LLVM  7.0.0svn
InstCombineSelect.cpp
Go to the documentation of this file.
1 //===- InstCombineSelect.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 implements the visitSelect function.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/APInt.h"
16 #include "llvm/ADT/Optional.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/IR/BasicBlock.h"
24 #include "llvm/IR/Constant.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/IRBuilder.h"
28 #include "llvm/IR/InstrTypes.h"
29 #include "llvm/IR/Instruction.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/IR/Operator.h"
34 #include "llvm/IR/PatternMatch.h"
35 #include "llvm/IR/Type.h"
36 #include "llvm/IR/User.h"
37 #include "llvm/IR/Value.h"
38 #include "llvm/Support/Casting.h"
40 #include "llvm/Support/KnownBits.h"
42 #include <cassert>
43 #include <utility>
44 
45 using namespace llvm;
46 using namespace PatternMatch;
47 
48 #define DEBUG_TYPE "instcombine"
49 
52  switch (SPF) {
53  default:
54  llvm_unreachable("unhandled!");
55 
56  case SPF_SMIN:
57  return SPF_SMAX;
58  case SPF_UMIN:
59  return SPF_UMAX;
60  case SPF_SMAX:
61  return SPF_SMIN;
62  case SPF_UMAX:
63  return SPF_UMIN;
64  }
65 }
66 
68  bool Ordered=false) {
69  switch (SPF) {
70  default:
71  llvm_unreachable("unhandled!");
72 
73  case SPF_SMIN:
74  return ICmpInst::ICMP_SLT;
75  case SPF_UMIN:
76  return ICmpInst::ICMP_ULT;
77  case SPF_SMAX:
78  return ICmpInst::ICMP_SGT;
79  case SPF_UMAX:
80  return ICmpInst::ICMP_UGT;
81  case SPF_FMINNUM:
82  return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT;
83  case SPF_FMAXNUM:
84  return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT;
85  }
86 }
87 
89  SelectPatternFlavor SPF, Value *A,
90  Value *B) {
93  return Builder.CreateSelect(Builder.CreateICmp(Pred, A, B), A, B);
94 }
95 
96 /// If one of the constants is zero (we know they can't both be) and we have an
97 /// icmp instruction with zero, and we have an 'and' with the non-constant value
98 /// and a power of two we can turn the select into a shift on the result of the
99 /// 'and'.
100 /// This folds:
101 /// select (icmp eq (and X, C1)), C2, C3
102 /// iff C1 is a power 2 and the difference between C2 and C3 is a power of 2.
103 /// To something like:
104 /// (shr (and (X, C1)), (log2(C1) - log2(C2-C3))) + C3
105 /// Or:
106 /// (shl (and (X, C1)), (log2(C2-C3) - log2(C1))) + C3
107 /// With some variations depending if C3 is larger than C2, or the shift
108 /// isn't needed, or the bit widths don't match.
109 static Value *foldSelectICmpAnd(Type *SelType, const ICmpInst *IC,
110  APInt TrueVal, APInt FalseVal,
111  InstCombiner::BuilderTy &Builder) {
112  assert(SelType->isIntOrIntVectorTy() && "Not an integer select?");
113 
114  // If this is a vector select, we need a vector compare.
115  if (SelType->isVectorTy() != IC->getType()->isVectorTy())
116  return nullptr;
117 
118  Value *V;
119  APInt AndMask;
120  bool CreateAnd = false;
121  ICmpInst::Predicate Pred = IC->getPredicate();
122  if (ICmpInst::isEquality(Pred)) {
123  if (!match(IC->getOperand(1), m_Zero()))
124  return nullptr;
125 
126  V = IC->getOperand(0);
127 
128  const APInt *AndRHS;
129  if (!match(V, m_And(m_Value(), m_Power2(AndRHS))))
130  return nullptr;
131 
132  AndMask = *AndRHS;
133  } else if (decomposeBitTestICmp(IC->getOperand(0), IC->getOperand(1),
134  Pred, V, AndMask)) {
135  assert(ICmpInst::isEquality(Pred) && "Not equality test?");
136 
137  if (!AndMask.isPowerOf2())
138  return nullptr;
139 
140  CreateAnd = true;
141  } else {
142  return nullptr;
143  }
144 
145  // If both select arms are non-zero see if we have a select of the form
146  // 'x ? 2^n + C : C'. Then we can offset both arms by C, use the logic
147  // for 'x ? 2^n : 0' and fix the thing up at the end.
148  APInt Offset(TrueVal.getBitWidth(), 0);
149  if (!TrueVal.isNullValue() && !FalseVal.isNullValue()) {
150  if ((TrueVal - FalseVal).isPowerOf2())
151  Offset = FalseVal;
152  else if ((FalseVal - TrueVal).isPowerOf2())
153  Offset = TrueVal;
154  else
155  return nullptr;
156 
157  // Adjust TrueVal and FalseVal to the offset.
158  TrueVal -= Offset;
159  FalseVal -= Offset;
160  }
161 
162  // Make sure one of the select arms is a power of 2.
163  if (!TrueVal.isPowerOf2() && !FalseVal.isPowerOf2())
164  return nullptr;
165 
166  // Determine which shift is needed to transform result of the 'and' into the
167  // desired result.
168  const APInt &ValC = !TrueVal.isNullValue() ? TrueVal : FalseVal;
169  unsigned ValZeros = ValC.logBase2();
170  unsigned AndZeros = AndMask.logBase2();
171 
172  if (CreateAnd) {
173  // Insert the AND instruction on the input to the truncate.
174  V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), AndMask));
175  }
176 
177  // If types don't match we can still convert the select by introducing a zext
178  // or a trunc of the 'and'.
179  if (ValZeros > AndZeros) {
180  V = Builder.CreateZExtOrTrunc(V, SelType);
181  V = Builder.CreateShl(V, ValZeros - AndZeros);
182  } else if (ValZeros < AndZeros) {
183  V = Builder.CreateLShr(V, AndZeros - ValZeros);
184  V = Builder.CreateZExtOrTrunc(V, SelType);
185  } else
186  V = Builder.CreateZExtOrTrunc(V, SelType);
187 
188  // Okay, now we know that everything is set up, we just don't know whether we
189  // have a icmp_ne or icmp_eq and whether the true or false val is the zero.
190  bool ShouldNotVal = !TrueVal.isNullValue();
191  ShouldNotVal ^= Pred == ICmpInst::ICMP_NE;
192  if (ShouldNotVal)
193  V = Builder.CreateXor(V, ValC);
194 
195  // Apply an offset if needed.
196  if (!Offset.isNullValue())
197  V = Builder.CreateAdd(V, ConstantInt::get(V->getType(), Offset));
198  return V;
199 }
200 
201 /// We want to turn code that looks like this:
202 /// %C = or %A, %B
203 /// %D = select %cond, %C, %A
204 /// into:
205 /// %C = select %cond, %B, 0
206 /// %D = or %A, %C
207 ///
208 /// Assuming that the specified instruction is an operand to the select, return
209 /// a bitmask indicating which operands of this instruction are foldable if they
210 /// equal the other incoming value of the select.
212  switch (I->getOpcode()) {
213  case Instruction::Add:
214  case Instruction::Mul:
215  case Instruction::And:
216  case Instruction::Or:
217  case Instruction::Xor:
218  return 3; // Can fold through either operand.
219  case Instruction::Sub: // Can only fold on the amount subtracted.
220  case Instruction::Shl: // Can only fold on the shift amount.
221  case Instruction::LShr:
222  case Instruction::AShr:
223  return 1;
224  default:
225  return 0; // Cannot fold
226  }
227 }
228 
229 /// For the same transformation as the previous function, return the identity
230 /// constant that goes into the select.
232  switch (I->getOpcode()) {
233  default: llvm_unreachable("This cannot happen!");
234  case Instruction::Add:
235  case Instruction::Sub:
236  case Instruction::Or:
237  case Instruction::Xor:
238  case Instruction::Shl:
239  case Instruction::LShr:
240  case Instruction::AShr:
242  case Instruction::And:
244  case Instruction::Mul:
245  return APInt(I->getType()->getScalarSizeInBits(), 1);
246  }
247 }
248 
249 /// We have (select c, TI, FI), and we know that TI and FI have the same opcode.
250 Instruction *InstCombiner::foldSelectOpOp(SelectInst &SI, Instruction *TI,
251  Instruction *FI) {
252  // Don't break up min/max patterns. The hasOneUse checks below prevent that
253  // for most cases, but vector min/max with bitcasts can be transformed. If the
254  // one-use restrictions are eased for other patterns, we still don't want to
255  // obfuscate min/max.
256  if ((match(&SI, m_SMin(m_Value(), m_Value())) ||
257  match(&SI, m_SMax(m_Value(), m_Value())) ||
258  match(&SI, m_UMin(m_Value(), m_Value())) ||
259  match(&SI, m_UMax(m_Value(), m_Value()))))
260  return nullptr;
261 
262  // If this is a cast from the same type, merge.
263  if (TI->getNumOperands() == 1 && TI->isCast()) {
264  Type *FIOpndTy = FI->getOperand(0)->getType();
265  if (TI->getOperand(0)->getType() != FIOpndTy)
266  return nullptr;
267 
268  // The select condition may be a vector. We may only change the operand
269  // type if the vector width remains the same (and matches the condition).
270  Type *CondTy = SI.getCondition()->getType();
271  if (CondTy->isVectorTy()) {
272  if (!FIOpndTy->isVectorTy())
273  return nullptr;
274  if (CondTy->getVectorNumElements() != FIOpndTy->getVectorNumElements())
275  return nullptr;
276 
277  // TODO: If the backend knew how to deal with casts better, we could
278  // remove this limitation. For now, there's too much potential to create
279  // worse codegen by promoting the select ahead of size-altering casts
280  // (PR28160).
281  //
282  // Note that ValueTracking's matchSelectPattern() looks through casts
283  // without checking 'hasOneUse' when it matches min/max patterns, so this
284  // transform may end up happening anyway.
285  if (TI->getOpcode() != Instruction::BitCast &&
286  (!TI->hasOneUse() || !FI->hasOneUse()))
287  return nullptr;
288  } else if (!TI->hasOneUse() || !FI->hasOneUse()) {
289  // TODO: The one-use restrictions for a scalar select could be eased if
290  // the fold of a select in visitLoadInst() was enhanced to match a pattern
291  // that includes a cast.
292  return nullptr;
293  }
294 
295  // Fold this by inserting a select from the input values.
296  Value *NewSI =
297  Builder.CreateSelect(SI.getCondition(), TI->getOperand(0),
298  FI->getOperand(0), SI.getName() + ".v", &SI);
299  return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
300  TI->getType());
301  }
302 
303  // Only handle binary operators (including two-operand getelementptr) with
304  // one-use here. As with the cast case above, it may be possible to relax the
305  // one-use constraint, but that needs be examined carefully since it may not
306  // reduce the total number of instructions.
307  if (TI->getNumOperands() != 2 || FI->getNumOperands() != 2 ||
308  (!isa<BinaryOperator>(TI) && !isa<GetElementPtrInst>(TI)) ||
309  !TI->hasOneUse() || !FI->hasOneUse())
310  return nullptr;
311 
312  // Figure out if the operations have any operands in common.
313  Value *MatchOp, *OtherOpT, *OtherOpF;
314  bool MatchIsOpZero;
315  if (TI->getOperand(0) == FI->getOperand(0)) {
316  MatchOp = TI->getOperand(0);
317  OtherOpT = TI->getOperand(1);
318  OtherOpF = FI->getOperand(1);
319  MatchIsOpZero = true;
320  } else if (TI->getOperand(1) == FI->getOperand(1)) {
321  MatchOp = TI->getOperand(1);
322  OtherOpT = TI->getOperand(0);
323  OtherOpF = FI->getOperand(0);
324  MatchIsOpZero = false;
325  } else if (!TI->isCommutative()) {
326  return nullptr;
327  } else if (TI->getOperand(0) == FI->getOperand(1)) {
328  MatchOp = TI->getOperand(0);
329  OtherOpT = TI->getOperand(1);
330  OtherOpF = FI->getOperand(0);
331  MatchIsOpZero = true;
332  } else if (TI->getOperand(1) == FI->getOperand(0)) {
333  MatchOp = TI->getOperand(1);
334  OtherOpT = TI->getOperand(0);
335  OtherOpF = FI->getOperand(1);
336  MatchIsOpZero = true;
337  } else {
338  return nullptr;
339  }
340 
341  // If we reach here, they do have operations in common.
