LLVM  10.0.0svn
InstCombineCompares.cpp
Go to the documentation of this file.
1 //===- InstCombineCompares.cpp --------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the visitICmp and visitFCmp functions.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/APSInt.h"
15 #include "llvm/ADT/SetVector.h"
16 #include "llvm/ADT/Statistic.h"
20 #include "llvm/IR/ConstantRange.h"
21 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/IR/PatternMatch.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/KnownBits.h"
27 
28 using namespace llvm;
29 using namespace PatternMatch;
30 
31 #define DEBUG_TYPE "instcombine"
32 
33 // How many times is a select replaced by one of its operands?
34 STATISTIC(NumSel, "Number of select opts");
35 
36 
37 /// Compute Result = In1+In2, returning true if the result overflowed for this
38 /// type.
39 static bool addWithOverflow(APInt &Result, const APInt &In1,
40  const APInt &In2, bool IsSigned = false) {
41  bool Overflow;
42  if (IsSigned)
43  Result = In1.sadd_ov(In2, Overflow);
44  else
45  Result = In1.uadd_ov(In2, Overflow);
46 
47  return Overflow;
48 }
49 
50 /// Compute Result = In1-In2, returning true if the result overflowed for this
51 /// type.
52 static bool subWithOverflow(APInt &Result, const APInt &In1,
53  const APInt &In2, bool IsSigned = false) {
54  bool Overflow;
55  if (IsSigned)
56  Result = In1.ssub_ov(In2, Overflow);
57  else
58  Result = In1.usub_ov(In2, Overflow);
59 
60  return Overflow;
61 }
62 
63 /// Given an icmp instruction, return true if any use of this comparison is a
64 /// branch on sign bit comparison.
65 static bool hasBranchUse(ICmpInst &I) {
66  for (auto *U : I.users())
67  if (isa<BranchInst>(U))
68  return true;
69  return false;
70 }
71 
72 /// Returns true if the exploded icmp can be expressed as a signed comparison
73 /// to zero and updates the predicate accordingly.
74 /// The signedness of the comparison is preserved.
75 /// TODO: Refactor with decomposeBitTestICmp()?
76 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
77  if (!ICmpInst::isSigned(Pred))
78  return false;
79 
80  if (C.isNullValue())
81  return ICmpInst::isRelational(Pred);
82 
83  if (C.isOneValue()) {
84  if (Pred == ICmpInst::ICMP_SLT) {
85  Pred = ICmpInst::ICMP_SLE;
86  return true;
87  }
88  } else if (C.isAllOnesValue()) {
89  if (Pred == ICmpInst::ICMP_SGT) {
90  Pred = ICmpInst::ICMP_SGE;
91  return true;
92  }
93  }
94 
95  return false;
96 }
97 
98 /// Given a signed integer type and a set of known zero and one bits, compute
99 /// the maximum and minimum values that could have the specified known zero and
100 /// known one bits, returning them in Min/Max.
101 /// TODO: Move to method on KnownBits struct?
103  APInt &Min, APInt &Max) {
104  assert(Known.getBitWidth() == Min.getBitWidth() &&
105  Known.getBitWidth() == Max.getBitWidth() &&
106  "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
107  APInt UnknownBits = ~(Known.Zero|Known.One);
108 
109  // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
110  // bit if it is unknown.
111  Min = Known.One;
112  Max = Known.One|UnknownBits;
113 
114  if (UnknownBits.isNegative()) { // Sign bit is unknown
115  Min.setSignBit();
116  Max.clearSignBit();
117  }
118 }
119 
120 /// Given an unsigned integer type and a set of known zero and one bits, compute
121 /// the maximum and minimum values that could have the specified known zero and
122 /// known one bits, returning them in Min/Max.
123 /// TODO: Move to method on KnownBits struct?
125  APInt &Min, APInt &Max) {
126  assert(Known.getBitWidth() == Min.getBitWidth() &&
127  Known.getBitWidth() == Max.getBitWidth() &&
128  "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
129  APInt UnknownBits = ~(Known.Zero|Known.One);
130 
131  // The minimum value is when the unknown bits are all zeros.
132  Min = Known.One;
133  // The maximum value is when the unknown bits are all ones.
134  Max = Known.One|UnknownBits;
135 }
136 
137 /// This is called when we see this pattern:
138 /// cmp pred (load (gep GV, ...)), cmpcst
139 /// where GV is a global variable with a constant initializer. Try to simplify
140 /// this into some simple computation that does not need the load. For example
141 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
142 ///
143 /// If AndCst is non-null, then the loaded value is masked with that constant
144 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
145 Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
146  GlobalVariable *GV,
147  CmpInst &ICI,
148  ConstantInt *AndCst) {
149  Constant *Init = GV->getInitializer();
150  if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
151  return nullptr;
152 
153  uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
154  // Don't blow up on huge arrays.
155  if (ArrayElementCount > MaxArraySizeForCombine)
156  return nullptr;
157 
158  // There are many forms of this optimization we can handle, for now, just do
159  // the simple index into a single-dimensional array.
160  //
161  // Require: GEP GV, 0, i {{, constant indices}}
162  if (GEP->getNumOperands() < 3 ||
163  !isa<ConstantInt>(GEP->getOperand(1)) ||
164  !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
165  isa<Constant>(GEP->getOperand(2)))
166  return nullptr;
167 
168  // Check that indices after the variable are constants and in-range for the
169  // type they index. Collect the indices. This is typically for arrays of
170  // structs.
171  SmallVector<unsigned, 4> LaterIndices;
172 
173  Type *EltTy = Init->getType()->getArrayElementType();
174  for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
175  ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
176  if (!Idx) return nullptr; // Variable index.
177 
178  uint64_t IdxVal = Idx->getZExtValue();
179  if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
180 
181  if (StructType *STy = dyn_cast<StructType>(EltTy))
182  EltTy = STy->getElementType(IdxVal);
183  else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
184  if (IdxVal >= ATy->getNumElements()) return nullptr;
185  EltTy = ATy->getElementType();
186  } else {
187  return nullptr; // Unknown type.
188  }
189 
190  LaterIndices.push_back(IdxVal);
191  }
192 
193  enum { Overdefined = -3, Undefined = -2 };
194 
195  // Variables for our state machines.
196 
197  // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
198  // "i == 47 | i == 87", where 47 is the first index the condition is true for,
199  // and 87 is the second (and last) index. FirstTrueElement is -2 when
200  // undefined, otherwise set to the first true element. SecondTrueElement is
201  // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
202  int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
203 
204  // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
205  // form "i != 47 & i != 87". Same state transitions as for true elements.
206  int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
207 
208  /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
209  /// define a state machine that triggers for ranges of values that the index
210  /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
211  /// This is -2 when undefined, -3 when overdefined, and otherwise the last
212  /// index in the range (inclusive). We use -2 for undefined here because we
213  /// use relative comparisons and don't want 0-1 to match -1.
214  int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
215 
216  // MagicBitvector - This is a magic bitvector where we set a bit if the
217  // comparison is true for element 'i'. If there are 64 elements or less in
218  // the array, this will fully represent all the comparison results.
219  uint64_t MagicBitvector = 0;
220 
221  // Scan the array and see if one of our patterns matches.
222  Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
223  for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
224  Constant *Elt = Init->getAggregateElement(i);
225  if (!Elt) return nullptr;
226 
227  // If this is indexing an array of structures, get the structure element.
228  if (!LaterIndices.empty())
229  Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
230 
231  // If the element is masked, handle it.
232  if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
233 
234  // Find out if the comparison would be true or false for the i'th element.
236  CompareRHS, DL, &TLI);
237  // If the result is undef for this element, ignore it.
238  if (isa<UndefValue>(C)) {
239  // Extend range state machines to cover this element in case there is an
240  // undef in the middle of the range.
