LLVM  6.0.0svn
InstCombineSimplifyDemanded.cpp
Go to the documentation of this file.
1 //===- InstCombineSimplifyDemanded.cpp ------------------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains logic for simplifying instructions based on information
11 // about how they are used.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "InstCombineInternal.h"
17 #include "llvm/IR/IntrinsicInst.h"
18 #include "llvm/IR/PatternMatch.h"
19 #include "llvm/Support/KnownBits.h"
20 
21 using namespace llvm;
22 using namespace llvm::PatternMatch;
23 
24 #define DEBUG_TYPE "instcombine"
25 
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");
33 
34  // The operand must be a constant integer or splat integer.
35  Value *Op = I->getOperand(OpNo);
36  const APInt *C;
37  if (!match(Op, m_APInt(C)))
38  return false;
39 
40  // If there are no bits set that aren't demanded, nothing to do.
41  if (C->isSubsetOf(Demanded))
42  return false;
43 
44  // This instruction is producing bits that are not demanded. Shrink the RHS.
45  I->setOperand(OpNo, ConstantInt::get(Op->getType(), *C & Demanded));
46 
47  return true;
48 }
49 
50 
51 
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));
58 
59  Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask, Known,
60  0, &Inst);
61  if (!V) return false;
62  if (V == &Inst) return true;
63  replaceInstUsesWith(Inst, V);
64  return true;
65 }
66 
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,
72  KnownBits &Known,
73  unsigned Depth) {
74  Use &U = I->getOperandUse(OpNo);
75  Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask, Known,
76  Depth, I);
77  if (!NewVal) return false;
78  U = NewVal;
79  return true;
80 }
81 
82 
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
93 /// expression.
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
99 /// be the same.
100 ///
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,
108  Instruction *CxtI) {
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();
113  assert(
114  (!VTy->isIntOrIntVectorTy() || VTy->getScalarSizeInBits() == BitWidth) &&
115  Known.getBitWidth() == BitWidth &&
116  "Value *V, DemandedMask and Known must have same BitWidth");
117 
118  if (isa<Constant>(V)) {
119  computeKnownBits(V, Known, Depth, CxtI);
120  return nullptr;
121  }
122 
123  Known.resetAll();
124  if (DemandedMask.isNullValue()) // Not demanding any bits from V.
125  return UndefValue::get(VTy);
126 
127  if (Depth == 6) // Limit search depth.
128  return nullptr;
129 
131  if (!I) {
132  computeKnownBits(V, Known, Depth, CxtI);
133  return nullptr; // Only analyze instructions.
134  }
135 
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);
141 
142  KnownBits LHSKnown(BitWidth), RHSKnown(BitWidth);
143 
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();
150 
151  switch (I->getOpcode()) {
152  default:
153  computeKnownBits(I, Known, Depth, CxtI);
154  break;
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,
159  Depth + 1))
160  return I;
161  assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
162  assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
163 
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;
168 
169  // If the client is only demanding bits that we know, return the known
170  // constant.
171  if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
172  return Constant::getIntegerValue(VTy, IKnownOne);
173 
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);
180 
181  // If the RHS is a constant, see if we can simplify it.
182  if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnown.Zero))
183  return I;
184 
185  Known.Zero = std::move(IKnownZero);
186  Known.One = std::move(IKnownOne);
187  break;
188  }
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,
193  Depth + 1))
194  return I;
195  assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
196  assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
197 
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;
202 
203  // If the client is only demanding bits that we know, return the known
204  // constant.
205  if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
206  return Constant::getIntegerValue(VTy, IKnownOne);
207 
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);
214 
215  // If the RHS is a constant, see if we can simplify it.
216  if (ShrinkDemandedConstant(I, 1, DemandedMask))
217  return I;
218 
219  Known.Zero = std::move(IKnownZero);
220  Known.One = std::move(IKnownOne);
221  break;
222  }
223  case Instruction::Xor: {
224  if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) ||
225  SimplifyDemandedBits(I, 0, DemandedMask, LHSKnown, Depth + 1))
226  return I;
227  assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
228  assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
229 
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);
236 
237  // If the client is only demanding bits that we know, return the known
238  // constant.
239  if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
240  return Constant::getIntegerValue(VTy, IKnownOne);
241 
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);
248 
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)) {
253  Instruction *Or =
254  BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
255  I->getName());
256  return InsertNewInstWith(Or, *I);
257  }
258 
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)) {
266  ~RHSKnown.One & DemandedMask);
267  Instruction *And = BinaryOperator::CreateAnd(I->getOperand(0), AndC);
268  return InsertNewInstWith(And, *I);
269  }
270 
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))
274  return I;
275 
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);
288 
289  Constant *AndC =
290  ConstantInt::get(I->getType(), NewMask & AndRHS->getValue());
291  Instruction *NewAnd = BinaryOperator::CreateAnd(I->getOperand(0), AndC);
292  InsertNewInstWith(NewAnd, *I);
293 
294  Constant *XorC =
295  ConstantInt::get(I->getType(), NewMask & XorRHS->getValue());
296  Instruction *NewXor = BinaryOperator::CreateXor(NewAnd, XorC);
297  return InsertNewInstWith(NewXor, *I);
298  }
299 
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);
304  break;
305  }
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.
309  Value *LHS, *RHS;
310  if (matchSelectPattern(I, LHS, RHS).Flavor != SPF_UNKNOWN)
311  return nullptr;
312 
313  if (SimplifyDemandedBits(I, 2, DemandedMask, RHSKnown, Depth + 1) ||
314  SimplifyDemandedBits(I, 1, DemandedMask, LHSKnown, Depth + 1))
315  return I;
316  assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
317  assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
318 
319  // If the operands are constants, see if we can simplify them.
320  if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
321  ShrinkDemandedConstant(I, 2, DemandedMask))
322  return I;
323 
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;
327  break;
328  case Instruction::ZExt:
329  case Instruction::Trunc: {
330  unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
331 
332  APInt InputDemandedMask = DemandedMask.zextOrTrunc(SrcBitWidth);
333  KnownBits InputKnown(SrcBitWidth);
334  if (SimplifyDemandedBits(I, 0, InputDemandedMask, InputKnown, Depth + 1))
335  return I;
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?");
341  break;
342  }
343  case Instruction::BitCast:
344  if (!I->getOperand(0)->getType()->isIntOrIntVectorTy())
345  return nullptr; // vector->int or fp->int?
346 
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.
352  return nullptr;
353  } else
354  // Don't touch a scalar-to-vector bitcast.
355  return nullptr;
356  } else if (I->getOperand(0)->getType()->isVectorTy())
357  // Don't touch a vector-to-scalar bitcast.
