LLVM  6.0.0svn
InstCombineCompares.cpp
Go to the documentation of this file.
1 //===- InstCombineCompares.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 visitICmp and visitFCmp functions.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/APSInt.h"
16 #include "llvm/ADT/SetVector.h"
17 #include "llvm/ADT/Statistic.h"
23 #include "llvm/IR/ConstantRange.h"
24 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/PatternMatch.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/KnownBits.h"
30 
31 using namespace llvm;
32 using namespace PatternMatch;
33 
34 #define DEBUG_TYPE "instcombine"
35 
36 // How many times is a select replaced by one of its operands?
37 STATISTIC(NumSel, "Number of select opts");
38 
39 
41  return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
42 }
43 
44 static bool hasAddOverflow(ConstantInt *Result,
45  ConstantInt *In1, ConstantInt *In2,
46  bool IsSigned) {
47  if (!IsSigned)
48  return Result->getValue().ult(In1->getValue());
49 
50  if (In2->isNegative())
51  return Result->getValue().sgt(In1->getValue());
52  return Result->getValue().slt(In1->getValue());
53 }
54 
55 /// Compute Result = In1+In2, returning true if the result overflowed for this
56 /// type.
57 static bool addWithOverflow(Constant *&Result, Constant *In1,
58  Constant *In2, bool IsSigned = false) {
59  Result = ConstantExpr::getAdd(In1, In2);
60 
61  if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
62  for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
64  if (hasAddOverflow(extractElement(Result, Idx),
65  extractElement(In1, Idx),
66  extractElement(In2, Idx),
67  IsSigned))
68  return true;
69  }
70  return false;
71  }
72 
73  return hasAddOverflow(cast<ConstantInt>(Result),
74  cast<ConstantInt>(In1), cast<ConstantInt>(In2),
75  IsSigned);
76 }
77 
78 static bool hasSubOverflow(ConstantInt *Result,
79  ConstantInt *In1, ConstantInt *In2,
80  bool IsSigned) {
81  if (!IsSigned)
82  return Result->getValue().ugt(In1->getValue());
83 
84  if (In2->isNegative())
85  return Result->getValue().slt(In1->getValue());
86 
87  return Result->getValue().sgt(In1->getValue());
88 }
89 
90 /// Compute Result = In1-In2, returning true if the result overflowed for this
91 /// type.
92 static bool subWithOverflow(Constant *&Result, Constant *In1,
93  Constant *In2, bool IsSigned = false) {
94  Result = ConstantExpr::getSub(In1, In2);
95 
96  if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
97  for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
99  if (hasSubOverflow(extractElement(Result, Idx),
100  extractElement(In1, Idx),
101  extractElement(In2, Idx),
102  IsSigned))
103  return true;
104  }
105  return false;
106  }
107 
108  return hasSubOverflow(cast<ConstantInt>(Result),
109  cast<ConstantInt>(In1), cast<ConstantInt>(In2),
110  IsSigned);
111 }
112 
113 /// Given an icmp instruction, return true if any use of this comparison is a
114 /// branch on sign bit comparison.
115 static bool hasBranchUse(ICmpInst &I) {
116  for (auto *U : I.users())
117  if (isa<BranchInst>(U))
118  return true;
119  return false;
120 }
121 
122 /// Given an exploded icmp instruction, return true if the comparison only
123 /// checks the sign bit. If it only checks the sign bit, set TrueIfSigned if the
124 /// result of the comparison is true when the input value is signed.
125 static bool isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS,
126  bool &TrueIfSigned) {
127  switch (Pred) {
128  case ICmpInst::ICMP_SLT: // True if LHS s< 0
129  TrueIfSigned = true;
130  return RHS.isNullValue();
131  case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
132  TrueIfSigned = true;
133  return RHS.isAllOnesValue();
134  case ICmpInst::ICMP_SGT: // True if LHS s> -1
135  TrueIfSigned = false;
136  return RHS.isAllOnesValue();
137  case ICmpInst::ICMP_UGT:
138  // True if LHS u> RHS and RHS == high-bit-mask - 1
139  TrueIfSigned = true;
140  return RHS.isMaxSignedValue();
141  case ICmpInst::ICMP_UGE:
142  // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
143  TrueIfSigned = true;
144  return RHS.isSignMask();
145  default:
146  return false;
147  }
148 }
149 
150 /// Returns true if the exploded icmp can be expressed as a signed comparison
151 /// to zero and updates the predicate accordingly.
152 /// The signedness of the comparison is preserved.
153 /// TODO: Refactor with decomposeBitTestICmp()?
154 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
155  if (!ICmpInst::isSigned(Pred))
156  return false;
157 
158  if (C.isNullValue())
159  return ICmpInst::isRelational(Pred);
160 
161  if (C.isOneValue()) {
162  if (Pred == ICmpInst::ICMP_SLT) {
163  Pred = ICmpInst::ICMP_SLE;
164  return true;
165  }
166  } else if (C.isAllOnesValue()) {
167  if (Pred == ICmpInst::ICMP_SGT) {
168  Pred = ICmpInst::ICMP_SGE;
169  return true;
170  }
171  }
172 
173  return false;
174 }
175 
176 /// Given a signed integer type and a set of known zero and one bits, compute
177 /// the maximum and minimum values that could have the specified known zero and
178 /// known one bits, returning them in Min/Max.
179 /// TODO: Move to method on KnownBits struct?
181  APInt &Min, APInt &Max) {
182  assert(Known.getBitWidth() == Min.getBitWidth() &&
183  Known.getBitWidth() == Max.getBitWidth() &&
184  "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
185  APInt UnknownBits = ~(Known.Zero|Known.One);
186 
187  // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
188  // bit if it is unknown.
189  Min = Known.One;
190  Max = Known.One|UnknownBits;
191 
192  if (UnknownBits.isNegative()) { // Sign bit is unknown
193  Min.setSignBit();
194  Max.clearSignBit();
195  }
196 }
197 
198 /// Given an unsigned integer type and a set of known zero and one bits, compute
199 /// the maximum and minimum values that could have the specified known zero and
200 /// known one bits, returning them in Min/Max.
201 /// TODO: Move to method on KnownBits struct?
203  APInt &Min, APInt &Max) {
204  assert(Known.getBitWidth() == Min.getBitWidth() &&
205  Known.getBitWidth() == Max.getBitWidth() &&
206  "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
207  APInt UnknownBits = ~(Known.Zero|Known.One);
208 
209  // The minimum value is when the unknown bits are all zeros.
210  Min = Known.One;
211  // The maximum value is when the unknown bits are all ones.
212  Max = Known.One|UnknownBits;
213 }
214 
215 /// This is called when we see this pattern:
216 /// cmp pred (load (gep GV, ...)), cmpcst
217 /// where GV is a global variable with a constant initializer. Try to simplify
218 /// this into some simple computation that does not need the load. For example
219 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
220 ///
221 /// If AndCst is non-null, then the loaded value is masked with that constant
222 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
223 Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
224  GlobalVariable *GV,
225  CmpInst &ICI,
226  ConstantInt *AndCst) {
227  Constant *Init = GV->getInitializer();
228  if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
229  return nullptr;
230 
231  uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
232  // Don't blow up on huge arrays.
233  if (ArrayElementCount > MaxArraySizeForCombine)
234  return nullptr;
235 
236  // There are many forms of this optimization we can handle, for now, just do
237  // the simple index into a single-dimensional array.
238  //
239  // Require: GEP GV, 0, i {{, constant indices}}
240  if (GEP->getNumOperands() < 3 ||
241  !isa<ConstantInt>(GEP->getOperand(1)) ||
242  !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
243  isa<Constant>(GEP->getOperand(2)))
244  return nullptr;
245 
246  // Check that indices after the variable are constants and in-range for the
247  // type they index. Collect the indices. This is typically for arrays of
248  // structs.
249  SmallVector<unsigned, 4> LaterIndices;
250 
251  Type *EltTy = Init->getType()->getArrayElementType();
252  for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
253  ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
254  if (!Idx) return nullptr; // Variable index.
255 
256  uint64_t IdxVal = Idx->getZExtValue();
257  if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
258 
259  if (StructType *STy = dyn_cast<StructType>(EltTy))
260  EltTy = STy->getElementType(IdxVal);
261  else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
262  if (IdxVal >= ATy->getNumElements()) return nullptr;
263  EltTy = ATy->getElementType();
264  } else {
265  return nullptr; // Unknown type.
266  }
267 
268  LaterIndices.push_back(IdxVal);
269  }
270 
271  enum { Overdefined = -3, Undefined = -2 };
272 
273  // Variables for our state machines.
274 
275  // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
276  // "i == 47 | i == 87", where 47 is the first index the condition is true for,
277  // and 87 is the second (and last) index. FirstTrueElement is -2 when
278  // undefined, otherwise set to the first true element. SecondTrueElement is
279  // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
280  int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
281 
282  // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
283  // form "i != 47 & i != 87". Same state transitions as for true elements.
284  int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
285 
286  /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
287  /// define a state machine that triggers for ranges of values that the index
288  /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
289  /// This is -2 when undefined, -3 when overdefined, and otherwise the last
290  /// index in the range (inclusive). We use -2 for undefined here because we
291  /// use relative comparisons and don't want 0-1 to match -1.
292  int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
293 
294  // MagicBitvector - This is a magic bitvector where we set a bit if the
295  // comparison is true for element 'i'. If there are 64 elements or less in
296  // the array, this will fully represent all the comparison results.
297  uint64_t MagicBitvector = 0;
298 
299  // Scan the array and see if one of our patterns matches.
300  Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
301  for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
302  Constant *Elt = Init->getAggregateElement(i);
303  if (!Elt) return nullptr;
304 
305  // If this is indexing an array of structures, get the structure element.
306  if (!LaterIndices.empty())
307  Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
308 
309  // If the element is masked, handle it.
310  if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
311 
312  // Find out if the comparison would be true or false for the i'th element.
314  CompareRHS, DL, &TLI);
315  // If the result is undef for this element, ignore it.
316  if (isa<UndefValue>(C)) {
317  // Extend range state machines to cover this element in case there is an
318  // undef in the middle of the range.
319  if (TrueRangeEnd == (int)i-1)
320  TrueRangeEnd = i;
321  if (FalseRangeEnd == (int)i-1)
322  FalseRangeEnd = i;
323  continue;
324  }
325 
326  // If we can't compute the result for any of the elements, we have to give
327  // up evaluating the entire conditional.
328  if (!isa<ConstantInt>(C)) return nullptr;
329 
330  // Otherwise, we know if the comparison is true or false for this element,
331  // update our state machines.
332  bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
333 
334  // State machine for single/double/range index comparison.
335  if (IsTrueForElt) {
336  // Update the TrueElement state machine.
337  if (FirstTrueElement == Undefined)
338  FirstTrueElement = TrueRangeEnd = i; // First true element.
339  else {
340  // Update double-compare state machine.
341  if (SecondTrueElement == Undefined)
342  SecondTrueElement = i;
343  else
344  SecondTrueElement = Overdefined;
345 
346  // Update range state machine.
347  if (TrueRangeEnd == (int)i-1)
348  TrueRangeEnd = i;
349  else
350  TrueRangeEnd = Overdefined;
351  }
352  } else {
353  // Update the FalseElement state machine.
354  if (FirstFalseElement == Undefined)
355  FirstFalseElement = FalseRangeEnd = i; // First false element.
356  else {
357  // Update double-compare state machine.
358  if (SecondFalseElement == Undefined)
359  SecondFalseElement = i;
360  else
361  SecondFalseElement = Overdefined;
362 
363  // Update range state machine.
364  if (FalseRangeEnd == (int)i-1)
365  FalseRangeEnd = i;
366  else
367  FalseRangeEnd = Overdefined;
368  }
369  }
370 
371  // If this element is in range, update our magic bitvector.
372  if (i < 64 && IsTrueForElt)
373  MagicBitvector |= 1ULL << i;
374 
375  // If all of our states become overdefined, bail out early. Since the
376  // predicate is expensive, only check it every 8 elements. This is only
377  // really useful for really huge arrays.
378  if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
379  SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
380  FalseRangeEnd == Overdefined)
381  return nullptr;
382  }
383 
384  // Now that we've scanned the entire array, emit our new comparison(s). We
385  // order the state machines in complexity of the generated code.
386  Value *Idx = GEP->getOperand(2);
387 
388  // If the index is larger than the pointer size of the target, truncate the
389  // index down like the GEP would do implicitly. We don't have to do this for
390  // an inbounds GEP because the index can't be out of range.
391  if (!GEP->isInBounds()) {
392  Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
393  unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
394  if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
395  Idx = Builder.CreateTrunc(Idx, IntPtrTy);
396  }
397 
398  // If the comparison is only true for one or two elements, emit direct
399  // comparisons.
400  if (SecondTrueElement != Overdefined) {
401  // None true -> false.
402  if (FirstTrueElement == Undefined)
403  return replaceInstUsesWith(ICI, Builder.getFalse());
404 
405  Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
406 
407  // True for one element -> 'i == 47'.
408  if (SecondTrueElement == Undefined)
409  return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
410 
411  // True for two elements -> 'i == 47 | i == 72'.
412  Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
413  Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
414  Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
415  return BinaryOperator::CreateOr(C1, C2);
416  }
417 
418  // If the comparison is only false for one or two elements, emit direct
419  // comparisons.
420  if (SecondFalseElement != Overdefined) {
421  // None false -> true.
422  if (FirstFalseElement == Undefined)
423  return replaceInstUsesWith(ICI, Builder.getTrue());
424 
425  Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
426 
427  // False for one element -> 'i != 47'.
428  if (SecondFalseElement == Undefined)
429  return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
430 
431  // False for two elements -> 'i != 47 & i != 72'.
432  Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
433  Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
434  Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
435  return BinaryOperator::CreateAnd(C1, C2);
436  }
437 
438  // If the comparison can be replaced with a range comparison for the elements
439  // where it is true, emit the range check.
440  if (TrueRangeEnd != Overdefined) {
441  assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
442 
443  // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
444  if (FirstTrueElement) {
445  Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
446  Idx = Builder.CreateAdd(Idx, Offs);
447  }
448 
449  Value *End = ConstantInt::get(Idx->getType(),
450  TrueRangeEnd-FirstTrueElement+1);
451  return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
452  }
453 
454  // False range check.
455  if (FalseRangeEnd != Overdefined) {
456  assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
457  // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
458  if (FirstFalseElement) {
459  Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
460  Idx = Builder.CreateAdd(Idx, Offs);
461  }
462 
463  Value *End = ConstantInt::get(Idx->getType(),
464  FalseRangeEnd-FirstFalseElement);
465  return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
466  }
467 
468  // If a magic bitvector captures the entire comparison state
469  // of this load, replace it with computation that does:
470  // ((magic_cst >> i) & 1) != 0
471  {
472  Type *Ty = nullptr;
473 
474  // Look for an appropriate type:
475  // - The type of Idx if the magic fits
476  // - The smallest fitting legal type if we have a DataLayout
477  // - Default to i32
478  if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
479  Ty = Idx->getType();
480  else
481  Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
482 
483  if (Ty) {
484  Value *V = Builder.CreateIntCast(Idx, Ty, false);
485  V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
486  V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
487  return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
488  }
489  }
490 
491  return nullptr;
492 }
493 
494 /// Return a value that can be used to compare the *offset* implied by a GEP to
495 /// zero. For example, if we have &A[i], we want to return 'i' for
496 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
497 /// are involved. The above expression would also be legal to codegen as
498 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
499 /// This latter form is less amenable to optimization though, and we are allowed
500 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
501 ///
502 /// If we can't emit an optimized form for this expression, this returns null.
503 ///
505  const DataLayout &DL) {
507 
508  // Check to see if this gep only has a single variable index. If so, and if
509  // any constant indices are a multiple of its scale, then we can compute this
510  // in terms of the scale of the variable index. For example, if the GEP
511  // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
512  // because the expression will cross zero at the same point.
513  unsigned i, e = GEP->getNumOperands();
514  int64_t Offset = 0;
515  for (i = 1; i != e; ++i, ++GTI) {
516  if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
517  // Compute the aggregate offset of constant indices.
518  if (CI->isZero()) continue;
519 
520  // Handle a struct index, which adds its field offset to the pointer.
521  if (StructType *STy = GTI.getStructTypeOrNull()) {
522  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
523  } else {
524  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
525  Offset += Size*CI->getSExtValue();
526  }
527  } else {
528  // Found our variable index.
529  break;
530  }
531  }
532 
533  // If there are no variable indices, we must have a constant offset, just
534  // evaluate it the general way.
535  if (i == e) return nullptr;
536 
537  Value *VariableIdx = GEP->getOperand(i);
538  // Determine the scale factor of the variable element. For example, this is
539  // 4 if the variable index is into an array of i32.
540  uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
541 
542  // Verify that there are no other variable indices. If so, emit the hard way.
543  for (++i, ++GTI; i != e; ++i, ++GTI) {
544  ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
545  if (!CI) return nullptr;
546 
547  // Compute the aggregate offset of constant indices.
548  if (CI->isZero()) continue;
549 
550  // Handle a struct index, which adds its field offset to the pointer.
551  if (StructType *STy = GTI.getStructTypeOrNull()) {
552  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
553  } else {
554  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
555  Offset += Size*CI->getSExtValue();
556  }
557  }
558 
559  // Okay, we know we have a single variable index, which must be a
560  // pointer/array/vector index. If there is no offset, life is simple, return
561  // the index.
562  Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
563  unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
564  if (Offset == 0) {
565  // Cast to intptrty in case a truncation occurs. If an extension is needed,
566  // we don't need to bother extending: the extension won't affect where the
567  // computation crosses zero.
568  if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
569  VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
570  }
571  return VariableIdx;
572  }
573 
574  // Otherwise, there is an index. The computation we will do will be modulo
575  // the pointer size, so get it.
576  uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
577 
578  Offset &= PtrSizeMask;
579  VariableScale &= PtrSizeMask;
580 
581  // To do this transformation, any constant index must be a multiple of the
582  // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
583  // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
584  // multiple of the variable scale.
585  int64_t NewOffs = Offset / (int64_t)VariableScale;
586  if (Offset != NewOffs*(int64_t)VariableScale)
587  return nullptr;
588 
589  // Okay, we can do this evaluation. Start by converting the index to intptr.
590  if (VariableIdx->getType() != IntPtrTy)
591  VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
592  true /*Signed*/);
593  Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
594  return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
595 }
596 
597 /// Returns true if we can rewrite Start as a GEP with pointer Base
598 /// and some integer offset. The nodes that need to be re-written
599 /// for this transformation will be added to Explored.
600 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
601  const DataLayout &DL,
602  SetVector<Value *> &Explored) {
603  SmallVector<Value *, 16> WorkList(1, Start);
604  Explored.insert(Base);
605 
606  // The following traversal gives us an order which can be used
607  // when doing the final transformation. Since in the final
608  // transformation we create the PHI replacement instructions first,
609  // we don't have to get them in any particular order.
610  //
611  // However, for other instructions we will have to traverse the
612  // operands of an instruction first, which means that we have to
613  // do a post-order traversal.
614  while (!WorkList.empty()) {
616 
617  while (!WorkList.empty()) {
618  if (Explored.size() >= 100)
619  return false;
620 
621  Value *V = WorkList.back();
622 
623  if (Explored.count(V) != 0) {
624  WorkList.pop_back();
625  continue;
626  }
627 
628  if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
629  !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
630  // We've found some value that we can't explore which is different from
631  // the base. Therefore we can't do this transformation.
632  return false;
633 
634  if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
635  auto *CI = dyn_cast<CastInst>(V);
636  if (!CI->isNoopCast(DL))
637  return false;
638 
639  if (Explored.count(CI->getOperand(0)) == 0)
640  WorkList.push_back(CI->getOperand(0));
641  }
642 
643  if (auto *GEP = dyn_cast<GEPOperator>(V)) {
644  // We're limiting the GEP to having one index. This will preserve
645  // the original pointer type. We could handle more cases in the
646  // future.
647  if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
648  GEP->getType() != Start->getType())
649  return false;
650 
651  if (Explored.count(GEP->getOperand(0)) == 0)
652  WorkList.push_back(GEP->getOperand(0));
653  }
654 
655  if (WorkList.back() == V) {
656  WorkList.pop_back();
657  // We've finished visiting this node, mark it as such.
658  Explored.insert(V);
659  }
660 
661  if (auto *PN = dyn_cast<PHINode>(V)) {
662  // We cannot transform PHIs on unsplittable basic blocks.
663  if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
664  return false;
665  Explored.insert(PN);
666  PHIs.insert(PN);
667  }
668  }
669 
670  // Explore the PHI nodes further.
671  for (auto *PN : PHIs)
672  for (Value *Op : PN->incoming_values())
673  if (Explored.count(Op) == 0)
674  WorkList.push_back(Op);
675  }
676 
677  // Make sure that we can do this. Since we can't insert GEPs in a basic
678  // block before a PHI node, we can't easily do this transformation if
679  // we have PHI node users of transformed instructions.
680  for (Value *Val : Explored) {
681  for (Value *Use : Val->uses()) {
682 
683  auto *PHI = dyn_cast<PHINode>(Use);
684  auto *Inst = dyn_cast<Instruction>(Val);
685 
686  if (Inst == Base || Inst == PHI || !Inst || !PHI ||
687  Explored.count(PHI) == 0)
688  continue;
689 
690  if (PHI->getParent() == Inst->getParent())
691  return false;
692  }
693  }
694  return true;
695 }
696 
697 // Sets the appropriate insert point on Builder where we can add
698 // a replacement Instruction for V (if that is possible).