241  if (TrueRangeEnd == (int)i-1)
242  TrueRangeEnd = i;
243  if (FalseRangeEnd == (int)i-1)
244  FalseRangeEnd = i;
245  continue;
246  }
247 
248  // If we can't compute the result for any of the elements, we have to give
249  // up evaluating the entire conditional.
250  if (!isa<ConstantInt>(C)) return nullptr;
251 
252  // Otherwise, we know if the comparison is true or false for this element,
253  // update our state machines.
254  bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
255 
256  // State machine for single/double/range index comparison.
257  if (IsTrueForElt) {
258  // Update the TrueElement state machine.
259  if (FirstTrueElement == Undefined)
260  FirstTrueElement = TrueRangeEnd = i; // First true element.
261  else {
262  // Update double-compare state machine.
263  if (SecondTrueElement == Undefined)
264  SecondTrueElement = i;
265  else
266  SecondTrueElement = Overdefined;
267 
268  // Update range state machine.
269  if (TrueRangeEnd == (int)i-1)
270  TrueRangeEnd = i;
271  else
272  TrueRangeEnd = Overdefined;
273  }
274  } else {
275  // Update the FalseElement state machine.
276  if (FirstFalseElement == Undefined)
277  FirstFalseElement = FalseRangeEnd = i; // First false element.
278  else {
279  // Update double-compare state machine.
280  if (SecondFalseElement == Undefined)
281  SecondFalseElement = i;
282  else
283  SecondFalseElement = Overdefined;
284 
285  // Update range state machine.
286  if (FalseRangeEnd == (int)i-1)
287  FalseRangeEnd = i;
288  else
289  FalseRangeEnd = Overdefined;
290  }
291  }
292 
293  // If this element is in range, update our magic bitvector.
294  if (i < 64 && IsTrueForElt)
295  MagicBitvector |= 1ULL << i;
296 
297  // If all of our states become overdefined, bail out early. Since the
298  // predicate is expensive, only check it every 8 elements. This is only
299  // really useful for really huge arrays.
300  if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
301  SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
302  FalseRangeEnd == Overdefined)
303  return nullptr;
304  }
305 
306  // Now that we've scanned the entire array, emit our new comparison(s). We
307  // order the state machines in complexity of the generated code.
308  Value *Idx = GEP->getOperand(2);
309 
310  // If the index is larger than the pointer size of the target, truncate the
311  // index down like the GEP would do implicitly. We don't have to do this for
312  // an inbounds GEP because the index can't be out of range.
313  if (!GEP->isInBounds()) {
314  Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
315  unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
316  if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
317  Idx = Builder.CreateTrunc(Idx, IntPtrTy);
318  }
319 
320  // If the comparison is only true for one or two elements, emit direct
321  // comparisons.
322  if (SecondTrueElement != Overdefined) {
323  // None true -> false.
324  if (FirstTrueElement == Undefined)
325  return replaceInstUsesWith(ICI, Builder.getFalse());
326 
327  Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
328 
329  // True for one element -> 'i == 47'.
330  if (SecondTrueElement == Undefined)
331  return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
332 
333  // True for two elements -> 'i == 47 | i == 72'.
334  Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
335  Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
336  Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
337  return BinaryOperator::CreateOr(C1, C2);
338  }
339 
340  // If the comparison is only false for one or two elements, emit direct
341  // comparisons.
342  if (SecondFalseElement != Overdefined) {
343  // None false -> true.
344  if (FirstFalseElement == Undefined)
345  return replaceInstUsesWith(ICI, Builder.getTrue());
346 
347  Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
348 
349  // False for one element -> 'i != 47'.
350  if (SecondFalseElement == Undefined)
351  return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
352 
353  // False for two elements -> 'i != 47 & i != 72'.
354  Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
355  Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
356  Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
357  return BinaryOperator::CreateAnd(C1, C2);
358  }
359 
360  // If the comparison can be replaced with a range comparison for the elements
361  // where it is true, emit the range check.
362  if (TrueRangeEnd != Overdefined) {
363  assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
364 
365  // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
366  if (FirstTrueElement) {
367  Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
368  Idx = Builder.CreateAdd(Idx, Offs);
369  }
370 
371  Value *End = ConstantInt::get(Idx->getType(),
372  TrueRangeEnd-FirstTrueElement+1);
373  return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
374  }
375 
376  // False range check.
377  if (FalseRangeEnd != Overdefined) {
378  assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
379  // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
380  if (FirstFalseElement) {
381  Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
382  Idx = Builder.CreateAdd(Idx, Offs);
383  }
384 
385  Value *End = ConstantInt::get(Idx->getType(),
386  FalseRangeEnd-FirstFalseElement);
387  return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
388  }
389 
390  // If a magic bitvector captures the entire comparison state
391  // of this load, replace it with computation that does:
392  // ((magic_cst >> i) & 1) != 0
393  {
394  Type *Ty = nullptr;
395 
396  // Look for an appropriate type:
397  // - The type of Idx if the magic fits
398  // - The smallest fitting legal type
399  if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
400  Ty = Idx->getType();
401  else
402  Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
403 
404  if (Ty) {
405  Value *V = Builder.CreateIntCast(Idx, Ty, false);
406  V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
407  V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
408  return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
409  }
410  }
411 
412  return nullptr;
413 }
414 
415 /// Return a value that can be used to compare the *offset* implied by a GEP to
416 /// zero. For example, if we have &A[i], we want to return 'i' for
417 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
418 /// are involved. The above expression would also be legal to codegen as
419 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
420 /// This latter form is less amenable to optimization though, and we are allowed
421 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
422 ///
423 /// If we can't emit an optimized form for this expression, this returns null.
424 ///
426  const DataLayout &DL) {
428 
429  // Check to see if this gep only has a single variable index. If so, and if
430  // any constant indices are a multiple of its scale, then we can compute this
431  // in terms of the scale of the variable index. For example, if the GEP
432  // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
433  // because the expression will cross zero at the same point.
434  unsigned i, e = GEP->getNumOperands();
435  int64_t Offset = 0;
436  for (i = 1; i != e; ++i, ++GTI) {
437  if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
438  // Compute the aggregate offset of constant indices.
439  if (CI->isZero()) continue;
440 
441  // Handle a struct index, which adds its field offset to the pointer.
442  if (StructType *STy = GTI.getStructTypeOrNull()) {
443  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
444  } else {
445  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
446  Offset += Size*CI->getSExtValue();
447  }
448  } else {
449  // Found our variable index.
450  break;
451  }
452  }
453 
454  // If there are no variable indices, we must have a constant offset, just
455  // evaluate it the general way.
456  if (i == e) return nullptr;
457 
458  Value *VariableIdx = GEP->getOperand(i);
459  // Determine the scale factor of the variable element. For example, this is
460  // 4 if the variable index is into an array of i32.
461  uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
462 
463  // Verify that there are no other variable indices. If so, emit the hard way.
464  for (++i, ++GTI; i != e; ++i, ++GTI) {
465  ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
466  if (!CI) return nullptr;
467 
468  // Compute the aggregate offset of constant indices.
469  if (CI->isZero()) continue;
470 
471  // Handle a struct index, which adds its field offset to the pointer.
472  if (StructType *STy = GTI.getStructTypeOrNull()) {
473  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
474  } else {
475  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
476  Offset += Size*CI->getSExtValue();
477  }
478  }
479 
480  // Okay, we know we have a single variable index, which must be a
481  // pointer/array/vector index. If there is no offset, life is simple, return
482  // the index.
483  Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
484  unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
485  if (Offset == 0) {
486  // Cast to intptrty in case a truncation occurs. If an extension is needed,
487  // we don't need to bother extending: the extension won't affect where the
488  // computation crosses zero.
489  if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
490  VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
491  }
492  return VariableIdx;
493  }
494 
495  // Otherwise, there is an index. The computation we will do will be modulo
496  // the pointer size.
497  Offset = SignExtend64(Offset, IntPtrWidth);
498  VariableScale = SignExtend64(VariableScale, IntPtrWidth);
499 
500  // To do this transformation, any constant index must be a multiple of the
501  // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
502  // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
503  // multiple of the variable scale.
