1478 case Intrinsic::x86_avx512_packsswb_512:
1479 case Intrinsic::x86_avx512_packusdw_512:
1480 case Intrinsic::x86_avx512_packuswb_512: {
1481 auto *Ty0 = II->getArgOperand(0)->getType();
1482 unsigned InnerVWidth = Ty0->getVectorNumElements();
1483 assert(VWidth == (InnerVWidth * 2) && "Unexpected input size");
1485 unsigned NumLanes = Ty0->getPrimitiveSizeInBits() / 128;
1486 unsigned VWidthPerLane = VWidth / NumLanes;
1487 unsigned InnerVWidthPerLane = InnerVWidth / NumLanes;
1489 // Per lane, pack the elements of the first input and then the second.
1491 // v8i16 PACK(v4i32 X, v4i32 Y) - (X[0..3],Y[0..3])
1492 // v32i8 PACK(v16i16 X, v16i16 Y) - (X[0..7],Y[0..7]),(X[8..15],Y[8..15])
1493 for (int OpNum = 0; OpNum != 2; ++OpNum) {
1494 APInt OpDemandedElts(InnerVWidth, 0);
1495 for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
1496 unsigned LaneIdx = Lane * VWidthPerLane;
1497 for (unsigned Elt = 0; Elt != InnerVWidthPerLane; ++Elt) {
1498 unsigned Idx = LaneIdx + Elt + InnerVWidthPerLane * OpNum;
1499 if (DemandedElts[Idx])
1500 OpDemandedElts.setBit((Lane * InnerVWidthPerLane) + Elt);
1504 // Demand elements from the operand.
1505 auto *Op = II->getArgOperand(OpNum);
1506 APInt OpUndefElts(InnerVWidth, 0);
1507 TmpV = SimplifyDemandedVectorElts(Op, OpDemandedElts, OpUndefElts,
1510 II->setArgOperand(OpNum, TmpV);
1514 // Pack the operand's UNDEF elements, one lane at a time.
1515 OpUndefElts = OpUndefElts.zext(VWidth);
1516 for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
1517 APInt LaneElts = OpUndefElts.lshr(InnerVWidthPerLane * Lane);
1518 LaneElts = LaneElts.getLoBits(InnerVWidthPerLane);
1519 LaneElts <<= InnerVWidthPerLane * (2 * Lane + OpNum);
1520 UndefElts |= LaneElts;
1527 case Intrinsic::x86_ssse3_pshuf_b_128:
1528 case Intrinsic::x86_avx2_pshuf_b:
1529 case Intrinsic::x86_avx512_pshuf_b_512:
1531 case Intrinsic::x86_avx_vpermilvar_ps:
1532 case Intrinsic::x86_avx_vpermilvar_ps_256:
1533 case Intrinsic::x86_avx512_vpermilvar_ps_512:
1534 case Intrinsic::x86_avx_vpermilvar_pd:
1535 case Intrinsic::x86_avx_vpermilvar_pd_256:
1536 case Intrinsic::x86_avx512_vpermilvar_pd_512:
1538 case Intrinsic::x86_avx2_permd:
1539 case Intrinsic::x86_avx2_permps: {
1540 Value *Op1 = II->getArgOperand(1);
1541 TmpV = SimplifyDemandedVectorElts(Op1, DemandedElts, UndefElts,
1543 if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1547 // SSE4A instructions leave the upper 64-bits of the 128-bit result
1548 // in an undefined state.