342  Value *NewSI = Builder.CreateSelect(SI.getCondition(), OtherOpT, OtherOpF,
343  SI.getName() + ".v", &SI);
344  Value *Op0 = MatchIsOpZero ? MatchOp : NewSI;
345  Value *Op1 = MatchIsOpZero ? NewSI : MatchOp;
346  if (auto *BO = dyn_cast<BinaryOperator>(TI)) {
347  return BinaryOperator::Create(BO->getOpcode(), Op0, Op1);
348  }
349  if (auto *TGEP = dyn_cast<GetElementPtrInst>(TI)) {
350  auto *FGEP = cast<GetElementPtrInst>(FI);
351  Type *ElementType = TGEP->getResultElementType();
352  return TGEP->isInBounds() && FGEP->isInBounds()
353  ? GetElementPtrInst::CreateInBounds(ElementType, Op0, {Op1})
354  : GetElementPtrInst::Create(ElementType, Op0, {Op1});
355  }
356  llvm_unreachable("Expected BinaryOperator or GEP");
357  return nullptr;
358 }
359 
360 static bool isSelect01(const APInt &C1I, const APInt &C2I) {
361  if (!C1I.isNullValue() && !C2I.isNullValue()) // One side must be zero.
362  return false;
363  return C1I.isOneValue() || C1I.isAllOnesValue() ||
364  C2I.isOneValue() || C2I.isAllOnesValue();
365 }
366 
367 /// Try to fold the select into one of the operands to allow further
368 /// optimization.
369 Instruction *InstCombiner::foldSelectIntoOp(SelectInst &SI, Value *TrueVal,
370  Value *FalseVal) {
371  // See the comment above GetSelectFoldableOperands for a description of the
372  // transformation we are doing here.
373  if (auto *TVI = dyn_cast<BinaryOperator>(TrueVal)) {
374  if (TVI->hasOneUse() && !isa<Constant>(FalseVal)) {
375  if (unsigned SFO = getSelectFoldableOperands(TVI)) {
376  unsigned OpToFold = 0;
377  if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
378  OpToFold = 1;
379  } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
380  OpToFold = 2;
381  }
382 
383  if (OpToFold) {
385  Value *OOp = TVI->getOperand(2-OpToFold);
386  // Avoid creating select between 2 constants unless it's selecting
387  // between 0, 1 and -1.
388  const APInt *OOpC;
389  bool OOpIsAPInt = match(OOp, m_APInt(OOpC));
390  if (!isa<Constant>(OOp) || (OOpIsAPInt && isSelect01(CI, *OOpC))) {
391  Value *C = ConstantInt::get(OOp->getType(), CI);
392  Value *NewSel = Builder.CreateSelect(SI.getCondition(), OOp, C);
393  NewSel->takeName(TVI);
394  BinaryOperator *BO = BinaryOperator::Create(TVI->getOpcode(),
395  FalseVal, NewSel);
396  BO->copyIRFlags(TVI);
397  return BO;
398  }
399  }
400  }
401  }
402  }
403 
404  if (auto *FVI = dyn_cast<BinaryOperator>(FalseVal)) {
405  if (FVI->hasOneUse() && !isa<Constant>(TrueVal)) {
406  if (unsigned SFO = getSelectFoldableOperands(FVI)) {
407  unsigned OpToFold = 0;
408  if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
409  OpToFold = 1;
410  } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
411  OpToFold = 2;
412  }
413 
414  if (OpToFold) {
416  Value *OOp = FVI->getOperand(2-OpToFold);
417  // Avoid creating select between 2 constants unless it's selecting
418  // between 0, 1 and -1.
419  const APInt *OOpC;
420  bool OOpIsAPInt = match(OOp, m_APInt(OOpC));
421  if (!isa<Constant>(OOp) || (OOpIsAPInt && isSelect01(CI, *OOpC))) {
422  Value *C = ConstantInt::get(OOp->getType(), CI);
423  Value *NewSel = Builder.CreateSelect(SI.getCondition(), C, OOp);
424  NewSel->takeName(FVI);
425  BinaryOperator *BO = BinaryOperator::Create(FVI->getOpcode(),
426  TrueVal, NewSel);
427  BO->copyIRFlags(FVI);
428  return BO;
429  }
430  }
431  }
432  }
433  }
434 
435  return nullptr;
436 }
437 
438 /// We want to turn:
439 /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2))
440 /// into:
441 /// (or (shl (and X, C1), C3), Y)
442 /// iff:
443 /// C1 and C2 are both powers of 2
444 /// where:
445 /// C3 = Log(C2) - Log(C1)
446 ///
447 /// This transform handles cases where:
448 /// 1. The icmp predicate is inverted
449 /// 2. The select operands are reversed
450 /// 3. The magnitude of C2 and C1 are flipped
451 static Value *foldSelectICmpAndOr(const ICmpInst *IC, Value *TrueVal,
452  Value *FalseVal,
453  InstCombiner::BuilderTy &Builder) {
454  // Only handle integer compares. Also, if this is a vector select, we need a
455  // vector compare.
456  if (!TrueVal->getType()->isIntOrIntVectorTy() ||
457  TrueVal->getType()->isVectorTy() != IC->getType()->isVectorTy())
458  return nullptr;
459 
460  Value *CmpLHS = IC->getOperand(0);
461  Value *CmpRHS = IC->getOperand(1);
462 
463  Value *V;
464  unsigned C1Log;
465  bool IsEqualZero;
466  bool NeedAnd = false;
467  if (IC->isEquality()) {
468  if (!match(CmpRHS, m_Zero()))
469  return nullptr;
470 
471  const APInt *C1;
472  if (!match(CmpLHS, m_And(m_Value(), m_Power2(C1))))
473  return nullptr;
474 
475  V = CmpLHS;
476  C1Log = C1->logBase2();
477  IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_EQ;
478  } else if (IC->getPredicate() == ICmpInst::ICMP_SLT ||
479  IC->getPredicate() == ICmpInst::ICMP_SGT) {
480  // We also need to recognize (icmp slt (trunc (X)), 0) and
481  // (icmp sgt (trunc (X)), -1).
482  IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_SGT;
483  if ((IsEqualZero && !match(CmpRHS, m_AllOnes())) ||
484  (!IsEqualZero && !match(CmpRHS, m_Zero())))
485  return nullptr;
486 
487  if (!match(CmpLHS, m_OneUse(m_Trunc(m_Value(V)))))
488  return nullptr;
489 
490  C1Log = CmpLHS->getType()->getScalarSizeInBits() - 1;
491  NeedAnd = true;
492  } else {
493  return nullptr;
494  }
495 
496  const APInt *C2;
497  bool OrOnTrueVal = false;
498  bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2)));
499  if (!OrOnFalseVal)
500  OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2)));
501 
502  if (!OrOnFalseVal && !OrOnTrueVal)
503  return nullptr;
504 
505  Value *Y = OrOnFalseVal ? TrueVal : FalseVal;
506 
507  unsigned C2Log = C2->logBase2();
508 
509  bool NeedXor = (!IsEqualZero && OrOnFalseVal) || (IsEqualZero && OrOnTrueVal);
510  bool NeedShift = C1Log != C2Log;
511  bool NeedZExtTrunc = Y->getType()->getScalarSizeInBits() !=
513 
514  // Make sure we don't create more instructions than we save.
515  Value *Or = OrOnFalseVal ? FalseVal : TrueVal;
516  if ((NeedShift + NeedXor + NeedZExtTrunc) >
517  (IC->hasOneUse() + Or->hasOneUse()))
518  return nullptr;
519 
520  if (NeedAnd) {
521  // Insert the AND instruction on the input to the truncate.
523  V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), C1));
524  }
525 
526  if (C2Log > C1Log) {
527  V = Builder.CreateZExtOrTrunc(V, Y->getType());
528  V = Builder.CreateShl(V, C2Log - C1Log);
529  } else if (C1Log > C2Log) {
530  V = Builder.CreateLShr(V, C1Log - C2Log);
531  V = Builder.CreateZExtOrTrunc(V, Y->getType());
532  } else
533  V = Builder.CreateZExtOrTrunc(V, Y->getType());
534 
535  if (NeedXor)
536  V = Builder.CreateXor(V, *C2);
537 
538  return Builder.CreateOr(V, Y);
539 }
540 
541 /// Transform patterns such as: (a > b) ? a - b : 0
542 /// into: ((a > b) ? a : b) - b)
543 /// This produces a canonical max pattern that is more easily recognized by the
544 /// backend and converted into saturated subtraction instructions if those
545 /// exist.
546 /// There are 8 commuted/swapped variants of this pattern.
547 /// TODO: Also support a - UMIN(a,b) patterns.
549  const Value *TrueVal,
550  const Value *FalseVal,
551  InstCombiner::BuilderTy &Builder) {
552  ICmpInst::Predicate Pred = ICI->getPredicate();
553  if (!ICmpInst::isUnsigned(Pred))
554  return nullptr;
555 
556  // (b > a) ? 0 : a - b -> (b <= a) ? a - b : 0
557  if (match(TrueVal, m_Zero())) {
558  Pred = ICmpInst::getInversePredicate(Pred);
559  std::swap(TrueVal, FalseVal);
560  }
561  if (!match(FalseVal, m_Zero()))
562  return nullptr;
563 
564  Value *A = ICI->getOperand(0);
565  Value *B = ICI->getOperand(1);
566  if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_ULT) {
567  // (b < a) ? a - b : 0 -> (a > b) ? a - b : 0
568  std::swap(A, B);
569  Pred = ICmpInst::getSwappedPredicate(Pred);
570  }
571 
572  assert((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_UGT) &&
573  "Unexpected isUnsigned predicate!");
574 
575  // Account for swapped form of subtraction: ((a > b) ? b - a : 0).
576  bool IsNegative = false;
577  if (match(TrueVal, m_Sub(m_Specific(B), m_Specific(A))))
578  IsNegative = true;
579  else if (!match(TrueVal, m_Sub(m_Specific(A), m_Specific(B))))
580  return nullptr;
581 
582  // If sub is used anywhere else, we wouldn't be able to eliminate it
583  // afterwards.
584  if (!TrueVal->hasOneUse())
585  return nullptr;
586 
587  // All checks passed, convert to canonical unsigned saturated subtraction
588  // form: sub(max()).
589  // (a > b) ? a - b : 0 -> ((a > b) ? a : b) - b)
590  Value *Max = Builder.CreateSelect(Builder.CreateICmp(Pred, A, B), A, B);
591  return IsNegative ? Builder.CreateSub(B, Max) : Builder.CreateSub(Max, B);
592 }
593 
594 /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single
595 /// call to cttz/ctlz with flag 'is_zero_undef' cleared.
596 ///
597 /// For example, we can fold the following code sequence:
598 /// \code
599 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true)
600 /// %1 = icmp ne i32 %x, 0
601 /// %2 = select i1 %1, i32 %0, i32 32
602 /// \code
603 ///
604 /// into:
605 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false)
606 static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal,
607  InstCombiner::BuilderTy &Builder) {
608  ICmpInst::Predicate Pred = ICI->getPredicate();
609  Value *CmpLHS = ICI->getOperand(0);
610  Value *CmpRHS = ICI->getOperand(1);
611 
612  // Check if the condition value compares a value for equality against zero.
613  if (!ICI->isEquality() || !match(CmpRHS, m_Zero()))
614  return nullptr;
615 
616  Value *Count = FalseVal;
617  Value *ValueOnZero = TrueVal;
618  if (Pred == ICmpInst::ICMP_NE)
619  std::swap(Count, ValueOnZero);
620 
621  // Skip zero extend/truncate.
622  Value *V = nullptr;
623  if (match(Count, m_ZExt(m_Value(V))) ||
624  match(Count, m_Trunc(m_Value(V))))
625  Count = V;
626 
627  // Check if the value propagated on zero is a constant number equal to the
628  // sizeof in bits of 'Count'.
629  unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits();
630  if (!match(ValueOnZero, m_SpecificInt(SizeOfInBits)))
631  return nullptr;
632 
633  // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the
634  // input to the cttz/ctlz is used as LHS for the compare instruction.