699 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
700  bool Before = true) {
701  if (auto *PHI = dyn_cast<PHINode>(V)) {
702  Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
703  return;
704  }
705  if (auto *I = dyn_cast<Instruction>(V)) {
706  if (!Before)
707  I = &*std::next(I->getIterator());
708  Builder.SetInsertPoint(I);
709  return;
710  }
711  if (auto *A = dyn_cast<Argument>(V)) {
712  // Set the insertion point in the entry block.
713  BasicBlock &Entry = A->getParent()->getEntryBlock();
714  Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
715  return;
716  }
717  // Otherwise, this is a constant and we don't need to set a new
718  // insertion point.
719  assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
720 }
721 
722 /// Returns a re-written value of Start as an indexed GEP using Base as a
723 /// pointer.
725  const DataLayout &DL,
726  SetVector<Value *> &Explored) {
727  // Perform all the substitutions. This is a bit tricky because we can
728  // have cycles in our use-def chains.
729  // 1. Create the PHI nodes without any incoming values.
730  // 2. Create all the other values.
731  // 3. Add the edges for the PHI nodes.
732  // 4. Emit GEPs to get the original pointers.
733  // 5. Remove the original instructions.
734  Type *IndexType = IntegerType::get(
735  Base->getContext(), DL.getPointerTypeSizeInBits(Start->getType()));
736 
738  NewInsts[Base] = ConstantInt::getNullValue(IndexType);
739 
740  // Create the new PHI nodes, without adding any incoming values.
741  for (Value *Val : Explored) {
742  if (Val == Base)
743  continue;
744  // Create empty phi nodes. This avoids cyclic dependencies when creating
745  // the remaining instructions.
746  if (auto *PHI = dyn_cast<PHINode>(Val))
747  NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
748  PHI->getName() + ".idx", PHI);
749  }
750  IRBuilder<> Builder(Base->getContext());
751 
752  // Create all the other instructions.
753  for (Value *Val : Explored) {
754 
755  if (NewInsts.find(Val) != NewInsts.end())
756  continue;
757 
758  if (auto *CI = dyn_cast<CastInst>(Val)) {
759  NewInsts[CI] = NewInsts[CI->getOperand(0)];
760  continue;
761  }
762  if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
763  Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
764  : GEP->getOperand(1);
765  setInsertionPoint(Builder, GEP);
766  // Indices might need to be sign extended. GEPs will magically do
767  // this, but we need to do it ourselves here.
768  if (Index->getType()->getScalarSizeInBits() !=
769  NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
770  Index = Builder.CreateSExtOrTrunc(
771  Index, NewInsts[GEP->getOperand(0)]->getType(),
772  GEP->getOperand(0)->getName() + ".sext");
773  }
774 
775  auto *Op = NewInsts[GEP->getOperand(0)];
776  if (isa<ConstantInt>(Op) && dyn_cast<ConstantInt>(Op)->isZero())
777  NewInsts[GEP] = Index;
778  else
779  NewInsts[GEP] = Builder.CreateNSWAdd(
780  Op, Index, GEP->getOperand(0)->getName() + ".add");
781  continue;
782  }
783  if (isa<PHINode>(Val))
784  continue;
785 
786  llvm_unreachable("Unexpected instruction type");
787  }
788 
789  // Add the incoming values to the PHI nodes.
790  for (Value *Val : Explored) {
791  if (Val == Base)
792  continue;
793  // All the instructions have been created, we can now add edges to the
794  // phi nodes.
795  if (auto *PHI = dyn_cast<PHINode>(Val)) {
796  PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
797  for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
798  Value *NewIncoming = PHI->getIncomingValue(I);
799 
800  if (NewInsts.find(NewIncoming) != NewInsts.end())
801  NewIncoming = NewInsts[NewIncoming];
802 
803  NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
804  }
805  }
806  }
807 
808  for (Value *Val : Explored) {
809  if (Val == Base)
810  continue;
811 
812  // Depending on the type, for external users we have to emit
813  // a GEP or a GEP + ptrtoint.
814  setInsertionPoint(Builder, Val, false);
815 
816  // If required, create an inttoptr instruction for Base.
817  Value *NewBase = Base;
818  if (!Base->getType()->isPointerTy())
819  NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
820  Start->getName() + "to.ptr");
821 
822  Value *GEP = Builder.CreateInBoundsGEP(
823  Start->getType()->getPointerElementType(), NewBase,
824  makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
825 
826  if (!Val->getType()->isPointerTy()) {
827  Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
828  Val->getName() + ".conv");
829  GEP = Cast;
830  }
831  Val->replaceAllUsesWith(GEP);
832  }
833 
834  return NewInsts[Start];
835 }
836 
837 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
838 /// the input Value as a constant indexed GEP. Returns a pair containing
839 /// the GEPs Pointer and Index.
840 static std::pair<Value *, Value *>
842  Type *IndexType = IntegerType::get(V->getContext(),
844 
845  Constant *Index = ConstantInt::getNullValue(IndexType);
846  while (true) {
847  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
848  // We accept only inbouds GEPs here to exclude the possibility of
849  // overflow.
850  if (!GEP->isInBounds())
851  break;
852  if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
853  GEP->getType() == V->getType()) {
854  V = GEP->getOperand(0);
855  Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
856  Index = ConstantExpr::getAdd(
857  Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
858  continue;
859  }
860  break;
861  }
862  if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
863  if (!CI->isNoopCast(DL))
864  break;
865  V = CI->getOperand(0);
866  continue;
867  }
868  if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
869  if (!CI->isNoopCast(DL))
870  break;
871  V = CI->getOperand(0);
872  continue;
873  }
874  break;
875  }
876  return {V, Index};
877 }
878 
879 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
880 /// We can look through PHIs, GEPs and casts in order to determine a common base
881 /// between GEPLHS and RHS.
883  ICmpInst::Predicate Cond,
884  const DataLayout &DL) {
885  if (!GEPLHS->hasAllConstantIndices())
886  return nullptr;
887 
888  // Make sure the pointers have the same type.
889  if (GEPLHS->getType() != RHS->getType())
890  return nullptr;
891 
892  Value *PtrBase, *Index;
893  std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
894 
895  // The set of nodes that will take part in this transformation.
896  SetVector<Value *> Nodes;
897 
898  if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
899  return nullptr;
900 
901  // We know we can re-write this as
902  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
903  // Since we've only looked through inbouds GEPs we know that we
904  // can't have overflow on either side. We can therefore re-write
905  // this as:
906  // OFFSET1 cmp OFFSET2
907  Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
908 
909  // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
910  // GEP having PtrBase as the pointer base, and has returned in NewRHS the
911  // offset. Since Index is the offset of LHS to the base pointer, we will now
912  // compare the offsets instead of comparing the pointers.
913  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
914 }
915 
916 /// Fold comparisons between a GEP instruction and something else. At this point
917 /// we know that the GEP is on the LHS of the comparison.
918 Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
919  ICmpInst::Predicate Cond,
920  Instruction &I) {
921  // Don't transform signed compares of GEPs into index compares. Even if the
922  // GEP is inbounds, the final add of the base pointer can have signed overflow
923  // and would change the result of the icmp.
924  // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
925  // the maximum signed value for the pointer type.
926  if (ICmpInst::isSigned(Cond))
927  return nullptr;
928 
929  // Look through bitcasts and addrspacecasts. We do not however want to remove
930  // 0 GEPs.
931  if (!isa<GetElementPtrInst>(RHS))
932  RHS = RHS->stripPointerCasts();
933 
934  Value *PtrBase = GEPLHS->getOperand(0);
935  if (PtrBase == RHS && GEPLHS->isInBounds()) {
936  // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
937  // This transformation (ignoring the base and scales) is valid because we
938  // know pointers can't overflow since the gep is inbounds. See if we can
939  // output an optimized form.
940  Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
941 
942  // If not, synthesize the offset the hard way.
943  if (!Offset)
944  Offset = EmitGEPOffset(GEPLHS);
945  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
946  Constant::getNullValue(Offset->getType()));
947  } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
948  // If the base pointers are different, but the indices are the same, just
949  // compare the base pointer.
950  if (PtrBase != GEPRHS->getOperand(0)) {
951  bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
952  IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
953  GEPRHS->getOperand(0)->getType();
954  if (IndicesTheSame)
955  for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
956  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
957  IndicesTheSame = false;
958  break;
959  }
960 
961  // If all indices are the same, just compare the base pointers.
962  if (IndicesTheSame)
963  return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
964 
965  // If we're comparing GEPs with two base pointers that only differ in type
966  // and both GEPs have only constant indices or just one use, then fold
967  // the compare with the adjusted indices.
968  if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
969  (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
970  (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
971  PtrBase->stripPointerCasts() ==
972  GEPRHS->getOperand(0)->stripPointerCasts()) {
973  Value *LOffset = EmitGEPOffset(GEPLHS);
974  Value *ROffset = EmitGEPOffset(GEPRHS);
975 
976  // If we looked through an addrspacecast between different sized address
977  // spaces, the LHS and RHS pointers are different sized
978  // integers. Truncate to the smaller one.
979  Type *LHSIndexTy = LOffset->getType();
980  Type *RHSIndexTy = ROffset->getType();
981  if (LHSIndexTy != RHSIndexTy) {
982  if (LHSIndexTy->getPrimitiveSizeInBits() <
983  RHSIndexTy->getPrimitiveSizeInBits()) {
984  ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
985  } else
986  LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
987  }
988 
989  Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
990  LOffset, ROffset);
991  return replaceInstUsesWith(I, Cmp);
992  }
993 
994  // Otherwise, the base pointers are different and the indices are
995  // different. Try convert this to an indexed compare by looking through
996  // PHIs/casts.
997  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
998  }
999 
1000  // If one of the GEPs has all zero indices, recurse.
1001  if (GEPLHS->hasAllZeroIndices())
1002  return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
1004 
1005  // If the other GEP has all zero indices, recurse.
1006  if (GEPRHS->hasAllZeroIndices())
1007  return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
1008 
1009  bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
1010  if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
1011  // If the GEPs only differ by one index, compare it.
1012  unsigned NumDifferences = 0; // Keep track of # differences.
1013  unsigned DiffOperand = 0; // The operand that differs.
1014  for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
1015  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
1016  if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
1017  GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
1018  // Irreconcilable differences.
1019  NumDifferences = 2;
1020  break;
1021  } else {
1022  if (NumDifferences++) break;
1023  DiffOperand = i;
1024  }
1025  }
1026 
1027  if (NumDifferences == 0) // SAME GEP?
1028  return replaceInstUsesWith(I, // No comparison is needed here.
1029  Builder.getInt1(ICmpInst::isTrueWhenEqual(Cond)));
1030 
1031  else if (NumDifferences == 1 && GEPsInBounds) {
1032  Value *LHSV = GEPLHS->getOperand(DiffOperand);
1033  Value *RHSV = GEPRHS->getOperand(DiffOperand);
1034  // Make sure we do a signed comparison here.
1035  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
1036  }
1037  }
1038 
1039  // Only lower this if the icmp is the only user of the GEP or if we expect
1040  // the result to fold to a constant!
1041  if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
1042  (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
1043  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
1044  Value *L = EmitGEPOffset(GEPLHS);
1045  Value *R = EmitGEPOffset(GEPRHS);
1046  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
1047  }
1048  }
1049 
1050  // Try convert this to an indexed compare by looking through PHIs/casts as a
1051  // last resort.
1052  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1053 }
1054 
1055 Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI,
1056  const AllocaInst *Alloca,
1057  const Value *Other) {
1058  assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1059 
1060  // It would be tempting to fold away comparisons between allocas and any
1061  // pointer not based on that alloca (e.g. an argument). However, even
1062  // though such pointers cannot alias, they can still compare equal.
1063  //
1064  // But LLVM doesn't specify where allocas get their memory, so if the alloca
1065  // doesn't escape we can argue that it's impossible to guess its value, and we
1066  // can therefore act as if any such guesses are wrong.
1067  //
1068  // The code below checks that the alloca doesn't escape, and that it's only
1069  // used in a comparison once (the current instruction). The
1070  // single-comparison-use condition ensures that we're trivially folding all
1071  // comparisons against the alloca consistently, and avoids the risk of
1072  // erroneously folding a comparison of the pointer with itself.
1073 
1074  unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1075 
1077  for (const Use &U : Alloca->uses()) {
1078  if (Worklist.size() >= MaxIter)
1079  return nullptr;
1080  Worklist.push_back(&U);
1081  }
1082 
1083  unsigned NumCmps = 0;
1084  while (!Worklist.empty()) {
1085  assert(Worklist.size() <= MaxIter);
1086  const Use *U = Worklist.pop_back_val();
1087  const Value *V = U->getUser();
1088  --MaxIter;
1089 
1090  if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1091  isa<SelectInst>(V)) {
1092  // Track the uses.
1093  } else if (isa<LoadInst>(V)) {
1094  // Loading from the pointer doesn't escape it.
1095  continue;
1096  } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1097  // Storing *to* the pointer is fine, but storing the pointer escapes it.
1098  if (SI->getValueOperand() == U->get())
1099  return nullptr;
1100  continue;
1101  } else if (isa<ICmpInst>(V)) {
1102  if (NumCmps++)
1103  return nullptr; // Found more than one cmp.
1104  continue;
1105  } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1106  switch (Intrin->getIntrinsicID()) {
1107  // These intrinsics don't escape or compare the pointer. Memset is safe
1108  // because we don't allow ptrtoint. Memcpy and memmove are safe because
1109  // we don't allow stores, so src cannot point to V.
1110  case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1111  case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1112  continue;
1113  default:
1114  return nullptr;
1115  }
1116  } else {
1117  return nullptr;
1118  }
1119  for (const Use &U : V->uses()) {
1120  if (Worklist.size() >= MaxIter)
1121  return nullptr;
1122  Worklist.push_back(&U);
1123  }
1124  }
1125 
1126  Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1127  return replaceInstUsesWith(
1128  ICI,
1130 }
1131 
1132 /// Fold "icmp pred (X+CI), X".
1133 Instruction *InstCombiner::foldICmpAddOpConst(Value *X, ConstantInt *CI,
1134  ICmpInst::Predicate Pred) {
1135  // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1136  // so the values can never be equal. Similarly for all other "or equals"
1137  // operators.
1138 
1139  // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1140  // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1141  // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1142  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1143  Value *R =
1145  return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1146  }
1147 
1148  // (X+1) >u X --> X <u (0-1) --> X != 255
1149  // (X+2) >u X --> X <u (0-2) --> X <u 254
1150  // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1151  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1152  return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
1153 
1154  unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
1156  APInt::getSignedMaxValue(BitWidth));
1157 
1158  // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1159  // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1160  // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1161  // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1162  // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1163  // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1164  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1165  return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
1166 
1167  // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1168  // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1169  // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1170  // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1171  // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1172  // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1173 
1174  assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1175  Constant *C = Builder.getInt(CI->getValue() - 1);
1176  return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
1177 }
1178 
1179 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1180 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1181 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1182 Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A,
1183  const APInt &AP1,
1184  const APInt &AP2) {
1185  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1186 
1187  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1188  if (I.getPredicate() == I.ICMP_NE)
1189  Pred = CmpInst::getInversePredicate(Pred);
1190  return new ICmpInst(Pred, LHS, RHS);
1191  };
1192 
1193  // Don't bother doing any work for cases which InstSimplify handles.
1194  if (AP2.isNullValue())
1195  return nullptr;
1196 
1197  bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1198  if (IsAShr) {
1199  if (AP2.isAllOnesValue())
1200  return nullptr;
1201  if (AP2.isNegative() != AP1.isNegative())
1202  return nullptr;
1203  if (AP2.sgt(AP1))
1204  return nullptr;
1205  }
1206 
1207  if (!AP1)
1208  // 'A' must be large enough to shift out the highest set bit.
1209  return getICmp(I.ICMP_UGT, A,
1210  ConstantInt::get(A->getType(), AP2.logBase2()));
1211 
1212  if (AP1 == AP2)
1213  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1214 
1215  int Shift;
1216  if (IsAShr && AP1.isNegative())
1217  Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1218  else
1219  Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1220 
1221  if (Shift > 0) {
1222  if (IsAShr && AP1 == AP2.ashr(Shift)) {
1223  // There are multiple solutions if we are comparing against -1 and the LHS
1224  // of the ashr is not a power of two.
1225  if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1226  return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1227  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1228  } else if (AP1 == AP2.lshr(Shift)) {
1229  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1230  }
1231  }
1232 
1233  // Shifting const2 will never be equal to const1.
1234  // FIXME: This should always be handled by InstSimplify?
1235  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1236  return replaceInstUsesWith(I, TorF);
1237 }
1238 
1239 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1240 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1241 Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A,
1242  const APInt &AP1,
1243  const APInt &AP2) {
1244  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1245 
1246  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1247  if (I.getPredicate() == I.ICMP_NE)
1248  Pred = CmpInst::getInversePredicate(Pred);
1249  return new ICmpInst(Pred, LHS, RHS);
1250  };
1251 
1252  // Don't bother doing any work for cases which InstSimplify handles.
1253  if (AP2.isNullValue())
1254  return nullptr;
1255 
1256  unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1257 
1258  if (!AP1 && AP2TrailingZeros != 0)
1259  return getICmp(
1260  I.ICMP_UGE, A,
1261  ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1262 
1263  if (AP1 == AP2)
1264  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1265 
1266  // Get the distance between the lowest bits that are set.
1267  int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1268 
1269  if (Shift > 0 && AP2.shl(Shift) == AP1)
1270  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1271 
1272  // Shifting const2 will never be equal to const1.
1273  // FIXME: This should always be handled by InstSimplify?
1274  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1275  return replaceInstUsesWith(I, TorF);
1276 }
1277 
1278 /// The caller has matched a pattern of the form:
1279 /// I = icmp ugt (add (add A, B), CI2), CI1
1280 /// If this is of the form:
1281 /// sum = a + b
1282 /// if (sum+128 >u 255)
1283 /// Then replace it with llvm.sadd.with.overflow.i8.
1284 ///
1286  ConstantInt *CI2, ConstantInt *CI1,
1287  InstCombiner &IC) {
1288  // The transformation we're trying to do here is to transform this into an
1289  // llvm.sadd.with.overflow. To do this, we have to replace the original add
1290  // with a narrower add, and discard the add-with-constant that is part of the
1291  // range check (if we can't eliminate it, this isn't profitable).
1292 
1293  // In order to eliminate the add-with-constant, the compare can be its only
1294  // use.
1295  Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1296  if (!AddWithCst->hasOneUse())
1297  return nullptr;
1298 
1299  // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1300  if (!CI2->getValue().isPowerOf2())
1301  return nullptr;
1302  unsigned NewWidth = CI2->getValue().countTrailingZeros();
1303  if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1304  return nullptr;
1305 
1306  // The width of the new add formed is 1 more than the bias.
1307  ++NewWidth;
1308 
1309  // Check to see that CI1 is an all-ones value with NewWidth bits.
1310  if (CI1->getBitWidth() == NewWidth ||
1311  CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1312  return nullptr;
1313 
1314  // This is only really a signed overflow check if the inputs have been
1315  // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1316  // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1317  unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1318  if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1319  IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1320  return nullptr;
1321 
1322  // In order to replace the original add with a narrower
1323  // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1324  // and truncates that discard the high bits of the add. Verify that this is
1325  // the case.
1326  Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1327  for (User *U : OrigAdd->users()) {
1328  if (U == AddWithCst)
1329  continue;
1330 
1331  // Only accept truncates for now. We would really like a nice recursive
1332  // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1333  // chain to see which bits of a value are actually demanded. If the
1334  // original add had another add which was then immediately truncated, we
1335  // could still do the transformation.
1336  TruncInst *TI = dyn_cast<TruncInst>(U);
1337  if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1338  return nullptr;
1339  }
1340 
1341  // If the pattern matches, truncate the inputs to the narrower type and
1342  // use the sadd_with_overflow intrinsic to efficiently compute both the
1343  // result and the overflow bit.
1344  Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1346  Intrinsic::sadd_with_overflow, NewType);
1347 
1348  InstCombiner::BuilderTy &Builder = IC.Builder;
1349 
1350  // Put the new code above the original add, in case there are any uses of the
1351  // add between the add and the compare.
1352  Builder.SetInsertPoint(OrigAdd);
1353 
1354  Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1355  Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1356  CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1357  Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1358  Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1359 
1360  // The inner add was the result of the narrow add, zero extended to the
1361  // wider type. Replace it with the result computed by the intrinsic.
1362  IC.replaceInstUsesWith(*OrigAdd, ZExt);
1363 
1364  // The original icmp gets replaced with the overflow value.
1365  return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1366 }
1367 
1368 // Fold icmp Pred X, C.
1369 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
1370  CmpInst::Predicate Pred = Cmp.getPredicate();
1371  Value *X = Cmp.getOperand(0);
1372 
1373  const APInt *C;
1374  if (!match(Cmp.getOperand(1), m_APInt(C)))
1375  return nullptr;
1376 
1377  Value *A = nullptr, *B = nullptr;
1378 
1379  // Match the following pattern, which is a common idiom when writing
1380  // overflow-safe integer arithmetic functions. The source performs an addition
1381  // in wider type and explicitly checks for overflow using comparisons against
1382  // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1383  //
1384  // TODO: This could probably be generalized to handle other overflow-safe
1385  // operations if we worked out the formulas to compute the appropriate magic
1386  // constants.
1387  //
1388  // sum = a + b
1389  // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1390  {
1391  ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1392  if (Pred == ICmpInst::ICMP_UGT &&
1393  match(X, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1395  Cmp, A, B, CI2, cast<ConstantInt>(Cmp.getOperand(1)), *this))
1396  return Res;
1397  }
1398 
1399  // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1400  if (C->isNullValue() && Pred == ICmpInst::ICMP_SGT) {
1402  if (SPR.Flavor == SPF_SMIN) {
1403  if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1404  return new ICmpInst(Pred, B, Cmp.getOperand(1));
1405  if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1406  return new ICmpInst(Pred, A, Cmp.getOperand(1));
1407  }
1408  }
1409 
1410  // FIXME: Use m_APInt to allow folds for splat constants.