504  int64_t NewOffs = Offset / (int64_t)VariableScale;
505  if (Offset != NewOffs*(int64_t)VariableScale)
506  return nullptr;
507 
508  // Okay, we can do this evaluation. Start by converting the index to intptr.
509  if (VariableIdx->getType() != IntPtrTy)
510  VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
511  true /*Signed*/);
512  Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
513  return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
514 }
515 
516 /// Returns true if we can rewrite Start as a GEP with pointer Base
517 /// and some integer offset. The nodes that need to be re-written
518 /// for this transformation will be added to Explored.
519 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
520  const DataLayout &DL,
521  SetVector<Value *> &Explored) {
522  SmallVector<Value *, 16> WorkList(1, Start);
523  Explored.insert(Base);
524 
525  // The following traversal gives us an order which can be used
526  // when doing the final transformation. Since in the final
527  // transformation we create the PHI replacement instructions first,
528  // we don't have to get them in any particular order.
529  //
530  // However, for other instructions we will have to traverse the
531  // operands of an instruction first, which means that we have to
532  // do a post-order traversal.
533  while (!WorkList.empty()) {
535 
536  while (!WorkList.empty()) {
537  if (Explored.size() >= 100)
538  return false;
539 
540  Value *V = WorkList.back();
541 
542  if (Explored.count(V) != 0) {
543  WorkList.pop_back();
544  continue;
545  }
546 
547  if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
548  !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
549  // We've found some value that we can't explore which is different from
550  // the base. Therefore we can't do this transformation.
551  return false;
552 
553  if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
554  auto *CI = dyn_cast<CastInst>(V);
555  if (!CI->isNoopCast(DL))
556  return false;
557 
558  if (Explored.count(CI->getOperand(0)) == 0)
559  WorkList.push_back(CI->getOperand(0));
560  }
561 
562  if (auto *GEP = dyn_cast<GEPOperator>(V)) {
563  // We're limiting the GEP to having one index. This will preserve
564  // the original pointer type. We could handle more cases in the
565  // future.
566  if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
567  GEP->getType() != Start->getType())
568  return false;
569 
570  if (Explored.count(GEP->getOperand(0)) == 0)
571  WorkList.push_back(GEP->getOperand(0));
572  }
573 
574  if (WorkList.back() == V) {
575  WorkList.pop_back();
576  // We've finished visiting this node, mark it as such.
577  Explored.insert(V);
578  }
579 
580  if (auto *PN = dyn_cast<PHINode>(V)) {
581  // We cannot transform PHIs on unsplittable basic blocks.
582  if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
583  return false;
584  Explored.insert(PN);
585  PHIs.insert(PN);
586  }
587  }
588 
589  // Explore the PHI nodes further.
590  for (auto *PN : PHIs)
591  for (Value *Op : PN->incoming_values())
592  if (Explored.count(Op) == 0)
593  WorkList.push_back(Op);
594  }
595 
596  // Make sure that we can do this. Since we can't insert GEPs in a basic
597  // block before a PHI node, we can't easily do this transformation if
598  // we have PHI node users of transformed instructions.
599  for (Value *Val : Explored) {
600  for (Value *Use : Val->uses()) {
601 
602  auto *PHI = dyn_cast<PHINode>(Use);
603  auto *Inst = dyn_cast<Instruction>(Val);
604 
605  if (Inst == Base || Inst == PHI || !Inst || !PHI ||
606  Explored.count(PHI) == 0)
607  continue;
608 
609  if (PHI->getParent() == Inst->getParent())
610  return false;
611  }
612  }
613  return true;
614 }
615 
616 // Sets the appropriate insert point on Builder where we can add
617 // a replacement Instruction for V (if that is possible).
618 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
619  bool Before = true) {
620  if (auto *PHI = dyn_cast<PHINode>(V)) {
621  Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
622  return;
623  }
624  if (auto *I = dyn_cast<Instruction>(V)) {
625  if (!Before)
626  I = &*std::next(I->getIterator());
627  Builder.SetInsertPoint(I);
628  return;
629  }
630  if (auto *A = dyn_cast<Argument>(V)) {
631  // Set the insertion point in the entry block.
632  BasicBlock &Entry = A->getParent()->getEntryBlock();
633  Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
634  return;
635  }
636  // Otherwise, this is a constant and we don't need to set a new
637  // insertion point.
638  assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
639 }
640 
641 /// Returns a re-written value of Start as an indexed GEP using Base as a
642 /// pointer.
644  const DataLayout &DL,
645  SetVector<Value *> &Explored) {
646  // Perform all the substitutions. This is a bit tricky because we can
647  // have cycles in our use-def chains.
648  // 1. Create the PHI nodes without any incoming values.
649  // 2. Create all the other values.
650  // 3. Add the edges for the PHI nodes.
651  // 4. Emit GEPs to get the original pointers.
652  // 5. Remove the original instructions.
653  Type *IndexType = IntegerType::get(
654  Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
655 
657  NewInsts[Base] = ConstantInt::getNullValue(IndexType);
658 
659  // Create the new PHI nodes, without adding any incoming values.
660  for (Value *Val : Explored) {
661  if (Val == Base)
662  continue;
663  // Create empty phi nodes. This avoids cyclic dependencies when creating
664  // the remaining instructions.
665  if (auto *PHI = dyn_cast<PHINode>(Val))
666  NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
667  PHI->getName() + ".idx", PHI);
668  }
669  IRBuilder<> Builder(Base->getContext());
670 
671  // Create all the other instructions.
672  for (Value *Val : Explored) {
673 
674  if (NewInsts.find(Val) != NewInsts.end())
675  continue;
676 
677  if (auto *CI = dyn_cast<CastInst>(Val)) {
678  // Don't get rid of the intermediate variable here; the store can grow
679  // the map which will invalidate the reference to the input value.
680  Value *V = NewInsts[CI->getOperand(0)];
681  NewInsts[CI] = V;
682  continue;
683  }
684  if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
685  Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
686  : GEP->getOperand(1);
687  setInsertionPoint(Builder, GEP);
688  // Indices might need to be sign extended. GEPs will magically do
689  // this, but we need to do it ourselves here.
690  if (Index->getType()->getScalarSizeInBits() !=
691  NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
692  Index = Builder.CreateSExtOrTrunc(
693  Index, NewInsts[GEP->getOperand(0)]->getType(),
694  GEP->getOperand(0)->getName() + ".sext");
695  }
696 
697  auto *Op = NewInsts[GEP->getOperand(0)];
698  if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
699  NewInsts[GEP] = Index;
700  else
701  NewInsts[GEP] = Builder.CreateNSWAdd(
702  Op, Index, GEP->getOperand(0)->getName() + ".add");
703  continue;
704  }
705  if (isa<PHINode>(Val))
706  continue;
707 
708  llvm_unreachable("Unexpected instruction type");
709  }
710 
711  // Add the incoming values to the PHI nodes.
712  for (Value *Val : Explored) {
713  if (Val == Base)
714  continue;
715  // All the instructions have been created, we can now add edges to the
716  // phi nodes.
717  if (auto *PHI = dyn_cast<PHINode>(Val)) {
718  PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
719  for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
720  Value *NewIncoming = PHI->getIncomingValue(I);
721 
722  if (NewInsts.find(NewIncoming) != NewInsts.end())
723  NewIncoming = NewInsts[NewIncoming];
724 
725  NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
726  }
727  }
728  }
729 
730  for (Value *Val : Explored) {
731  if (Val == Base)
732  continue;
733 
734  // Depending on the type, for external users we have to emit
735  // a GEP or a GEP + ptrtoint.
736  setInsertionPoint(Builder, Val, false);
737 
738  // If required, create an inttoptr instruction for Base.