358  return nullptr;
359 
360  if (SimplifyDemandedBits(I, 0, DemandedMask, Known, Depth + 1))
361  return I;
362  assert(!Known.hasConflict() && "Bits known to be one AND zero?");
363  break;
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();
367 
368  APInt InputDemandedBits = DemandedMask.trunc(SrcBitWidth);
369 
370  // If any of the sign extended bits are demanded, we know that the sign
371  // bit is demanded.
372  if (DemandedMask.getActiveBits() > SrcBitWidth)
373  InputDemandedBits.setBit(SrcBitWidth-1);
374 
375  KnownBits InputKnown(SrcBitWidth);
376  if (SimplifyDemandedBits(I, 0, InputDemandedBits, InputKnown, Depth + 1))
377  return I;
378 
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);
386  }
387 
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?");
392  break;
393  }
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  if (NLZ > 0) {
400  // Right fill the mask of bits for this ADD/SUB to demand the most
401  // significant bit and all those below it.
402  APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
403  if (ShrinkDemandedConstant(I, 0, DemandedFromOps) ||
404  SimplifyDemandedBits(I, 0, DemandedFromOps, LHSKnown, Depth + 1) ||
405  ShrinkDemandedConstant(I, 1, DemandedFromOps) ||
406  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);
413  return I;
414  }
415 
416  // If we are known to be adding/subtracting zeros to every bit below
417  // the highest demanded bit, we just return the other side.
418  if (DemandedFromOps.isSubsetOf(RHSKnown.Zero))
419  return I->getOperand(0);
420  // We can't do this with the LHS for subtraction, unless we are only
421  // demanding the LSB.
422  if ((I->getOpcode() == Instruction::Add ||
423  DemandedFromOps.isOneValue()) &&
424  DemandedFromOps.isSubsetOf(LHSKnown.Zero))
425  return I->getOperand(1);
426  }
427 
428  // Otherwise just hand the add/sub off to computeKnownBits to fill in
429  // the known zeros and ones.
430  computeKnownBits(V, Known, Depth, CxtI);
431  break;
432  }
433  case Instruction::Shl: {
434  const APInt *SA;
435  if (match(I->getOperand(1), m_APInt(SA))) {
436  const APInt *ShrAmt;
437  if (match(I->getOperand(0), m_Shr(m_Value(), m_APInt(ShrAmt)))) {
438  Instruction *Shr = cast<Instruction>(I->getOperand(0));
439  if (Value *R = simplifyShrShlDemandedBits(
440  Shr, *ShrAmt, I, *SA, DemandedMask, Known))
441  return R;
442  }
443 
444  uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
445  APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
446 
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);
453 
454  if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1))
455  return I;
456  assert(!Known.hasConflict() && "Bits known to be one AND zero?");
457  Known.Zero <<= ShiftAmt;
458  Known.One <<= ShiftAmt;
459  // low bits known zero.
460  if (ShiftAmt)
461  Known.Zero.setLowBits(ShiftAmt);
462  }
463  break;
464  }
465  case Instruction::LShr: {
466  const APInt *SA;
467  if (match(I->getOperand(1), m_APInt(SA))) {
468  uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
469 
470  // Unsigned shift right.
471  APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
472 
473  // If the shift is exact, then it does demand the low bits (and knows that
474  // they are zero).
475  if (cast<LShrOperator>(I)->isExact())
476  DemandedMaskIn.setLowBits(ShiftAmt);
477 
478  if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1))
479  return I;
480  assert(!Known.hasConflict() && "Bits known to be one AND zero?");
481  Known.Zero.lshrInPlace(ShiftAmt);
482  Known.One.lshrInPlace(ShiftAmt);
483  if (ShiftAmt)
484  Known.Zero.setHighBits(ShiftAmt); // high bits known zero.
485  }
486  break;
487  }
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);
498  }
499 
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);
504 
505  const APInt *SA;
506  if (match(I->getOperand(1), m_APInt(SA))) {
507  uint32_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
508 
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
512  // demanded.
513  if (DemandedMask.countLeadingZeros() <= ShiftAmt)
514  DemandedMaskIn.setSignBit();
515 
516  // If the shift is exact, then it does demand the low bits (and knows that
517  // they are zero).
518  if (cast<AShrOperator>(I)->isExact())
519  DemandedMaskIn.setLowBits(ShiftAmt);
520 
521  if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1))
522  return I;
523 
524  assert(!Known.hasConflict() && "Bits known to be one AND zero?");
525  // Compute the new bits that are at the top now.
526  APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
527  Known.Zero.lshrInPlace(ShiftAmt);
528  Known.One.lshrInPlace(ShiftAmt);
529 
530  // Handle the sign bits.
531  APInt SignMask(APInt::getSignMask(BitWidth));
532  // Adjust to where it is now in the mask.
533  SignMask.lshrInPlace(ShiftAmt);
534 
535  // If the input sign bit is known to be zero, or if none of the top bits
536  // are demanded, turn this into an unsigned shift right.
537  if (BitWidth <= ShiftAmt || Known.Zero[BitWidth-ShiftAmt-1] ||
538  !DemandedMask.intersects(HighBits)) {
539  BinaryOperator *LShr = BinaryOperator::CreateLShr(I->getOperand(0),
540  I->getOperand(1));
541  LShr->setIsExact(cast<BinaryOperator>(I)->isExact());
542  return InsertNewInstWith(LShr, *I);
543  } else if (Known.One.intersects(SignMask)) { // New bits are known one.
544  Known.One |= HighBits;
545  }
546  }
547  break;
548  }
549  case Instruction::SRem:
550  if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
551  // X % -1 demands all the bits because we don't want to introduce
552  // INT_MIN % -1 (== undef) by accident.
553  if (Rem->isMinusOne())
554  break;
555  APInt RA = Rem->getValue().abs();
556  if (RA.isPowerOf2()) {
557  if (DemandedMask.ult(RA)) // srem won't affect demanded bits
558  return I->getOperand(0);
559 
560  APInt LowBits = RA - 1;
561  APInt Mask2 = LowBits | APInt::getSignMask(BitWidth);
562  if (SimplifyDemandedBits(I, 0, Mask2, LHSKnown, Depth + 1))
563  return I;
564 
565  // The low bits of LHS are unchanged by the srem.
566  Known.Zero = LHSKnown.Zero & LowBits;
567  Known.One = LHSKnown.One & LowBits;
568 
569  // If LHS is non-negative or has all low bits zero, then the upper bits
570  // are all zero.
571  if (LHSKnown.isNonNegative() || LowBits.isSubsetOf(LHSKnown.Zero))
572  Known.Zero |= ~LowBits;
573 
574  // If LHS is negative and not all low bits are zero, then the upper bits
575  // are all one.