1411  ConstantInt *CI = dyn_cast<ConstantInt>(Cmp.getOperand(1));
1412  if (!CI)
1413  return nullptr;
1414 
1415  // Canonicalize icmp instructions based on dominating conditions.
1416  BasicBlock *Parent = Cmp.getParent();
1417  BasicBlock *Dom = Parent->getSinglePredecessor();
1418  auto *BI = Dom ? dyn_cast<BranchInst>(Dom->getTerminator()) : nullptr;
1419  ICmpInst::Predicate Pred2;
1420  BasicBlock *TrueBB, *FalseBB;
1421  ConstantInt *CI2;
1422  if (BI && match(BI, m_Br(m_ICmp(Pred2, m_Specific(X), m_ConstantInt(CI2)),
1423  TrueBB, FalseBB)) &&
1424  TrueBB != FalseBB) {
1425  ConstantRange CR =
1427  ConstantRange DominatingCR =
1428  (Parent == TrueBB)
1431  CmpInst::getInversePredicate(Pred2), CI2->getValue());
1432  ConstantRange Intersection = DominatingCR.intersectWith(CR);
1433  ConstantRange Difference = DominatingCR.difference(CR);
1434  if (Intersection.isEmptySet())
1435  return replaceInstUsesWith(Cmp, Builder.getFalse());
1436  if (Difference.isEmptySet())
1437  return replaceInstUsesWith(Cmp, Builder.getTrue());
1438 
1439  // If this is a normal comparison, it demands all bits. If it is a sign
1440  // bit comparison, it only demands the sign bit.
1441  bool UnusedBit;
1442  bool IsSignBit = isSignBitCheck(Pred, CI->getValue(), UnusedBit);
1443 
1444  // Canonicalizing a sign bit comparison that gets used in a branch,
1445  // pessimizes codegen by generating branch on zero instruction instead
1446  // of a test and branch. So we avoid canonicalizing in such situations
1447  // because test and branch instruction has better branch displacement
1448  // than compare and branch instruction.
1449  if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1450  return nullptr;
1451 
1452  if (auto *AI = Intersection.getSingleElement())
1453  return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*AI));
1454  if (auto *AD = Difference.getSingleElement())
1455  return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*AD));
1456  }
1457 
1458  return nullptr;
1459 }
1460 
1461 /// Fold icmp (trunc X, Y), C.
1462 Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
1463  TruncInst *Trunc,
1464  const APInt *C) {
1465  ICmpInst::Predicate Pred = Cmp.getPredicate();
1466  Value *X = Trunc->getOperand(0);
1467  if (C->isOneValue() && C->getBitWidth() > 1) {
1468  // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1469  Value *V = nullptr;
1470  if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1471  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1472  ConstantInt::get(V->getType(), 1));
1473  }
1474 
1475  if (Cmp.isEquality() && Trunc->hasOneUse()) {
1476  // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1477  // of the high bits truncated out of x are known.
1478  unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1479  SrcBits = X->getType()->getScalarSizeInBits();
1480  KnownBits Known = computeKnownBits(X, 0, &Cmp);
1481 
1482  // If all the high bits are known, we can do this xform.
1483  if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1484  // Pull in the high bits from known-ones set.
1485  APInt NewRHS = C->zext(SrcBits);
1486  NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1487  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1488  }
1489  }
1490 
1491  return nullptr;
1492 }
1493 
1494 /// Fold icmp (xor X, Y), C.
1495 Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp,
1496  BinaryOperator *Xor,
1497  const APInt *C) {
1498  Value *X = Xor->getOperand(0);
1499  Value *Y = Xor->getOperand(1);
1500  const APInt *XorC;
1501  if (!match(Y, m_APInt(XorC)))
1502  return nullptr;
1503 
1504  // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1505  // fold the xor.
1506  ICmpInst::Predicate Pred = Cmp.getPredicate();
1507  if ((Pred == ICmpInst::ICMP_SLT && C->isNullValue()) ||
1508  (Pred == ICmpInst::ICMP_SGT && C->isAllOnesValue())) {
1509 
1510  // If the sign bit of the XorCst is not set, there is no change to
1511  // the operation, just stop using the Xor.
1512  if (!XorC->isNegative()) {
1513  Cmp.setOperand(0, X);
1514  Worklist.Add(Xor);
1515  return &Cmp;
1516  }
1517 
1518  // Was the old condition true if the operand is positive?
1519  bool isTrueIfPositive = Pred == ICmpInst::ICMP_SGT;
1520 
1521  // If so, the new one isn't.
1522  isTrueIfPositive ^= true;
1523 
1524  Constant *CmpConstant = cast<Constant>(Cmp.getOperand(1));
1525  if (isTrueIfPositive)
1526  return new ICmpInst(ICmpInst::ICMP_SGT, X, SubOne(CmpConstant));
1527  else
1528  return new ICmpInst(ICmpInst::ICMP_SLT, X, AddOne(CmpConstant));
1529  }
1530 
1531  if (Xor->hasOneUse()) {
1532  // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1533  if (!Cmp.isEquality() && XorC->isSignMask()) {
1534  Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1535  : Cmp.getSignedPredicate();
1536  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), *C ^ *XorC));
1537  }
1538 
1539  // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1540  if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1541  Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1542  : Cmp.getSignedPredicate();
1543  Pred = Cmp.getSwappedPredicate(Pred);
1544  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), *C ^ *XorC));
1545  }
1546  }
1547 
1548  // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1549  // iff -C is a power of 2
1550  if (Pred == ICmpInst::ICMP_UGT && *XorC == ~(*C) && (*C + 1).isPowerOf2())
1551  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1552 
1553  // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1554  // iff -C is a power of 2
1555  if (Pred == ICmpInst::ICMP_ULT && *XorC == -(*C) && C->isPowerOf2())
1556  return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
1557 
1558  return nullptr;
1559 }
1560 
1561 /// Fold icmp (and (sh X, Y), C2), C1.
1562 Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
1563  const APInt *C1, const APInt *C2) {
1564  BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1565  if (!Shift || !Shift->isShift())
1566  return nullptr;
1567 
1568  // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1569  // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1570  // code produced by the clang front-end, for bitfield access.
1571  // This seemingly simple opportunity to fold away a shift turns out to be
1572  // rather complicated. See PR17827 for details.
1573  unsigned ShiftOpcode = Shift->getOpcode();
1574  bool IsShl = ShiftOpcode == Instruction::Shl;
1575  const APInt *C3;
1576  if (match(Shift->getOperand(1), m_APInt(C3))) {
1577  bool CanFold = false;
1578  if (ShiftOpcode == Instruction::AShr) {
1579  // There may be some constraints that make this possible, but nothing
1580  // simple has been discovered yet.
1581  CanFold = false;
1582  } else if (ShiftOpcode == Instruction::Shl) {
1583  // For a left shift, we can fold if the comparison is not signed. We can
1584  // also fold a signed comparison if the mask value and comparison value
1585  // are not negative. These constraints may not be obvious, but we can
1586  // prove that they are correct using an SMT solver.
1587  if (!Cmp.isSigned() || (!C2->isNegative() && !C1->isNegative()))
1588  CanFold = true;
1589  } else if (ShiftOpcode == Instruction::LShr) {
1590  // For a logical right shift, we can fold if the comparison is not signed.
1591  // We can also fold a signed comparison if the shifted mask value and the
1592  // shifted comparison value are not negative. These constraints may not be
1593  // obvious, but we can prove that they are correct using an SMT solver.
1594  if (!Cmp.isSigned() ||
1595  (!C2->shl(*C3).isNegative() && !C1->shl(*C3).isNegative()))
1596  CanFold = true;
1597  }
1598 
1599  if (CanFold) {
1600  APInt NewCst = IsShl ? C1->lshr(*C3) : C1->shl(*C3);
1601  APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3);
1602  // Check to see if we are shifting out any of the bits being compared.
1603  if (SameAsC1 != *C1) {
1604  // If we shifted bits out, the fold is not going to work out. As a
1605  // special case, check to see if this means that the result is always
1606  // true or false now.
1607  if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1608  return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1609  if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1610  return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1611  } else {
1612  Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst));
1613  APInt NewAndCst = IsShl ? C2->lshr(*C3) : C2->shl(*C3);
1614  And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst));
1615  And->setOperand(0, Shift->getOperand(0));
1616  Worklist.Add(Shift); // Shift is dead.
1617  return &Cmp;
1618  }
1619  }
1620  }
1621 
1622  // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1623  // preferable because it allows the C2 << Y expression to be hoisted out of a
1624  // loop if Y is invariant and X is not.
1625  if (Shift->hasOneUse() && C1->isNullValue() && Cmp.isEquality() &&
1626  !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1627  // Compute C2 << Y.
1628  Value *NewShift =
1629  IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1630  : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1631 
1632  // Compute X & (C2 << Y).
1633  Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1634  Cmp.setOperand(0, NewAnd);
1635  return &Cmp;
1636  }
1637 
1638  return nullptr;
1639 }
1640 
1641 /// Fold icmp (and X, C2), C1.
1642 Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp,
1643  BinaryOperator *And,
1644  const APInt *C1) {
1645  const APInt *C2;
1646  if (!match(And->getOperand(1), m_APInt(C2)))
1647  return nullptr;
1648 
1649  if (!And->hasOneUse() || !And->getOperand(0)->hasOneUse())
1650  return nullptr;
1651 
1652  // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1653  // the input width without changing the value produced, eliminate the cast:
1654  //
1655  // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1656  //
1657  // We can do this transformation if the constants do not have their sign bits
1658  // set or if it is an equality comparison. Extending a relational comparison
1659  // when we're checking the sign bit would not work.
1660  Value *W;
1661  if (match(And->getOperand(0), m_Trunc(m_Value(W))) &&
1662  (Cmp.isEquality() || (!C1->isNegative() && !C2->isNegative()))) {
1663  // TODO: Is this a good transform for vectors? Wider types may reduce
1664  // throughput. Should this transform be limited (even for scalars) by using
1665  // shouldChangeType()?
1666  if (!Cmp.getType()->isVectorTy()) {
1667  Type *WideType = W->getType();
1668  unsigned WideScalarBits = WideType->getScalarSizeInBits();
1669  Constant *ZextC1 = ConstantInt::get(WideType, C1->zext(WideScalarBits));
1670  Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1671  Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1672  return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1673  }
1674  }
1675 
1676  if (Instruction *I = foldICmpAndShift(Cmp, And, C1, C2))
1677  return I;
1678 
1679  // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1680  // (icmp pred (and A, (or (shl 1, B), 1), 0))
1681  //
1682  // iff pred isn't signed
1683  if (!Cmp.isSigned() && C1->isNullValue() &&
1684  match(And->getOperand(1), m_One())) {
1685  Constant *One = cast<Constant>(And->getOperand(1));
1686  Value *Or = And->getOperand(0);
1687  Value *A, *B, *LShr;
1688  if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1689  match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1690  unsigned UsesRemoved = 0;
1691  if (And->hasOneUse())
1692  ++UsesRemoved;
1693  if (Or->hasOneUse())
1694  ++UsesRemoved;
1695  if (LShr->hasOneUse())
1696  ++UsesRemoved;
1697 
1698  // Compute A & ((1 << B) | 1)
1699  Value *NewOr = nullptr;
1700  if (auto *C = dyn_cast<Constant>(B)) {
1701  if (UsesRemoved >= 1)
1702  NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1703  } else {
1704  if (UsesRemoved >= 3)
1705  NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1706  /*HasNUW=*/true),
1707  One, Or->getName());
1708  }
1709  if (NewOr) {
1710  Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1711  Cmp.setOperand(0, NewAnd);
1712  return &Cmp;
1713  }
1714  }
1715  }
1716 
1717  // (X & C2) > C1 --> (X & C2) != 0, if any bit set in (X & C2) will produce a
1718  // result greater than C1. Also handle (X & C2) < C1 --> (X & C2) == 0.
1719  if (!C2->isNullValue()) {
1720  unsigned NumTZ = C2->countTrailingZeros();
1721  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT &&
1722  NumTZ >= C1->getActiveBits()) {
1723  Constant *Zero = Constant::getNullValue(And->getType());
1724  return new ICmpInst(ICmpInst::ICMP_NE, And, Zero);
1725  }
1726  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
1727  NumTZ >= C1->ceilLogBase2()) {
1728  Constant *Zero = Constant::getNullValue(And->getType());
1729  return new ICmpInst(ICmpInst::ICMP_EQ, And, Zero);
1730  }
1731  }
1732 
1733  return nullptr;
1734 }
1735 
1736 /// Fold icmp (and X, Y), C.
1737 Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp,
1738  BinaryOperator *And,
1739  const APInt *C) {
1740  if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1741  return I;
1742 
1743  // TODO: These all require that Y is constant too, so refactor with the above.
1744 
1745  // Try to optimize things like "A[i] & 42 == 0" to index computations.
1746  Value *X = And->getOperand(0);
1747  Value *Y = And->getOperand(1);
1748  if (auto *LI = dyn_cast<LoadInst>(X))
1749  if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1750  if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1751  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1752  !LI->isVolatile() && isa<ConstantInt>(Y)) {
1753  ConstantInt *C2 = cast<ConstantInt>(Y);
1754  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1755  return Res;
1756  }
1757 
1758  if (!Cmp.isEquality())
1759  return nullptr;
1760 
1761  // X & -C == -C -> X > u ~C
1762  // X & -C != -C -> X <= u ~C
1763  // iff C is a power of 2
1764  if (Cmp.getOperand(1) == Y && (-(*C)).isPowerOf2()) {
1765  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1767  return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1768  }
1769 
1770  // (X & C2) == 0 -> (trunc X) >= 0
1771  // (X & C2) != 0 -> (trunc X) < 0
1772  // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1773  const APInt *C2;
1774  if (And->hasOneUse() && C->isNullValue() && match(Y, m_APInt(C2))) {
1775  int32_t ExactLogBase2 = C2->exactLogBase2();
1776  if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1777  Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1778  if (And->getType()->isVectorTy())
1779  NTy = VectorType::get(NTy, And->getType()->getVectorNumElements());
1780  Value *Trunc = Builder.CreateTrunc(X, NTy);
1781  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1783  return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1784  }
1785  }
1786 
1787  return nullptr;
1788 }
1789 
1790 /// Fold icmp (or X, Y), C.
1791 Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
1792  const APInt *C) {
1793  ICmpInst::Predicate Pred = Cmp.getPredicate();
1794  if (C->isOneValue()) {
1795  // icmp slt signum(V) 1 --> icmp slt V, 1
1796  Value *V = nullptr;
1797  if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1798  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1799  ConstantInt::get(V->getType(), 1));
1800  }
1801 
1802  // X | C == C --> X <=u C
1803  // X | C != C --> X >u C
1804  // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1805  if (Cmp.isEquality() && Cmp.getOperand(1) == Or->getOperand(1) &&
1806  (*C + 1).isPowerOf2()) {
1808  return new ICmpInst(Pred, Or->getOperand(0), Or->getOperand(1));
1809  }
1810 
1811  if (!Cmp.isEquality() || !C->isNullValue() || !Or->hasOneUse())
1812  return nullptr;
1813 
1814  Value *P, *Q;
1815  if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1816  // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1817  // -> and (icmp eq P, null), (icmp eq Q, null).
1818  Value *CmpP =
1819  Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1820  Value *CmpQ =
1821  Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1822  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1823  return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1824  }
1825 
1826  // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1827  // a shorter form that has more potential to be folded even further.
1828  Value *X1, *X2, *X3, *X4;
1829  if (match(Or->getOperand(0), m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1830  match(Or->getOperand(1), m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1831  // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1832  // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1833  Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1834  Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1835  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1836  return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1837  }
1838 
1839  return nullptr;
1840 }
1841 
1842 /// Fold icmp (mul X, Y), C.
1843 Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp,
1844  BinaryOperator *Mul,
1845  const APInt *C) {
1846  const APInt *MulC;
1847  if (!match(Mul->getOperand(1), m_APInt(MulC)))
1848  return nullptr;
1849 
1850  // If this is a test of the sign bit and the multiply is sign-preserving with
1851  // a constant operand, use the multiply LHS operand instead.
1852  ICmpInst::Predicate Pred = Cmp.getPredicate();
1853  if (isSignTest(Pred, *C) && Mul->hasNoSignedWrap()) {
1854  if (MulC->isNegative())
1855  Pred = ICmpInst::getSwappedPredicate(Pred);
1856  return new ICmpInst(Pred, Mul->getOperand(0),
1858  }
1859 
1860  return nullptr;
1861 }
1862 
1863 /// Fold icmp (shl 1, Y), C.
1865  const APInt *C) {
1866  Value *Y;
1867  if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1868  return nullptr;
1869 
1870  Type *ShiftType = Shl->getType();
1871  uint32_t TypeBits = C->getBitWidth();
1872  bool CIsPowerOf2 = C->isPowerOf2();
1873  ICmpInst::Predicate Pred = Cmp.getPredicate();
1874  if (Cmp.isUnsigned()) {
1875  // (1 << Y) pred C -> Y pred Log2(C)
1876  if (!CIsPowerOf2) {
1877  // (1 << Y) < 30 -> Y <= 4
1878  // (1 << Y) <= 30 -> Y <= 4
1879  // (1 << Y) >= 30 -> Y > 4
1880  // (1 << Y) > 30 -> Y > 4
1881  if (Pred == ICmpInst::ICMP_ULT)
1882  Pred = ICmpInst::ICMP_ULE;
1883  else if (Pred == ICmpInst::ICMP_UGE)
1884  Pred = ICmpInst::ICMP_UGT;
1885  }
1886 
1887  // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
1888  // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
1889  unsigned CLog2 = C->logBase2();
1890  if (CLog2 == TypeBits - 1) {
1891  if (Pred == ICmpInst::ICMP_UGE)
1892  Pred = ICmpInst::ICMP_EQ;
1893  else if (Pred == ICmpInst::ICMP_ULT)
1894  Pred = ICmpInst::ICMP_NE;
1895  }
1896  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
1897  } else if (Cmp.isSigned()) {
1898  Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
1899  if (C->isAllOnesValue()) {
1900  // (1 << Y) <= -1 -> Y == 31
1901  if (Pred == ICmpInst::ICMP_SLE)
1902  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1903 
1904  // (1 << Y) > -1 -> Y != 31
1905  if (Pred == ICmpInst::ICMP_SGT)
1906  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1907  } else if (!(*C)) {
1908  // (1 << Y) < 0 -> Y == 31
1909  // (1 << Y) <= 0 -> Y == 31
1910  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1911  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1912 
1913  // (1 << Y) >= 0 -> Y != 31
1914  // (1 << Y) > 0 -> Y != 31
1915  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1916  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1917  }
1918  } else if (Cmp.isEquality() && CIsPowerOf2) {
1919  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C->logBase2()));
1920  }
1921 
1922  return nullptr;
1923 }
1924 
1925 /// Fold icmp (shl X, Y), C.
1926 Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp,
1927  BinaryOperator *Shl,
1928  const APInt *C) {
1929  const APInt *ShiftVal;
1930  if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
1931  return foldICmpShlConstConst(Cmp, Shl->getOperand(1), *C, *ShiftVal);
1932 
1933  const APInt *ShiftAmt;
1934  if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
1935  return foldICmpShlOne(Cmp, Shl, C);
1936 
1937  // Check that the shift amount is in range. If not, don't perform undefined
1938  // shifts. When the shift is visited, it will be simplified.
1939  unsigned TypeBits = C->getBitWidth();
1940  if (ShiftAmt->uge(TypeBits))
1941  return nullptr;
1942 
1943  ICmpInst::Predicate Pred = Cmp.getPredicate();
1944  Value *X = Shl->getOperand(0);
1945  Type *ShType = Shl->getType();
1946 
1947  // NSW guarantees that we are only shifting out sign bits from the high bits,
1948  // so we can ASHR the compare constant without needing a mask and eliminate
1949  // the shift.
1950  if (Shl->hasNoSignedWrap()) {
1951  if (Pred == ICmpInst::ICMP_SGT) {
1952  // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
1953  APInt ShiftedC = C->ashr(*ShiftAmt);
1954  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1955  }
1956  if (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) {
1957  // This is the same code as the SGT case, but assert the pre-condition
1958  // that is needed for this to work with equality predicates.
1959  assert(C->ashr(*ShiftAmt).shl(*ShiftAmt) == *C &&
1960  "Compare known true or false was not folded");
1961  APInt ShiftedC = C->ashr(*ShiftAmt);
1962  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1963  }
1964  if (Pred == ICmpInst::ICMP_SLT) {
1965  // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
1966  // (X << S) <=s C is equiv to X <=s (C >> S) for all C
1967  // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
1968  // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
1969  assert(!C->isMinSignedValue() && "Unexpected icmp slt");
1970  APInt ShiftedC = (*C - 1).ashr(*ShiftAmt) + 1;
1971  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1972  }
1973  // If this is a signed comparison to 0 and the shift is sign preserving,
1974  // use the shift LHS operand instead; isSignTest may change 'Pred', so only
1975  // do that if we're sure to not continue on in this function.
1976  if (isSignTest(Pred, *C))
1977  return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
1978  }
1979 
1980  // NUW guarantees that we are only shifting out zero bits from the high bits,
1981  // so we can LSHR the compare constant without needing a mask and eliminate
1982  // the shift.