635  if (match(Count, m_Intrinsic<Intrinsic::cttz>(m_Specific(CmpLHS))) ||
636  match(Count, m_Intrinsic<Intrinsic::ctlz>(m_Specific(CmpLHS)))) {
637  IntrinsicInst *II = cast<IntrinsicInst>(Count);
638  // Explicitly clear the 'undef_on_zero' flag.
639  IntrinsicInst *NewI = cast<IntrinsicInst>(II->clone());
640  NewI->setArgOperand(1, ConstantInt::getFalse(NewI->getContext()));
641  Builder.Insert(NewI);
642  return Builder.CreateZExtOrTrunc(NewI, ValueOnZero->getType());
643  }
644 
645  return nullptr;
646 }
647 
648 /// Return true if we find and adjust an icmp+select pattern where the compare
649 /// is with a constant that can be incremented or decremented to match the
650 /// minimum or maximum idiom.
651 static bool adjustMinMax(SelectInst &Sel, ICmpInst &Cmp) {
652  ICmpInst::Predicate Pred = Cmp.getPredicate();
653  Value *CmpLHS = Cmp.getOperand(0);
654  Value *CmpRHS = Cmp.getOperand(1);
655  Value *TrueVal = Sel.getTrueValue();
656  Value *FalseVal = Sel.getFalseValue();
657 
658  // We may move or edit the compare, so make sure the select is the only user.
659  const APInt *CmpC;
660  if (!Cmp.hasOneUse() || !match(CmpRHS, m_APInt(CmpC)))
661  return false;
662 
663  // These transforms only work for selects of integers or vector selects of
664  // integer vectors.
665  Type *SelTy = Sel.getType();
666  auto *SelEltTy = dyn_cast<IntegerType>(SelTy->getScalarType());
667  if (!SelEltTy || SelTy->isVectorTy() != Cmp.getType()->isVectorTy())
668  return false;
669 
670  Constant *AdjustedRHS;
671  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT)
672  AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC + 1);
673  else if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT)
674  AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC - 1);
675  else
676  return false;
677 
678  // X > C ? X : C+1 --> X < C+1 ? C+1 : X
679  // X < C ? X : C-1 --> X > C-1 ? C-1 : X
680  if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
681  (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
682  ; // Nothing to do here. Values match without any sign/zero extension.
683  }
684  // Types do not match. Instead of calculating this with mixed types, promote
685  // all to the larger type. This enables scalar evolution to analyze this
686  // expression.
687  else if (CmpRHS->getType()->getScalarSizeInBits() < SelEltTy->getBitWidth()) {
688  Constant *SextRHS = ConstantExpr::getSExt(AdjustedRHS, SelTy);
689 
690  // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X
691  // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X
692  // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X
693  // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X
694  if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && SextRHS == FalseVal) {
695  CmpLHS = TrueVal;
696  AdjustedRHS = SextRHS;
697  } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) &&
698  SextRHS == TrueVal) {
699  CmpLHS = FalseVal;
700  AdjustedRHS = SextRHS;
701  } else if (Cmp.isUnsigned()) {
702  Constant *ZextRHS = ConstantExpr::getZExt(AdjustedRHS, SelTy);
703  // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X
704  // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X
705  // zext + signed compare cannot be changed:
706  // 0xff <s 0x00, but 0x00ff >s 0x0000
707  if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && ZextRHS == FalseVal) {
708  CmpLHS = TrueVal;
709  AdjustedRHS = ZextRHS;
710  } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) &&
711  ZextRHS == TrueVal) {
712  CmpLHS = FalseVal;
713  AdjustedRHS = ZextRHS;
714  } else {
715  return false;
716  }
717  } else {
718  return false;
719  }
720  } else {
721  return false;
722  }
723 
724  Pred = ICmpInst::getSwappedPredicate(Pred);
725  CmpRHS = AdjustedRHS;
726  std::swap(FalseVal, TrueVal);
727  Cmp.setPredicate(Pred);
728  Cmp.setOperand(0, CmpLHS);
729  Cmp.setOperand(1, CmpRHS);
730  Sel.setOperand(1, TrueVal);
731  Sel.setOperand(2, FalseVal);
732  Sel.swapProfMetadata();
733 
734  // Move the compare instruction right before the select instruction. Otherwise
735  // the sext/zext value may be defined after the compare instruction uses it.
736  Cmp.moveBefore(&Sel);
737 
738  return true;
739 }
740 
741 /// If this is an integer min/max (icmp + select) with a constant operand,
742 /// create the canonical icmp for the min/max operation and canonicalize the
743 /// constant to the 'false' operand of the select:
744 /// select (icmp Pred X, C1), C2, X --> select (icmp Pred' X, C2), X, C2
745 /// Note: if C1 != C2, this will change the icmp constant to the existing
746 /// constant operand of the select.
747 static Instruction *
748 canonicalizeMinMaxWithConstant(SelectInst &Sel, ICmpInst &Cmp,
749  InstCombiner::BuilderTy &Builder) {
750  if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1)))
751  return nullptr;
752 
753  // Canonicalize the compare predicate based on whether we have min or max.
754  Value *LHS, *RHS;
755  ICmpInst::Predicate NewPred;
756  SelectPatternResult SPR = matchSelectPattern(&Sel, LHS, RHS);
757  switch (SPR.Flavor) {
758  case SPF_SMIN: NewPred = ICmpInst::ICMP_SLT; break;
759  case SPF_UMIN: NewPred = ICmpInst::ICMP_ULT; break;
760  case SPF_SMAX: NewPred = ICmpInst::ICMP_SGT; break;
761  case SPF_UMAX: NewPred = ICmpInst::ICMP_UGT; break;
762  default: return nullptr;
763  }
764 
765  // Is this already canonical?
766  if (Cmp.getOperand(0) == LHS && Cmp.getOperand(1) == RHS &&
767  Cmp.getPredicate() == NewPred)
768  return nullptr;
769 
770  // Create the canonical compare and plug it into the select.
771  Sel.setCondition(Builder.CreateICmp(NewPred, LHS, RHS));
772 
773  // If the select operands did not change, we're done.
774  if (Sel.getTrueValue() == LHS && Sel.getFalseValue() == RHS)
775  return &Sel;
776 
777  // If we are swapping the select operands, swap the metadata too.
778  assert(Sel.getTrueValue() == RHS && Sel.getFalseValue() == LHS &&
779  "Unexpected results from matchSelectPattern");
780  Sel.setTrueValue(LHS);
781  Sel.setFalseValue(RHS);
782  Sel.swapProfMetadata();
783  return &Sel;
784 }
785 
786 /// Visit a SelectInst that has an ICmpInst as its first operand.
787 Instruction *InstCombiner::foldSelectInstWithICmp(SelectInst &SI,
788  ICmpInst *ICI) {
789  Value *TrueVal = SI.getTrueValue();
790  Value *FalseVal = SI.getFalseValue();
791 
792  if (Instruction *NewSel = canonicalizeMinMaxWithConstant(SI, *ICI, Builder))
793  return NewSel;
794 
795  bool Changed = adjustMinMax(SI, *ICI);
796 
797  ICmpInst::Predicate Pred = ICI->getPredicate();
798  Value *CmpLHS = ICI->getOperand(0);
799  Value *CmpRHS = ICI->getOperand(1);
800 
801  // Transform (X >s -1) ? C1 : C2 --> ((X >>s 31) & (C2 - C1)) + C1
802  // and (X <s 0) ? C2 : C1 --> ((X >>s 31) & (C2 - C1)) + C1
803  // FIXME: Type and constness constraints could be lifted, but we have to
804  // watch code size carefully. We should consider xor instead of
805  // sub/add when we decide to do that.
806  // TODO: Merge this with foldSelectICmpAnd somehow.
807  if (CmpLHS->getType()->isIntOrIntVectorTy() &&
808  CmpLHS->getType() == TrueVal->getType()) {
809  const APInt *C1, *C2;
810  if (match(TrueVal, m_APInt(C1)) && match(FalseVal, m_APInt(C2))) {
811  ICmpInst::Predicate Pred = ICI->getPredicate();
812  Value *X;
813  APInt Mask;
814  if (decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask, false)) {
815  if (Mask.isSignMask()) {
816  assert(X == CmpLHS && "Expected to use the compare input directly");
817  assert(ICmpInst::isEquality(Pred) && "Expected equality predicate");
818 
819  if (Pred == ICmpInst::ICMP_NE)
820  std::swap(C1, C2);
821 
822  // This shift results in either -1 or 0.
823  Value *AShr = Builder.CreateAShr(X, Mask.getBitWidth() - 1);
824 
825  // Check if we can express the operation with a single or.
826  if (C2->isAllOnesValue())
827  return replaceInstUsesWith(SI, Builder.CreateOr(AShr, *C1));
828 
829  Value *And = Builder.CreateAnd(AShr, *C2 - *C1);
830  return replaceInstUsesWith(SI, Builder.CreateAdd(And,
831  ConstantInt::get(And->getType(), *C1)));
832  }
833  }
834  }
835  }
836 
837  {
838  const APInt *TrueValC, *FalseValC;
839  if (match(TrueVal, m_APInt(TrueValC)) &&
840  match(FalseVal, m_APInt(FalseValC)))
841  if (Value *V = foldSelectICmpAnd(SI.getType(), ICI, *TrueValC,
842  *FalseValC, Builder))
843  return replaceInstUsesWith(SI, V);
844  }
845 
846  // NOTE: if we wanted to, this is where to detect integer MIN/MAX
847 
848  if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) {
849  if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) {
850  // Transform (X == C) ? X : Y -> (X == C) ? C : Y
851  SI.setOperand(1, CmpRHS);
852  Changed = true;
853  } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) {
854  // Transform (X != C) ? Y : X -> (X != C) ? Y : C
855  SI.setOperand(2, CmpRHS);
856  Changed = true;
857  }
858  }
859 
860  // FIXME: This code is nearly duplicated in InstSimplify. Using/refactoring
861  // decomposeBitTestICmp() might help.
862  {
863  unsigned BitWidth =
864  DL.getTypeSizeInBits(TrueVal->getType()->getScalarType());
865  APInt MinSignedValue = APInt::getSignedMinValue(BitWidth);
866  Value *X;
867  const APInt *Y, *C;
868  bool TrueWhenUnset;
869  bool IsBitTest = false;
870  if (ICmpInst::isEquality(Pred) &&
871  match(CmpLHS, m_And(m_Value(X), m_Power2(Y))) &&
872  match(CmpRHS, m_Zero())) {
873  IsBitTest = true;
874  TrueWhenUnset = Pred == ICmpInst::ICMP_EQ;
875  } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
876  X = CmpLHS;
877  Y = &MinSignedValue;
878  IsBitTest = true;
879  TrueWhenUnset = false;
880  } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
881  X = CmpLHS;
882  Y = &MinSignedValue;
883  IsBitTest = true;
884  TrueWhenUnset = true;
885  }
886  if (IsBitTest) {
887  Value *V = nullptr;
888  // (X & Y) == 0 ? X : X ^ Y --> X & ~Y
889  if (TrueWhenUnset && TrueVal == X &&
890  match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
891  V = Builder.CreateAnd(X, ~(*Y));
892  // (X & Y) != 0 ? X ^ Y : X --> X & ~Y
893  else if (!TrueWhenUnset && FalseVal == X &&
894  match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
895  V = Builder.CreateAnd(X, ~(*Y));
896  // (X & Y) == 0 ? X ^ Y : X --> X | Y
897  else if (TrueWhenUnset && FalseVal == X &&
898  match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
899  V = Builder.CreateOr(X, *Y);
900  // (X & Y) != 0 ? X : X ^ Y --> X | Y
901  else if (!TrueWhenUnset && TrueVal == X &&
902  match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
903  V = Builder.CreateOr(X, *Y);
904 
905  if (V)
906  return replaceInstUsesWith(SI, V);
907  }
908  }
909 
910  if (Value *V = foldSelectICmpAndOr(ICI, TrueVal, FalseVal, Builder))
911  return replaceInstUsesWith(SI, V);
912 
913  if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, Builder))
914  return replaceInstUsesWith(SI, V);
915 
916  if (Value *V = canonicalizeSaturatedSubtract(ICI, TrueVal, FalseVal, Builder))
917  return replaceInstUsesWith(SI, V);
918 
919  return Changed ? &SI : nullptr;
920 }
921 
922 /// SI is a select whose condition is a PHI node (but the two may be in
923 /// different blocks). See if the true/false values (V) are live in all of the
924 /// predecessor blocks of the PHI. For example, cases like this can't be mapped:
925 ///
926 /// X = phi [ C1, BB1], [C2, BB2]
927 /// Y = add
928 /// Z = select X, Y, 0
929 ///
930 /// because Y is not live in BB1/BB2.