739  Value *NewBase = Base;
740  if (!Base->getType()->isPointerTy())
741  NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
742  Start->getName() + "to.ptr");
743 
744  Value *GEP = Builder.CreateInBoundsGEP(
745  Start->getType()->getPointerElementType(), NewBase,
746  makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
747 
748  if (!Val->getType()->isPointerTy()) {
749  Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
750  Val->getName() + ".conv");
751  GEP = Cast;
752  }
753  Val->replaceAllUsesWith(GEP);
754  }
755 
756  return NewInsts[Start];
757 }
758 
759 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
760 /// the input Value as a constant indexed GEP. Returns a pair containing
761 /// the GEPs Pointer and Index.
762 static std::pair<Value *, Value *>
764  Type *IndexType = IntegerType::get(V->getContext(),
766 
768  while (true) {
769  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
770  // We accept only inbouds GEPs here to exclude the possibility of
771  // overflow.
772  if (!GEP->isInBounds())
773  break;
774  if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
775  GEP->getType() == V->getType()) {
776  V = GEP->getOperand(0);
777  Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
778  Index = ConstantExpr::getAdd(
779  Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
780  continue;
781  }
782  break;
783  }
784  if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
785  if (!CI->isNoopCast(DL))
786  break;
787  V = CI->getOperand(0);
788  continue;
789  }
790  if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
791  if (!CI->isNoopCast(DL))
792  break;
793  V = CI->getOperand(0);
794  continue;
795  }
796  break;
797  }
798  return {V, Index};
799 }
800 
801 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
802 /// We can look through PHIs, GEPs and casts in order to determine a common base
803 /// between GEPLHS and RHS.
805  ICmpInst::Predicate Cond,
806  const DataLayout &DL) {
807  // FIXME: Support vector of pointers.
808  if (GEPLHS->getType()->isVectorTy())
809  return nullptr;
810 
811  if (!GEPLHS->hasAllConstantIndices())
812  return nullptr;
813 
814  // Make sure the pointers have the same type.
815  if (GEPLHS->getType() != RHS->getType())
816  return nullptr;
817 
818  Value *PtrBase, *Index;
819  std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
820 
821  // The set of nodes that will take part in this transformation.
822  SetVector<Value *> Nodes;
823 
824  if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
825  return nullptr;
826 
827  // We know we can re-write this as
828  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
829  // Since we've only looked through inbouds GEPs we know that we
830  // can't have overflow on either side. We can therefore re-write
831  // this as:
832  // OFFSET1 cmp OFFSET2
833  Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
834 
835  // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
836  // GEP having PtrBase as the pointer base, and has returned in NewRHS the
837  // offset. Since Index is the offset of LHS to the base pointer, we will now
838  // compare the offsets instead of comparing the pointers.
839  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
840 }
841 
842 /// Fold comparisons between a GEP instruction and something else. At this point
843 /// we know that the GEP is on the LHS of the comparison.
844 Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
845  ICmpInst::Predicate Cond,
846  Instruction &I) {
847  // Don't transform signed compares of GEPs into index compares. Even if the
848  // GEP is inbounds, the final add of the base pointer can have signed overflow
849  // and would change the result of the icmp.
850  // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
851  // the maximum signed value for the pointer type.
852  if (ICmpInst::isSigned(Cond))
853  return nullptr;
854 
855  // Look through bitcasts and addrspacecasts. We do not however want to remove
856  // 0 GEPs.
857  if (!isa<GetElementPtrInst>(RHS))
858  RHS = RHS->stripPointerCasts();
859 
860  Value *PtrBase = GEPLHS->getOperand(0);
861  // FIXME: Support vector pointer GEPs.
862  if (PtrBase == RHS && GEPLHS->isInBounds() &&
863  !GEPLHS->getType()->isVectorTy()) {
864  // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
865  // This transformation (ignoring the base and scales) is valid because we
866  // know pointers can't overflow since the gep is inbounds. See if we can
867  // output an optimized form.
868  Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
869 
870  // If not, synthesize the offset the hard way.
871  if (!Offset)
872  Offset = EmitGEPOffset(GEPLHS);
873  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
874  Constant::getNullValue(Offset->getType()));
875  }
876 
877  if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
878  isa<Constant>(RHS) && cast<Constant>(RHS)->isNullValue() &&
880  RHS->getType()->getPointerAddressSpace())) {
881  // For most address spaces, an allocation can't be placed at null, but null
882  // itself is treated as a 0 size allocation in the in bounds rules. Thus,
883  // the only valid inbounds address derived from null, is null itself.
884  // Thus, we have four cases to consider:
885  // 1) Base == nullptr, Offset == 0 -> inbounds, null
886  // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
887  // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
888  // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
889  //
890  // (Note if we're indexing a type of size 0, that simply collapses into one
891  // of the buckets above.)
892  //
893  // In general, we're allowed to make values less poison (i.e. remove
894  // sources of full UB), so in this case, we just select between the two
895  // non-poison cases (1 and 4 above).
896  //
897  // For vectors, we apply the same reasoning on a per-lane basis.
898  auto *Base = GEPLHS->getPointerOperand();
899  if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
900  int NumElts = GEPLHS->getType()->getVectorNumElements();
901  Base = Builder.CreateVectorSplat(NumElts, Base);
902  }
903  return new ICmpInst(Cond, Base,
905  cast<Constant>(RHS), Base->getType()));
906  } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
907  // If the base pointers are different, but the indices are the same, just
908  // compare the base pointer.
909  if (PtrBase != GEPRHS->getOperand(0)) {
910  bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
911  IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
912  GEPRHS->getOperand(0)->getType();
913  if (IndicesTheSame)
914  for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
915  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
916  IndicesTheSame = false;
917  break;
918  }
919 
920  // If all indices are the same, just compare the base pointers.
921  Type *BaseType = GEPLHS->getOperand(0)->getType();
922  if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
923  return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
924 
925  // If we're comparing GEPs with two base pointers that only differ in type
926  // and both GEPs have only constant indices or just one use, then fold
927  // the compare with the adjusted indices.
928  // FIXME: Support vector of pointers.
929  if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
930  (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
931  (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
932  PtrBase->stripPointerCasts() ==
933  GEPRHS->getOperand(0)->stripPointerCasts() &&
934  !GEPLHS->getType()->isVectorTy()) {
935  Value *LOffset = EmitGEPOffset(GEPLHS);
936  Value *ROffset = EmitGEPOffset(GEPRHS);
937 
938  // If we looked through an addrspacecast between different sized address
939  // spaces, the LHS and RHS pointers are different sized
940  // integers. Truncate to the smaller one.
941  Type *LHSIndexTy = LOffset->getType();
942  Type *RHSIndexTy = ROffset->getType();
943  if (LHSIndexTy != RHSIndexTy) {
944  if (LHSIndexTy->getPrimitiveSizeInBits() <
945  RHSIndexTy->getPrimitiveSizeInBits()) {
946  ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
947  } else
948  LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
949  }
950 
951  Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
952  LOffset, ROffset);
953  return replaceInstUsesWith(I, Cmp);
954  }
955 
956  // Otherwise, the base pointers are different and the indices are
957  // different. Try convert this to an indexed compare by looking through
958  // PHIs/casts.
959  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
960  }
961 
962  // If one of the GEPs has all zero indices, recurse.
963  // FIXME: Handle vector of pointers.
964  if (!GEPLHS->getType()->isVectorTy() && GEPLHS->hasAllZeroIndices())
965  return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
967 
968  // If the other GEP has all zero indices, recurse.
969  // FIXME: Handle vector of pointers.
970  if (!GEPRHS->getType()->isVectorTy() && GEPRHS->hasAllZeroIndices())
971  return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
972 
973  bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
974  if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
975  // If the GEPs only differ by one index, compare it.
976  unsigned NumDifferences = 0; // Keep track of # differences.
977  unsigned DiffOperand = 0; // The operand that differs.