576  if (LHSKnown.isNegative() && LowBits.intersects(LHSKnown.One))
577  Known.One |= ~LowBits;
578 
579  assert(!Known.hasConflict() && "Bits known to be one AND zero?");
580  break;
581  }
582  }
583 
584  // The sign bit is the LHS's sign bit, except when the result of the
585  // remainder is zero.
586  if (DemandedMask.isSignBitSet()) {
587  computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1, CxtI);
588  // If it's known zero, our sign bit is also zero.
589  if (LHSKnown.isNonNegative())
590  Known.makeNonNegative();
591  }
592  break;
593  case Instruction::URem: {
594  KnownBits Known2(BitWidth);
595  APInt AllOnes = APInt::getAllOnesValue(BitWidth);
596  if (SimplifyDemandedBits(I, 0, AllOnes, Known2, Depth + 1) ||
597  SimplifyDemandedBits(I, 1, AllOnes, Known2, Depth + 1))
598  return I;
599 
600  unsigned Leaders = Known2.countMinLeadingZeros();
601  Known.Zero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
602  break;
603  }
604  case Instruction::Call:
605  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
606  switch (II->getIntrinsicID()) {
607  default: break;
608  case Intrinsic::bswap: {
609  // If the only bits demanded come from one byte of the bswap result,
610  // just shift the input byte into position to eliminate the bswap.
611  unsigned NLZ = DemandedMask.countLeadingZeros();
612  unsigned NTZ = DemandedMask.countTrailingZeros();
613 
614  // Round NTZ down to the next byte. If we have 11 trailing zeros, then
615  // we need all the bits down to bit 8. Likewise, round NLZ. If we
616  // have 14 leading zeros, round to 8.
617  NLZ &= ~7;
618  NTZ &= ~7;
619  // If we need exactly one byte, we can do this transformation.
620  if (BitWidth-NLZ-NTZ == 8) {
621  unsigned ResultBit = NTZ;
622  unsigned InputBit = BitWidth-NTZ-8;
623 
624  // Replace this with either a left or right shift to get the byte into
625  // the right place.
626  Instruction *NewVal;
627  if (InputBit > ResultBit)
628  NewVal = BinaryOperator::CreateLShr(II->getArgOperand(0),
629  ConstantInt::get(I->getType(), InputBit-ResultBit));
630  else
631  NewVal = BinaryOperator::CreateShl(II->getArgOperand(0),
632  ConstantInt::get(I->getType(), ResultBit-InputBit));
633  NewVal->takeName(I);
634  return InsertNewInstWith(NewVal, *I);
635  }
636 
637  // TODO: Could compute known zero/one bits based on the input.
638  break;
639  }
640  case Intrinsic::x86_mmx_pmovmskb:
641  case Intrinsic::x86_sse_movmsk_ps:
642  case Intrinsic::x86_sse2_movmsk_pd:
643  case Intrinsic::x86_sse2_pmovmskb_128:
644  case Intrinsic::x86_avx_movmsk_ps_256:
645  case Intrinsic::x86_avx_movmsk_pd_256:
646  case Intrinsic::x86_avx2_pmovmskb: {
647  // MOVMSK copies the vector elements' sign bits to the low bits
648  // and zeros the high bits.
649  unsigned ArgWidth;
650  if (II->getIntrinsicID() == Intrinsic::x86_mmx_pmovmskb) {
651  ArgWidth = 8; // Arg is x86_mmx, but treated as <8 x i8>.
652  } else {
653  auto Arg = II->getArgOperand(0);
654  auto ArgType = cast<VectorType>(Arg->getType());
655  ArgWidth = ArgType->getNumElements();
656  }
657 
658  // If we don't need any of low bits then return zero,
659  // we know that DemandedMask is non-zero already.
660  APInt DemandedElts = DemandedMask.zextOrTrunc(ArgWidth);
661  if (DemandedElts.isNullValue())
662  return ConstantInt::getNullValue(VTy);
663 
664  // We know that the upper bits are set to zero.
665  Known.Zero.setBitsFrom(ArgWidth);
666  return nullptr;
667  }
668  case Intrinsic::x86_sse42_crc32_64_64:
669  Known.Zero.setBitsFrom(32);
670  return nullptr;
671  }
672  }
673  computeKnownBits(V, Known, Depth, CxtI);
674  break;
675  }
676 
677  // If the client is only demanding bits that we know, return the known
678  // constant.
679  if (DemandedMask.isSubsetOf(Known.Zero|Known.One))
680  return Constant::getIntegerValue(VTy, Known.One);
681  return nullptr;
682 }
683 
684 /// Helper routine of SimplifyDemandedUseBits. It computes Known
685 /// bits. It also tries to handle simplifications that can be done based on
686 /// DemandedMask, but without modifying the Instruction.
687 Value *InstCombiner::SimplifyMultipleUseDemandedBits(Instruction *I,
688  const APInt &DemandedMask,
689  KnownBits &Known,
690  unsigned Depth,
691  Instruction *CxtI) {
692  unsigned BitWidth = DemandedMask.getBitWidth();
693  Type *ITy = I->getType();
694 
695  KnownBits LHSKnown(BitWidth);
696  KnownBits RHSKnown(BitWidth);
697 
698  // Despite the fact that we can't simplify this instruction in all User's
699  // context, we can at least compute the known bits, and we can
700  // do simplifications that apply to *just* the one user if we know that
701  // this instruction has a simpler value in that context.
702  switch (I->getOpcode()) {
703  case Instruction::And: {
704  // If either the LHS or the RHS are Zero, the result is zero.
705  computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
706  computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1,
707  CxtI);
708 
709  // Output known-0 are known to be clear if zero in either the LHS | RHS.
710  APInt IKnownZero = RHSKnown.Zero | LHSKnown.Zero;
711  // Output known-1 bits are only known if set in both the LHS & RHS.
712  APInt IKnownOne = RHSKnown.One & LHSKnown.One;
713 
714  // If the client is only demanding bits that we know, return the known
715  // constant.
716  if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
717  return Constant::getIntegerValue(ITy, IKnownOne);
718 
719  // If all of the demanded bits are known 1 on one side, return the other.
720  // These bits cannot contribute to the result of the 'and' in this
721  // context.
722  if (DemandedMask.isSubsetOf(LHSKnown.Zero | RHSKnown.One))
723  return I->getOperand(0);
724  if (DemandedMask.isSubsetOf(RHSKnown.Zero | LHSKnown.One))
725  return I->getOperand(1);
726 
727  Known.Zero = std::move(IKnownZero);
728  Known.One = std::move(IKnownOne);
729  break;
730  }
731  case Instruction::Or: {
732  // We can simplify (X|Y) -> X or Y in the user's context if we know that
733  // only bits from X or Y are demanded.