1983  if (Shl->hasNoUnsignedWrap()) {
1984  if (Pred == ICmpInst::ICMP_UGT) {
1985  // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
1986  APInt ShiftedC = C->lshr(*ShiftAmt);
1987  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1988  }
1989  if (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) {
1990  // This is the same code as the UGT case, but assert the pre-condition
1991  // that is needed for this to work with equality predicates.
1992  assert(C->lshr(*ShiftAmt).shl(*ShiftAmt) == *C &&
1993  "Compare known true or false was not folded");
1994  APInt ShiftedC = C->lshr(*ShiftAmt);
1995  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1996  }
1997  if (Pred == ICmpInst::ICMP_ULT) {
1998  // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
1999  // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2000  // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2001  // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2002  assert(C->ugt(0) && "ult 0 should have been eliminated");
2003  APInt ShiftedC = (*C - 1).lshr(*ShiftAmt) + 1;
2004  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2005  }
2006  }
2007 
2008  if (Cmp.isEquality() && Shl->hasOneUse()) {
2009  // Strength-reduce the shift into an 'and'.
2011  ShType,
2012  APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2013  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2014  Constant *LShrC = ConstantInt::get(ShType, C->lshr(*ShiftAmt));
2015  return new ICmpInst(Pred, And, LShrC);
2016  }
2017 
2018  // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2019  bool TrueIfSigned = false;
2020  if (Shl->hasOneUse() && isSignBitCheck(Pred, *C, TrueIfSigned)) {
2021  // (X << 31) <s 0 --> (X & 1) != 0
2023  ShType,
2024  APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2025  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2026  return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2027  And, Constant::getNullValue(ShType));
2028  }
2029 
2030  // Transform (icmp pred iM (shl iM %v, N), C)
2031  // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2032  // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2033  // This enables us to get rid of the shift in favor of a trunc that may be
2034  // free on the target. It has the additional benefit of comparing to a
2035  // smaller constant that may be more target-friendly.
2036  unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2037  if (Shl->hasOneUse() && Amt != 0 && C->countTrailingZeros() >= Amt &&
2038  DL.isLegalInteger(TypeBits - Amt)) {
2039  Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2040  if (ShType->isVectorTy())
2041  TruncTy = VectorType::get(TruncTy, ShType->getVectorNumElements());
2042  Constant *NewC =
2043  ConstantInt::get(TruncTy, C->ashr(*ShiftAmt).trunc(TypeBits - Amt));
2044  return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2045  }
2046 
2047  return nullptr;
2048 }
2049 
2050 /// Fold icmp ({al}shr X, Y), C.
2051 Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp,
2052  BinaryOperator *Shr,
2053  const APInt *C) {
2054  // An exact shr only shifts out zero bits, so:
2055  // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2056  Value *X = Shr->getOperand(0);
2057  CmpInst::Predicate Pred = Cmp.getPredicate();
2058  if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2059  C->isNullValue())
2060  return new ICmpInst(Pred, X, Cmp.getOperand(1));
2061 
2062  const APInt *ShiftVal;
2063  if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2064  return foldICmpShrConstConst(Cmp, Shr->getOperand(1), *C, *ShiftVal);
2065 
2066  const APInt *ShiftAmt;
2067  if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2068  return nullptr;
2069 
2070  // Check that the shift amount is in range. If not, don't perform undefined
2071  // shifts. When the shift is visited it will be simplified.
2072  unsigned TypeBits = C->getBitWidth();
2073  unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2074  if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2075  return nullptr;
2076 
2077  bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2078  if (!Cmp.isEquality()) {
2079  // If we have an unsigned comparison and an ashr, we can't simplify this.
2080  // Similarly for signed comparisons with lshr.
2081  if (Cmp.isSigned() != IsAShr)
2082  return nullptr;
2083 
2084  // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
2085  // by a power of 2. Since we already have logic to simplify these,
2086  // transform to div and then simplify the resultant comparison.
2087  if (IsAShr && (!Shr->isExact() || ShAmtVal == TypeBits - 1))
2088  return nullptr;
2089 
2090  // Revisit the shift (to delete it).
2091  Worklist.Add(Shr);
2092 
2093  Constant *DivCst = ConstantInt::get(
2094  Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
2095 
2096  Value *Tmp = IsAShr ? Builder.CreateSDiv(X, DivCst, "", Shr->isExact())
2097  : Builder.CreateUDiv(X, DivCst, "", Shr->isExact());
2098 
2099  Cmp.setOperand(0, Tmp);
2100 
2101  // If the builder folded the binop, just return it.
2102  BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
2103  if (!TheDiv)
2104  return &Cmp;
2105 
2106  // Otherwise, fold this div/compare.
2107  assert(TheDiv->getOpcode() == Instruction::SDiv ||
2108  TheDiv->getOpcode() == Instruction::UDiv);
2109 
2110  Instruction *Res = foldICmpDivConstant(Cmp, TheDiv, C);
2111  assert(Res && "This div/cst should have folded!");
2112  return Res;
2113  }
2114 
2115  // Handle equality comparisons of shift-by-constant.
2116 
2117  // If the comparison constant changes with the shift, the comparison cannot
2118  // succeed (bits of the comparison constant cannot match the shifted value).
2119  // This should be known by InstSimplify and already be folded to true/false.
2120  assert(((IsAShr && C->shl(ShAmtVal).ashr(ShAmtVal) == *C) ||
2121  (!IsAShr && C->shl(ShAmtVal).lshr(ShAmtVal) == *C)) &&
2122  "Expected icmp+shr simplify did not occur.");
2123 
2124  // Check if the bits shifted out are known to be zero. If so, we can compare
2125  // against the unshifted value:
2126  // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2127  Constant *ShiftedCmpRHS = ConstantInt::get(Shr->getType(), *C << ShAmtVal);
2128  if (Shr->hasOneUse()) {
2129  if (Shr->isExact())
2130  return new ICmpInst(Pred, X, ShiftedCmpRHS);
2131 
2132  // Otherwise strength reduce the shift into an 'and'.
2133  APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2134  Constant *Mask = ConstantInt::get(Shr->getType(), Val);
2135  Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2136  return new ICmpInst(Pred, And, ShiftedCmpRHS);
2137  }
2138 
2139  return nullptr;
2140 }
2141 
2142 /// Fold icmp (udiv X, Y), C.
2143 Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp,
2144  BinaryOperator *UDiv,
2145  const APInt *C) {
2146  const APInt *C2;
2147  if (!match(UDiv->getOperand(0), m_APInt(C2)))
2148  return nullptr;
2149 
2150  assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2151 
2152  // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2153  Value *Y = UDiv->getOperand(1);
2154  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2155  assert(!C->isMaxValue() &&
2156  "icmp ugt X, UINT_MAX should have been simplified already.");
2157  return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2158  ConstantInt::get(Y->getType(), C2->udiv(*C + 1)));
2159  }
2160 
2161  // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2162  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2163  assert(*C != 0 && "icmp ult X, 0 should have been simplified already.");
2164  return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2165  ConstantInt::get(Y->getType(), C2->udiv(*C)));
2166  }
2167 
2168  return nullptr;
2169 }
2170 
2171 /// Fold icmp ({su}div X, Y), C.
2172 Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp,
2173  BinaryOperator *Div,
2174  const APInt *C) {
2175  // Fold: icmp pred ([us]div X, C2), C -> range test
2176  // Fold this div into the comparison, producing a range check.
2177  // Determine, based on the divide type, what the range is being
2178  // checked. If there is an overflow on the low or high side, remember
2179  // it, otherwise compute the range [low, hi) bounding the new value.
2180  // See: InsertRangeTest above for the kinds of replacements possible.
2181  const APInt *C2;
2182  if (!match(Div->getOperand(1), m_APInt(C2)))
2183  return nullptr;
2184 
2185  // FIXME: If the operand types don't match the type of the divide
2186  // then don't attempt this transform. The code below doesn't have the
2187  // logic to deal with a signed divide and an unsigned compare (and
2188  // vice versa). This is because (x /s C2) <s C produces different
2189  // results than (x /s C2) <u C or (x /u C2) <s C or even
2190  // (x /u C2) <u C. Simply casting the operands and result won't
2191  // work. :( The if statement below tests that condition and bails
2192  // if it finds it.
2193  bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2194  if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2195  return nullptr;
2196 
2197  // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2198  // INT_MIN will also fail if the divisor is 1. Although folds of all these
2199  // division-by-constant cases should be present, we can not assert that they
2200  // have happened before we reach this icmp instruction.
2201  if (C2->isNullValue() || C2->isOneValue() ||
2202  (DivIsSigned && C2->isAllOnesValue()))
2203  return nullptr;
2204 
2205  // TODO: We could do all of the computations below using APInt.
2206  Constant *CmpRHS = cast<Constant>(Cmp.getOperand(1));
2207  Constant *DivRHS = cast<Constant>(Div->getOperand(1));
2208 
2209  // Compute Prod = CmpRHS * DivRHS. We are essentially solving an equation of
2210  // form X / C2 = C. We solve for X by multiplying C2 (DivRHS) and C (CmpRHS).
2211  // By solving for X, we can turn this into a range check instead of computing
2212  // a divide.
2213  Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
2214 
2215  // Determine if the product overflows by seeing if the product is not equal to
2216  // the divide. Make sure we do the same kind of divide as in the LHS
2217  // instruction that we're folding.
2218  bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS)
2219  : ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
2220 
2221  ICmpInst::Predicate Pred = Cmp.getPredicate();
2222 
2223  // If the division is known to be exact, then there is no remainder from the
2224  // divide, so the covered range size is unit, otherwise it is the divisor.
2225  Constant *RangeSize =
2226  Div->isExact() ? ConstantInt::get(Div->getType(), 1) : DivRHS;
2227 
2228  // Figure out the interval that is being checked. For example, a comparison
2229  // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2230  // Compute this interval based on the constants involved and the signedness of
2231  // the compare/divide. This computes a half-open interval, keeping track of
2232  // whether either value in the interval overflows. After analysis each
2233  // overflow variable is set to 0 if it's corresponding bound variable is valid
2234  // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2235  int LoOverflow = 0, HiOverflow = 0;
2236  Constant *LoBound = nullptr, *HiBound = nullptr;
2237 
2238  if (!DivIsSigned) { // udiv
2239  // e.g. X/5 op 3 --> [15, 20)
2240  LoBound = Prod;
2241  HiOverflow = LoOverflow = ProdOV;
2242  if (!HiOverflow) {
2243  // If this is not an exact divide, then many values in the range collapse
2244  // to the same result value.
2245  HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2246  }
2247  } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2248  if (C->isNullValue()) { // (X / pos) op 0
2249  // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2250  LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
2251  HiBound = RangeSize;
2252  } else if (C->isStrictlyPositive()) { // (X / pos) op pos
2253  LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2254  HiOverflow = LoOverflow = ProdOV;
2255  if (!HiOverflow)
2256  HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2257  } else { // (X / pos) op neg
2258  // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2259  HiBound = AddOne(Prod);
2260  LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2261  if (!LoOverflow) {
2262  Constant *DivNeg = ConstantExpr::getNeg(RangeSize);
2263  LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2264  }
2265  }
2266  } else if (C2->isNegative()) { // Divisor is < 0.
2267  if (Div->isExact())
2268  RangeSize = ConstantExpr::getNeg(RangeSize);
2269  if (C->isNullValue()) { // (X / neg) op 0
2270  // e.g. X/-5 op 0 --> [-4, 5)
2271  LoBound = AddOne(RangeSize);
2272  HiBound = ConstantExpr::getNeg(RangeSize);
2273  if (HiBound == DivRHS) { // -INTMIN = INTMIN
2274  HiOverflow = 1; // [INTMIN+1, overflow)
2275  HiBound = nullptr; // e.g. X/INTMIN = 0 --> X > INTMIN
2276  }
2277  } else if (C->isStrictlyPositive()) { // (X / neg) op pos
2278  // e.g. X/-5 op 3 --> [-19, -14)
2279  HiBound = AddOne(Prod);
2280  HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2281  if (!LoOverflow)
2282  LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2283  } else { // (X / neg) op neg
2284  LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2285  LoOverflow = HiOverflow = ProdOV;
2286  if (!HiOverflow)
2287  HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2288  }
2289 
2290  // Dividing by a negative swaps the condition. LT <-> GT
2291  Pred = ICmpInst::getSwappedPredicate(Pred);
2292  }
2293 
2294  Value *X = Div->getOperand(0);
2295  switch (Pred) {
2296  default: llvm_unreachable("Unhandled icmp opcode!");
2297  case ICmpInst::ICMP_EQ:
2298  if (LoOverflow && HiOverflow)
2299  return replaceInstUsesWith(Cmp, Builder.getFalse());
2300  if (HiOverflow)
2301  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2302  ICmpInst::ICMP_UGE, X, LoBound);
2303  if (LoOverflow)
2304  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2305  ICmpInst::ICMP_ULT, X, HiBound);
2306  return replaceInstUsesWith(
2307  Cmp, insertRangeTest(X, LoBound->getUniqueInteger(),
2308  HiBound->getUniqueInteger(), DivIsSigned, true));
2309  case ICmpInst::ICMP_NE:
2310  if (LoOverflow && HiOverflow)
2311  return replaceInstUsesWith(Cmp, Builder.getTrue());
2312  if (HiOverflow)
2313  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2314  ICmpInst::ICMP_ULT, X, LoBound);
2315  if (LoOverflow)
2316  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2317  ICmpInst::ICMP_UGE, X, HiBound);
2318  return replaceInstUsesWith(Cmp,
2319  insertRangeTest(X, LoBound->getUniqueInteger(),
2320  HiBound->getUniqueInteger(),
2321  DivIsSigned, false));
2322  case ICmpInst::ICMP_ULT:
2323  case ICmpInst::ICMP_SLT:
2324  if (LoOverflow == +1) // Low bound is greater than input range.
2325  return replaceInstUsesWith(Cmp, Builder.getTrue());
2326  if (LoOverflow == -1) // Low bound is less than input range.
2327  return replaceInstUsesWith(Cmp, Builder.getFalse());
2328  return new ICmpInst(Pred, X, LoBound);
2329  case ICmpInst::ICMP_UGT:
2330  case ICmpInst::ICMP_SGT:
2331  if (HiOverflow == +1) // High bound greater than input range.
2332  return replaceInstUsesWith(Cmp, Builder.getFalse());
2333  if (HiOverflow == -1) // High bound less than input range.
2334  return replaceInstUsesWith(Cmp, Builder.getTrue());
2335  if (Pred == ICmpInst::ICMP_UGT)
2336  return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
2337  return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
2338  }
2339 
2340  return nullptr;
2341 }
2342 
2343 /// Fold icmp (sub X, Y), C.
2344 Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp,
2345  BinaryOperator *Sub,
2346  const APInt *C) {
2347  Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2348  ICmpInst::Predicate Pred = Cmp.getPredicate();
2349 
2350  // The following transforms are only worth it if the only user of the subtract
2351  // is the icmp.
2352  if (!Sub->hasOneUse())
2353  return nullptr;
2354 
2355  if (Sub->hasNoSignedWrap()) {
2356  // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2357  if (Pred == ICmpInst::ICMP_SGT && C->isAllOnesValue())
2358  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2359 
2360  // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2361  if (Pred == ICmpInst::ICMP_SGT && C->isNullValue())
2362  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2363 
2364  // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2365  if (Pred == ICmpInst::ICMP_SLT && C->isNullValue())
2366  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2367 
2368  // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2369  if (Pred == ICmpInst::ICMP_SLT && C->isOneValue())
2370  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2371  }
2372 
2373  const APInt *C2;
2374  if (!match(X, m_APInt(C2)))
2375  return nullptr;
2376 
2377  // C2 - Y <u C -> (Y | (C - 1)) == C2
2378  // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2379  if (Pred == ICmpInst::ICMP_ULT && C->isPowerOf2() &&
2380  (*C2 & (*C - 1)) == (*C - 1))
2381  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, *C - 1), X);
2382 
2383  // C2 - Y >u C -> (Y | C) != C2
2384  // iff C2 & C == C and C + 1 is a power of 2
2385  if (Pred == ICmpInst::ICMP_UGT && (*C + 1).isPowerOf2() && (*C2 & *C) == *C)
2386  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, *C), X);
2387 
2388  return nullptr;
2389 }
2390 
2391 /// Fold icmp (add X, Y), C.
2392 Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp,
2394  const APInt *C) {
2395  Value *Y = Add->getOperand(1);
2396  const APInt *C2;
2397  if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2398  return nullptr;
2399 
2400  // Fold icmp pred (add X, C2), C.
2401  Value *X = Add->getOperand(0);
2402  Type *Ty = Add->getType();
2403  CmpInst::Predicate Pred = Cmp.getPredicate();
2404 
2405  // If the add does not wrap, we can always adjust the compare by subtracting
2406  // the constants. Equality comparisons are handled elsewhere. SGE/SLE are
2407  // canonicalized to SGT/SLT.
2408  if (Add->hasNoSignedWrap() &&
2409  (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) {
2410  bool Overflow;
2411  APInt NewC = C->ssub_ov(*C2, Overflow);
2412  // If there is overflow, the result must be true or false.
2413  // TODO: Can we assert there is no overflow because InstSimplify always
2414  // handles those cases?
2415  if (!Overflow)
2416  // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2417  return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2418  }
2419 
2420  auto CR = ConstantRange::makeExactICmpRegion(Pred, *C).subtract(*C2);
2421  const APInt &Upper = CR.getUpper();
2422  const APInt &Lower = CR.getLower();
2423  if (Cmp.isSigned()) {
2424  if (Lower.isSignMask())
2425  return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2426  if (Upper.isSignMask())
2427  return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2428  } else {
2429  if (Lower.isMinValue())
2430  return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2431  if (Upper.isMinValue())
2432  return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2433  }
2434 
2435  if (!Add->hasOneUse())
2436  return nullptr;
2437 
2438  // X+C <u C2 -> (X & -C2) == C
2439  // iff C & (C2-1) == 0
2440  // C2 is a power of 2
2441  if (Pred == ICmpInst::ICMP_ULT && C->isPowerOf2() && (*C2 & (*C - 1)) == 0)
2442  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -(*C)),
2443  ConstantExpr::getNeg(cast<Constant>(Y)));
2444 
2445  // X+C >u C2 -> (X & ~C2) != C
2446  // iff C & C2 == 0
2447  // C2+1 is a power of 2
2448  if (Pred == ICmpInst::ICMP_UGT && (*C + 1).isPowerOf2() && (*C2 & *C) == 0)
2449  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~(*C)),
2450  ConstantExpr::getNeg(cast<Constant>(Y)));
2451 
2452  return nullptr;
2453 }
2454 
2455 bool InstCombiner::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2456  Value *&RHS, ConstantInt *&Less,
2457  ConstantInt *&Equal,
2458  ConstantInt *&Greater) {
2459  // TODO: Generalize this to work with other comparison idioms or ensure
2460  // they get canonicalized into this form.
2461 
2462  // select i1 (a == b), i32 Equal, i32 (select i1 (a < b), i32 Less, i32
2463  // Greater), where Equal, Less and Greater are placeholders for any three
2464  // constants.
2465  ICmpInst::Predicate PredA, PredB;
2466  if (match(SI->getTrueValue(), m_ConstantInt(Equal)) &&
2467  match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) &&
2468  PredA == ICmpInst::ICMP_EQ &&
2469  match(SI->getFalseValue(),
2470  m_Select(m_ICmp(PredB, m_Specific(LHS), m_Specific(RHS)),
2471  m_ConstantInt(Less), m_ConstantInt(Greater))) &&
2472  PredB == ICmpInst::ICMP_SLT) {
2473  return true;
2474  }
2475  return false;
2476 }
2477 
2478 Instruction *InstCombiner::foldICmpSelectConstant(ICmpInst &Cmp,
2479  SelectInst *Select,
2480  ConstantInt *C) {
2481 
2482  assert(C && "Cmp RHS should be a constant int!");
2483  // If we're testing a constant value against the result of a three way
2484  // comparison, the result can be expressed directly in terms of the
2485  // original values being compared. Note: We could possibly be more
2486  // aggressive here and remove the hasOneUse test. The original select is
2487  // really likely to simplify or sink when we remove a test of the result.
2488  Value *OrigLHS, *OrigRHS;
2489  ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2490  if (Cmp.hasOneUse() &&
2491  matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2492  C3GreaterThan)) {
2493  assert(C1LessThan && C2Equal && C3GreaterThan);
2494 
2495  bool TrueWhenLessThan =
2496  ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2497  ->isAllOnesValue();
2498  bool TrueWhenEqual =
2499  ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2500  ->isAllOnesValue();
2501  bool TrueWhenGreaterThan =
2502  ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2503  ->isAllOnesValue();
2504 
2505  // This generates the new instruction that will replace the original Cmp
2506  // Instruction. Instead of enumerating the various combinations when
2507  // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2508  // false, we rely on chaining of ORs and future passes of InstCombine to
2509  // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2510 
2511  // When none of the three constants satisfy the predicate for the RHS (C),
2512  // the entire original Cmp can be simplified to a false.
2513  Value *Cond = Builder.getFalse();
2514  if (TrueWhenLessThan)
2515  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT, OrigLHS, OrigRHS));
2516  if (TrueWhenEqual)
2517  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ, OrigLHS, OrigRHS));
2518  if (TrueWhenGreaterThan)
2519  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT, OrigLHS, OrigRHS));
2520 
2521  return replaceInstUsesWith(Cmp, Cond);
2522  }
2523  return nullptr;
2524 }
2525 
2526 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2527 /// where X is some kind of instruction.