1549 case Intrinsic::x86_sse4a_extrq:
1550 case Intrinsic::x86_sse4a_extrqi:
1551 case Intrinsic::x86_sse4a_insertq:
1552 case Intrinsic::x86_sse4a_insertqi:
1553 UndefElts.setHighBits(VWidth / 2);
1555 case Intrinsic::amdgcn_buffer_load:
1556 case Intrinsic::amdgcn_buffer_load_format:
1557 case Intrinsic::amdgcn_image_sample:
1558 case Intrinsic::amdgcn_image_sample_cl:
1559 case Intrinsic::amdgcn_image_sample_d:
1560 case Intrinsic::amdgcn_image_sample_d_cl:
1561 case Intrinsic::amdgcn_image_sample_l:
1562 case Intrinsic::amdgcn_image_sample_b:
1563 case Intrinsic::amdgcn_image_sample_b_cl:
1564 case Intrinsic::amdgcn_image_sample_lz:
1565 case Intrinsic::amdgcn_image_sample_cd:
1566 case Intrinsic::amdgcn_image_sample_cd_cl:
1568 case Intrinsic::amdgcn_image_sample_c:
1569 case Intrinsic::amdgcn_image_sample_c_cl:
1570 case Intrinsic::amdgcn_image_sample_c_d:
1571 case Intrinsic::amdgcn_image_sample_c_d_cl:
1572 case Intrinsic::amdgcn_image_sample_c_l:
1573 case Intrinsic::amdgcn_image_sample_c_b:
1574 case Intrinsic::amdgcn_image_sample_c_b_cl:
1575 case Intrinsic::amdgcn_image_sample_c_lz:
1576 case Intrinsic::amdgcn_image_sample_c_cd:
1577 case Intrinsic::amdgcn_image_sample_c_cd_cl:
1579 case Intrinsic::amdgcn_image_sample_o:
1580 case Intrinsic::amdgcn_image_sample_cl_o:
1581 case Intrinsic::amdgcn_image_sample_d_o:
1582 case Intrinsic::amdgcn_image_sample_d_cl_o:
1583 case Intrinsic::amdgcn_image_sample_l_o:
1584 case Intrinsic::amdgcn_image_sample_b_o:
1585 case Intrinsic::amdgcn_image_sample_b_cl_o:
1586 case Intrinsic::amdgcn_image_sample_lz_o:
1587 case Intrinsic::amdgcn_image_sample_cd_o:
1588 case Intrinsic::amdgcn_image_sample_cd_cl_o:
1590 case Intrinsic::amdgcn_image_sample_c_o:
1591 case Intrinsic::amdgcn_image_sample_c_cl_o:
1592 case Intrinsic::amdgcn_image_sample_c_d_o:
1593 case Intrinsic::amdgcn_image_sample_c_d_cl_o:
1594 case Intrinsic::amdgcn_image_sample_c_l_o:
1595 case Intrinsic::amdgcn_image_sample_c_b_o:
1596 case Intrinsic::amdgcn_image_sample_c_b_cl_o:
1597 case Intrinsic::amdgcn_image_sample_c_lz_o:
1598 case Intrinsic::amdgcn_image_sample_c_cd_o:
1599 case Intrinsic::amdgcn_image_sample_c_cd_cl_o:
1601 case Intrinsic::amdgcn_image_getlod: {
1602 if (VWidth == 1 || !DemandedElts.isMask())
1605 // TODO: Handle 3 vectors when supported in code gen.
1606 unsigned NewNumElts = PowerOf2Ceil(DemandedElts.countTrailingOnes());
1607 if (NewNumElts == VWidth)
1610 Module *M = II->getParent()->getParent()->getParent();
1611 Type *EltTy = V->getType()->getVectorElementType();
1613 Type *NewTy = (NewNumElts == 1) ? EltTy :
1614 VectorType::get(EltTy, NewNumElts);
1616 auto IID = II->getIntrinsicID();
1618 bool IsBuffer = IID == Intrinsic::amdgcn_buffer_load ||
1619 IID == Intrinsic::amdgcn_buffer_load_format;
1621 Function *NewIntrin = IsBuffer ?
1622 Intrinsic::getDeclaration(M, IID, NewTy) :
1623 // Samplers have 3 mangled types.
1624 Intrinsic::getDeclaration(M, IID,
1625 { NewTy, II->getArgOperand(0)->getType(),
1626 II->getArgOperand(1)->getType()});
1628 SmallVector<Value *, 5> Args;
1629 for (unsigned I = 0, E = II->getNumArgOperands(); I != E; ++I)
1630 Args.push_back(II->getArgOperand(I));
1632 IRBuilderBase::InsertPointGuard Guard(Builder);
1633 Builder.SetInsertPoint(II);
1635 CallInst *NewCall = Builder.CreateCall(NewIntrin, Args);
1636 NewCall->takeName(II);
1637 NewCall->copyMetadata(*II);
1640 ConstantInt *DMask = dyn_cast<ConstantInt>(NewCall->getArgOperand(3));
1642 unsigned DMaskVal = DMask->getZExtValue() & 0xf;
1644 unsigned PopCnt = 0;
1645 unsigned NewDMask = 0;
1646 for (unsigned I = 0; I < 4; ++I) {
1647 const unsigned Bit = 1 << I;
1648 if (!!(DMaskVal & Bit)) {
1649 if (++PopCnt > NewNumElts)
1656 NewCall->setArgOperand(3, ConstantInt::get(DMask->getType(), NewDMask));
1661 if (NewNumElts == 1) {
1662 return Builder.CreateInsertElement(UndefValue::get(V->getType()),
1663 NewCall, static_cast<uint64_t>(0));
1666 SmallVector<uint32_t, 8> EltMask;
1667 for (unsigned I = 0; I < VWidth; ++I)
1668 EltMask.push_back(I);
1670 Value *Shuffle = Builder.CreateShuffleVector(
1671 NewCall, UndefValue::get(NewTy), EltMask);
1680 return MadeChange ? I : nullptr;