734 
735  // If either the LHS or the RHS are One, the result is One.
736  computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
737  computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1,
738  CxtI);
739 
740  // Output known-0 bits are only known if clear in both the LHS & RHS.
741  APInt IKnownZero = RHSKnown.Zero & LHSKnown.Zero;
742  // Output known-1 are known to be set if set in either the LHS | RHS.
743  APInt IKnownOne = RHSKnown.One | LHSKnown.One;
744 
745  // If the client is only demanding bits that we know, return the known
746  // constant.
747  if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
748  return Constant::getIntegerValue(ITy, IKnownOne);
749 
750  // If all of the demanded bits are known zero on one side, return the
751  // other. These bits cannot contribute to the result of the 'or' in this
752  // context.
753  if (DemandedMask.isSubsetOf(LHSKnown.One | RHSKnown.Zero))
754  return I->getOperand(0);
755  if (DemandedMask.isSubsetOf(RHSKnown.One | LHSKnown.Zero))
756  return I->getOperand(1);
757 
758  Known.Zero = std::move(IKnownZero);
759  Known.One = std::move(IKnownOne);
760  break;
761  }
762  case Instruction::Xor: {
763  // We can simplify (X^Y) -> X or Y in the user's context if we know that
764  // only bits from X or Y are demanded.
765 
766  computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
767  computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1,
768  CxtI);
769 
770  // Output known-0 bits are known if clear or set in both the LHS & RHS.
771  APInt IKnownZero = (RHSKnown.Zero & LHSKnown.Zero) |
772  (RHSKnown.One & LHSKnown.One);
773  // Output known-1 are known to be set if set in only one of the LHS, RHS.
774  APInt IKnownOne = (RHSKnown.Zero & LHSKnown.One) |
775  (RHSKnown.One & LHSKnown.Zero);
776 
777  // If the client is only demanding bits that we know, return the known
778  // constant.
779  if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
780  return Constant::getIntegerValue(ITy, IKnownOne);
781 
782  // If all of the demanded bits are known zero on one side, return the
783  // other.
784  if (DemandedMask.isSubsetOf(RHSKnown.Zero))
785  return I->getOperand(0);
786  if (DemandedMask.isSubsetOf(LHSKnown.Zero))
787  return I->getOperand(1);
788 
789  // Output known-0 bits are known if clear or set in both the LHS & RHS.
790  Known.Zero = std::move(IKnownZero);
791  // Output known-1 are known to be set if set in only one of the LHS, RHS.
792  Known.One = std::move(IKnownOne);
793  break;
794  }
795  default:
796  // Compute the Known bits to simplify things downstream.
797  computeKnownBits(I, Known, Depth, CxtI);
798 
799  // If this user is only demanding bits that we know, return the known
800  // constant.
801  if (DemandedMask.isSubsetOf(Known.Zero|Known.One))
802  return Constant::getIntegerValue(ITy, Known.One);
803 
804  break;
805  }
806 
807  return nullptr;
808 }
809 
810 
811 /// Helper routine of SimplifyDemandedUseBits. It tries to simplify
812 /// "E1 = (X lsr C1) << C2", where the C1 and C2 are constant, into
813 /// "E2 = X << (C2 - C1)" or "E2 = X >> (C1 - C2)", depending on the sign
814 /// of "C2-C1".
815 ///
816 /// Suppose E1 and E2 are generally different in bits S={bm, bm+1,
817 /// ..., bn}, without considering the specific value X is holding.
818 /// This transformation is legal iff one of following conditions is hold:
819 /// 1) All the bit in S are 0, in this case E1 == E2.
820 /// 2) We don't care those bits in S, per the input DemandedMask.
821 /// 3) Combination of 1) and 2). Some bits in S are 0, and we don't care the
822 /// rest bits.
823 ///
824 /// Currently we only test condition 2).
825 ///
826 /// As with SimplifyDemandedUseBits, it returns NULL if the simplification was
827 /// not successful.
828 Value *
829 InstCombiner::simplifyShrShlDemandedBits(Instruction *Shr, const APInt &ShrOp1,
830  Instruction *Shl, const APInt &ShlOp1,
831  const APInt &DemandedMask,
832  KnownBits &Known) {
833  if (!ShlOp1 || !ShrOp1)
834  return nullptr; // No-op.
835 
836  Value *VarX = Shr->getOperand(0);
837  Type *Ty = VarX->getType();
838  unsigned BitWidth = Ty->getScalarSizeInBits();
839  if (ShlOp1.uge(BitWidth) || ShrOp1.uge(BitWidth))
840  return nullptr; // Undef.
841 
842  unsigned ShlAmt = ShlOp1.getZExtValue();
843  unsigned ShrAmt = ShrOp1.getZExtValue();
844 
845  Known.One.clearAllBits();
846  Known.Zero.setLowBits(ShlAmt - 1);
847  Known.Zero &= DemandedMask;
848 
849  APInt BitMask1(APInt::getAllOnesValue(BitWidth));
850  APInt BitMask2(APInt::getAllOnesValue(BitWidth));
851 
852  bool isLshr = (Shr->getOpcode() == Instruction::LShr);
853  BitMask1 = isLshr ? (BitMask1.lshr(ShrAmt) << ShlAmt) :
854  (BitMask1.ashr(ShrAmt) << ShlAmt);
855 
856  if (ShrAmt <= ShlAmt) {
857  BitMask2 <<= (ShlAmt - ShrAmt);
858  } else {
859  BitMask2 = isLshr ? BitMask2.lshr(ShrAmt - ShlAmt):
860  BitMask2.ashr(ShrAmt - ShlAmt);
861  }
862 
863  // Check if condition-2 (see the comment to this function) is satified.
864  if ((BitMask1 & DemandedMask) == (BitMask2 & DemandedMask)) {
865  if (ShrAmt == ShlAmt)
866  return VarX;
867 
868  if (!Shr->hasOneUse())
869  return nullptr;
870 
871  BinaryOperator *New;
872  if (ShrAmt < ShlAmt) {
873  Constant *Amt = ConstantInt::get(VarX->getType(), ShlAmt - ShrAmt);
874  New = BinaryOperator::CreateShl(VarX, Amt);
875  BinaryOperator *Orig = cast<BinaryOperator>(Shl);
876  New->setHasNoSignedWrap(Orig->hasNoSignedWrap());
877  New->setHasNoUnsignedWrap(Orig->hasNoUnsignedWrap());
878  } else {
879  Constant *Amt = ConstantInt::get(VarX->getType(), ShrAmt - ShlAmt);
880  New = isLshr ? BinaryOperator::CreateLShr(VarX, Amt) :
881  BinaryOperator::CreateAShr(VarX, Amt);
882  if (cast<BinaryOperator>(Shr)->isExact())
883  New->setIsExact(true);
884  }
885 
886  return InsertNewInstWith(New, *Shl);
887  }
888 
889  return nullptr;
890 }
891 
892 /// The specified value produces a vector with any number of elements.