2528 Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) {
2529  const APInt *C;
2530  if (!match(Cmp.getOperand(1), m_APInt(C)))
2531  return nullptr;
2532 
2533  if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2534  switch (BO->getOpcode()) {
2535  case Instruction::Xor:
2536  if (Instruction *I = foldICmpXorConstant(Cmp, BO, C))
2537  return I;
2538  break;
2539  case Instruction::And:
2540  if (Instruction *I = foldICmpAndConstant(Cmp, BO, C))
2541  return I;
2542  break;
2543  case Instruction::Or:
2544  if (Instruction *I = foldICmpOrConstant(Cmp, BO, C))
2545  return I;
2546  break;
2547  case Instruction::Mul:
2548  if (Instruction *I = foldICmpMulConstant(Cmp, BO, C))
2549  return I;
2550  break;
2551  case Instruction::Shl:
2552  if (Instruction *I = foldICmpShlConstant(Cmp, BO, C))
2553  return I;
2554  break;
2555  case Instruction::LShr:
2556  case Instruction::AShr:
2557  if (Instruction *I = foldICmpShrConstant(Cmp, BO, C))
2558  return I;
2559  break;
2560  case Instruction::UDiv:
2561  if (Instruction *I = foldICmpUDivConstant(Cmp, BO, C))
2562  return I;
2564  case Instruction::SDiv:
2565  if (Instruction *I = foldICmpDivConstant(Cmp, BO, C))
2566  return I;
2567  break;
2568  case Instruction::Sub:
2569  if (Instruction *I = foldICmpSubConstant(Cmp, BO, C))
2570  return I;
2571  break;
2572  case Instruction::Add:
2573  if (Instruction *I = foldICmpAddConstant(Cmp, BO, C))
2574  return I;
2575  break;
2576  default:
2577  break;
2578  }
2579  // TODO: These folds could be refactored to be part of the above calls.
2580  if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, C))
2581  return I;
2582  }
2583 
2584  // Match against CmpInst LHS being instructions other than binary operators.
2585 
2586  if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2587  // For now, we only support constant integers while folding the
2588  // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2589  // similar to the cases handled by binary ops above.
2590  if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2591  if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2592  return I;
2593  }
2594 
2595  if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2596  if (Instruction *I = foldICmpTruncConstant(Cmp, TI, C))
2597  return I;
2598  }
2599 
2600  if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, C))
2601  return I;
2602 
2603  return nullptr;
2604 }
2605 
2606 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2607 /// icmp eq/ne BO, C.
2608 Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
2609  BinaryOperator *BO,
2610  const APInt *C) {
2611  // TODO: Some of these folds could work with arbitrary constants, but this
2612  // function is limited to scalar and vector splat constants.
2613  if (!Cmp.isEquality())
2614  return nullptr;
2615 
2616  ICmpInst::Predicate Pred = Cmp.getPredicate();
2617  bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2618  Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2619  Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2620 
2621  switch (BO->getOpcode()) {
2622  case Instruction::SRem:
2623  // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2624  if (C->isNullValue() && BO->hasOneUse()) {
2625  const APInt *BOC;
2626  if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2627  Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
2628  return new ICmpInst(Pred, NewRem,
2630  }
2631  }
2632  break;
2633  case Instruction::Add: {
2634  // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
2635  const APInt *BOC;
2636  if (match(BOp1, m_APInt(BOC))) {
2637  if (BO->hasOneUse()) {
2638  Constant *SubC = ConstantExpr::getSub(RHS, cast<Constant>(BOp1));
2639  return new ICmpInst(Pred, BOp0, SubC);
2640  }
2641  } else if (C->isNullValue()) {
2642  // Replace ((add A, B) != 0) with (A != -B) if A or B is
2643  // efficiently invertible, or if the add has just this one use.
2644  if (Value *NegVal = dyn_castNegVal(BOp1))
2645  return new ICmpInst(Pred, BOp0, NegVal);
2646  if (Value *NegVal = dyn_castNegVal(BOp0))
2647  return new ICmpInst(Pred, NegVal, BOp1);
2648  if (BO->hasOneUse()) {
2649  Value *Neg = Builder.CreateNeg(BOp1);
2650  Neg->takeName(BO);
2651  return new ICmpInst(Pred, BOp0, Neg);
2652  }
2653  }
2654  break;
2655  }
2656  case Instruction::Xor:
2657  if (BO->hasOneUse()) {
2658  if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2659  // For the xor case, we can xor two constants together, eliminating
2660  // the explicit xor.
2661  return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
2662  } else if (C->isNullValue()) {
2663  // Replace ((xor A, B) != 0) with (A != B)
2664  return new ICmpInst(Pred, BOp0, BOp1);
2665  }
2666  }
2667  break;
2668  case Instruction::Sub:
2669  if (BO->hasOneUse()) {
2670  const APInt *BOC;
2671  if (match(BOp0, m_APInt(BOC))) {
2672  // Replace ((sub BOC, B) != C) with (B != BOC-C).
2673  Constant *SubC = ConstantExpr::getSub(cast<Constant>(BOp0), RHS);
2674  return new ICmpInst(Pred, BOp1, SubC);
2675  } else if (C->isNullValue()) {
2676  // Replace ((sub A, B) != 0) with (A != B).
2677  return new ICmpInst(Pred, BOp0, BOp1);
2678  }
2679  }
2680  break;
2681  case Instruction::Or: {
2682  const APInt *BOC;
2683  if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
2684  // Comparing if all bits outside of a constant mask are set?
2685  // Replace (X | C) == -1 with (X & ~C) == ~C.
2686  // This removes the -1 constant.
2687  Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
2688  Value *And = Builder.CreateAnd(BOp0, NotBOC);
2689  return new ICmpInst(Pred, And, NotBOC);
2690  }
2691  break;
2692  }
2693  case Instruction::And: {
2694  const APInt *BOC;
2695  if (match(BOp1, m_APInt(BOC))) {
2696  // If we have ((X & C) == C), turn it into ((X & C) != 0).
2697  if (C == BOC && C->isPowerOf2())
2698  return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
2699  BO, Constant::getNullValue(RHS->getType()));
2700 
2701  // Don't perform the following transforms if the AND has multiple uses
2702  if (!BO->hasOneUse())
2703  break;
2704 
2705  // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
2706  if (BOC->isSignMask()) {
2707  Constant *Zero = Constant::getNullValue(BOp0->getType());
2708  auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
2709  return new ICmpInst(NewPred, BOp0, Zero);
2710  }
2711 
2712  // ((X & ~7) == 0) --> X < 8
2713  if (C->isNullValue() && (~(*BOC) + 1).isPowerOf2()) {
2714  Constant *NegBOC = ConstantExpr::getNeg(cast<Constant>(BOp1));
2715  auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
2716  return new ICmpInst(NewPred, BOp0, NegBOC);
2717  }
2718  }
2719  break;
2720  }
2721  case Instruction::Mul:
2722  if (C->isNullValue() && BO->hasNoSignedWrap()) {
2723  const APInt *BOC;
2724  if (match(BOp1, m_APInt(BOC)) && !BOC->isNullValue()) {
2725  // The trivial case (mul X, 0) is handled by InstSimplify.
2726  // General case : (mul X, C) != 0 iff X != 0
2727  // (mul X, C) == 0 iff X == 0
2728  return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType()));
2729  }
2730  }
2731  break;
2732  case Instruction::UDiv:
2733  if (C->isNullValue()) {
2734  // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
2735  auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
2736  return new ICmpInst(NewPred, BOp1, BOp0);
2737  }
2738  break;
2739  default:
2740  break;
2741  }
2742  return nullptr;
2743 }
2744 
2745 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
2746 Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
2747  const APInt *C) {
2749  if (!II || !Cmp.isEquality())
2750  return nullptr;
2751 
2752  // Handle icmp {eq|ne} <intrinsic>, Constant.
2753  Type *Ty = II->getType();
2754  switch (II->getIntrinsicID()) {
2755  case Intrinsic::bswap:
2756  Worklist.Add(II);
2757  Cmp.setOperand(0, II->getArgOperand(0));
2758  Cmp.setOperand(1, ConstantInt::get(Ty, C->byteSwap()));
2759  return &Cmp;
2760 
2761  case Intrinsic::ctlz:
2762  case Intrinsic::cttz:
2763  // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
2764  if (*C == C->getBitWidth()) {
2765  Worklist.Add(II);
2766  Cmp.setOperand(0, II->getArgOperand(0));
2768  return &Cmp;
2769  }
2770  break;
2771 
2772  case Intrinsic::ctpop: {
2773  // popcount(A) == 0 -> A == 0 and likewise for !=
2774  // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
2775  bool IsZero = C->isNullValue();
2776  if (IsZero || *C == C->getBitWidth()) {
2777  Worklist.Add(II);
2778  Cmp.setOperand(0, II->getArgOperand(0));
2779  auto *NewOp =
2781  Cmp.setOperand(1, NewOp);
2782  return &Cmp;
2783  }
2784  break;
2785  }
2786  default:
2787  break;
2788  }
2789 
2790  return nullptr;
2791 }
2792 
2793 /// Handle icmp with constant (but not simple integer constant) RHS.
2794 Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) {
2795  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2796  Constant *RHSC = dyn_cast<Constant>(Op1);
2797  Instruction *LHSI = dyn_cast<Instruction>(Op0);
2798  if (!RHSC || !LHSI)
2799  return nullptr;
2800 
2801  switch (LHSI->getOpcode()) {
2802  case Instruction::GetElementPtr:
2803  // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2804  if (RHSC->isNullValue() &&
2805  cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2806  return new ICmpInst(
2807  I.getPredicate(), LHSI->getOperand(0),
2808  Constant::getNullValue(LHSI->getOperand(0)->getType()));
2809  break;
2810  case Instruction::PHI:
2811  // Only fold icmp into the PHI if the phi and icmp are in the same
2812  // block. If in the same block, we're encouraging jump threading. If
2813  // not, we are just pessimizing the code by making an i1 phi.
2814  if (LHSI->getParent() == I.getParent())
2815  if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
2816  return NV;
2817  break;
2818  case Instruction::Select: {
2819  // If either operand of the select is a constant, we can fold the
2820  // comparison into the select arms, which will cause one to be
2821  // constant folded and the select turned into a bitwise or.
2822  Value *Op1 = nullptr, *Op2 = nullptr;
2823  ConstantInt *CI = nullptr;
2824  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2825  Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2826  CI = dyn_cast<ConstantInt>(Op1);
2827  }
2828  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2829  Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2830  CI = dyn_cast<ConstantInt>(Op2);
2831  }
2832 
2833  // We only want to perform this transformation if it will not lead to
2834  // additional code. This is true if either both sides of the select
2835  // fold to a constant (in which case the icmp is replaced with a select
2836  // which will usually simplify) or this is the only user of the
2837  // select (in which case we are trading a select+icmp for a simpler
2838  // select+icmp) or all uses of the select can be replaced based on
2839  // dominance information ("Global cases").
2840  bool Transform = false;
2841  if (Op1 && Op2)
2842  Transform = true;
2843  else if (Op1 || Op2) {
2844  // Local case
2845  if (LHSI->hasOneUse())
2846  Transform = true;
2847  // Global cases
2848  else if (CI && !CI->isZero())
2849  // When Op1 is constant try replacing select with second operand.
2850  // Otherwise Op2 is constant and try replacing select with first
2851  // operand.
2852  Transform =
2853  replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
2854  }
2855  if (Transform) {
2856  if (!Op1)
2857  Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
2858  I.getName());
2859  if (!Op2)
2860  Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
2861  I.getName());
2862  return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2863  }
2864  break;
2865  }
2866  case Instruction::IntToPtr:
2867  // icmp pred inttoptr(X), null -> icmp pred X, 0
2868  if (RHSC->isNullValue() &&
2869  DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
2870  return new ICmpInst(
2871  I.getPredicate(), LHSI->getOperand(0),
2872  Constant::getNullValue(LHSI->getOperand(0)->getType()));
2873  break;
2874 
2875  case Instruction::Load:
2876  // Try to optimize things like "A[i] > 4" to index computations.
2877  if (GetElementPtrInst *GEP =
2878  dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2879  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2880  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2881  !cast<LoadInst>(LHSI)->isVolatile())
2882  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
2883  return Res;
2884  }
2885  break;
2886  }
2887 
2888  return nullptr;
2889 }
2890 
2891 /// Try to fold icmp (binop), X or icmp X, (binop).
2892 /// TODO: A large part of this logic is duplicated in InstSimplify's
2893 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
2894 /// duplication.
2895 Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I) {
2896  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2897 
2898  // Special logic for binary operators.
2899  BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2900  BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2901  if (!BO0 && !BO1)
2902  return nullptr;
2903 
2904  const CmpInst::Predicate Pred = I.getPredicate();
2905  bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2906  if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2907  NoOp0WrapProblem =
2908  ICmpInst::isEquality(Pred) ||
2909  (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2910  (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2911  if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2912  NoOp1WrapProblem =
2913  ICmpInst::isEquality(Pred) ||
2914  (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2915  (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2916 
2917  // Analyze the case when either Op0 or Op1 is an add instruction.
2918  // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2919  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2920  if (BO0 && BO0->getOpcode() == Instruction::Add) {
2921  A = BO0->getOperand(0);
2922  B = BO0->getOperand(1);
2923  }
2924  if (BO1 && BO1->getOpcode() == Instruction::Add) {
2925  C = BO1->getOperand(0);
2926  D = BO1->getOperand(1);
2927  }
2928 
2929  // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2930  if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2931  return new ICmpInst(Pred, A == Op1 ? B : A,
2932  Constant::getNullValue(Op1->getType()));
2933 
2934  // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2935  if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2936  return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2937  C == Op0 ? D : C);
2938 
2939  // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2940  if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
2941  NoOp1WrapProblem &&
2942  // Try not to increase register pressure.
2943  BO0->hasOneUse() && BO1->hasOneUse()) {
2944  // Determine Y and Z in the form icmp (X+Y), (X+Z).
2945  Value *Y, *Z;
2946  if (A == C) {
2947  // C + B == C + D -> B == D
2948  Y = B;
2949  Z = D;
2950  } else if (A == D) {
2951  // D + B == C + D -> B == C
2952  Y = B;
2953  Z = C;
2954  } else if (B == C) {
2955  // A + C == C + D -> A == D
2956  Y = A;
2957  Z = D;
2958  } else {
2959  assert(B == D);
2960  // A + D == C + D -> A == C
2961  Y = A;
2962  Z = C;
2963  }
2964  return new ICmpInst(Pred, Y, Z);
2965  }
2966 
2967  // icmp slt (X + -1), Y -> icmp sle X, Y
2968  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
2969  match(B, m_AllOnes()))
2970  return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
2971 
2972  // icmp sge (X + -1), Y -> icmp sgt X, Y
2973  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
2974  match(B, m_AllOnes()))
2975  return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
2976 
2977  // icmp sle (X + 1), Y -> icmp slt X, Y
2978  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
2979  return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
2980 
2981  // icmp sgt (X + 1), Y -> icmp sge X, Y
2982  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
2983  return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
2984 
2985  // icmp sgt X, (Y + -1) -> icmp sge X, Y
2986  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
2987  match(D, m_AllOnes()))
2988  return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
2989 
2990  // icmp sle X, (Y + -1) -> icmp slt X, Y
2991  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
2992  match(D, m_AllOnes()))
2993  return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
2994 
2995  // icmp sge X, (Y + 1) -> icmp sgt X, Y
2996  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
2997  return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
2998 
2999  // icmp slt X, (Y + 1) -> icmp sle X, Y
3000  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
3001  return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3002 
3003  // TODO: The subtraction-related identities shown below also hold, but
3004  // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3005  // wouldn't happen even if they were implemented.
3006  //
3007  // icmp ult (X - 1), Y -> icmp ule X, Y
3008  // icmp uge (X - 1), Y -> icmp ugt X, Y
3009  // icmp ugt X, (Y - 1) -> icmp uge X, Y
3010  // icmp ule X, (Y - 1) -> icmp ult X, Y
3011 
3012  // icmp ule (X + 1), Y -> icmp ult X, Y
3013  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
3014  return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3015 
3016  // icmp ugt (X + 1), Y -> icmp uge X, Y
3017  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
3018  return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3019 
3020  // icmp uge X, (Y + 1) -> icmp ugt X, Y
3021  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
3022  return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3023 
3024  // icmp ult X, (Y + 1) -> icmp ule X, Y
3025  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
3026  return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3027 
3028  // if C1 has greater magnitude than C2:
3029  // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3030  // s.t. C3 = C1 - C2
3031  //
3032  // if C2 has greater magnitude than C1:
3033  // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3034  // s.t. C3 = C2 - C1
3035  if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3036  (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3037  if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3038  if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3039  const APInt &AP1 = C1->getValue();
3040  const APInt &AP2 = C2->getValue();
3041  if (AP1.isNegative() == AP2.isNegative()) {
3042  APInt AP1Abs = C1->getValue().abs();
3043  APInt AP2Abs = C2->getValue().abs();
3044  if (AP1Abs.uge(AP2Abs)) {
3045  ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3046  Value *NewAdd = Builder.CreateNSWAdd(A, C3);
3047  return new ICmpInst(Pred, NewAdd, C);
3048  } else {
3049  ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3050  Value *NewAdd = Builder.CreateNSWAdd(C, C3);
3051  return new ICmpInst(Pred, A, NewAdd);
3052  }
3053  }
3054  }
3055 
3056  // Analyze the case when either Op0 or Op1 is a sub instruction.
3057  // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3058  A = nullptr;
3059  B = nullptr;
3060  C = nullptr;
3061  D = nullptr;
3062  if (BO0 && BO0->getOpcode() == Instruction::Sub) {
3063  A = BO0->getOperand(0);
3064  B = BO0->getOperand(1);
3065  }
3066  if (BO1 && BO1->getOpcode() == Instruction::Sub) {
3067  C = BO1->getOperand(0);
3068  D = BO1->getOperand(1);
3069  }
3070 
3071  // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3072  if (A == Op1 && NoOp0WrapProblem)
3073  return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3074 
3075  // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3076  if (C == Op0 && NoOp1WrapProblem)
3077  return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3078 
3079  // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3080  if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3081  // Try not to increase register pressure.
3082  BO0->hasOneUse() && BO1->hasOneUse())
3083  return new ICmpInst(Pred, A, C);
3084 
3085  // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3086  if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3087  // Try not to increase register pressure.
3088  BO0->hasOneUse() && BO1->hasOneUse())
3089  return new ICmpInst(Pred, D, B);
3090 
3091  // icmp (0-X) < cst --> x > -cst
3092  if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3093  Value *X;
3094  if (match(BO0, m_Neg(m_Value(X))))
3095  if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3096  if (!RHSC->isMinValue(/*isSigned=*/true))
3097  return new ICmpInst(I.getSwappedPredicate(), X,
3098  ConstantExpr::getNeg(RHSC));
3099  }
3100 
3101  BinaryOperator *SRem = nullptr;
3102  // icmp (srem X, Y), Y
3103  if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
3104  SRem = BO0;
3105  // icmp Y, (srem X, Y)
3106  else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3107  Op0 == BO1->getOperand(1))
3108  SRem = BO1;
3109  if (SRem) {
3110  // We don't check hasOneUse to avoid increasing register pressure because
3111  // the value we use is the same value this instruction was already using.
3112  switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3113  default:
3114  break;
3115  case ICmpInst::ICMP_EQ:
3116  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3117  case ICmpInst::ICMP_NE:
3118  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3119  case ICmpInst::ICMP_SGT:
3120  case ICmpInst::ICMP_SGE:
3121  return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3123  case ICmpInst::ICMP_SLT:
3124  case ICmpInst::ICMP_SLE:
3125  return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3126  Constant::getNullValue(SRem->getType()));
3127  }
3128  }
3129 
3130  if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
3131  BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
3132  switch (BO0->getOpcode()) {
3133  default:
3134  break;
3135  case Instruction::Add:
3136  case Instruction::Sub:
3137  case Instruction::Xor: {
3138  if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3139  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3140 
3141  const APInt *C;
3142  if (match(BO0->getOperand(1), m_APInt(C))) {
3143  // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
3144  if (C->isSignMask()) {
3145  ICmpInst::Predicate NewPred =
3147  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3148  }
3149 
3150  // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
3151  if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
3152  ICmpInst::Predicate NewPred =
3154  NewPred = I.getSwappedPredicate(NewPred);
3155  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3156  }
3157  }
3158  break;
3159  }
3160  case Instruction::Mul: {
3161  if (!I.isEquality())
3162  break;
3163 
3164  const APInt *C;
3165  if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
3166  !C->isOneValue()) {
3167  // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
3168  // Mask = -1 >> count-trailing-zeros(C).
3169  if (unsigned TZs = C->countTrailingZeros()) {
3171  BO0->getType(),
3172  APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
3173  Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
3174  Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
3175  return new ICmpInst(Pred, And1, And2);
3176  }
3177  // If there are no trailing zeros in the multiplier, just eliminate
3178  // the multiplies (no masking is needed):
3179  // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
3180  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3181  }
3182  break;
3183  }
3184  case Instruction::UDiv:
3185  case Instruction::LShr:
3186  if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
3187  break;
3188  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3189 
3190  case Instruction::SDiv:
3191  if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
3192  break;
3193  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3194 
3195  case Instruction::AShr:
3196  if (!BO0->isExact() || !BO1->isExact())
3197  break;
3198  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3199 
3200  case Instruction::Shl: {
3201  bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3202  bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3203  if (!NUW && !NSW)
3204  break;
3205  if (!NSW && I.isSigned())
3206  break;
3207  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3208  }
3209  }
3210  }
3211 
3212  if (BO0) {
3213  // Transform A & (L - 1) `ult` L --> L != 0
3214  auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
3215  auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
3216 
3217  if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
3218  auto *Zero = Constant::getNullValue(BO0->getType());
3219  return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
3220  }
3221  }
3222 
3223  return nullptr;
3224 }
3225 
3226 /// Fold icmp Pred min|max(X, Y), X.