931 static bool canSelectOperandBeMappingIntoPredBlock(const Value *V,
932  const SelectInst &SI) {
933  // If the value is a non-instruction value like a constant or argument, it
934  // can always be mapped.
935  const Instruction *I = dyn_cast<Instruction>(V);
936  if (!I) return true;
937 
938  // If V is a PHI node defined in the same block as the condition PHI, we can
939  // map the arguments.
940  const PHINode *CondPHI = cast<PHINode>(SI.getCondition());
941 
942  if (const PHINode *VP = dyn_cast<PHINode>(I))
943  if (VP->getParent() == CondPHI->getParent())
944  return true;
945 
946  // Otherwise, if the PHI and select are defined in the same block and if V is
947  // defined in a different block, then we can transform it.
948  if (SI.getParent() == CondPHI->getParent() &&
949  I->getParent() != CondPHI->getParent())
950  return true;
951 
952  // Otherwise we have a 'hard' case and we can't tell without doing more
953  // detailed dominator based analysis, punt.
954  return false;
955 }
956 
957 /// We have an SPF (e.g. a min or max) of an SPF of the form:
958 /// SPF2(SPF1(A, B), C)
959 Instruction *InstCombiner::foldSPFofSPF(Instruction *Inner,
960  SelectPatternFlavor SPF1,
961  Value *A, Value *B,
962  Instruction &Outer,
963  SelectPatternFlavor SPF2, Value *C) {
964  if (Outer.getType() != Inner->getType())
965  return nullptr;
966 
967  if (C == A || C == B) {
968  // MAX(MAX(A, B), B) -> MAX(A, B)
969  // MIN(MIN(a, b), a) -> MIN(a, b)
970  if (SPF1 == SPF2)
971  return replaceInstUsesWith(Outer, Inner);
972 
973  // MAX(MIN(a, b), a) -> a
974  // MIN(MAX(a, b), a) -> a
975  if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) ||
976  (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) ||
977  (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) ||
978  (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN))
979  return replaceInstUsesWith(Outer, C);
980  }
981 
982  if (SPF1 == SPF2) {
983  const APInt *CB, *CC;
984  if (match(B, m_APInt(CB)) && match(C, m_APInt(CC))) {
985  // MIN(MIN(A, 23), 97) -> MIN(A, 23)
986  // MAX(MAX(A, 97), 23) -> MAX(A, 97)
987  if ((SPF1 == SPF_UMIN && CB->ule(*CC)) ||
988  (SPF1 == SPF_SMIN && CB->sle(*CC)) ||
989  (SPF1 == SPF_UMAX && CB->uge(*CC)) ||
990  (SPF1 == SPF_SMAX && CB->sge(*CC)))
991  return replaceInstUsesWith(Outer, Inner);
992 
993  // MIN(MIN(A, 97), 23) -> MIN(A, 23)
994  // MAX(MAX(A, 23), 97) -> MAX(A, 97)
995  if ((SPF1 == SPF_UMIN && CB->ugt(*CC)) ||
996  (SPF1 == SPF_SMIN && CB->sgt(*CC)) ||
997  (SPF1 == SPF_UMAX && CB->ult(*CC)) ||
998  (SPF1 == SPF_SMAX && CB->slt(*CC))) {
999  Outer.replaceUsesOfWith(Inner, A);
1000  return &Outer;
1001  }
1002  }
1003  }
1004 
1005  // ABS(ABS(X)) -> ABS(X)
1006  // NABS(NABS(X)) -> NABS(X)
1007  if (SPF1 == SPF2 && (SPF1 == SPF_ABS || SPF1 == SPF_NABS)) {
1008  return replaceInstUsesWith(Outer, Inner);
1009  }
1010 
1011  // ABS(NABS(X)) -> ABS(X)
1012  // NABS(ABS(X)) -> NABS(X)
1013  if ((SPF1 == SPF_ABS && SPF2 == SPF_NABS) ||
1014  (SPF1 == SPF_NABS && SPF2 == SPF_ABS)) {
1015  SelectInst *SI = cast<SelectInst>(Inner);
1016  Value *NewSI =
1017  Builder.CreateSelect(SI->getCondition(), SI->getFalseValue(),
1018  SI->getTrueValue(), SI->getName(), SI);
1019  return replaceInstUsesWith(Outer, NewSI);
1020  }
1021 
1022  auto IsFreeOrProfitableToInvert =
1023  [&](Value *V, Value *&NotV, bool &ElidesXor) {
1024  if (match(V, m_Not(m_Value(NotV)))) {
1025  // If V has at most 2 uses then we can get rid of the xor operation
1026  // entirely.
1027  ElidesXor |= !V->hasNUsesOrMore(3);
1028  return true;
1029  }
1030 
1031  if (IsFreeToInvert(V, !V->hasNUsesOrMore(3))) {
1032  NotV = nullptr;
1033  return true;
1034  }
1035 
1036  return false;
1037  };
1038 
1039  Value *NotA, *NotB, *NotC;
1040  bool ElidesXor = false;
1041 
1042  // MIN(MIN(~A, ~B), ~C) == ~MAX(MAX(A, B), C)
1043  // MIN(MAX(~A, ~B), ~C) == ~MAX(MIN(A, B), C)
1044  // MAX(MIN(~A, ~B), ~C) == ~MIN(MAX(A, B), C)
1045  // MAX(MAX(~A, ~B), ~C) == ~MIN(MIN(A, B), C)
1046  //
1047  // This transform is performance neutral if we can elide at least one xor from
1048  // the set of three operands, since we'll be tacking on an xor at the very
1049  // end.
1050  if (SelectPatternResult::isMinOrMax(SPF1) &&
1052  IsFreeOrProfitableToInvert(A, NotA, ElidesXor) &&
1053  IsFreeOrProfitableToInvert(B, NotB, ElidesXor) &&
1054  IsFreeOrProfitableToInvert(C, NotC, ElidesXor) && ElidesXor) {
1055  if (!NotA)
1056  NotA = Builder.CreateNot(A);
1057  if (!NotB)
1058  NotB = Builder.CreateNot(B);
1059  if (!NotC)
1060  NotC = Builder.CreateNot(C);
1061 
1062  Value *NewInner = generateMinMaxSelectPattern(
1063  Builder, getInverseMinMaxSelectPattern(SPF1), NotA, NotB);
1064  Value *NewOuter = Builder.CreateNot(generateMinMaxSelectPattern(
1065  Builder, getInverseMinMaxSelectPattern(SPF2), NewInner, NotC));
1066  return replaceInstUsesWith(Outer, NewOuter);
1067  }
1068 
1069  return nullptr;
1070 }
1071 
1072 /// Turn select C, (X + Y), (X - Y) --> (X + (select C, Y, (-Y))).
1073 /// This is even legal for FP.
1074 static Instruction *foldAddSubSelect(SelectInst &SI,
1075  InstCombiner::BuilderTy &Builder) {
1076  Value *CondVal = SI.getCondition();
1077  Value *TrueVal = SI.getTrueValue();
1078  Value *FalseVal = SI.getFalseValue();
1079  auto *TI = dyn_cast<Instruction>(TrueVal);
1080  auto *FI = dyn_cast<Instruction>(FalseVal);
1081  if (!TI || !FI || !TI->hasOneUse() || !FI->hasOneUse())
1082  return nullptr;
1083 
1084  Instruction *AddOp = nullptr, *SubOp = nullptr;
1085  if ((TI->getOpcode() == Instruction::Sub &&
1086  FI->getOpcode() == Instruction::Add) ||
1087  (TI->getOpcode() == Instruction::FSub &&
1088  FI->getOpcode() == Instruction::FAdd)) {
1089  AddOp = FI;
1090  SubOp = TI;
1091  } else if ((FI->getOpcode() == Instruction::Sub &&
1092  TI->getOpcode() == Instruction::Add) ||
1093  (FI->getOpcode() == Instruction::FSub &&
1094  TI->getOpcode() == Instruction::FAdd)) {
1095  AddOp = TI;
1096  SubOp = FI;
1097  }
1098 
1099  if (AddOp) {
1100  Value *OtherAddOp = nullptr;
1101  if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
1102  OtherAddOp = AddOp->getOperand(1);
1103  } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
1104  OtherAddOp = AddOp->getOperand(0);
1105  }
1106 
1107  if (OtherAddOp) {
1108  // So at this point we know we have (Y -> OtherAddOp):
1109  // select C, (add X, Y), (sub X, Z)
1110  Value *NegVal; // Compute -Z
1111  if (SI.getType()->isFPOrFPVectorTy()) {
1112  NegVal = Builder.CreateFNeg(SubOp->getOperand(1));
1113  if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) {
1114  FastMathFlags Flags = AddOp->getFastMathFlags();
1115  Flags &= SubOp->getFastMathFlags();
1116  NegInst->setFastMathFlags(Flags);
1117  }
1118  } else {
1119  NegVal = Builder.CreateNeg(SubOp->getOperand(1));
1120  }
1121 
1122  Value *NewTrueOp = OtherAddOp;
1123  Value *NewFalseOp = NegVal;
1124  if (AddOp != TI)
1125  std::swap(NewTrueOp, NewFalseOp);
1126  Value *NewSel = Builder.CreateSelect(CondVal, NewTrueOp, NewFalseOp,
1127  SI.getName() + ".p", &SI);
1128 
1129  if (SI.getType()->isFPOrFPVectorTy()) {
1130  Instruction *RI =
1131  BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel);
1132 
1133  FastMathFlags Flags = AddOp->getFastMathFlags();
1134  Flags &= SubOp->getFastMathFlags();
1135  RI->setFastMathFlags(Flags);
1136  return RI;
1137  } else
1138  return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
1139  }
1140  }
1141  return nullptr;
1142 }
1143 
1144 Instruction *InstCombiner::foldSelectExtConst(SelectInst &Sel) {
1145  Instruction *ExtInst;
1146  if (!match(Sel.getTrueValue(), m_Instruction(ExtInst)) &&
1147  !match(Sel.getFalseValue(), m_Instruction(ExtInst)))
1148  return nullptr;
1149 
1150  auto ExtOpcode = ExtInst->getOpcode();
1151  if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt)
1152  return nullptr;
1153 
1154  // TODO: Handle larger types? That requires adjusting FoldOpIntoSelect too.
1155  Value *X = ExtInst->getOperand(0);
1156  Type *SmallType = X->getType();
1157  if (!SmallType->isIntOrIntVectorTy(1))
1158  return nullptr;
1159 
1160  Constant *C;
1161  if (!match(Sel.getTrueValue(), m_Constant(C)) &&
1162  !match(Sel.getFalseValue(), m_Constant(C)))
1163  return nullptr;
1164 
1165  // If the constant is the same after truncation to the smaller type and
1166  // extension to the original type, we can narrow the select.
1167  Value *Cond = Sel.getCondition();
1168  Type *SelType = Sel.getType();
1169  Constant *TruncC = ConstantExpr::getTrunc(C, SmallType);
1170  Constant *ExtC = ConstantExpr::getCast(ExtOpcode, TruncC, SelType);
1171  if (ExtC == C) {
1172  Value *TruncCVal = cast<Value>(TruncC);
1173  if (ExtInst == Sel.getFalseValue())
1174  std::swap(X, TruncCVal);
1175 
1176  // select Cond, (ext X), C --> ext(select Cond, X, C')
1177  // select Cond, C, (ext X) --> ext(select Cond, C', X)
1178  Value *NewSel = Builder.CreateSelect(Cond, X, TruncCVal, "narrow", &Sel);
1179  return CastInst::Create(Instruction::CastOps(ExtOpcode), NewSel, SelType);
1180  }
1181 
1182  // If one arm of the select is the extend of the condition, replace that arm
1183  // with the extension of the appropriate known bool value.