893 /// DemandedElts contains the set of elements that are actually used by the
894 /// caller. This method analyzes which elements of the operand are undef and
895 /// returns that information in UndefElts.
896 ///
897 /// If the information about demanded elements can be used to simplify the
898 /// operation, the operation is simplified, then the resultant value is
899 /// returned. This returns null if no change was made.
900 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
901  APInt &UndefElts,
902  unsigned Depth) {
903  unsigned VWidth = V->getType()->getVectorNumElements();
904  APInt EltMask(APInt::getAllOnesValue(VWidth));
905  assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
906 
907  if (isa<UndefValue>(V)) {
908  // If the entire vector is undefined, just return this info.
909  UndefElts = EltMask;
910  return nullptr;
911  }
912 
913  if (DemandedElts.isNullValue()) { // If nothing is demanded, provide undef.
914  UndefElts = EltMask;
915  return UndefValue::get(V->getType());
916  }
917 
918  UndefElts = 0;
919 
920  // Handle ConstantAggregateZero, ConstantVector, ConstantDataSequential.
921  if (Constant *C = dyn_cast<Constant>(V)) {
922  // Check if this is identity. If so, return 0 since we are not simplifying
923  // anything.
924  if (DemandedElts.isAllOnesValue())
925  return nullptr;
926 
927  Type *EltTy = cast<VectorType>(V->getType())->getElementType();
928  Constant *Undef = UndefValue::get(EltTy);
929 
931  for (unsigned i = 0; i != VWidth; ++i) {
932  if (!DemandedElts[i]) { // If not demanded, set to undef.
933  Elts.push_back(Undef);
934  UndefElts.setBit(i);
935  continue;
936  }
937 
938  Constant *Elt = C->getAggregateElement(i);
939  if (!Elt) return nullptr;
940 
941  if (isa<UndefValue>(Elt)) { // Already undef.
942  Elts.push_back(Undef);
943  UndefElts.setBit(i);
944  } else { // Otherwise, defined.
945  Elts.push_back(Elt);
946  }
947  }
948 
949  // If we changed the constant, return it.
950  Constant *NewCV = ConstantVector::get(Elts);
951  return NewCV != C ? NewCV : nullptr;
952  }
953 
954  // Limit search depth.
955  if (Depth == 10)
956  return nullptr;
957 
958  // If multiple users are using the root value, proceed with
959  // simplification conservatively assuming that all elements
960  // are needed.
961  if (!V->hasOneUse()) {
962  // Quit if we find multiple users of a non-root value though.
963  // They'll be handled when it's their turn to be visited by
964  // the main instcombine process.
965  if (Depth != 0)
966  // TODO: Just compute the UndefElts information recursively.
967  return nullptr;
968 
969  // Conservatively assume that all elements are needed.
970  DemandedElts = EltMask;
971  }
972 
974  if (!I) return nullptr; // Only analyze instructions.
975 
976  bool MadeChange = false;
977  APInt UndefElts2(VWidth, 0);
978  APInt UndefElts3(VWidth, 0);
979  Value *TmpV;
980  switch (I->getOpcode()) {
981  default: break;
982 
983  case Instruction::InsertElement: {
984  // If this is a variable index, we don't know which element it overwrites.
985  // demand exactly the same input as we produce.
987  if (!Idx) {
988  // Note that we can't propagate undef elt info, because we don't know
989  // which elt is getting updated.
990  TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
991  UndefElts2, Depth + 1);
992  if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
993  break;
994  }
995 
996  // If this is inserting an element that isn't demanded, remove this
997  // insertelement.
998  unsigned IdxNo = Idx->getZExtValue();
999  if (IdxNo >= VWidth || !DemandedElts[IdxNo]) {
1000  Worklist.Add(I);
1001  return I->getOperand(0);
1002  }
1003 
1004  // Otherwise, the element inserted overwrites whatever was there, so the
1005  // input demanded set is simpler than the output set.
1006  APInt DemandedElts2 = DemandedElts;
1007  DemandedElts2.clearBit(IdxNo);
1008  TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1009  UndefElts, Depth + 1);
1010  if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1011 
1012  // The inserted element is defined.
1013  UndefElts.clearBit(IdxNo);
1014  break;
1015  }
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);
1029  else
1030  RightDemanded.setBit(MaskVal - LHSVWidth);
1031  }
1032  }
1033  }
1034 
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; }
1039 
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; }
1044 
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);
1061  } else {
1062  LHSIdx = LHSIdx == -1u ? i : LHSVWidth;
1063  LHSValIdx = LHSValIdx == -1u ? MaskVal : LHSVWidth;
1064  LHSUniform = LHSUniform && (MaskVal == i);
1065  }
1066  } else {
1067  if (RHSUndefElts[MaskVal - LHSVWidth]) {
1068  NewUndefElts = true;
1069  UndefElts.setBit(i);
1070  } else {
1071  RHSIdx = RHSIdx == -1u ? i : LHSVWidth;
1072  RHSValIdx = RHSValIdx == -1u ? MaskVal - LHSVWidth : LHSVWidth;
1073  RHSUniform = RHSUniform && (MaskVal - LHSVWidth == i);
1074  }
1075  }
1076  }
1077 
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;
1085  unsigned Idx = -1u;
1086 
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);
1092  Idx = LHSIdx;
1093  }
1094  }
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);
1099  Idx = RHSIdx;
1100  }
1101  }
1102  // Found constant vector with single element - convert to insertelement.
1103  if (Op && Value) {
1105  Op, Value, ConstantInt::get(Type::getInt32Ty(I->getContext()), Idx),
1106  Shuffle->getName());
1107  InsertNewInstWith(New, *Shuffle);
1108  return New;
1109  }
1110  }
1111  if (NewUndefElts) {
1112  // Add additional discovered undefs.