3228  ICmpInst::Predicate Pred = Cmp.getPredicate();
3229  Value *Op0 = Cmp.getOperand(0);
3230  Value *X = Cmp.getOperand(1);
3231 
3232  // Canonicalize minimum or maximum operand to LHS of the icmp.
3233  if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
3234  match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
3235  match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
3236  match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
3237  std::swap(Op0, X);
3238  Pred = Cmp.getSwappedPredicate();
3239  }
3240 
3241  Value *Y;
3242  if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
3243  // smin(X, Y) == X --> X s<= Y
3244  // smin(X, Y) s>= X --> X s<= Y
3245  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
3246  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
3247 
3248  // smin(X, Y) != X --> X s> Y
3249  // smin(X, Y) s< X --> X s> Y
3250  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
3251  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
3252 
3253  // These cases should be handled in InstSimplify:
3254  // smin(X, Y) s<= X --> true
3255  // smin(X, Y) s> X --> false
3256  return nullptr;
3257  }
3258 
3259  if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
3260  // smax(X, Y) == X --> X s>= Y
3261  // smax(X, Y) s<= X --> X s>= Y
3262  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
3263  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
3264 
3265  // smax(X, Y) != X --> X s< Y
3266  // smax(X, Y) s> X --> X s< Y
3267  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
3268  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
3269 
3270  // These cases should be handled in InstSimplify:
3271  // smax(X, Y) s>= X --> true
3272  // smax(X, Y) s< X --> false
3273  return nullptr;
3274  }
3275 
3276  if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
3277  // umin(X, Y) == X --> X u<= Y
3278  // umin(X, Y) u>= X --> X u<= Y
3279  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
3280  return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
3281 
3282  // umin(X, Y) != X --> X u> Y
3283  // umin(X, Y) u< X --> X u> Y
3284  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
3285  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
3286 
3287  // These cases should be handled in InstSimplify:
3288  // umin(X, Y) u<= X --> true
3289  // umin(X, Y) u> X --> false
3290  return nullptr;
3291  }
3292 
3293  if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
3294  // umax(X, Y) == X --> X u>= Y
3295  // umax(X, Y) u<= X --> X u>= Y
3296  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
3297  return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
3298 
3299  // umax(X, Y) != X --> X u< Y
3300  // umax(X, Y) u> X --> X u< Y
3301  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
3302  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
3303 
3304  // These cases should be handled in InstSimplify:
3305  // umax(X, Y) u>= X --> true
3306  // umax(X, Y) u< X --> false
3307  return nullptr;
3308  }
3309 
3310  return nullptr;
3311 }
3312 
3313 Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
3314  if (!I.isEquality())
3315  return nullptr;
3316 
3317  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3318  const CmpInst::Predicate Pred = I.getPredicate();
3319  Value *A, *B, *C, *D;
3320  if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3321  if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3322  Value *OtherVal = A == Op1 ? B : A;
3323  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3324  }
3325 
3326  if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3327  // A^c1 == C^c2 --> A == C^(c1^c2)
3328  ConstantInt *C1, *C2;
3329  if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
3330  Op1->hasOneUse()) {
3331  Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
3332  Value *Xor = Builder.CreateXor(C, NC);
3333  return new ICmpInst(Pred, A, Xor);
3334  }
3335 
3336  // A^B == A^D -> B == D
3337  if (A == C)
3338  return new ICmpInst(Pred, B, D);
3339  if (A == D)
3340  return new ICmpInst(Pred, B, C);
3341  if (B == C)
3342  return new ICmpInst(Pred, A, D);
3343  if (B == D)
3344  return new ICmpInst(Pred, A, C);
3345  }
3346  }
3347 
3348  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
3349  // A == (A^B) -> B == 0
3350  Value *OtherVal = A == Op0 ? B : A;
3351  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3352  }
3353 
3354  // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3355  if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3356  match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3357  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3358 
3359  if (A == C) {
3360  X = B;
3361  Y = D;
3362  Z = A;
3363  } else if (A == D) {
3364  X = B;
3365  Y = C;
3366  Z = A;
3367  } else if (B == C) {
3368  X = A;
3369  Y = D;
3370  Z = B;
3371  } else if (B == D) {
3372  X = A;
3373  Y = C;
3374  Z = B;
3375  }
3376 
3377  if (X) { // Build (X^Y) & Z
3378  Op1 = Builder.CreateXor(X, Y);
3379  Op1 = Builder.CreateAnd(Op1, Z);
3380  I.setOperand(0, Op1);
3381  I.setOperand(1, Constant::getNullValue(Op1->getType()));
3382  return &I;
3383  }
3384  }
3385 
3386  // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3387  // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3388  ConstantInt *Cst1;
3389  if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
3390  match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3391  (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3392  match(Op1, m_ZExt(m_Value(A))))) {
3393  APInt Pow2 = Cst1->getValue() + 1;
3394  if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3395  Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3396  return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
3397  }
3398 
3399  // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3400  // For lshr and ashr pairs.
3401  if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3402  match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3403  (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3404  match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3405  unsigned TypeBits = Cst1->getBitWidth();
3406  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3407  if (ShAmt < TypeBits && ShAmt != 0) {
3408  ICmpInst::Predicate NewPred =
3410  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3411  APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3412  return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
3413  }
3414  }
3415 
3416  // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
3417  if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
3418  match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
3419  unsigned TypeBits = Cst1->getBitWidth();
3420  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3421  if (ShAmt < TypeBits && ShAmt != 0) {
3422  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3423  APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
3424  Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
3425  I.getName() + ".mask");
3426  return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
3427  }
3428  }
3429 
3430  // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3431  // "icmp (and X, mask), cst"
3432  uint64_t ShAmt = 0;
3433  if (Op0->hasOneUse() &&
3434  match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
3435  match(Op1, m_ConstantInt(Cst1)) &&
3436  // Only do this when A has multiple uses. This is most important to do
3437  // when it exposes other optimizations.
3438  !A->hasOneUse()) {
3439  unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3440 
3441  if (ShAmt < ASize) {
3442  APInt MaskV =
3444  MaskV <<= ShAmt;
3445 
3446  APInt CmpV = Cst1->getValue().zext(ASize);
3447  CmpV <<= ShAmt;
3448 
3449  Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
3450  return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
3451  }
3452  }
3453 
3454  // If both operands are byte-swapped or bit-reversed, just compare the
3455  // original values.
3456  // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
3457  // and handle more intrinsics.
3458  if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
3459  (match(Op0, m_BitReverse(m_Value(A))) &&
3460  match(Op1, m_BitReverse(m_Value(B)))))
3461  return new ICmpInst(Pred, A, B);
3462 
3463  return nullptr;
3464 }
3465 
3466 /// Handle icmp (cast x to y), (cast/cst). We only handle extending casts so
3467 /// far.
3468 Instruction *InstCombiner::foldICmpWithCastAndCast(ICmpInst &ICmp) {
3469  const CastInst *LHSCI = cast<CastInst>(ICmp.getOperand(0));
3470  Value *LHSCIOp = LHSCI->getOperand(0);
3471  Type *SrcTy = LHSCIOp->getType();
3472  Type *DestTy = LHSCI->getType();
3473  Value *RHSCIOp;
3474 
3475  // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
3476  // integer type is the same size as the pointer type.
3477  if (LHSCI->getOpcode() == Instruction::PtrToInt &&
3478  DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
3479  Value *RHSOp = nullptr;
3480  if (auto *RHSC = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
3481  Value *RHSCIOp = RHSC->getOperand(0);
3482  if (RHSCIOp->getType()->getPointerAddressSpace() ==
3483  LHSCIOp->getType()->getPointerAddressSpace()) {
3484  RHSOp = RHSC->getOperand(0);
3485  // If the pointer types don't match, insert a bitcast.
3486  if (LHSCIOp->getType() != RHSOp->getType())
3487  RHSOp = Builder.CreateBitCast(RHSOp, LHSCIOp->getType());
3488  }
3489  } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
3490  RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
3491  }
3492 
3493  if (RHSOp)
3494  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSOp);
3495  }
3496 
3497  // The code below only handles extension cast instructions, so far.
3498  // Enforce this.
3499  if (LHSCI->getOpcode() != Instruction::ZExt &&
3500  LHSCI->getOpcode() != Instruction::SExt)
3501  return nullptr;
3502 
3503  bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
3504  bool isSignedCmp = ICmp.isSigned();
3505 
3506  if (auto *CI = dyn_cast<CastInst>(ICmp.getOperand(1))) {
3507  // Not an extension from the same type?
3508  RHSCIOp = CI->getOperand(0);
3509  if (RHSCIOp->getType() != LHSCIOp->getType())
3510  return nullptr;
3511 
3512  // If the signedness of the two casts doesn't agree (i.e. one is a sext
3513  // and the other is a zext), then we can't handle this.
3514  if (CI->getOpcode() != LHSCI->getOpcode())
3515  return nullptr;
3516 
3517  // Deal with equality cases early.
3518  if (ICmp.isEquality())
3519  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3520 
3521  // A signed comparison of sign extended values simplifies into a
3522  // signed comparison.
3523  if (isSignedCmp && isSignedExt)
3524  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3525 
3526  // The other three cases all fold into an unsigned comparison.
3527  return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
3528  }
3529 
3530  // If we aren't dealing with a constant on the RHS, exit early.
3531  auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
3532  if (!C)
3533  return nullptr;
3534 
3535  // Compute the constant that would happen if we truncated to SrcTy then
3536  // re-extended to DestTy.
3537  Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
3538  Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
3539 
3540  // If the re-extended constant didn't change...
3541  if (Res2 == C) {
3542  // Deal with equality cases early.
3543  if (ICmp.isEquality())
3544  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3545 
3546  // A signed comparison of sign extended values simplifies into a
3547  // signed comparison.
3548  if (isSignedExt && isSignedCmp)
3549  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3550 
3551  // The other three cases all fold into an unsigned comparison.
3552  return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, Res1);
3553  }
3554 
3555  // The re-extended constant changed, partly changed (in the case of a vector),
3556  // or could not be determined to be equal (in the case of a constant
3557  // expression), so the constant cannot be represented in the shorter type.
3558  // Consequently, we cannot emit a simple comparison.
3559  // All the cases that fold to true or false will have already been handled
3560  // by SimplifyICmpInst, so only deal with the tricky case.
3561 
3562  if (isSignedCmp || !isSignedExt || !isa<ConstantInt>(C))
3563  return nullptr;
3564 
3565  // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
3566  // should have been folded away previously and not enter in here.
3567 
3568  // We're performing an unsigned comp with a sign extended value.
3569  // This is true if the input is >= 0. [aka >s -1]
3570  Constant *NegOne = Constant::getAllOnesValue(SrcTy);
3571  Value *Result = Builder.CreateICmpSGT(LHSCIOp, NegOne, ICmp.getName());
3572 
3573  // Finally, return the value computed.
3574  if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
3575  return replaceInstUsesWith(ICmp, Result);
3576 
3577  assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
3578  return BinaryOperator::CreateNot(Result);
3579 }
3580 
3581 bool InstCombiner::OptimizeOverflowCheck(OverflowCheckFlavor OCF, Value *LHS,
3582  Value *RHS, Instruction &OrigI,
3583  Value *&Result, Constant *&Overflow) {
3584  if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
3585  std::swap(LHS, RHS);
3586 
3587  auto SetResult = [&](Value *OpResult, Constant *OverflowVal, bool ReuseName) {
3588  Result = OpResult;
3589  Overflow = OverflowVal;
3590  if (ReuseName)
3591  Result->takeName(&OrigI);
3592  return true;
3593  };
3594 
3595  // If the overflow check was an add followed by a compare, the insertion point
3596  // may be pointing to the compare. We want to insert the new instructions
3597  // before the add in case there are uses of the add between the add and the
3598  // compare.
3599  Builder.SetInsertPoint(&OrigI);
3600 
3601  switch (OCF) {
3602  case OCF_INVALID:
3603  llvm_unreachable("bad overflow check kind!");
3604 
3605  case OCF_UNSIGNED_ADD: {
3606  OverflowResult OR = computeOverflowForUnsignedAdd(LHS, RHS, &OrigI);
3608  return SetResult(Builder.CreateNUWAdd(LHS, RHS), Builder.getFalse(),
3609  true);
3610 
3612  return SetResult(Builder.CreateAdd(LHS, RHS), Builder.getTrue(), true);
3613 
3614  // Fall through uadd into sadd
3616  }
3617  case OCF_SIGNED_ADD: {
3618  // X + 0 -> {X, false}
3619  if (match(RHS, m_Zero()))
3620  return SetResult(LHS, Builder.getFalse(), false);
3621 
3622  // We can strength reduce this signed add into a regular add if we can prove
3623  // that it will never overflow.
3624  if (OCF == OCF_SIGNED_ADD)
3625  if (willNotOverflowSignedAdd(LHS, RHS, OrigI))
3626  return SetResult(Builder.CreateNSWAdd(LHS, RHS), Builder.getFalse(),
3627  true);
3628  break;
3629  }
3630 
3631  case OCF_UNSIGNED_SUB:
3632  case OCF_SIGNED_SUB: {
3633  // X - 0 -> {X, false}
3634  if (match(RHS, m_Zero()))
3635  return SetResult(LHS, Builder.getFalse(), false);
3636 
3637  if (OCF == OCF_SIGNED_SUB) {
3638  if (willNotOverflowSignedSub(LHS, RHS, OrigI))
3639  return SetResult(Builder.CreateNSWSub(LHS, RHS), Builder.getFalse(),
3640  true);
3641  } else {
3642  if (willNotOverflowUnsignedSub(LHS, RHS, OrigI))
3643  return SetResult(Builder.CreateNUWSub(LHS, RHS), Builder.getFalse(),
3644  true);
3645  }
3646  break;
3647  }
3648 
3649  case OCF_UNSIGNED_MUL: {
3650  OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, &OrigI);
3652  return SetResult(Builder.CreateNUWMul(LHS, RHS), Builder.getFalse(),
3653  true);
3655  return SetResult(Builder.CreateMul(LHS, RHS), Builder.getTrue(), true);
3657  }
3658  case OCF_SIGNED_MUL:
3659  // X * undef -> undef
3660  if (isa<UndefValue>(RHS))
3661  return SetResult(RHS, UndefValue::get(Builder.getInt1Ty()), false);
3662 
3663  // X * 0 -> {0, false}
3664  if (match(RHS, m_Zero()))
3665  return SetResult(RHS, Builder.getFalse(), false);
3666 
3667  // X * 1 -> {X, false}
3668  if (match(RHS, m_One()))
3669  return SetResult(LHS, Builder.getFalse(), false);
3670 
3671  if (OCF == OCF_SIGNED_MUL)
3672  if (willNotOverflowSignedMul(LHS, RHS, OrigI))
3673  return SetResult(Builder.CreateNSWMul(LHS, RHS), Builder.getFalse(),
3674  true);
3675  break;
3676  }
3677 
3678  return false;
3679 }
3680 
3681 /// \brief Recognize and process idiom involving test for multiplication
3682 /// overflow.
3683 ///
3684 /// The caller has matched a pattern of the form:
3685 /// I = cmp u (mul(zext A, zext B), V
3686 /// The function checks if this is a test for overflow and if so replaces
3687 /// multiplication with call to 'mul.with.overflow' intrinsic.
3688 ///
3689 /// \param I Compare instruction.
3690 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
3691 /// the compare instruction. Must be of integer type.
3692 /// \param OtherVal The other argument of compare instruction.
3693 /// \returns Instruction which must replace the compare instruction, NULL if no
3694 /// replacement required.
3696  Value *OtherVal, InstCombiner &IC) {
3697  // Don't bother doing this transformation for pointers, don't do it for
3698  // vectors.
3699  if (!isa<IntegerType>(MulVal->getType()))
3700  return nullptr;
3701 
3702  assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
3703  assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
3704  auto *MulInstr = dyn_cast<Instruction>(MulVal);
3705  if (!MulInstr)
3706  return nullptr;
3707  assert(MulInstr->getOpcode() == Instruction::Mul);
3708 
3709  auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
3710  *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
3711  assert(LHS->getOpcode() == Instruction::ZExt);
3712  assert(RHS->getOpcode() == Instruction::ZExt);
3713  Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
3714 
3715  // Calculate type and width of the result produced by mul.with.overflow.
3716  Type *TyA = A->getType(), *TyB = B->getType();
3717  unsigned WidthA = TyA->getPrimitiveSizeInBits(),
3718  WidthB = TyB->getPrimitiveSizeInBits();
3719  unsigned MulWidth;
3720  Type *MulType;
3721  if (WidthB > WidthA) {
3722  MulWidth = WidthB;
3723  MulType = TyB;
3724  } else {
3725  MulWidth = WidthA;
3726  MulType = TyA;
3727  }
3728 
3729  // In order to replace the original mul with a narrower mul.with.overflow,
3730  // all uses must ignore upper bits of the product. The number of used low
3731  // bits must be not greater than the width of mul.with.overflow.
3732  if (MulVal->hasNUsesOrMore(2))
3733  for (User *U : MulVal->users()) {
3734  if (U == &I)
3735  continue;
3736  if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
3737  // Check if truncation ignores bits above MulWidth.
3738  unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
3739  if (TruncWidth > MulWidth)
3740  return nullptr;
3741  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
3742  // Check if AND ignores bits above MulWidth.
3743  if (BO->getOpcode() != Instruction::And)
3744  return nullptr;
3745  if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3746  const APInt &CVal = CI->getValue();
3747  if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
3748  return nullptr;
3749  } else {
3750  // In this case we could have the operand of the binary operation
3751  // being defined in another block, and performing the replacement
3752  // could break the dominance relation.
3753  return nullptr;
3754  }
3755  } else {
3756  // Other uses prohibit this transformation.
3757  return nullptr;
3758  }
3759  }
3760 
3761  // Recognize patterns
3762  switch (I.getPredicate()) {
3763  case ICmpInst::ICMP_EQ:
3764  case ICmpInst::ICMP_NE:
3765  // Recognize pattern:
3766  // mulval = mul(zext A, zext B)
3767  // cmp eq/neq mulval, zext trunc mulval
3768  if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
3769  if (Zext->hasOneUse()) {
3770  Value *ZextArg = Zext->getOperand(0);
3771  if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
3772  if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
3773  break; //Recognized
3774  }
3775 
3776  // Recognize pattern:
3777  // mulval = mul(zext A, zext B)
3778  // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
3779  ConstantInt *CI;
3780  Value *ValToMask;
3781  if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
3782  if (ValToMask != MulVal)
3783  return nullptr;
3784  const APInt &CVal = CI->getValue() + 1;
3785  if (CVal.isPowerOf2()) {
3786  unsigned MaskWidth = CVal.logBase2();
3787  if (MaskWidth == MulWidth)
3788  break; // Recognized
3789  }
3790  }
3791  return nullptr;
3792 
3793  case ICmpInst::ICMP_UGT:
3794  // Recognize pattern:
3795  // mulval = mul(zext A, zext B)
3796  // cmp ugt mulval, max
3797  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3798  APInt MaxVal = APInt::getMaxValue(MulWidth);
3799  MaxVal = MaxVal.zext(CI->getBitWidth());
3800  if (MaxVal.eq(CI->getValue()))
3801  break; // Recognized
3802  }
3803  return nullptr;
3804 
3805  case ICmpInst::ICMP_UGE:
3806  // Recognize pattern:
3807  // mulval = mul(zext A, zext B)
3808  // cmp uge mulval, max+1
3809  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3810  APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
3811  if (MaxVal.eq(CI->getValue()))
3812  break; // Recognized
3813  }
3814  return nullptr;
3815 
3816  case ICmpInst::ICMP_ULE:
3817  // Recognize pattern:
3818  // mulval = mul(zext A, zext B)
3819  // cmp ule mulval, max
3820  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3821  APInt MaxVal = APInt::getMaxValue(MulWidth);
3822  MaxVal = MaxVal.zext(CI->getBitWidth());
3823  if (MaxVal.eq(CI->getValue()))
3824  break; // Recognized
3825  }
3826  return nullptr;
3827 
3828  case ICmpInst::ICMP_ULT:
3829  // Recognize pattern:
3830  // mulval = mul(zext A, zext B)
3831  // cmp ule mulval, max + 1
3832  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3833  APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
3834  if (MaxVal.eq(CI->getValue()))
3835  break; // Recognized
3836  }
3837  return nullptr;
3838 
3839  default:
3840  return nullptr;
3841  }
3842 
3843  InstCombiner::BuilderTy &Builder = IC.Builder;
3844  Builder.SetInsertPoint(MulInstr);
3845 
3846  // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
3847  Value *MulA = A, *MulB = B;
3848  if (WidthA < MulWidth)
3849  MulA = Builder.CreateZExt(A, MulType);
3850  if (WidthB < MulWidth)
3851  MulB = Builder.CreateZExt(B, MulType);
3853  Intrinsic::umul_with_overflow, MulType);
3854  CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
3855  IC.Worklist.Add(MulInstr);
3856 
3857  // If there are uses of mul result other than the comparison, we know that
3858  // they are truncation or binary AND. Change them to use result of
3859  // mul.with.overflow and adjust properly mask/size.