1184  if (Cond == X) {
1185  if (ExtInst == Sel.getTrueValue()) {
1186  // select X, (sext X), C --> select X, -1, C
1187  // select X, (zext X), C --> select X, 1, C
1188  Constant *One = ConstantInt::getTrue(SmallType);
1189  Constant *AllOnesOrOne = ConstantExpr::getCast(ExtOpcode, One, SelType);
1190  return SelectInst::Create(Cond, AllOnesOrOne, C, "", nullptr, &Sel);
1191  } else {
1192  // select X, C, (sext X) --> select X, C, 0
1193  // select X, C, (zext X) --> select X, C, 0
1194  Constant *Zero = ConstantInt::getNullValue(SelType);
1195  return SelectInst::Create(Cond, C, Zero, "", nullptr, &Sel);
1196  }
1197  }
1198 
1199  return nullptr;
1200 }
1201 
1202 /// Try to transform a vector select with a constant condition vector into a
1203 /// shuffle for easier combining with other shuffles and insert/extract.
1204 static Instruction *canonicalizeSelectToShuffle(SelectInst &SI) {
1205  Value *CondVal = SI.getCondition();
1206  Constant *CondC;
1207  if (!CondVal->getType()->isVectorTy() || !match(CondVal, m_Constant(CondC)))
1208  return nullptr;
1209 
1210  unsigned NumElts = CondVal->getType()->getVectorNumElements();
1212  Mask.reserve(NumElts);
1213  Type *Int32Ty = Type::getInt32Ty(CondVal->getContext());
1214  for (unsigned i = 0; i != NumElts; ++i) {
1215  Constant *Elt = CondC->getAggregateElement(i);
1216  if (!Elt)
1217  return nullptr;
1218 
1219  if (Elt->isOneValue()) {
1220  // If the select condition element is true, choose from the 1st vector.
1221  Mask.push_back(ConstantInt::get(Int32Ty, i));
1222  } else if (Elt->isNullValue()) {
1223  // If the select condition element is false, choose from the 2nd vector.
1224  Mask.push_back(ConstantInt::get(Int32Ty, i + NumElts));
1225  } else if (isa<UndefValue>(Elt)) {
1226  // Undef in a select condition (choose one of the operands) does not mean
1227  // the same thing as undef in a shuffle mask (any value is acceptable), so
1228  // give up.
1229  return nullptr;
1230  } else {
1231  // Bail out on a constant expression.
1232  return nullptr;
1233  }
1234  }
1235 
1236  return new ShuffleVectorInst(SI.getTrueValue(), SI.getFalseValue(),
1237  ConstantVector::get(Mask));
1238 }
1239 
1240 /// Reuse bitcasted operands between a compare and select:
1241 /// select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
1242 /// bitcast (select (cmp (bitcast C), (bitcast D)), (bitcast C), (bitcast D))
1243 static Instruction *foldSelectCmpBitcasts(SelectInst &Sel,
1244  InstCombiner::BuilderTy &Builder) {
1245  Value *Cond = Sel.getCondition();
1246  Value *TVal = Sel.getTrueValue();
1247  Value *FVal = Sel.getFalseValue();
1248 
1249  CmpInst::Predicate Pred;
1250  Value *A, *B;
1251  if (!match(Cond, m_Cmp(Pred, m_Value(A), m_Value(B))))
1252  return nullptr;
1253 
1254  // The select condition is a compare instruction. If the select's true/false
1255  // values are already the same as the compare operands, there's nothing to do.
1256  if (TVal == A || TVal == B || FVal == A || FVal == B)
1257  return nullptr;
1258 
1259  Value *C, *D;
1260  if (!match(A, m_BitCast(m_Value(C))) || !match(B, m_BitCast(m_Value(D))))
1261  return nullptr;
1262 
1263  // select (cmp (bitcast C), (bitcast D)), (bitcast TSrc), (bitcast FSrc)
1264  Value *TSrc, *FSrc;
1265  if (!match(TVal, m_BitCast(m_Value(TSrc))) ||
1266  !match(FVal, m_BitCast(m_Value(FSrc))))
1267  return nullptr;
1268 
1269  // If the select true/false values are *different bitcasts* of the same source
1270  // operands, make the select operands the same as the compare operands and
1271  // cast the result. This is the canonical select form for min/max.
1272  Value *NewSel;
1273  if (TSrc == C && FSrc == D) {
1274  // select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
1275  // bitcast (select (cmp A, B), A, B)
1276  NewSel = Builder.CreateSelect(Cond, A, B, "", &Sel);
1277  } else if (TSrc == D && FSrc == C) {
1278  // select (cmp (bitcast C), (bitcast D)), (bitcast' D), (bitcast' C) -->
1279  // bitcast (select (cmp A, B), B, A)
1280  NewSel = Builder.CreateSelect(Cond, B, A, "", &Sel);
1281  } else {
1282  return nullptr;
1283  }
1284  return CastInst::CreateBitOrPointerCast(NewSel, Sel.getType());
1285 }
1286 
1287 /// Try to eliminate select instructions that test the returned flag of cmpxchg
1288 /// instructions.
1289 ///
1290 /// If a select instruction tests the returned flag of a cmpxchg instruction and
1291 /// selects between the returned value of the cmpxchg instruction its compare
1292 /// operand, the result of the select will always be equal to its false value.
1293 /// For example:
1294 ///
1295 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst
1296 /// %1 = extractvalue { i64, i1 } %0, 1
1297 /// %2 = extractvalue { i64, i1 } %0, 0
1298 /// %3 = select i1 %1, i64 %compare, i64 %2
1299 /// ret i64 %3
1300 ///
1301 /// The returned value of the cmpxchg instruction (%2) is the original value
1302 /// located at %ptr prior to any update. If the cmpxchg operation succeeds, %2
1303 /// must have been equal to %compare. Thus, the result of the select is always
1304 /// equal to %2, and the code can be simplified to:
1305 ///
1306 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst
1307 /// %1 = extractvalue { i64, i1 } %0, 0
1308 /// ret i64 %1
1309 ///
1310 static Instruction *foldSelectCmpXchg(SelectInst &SI) {
1311  // A helper that determines if V is an extractvalue instruction whose
1312  // aggregate operand is a cmpxchg instruction and whose single index is equal
1313  // to I. If such conditions are true, the helper returns the cmpxchg
1314  // instruction; otherwise, a nullptr is returned.
1315  auto isExtractFromCmpXchg = [](Value *V, unsigned I) -> AtomicCmpXchgInst * {
1316  auto *Extract = dyn_cast<ExtractValueInst>(V);
1317  if (!Extract)
1318  return nullptr;
1319  if (Extract->getIndices()[0] != I)
1320  return nullptr;
1321  return dyn_cast<AtomicCmpXchgInst>(Extract->getAggregateOperand());
1322  };
1323 
1324  // If the select has a single user, and this user is a select instruction that
1325  // we can simplify, skip the cmpxchg simplification for now.
1326  if (SI.hasOneUse())
1327  if (auto *Select = dyn_cast<SelectInst>(SI.user_back()))
1328  if (Select->getCondition() == SI.getCondition())
1329  if (Select->getFalseValue() == SI.getTrueValue() ||
1330  Select->getTrueValue() == SI.getFalseValue())
1331  return nullptr;
1332 
1333  // Ensure the select condition is the returned flag of a cmpxchg instruction.
1334  auto *CmpXchg = isExtractFromCmpXchg(SI.getCondition(), 1);
1335  if (!CmpXchg)
1336  return nullptr;
1337 
1338  // Check the true value case: The true value of the select is the returned
1339  // value of the same cmpxchg used by the condition, and the false value is the
1340  // cmpxchg instruction's compare operand.
1341  if (auto *X = isExtractFromCmpXchg(SI.getTrueValue(), 0))
1342  if (X == CmpXchg && X->getCompareOperand() == SI.getFalseValue()) {
1343  SI.setTrueValue(SI.getFalseValue());
1344  return &SI;
1345  }
1346 
1347  // Check the false value case: The false value of the select is the returned
1348  // value of the same cmpxchg used by the condition, and the true value is the
1349  // cmpxchg instruction's compare operand.
1350  if (auto *X = isExtractFromCmpXchg(SI.getFalseValue(), 0))
1351  if (X == CmpXchg && X->getCompareOperand() == SI.getTrueValue()) {
1352  SI.setTrueValue(SI.getFalseValue());
1353  return &SI;
1354  }
1355 
1356  return nullptr;
1357 }
1358 
1359 /// Reduce a sequence of min/max with a common operand.
1360 static Instruction *factorizeMinMaxTree(SelectPatternFlavor SPF, Value *LHS,
1361  Value *RHS,
1362  InstCombiner::BuilderTy &Builder) {
1363  assert(SelectPatternResult::isMinOrMax(SPF) && "Expected a min/max");
1364  // TODO: Allow FP min/max with nnan/nsz.
1365  if (!LHS->getType()->isIntOrIntVectorTy())
1366  return nullptr;
1367 
1368  // Match 3 of the same min/max ops. Example: umin(umin(), umin()).
1369  Value *A, *B, *C, *D;
1370  SelectPatternResult L = matchSelectPattern(LHS, A, B);
1371  SelectPatternResult R = matchSelectPattern(RHS, C, D);
1372  if (SPF != L.Flavor || L.Flavor != R.Flavor)
1373  return nullptr;
1374 
1375  // Look for a common operand. The use checks are different than usual because
1376  // a min/max pattern typically has 2 uses of each op: 1 by the cmp and 1 by
1377  // the select.
1378  Value *MinMaxOp = nullptr;
1379  Value *ThirdOp = nullptr;
1380  if (LHS->getNumUses() <= 2 && RHS->getNumUses() > 2) {
1381  // If the LHS is only used in this chain and the RHS is used outside of it,
1382  // reuse the RHS min/max because that will eliminate the LHS.
1383  if (D == A || C == A) {
1384  // min(min(a, b), min(c, a)) --> min(min(c, a), b)
1385  // min(min(a, b), min(a, d)) --> min(min(a, d), b)
1386  MinMaxOp = RHS;
1387  ThirdOp = B;
1388  } else if (D == B || C == B) {
1389  // min(min(a, b), min(c, b)) --> min(min(c, b), a)
1390  // min(min(a, b), min(b, d)) --> min(min(b, d), a)
1391  MinMaxOp = RHS;
1392  ThirdOp = A;
1393  }
1394  } else if (RHS->getNumUses() <= 2) {
1395  // Reuse the LHS. This will eliminate the RHS.
1396  if (D == A || D == B) {
1397  // min(min(a, b), min(c, a)) --> min(min(a, b), c)
1398  // min(min(a, b), min(c, b)) --> min(min(a, b), c)
1399  MinMaxOp = LHS;
1400  ThirdOp = C;
1401  } else if (C == A || C == B) {
1402  // min(min(a, b), min(b, d)) --> min(min(a, b), d)
1403  // min(min(a, b), min(c, b)) --> min(min(a, b), d)
1404  MinMaxOp = LHS;
1405  ThirdOp = D;
1406  }
1407  }
1408  if (!MinMaxOp || !ThirdOp)
1409  return nullptr;
1410 
1412  Value *CmpABC = Builder.CreateICmp(P, MinMaxOp, ThirdOp);
1413  return SelectInst::Create(CmpABC, MinMaxOp, ThirdOp);
1414 }
1415 
1417  Value *CondVal = SI.getCondition();
1418  Value *TrueVal = SI.getTrueValue();
1419  Value *FalseVal = SI.getFalseValue();
1420  Type *SelType = SI.getType();
1421 
1422  // FIXME: Remove this workaround when freeze related patches are done.
1423  // For select with undef operand which feeds into an equality comparison,
1424  // don't simplify it so loop unswitch can know the equality comparison
1425  // may have an undef operand. This is a workaround for PR31652 caused by
1426  // descrepancy about branch on undef between LoopUnswitch and GVN.
1427  if (isa<UndefValue>(TrueVal) || isa<UndefValue>(FalseVal)) {
1428  if (llvm::any_of(SI.users(), [&](User *U) {
1429  ICmpInst *CI = dyn_cast<ICmpInst>(U);
1430  if (CI && CI->isEquality())
1431  return true;
1432  return false;
1433  })) {
1434  return nullptr;
1435  }
1436  }
1437 
1438  if (Value *V = SimplifySelectInst(CondVal, TrueVal, FalseVal,
1439  SQ.getWithInstruction(&SI)))
1440  return replaceInstUsesWith(SI, V);
1441 
1442  if (Instruction *I = canonicalizeSelectToShuffle(SI))
1443  return I;
1444 
1445  // Canonicalize a one-use integer compare with a non-canonical predicate by
1446  // inverting the predicate and swapping the select operands. This matches a
1447  // compare canonicalization for conditional branches.
1448  // TODO: Should we do the same for FP compares?