1114  for (unsigned i = 0; i < VWidth; ++i) {
1115  if (UndefElts[i])
1117  else
1119  Shuffle->getMaskValue(i)));
1120  }
1121  I->setOperand(2, ConstantVector::get(Elts));
1122  MadeChange = true;
1123  }
1124  break;
1125  }
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))
1135  continue;
1136  if (CElt->isNullValue())
1137  LeftDemanded.clearBit(i);
1138  else
1139  RightDemanded.clearBit(i);
1140  }
1141  }
1142 
1143  TmpV = SimplifyDemandedVectorElts(I->getOperand(1), LeftDemanded, UndefElts,
1144  Depth + 1);
1145  if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1146 
1147  TmpV = SimplifyDemandedVectorElts(I->getOperand(2), RightDemanded,
1148  UndefElts2, Depth + 1);
1149  if (TmpV) { I->setOperand(2, TmpV); MadeChange = true; }
1150 
1151  // Output elements are undefined if both are undefined.
1152  UndefElts &= UndefElts2;
1153  break;
1154  }
1155  case Instruction::BitCast: {
1156  // Vector->vector casts only.
1157  VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1158  if (!VTy) break;
1159  unsigned InVWidth = VTy->getNumElements();
1160  APInt InputDemandedElts(InVWidth, 0);
1161  UndefElts2 = APInt(InVWidth, 0);
1162  unsigned Ratio;
1163 
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.
1167  Ratio = 1;
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);
1185  } else {
1186  // Unsupported so far.
1187  break;
1188  }
1189 
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);
1193  if (TmpV) {
1194  I->setOperand(0, TmpV);
1195  MadeChange = true;
1196  }
1197 
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);
1215  }
1216  } else {
1217  llvm_unreachable("Unimp");
1218  }
1219  break;
1220  }
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,
1229  Depth + 1);
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; }
1234 
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;
1238  break;
1239  case Instruction::FPTrunc:
1240  case Instruction::FPExt:
1241  TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts, UndefElts,
1242  Depth + 1);
1243  if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1244  break;
1245 
1246  case Instruction::Call: {
1248  if (!II) break;
1249  switch (II->getIntrinsicID()) {
1250  default: break;
1251 
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]) {
1259  Worklist.Add(II);
1260  return ConstantAggregateZero::get(II->getType());
1261  }
1262 
1263  // Only the lower element is used.
1264  DemandedElts = 1;
1265  TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts,
1266  UndefElts, Depth + 1);
1267  if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1268 
1269  // Only the lower element is undefined. The high elements are zero.
1270  UndefElts = UndefElts[0];
1271  break;
1272 
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; }
1281 
1282  // If lowest element of a scalar op isn't used then use Arg0.
1283  if (!DemandedElts[0]) {
1284  Worklist.Add(II);
1285  return II->getArgOperand(0);
1286  }
1287  // TODO: If only low elt lower SQRT to FSQRT (with rounding/exceptions
1288  // checks).
1289  break;
1290 
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
1293  // operands.
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; }
1303 
1304  // If lowest element of a scalar op isn't used then use Arg0.
1305  if (!DemandedElts[0]) {
1306  Worklist.Add(II);
1307  return II->getArgOperand(0);
1308  }
1309 
1310  // Only lower element is used for operand 1.
1311  DemandedElts = 1;
1312  TmpV = SimplifyDemandedVectorElts(II->getArgOperand(1), DemandedElts,
1313  UndefElts2, Depth + 1);
1314  if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1315 
1316  // Lower element is undefined if both lower elements are undefined.
1317  // Consider things like undef&0. The result is known zero, not undef.
1318  if (!UndefElts2[0])
1319  UndefElts.clearBit(0);
1320 
1321  break;
1322  }
1323 
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; }
1334 
1335  // If lowest element of a scalar op isn't used then use Arg0.
1336  if (!DemandedElts[0]) {
1337  Worklist.Add(II);
1338  return II->getArgOperand(0);
1339  }
1340 
1341  // Only lower element is used for operand 1.
1342  DemandedElts = 1;
1343  TmpV = SimplifyDemandedVectorElts(II->getArgOperand(1), DemandedElts,
1344  UndefElts2, Depth + 1);
1345  if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1346 
1347  // Take the high undef elements from operand 0 and take the lower element
1348  // from operand 1.
1349  UndefElts.clearBit(0);
1350  UndefElts |= UndefElts2[0];
1351  break;
1352  }
1353 
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
1356  // three inputs.
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; }
1384 
1385  // If lowest element of a scalar op isn't used then use Arg0.
1386  if (!DemandedElts[0]) {
1387  Worklist.Add(II);
1388  return II->getArgOperand(0);
1389  }
1390 
1391  // Only lower element is used for operand 1 and 2.
1392  DemandedElts = 1;
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; }
1399 
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);
1404 
1405  break;
1406 
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; }
1417 
1418  // If lowest element of a scalar op isn't used then use Arg2.
1419  if (!DemandedElts[0]) {
1420  Worklist.Add(II);
1421  return II->getArgOperand(2);
1422  }
1423 
1424  // Only lower element is used for operand 0 and 1.
1425  DemandedElts = 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; }
1432 
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);
1437 
1438  break;
1439 
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");
1450 
1451  APInt InnerDemandedElts(InnerVWidth, 0);
1452  for (unsigned i = 0; i != VWidth; ++i)
1453  if (DemandedElts[i])
1454  InnerDemandedElts.setBit(i * 2);
1455 
1456  UndefElts2 = APInt(InnerVWidth, 0);
1457  TmpV = SimplifyDemandedVectorElts(Op0, InnerDemandedElts, UndefElts2,
1458  Depth + 1);
1459  if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1460 
1461  UndefElts3 = APInt(InnerVWidth, 0);
1462  TmpV = SimplifyDemandedVectorElts(Op1, InnerDemandedElts, UndefElts3,
1463  Depth + 1);
1464  if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1465 
1466  break;
1467  }
1468 
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");
1484 
1485  unsigned NumLanes = Ty0->getPrimitiveSizeInBits() / 128;
1486  unsigned VWidthPerLane = VWidth / NumLanes;
1487  unsigned InnerVWidthPerLane = InnerVWidth / NumLanes;
1488 
1489  // Per lane, pack the elements of the first input and then the second.
1490  // e.g.
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);
1501  }
1502  }
1503 
1504  // Demand elements from the operand.
1505  auto *Op = II->getArgOperand(OpNum);
1506  APInt OpUndefElts(InnerVWidth, 0);
1507  TmpV = SimplifyDemandedVectorElts(Op, OpDemandedElts, OpUndefElts,
1508  Depth + 1);
1509  if (TmpV) {
1510  II->setArgOperand(OpNum, TmpV);
1511  MadeChange = true;
1512  }
1513 
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;
1521  }
1522  }
1523  break;
1524  }
1525 
1526  // PSHUFB
1527  case Intrinsic::x86_ssse3_pshuf_b_128:
1528  case Intrinsic::x86_avx2_pshuf_b:
1529  case Intrinsic::x86_avx512_pshuf_b_512:
1530  // PERMILVAR
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:
1537  // PERMV
1538  case Intrinsic::x86_avx2_permd:
1539  case Intrinsic::x86_avx2_permps: {
1540  Value *Op1 = II->getArgOperand(1);
1541  TmpV = SimplifyDemandedVectorElts(Op1, DemandedElts, UndefElts,
1542  Depth + 1);
1543  if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1544  break;
1545  }
1546 
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);
1554  break;
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:
1567 
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:
1578 
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:
1589 
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:
1600 
1601  case Intrinsic::amdgcn_image_getlod: {
1602  if (VWidth == 1 || !DemandedElts.isMask())
1603  return nullptr;
1604 
1605  // TODO: Handle 3 vectors when supported in code gen.