3860  if (MulVal->hasNUsesOrMore(2)) {
3861  Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
3862  for (User *U : MulVal->users()) {
3863  if (U == &I || U == OtherVal)
3864  continue;
3865  if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
3866  if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
3867  IC.replaceInstUsesWith(*TI, Mul);
3868  else
3869  TI->setOperand(0, Mul);
3870  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
3871  assert(BO->getOpcode() == Instruction::And);
3872  // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
3873  ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
3874  APInt ShortMask = CI->getValue().trunc(MulWidth);
3875  Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
3876  Instruction *Zext =
3877  cast<Instruction>(Builder.CreateZExt(ShortAnd, BO->getType()));
3878  IC.Worklist.Add(Zext);
3879  IC.replaceInstUsesWith(*BO, Zext);
3880  } else {
3881  llvm_unreachable("Unexpected Binary operation");
3882  }
3883  IC.Worklist.Add(cast<Instruction>(U));
3884  }
3885  }
3886  if (isa<Instruction>(OtherVal))
3887  IC.Worklist.Add(cast<Instruction>(OtherVal));
3888 
3889  // The original icmp gets replaced with the overflow value, maybe inverted
3890  // depending on predicate.
3891  bool Inverse = false;
3892  switch (I.getPredicate()) {
3893  case ICmpInst::ICMP_NE:
3894  break;
3895  case ICmpInst::ICMP_EQ:
3896  Inverse = true;
3897  break;
3898  case ICmpInst::ICMP_UGT:
3899  case ICmpInst::ICMP_UGE:
3900  if (I.getOperand(0) == MulVal)
3901  break;
3902  Inverse = true;
3903  break;
3904  case ICmpInst::ICMP_ULT:
3905  case ICmpInst::ICMP_ULE:
3906  if (I.getOperand(1) == MulVal)
3907  break;
3908  Inverse = true;
3909  break;
3910  default:
3911  llvm_unreachable("Unexpected predicate");
3912  }
3913  if (Inverse) {
3914  Value *Res = Builder.CreateExtractValue(Call, 1);
3915  return BinaryOperator::CreateNot(Res);
3916  }
3917 
3918  return ExtractValueInst::Create(Call, 1);
3919 }
3920 
3921 /// When performing a comparison against a constant, it is possible that not all
3922 /// the bits in the LHS are demanded. This helper method computes the mask that
3923 /// IS demanded.
3924 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
3925  const APInt *RHS;
3926  if (!match(I.getOperand(1), m_APInt(RHS)))
3927  return APInt::getAllOnesValue(BitWidth);
3928 
3929  // If this is a normal comparison, it demands all bits. If it is a sign bit
3930  // comparison, it only demands the sign bit.
3931  bool UnusedBit;
3932  if (isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
3933  return APInt::getSignMask(BitWidth);
3934 
3935  switch (I.getPredicate()) {
3936  // For a UGT comparison, we don't care about any bits that
3937  // correspond to the trailing ones of the comparand. The value of these
3938  // bits doesn't impact the outcome of the comparison, because any value
3939  // greater than the RHS must differ in a bit higher than these due to carry.
3940  case ICmpInst::ICMP_UGT:
3941  return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
3942 
3943  // Similarly, for a ULT comparison, we don't care about the trailing zeros.
3944  // Any value less than the RHS must differ in a higher bit because of carries.
3945  case ICmpInst::ICMP_ULT:
3946  return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
3947 
3948  default:
3949  return APInt::getAllOnesValue(BitWidth);
3950  }
3951 }
3952 
3953 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
3954 /// should be swapped.
3955 /// The decision is based on how many times these two operands are reused
3956 /// as subtract operands and their positions in those instructions.
3957 /// The rational is that several architectures use the same instruction for
3958 /// both subtract and cmp, thus it is better if the order of those operands
3959 /// match.
3960 /// \return true if Op0 and Op1 should be swapped.
3961 static bool swapMayExposeCSEOpportunities(const Value * Op0,
3962  const Value * Op1) {
3963  // Filter out pointer value as those cannot appears directly in subtract.
3964  // FIXME: we may want to go through inttoptrs or bitcasts.
3965  if (Op0->getType()->isPointerTy())
3966  return false;
3967  // Count every uses of both Op0 and Op1 in a subtract.
3968  // Each time Op0 is the first operand, count -1: swapping is bad, the
3969  // subtract has already the same layout as the compare.
3970  // Each time Op0 is the second operand, count +1: swapping is good, the
3971  // subtract has a different layout as the compare.
3972  // At the end, if the benefit is greater than 0, Op0 should come second to
3973  // expose more CSE opportunities.
3974  int GlobalSwapBenefits = 0;
3975  for (const User *U : Op0->users()) {
3976  const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
3977  if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
3978  continue;
3979  // If Op0 is the first argument, this is not beneficial to swap the
3980  // arguments.
3981  int LocalSwapBenefits = -1;
3982  unsigned Op1Idx = 1;
3983  if (BinOp->getOperand(Op1Idx) == Op0) {
3984  Op1Idx = 0;
3985  LocalSwapBenefits = 1;
3986  }
3987  if (BinOp->getOperand(Op1Idx) != Op1)
3988  continue;
3989  GlobalSwapBenefits += LocalSwapBenefits;
3990  }
3991  return GlobalSwapBenefits > 0;
3992 }
3993 
3994 /// \brief Check that one use is in the same block as the definition and all
3995 /// other uses are in blocks dominated by a given block.
3996 ///
3997 /// \param DI Definition
3998 /// \param UI Use
3999 /// \param DB Block that must dominate all uses of \p DI outside
4000 /// the parent block
4001 /// \return true when \p UI is the only use of \p DI in the parent block
4002 /// and all other uses of \p DI are in blocks dominated by \p DB.
4003 ///
4005  const Instruction *UI,
4006  const BasicBlock *DB) const {
4007  assert(DI && UI && "Instruction not defined\n");
4008  // Ignore incomplete definitions.
4009  if (!DI->getParent())
4010  return false;
4011  // DI and UI must be in the same block.
4012  if (DI->getParent() != UI->getParent())
4013  return false;
4014  // Protect from self-referencing blocks.
4015  if (DI->getParent() == DB)
4016  return false;
4017  for (const User *U : DI->users()) {
4018  auto *Usr = cast<Instruction>(U);
4019  if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4020  return false;
4021  }
4022  return true;
4023 }
4024 
4025 /// Return true when the instruction sequence within a block is select-cmp-br.
4026 static bool isChainSelectCmpBranch(const SelectInst *SI) {
4027  const BasicBlock *BB = SI->getParent();
4028  if (!BB)
4029  return false;
4030  auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4031  if (!BI || BI->getNumSuccessors() != 2)
4032  return false;
4033  auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4034  if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4035  return false;
4036  return true;
4037 }
4038 
4039 /// \brief True when a select result is replaced by one of its operands
4040 /// in select-icmp sequence. This will eventually result in the elimination
4041 /// of the select.
4042 ///
4043 /// \param SI Select instruction
4044 /// \param Icmp Compare instruction
4045 /// \param SIOpd Operand that replaces the select
4046 ///
4047 /// Notes:
4048 /// - The replacement is global and requires dominator information
4049 /// - The caller is responsible for the actual replacement
4050 ///
4051 /// Example:
4052 ///
4053 /// entry:
4054 /// %4 = select i1 %3, %C* %0, %C* null
4055 /// %5 = icmp eq %C* %4, null
4056 /// br i1 %5, label %9, label %7
4057 /// ...
4058 /// ; <label>:7 ; preds = %entry
4059 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4060 /// ...
4061 ///
4062 /// can be transformed to
4063 ///
4064 /// %5 = icmp eq %C* %0, null
4065 /// %6 = select i1 %3, i1 %5, i1 true
4066 /// br i1 %6, label %9, label %7
4067 /// ...
4068 /// ; <label>:7 ; preds = %entry
4069 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
4070 ///
4071 /// Similar when the first operand of the select is a constant or/and
4072 /// the compare is for not equal rather than equal.
4073 ///
4074 /// NOTE: The function is only called when the select and compare constants
4075 /// are equal, the optimization can work only for EQ predicates. This is not a
4076 /// major restriction since a NE compare should be 'normalized' to an equal
4077 /// compare, which usually happens in the combiner and test case
4078 /// select-cmp-br.ll checks for it.
4080  const ICmpInst *Icmp,
4081  const unsigned SIOpd) {
4082  assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
4083  if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4084  BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4085  // The check for the single predecessor is not the best that can be
4086  // done. But it protects efficiently against cases like when SI's
4087  // home block has two successors, Succ and Succ1, and Succ1 predecessor
4088  // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4089  // replaced can be reached on either path. So the uniqueness check
4090  // guarantees that the path all uses of SI (outside SI's parent) are on
4091  // is disjoint from all other paths out of SI. But that information
4092  // is more expensive to compute, and the trade-off here is in favor
4093  // of compile-time. It should also be noticed that we check for a single
4094  // predecessor and not only uniqueness. This to handle the situation when
4095  // Succ and Succ1 points to the same basic block.
4096  if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
4097  NumSel++;
4098  SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
4099  return true;
4100  }
4101  }
4102  return false;
4103 }
4104 
4105 /// Try to fold the comparison based on range information we can get by checking
4106 /// whether bits are known to be zero or one in the inputs.
4107 Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) {
4108  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4109  Type *Ty = Op0->getType();
4110  ICmpInst::Predicate Pred = I.getPredicate();
4111 
4112  // Get scalar or pointer size.
4113  unsigned BitWidth = Ty->isIntOrIntVectorTy()
4114  ? Ty->getScalarSizeInBits()
4115  : DL.getTypeSizeInBits(Ty->getScalarType());
4116 
4117  if (!BitWidth)
4118  return nullptr;
4119 
4120  KnownBits Op0Known(BitWidth);
4121  KnownBits Op1Known(BitWidth);
4122 
4123  if (SimplifyDemandedBits(&I, 0,
4124  getDemandedBitsLHSMask(I, BitWidth),
4125  Op0Known, 0))
4126  return &I;
4127 
4128  if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
4129  Op1Known, 0))
4130  return &I;
4131 
4132  // Given the known and unknown bits, compute a range that the LHS could be
4133  // in. Compute the Min, Max and RHS values based on the known bits. For the
4134  // EQ and NE we use unsigned values.
4135  APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
4136  APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
4137  if (I.isSigned()) {
4138  computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4139  computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4140  } else {
4141  computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4142  computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4143  }
4144 
4145  // If Min and Max are known to be the same, then SimplifyDemandedBits
4146  // figured out that the LHS is a constant. Constant fold this now, so that
4147  // code below can assume that Min != Max.
4148  if (!isa<Constant>(Op0) && Op0Min == Op0Max)
4149  return new ICmpInst(Pred, ConstantInt::get(Op0->getType(), Op0Min), Op1);
4150  if (!isa<Constant>(Op1) && Op1Min == Op1Max)
4151  return new ICmpInst(Pred, Op0, ConstantInt::get(Op1->getType(), Op1Min));
4152 
4153  // Based on the range information we know about the LHS, see if we can
4154  // simplify this comparison. For example, (x&4) < 8 is always true.
4155  switch (Pred) {
4156  default:
4157  llvm_unreachable("Unknown icmp opcode!");
4158  case ICmpInst::ICMP_EQ:
4159  case ICmpInst::ICMP_NE: {
4160  if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
4161  return Pred == CmpInst::ICMP_EQ
4162  ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
4163  : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4164  }
4165 
4166  // If all bits are known zero except for one, then we know at most one bit
4167  // is set. If the comparison is against zero, then this is a check to see if
4168  // *that* bit is set.
4169  APInt Op0KnownZeroInverted = ~Op0Known.Zero;
4170  if (Op1Known.isZero()) {
4171  // If the LHS is an AND with the same constant, look through it.
4172  Value *LHS = nullptr;
4173  const APInt *LHSC;
4174  if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
4175  *LHSC != Op0KnownZeroInverted)
4176  LHS = Op0;
4177 
4178  Value *X;
4179  if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
4180  APInt ValToCheck = Op0KnownZeroInverted;
4181  Type *XTy = X->getType();
4182  if (ValToCheck.isPowerOf2()) {
4183  // ((1 << X) & 8) == 0 -> X != 3
4184  // ((1 << X) & 8) != 0 -> X == 3
4185  auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4186  auto NewPred = ICmpInst::getInversePredicate(Pred);
4187  return new ICmpInst(NewPred, X, CmpC);
4188  } else if ((++ValToCheck).isPowerOf2()) {
4189  // ((1 << X) & 7) == 0 -> X >= 3
4190  // ((1 << X) & 7) != 0 -> X < 3
4191  auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4192  auto NewPred =
4194  return new ICmpInst(NewPred, X, CmpC);
4195  }
4196  }
4197 
4198  // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
4199  const APInt *CI;
4200  if (Op0KnownZeroInverted.isOneValue() &&
4201  match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
4202  // ((8 >>u X) & 1) == 0 -> X != 3
4203  // ((8 >>u X) & 1) != 0 -> X == 3
4204  unsigned CmpVal = CI->countTrailingZeros();
4205  auto NewPred = ICmpInst::getInversePredicate(Pred);
4206  return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
4207  }
4208  }
4209  break;
4210  }
4211  case ICmpInst::ICMP_ULT: {
4212  if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
4213  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4214  if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
4215  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4216  if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
4217  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4218 
4219  if (Op1Min == Op1Max) {
4220  // A <u C -> A == C-1 if min(A)+1 == C
4221  if (Op1Min == Op0Min + 1)
4222  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4223  ConstantInt::get(Op0->getType(), Op1Min - 1));
4224  }
4225  break;
4226  }
4227  case ICmpInst::ICMP_UGT: {
4228  if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
4229  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4230  if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
4231  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4232  if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
4233  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4234 
4235  if (Op1Min == Op1Max) {
4236  // A >u C -> A == C+1 if max(a)-1 == C
4237  if (Op1Min == Op0Max - 1)
4238  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4239  ConstantInt::get(Op1->getType(), Op1Min + 1));
4240  }
4241  break;
4242  }
4243  case ICmpInst::ICMP_SLT:
4244  if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
4245  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4246  if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
4247  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4248  if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
4249  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4250  if (Op1Min == Op1Max) { // Constant RHS
4251  if (Op1Min == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
4252  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4253  ConstantInt::get(Op1->getType(), Op1Min - 1));
4254  }
4255  break;
4256  case ICmpInst::ICMP_SGT:
4257  if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
4258  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4259  if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
4260  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4261  if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
4262  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4263  if (Op1Min == Op1Max) { // Constant RHS
4264  if (Op1Min == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
4265  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4266  ConstantInt::get(Op1->getType(), Op1Min + 1));
4267  }
4268  break;
4269  case ICmpInst::ICMP_SGE:
4270  assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
4271  if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
4272  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4273  if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
4274  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4275  if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
4276  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4277  break;
4278  case ICmpInst::ICMP_SLE:
4279  assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
4280  if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
4281  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4282  if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
4283  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4284  if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
4285  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4286  break;
4287  case ICmpInst::ICMP_UGE:
4288  assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
4289  if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
4290  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4291  if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
4292  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4293  if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
4294  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4295  break;
4296  case ICmpInst::ICMP_ULE:
4297  assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
4298  if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
4299  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4300  if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
4301  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4302  if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
4303  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4304  break;
4305  }
4306 
4307  // Turn a signed comparison into an unsigned one if both operands are known to
4308  // have the same sign.
4309  if (I.isSigned() &&
4310  ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
4311  (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
4312  return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
4313 
4314  return nullptr;
4315 }
4316 
4317 /// If we have an icmp le or icmp ge instruction with a constant operand, turn
4318 /// it into the appropriate icmp lt or icmp gt instruction. This transform
4319 /// allows them to be folded in visitICmpInst.
4321  ICmpInst::Predicate Pred = I.getPredicate();
4322  if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGE &&
4323  Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_UGE)
4324  return nullptr;
4325 
4326  Value *Op0 = I.getOperand(0);
4327  Value *Op1 = I.getOperand(1);
4328  auto *Op1C = dyn_cast<Constant>(Op1);
4329  if (!Op1C)
4330  return nullptr;
4331 
4332  // Check if the constant operand can be safely incremented/decremented without
4333  // overflowing/underflowing. For scalars, SimplifyICmpInst has already handled
4334  // the edge cases for us, so we just assert on them. For vectors, we must
4335  // handle the edge cases.
4336  Type *Op1Type = Op1->getType();
4337  bool IsSigned = I.isSigned();
4338  bool IsLE = (Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_ULE);
4339  auto *CI = dyn_cast<ConstantInt>(Op1C);
4340  if (CI) {
4341  // A <= MAX -> TRUE ; A >= MIN -> TRUE
4342  assert(IsLE ? !CI->isMaxValue(IsSigned) : !CI->isMinValue(IsSigned));
4343  } else if (Op1Type->isVectorTy()) {
4344  // TODO? If the edge cases for vectors were guaranteed to be handled as they
4345  // are for scalar, we could remove the min/max checks. However, to do that,
4346  // we would have to use insertelement/shufflevector to replace edge values.
4347  unsigned NumElts = Op1Type->getVectorNumElements();
4348  for (unsigned i = 0; i != NumElts; ++i) {
4349  Constant *Elt = Op1C->getAggregateElement(i);
4350  if (!Elt)
4351  return nullptr;
4352 
4353  if (isa<UndefValue>(Elt))
4354  continue;
4355 
4356  // Bail out if we can't determine if this constant is min/max or if we
4357  // know that this constant is min/max.
4358  auto *CI = dyn_cast<ConstantInt>(Elt);
4359  if (!CI || (IsLE ? CI->isMaxValue(IsSigned) : CI->isMinValue(IsSigned)))
4360  return nullptr;
4361  }
4362  } else {
4363  // ConstantExpr?
4364  return nullptr;
4365  }
4366 
4367  // Increment or decrement the constant and set the new comparison predicate:
4368  // ULE -> ULT ; UGE -> UGT ; SLE -> SLT ; SGE -> SGT
4369  Constant *OneOrNegOne = ConstantInt::get(Op1Type, IsLE ? 1 : -1, true);
4371  NewPred = IsSigned ? ICmpInst::getSignedPredicate(NewPred) : NewPred;
4372  return new ICmpInst(NewPred, Op0, ConstantExpr::getAdd(Op1C, OneOrNegOne));
4373 }
4374 
4375 /// Integer compare with boolean values can always be turned into bitwise ops.
4377  InstCombiner::BuilderTy &Builder) {
4378  Value *A = I.getOperand(0), *B = I.getOperand(1);
4379  assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only");
4380 
4381  // A boolean compared to true/false can be simplified to Op0/true/false in
4382  // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
4383  // Cases not handled by InstSimplify are always 'not' of Op0.
4384  if (match(B, m_Zero())) {
4385  switch (I.getPredicate()) {
4386  case CmpInst::ICMP_EQ: // A == 0 -> !A
4387  case CmpInst::ICMP_ULE: // A <=u 0 -> !A
4388  case CmpInst::ICMP_SGE: // A >=s 0 -> !A
4389  return BinaryOperator::CreateNot(A);
4390  default:
4391  llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4392  }
4393  } else if (match(B, m_One())) {
4394  switch (I.getPredicate()) {
4395  case CmpInst::ICMP_NE: // A != 1 -> !A
4396  case CmpInst::ICMP_ULT: // A <u 1 -> !A
4397  case CmpInst::ICMP_SGT: // A >s -1 -> !A
4398  return BinaryOperator::CreateNot(A);
4399  default:
4400  llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4401  }
4402  }
4403 
4404  switch (I.getPredicate()) {
4405  default:
4406  llvm_unreachable("Invalid icmp instruction!");
4407  case ICmpInst::ICMP_EQ:
4408  // icmp eq i1 A, B -> ~(A ^ B)
4409  return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
4410 
4411  case ICmpInst::ICMP_NE:
4412  // icmp ne i1 A, B -> A ^ B
4413  return BinaryOperator::CreateXor(A, B);
4414 
4415  case ICmpInst::ICMP_UGT:
4416  // icmp ugt -> icmp ult
4417  std::swap(A, B);
4419  case ICmpInst::ICMP_ULT:
4420  // icmp ult i1 A, B -> ~A & B
4421  return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
4422 
4423  case ICmpInst::ICMP_SGT:
4424  // icmp sgt -> icmp slt
4425  std::swap(A, B);
4427  case ICmpInst::ICMP_SLT:
4428  // icmp slt i1 A, B -> A & ~B
4429  return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
4430 
4431  case ICmpInst::ICMP_UGE:
4432  // icmp uge -> icmp ule
4433  std::swap(A, B);
4435  case ICmpInst::ICMP_ULE:
4436  // icmp ule i1 A, B -> ~A | B
4437  return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
4438 
4439  case ICmpInst::ICMP_SGE:
4440  // icmp sge -> icmp sle
4441  std::swap(A, B);
4443  case ICmpInst::ICMP_SLE:
4444  // icmp sle i1 A, B -> A | ~B
4445  return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
4446  }
4447 }
4448 
4450  bool Changed = false;
4451  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4452  unsigned Op0Cplxity = getComplexity(Op0);
4453  unsigned Op1Cplxity = getComplexity(Op1);
4454 
4455  /// Orders the operands of the compare so that they are listed from most
4456  /// complex to least complex. This puts constants before unary operators,
4457  /// before binary operators.