978  for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
979  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
980  Type *LHSType = GEPLHS->getOperand(i)->getType();
981  Type *RHSType = GEPRHS->getOperand(i)->getType();
982  // FIXME: Better support for vector of pointers.
983  if (LHSType->getPrimitiveSizeInBits() !=
984  RHSType->getPrimitiveSizeInBits() ||
985  (GEPLHS->getType()->isVectorTy() &&
986  (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
987  // Irreconcilable differences.
988  NumDifferences = 2;
989  break;
990  }
991 
992  if (NumDifferences++) break;
993  DiffOperand = i;
994  }
995 
996  if (NumDifferences == 0) // SAME GEP?
997  return replaceInstUsesWith(I, // No comparison is needed here.
999 
1000  else if (NumDifferences == 1 && GEPsInBounds) {
1001  Value *LHSV = GEPLHS->getOperand(DiffOperand);
1002  Value *RHSV = GEPRHS->getOperand(DiffOperand);
1003  // Make sure we do a signed comparison here.
1004  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
1005  }
1006  }
1007 
1008  // Only lower this if the icmp is the only user of the GEP or if we expect
1009  // the result to fold to a constant!
1010  if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
1011  (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
1012  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
1013  Value *L = EmitGEPOffset(GEPLHS);
1014  Value *R = EmitGEPOffset(GEPRHS);
1015  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
1016  }
1017  }
1018 
1019  // Try convert this to an indexed compare by looking through PHIs/casts as a
1020  // last resort.
1021  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1022 }
1023 
1024 Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI,
1025  const AllocaInst *Alloca,
1026  const Value *Other) {
1027  assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1028 
1029  // It would be tempting to fold away comparisons between allocas and any
1030  // pointer not based on that alloca (e.g. an argument). However, even
1031  // though such pointers cannot alias, they can still compare equal.
1032  //
1033  // But LLVM doesn't specify where allocas get their memory, so if the alloca
1034  // doesn't escape we can argue that it's impossible to guess its value, and we
1035  // can therefore act as if any such guesses are wrong.
1036  //
1037  // The code below checks that the alloca doesn't escape, and that it's only
1038  // used in a comparison once (the current instruction). The
1039  // single-comparison-use condition ensures that we're trivially folding all
1040  // comparisons against the alloca consistently, and avoids the risk of
1041  // erroneously folding a comparison of the pointer with itself.
1042 
1043  unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1044 
1046  for (const Use &U : Alloca->uses()) {
1047  if (Worklist.size() >= MaxIter)
1048  return nullptr;
1049  Worklist.push_back(&U);
1050  }
1051 
1052  unsigned NumCmps = 0;
1053  while (!Worklist.empty()) {
1054  assert(Worklist.size() <= MaxIter);
1055  const Use *U = Worklist.pop_back_val();
1056  const Value *V = U->getUser();
1057  --MaxIter;
1058 
1059  if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1060  isa<SelectInst>(V)) {
1061  // Track the uses.
1062  } else if (isa<LoadInst>(V)) {
1063  // Loading from the pointer doesn't escape it.
1064  continue;
1065  } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1066  // Storing *to* the pointer is fine, but storing the pointer escapes it.
1067  if (SI->getValueOperand() == U->get())
1068  return nullptr;
1069  continue;
1070  } else if (isa<ICmpInst>(V)) {
1071  if (NumCmps++)
1072  return nullptr; // Found more than one cmp.
1073  continue;
1074  } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1075  switch (Intrin->getIntrinsicID()) {
1076  // These intrinsics don't escape or compare the pointer. Memset is safe
1077  // because we don't allow ptrtoint. Memcpy and memmove are safe because
1078  // we don't allow stores, so src cannot point to V.
1079  case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1080  case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1081  continue;
1082  default:
1083  return nullptr;
1084  }
1085  } else {
1086  return nullptr;
1087  }
1088  for (const Use &U : V->uses()) {
1089  if (Worklist.size() >= MaxIter)
1090  return nullptr;
1091  Worklist.push_back(&U);
1092  }
1093  }
1094 
1095  Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1096  return replaceInstUsesWith(
1097  ICI,
1099 }
1100 
1101 /// Fold "icmp pred (X+C), X".
1102 Instruction *InstCombiner::foldICmpAddOpConst(Value *X, const APInt &C,
1103  ICmpInst::Predicate Pred) {
1104  // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1105  // so the values can never be equal. Similarly for all other "or equals"
1106  // operators.
1107  assert(!!C && "C should not be zero!");
1108 
1109  // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1110  // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1111  // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1112  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1113  Constant *R = ConstantInt::get(X->getType(),
1114  APInt::getMaxValue(C.getBitWidth()) - C);
1115  return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1116  }
1117 
1118  // (X+1) >u X --> X <u (0-1) --> X != 255
1119  // (X+2) >u X --> X <u (0-2) --> X <u 254
1120  // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1121  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1122  return new ICmpInst(ICmpInst::ICMP_ULT, X,
1123  ConstantInt::get(X->getType(), -C));
1124 
1126 
1127  // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1128  // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1129  // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1130  // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1131  // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1132  // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1133  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1134  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1135  ConstantInt::get(X->getType(), SMax - C));
1136 
1137  // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1138  // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1139  // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1140  // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1141  // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1142  // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1143 
1144  assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1145  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1146  ConstantInt::get(X->getType(), SMax - (C - 1)));
1147 }
1148 
1149 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1150 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1151 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1152 Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A,
1153  const APInt &AP1,
1154  const APInt &AP2) {
1155  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1156 
1157  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1158  if (I.getPredicate() == I.ICMP_NE)
1159  Pred = CmpInst::getInversePredicate(Pred);
1160  return new ICmpInst(Pred, LHS, RHS);
1161  };
1162 
1163  // Don't bother doing any work for cases which InstSimplify handles.
1164  if (AP2.isNullValue())
1165  return nullptr;
1166 
1167  bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1168  if (IsAShr) {
1169  if (AP2.isAllOnesValue())
1170  return nullptr;
1171  if (AP2.isNegative() != AP1.isNegative())
1172  return nullptr;
1173  if (AP2.sgt(AP1))
1174  return nullptr;
1175  }
1176 
1177  if (!AP1)
1178  // 'A' must be large enough to shift out the highest set bit.
1179  return getICmp(I.ICMP_UGT, A,
1180  ConstantInt::get(A->getType(), AP2.logBase2()));
1181 
1182  if (AP1 == AP2)
1183  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1184 
1185  int Shift;
1186  if (IsAShr && AP1.isNegative())
1187  Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1188  else
1189  Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1190 
1191  if (Shift > 0) {
1192  if (IsAShr && AP1 == AP2.ashr(Shift)) {
1193  // There are multiple solutions if we are comparing against -1 and the LHS
1194  // of the ashr is not a power of two.
1195  if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1196  return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1197  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1198  } else if (AP1 == AP2.lshr(Shift)) {
1199  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1200  }
1201  }
1202 
1203  // Shifting const2 will never be equal to const1.
1204  // FIXME: This should always be handled by InstSimplify?
1205  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1206  return replaceInstUsesWith(I, TorF);
1207 }
1208 
1209 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1210 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1211 Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A,
1212  const APInt &AP1,
1213  const APInt &AP2) {
1214  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1215 
1216  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1217  if (I.getPredicate() == I.ICMP_NE)
1218  Pred = CmpInst::getInversePredicate(Pred);
1219  return new ICmpInst(Pred, LHS, RHS);
1220  };
1221 
1222  // Don't bother doing any work for cases which InstSimplify handles.
1223  if (AP2.isNullValue())
1224  return nullptr;
1225 
1226  unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1227 
1228  if (!AP1 && AP2TrailingZeros != 0)
1229  return getICmp(
1230  I.ICMP_UGE, A,
1231  ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1232 
1233  if (AP1 == AP2)
1234  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1235 
1236  // Get the distance between the lowest bits that are set.