1606  unsigned NewNumElts = PowerOf2Ceil(DemandedElts.countTrailingOnes());
1607  if (NewNumElts == VWidth)
1608  return nullptr;
1609 
1610  Module *M = II->getParent()->getParent()->getParent();
1611  Type *EltTy = V->getType()->getVectorElementType();
1612 
1613  Type *NewTy = (NewNumElts == 1) ? EltTy :
1614  VectorType::get(EltTy, NewNumElts);
1615 
1616  auto IID = II->getIntrinsicID();
1617 
1618  bool IsBuffer = IID == Intrinsic::amdgcn_buffer_load ||
1619  IID == Intrinsic::amdgcn_buffer_load_format;
1620 
1621  Function *NewIntrin = IsBuffer ?
1622  Intrinsic::getDeclaration(M, IID, NewTy) :
1623  // Samplers have 3 mangled types.
1625  { NewTy, II->getArgOperand(0)->getType(),
1626  II->getArgOperand(1)->getType()});
1627 
1629  for (unsigned I = 0, E = II->getNumArgOperands(); I != E; ++I)
1630  Args.push_back(II->getArgOperand(I));
1631 
1632  IRBuilderBase::InsertPointGuard Guard(Builder);
1633  Builder.SetInsertPoint(II);
1634 
1635  CallInst *NewCall = Builder.CreateCall(NewIntrin, Args);
1636  NewCall->takeName(II);
1637  NewCall->copyMetadata(*II);
1638 
1639  if (!IsBuffer) {
1640  ConstantInt *DMask = dyn_cast<ConstantInt>(NewCall->getArgOperand(3));
1641  if (DMask) {
1642  unsigned DMaskVal = DMask->getZExtValue() & 0xf;
1643 
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)
1650  break;
1651 
1652  NewDMask |= Bit;
1653  }
1654  }
1655 
1656  NewCall->setArgOperand(3, ConstantInt::get(DMask->getType(), NewDMask));
1657  }
1658  }
1659 
1660 
1661  if (NewNumElts == 1) {
1662  return Builder.CreateInsertElement(UndefValue::get(V->getType()),
1663  NewCall, static_cast<uint64_t>(0));
1664  }
1665 
1666  SmallVector<uint32_t, 8> EltMask;
1667  for (unsigned I = 0; I < VWidth; ++I)
1668  EltMask.push_back(I);
1669 
1670  Value *Shuffle = Builder.CreateShuffleVector(
1671  NewCall, UndefValue::get(NewTy), EltMask);
1672 
1673  MadeChange = true;
1674  return Shuffle;
1675  }
1676  }
1677  break;
1678  }
1679  }
1680  return MadeChange ? I : nullptr;
1681 }
void clearAllBits()
Set every bit to 0.
Definition: APInt.h:1431
APInt abs() const
Get the absolute value;.
Definition: APInt.h:1779
Type * getVectorElementType() const
Definition: Type.h:368
uint64_t CallInst * C
BinOpPred_match< LHS, RHS, is_right_shift_op > m_Shr(const LHS &L, const RHS &R)
Matches logical shift operations.
Definition: PatternMatch.h:734
IntegerType * getType() const
getType - Specialize the getType() method to always return an IntegerType, which reduces the amount o...
Definition: Constants.h:172
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:72
void setSignBit()
Set the sign bit to 1.
Definition: APInt.h:1392
bool isSignMask() const
Check if the APInt&#39;s value is returned by getSignMask.
Definition: APInt.h:466
uint64_t getZExtValue() const
Get zero extended value.
Definition: APInt.h:1542
static APInt getAllOnesValue(unsigned numBits)
Get the all-ones value.
Definition: APInt.h:555
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
A Module instance is used to store all the information related to an LLVM module. ...
Definition: Module.h:63
bool hasConflict() const
Returns true if there is conflicting information.
Definition: KnownBits.h:47
This class represents zero extension of integer types.
static ConstantAggregateZero * get(Type *Ty)
Definition: Constants.cpp:1237
APInt zext(unsigned width) const
Zero extend to a new width.
Definition: APInt.cpp:865
This class represents a function call, abstracting a target machine&#39;s calling convention.
static APInt getLowBitsSet(unsigned numBits, unsigned loBitsSet)
Get a value with low bits set.
Definition: APInt.h:641
bool isSubsetOf(const APInt &RHS) const
This operation checks that all bits set in this APInt are also set in RHS.
Definition: APInt.h:1308
This instruction constructs a fixed permutation of two input vectors.
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:697
const Use & getOperandUse(unsigned i) const
Definition: User.h:167
APInt trunc(unsigned width) const
Truncate to new width.
Definition: APInt.cpp:818
void setAllBits()
Set every bit to 1.
Definition: APInt.h:1369
APInt zextOrTrunc(unsigned width) const
Zero extend or truncate to width.
Definition: APInt.cpp:883
void setBitsFrom(unsigned loBit)
Set the top bits starting from loBit.
Definition: APInt.h:1416
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:227
unsigned getBitWidth() const
Get the bit width of this value.
Definition: KnownBits.h:40
bool hasNoSignedWrap() const
Determine whether the no signed wrap flag is set.
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1488
Value * get() const
Definition: Use.h:108
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:207
unsigned countTrailingZeros() const
Count the number of trailing zero bits.
Definition: APInt.h:1611
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
static InsertElementInst * Create(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
APInt getLoBits(unsigned numBits) const
Compute an APInt containing numBits lowbits from this APInt.
Definition: APInt.cpp:515
void setBit(unsigned BitPosition)
Set a given bit to 1.
Definition: APInt.h:1382
void setHighBits(unsigned hiBits)
Set the top hiBits bits.
Definition: APInt.h:1426
unsigned getNumArgOperands() const
Return the number of call arguments.
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:560
APInt shl(unsigned shiftAmt) const
Left-shift function.
Definition: APInt.h:981
A Use represents the edge between a Value definition and its users.