4458  if (Op0Cplxity < Op1Cplxity ||
4459  (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
4460  I.swapOperands();
4461  std::swap(Op0, Op1);
4462  Changed = true;
4463  }
4464 
4465  if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1,
4466  SQ.getWithInstruction(&I)))
4467  return replaceInstUsesWith(I, V);
4468 
4469  // Comparing -val or val with non-zero is the same as just comparing val
4470  // ie, abs(val) != 0 -> val != 0
4471  if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
4472  Value *Cond, *SelectTrue, *SelectFalse;
4473  if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
4474  m_Value(SelectFalse)))) {
4475  if (Value *V = dyn_castNegVal(SelectTrue)) {
4476  if (V == SelectFalse)
4477  return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4478  }
4479  else if (Value *V = dyn_castNegVal(SelectFalse)) {
4480  if (V == SelectTrue)
4481  return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4482  }
4483  }
4484  }
4485 
4486  if (Op0->getType()->isIntOrIntVectorTy(1))
4487  if (Instruction *Res = canonicalizeICmpBool(I, Builder))
4488  return Res;
4489 
4490  if (ICmpInst *NewICmp = canonicalizeCmpWithConstant(I))
4491  return NewICmp;
4492 
4493  if (Instruction *Res = foldICmpWithConstant(I))
4494  return Res;
4495 
4496  if (Instruction *Res = foldICmpUsingKnownBits(I))
4497  return Res;
4498 
4499  // Test if the ICmpInst instruction is used exclusively by a select as
4500  // part of a minimum or maximum operation. If so, refrain from doing
4501  // any other folding. This helps out other analyses which understand
4502  // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
4503  // and CodeGen. And in this case, at least one of the comparison
4504  // operands has at least one user besides the compare (the select),
4505  // which would often largely negate the benefit of folding anyway.
4506  if (I.hasOneUse())
4507  if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
4508  if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
4509  (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
4510  return nullptr;
4511 
4512  // FIXME: We only do this after checking for min/max to prevent infinite
4513  // looping caused by a reverse canonicalization of these patterns for min/max.
4514  // FIXME: The organization of folds is a mess. These would naturally go into
4515  // canonicalizeCmpWithConstant(), but we can't move all of the above folds
4516  // down here after the min/max restriction.
4517  ICmpInst::Predicate Pred = I.getPredicate();
4518  const APInt *C;
4519  if (match(Op1, m_APInt(C))) {
4520  // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
4521  if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
4522  Constant *Zero = Constant::getNullValue(Op0->getType());
4523  return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
4524  }
4525 
4526  // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
4527  if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
4528  Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
4529  return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
4530  }
4531  }
4532 
4533  if (Instruction *Res = foldICmpInstWithConstant(I))
4534  return Res;
4535 
4536  if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
4537  return Res;
4538 
4539  // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4540  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
4541  if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I))
4542  return NI;
4543  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
4544  if (Instruction *NI = foldGEPICmp(GEP, Op0,
4546  return NI;
4547 
4548  // Try to optimize equality comparisons against alloca-based pointers.
4549  if (Op0->getType()->isPointerTy() && I.isEquality()) {
4550  assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?");
4551  if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op0, DL)))
4552  if (Instruction *New = foldAllocaCmp(I, Alloca, Op1))
4553  return New;
4554  if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op1, DL)))
4555  if (Instruction *New = foldAllocaCmp(I, Alloca, Op0))
4556  return New;
4557  }
4558 
4559  // Test to see if the operands of the icmp are casted versions of other
4560  // values. If the ptr->ptr cast can be stripped off both arguments, we do so
4561  // now.
4562  if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
4563  if (Op0->getType()->isPointerTy() &&
4564  (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
4565  // We keep moving the cast from the left operand over to the right
4566  // operand, where it can often be eliminated completely.
4567  Op0 = CI->getOperand(0);
4568 
4569  // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
4570  // so eliminate it as well.
4571  if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
4572  Op1 = CI2->getOperand(0);
4573 
4574  // If Op1 is a constant, we can fold the cast into the constant.
4575  if (Op0->getType() != Op1->getType()) {
4576  if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
4577  Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
4578  } else {
4579  // Otherwise, cast the RHS right before the icmp
4580  Op1 = Builder.CreateBitCast(Op1, Op0->getType());
4581  }
4582  }
4583  return new ICmpInst(I.getPredicate(), Op0, Op1);
4584  }
4585  }
4586 
4587  if (isa<CastInst>(Op0)) {
4588  // Handle the special case of: icmp (cast bool to X), <cst>
4589  // This comes up when you have code like
4590  // int X = A < B;
4591  // if (X) ...
4592  // For generality, we handle any zero-extension of any operand comparison
4593  // with a constant or another cast from the same type.
4594  if (isa<Constant>(Op1) || isa<CastInst>(Op1))
4595  if (Instruction *R = foldICmpWithCastAndCast(I))
4596  return R;
4597  }
4598 
4599  if (Instruction *Res = foldICmpBinOp(I))
4600  return Res;
4601 
4602  if (Instruction *Res = foldICmpWithMinMax(I))
4603  return Res;
4604 
4605  {
4606  Value *A, *B;
4607  // Transform (A & ~B) == 0 --> (A & B) != 0
4608  // and (A & ~B) != 0 --> (A & B) == 0
4609  // if A is a power of 2.
4610  if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
4611  match(Op1, m_Zero()) &&
4612  isKnownToBeAPowerOfTwo(A, false, 0, &I) && I.isEquality())
4613  return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(A, B),
4614  Op1);
4615 
4616  // ~X < ~Y --> Y < X
4617  // ~X < C --> X > ~C
4618  if (match(Op0, m_Not(m_Value(A)))) {
4619  if (match(Op1, m_Not(m_Value(B))))
4620  return new ICmpInst(I.getPredicate(), B, A);
4621 
4622  const APInt *C;
4623  if (match(Op1, m_APInt(C)))
4624  return new ICmpInst(I.getSwappedPredicate(), A,
4625  ConstantInt::get(Op1->getType(), ~(*C)));
4626  }
4627 
4628  Instruction *AddI = nullptr;
4629  if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
4630  m_Instruction(AddI))) &&
4631  isa<IntegerType>(A->getType())) {
4632  Value *Result;
4633  Constant *Overflow;
4634  if (OptimizeOverflowCheck(OCF_UNSIGNED_ADD, A, B, *AddI, Result,
4635  Overflow)) {
4636  replaceInstUsesWith(*AddI, Result);
4637  return replaceInstUsesWith(I, Overflow);
4638  }
4639  }
4640 
4641  // (zext a) * (zext b) --> llvm.umul.with.overflow.
4642  if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
4643  if (Instruction *R = processUMulZExtIdiom(I, Op0, Op1, *this))
4644  return R;
4645  }
4646  if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
4647  if (Instruction *R = processUMulZExtIdiom(I, Op1, Op0, *this))
4648  return R;
4649  }
4650  }
4651 
4652  if (Instruction *Res = foldICmpEquality(I))
4653  return Res;
4654 
4655  // The 'cmpxchg' instruction returns an aggregate containing the old value and
4656  // an i1 which indicates whether or not we successfully did the swap.
4657  //
4658  // Replace comparisons between the old value and the expected value with the
4659  // indicator that 'cmpxchg' returns.
4660  //
4661  // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
4662  // spuriously fail. In those cases, the old value may equal the expected
4663  // value but it is possible for the swap to not occur.
4664  if (I.getPredicate() == ICmpInst::ICMP_EQ)
4665  if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
4666  if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
4667  if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
4668  !ACXI->isWeak())
4669  return ExtractValueInst::Create(ACXI, 1);
4670 
4671  {
4672  Value *X; ConstantInt *Cst;
4673  // icmp X+Cst, X
4674  if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
4675  return foldICmpAddOpConst(X, Cst, I.getPredicate());
4676 
4677  // icmp X, X+Cst
4678  if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
4679  return foldICmpAddOpConst(X, Cst, I.getSwappedPredicate());
4680  }
4681  return Changed ? &I : nullptr;
4682 }
4683 
4684 /// Fold fcmp ([us]itofp x, cst) if possible.
4685 Instruction *InstCombiner::foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI,
4686  Constant *RHSC) {
4687  if (!isa<ConstantFP>(RHSC)) return nullptr;
4688  const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
4689 
4690  // Get the width of the mantissa. We don't want to hack on conversions that
4691  // might lose information from the integer, e.g. "i64 -> float"
4692  int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
4693  if (MantissaWidth == -1) return nullptr; // Unknown.
4694 
4695  IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
4696 
4697  bool LHSUnsigned = isa<UIToFPInst>(LHSI);
4698 
4699  if (I.isEquality()) {
4701  bool IsExact = false;
4702  APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned);
4703  RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
4704 
4705  // If the floating point constant isn't an integer value, we know if we will
4706  // ever compare equal / not equal to it.
4707  if (!IsExact) {
4708  // TODO: Can never be -0.0 and other non-representable values
4709  APFloat RHSRoundInt(RHS);
4711  if (RHS.compare(RHSRoundInt) != APFloat::cmpEqual) {
4712  if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
4713  return replaceInstUsesWith(I, Builder.getFalse());
4714 
4716  return replaceInstUsesWith(I, Builder.getTrue());
4717  }
4718  }
4719 
4720  // TODO: If the constant is exactly representable, is it always OK to do
4721  // equality compares as integer?
4722  }
4723 
4724  // Check to see that the input is converted from an integer type that is small
4725  // enough that preserves all bits. TODO: check here for "known" sign bits.
4726  // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
4727  unsigned InputSize = IntTy->getScalarSizeInBits();
4728 
4729  // Following test does NOT adjust InputSize downwards for signed inputs,
4730  // because the most negative value still requires all the mantissa bits
4731  // to distinguish it from one less than that value.
4732  if ((int)InputSize > MantissaWidth) {
4733  // Conversion would lose accuracy. Check if loss can impact comparison.
4734  int Exp = ilogb(RHS);
4735  if (Exp == APFloat::IEK_Inf) {
4736  int MaxExponent = ilogb(APFloat::getLargest(RHS.getSemantics()));
4737  if (MaxExponent < (int)InputSize - !LHSUnsigned)
4738  // Conversion could create infinity.
4739  return nullptr;
4740  } else {
4741  // Note that if RHS is zero or NaN, then Exp is negative
4742  // and first condition is trivially false.
4743  if (MantissaWidth <= Exp && Exp <= (int)InputSize - !LHSUnsigned)
4744  // Conversion could affect comparison.
4745  return nullptr;
4746  }
4747  }
4748 
4749  // Otherwise, we can potentially simplify the comparison. We know that it
4750  // will always come through as an integer value and we know the constant is
4751  // not a NAN (it would have been previously simplified).
4752  assert(!RHS.isNaN() && "NaN comparison not already folded!");
4753 
4754  ICmpInst::Predicate Pred;
4755  switch (I.getPredicate()) {
4756  default: llvm_unreachable("Unexpected predicate!");
4757  case FCmpInst::FCMP_UEQ:
4758  case FCmpInst::FCMP_OEQ:
4759  Pred = ICmpInst::ICMP_EQ;
4760  break;
4761  case FCmpInst::FCMP_UGT:
4762  case FCmpInst::FCMP_OGT:
4763  Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
4764  break;
4765  case FCmpInst::FCMP_UGE:
4766  case FCmpInst::FCMP_OGE:
4767  Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
4768  break;
4769  case FCmpInst::FCMP_ULT:
4770  case FCmpInst::FCMP_OLT:
4771  Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
4772  break;
4773  case FCmpInst::FCMP_ULE:
4774  case FCmpInst::FCMP_OLE:
4775  Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
4776  break;
4777  case FCmpInst::FCMP_UNE:
4778  case FCmpInst::FCMP_ONE:
4779  Pred = ICmpInst::ICMP_NE;
4780  break;
4781  case FCmpInst::FCMP_ORD:
4782  return replaceInstUsesWith(I, Builder.getTrue());
4783  case FCmpInst::FCMP_UNO:
4784  return replaceInstUsesWith(I, Builder.getFalse());
4785  }
4786 
4787  // Now we know that the APFloat is a normal number, zero or inf.
4788 
4789  // See if the FP constant is too large for the integer. For example,
4790  // comparing an i8 to 300.0.
4791  unsigned IntWidth = IntTy->getScalarSizeInBits();
4792 
4793  if (!LHSUnsigned) {
4794  // If the RHS value is > SignedMax, fold the comparison. This handles +INF
4795  // and large values.
4796  APFloat SMax(RHS.getSemantics());
4797  SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
4799  if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
4800  if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
4801  Pred == ICmpInst::ICMP_SLE)
4802  return replaceInstUsesWith(I, Builder.getTrue());
4803  return replaceInstUsesWith(I, Builder.getFalse());
4804  }
4805  } else {
4806  // If the RHS value is > UnsignedMax, fold the comparison. This handles
4807  // +INF and large values.
4808  APFloat UMax(RHS.getSemantics());
4809  UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
4811  if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
4812  if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
4813  Pred == ICmpInst::ICMP_ULE)
4814  return replaceInstUsesWith(I, Builder.getTrue());
4815  return replaceInstUsesWith(I, Builder.getFalse());
4816  }
4817  }
4818 
4819  if (!LHSUnsigned) {
4820  // See if the RHS value is < SignedMin.
4821  APFloat SMin(RHS.getSemantics());
4822  SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
4824  if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
4825  if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
4826  Pred == ICmpInst::ICMP_SGE)
4827  return replaceInstUsesWith(I, Builder.getTrue());
4828  return replaceInstUsesWith(I, Builder.getFalse());
4829  }
4830  } else {
4831  // See if the RHS value is < UnsignedMin.
4832  APFloat SMin(RHS.getSemantics());
4833  SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
4835  if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
4836  if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
4837  Pred == ICmpInst::ICMP_UGE)
4838  return replaceInstUsesWith(I, Builder.getTrue());
4839  return replaceInstUsesWith(I, Builder.getFalse());
4840  }
4841  }
4842 
4843  // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
4844  // [0, UMAX], but it may still be fractional. See if it is fractional by
4845  // casting the FP value to the integer value and back, checking for equality.
4846  // Don't do this for zero, because -0.0 is not fractional.
4847  Constant *RHSInt = LHSUnsigned
4848  ? ConstantExpr::getFPToUI(RHSC, IntTy)
4849  : ConstantExpr::getFPToSI(RHSC, IntTy);
4850  if (!RHS.isZero()) {
4851  bool Equal = LHSUnsigned
4852  ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
4853  : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
4854  if (!Equal) {
4855  // If we had a comparison against a fractional value, we have to adjust
4856  // the compare predicate and sometimes the value. RHSC is rounded towards
4857  // zero at this point.
4858  switch (Pred) {
4859  default: llvm_unreachable("Unexpected integer comparison!");
4860  case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
4861  return replaceInstUsesWith(I, Builder.getTrue());
4862  case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
4863  return replaceInstUsesWith(I, Builder.getFalse());
4864  case ICmpInst::ICMP_ULE:
4865  // (float)int <= 4.4 --> int <= 4
4866  // (float)int <= -4.4 --> false
4867  if (RHS.isNegative())
4868  return replaceInstUsesWith(I, Builder.getFalse());
4869  break;
4870  case ICmpInst::ICMP_SLE:
4871  // (float)int <= 4.4 --> int <= 4
4872  // (float)int <= -4.4 --> int < -4
4873  if (RHS.isNegative())
4874  Pred = ICmpInst::ICMP_SLT;
4875  break;
4876  case ICmpInst::ICMP_ULT:
4877  // (float)int < -4.4 --> false
4878  // (float)int < 4.4 --> int <= 4
4879  if (RHS.isNegative())
4880  return replaceInstUsesWith(I, Builder.getFalse());
4881  Pred = ICmpInst::ICMP_ULE;
4882  break;
4883  case ICmpInst::ICMP_SLT:
4884  // (float)int < -4.4 --> int < -4
4885  // (float)int < 4.4 --> int <= 4
4886  if (!RHS.isNegative())
4887  Pred = ICmpInst::ICMP_SLE;
4888  break;
4889  case ICmpInst::ICMP_UGT:
4890  // (float)int > 4.4 --> int > 4
4891  // (float)int > -4.4 --> true
4892  if (RHS.isNegative())
4893  return replaceInstUsesWith(I, Builder.getTrue());
4894  break;
4895  case ICmpInst::ICMP_SGT:
4896  // (float)int > 4.4 --> int > 4
4897  // (float)int > -4.4 --> int >= -4
4898  if (RHS.isNegative())
4899  Pred = ICmpInst::ICMP_SGE;
4900  break;
4901  case ICmpInst::ICMP_UGE:
4902  // (float)int >= -4.4 --> true
4903  // (float)int >= 4.4 --> int > 4
4904  if (RHS.isNegative())
4905  return replaceInstUsesWith(I, Builder.getTrue());
4906  Pred = ICmpInst::ICMP_UGT;
4907  break;
4908  case ICmpInst::ICMP_SGE:
4909  // (float)int >= -4.4 --> int >= -4
4910  // (float)int >= 4.4 --> int > 4
4911  if (!RHS.isNegative())
4912  Pred = ICmpInst::ICMP_SGT;
4913  break;
4914  }
4915  }
4916  }
4917 
4918  // Lower this FP comparison into an appropriate integer version of the
4919  // comparison.
4920  return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
4921 }
4922 
4924  bool Changed = false;
4925 
4926  /// Orders the operands of the compare so that they are listed from most
4927  /// complex to least complex. This puts constants before unary operators,
4928  /// before binary operators.
4929  if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
4930  I.swapOperands();
4931  Changed = true;
4932  }
4933 
4934  const CmpInst::Predicate Pred = I.getPredicate();
4935  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4936  if (Value *V = SimplifyFCmpInst(Pred, Op0, Op1, I.getFastMathFlags(),
4937  SQ.getWithInstruction(&I)))
4938  return replaceInstUsesWith(I, V);
4939 
4940  // Simplify 'fcmp pred X, X'
4941  if (Op0 == Op1) {
4942  switch (Pred) {
4943  default: break;
4944  case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4945  case FCmpInst::FCMP_ULT: // True if unordered or less than
4946  case FCmpInst::FCMP_UGT: // True if unordered or greater than
4947  case FCmpInst::FCMP_UNE: // True if unordered or not equal
4948  // Canonicalize these to be 'fcmp uno %X, 0.0'.
4951  return &I;
4952 
4953  case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4954  case FCmpInst::FCMP_OEQ: // True if ordered and equal
4955  case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4956  case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4957  // Canonicalize these to be 'fcmp ord %X, 0.0'.
4960  return &I;
4961  }
4962  }
4963 
4964  // If we're just checking for a NaN (ORD/UNO) and have a non-NaN operand,
4965  // then canonicalize the operand to 0.0.
4966  if (Pred == CmpInst::FCMP_ORD || Pred == CmpInst::FCMP_UNO) {
4967  if (!match(Op0, m_Zero()) && isKnownNeverNaN(Op0)) {
4969  return &I;
4970  }
4971  if (!match(Op1, m_Zero()) && isKnownNeverNaN(Op1)) {
4973  return &I;
4974  }
4975  }
4976 
4977  // Test if the FCmpInst instruction is used exclusively by a select as
4978  // part of a minimum or maximum operation. If so, refrain from doing
4979  // any other folding. This helps out other analyses which understand
4980  // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
4981  // and CodeGen. And in this case, at least one of the comparison
4982  // operands has at least one user besides the compare (the select),
4983  // which would often largely negate the benefit of folding anyway.
4984  if (I.hasOneUse())
4985  if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
4986  if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
4987  (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
4988  return nullptr;
4989 
4990  // Handle fcmp with constant RHS
4991  if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4992  if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4993  switch (LHSI->getOpcode()) {
4994  case Instruction::FPExt: {
4995  // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
4996  FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
4997  ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
4998  if (!RHSF)
4999  break;
5000 
5001  const fltSemantics *Sem;
5002  // FIXME: This shouldn't be here.
5003  if (LHSExt->getSrcTy()->isHalfTy())
5004  Sem = &APFloat::IEEEhalf();
5005  else if (LHSExt->getSrcTy()->isFloatTy())
5006  Sem = &APFloat::IEEEsingle();
5007  else if (LHSExt->getSrcTy()->isDoubleTy())
5008  Sem = &APFloat::IEEEdouble();
5009  else if (LHSExt->getSrcTy()->isFP128Ty())
5010  Sem = &APFloat::IEEEquad();
5011  else if (LHSExt->getSrcTy()->isX86_FP80Ty())
5012  Sem = &APFloat::x87DoubleExtended();
5013  else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
5014  Sem = &APFloat::PPCDoubleDouble();
5015  else
5016  break;
5017 
5018  bool Lossy;
5019  APFloat F = RHSF->getValueAPF();
5020  F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
5021 
5022  // Avoid lossy conversions and denormals. Zero is a special case
5023  // that's OK to convert.
5024  APFloat Fabs = F;
5025  Fabs.clearSign();
5026  if (!Lossy &&
5027  ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
5028  APFloat::cmpLessThan) || Fabs.isZero()))
5029 
5030  return new FCmpInst(Pred, LHSExt->getOperand(0),
5031  ConstantFP::get(RHSC->getContext(), F));
5032  break;
5033  }
5034  case Instruction::PHI:
5035  // Only fold fcmp into the PHI if the phi and fcmp are in the same
5036  // block. If in the same block, we're encouraging jump threading. If
5037  // not, we are just pessimizing the code by making an i1 phi.
5038  if (LHSI->getParent() == I.getParent())
5039  if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
5040  return NV;
5041  break;
5042  case Instruction::SIToFP:
5043  case Instruction::UIToFP:
5044  if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC))
5045  return NV;
5046  break;
5047  case Instruction::FSub: {
5048  // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
5049  Value *Op;
5050  if (match(LHSI, m_FNeg(m_Value(Op))))
5051  return new FCmpInst(I.getSwappedPredicate(), Op,
5052  ConstantExpr::getFNeg(RHSC));
5053  break;
5054  }
5055  case Instruction::Load:
5056  if (GetElementPtrInst *GEP =
5057  dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
5058  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
5059  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
5060  !cast<LoadInst>(LHSI)->isVolatile())
5061  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
5062  return Res;
5063  }
5064  break;
5065  case Instruction::Call: {
5066  if (!RHSC->isNullValue())