1237  int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1238 
1239  if (Shift > 0 && AP2.shl(Shift) == AP1)
1240  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1241 
1242  // Shifting const2 will never be equal to const1.
1243  // FIXME: This should always be handled by InstSimplify?
1244  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1245  return replaceInstUsesWith(I, TorF);
1246 }
1247 
1248 /// The caller has matched a pattern of the form:
1249 /// I = icmp ugt (add (add A, B), CI2), CI1
1250 /// If this is of the form:
1251 /// sum = a + b
1252 /// if (sum+128 >u 255)
1253 /// Then replace it with llvm.sadd.with.overflow.i8.
1254 ///
1256  ConstantInt *CI2, ConstantInt *CI1,
1257  InstCombiner &IC) {
1258  // The transformation we're trying to do here is to transform this into an
1259  // llvm.sadd.with.overflow. To do this, we have to replace the original add
1260  // with a narrower add, and discard the add-with-constant that is part of the
1261  // range check (if we can't eliminate it, this isn't profitable).
1262 
1263  // In order to eliminate the add-with-constant, the compare can be its only
1264  // use.
1265  Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1266  if (!AddWithCst->hasOneUse())
1267  return nullptr;
1268 
1269  // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1270  if (!CI2->getValue().isPowerOf2())
1271  return nullptr;
1272  unsigned NewWidth = CI2->getValue().countTrailingZeros();
1273  if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1274  return nullptr;
1275 
1276  // The width of the new add formed is 1 more than the bias.
1277  ++NewWidth;
1278 
1279  // Check to see that CI1 is an all-ones value with NewWidth bits.
1280  if (CI1->getBitWidth() == NewWidth ||
1281  CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1282  return nullptr;
1283 
1284  // This is only really a signed overflow check if the inputs have been
1285  // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1286  // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1287  unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1288  if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1289  IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1290  return nullptr;
1291 
1292  // In order to replace the original add with a narrower
1293  // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1294  // and truncates that discard the high bits of the add. Verify that this is
1295  // the case.
1296  Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1297  for (User *U : OrigAdd->users()) {
1298  if (U == AddWithCst)
1299  continue;
1300 
1301  // Only accept truncates for now. We would really like a nice recursive
1302  // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1303  // chain to see which bits of a value are actually demanded. If the
1304  // original add had another add which was then immediately truncated, we
1305  // could still do the transformation.
1306  TruncInst *TI = dyn_cast<TruncInst>(U);
1307  if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1308  return nullptr;
1309  }
1310 
1311  // If the pattern matches, truncate the inputs to the narrower type and
1312  // use the sadd_with_overflow intrinsic to efficiently compute both the
1313  // result and the overflow bit.
1314  Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1316  I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1317 
1318  InstCombiner::BuilderTy &Builder = IC.Builder;
1319 
1320  // Put the new code above the original add, in case there are any uses of the
1321  // add between the add and the compare.
1322  Builder.SetInsertPoint(OrigAdd);
1323 
1324  Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1325  Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1326  CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1327  Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1328  Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1329 
1330  // The inner add was the result of the narrow add, zero extended to the
1331  // wider type. Replace it with the result computed by the intrinsic.
1332  IC.replaceInstUsesWith(*OrigAdd, ZExt);
1333 
1334  // The original icmp gets replaced with the overflow value.
1335  return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1336 }
1337 
1338 /// If we have:
1339 /// icmp eq/ne (urem/srem %x, %y), 0
1340 /// iff %y is a power-of-two, we can replace this with a bit test:
1341 /// icmp eq/ne (and %x, (add %y, -1)), 0
1342 Instruction *InstCombiner::foldIRemByPowerOfTwoToBitTest(ICmpInst &I) {
1343  // This fold is only valid for equality predicates.
1344  if (!I.isEquality())
1345  return nullptr;
1346  ICmpInst::Predicate Pred;
1347  Value *X, *Y, *Zero;
1348  if (!match(&I, m_ICmp(Pred, m_OneUse(m_IRem(m_Value(X), m_Value(Y))),
1349  m_CombineAnd(m_Zero(), m_Value(Zero)))))
1350  return nullptr;
1351  if (!isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, 0, &I))
1352  return nullptr;
1353  // This may increase instruction count, we don't enforce that Y is a constant.
1354  Value *Mask = Builder.CreateAdd(Y, Constant::getAllOnesValue(Y->getType()));
1355  Value *Masked = Builder.CreateAnd(X, Mask);
1356  return ICmpInst::Create(Instruction::ICmp, Pred, Masked, Zero);
1357 }
1358 
1359 /// Fold equality-comparison between zero and any (maybe truncated) right-shift
1360 /// by one-less-than-bitwidth into a sign test on the original value.
1361 Instruction *InstCombiner::foldSignBitTest(ICmpInst &I) {
1362  Instruction *Val;
1363  ICmpInst::Predicate Pred;
1364  if (!I.isEquality() || !match(&I, m_ICmp(Pred, m_Instruction(Val), m_Zero())))
1365  return nullptr;
1366 
1367  Value *X;
1368  Type *XTy;
1369 
1370  Constant *C;
1371  if (match(Val, m_TruncOrSelf(m_Shr(m_Value(X), m_Constant(C))))) {
1372  XTy = X->getType();
1373  unsigned XBitWidth = XTy->getScalarSizeInBits();
1374  if (!match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1375  APInt(XBitWidth, XBitWidth - 1))))
1376  return nullptr;
1377  } else if (isa<BinaryOperator>(Val) &&
1378  (X = reassociateShiftAmtsOfTwoSameDirectionShifts(
1379  cast<BinaryOperator>(Val), SQ.getWithInstruction(Val),
1380  /*AnalyzeForSignBitExtraction=*/true))) {
1381  XTy = X->getType();
1382  } else
1383  return nullptr;
1384 
1385  return ICmpInst::Create(Instruction::ICmp,
1388  X, ConstantInt::getNullValue(XTy));
1389 }
1390 
1391 // Handle icmp pred X, 0
1392 Instruction *InstCombiner::foldICmpWithZero(ICmpInst &Cmp) {
1393  CmpInst::Predicate Pred = Cmp.getPredicate();
1394  if (!match(Cmp.getOperand(1), m_Zero()))
1395  return nullptr;
1396 
1397  // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1398  if (Pred == ICmpInst::ICMP_SGT) {
1399  Value *A, *B;
1401  if (SPR.Flavor == SPF_SMIN) {
1402  if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1403  return new ICmpInst(Pred, B, Cmp.getOperand(1));
1404  if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1405  return new ICmpInst(Pred, A, Cmp.getOperand(1));
1406  }
1407  }
1408 
1409  if (Instruction *New = foldIRemByPowerOfTwoToBitTest(Cmp))
1410  return New;
1411 
1412  // Given:
1413  // icmp eq/ne (urem %x, %y), 0
1414  // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1415  // icmp eq/ne %x, 0
1416  Value *X, *Y;
1417  if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1418  ICmpInst::isEquality(Pred)) {
1419  KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1420  KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1421  if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1422  return new ICmpInst(Pred, X, Cmp.getOperand(1));
1423  }
1424 
1425  return nullptr;
1426 }
1427 
1428 /// Fold icmp Pred X, C.
1429 /// TODO: This code structure does not make sense. The saturating add fold
1430 /// should be moved to some other helper and extended as noted below (it is also
1431 /// possible that code has been made unnecessary - do we canonicalize IR to
1432 /// overflow/saturating intrinsics or not?).
1433 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
1434  // Match the following pattern, which is a common idiom when writing
1435  // overflow-safe integer arithmetic functions. The source performs an addition
1436  // in wider type and explicitly checks for overflow using comparisons against
1437  // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1438  //
1439  // TODO: This could probably be generalized to handle other overflow-safe
1440  // operations if we worked out the formulas to compute the appropriate magic
1441  // constants.