Definition: Use.h:56
uint64_t getNumElements() const
Definition: DerivedTypes.h:359
void lshrInPlace(unsigned ShiftAmt)
Logical right-shift this APInt by ShiftAmt in place.
Definition: APInt.h:966
unsigned getActiveBits() const
Compute the number of active bits in the value.
Definition: APInt.h:1512
bool isNullValue() const
Return true if this is the value that would be returned by getNullValue.
Definition: Constants.cpp:86
void setIsExact(bool b=true)
Set or clear the exact flag on this instruction, which must be an operator which supports this flag...
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
void clearBit(unsigned BitPosition)
Set a given bit to 0.
Definition: APInt.h:1441
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:138
KnownBits zextOrTrunc(unsigned BitWidth)
Zero extends or truncates the underlying known Zero and One bits.
Definition: KnownBits.h:133
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:121
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:203
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:290
Function * getDeclaration(Module *M, ID id, ArrayRef< Type *> Tys=None)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
Definition: Function.cpp:975
Value * getOperand(unsigned i) const
Definition: User.h:154
SelectPatternResult matchSelectPattern(Value *V, Value *&LHS, Value *&RHS, Instruction::CastOps *CastOp=nullptr)
Pattern match integer [SU]MIN, [SU]MAX and ABS idioms, returning the kind and providing the out param...
Constant * getAggregateElement(unsigned Elt) const
For aggregates (struct/array/vector) return the constant that corresponds to the specified element if...
Definition: Constants.cpp:277
static APInt getHighBitsSet(unsigned numBits, unsigned hiBitsSet)
Get a value with high bits set.
Definition: APInt.h:629
bool isAllOnesValue() const
Determine if all bits are set.
Definition: APInt.h:389
uint64_t getZExtValue() const
Return the constant as a 64-bit unsigned integer value after it has been zero extended as appropriate...
Definition: Constants.h:149
* if(!EatIfPresent(lltok::kw_thread_local)) return false
ParseOptionalThreadLocal := /*empty.
apint_match m_APInt(const APInt *&Res)
Match a ConstantInt or splatted ConstantVector, binding the specified pointer to the contained APInt...
Definition: PatternMatch.h:240
unsigned countPopulation() const
Count the number of bits set.
Definition: APInt.h:1637
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
bool ult(const APInt &RHS) const
Unsigned less than comparison.
Definition: APInt.h:1164
This is an important base class in LLVM.
Definition: Constant.h:42
void resetAll()
Resets the known state of all bits.
Definition: KnownBits.h:66
constexpr char Args[]
Key for Kernel::Metadata::mArgs.
bool isOneValue() const
Determine if this is a value of 1.
Definition: APInt.h:404
Constant Vector Declarations.
Definition: Constants.h:491
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1320
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
APInt lshr(unsigned shiftAmt) const
Logical right-shift function.
Definition: APInt.h:959
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:244
Intrinsic::ID getIntrinsicID() const
Return the intrinsic ID of this intrinsic.
Definition: IntrinsicInst.h:51
void makeNonNegative()
Make this value negative.
Definition: KnownBits.h:108
static APInt getSignMask(unsigned BitWidth)
Get the SignMask for a specific bit width.
Definition: APInt.h:548
APInt ashr(unsigned ShiftAmt) const
Arithmetic right-shift function.
Definition: APInt.h:935
void setHasNoSignedWrap(bool b=true)
Set or clear the nsw flag on this instruction, which must be an operator which supports this flag...
#define E
Definition: LargeTest.cpp:27
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
unsigned getScalarSizeInBits() const LLVM_READONLY
If this is a vector type, return the getPrimitiveSizeInBits value for the element type...
Definition: Type.cpp:130
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:864
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:560
bool uge(const APInt &RHS) const
Unsigned greater or equal comparison.
Definition: APInt.h:1272
void setOperand(unsigned i, Value *Val)
Definition: User.h:159
unsigned getVectorNumElements() const
Definition: DerivedTypes.h:462
Class to represent vector types.
Definition: DerivedTypes.h:393
Class for arbitrary precision integers.
Definition: APInt.h:69
bool isPowerOf2() const
Check if this APInt&#39;s value is a power of two greater than zero.
Definition: APInt.h:457
static int getMaskValue(Constant *Mask, unsigned Elt)
Return the shuffle mask value for the specified element of the mask.
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:1300
static IntegerType * getInt32Ty(LLVMContext &C)
Definition: Type.cpp:176
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:475
Value * getArgOperand(unsigned i) const
getArgOperand/setArgOperand - Return/set the i-th call argument.
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:218
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:108
#define I(x, y, z)
Definition: MD5.cpp:58
void setArgOperand(unsigned i, Value *v)
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
bool hasNoUnsignedWrap() const
Determine whether the no unsigned wrap flag is set.
void setHasNoUnsignedWrap(bool b=true)
Set or clear the nsw flag on this instruction, which must be an operator which supports this flag...
Definition: Instruction.cpp:98
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:545
LLVM Value Representation.
Definition: Value.h:73
This file provides internal interfaces used to implement the InstCombine.
void copyMetadata(const Instruction &SrcInst, ArrayRef< unsigned > WL=ArrayRef< unsigned >())
Copy metadata from SrcInst to this instruction.
static VectorType * get(Type *ElementType, unsigned NumElements)
This static method is the primary way to construct an VectorType.
Definition: Type.cpp:593
KnownBits sext(unsigned BitWidth)
Sign extends the underlying known Zero and One bits.
Definition: KnownBits.h:127
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.
bool isSignBitSet() const
Determine if sign bit of this APInt is set.
Definition: APInt.h:369
unsigned countMinLeadingZeros() const
Returns the minimum number of leading zero bits.
Definition: KnownBits.h:148
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:408
bool isNonNegative() const
Returns true if this value is known to be non-negative.
Definition: KnownBits.h:99
unsigned countLeadingZeros() const
The APInt version of the countLeadingZeros functions in MathExtras.h.
Definition: APInt.h:1575
VectorType * getType() const
Overload to return most specific vector type.
void computeKnownBits(const Value *V, KnownBits &Known, const DataLayout &DL, unsigned Depth=0, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, OptimizationRemarkEmitter *ORE=nullptr)
Determine which bits of V are known to be either zero or one and return them in the KnownZero/KnownOn...
static Constant * get(ArrayRef< Constant *> V)
Definition: Constants.cpp:984
void setLowBits(unsigned loBits)
Set the bottom loBits bits.
Definition: APInt.h:1421
bool isNullValue() const
Determine if all bits are clear.
Definition: APInt.h:399
uint64_t PowerOf2Ceil(uint64_t A)
Returns the power of two which is greater than or equal to the given value.
Definition: MathExtras.h:651
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:66