1449  CmpInst::Predicate Pred;
1450  if (match(CondVal, m_OneUse(m_ICmp(Pred, m_Value(), m_Value()))) &&
1451  !isCanonicalPredicate(Pred)) {
1452  // Swap true/false values and condition.
1453  CmpInst *Cond = cast<CmpInst>(CondVal);
1455  SI.setOperand(1, FalseVal);
1456  SI.setOperand(2, TrueVal);
1457  SI.swapProfMetadata();
1458  Worklist.Add(Cond);
1459  return &SI;
1460  }
1461 
1462  if (SelType->isIntOrIntVectorTy(1) &&
1463  TrueVal->getType() == CondVal->getType()) {
1464  if (match(TrueVal, m_One())) {
1465  // Change: A = select B, true, C --> A = or B, C
1466  return BinaryOperator::CreateOr(CondVal, FalseVal);
1467  }
1468  if (match(TrueVal, m_Zero())) {
1469  // Change: A = select B, false, C --> A = and !B, C
1470  Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
1471  return BinaryOperator::CreateAnd(NotCond, FalseVal);
1472  }
1473  if (match(FalseVal, m_Zero())) {
1474  // Change: A = select B, C, false --> A = and B, C
1475  return BinaryOperator::CreateAnd(CondVal, TrueVal);
1476  }
1477  if (match(FalseVal, m_One())) {
1478  // Change: A = select B, C, true --> A = or !B, C
1479  Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
1480  return BinaryOperator::CreateOr(NotCond, TrueVal);
1481  }
1482 
1483  // select a, a, b -> a | b
1484  // select a, b, a -> a & b
1485  if (CondVal == TrueVal)
1486  return BinaryOperator::CreateOr(CondVal, FalseVal);
1487  if (CondVal == FalseVal)
1488  return BinaryOperator::CreateAnd(CondVal, TrueVal);
1489 
1490  // select a, ~a, b -> (~a) & b
1491  // select a, b, ~a -> (~a) | b
1492  if (match(TrueVal, m_Not(m_Specific(CondVal))))
1493  return BinaryOperator::CreateAnd(TrueVal, FalseVal);
1494  if (match(FalseVal, m_Not(m_Specific(CondVal))))
1495  return BinaryOperator::CreateOr(TrueVal, FalseVal);
1496  }
1497 
1498  // Selecting between two integer or vector splat integer constants?
1499  //
1500  // Note that we don't handle a scalar select of vectors:
1501  // select i1 %c, <2 x i8> <1, 1>, <2 x i8> <0, 0>
1502  // because that may need 3 instructions to splat the condition value:
1503  // extend, insertelement, shufflevector.
1504  if (SelType->isIntOrIntVectorTy() &&
1505  CondVal->getType()->isVectorTy() == SelType->isVectorTy()) {
1506  // select C, 1, 0 -> zext C to int
1507  if (match(TrueVal, m_One()) && match(FalseVal, m_Zero()))
1508  return new ZExtInst(CondVal, SelType);
1509 
1510  // select C, -1, 0 -> sext C to int
1511  if (match(TrueVal, m_AllOnes()) && match(FalseVal, m_Zero()))
1512  return new SExtInst(CondVal, SelType);
1513 
1514  // select C, 0, 1 -> zext !C to int
1515  if (match(TrueVal, m_Zero()) && match(FalseVal, m_One())) {
1516  Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
1517  return new ZExtInst(NotCond, SelType);
1518  }
1519 
1520  // select C, 0, -1 -> sext !C to int
1521  if (match(TrueVal, m_Zero()) && match(FalseVal, m_AllOnes())) {
1522  Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
1523  return new SExtInst(NotCond, SelType);
1524  }
1525  }
1526 
1527  // See if we are selecting two values based on a comparison of the two values.
1528  if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
1529  if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
1530  // Transform (X == Y) ? X : Y -> Y
1531  if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
1532  // This is not safe in general for floating point:
1533  // consider X== -0, Y== +0.
1534  // It becomes safe if either operand is a nonzero constant.
1535  ConstantFP *CFPt, *CFPf;
1536  if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
1537  !CFPt->getValueAPF().isZero()) ||
1538  ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
1539  !CFPf->getValueAPF().isZero()))
1540  return replaceInstUsesWith(SI, FalseVal);
1541  }
1542  // Transform (X une Y) ? X : Y -> X
1543  if (FCI->getPredicate() == FCmpInst::FCMP_UNE) {
1544  // This is not safe in general for floating point:
1545  // consider X== -0, Y== +0.
1546  // It becomes safe if either operand is a nonzero constant.
1547  ConstantFP *CFPt, *CFPf;
1548  if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
1549  !CFPt->getValueAPF().isZero()) ||
1550  ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
1551  !CFPf->getValueAPF().isZero()))
1552  return replaceInstUsesWith(SI, TrueVal);
1553  }
1554 
1555  // Canonicalize to use ordered comparisons by swapping the select
1556  // operands.
1557  //
1558  // e.g.
1559  // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X
1560  if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) {
1561  FCmpInst::Predicate InvPred = FCI->getInversePredicate();
1562  IRBuilder<>::FastMathFlagGuard FMFG(Builder);
1563  Builder.setFastMathFlags(FCI->getFastMathFlags());
1564  Value *NewCond = Builder.CreateFCmp(InvPred, TrueVal, FalseVal,
1565  FCI->getName() + ".inv");
1566 
1567  return SelectInst::Create(NewCond, FalseVal, TrueVal,
1568  SI.getName() + ".p");
1569  }
1570 
1571  // NOTE: if we wanted to, this is where to detect MIN/MAX
1572  } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
1573  // Transform (X == Y) ? Y : X -> X
1574  if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
1575  // This is not safe in general for floating point:
1576  // consider X== -0, Y== +0.
1577  // It becomes safe if either operand is a nonzero constant.
1578  ConstantFP *CFPt, *CFPf;
1579  if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
1580  !CFPt->getValueAPF().isZero()) ||
1581  ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
1582  !CFPf->getValueAPF().isZero()))
1583  return replaceInstUsesWith(SI, FalseVal);
1584  }
1585  // Transform (X une Y) ? Y : X -> Y
1586  if (FCI->getPredicate() == FCmpInst::FCMP_UNE) {
1587  // This is not safe in general for floating point:
1588  // consider X== -0, Y== +0.
1589  // It becomes safe if either operand is a nonzero constant.
1590  ConstantFP *CFPt, *CFPf;
1591  if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
1592  !CFPt->getValueAPF().isZero()) ||
1593  ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
1594  !CFPf->getValueAPF().isZero()))
1595  return replaceInstUsesWith(SI, TrueVal);
1596  }
1597 
1598  // Canonicalize to use ordered comparisons by swapping the select
1599  // operands.
1600  //
1601  // e.g.
1602  // (X ugt Y) ? X : Y -> (X ole Y) ? X : Y
1603  if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) {
1604  FCmpInst::Predicate InvPred = FCI->getInversePredicate();
1605  IRBuilder<>::FastMathFlagGuard FMFG(Builder);
1606  Builder.setFastMathFlags(FCI->getFastMathFlags());
1607  Value *NewCond = Builder.CreateFCmp(InvPred, FalseVal, TrueVal,
1608  FCI->getName() + ".inv");
1609 
1610  return SelectInst::Create(NewCond, FalseVal, TrueVal,
1611  SI.getName() + ".p");
1612  }
1613 
1614  // NOTE: if we wanted to, this is where to detect MIN/MAX
1615  }
1616  // NOTE: if we wanted to, this is where to detect ABS
1617  }
1618 
1619  // See if we are selecting two values based on a comparison of the two values.
1620  if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
1621  if (Instruction *Result = foldSelectInstWithICmp(SI, ICI))
1622  return Result;
1623 
1624  if (Instruction *Add = foldAddSubSelect(SI, Builder))
1625  return Add;
1626 
1627  // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
1628  auto *TI = dyn_cast<Instruction>(TrueVal);
1629  auto *FI = dyn_cast<Instruction>(FalseVal);
1630  if (TI && FI && TI->getOpcode() == FI->getOpcode())
1631  if (Instruction *IV = foldSelectOpOp(SI, TI, FI))
1632  return IV;
1633 
1634  if (Instruction *I = foldSelectExtConst(SI))
1635  return I;
1636 
1637  // See if we can fold the select into one of our operands.
1638  if (SelType->isIntOrIntVectorTy() || SelType->isFPOrFPVectorTy()) {
1639  if (Instruction *FoldI = foldSelectIntoOp(SI, TrueVal, FalseVal))
1640  return FoldI;
1641 
1642  Value *LHS, *RHS, *LHS2, *RHS2;
1643  Instruction::CastOps CastOp;
1644  SelectPatternResult SPR = matchSelectPattern(&SI, LHS, RHS, &CastOp);
1645  auto SPF = SPR.Flavor;
1646 
1648  // Canonicalize so that
1649  // - type casts are outside select patterns.
1650  // - float clamp is transformed to min/max pattern
1651 
1652  bool IsCastNeeded = LHS->getType() != SelType;
1653  Value *CmpLHS = cast<CmpInst>(CondVal)->getOperand(0);
1654  Value *CmpRHS = cast<CmpInst>(CondVal)->getOperand(1);
1655  if (IsCastNeeded ||
1656  (LHS->getType()->isFPOrFPVectorTy() &&
1657  ((CmpLHS != LHS && CmpLHS != RHS) ||
1658  (CmpRHS != LHS && CmpRHS != RHS)))) {
1660 
1661  Value *Cmp;
1662  if (CmpInst::isIntPredicate(Pred)) {
1663  Cmp = Builder.CreateICmp(Pred, LHS, RHS);
1664  } else {
1665  IRBuilder<>::FastMathFlagGuard FMFG(Builder);
1666  auto FMF = cast<FPMathOperator>(SI.getCondition())->getFastMathFlags();
1667  Builder.setFastMathFlags(FMF);
1668  Cmp = Builder.CreateFCmp(Pred, LHS, RHS);
1669  }
1670 
1671  Value *NewSI = Builder.CreateSelect(Cmp, LHS, RHS, SI.getName(), &SI);
1672  if (!IsCastNeeded)
1673  return replaceInstUsesWith(SI, NewSI);
1674 
1675  Value *NewCast = Builder.CreateCast(CastOp, NewSI, SelType);
1676  return replaceInstUsesWith(SI, NewCast);
1677  }
1678 
1679  // MAX(~a, ~b) -> ~MIN(a, b)
1680  // MIN(~a, ~b) -> ~MAX(a, b)
1681  Value *A, *B;
1682  if (match(LHS, m_Not(m_Value(A))) && match(RHS, m_Not(m_Value(B))) &&
1683  (LHS->getNumUses() <= 2 || RHS->getNumUses() <= 2)) {
1684  CmpInst::Predicate InvertedPred =
1686  Value *InvertedCmp = Builder.CreateICmp(InvertedPred, A, B);
1687  Value *NewSel = Builder.CreateSelect(InvertedCmp, A, B);
1688  return BinaryOperator::CreateNot(NewSel);
1689  }
1690 
1691  if (Instruction *I = factorizeMinMaxTree(SPF, LHS, RHS, Builder))
1692  return I;
1693  }
1694 
1695  if (SPF) {
1696  // MAX(MAX(a, b), a) -> MAX(a, b)
1697  // MIN(MIN(a, b), a) -> MIN(a, b)
1698  // MAX(MIN(a, b), a) -> a
1699  // MIN(MAX(a, b), a) -> a
1700  // ABS(ABS(a)) -> ABS(a)
1701  // NABS(NABS(a)) -> NABS(a)
1702  if (SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor)
1703  if (Instruction *R = foldSPFofSPF(cast<Instruction>(LHS),SPF2,LHS2,RHS2,
1704  SI, SPF, RHS))
1705  return R;
1706  if (SelectPatternFlavor SPF2 = matchSelectPattern(RHS, LHS2, RHS2).Flavor)
1707  if (Instruction *R = foldSPFofSPF(cast<Instruction>(RHS),SPF2,LHS2,RHS2,
1708  SI, SPF, LHS))
1709  return R;
1710  }
1711 
1712  // TODO.
1713  // ABS(-X) -> ABS(X)
1714  }
1715 
1716  // See if we can fold the select into a phi node if the condition is a select.
1717  if (auto *PN = dyn_cast<PHINode>(SI.getCondition()))
1718  // The true/false values have to be live in the PHI predecessor's blocks.