1442  //
1443  // sum = a + b
1444  // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1445  CmpInst::Predicate Pred = Cmp.getPredicate();
1446  Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1447  Value *A, *B;
1448  ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1449  if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1450  match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1451  if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1452  return Res;
1453 
1454  return nullptr;
1455 }
1456 
1457 /// Canonicalize icmp instructions based on dominating conditions.
1458 Instruction *InstCombiner::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1459  // This is a cheap/incomplete check for dominance - just match a single
1460  // predecessor with a conditional branch.
1461  BasicBlock *CmpBB = Cmp.getParent();
1462  BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1463  if (!DomBB)
1464  return nullptr;
1465 
1466  Value *DomCond;
1467  BasicBlock *TrueBB, *FalseBB;
1468  if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1469  return nullptr;
1470 
1471  assert((TrueBB == CmpBB || FalseBB == CmpBB) &&
1472  "Predecessor block does not point to successor?");
1473 
1474  // The branch should get simplified. Don't bother simplifying this condition.
1475  if (TrueBB == FalseBB)
1476  return nullptr;
1477 
1478  // Try to simplify this compare to T/F based on the dominating condition.
1479  Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1480  if (Imp)
1481  return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1482 
1483  CmpInst::Predicate Pred = Cmp.getPredicate();
1484  Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1485  ICmpInst::Predicate DomPred;
1486  const APInt *C, *DomC;
1487  if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1488  match(Y, m_APInt(C))) {
1489  // We have 2 compares of a variable with constants. Calculate the constant
1490  // ranges of those compares to see if we can transform the 2nd compare:
1491  // DomBB:
1492  // DomCond = icmp DomPred X, DomC
1493  // br DomCond, CmpBB, FalseBB
1494  // CmpBB:
1495  // Cmp = icmp Pred X, C
1497  ConstantRange DominatingCR =
1498  (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1500  CmpInst::getInversePredicate(DomPred), *DomC);
1501  ConstantRange Intersection = DominatingCR.intersectWith(CR);
1502  ConstantRange Difference = DominatingCR.difference(CR);
1503  if (Intersection.isEmptySet())
1504  return replaceInstUsesWith(Cmp, Builder.getFalse());
1505  if (Difference.isEmptySet())
1506  return replaceInstUsesWith(Cmp, Builder.getTrue());
1507 
1508  // Canonicalizing a sign bit comparison that gets used in a branch,
1509  // pessimizes codegen by generating branch on zero instruction instead
1510  // of a test and branch. So we avoid canonicalizing in such situations
1511  // because test and branch instruction has better branch displacement
1512  // than compare and branch instruction.
1513  bool UnusedBit;
1514  bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1515  if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1516  return nullptr;
1517 
1518  if (const APInt *EqC = Intersection.getSingleElement())
1519  return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1520  if (const APInt *NeC = Difference.getSingleElement())
1521  return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1522  }
1523 
1524  return nullptr;
1525 }
1526 
1527 /// Fold icmp (trunc X, Y), C.
1528 Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
1529  TruncInst *Trunc,
1530  const APInt &C) {
1531  ICmpInst::Predicate Pred = Cmp.getPredicate();
1532  Value *X = Trunc->getOperand(0);
1533  if (C.isOneValue() && C.getBitWidth() > 1) {
1534  // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1535  Value *V = nullptr;
1536  if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1537  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1538  ConstantInt::get(V->getType(), 1));
1539  }
1540 
1541  if (Cmp.isEquality() && Trunc->hasOneUse()) {
1542  // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1543  // of the high bits truncated out of x are known.
1544  unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1545  SrcBits = X->getType()->getScalarSizeInBits();
1546  KnownBits Known = computeKnownBits(X, 0, &Cmp);
1547 
1548  // If all the high bits are known, we can do this xform.
1549  if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1550  // Pull in the high bits from known-ones set.
1551  APInt NewRHS = C.zext(SrcBits);
1552  NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1553  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1554  }
1555  }
1556 
1557  return nullptr;
1558 }
1559 
1560 /// Fold icmp (xor X, Y), C.
1561 Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp,
1562  BinaryOperator *Xor,
1563  const APInt &C) {
1564  Value *X = Xor->getOperand(0);
1565  Value *Y = Xor->getOperand(1);
1566  const APInt *XorC;
1567  if (!match(Y, m_APInt(XorC)))
1568  return nullptr;
1569 
1570  // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1571  // fold the xor.
1572  ICmpInst::Predicate Pred = Cmp.getPredicate();
1573  bool TrueIfSigned = false;
1574  if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1575 
1576  // If the sign bit of the XorCst is not set, there is no change to
1577  // the operation, just stop using the Xor.
1578  if (!XorC->isNegative()) {
1579  Cmp.setOperand(0, X);
1580  Worklist.Add(Xor);
1581  return &Cmp;
1582  }
1583 
1584  // Emit the opposite comparison.
1585  if (TrueIfSigned)
1586  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1588  else
1589  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1591  }
1592 
1593  if (Xor->hasOneUse()) {
1594  // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1595  if (!Cmp.isEquality() && XorC->isSignMask()) {
1596  Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1597  : Cmp.getSignedPredicate();
1598  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1599  }
1600 
1601  // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1602  if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1603  Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1604  : Cmp.getSignedPredicate();
1605  Pred = Cmp.getSwappedPredicate(Pred);
1606  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1607  }
1608  }
1609 
1610  // Mask constant magic can eliminate an 'xor' with unsigned compares.
1611  if (Pred == ICmpInst::ICMP_UGT) {
1612  // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1613  if (*XorC == ~C && (C + 1).isPowerOf2())
1614  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1615  // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1616  if (*XorC == C && (C + 1).isPowerOf2())
1617  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1618  }
1619  if (Pred == ICmpInst::ICMP_ULT) {
1620  // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1621  if (*XorC == -C && C.isPowerOf2())
1622  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1623  ConstantInt::get(X->getType(), ~C));
1624  // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1625  if (*XorC == C && (-C).isPowerOf2())
1626  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1627  ConstantInt::get(X->getType(), ~C));
1628  }
1629  return nullptr;
1630 }
1631 
1632 /// Fold icmp (and (sh X, Y), C2), C1.
1633 Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
1634  const APInt &C1, const APInt &C2) {
1635  BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1636  if (!Shift || !Shift->isShift())
1637  return nullptr;
1638 
1639  // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1640  // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1641  // code produced by the clang front-end, for bitfield access.
1642  // This seemingly simple opportunity to fold away a shift turns out to be
1643  // rather complicated. See PR17827 for details.
1644  unsigned ShiftOpcode = Shift->getOpcode();
1645  bool IsShl = ShiftOpcode == Instruction::Shl;
1646  const APInt *C3;
1647  if (match(Shift->getOperand(1), m_APInt(C3))) {
1648  bool CanFold = false;
1649  if (ShiftOpcode == Instruction::Shl) {
1650  // For a left shift, we can fold if the comparison is not signed. We can
1651  // also fold a signed comparison if the mask value and comparison value
1652  // are not negative. These constraints may not be obvious, but we can
1653  // prove that they are correct using an SMT solver.
1654  if (!Cmp.isSigned() || (!C2.isNegative() && !C1.isNegative()))
1655  CanFold = true;
1656  } else {
1657  bool IsAshr = ShiftOpcode == Instruction::AShr;
1658  // For a logical right shift, we can fold if the comparison is not signed.
1659  // We can also fold a signed comparison if the shifted mask value and the
1660  // shifted comparison value are not negative. These constraints may not be
1661  // obvious, but we can prove that they are correct using an SMT solver.
1662  // For an arithmetic shift right we can do the same, if we ensure
1663  // the And doesn't use any bits being shifted in. Normally these would
1664  // be turned into lshr by SimplifyDemandedBits, but not if there is an
1665  // additional user.