1719  if (canSelectOperandBeMappingIntoPredBlock(TrueVal, SI) &&
1720  canSelectOperandBeMappingIntoPredBlock(FalseVal, SI))
1721  if (Instruction *NV = foldOpIntoPhi(SI, PN))
1722  return NV;
1723 
1724  if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) {
1725  if (TrueSI->getCondition()->getType() == CondVal->getType()) {
1726  // select(C, select(C, a, b), c) -> select(C, a, c)
1727  if (TrueSI->getCondition() == CondVal) {
1728  if (SI.getTrueValue() == TrueSI->getTrueValue())
1729  return nullptr;
1730  SI.setOperand(1, TrueSI->getTrueValue());
1731  return &SI;
1732  }
1733  // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b)
1734  // We choose this as normal form to enable folding on the And and shortening
1735  // paths for the values (this helps GetUnderlyingObjects() for example).
1736  if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) {
1737  Value *And = Builder.CreateAnd(CondVal, TrueSI->getCondition());
1738  SI.setOperand(0, And);
1739  SI.setOperand(1, TrueSI->getTrueValue());
1740  return &SI;
1741  }
1742  }
1743  }
1744  if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) {
1745  if (FalseSI->getCondition()->getType() == CondVal->getType()) {
1746  // select(C, a, select(C, b, c)) -> select(C, a, c)
1747  if (FalseSI->getCondition() == CondVal) {
1748  if (SI.getFalseValue() == FalseSI->getFalseValue())
1749  return nullptr;
1750  SI.setOperand(2, FalseSI->getFalseValue());
1751  return &SI;
1752  }
1753  // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b)
1754  if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) {
1755  Value *Or = Builder.CreateOr(CondVal, FalseSI->getCondition());
1756  SI.setOperand(0, Or);
1757  SI.setOperand(2, FalseSI->getFalseValue());
1758  return &SI;
1759  }
1760  }
1761  }
1762 
1763  auto canMergeSelectThroughBinop = [](BinaryOperator *BO) {
1764  // The select might be preventing a division by 0.
1765  switch (BO->getOpcode()) {
1766  default:
1767  return true;
1768  case Instruction::SRem:
1769  case Instruction::URem:
1770  case Instruction::SDiv:
1771  case Instruction::UDiv:
1772  return false;
1773  }
1774  };
1775 
1776  // Try to simplify a binop sandwiched between 2 selects with the same
1777  // condition.
1778  // select(C, binop(select(C, X, Y), W), Z) -> select(C, binop(X, W), Z)
1779  BinaryOperator *TrueBO;
1780  if (match(TrueVal, m_OneUse(m_BinOp(TrueBO))) &&
1781  canMergeSelectThroughBinop(TrueBO)) {
1782  if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(0))) {
1783  if (TrueBOSI->getCondition() == CondVal) {
1784  TrueBO->setOperand(0, TrueBOSI->getTrueValue());
1785  Worklist.Add(TrueBO);
1786  return &SI;
1787  }
1788  }
1789  if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(1))) {
1790  if (TrueBOSI->getCondition() == CondVal) {
1791  TrueBO->setOperand(1, TrueBOSI->getTrueValue());
1792  Worklist.Add(TrueBO);
1793  return &SI;
1794  }
1795  }
1796  }
1797 
1798  // select(C, Z, binop(select(C, X, Y), W)) -> select(C, Z, binop(Y, W))
1799  BinaryOperator *FalseBO;
1800  if (match(FalseVal, m_OneUse(m_BinOp(FalseBO))) &&
1801  canMergeSelectThroughBinop(FalseBO)) {
1802  if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(0))) {
1803  if (FalseBOSI->getCondition() == CondVal) {
1804  FalseBO->setOperand(0, FalseBOSI->getFalseValue());
1805  Worklist.Add(FalseBO);
1806  return &SI;
1807  }
1808  }
1809  if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(1))) {
1810  if (FalseBOSI->getCondition() == CondVal) {
1811  FalseBO->setOperand(1, FalseBOSI->getFalseValue());
1812  Worklist.Add(FalseBO);
1813  return &SI;
1814  }
1815  }
1816  }
1817 
1818  if (BinaryOperator::isNot(CondVal)) {
1820  SI.setOperand(1, FalseVal);
1821  SI.setOperand(2, TrueVal);
1822  return &SI;
1823  }
1824 
1825  if (VectorType *VecTy = dyn_cast<VectorType>(SelType)) {
1826  unsigned VWidth = VecTy->getNumElements();
1827  APInt UndefElts(VWidth, 0);
1828  APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
1829  if (Value *V = SimplifyDemandedVectorElts(&SI, AllOnesEltMask, UndefElts)) {
1830  if (V != &SI)
1831  return replaceInstUsesWith(SI, V);
1832  return &SI;
1833  }
1834  }
1835 
1836  // See if we can determine the result of this select based on a dominating
1837  // condition.
1838  BasicBlock *Parent = SI.getParent();
1839  if (BasicBlock *Dom = Parent->getSinglePredecessor()) {
1840  auto *PBI = dyn_cast_or_null<BranchInst>(Dom->getTerminator());
1841  if (PBI && PBI->isConditional() &&
1842  PBI->getSuccessor(0) != PBI->getSuccessor(1) &&
1843  (PBI->getSuccessor(0) == Parent || PBI->getSuccessor(1) == Parent)) {
1844  bool CondIsTrue = PBI->getSuccessor(0) == Parent;
1845  Optional<bool> Implication = isImpliedCondition(
1846  PBI->getCondition(), SI.getCondition(), DL, CondIsTrue);
1847  if (Implication) {
1848  Value *V = *Implication ? TrueVal : FalseVal;
1849  return replaceInstUsesWith(SI, V);
1850  }
1851  }
1852  }
1853 
1854  // If we can compute the condition, there's no need for a select.
1855  // Like the above fold, we are attempting to reduce compile-time cost by
1856  // putting this fold here with limitations rather than in InstSimplify.
1857  // The motivation for this call into value tracking is to take advantage of
1858  // the assumption cache, so make sure that is populated.
1859  if (!CondVal->getType()->isVectorTy() && !AC.assumptions().empty()) {
1860  KnownBits Known(1);
1861  computeKnownBits(CondVal, Known, 0, &SI);
1862  if (Known.One.isOneValue())
1863  return replaceInstUsesWith(SI, TrueVal);
1864  if (Known.Zero.isOneValue())
1865  return replaceInstUsesWith(SI, FalseVal);
1866  }
1867 
1868  if (Instruction *BitCastSel = foldSelectCmpBitcasts(SI, Builder))
1869  return BitCastSel;
1870 
1871  // Simplify selects that test the returned flag of cmpxchg instructions.
1872  if (Instruction *Select = foldSelectCmpXchg(SI))
1873  return Select;
1874 
1875  return nullptr;
1876 }
MaxMin_match< ICmpInst, LHS, RHS, umin_pred_ty > m_UMin(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
Definition: PatternMatch.h:637
uint64_t CallInst * C
static bool isSelect01(const APInt &C1I, const APInt &C2I)
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:548
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:72
Value * CreateICmp(CmpInst::Predicate P, Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1677
This class is the base class for the comparison instructions.
Definition: InstrTypes.h:843
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
void setFastMathFlags(FastMathFlags FMF)
Convenience function for setting multiple fast-math flags on this instruction, which must be an opera...
bool isSignMask() const
Check if the APInt&#39;s value is returned by getSignMask.
Definition: APInt.h:466
bool isZero() const
Definition: APFloat.h:1143
static bool IsFreeToInvert(Value *V, bool WillInvertAllUses)
Return true if the specified value is free to invert (apply ~ to).
class_match< CmpInst > m_Cmp()
Matches any compare instruction and ignore it.
Definition: PatternMatch.h:80
This instruction extracts a struct member or array element value from an aggregate value...
static APInt getAllOnesValue(unsigned numBits)
Get the all-ones value.
Definition: APInt.h:555
BinaryOp_match< LHS, RHS, Instruction::Sub > m_Sub(const LHS &L, const RHS &R)
Definition: PatternMatch.h:577
DiagnosticInfoOptimizationBase::Argument NV
static BinaryOperator * CreateNot(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
Unsigned minimum.
Value * CreateZExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a ZExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:1434
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
BinaryOps getOpcode() const
Definition: InstrTypes.h:523
Value * CreateXor(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1138
bool decomposeBitTestICmp(Value *LHS, Value *RHS, CmpInst::Predicate &Pred, Value *&X, APInt &Mask, bool LookThroughTrunc=true)
Decompose an icmp into the form ((X & Mask) pred 0) if possible.
an instruction that atomically checks whether a specified value is in a memory location, and, if it is, stores a new value there.
Definition: Instructions.h:514
match_zero m_Zero()
Match an arbitrary zero/null constant.
Definition: PatternMatch.h:145
This class represents zero extension of integer types.
static GetElementPtrInst * Create(Type *PointeeType, Value *Ptr, ArrayRef< Value *> IdxList, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Definition: Instructions.h:863
bool slt(const APInt &RHS) const
Signed less than comparison.
Definition: APInt.h:1183
class_match< Constant > m_Constant()
Match an arbitrary Constant and ignore it.
Definition: PatternMatch.h:91
static bool isCanonicalPredicate(CmpInst::Predicate Pred)
Predicate canonicalization reduces the number of patterns that need to be matched by other transforms...
const Value * getTrueValue() const
unsigned less or equal
Definition: InstrTypes.h:879
unsigned less than
Definition: InstrTypes.h:878
0 1 0 0 True if ordered and less than
Definition: InstrTypes.h:859
This instruction constructs a fixed permutation of two input vectors.
static SelectInst * Create(Value *C, Value *S1, Value *S2, const Twine &NameStr="", Instruction *InsertBefore=nullptr, Instruction *MDFrom=nullptr)
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:738
static unsigned getSelectFoldableOperands(BinaryOperator *I)
We want to turn code that looks like this: C = or A, B D = select cond, C, A into: C = select cond...
1 1 1 0 True if unordered or not equal
Definition: InstrTypes.h:869
static CastInst * CreateBitOrPointerCast(Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Create a BitCast, a PtrToInt, or an IntToPTr cast instruction.
bool sgt(const APInt &RHS) const
Signed greather than comparison.
Definition: APInt.h:1253
static bool isEquality(Predicate P)
Return true if this predicate is either EQ or NE.
This class represents a sign extension of integer types.
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:227
void reserve(size_type N)
Definition: SmallVector.h:378
bool sle(const APInt &RHS) const
Signed less or equal comparison.
Definition: APInt.h:1218
void copyIRFlags(const Value *V, bool IncludeWrapFlags=true)
Convenience method to copy supported exact, fast-math, and (optionally) wrapping flags from V to this...
bool Ordered
Only applicable if Flavor is SPF_FMINNUM or SPF_FMAXNUM.
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1488
Signed maximum.
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:232
Value * CreateNot(Value *V, const Twine &Name="")
Definition: IRBuilder.h:1185
static GCMetadataPrinterRegistry::Add< OcamlGCMetadataPrinter > Y("ocaml", "ocaml 3.10-compatible collector")
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
static Value * foldSelectICmpAnd(Type *SelType, const ICmpInst *IC, APInt TrueVal, APInt FalseVal, InstCombiner::BuilderTy &Builder)
If one of the constants is zero (we know they can&#39;t both be) and we have an icmp instruction with zer...
BinaryOp_match< LHS, RHS, Instruction::Xor > m_Xor(const LHS &L, const RHS &R)
Definition: PatternMatch.h:649
This class represents the LLVM &#39;select&#39; instruction.
Predicate getInversePredicate() const
For example, EQ -> NE, UGT -> ULE, SLT -> SGE, OEQ -> UNE, UGT -> OLE, OLT -> UGE, etc.
Definition: InstrTypes.h:951
Absolute value.
bool isUnsigned() const
Determine if this instruction is using an unsigned comparison.
Definition: InstrTypes.h:1028
CastClass_match< OpTy, Instruction::Trunc > m_Trunc(const OpTy &Op)
Matches Trunc.
Definition: PatternMatch.h:975
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:707
static Constant * getSExt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1592
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:932
static SelectPatternFlavor getInverseMinMaxSelectPattern(SelectPatternFlavor SPF)
This file implements a class to represent arbitrary precision integral constant values and operations...
not_match< LHS > m_Not(const LHS &L)
static Constant * getZExt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1606
Instruction * clone() const
Create a copy of &#39;this&#39; instruction that is identical in all ways except the following: ...
bool isNullValue() const
Return true if this is the value that would be returned by getNullValue.
Definition: Constants.cpp:85
FastMathFlags getFastMathFlags() const
Convenience function for getting all the fast-math flags, which must be an operator which supports th...
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
CastClass_match< OpTy, Instruction::ZExt > m_ZExt(const OpTy &Op)
Matches ZExt.
Definition: PatternMatch.h:987
This instruction compares its operands according to the predicate given to the constructor.
MaxMin_match< ICmpInst, LHS, RHS, smin_pred_ty > m_SMin(const LHS &L, const RHS &R)
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:126
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:954
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:203
cst_pred_ty< is_power2 > m_Power2()
Match an integer or vector power-of-2.
Definition: PatternMatch.h:375
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:301
static BinaryOperator * CreateAdd(Value *S1, Value *S2, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp)
static const Value * getNotArgument(const Value *BinOp)
Value * getOperand(unsigned i) const
Definition: User.h:154
void replaceUsesOfWith(Value *From, Value *To)
Replace uses of one Value with another.
Definition: User.cpp:21
Value * CreateOr(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1122
Constant * getAggregateElement(unsigned Elt) const
For aggregates (struct/array/vector) return the constant that corresponds to the specified element if...
Definition: Constants.cpp:302
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return &#39;this&#39;.
Definition: Type.h:301
Value * CreateFCmp(CmpInst::Predicate P, Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1684
OneUse_match< T > m_OneUse(const T &SubPattern)
Definition: PatternMatch.h:63
#define P(N)
bool hasNUsesOrMore(unsigned N) const
Return true if this value has N users or more.
Definition: Value.cpp:140
Value * SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal, const SimplifyQuery &Q)
Given operands for a SelectInst, fold the result or return null.
bool isAllOnesValue() const
Determine if all bits are set.
Definition: APInt.h:389
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
MaxMin_match< ICmpInst, LHS, RHS, umax_pred_ty > m_UMax(const LHS &L, const RHS &R)
apint_match m_APInt(const APInt *&Res)
Match a ConstantInt or splatted ConstantVector, binding the specified pointer to the contained APInt...
Definition: PatternMatch.h:238
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:217
Value * CreateAShr(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Definition: IRBuilder.h:1088
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
BinaryOp_match< LHS, RHS, Instruction::Or > m_Or(const LHS &L, const RHS &R)
Definition: PatternMatch.h:643
bool ult(const APInt &RHS) const
Unsigned less than comparison.
Definition: APInt.h:1164
CastClass_match< OpTy, Instruction::BitCast > m_BitCast(const OpTy &Op)
Matches BitCast.
Definition: PatternMatch.h:963
This is an important base class in LLVM.
Definition: Constant.h:42
This file contains the declarations for the subclasses of Constant, which represent the different fla...
Value * CreateSelect(Value *C, Value *True, Value *False, const Twine &Name="", Instruction *MDFrom=nullptr)
Definition: IRBuilder.h:1732
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:264
bool isOneValue() const
Determine if this is a value of 1.
Definition: APInt.h:404
cst_pred_ty< is_all_ones > m_AllOnes()
Match an integer or vector with all bits set.
Definition: PatternMatch.h:333
static APInt getOneBitSet(unsigned numBits, unsigned BitNo)
Return an APInt with exactly one bit set in the result.
Definition: APInt.h:581
specificval_ty m_Specific(const Value *V)
Match if we have a specific specified value.
Definition: PatternMatch.h:445
bool any_of(R &&Range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:821
Value * CreateNeg(Value *V, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1163
This instruction compares its operands according to the predicate given to the constructor.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:853
static bool isNot(const Value *V)
class_match< BinaryOperator > m_BinOp()
Match an arbitrary binary operation and ignore it.
Definition: PatternMatch.h:75
Class to represent integer types.
Definition: DerivedTypes.h:40
const Value * getCondition() const
bool isCast() const
Definition: Instruction.h:132
void swapProfMetadata()
If the instruction has "branch_weights" MD_prof metadata and the MDNode has three operands (including...
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
signed greater than
Definition: InstrTypes.h:880
const APFloat & getValueAPF() const
Definition: Constants.h:299
CastClass_match< OpTy, Instruction::SExt > m_SExt(const OpTy &Op)
Matches SExt.
Definition: PatternMatch.h:981
Floating point maxnum.
0 0 1 0 True if ordered and greater than
Definition: InstrTypes.h:857
static Value * foldSelectICmpAndOr(const ICmpInst *IC, Value *TrueVal, Value *FalseVal, InstCombiner::BuilderTy &Builder)
We want to turn: (select (icmp eq (and X, C1), 0), Y, (or Y, C2)) into: (or (shl (and X...
unsigned getNumOperands() const
Definition: User.h:176
SelectPatternFlavor Flavor
unsigned getScalarSizeInBits() const LLVM_READONLY
If this is a vector type, return the getPrimitiveSizeInBits value for the element type...
Definition: Type.cpp:130
Unsigned maximum.
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:862
1 1 0 0 True if unordered or less than
Definition: InstrTypes.h:867
SelectPatternFlavor
Specific patterns of select instructions we can match.
Instruction * user_back()
Specialize the methods defined in Value, as we know that an instruction can only be used by other ins...
Definition: Instruction.h:64
Floating point minnum.
signed less than
Definition: InstrTypes.h:882
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
static Constant * getTrunc(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1578
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:585
bool uge(const APInt &RHS) const
Unsigned greater or equal comparison.
Definition: APInt.h:1272
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:541
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:452
Instruction * visitSelectInst(SelectInst &SI)
void setPredicate(Predicate P)
Set the predicate for this instruction to the specified value.
Definition: InstrTypes.h:932
void setOperand(unsigned i, Value *Val)
Definition: User.h:159
unsigned logBase2() const
Definition: APInt.h:1727
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:923
unsigned getVectorNumElements() const
Definition: DerivedTypes.h:462
bool isIntPredicate() const
Definition: InstrTypes.h:945
Class to represent vector types.
Definition: DerivedTypes.h:393
Class for arbitrary precision integers.
Definition: APInt.h:69
bool ule(const APInt &RHS) const
Unsigned less or equal comparison.
Definition: APInt.h:1202
bool isPowerOf2() const
Check if this APInt&#39;s value is a power of two greater than zero.
Definition: APInt.h:457
static BinaryOperator * Create(BinaryOps Op, Value *S1, Value *S2, const Twine &Name=Twine(), Instruction *InsertBefore=nullptr)
Construct a binary instruction, given the opcode and the two operands.
bool sge(const APInt &RHS) const
Signed greater or equal comparison.
Definition: APInt.h:1288
iterator_range< user_iterator > users()
Definition: Value.h:405
static Constant * getCast(unsigned ops, Constant *C, Type *Ty, bool OnlyIfReduced=false)
Convenience function for getting a Cast operation.
Definition: Constants.cpp:1471
Value * CreateShl(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1051
const Value * getFalseValue() const
void setCondition(Value *V)
static APInt getSelectFoldableConstant(BinaryOperator *I)
For the same transformation as the previous function, return the identity constant that goes into the...
static CastInst * Create(Instruction::CastOps, Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Provides a way to construct any of the CastInst subclasses using an opcode instead of the subclass&#39;s ...
bool ugt(const APInt &RHS) const
Unsigned greather than comparison.
Definition: APInt.h:1234
Predicate getPredicate() const
Return the predicate for this instruction.
Definition: InstrTypes.h:927
unsigned getNumUses() const
This method computes the number of uses of this Value.
Definition: Value.cpp:170
static cl::opt< bool > NeedAnd("extract-needand", cl::init(true), cl::Hidden, cl::desc("Require & in extract patterns"))
static Value * canonicalizeSaturatedSubtract(const ICmpInst *ICI, const Value *TrueVal, const Value *FalseVal, InstCombiner::BuilderTy &Builder)
Transform patterns such as: (a > b) ? a - b : 0 into: ((a > b) ? a : b) - b) This produces a canonica...
void setTrueValue(Value *V)
static IntegerType * getInt32Ty(LLVMContext &C)
Definition: Type.cpp:176
static bool isMinOrMax(SelectPatternFlavor SPF)
When implementing this min/max pattern as fcmp; select, does the fcmp have to be ordered?
unsigned greater or equal
Definition: InstrTypes.h:877
static bool isUnordered(Predicate predicate)
Determine if the predicate is an unordered operation.
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:224
bool isEquality() const
Return true if this predicate is either EQ or NE.
#define I(x, y, z)
Definition: MD5.cpp:58
void setArgOperand(unsigned i, Value *v)
MaxMin_match< ICmpInst, LHS, RHS, smax_pred_ty > m_SMax(const LHS &L, const RHS &R)
static Value * generateMinMaxSelectPattern(InstCombiner::BuilderTy &Builder, SelectPatternFlavor SPF, Value *A, Value *B)
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
void setFalseValue(Value *V)
1 0 1 0 True if unordered or greater than
Definition: InstrTypes.h:865
Signed minimum.
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1106
Optional< bool > isImpliedCondition(const Value *LHS, const Value *RHS, const DataLayout &DL, bool LHSIsTrue=true, unsigned Depth=0)
Return true if RHS is known to be implied true by LHS.
InstTy * Insert(InstTy *I, const Twine &Name="") const
Insert and return the specified instruction.
Definition: IRBuilder.h:758
bool isFPOrFPVectorTy() const
Return true if this is a FP type or a vector of FP.
Definition: Type.h:185
bool isOneValue() const
Returns true if the value is one.
Definition: Constants.cpp:126
Value * CreateCast(Instruction::CastOps Op, Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1519
static GetElementPtrInst * CreateInBounds(Value *Ptr, ArrayRef< Value *> IdxList, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Create an "inbounds" getelementptr.
Definition: Instructions.h:897
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
static APInt getSignedMinValue(unsigned numBits)
Gets minimum signed value of APInt for a specific bit width.
Definition: APInt.h:538
0 0 0 1 True if ordered and equal
Definition: InstrTypes.h:856
LLVM Value Representation.
Definition: Value.h:73
static CmpInst::Predicate getCmpPredicateForMinMax(SelectPatternFlavor SPF, bool Ordered=false)
This file provides internal interfaces used to implement the InstCombine.
SelectPatternResult matchSelectPattern(Value *V, Value *&LHS, Value *&RHS, Instruction::CastOps *CastOp=nullptr, unsigned Depth=0)
Pattern match integer [SU]MIN, [SU]MAX and ABS idioms, returning the kind and providing the out param...
cst_pred_ty< is_one > m_One()
Match an integer 1 or a vector with all elements equal to 1.
Definition: PatternMatch.h:397
std::underlying_type< E >::type Mask()
Get a bitmask with 1s in all places up to the high-order bit of E&#39;s largest value.
Definition: BitmaskEnum.h:81
void setFastMathFlags(FastMathFlags NewFMF)
Set the fast-math flags to be used with generated fp-math operators.
Definition: IRBuilder.h:221
void moveBefore(Instruction *MovePos)
Unlink this instruction from its current basic block and insert it into the basic block that MovePos ...
Definition: Instruction.cpp:86
Value * CreateLShr(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Definition: IRBuilder.h:1070
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:418
Convenience struct for specifying and reasoning about fast-math flags.
Definition: Operator.h:160
unsigned greater than
Definition: InstrTypes.h:876
Predicate getSwappedPredicate() const
For example, EQ->EQ, SLE->SGE, ULT->UGT, OEQ->OEQ, ULE->UGE, OLT->OGT, etc.
Definition: InstrTypes.h:967
static APInt getNullValue(unsigned numBits)
Get the &#39;0&#39; value.
Definition: APInt.h:562
specific_intval m_SpecificInt(uint64_t V)
Match a specific integer value or vector with all elements equal to the value.
Definition: PatternMatch.h:506
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:1020
Value * CreateFNeg(Value *V, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1178
bind_ty< Instruction > m_Instruction(Instruction *&I)
Match an instruction, capturing it if we match.
Definition: PatternMatch.h:422
bool isNullValue() const
Determine if all bits are clear.
Definition: APInt.h:399
IntegerType * Int32Ty
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:67
CmpClass_match< LHS, RHS, ICmpInst, ICmpInst::Predicate > m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R)
Definition: PatternMatch.h:900