1666  if (!IsAshr || (C2.shl(*C3).lshr(*C3) == C2)) {
1667  if (!Cmp.isSigned() ||
1668  (!C2.shl(*C3).isNegative() && !C1.shl(*C3).isNegative()))
1669  CanFold = true;
1670  }
1671  }
1672 
1673  if (CanFold) {
1674  APInt NewCst = IsShl ? C1.lshr(*C3) : C1.shl(*C3);
1675  APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3);
1676  // Check to see if we are shifting out any of the bits being compared.
1677  if (SameAsC1 != C1) {
1678  // If we shifted bits out, the fold is not going to work out. As a
1679  // special case, check to see if this means that the result is always
1680  // true or false now.
1681  if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1682  return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1683  if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1684  return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1685  } else {
1686  Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst));
1687  APInt NewAndCst = IsShl ? C2.lshr(*C3) : C2.shl(*C3);
1688  And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst));
1689  And->setOperand(0, Shift->getOperand(0));
1690  Worklist.Add(Shift); // Shift is dead.
1691  return &Cmp;
1692  }
1693  }
1694  }
1695 
1696  // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1697  // preferable because it allows the C2 << Y expression to be hoisted out of a
1698  // loop if Y is invariant and X is not.
1699  if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1700  !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1701  // Compute C2 << Y.
1702  Value *NewShift =
1703  IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1704  : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1705 
1706  // Compute X & (C2 << Y).
1707  Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1708  Cmp.setOperand(0, NewAnd);
1709  return &Cmp;
1710  }
1711 
1712  return nullptr;
1713 }
1714 
1715 /// Fold icmp (and X, C2), C1.
1716 Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp,
1717  BinaryOperator *And,
1718  const APInt &C1) {
1719  bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1720 
1721  // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1722  // TODO: We canonicalize to the longer form for scalars because we have
1723  // better analysis/folds for icmp, and codegen may be better with icmp.
1724  if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isNullValue() &&
1725  match(And->getOperand(1), m_One()))
1726  return new TruncInst(And->getOperand(0), Cmp.getType());
1727 
1728  const APInt *C2;
1729  Value *X;
1730  if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1731  return nullptr;
1732 
1733  // Don't perform the following transforms if the AND has multiple uses
1734  if (!And->hasOneUse())
1735  return nullptr;
1736 
1737  if (Cmp.isEquality() && C1.isNullValue()) {
1738  // Restrict this fold to single-use 'and' (PR10267).
1739  // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1740  if (C2->isSignMask()) {
1741  Constant *Zero = Constant::getNullValue(X->getType());
1742  auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1743  return new ICmpInst(NewPred, X, Zero);
1744  }
1745 
1746  // Restrict this fold only for single-use 'and' (PR10267).
1747  // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1748  if ((~(*C2) + 1).isPowerOf2()) {
1749  Constant *NegBOC =
1750  ConstantExpr::getNeg(cast<Constant>(And->getOperand(1)));
1751  auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1752  return new ICmpInst(NewPred, X, NegBOC);
1753  }
1754  }
1755 
1756  // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1757  // the input width without changing the value produced, eliminate the cast:
1758  //
1759  // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1760  //
1761  // We can do this transformation if the constants do not have their sign bits
1762  // set or if it is an equality comparison. Extending a relational comparison
1763  // when we're checking the sign bit would not work.
1764  Value *W;
1765  if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1766  (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1767  // TODO: Is this a good transform for vectors? Wider types may reduce
1768  // throughput. Should this transform be limited (even for scalars) by using
1769  // shouldChangeType()?
1770  if (!Cmp.getType()->isVectorTy()) {
1771  Type *WideType = W->getType();
1772  unsigned WideScalarBits = WideType->getScalarSizeInBits();
1773  Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1774  Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1775  Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1776  return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1777  }
1778  }
1779 
1780  if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1781  return I;
1782 
1783  // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1784  // (icmp pred (and A, (or (shl 1, B), 1), 0))
1785  //
1786  // iff pred isn't signed
1787  if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1788  match(And->getOperand(1), m_One())) {
1789  Constant *One = cast<Constant>(And->getOperand(1));
1790  Value *Or = And->getOperand(0);
1791  Value *A, *B, *LShr;
1792  if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1793  match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1794  unsigned UsesRemoved = 0;
1795  if (And->hasOneUse())
1796  ++UsesRemoved;
1797  if (Or->hasOneUse())
1798  ++UsesRemoved;
1799  if (LShr->hasOneUse())
1800  ++UsesRemoved;
1801 
1802  // Compute A & ((1 << B) | 1)
1803  Value *NewOr = nullptr;
1804  if (auto *C = dyn_cast<Constant>(B)) {
1805  if (UsesRemoved >= 1)
1806  NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1807  } else {
1808  if (UsesRemoved >= 3)
1809  NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1810  /*HasNUW=*/true),
1811  One, Or->getName());
1812  }
1813  if (NewOr) {
1814  Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1815  Cmp.setOperand(0, NewAnd);
1816  return &Cmp;
1817  }
1818  }
1819  }
1820 
1821  return nullptr;
1822 }
1823 
1824 /// Fold icmp (and X, Y), C.
1825 Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp,
1826  BinaryOperator *And,
1827  const APInt &C) {
1828  if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1829  return I;
1830 
1831  // TODO: These all require that Y is constant too, so refactor with the above.
1832 
1833  // Try to optimize things like "A[i] & 42 == 0" to index computations.
1834  Value *X = And->getOperand(0);
1835  Value *Y = And->getOperand(1);
1836  if (auto *LI = dyn_cast<LoadInst>(X))
1837  if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1838  if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1839  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1840  !LI->isVolatile() && isa<ConstantInt>(Y)) {
1841  ConstantInt *C2 = cast<ConstantInt>(Y);
1842  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1843  return Res;
1844  }
1845 
1846  if (!Cmp.isEquality())
1847  return nullptr;
1848 
1849  // X & -C == -C -> X > u ~C
1850  // X & -C != -C -> X <= u ~C
1851  // iff C is a power of 2
1852  if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1853  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1855  return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1856  }
1857 
1858  // (X & C2) == 0 -> (trunc X) >= 0
1859  // (X & C2) != 0 -> (trunc X) < 0
1860  // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1861  const APInt *C2;
1862  if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1863  int32_t ExactLogBase2 = C2->exactLogBase2();
1864  if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1865  Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1866  if (And->getType()->isVectorTy())
1867  NTy = VectorType::get(NTy, And->getType()->getVectorNumElements());
1868  Value *Trunc = Builder.CreateTrunc(X, NTy);
1869  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1871  return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1872  }
1873  }
1874 
1875  return nullptr;
1876 }
1877 
1878 /// Fold icmp (or X, Y), C.
1879 Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
1880  const APInt &C) {
1881  ICmpInst::Predicate Pred = Cmp.getPredicate();
1882  if (C.isOneValue()) {
1883  // icmp slt signum(V) 1 --> icmp slt V, 1
1884  Value *V = nullptr;
1885  if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1886  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1887  ConstantInt::get(V->getType(), 1));
1888  }
1889 
1890  Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1891  if (Cmp.isEquality() && Cmp.getOperand(1) == OrOp1) {
1892  // X | C == C --> X <=u C
1893  // X | C != C --> X >u C
1894  // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1895  if ((C + 1).isPowerOf2()) {
1897  return new ICmpInst(Pred, OrOp0, OrOp1);
1898  }
1899  // More general: are all bits outside of a mask constant set or not set?
1900  // X | C == C --> (X & ~C) == 0
1901  // X | C != C --> (X & ~C) != 0
1902  if (Or->hasOneUse()) {
1903  Value *A = Builder.CreateAnd(OrOp0, ~C);
1904  return new ICmpInst(Pred, A, ConstantInt::getNullValue(OrOp0->getType()));
1905  }
1906  }
1907 
1908  if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1909  return nullptr;
1910 
1911  Value *P, *Q;
1912  if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1913  // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1914  // -> and (icmp eq P, null), (icmp eq Q, null).
1915  Value *CmpP =
1916  Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1917  Value *CmpQ =
1918  Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1919  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1920  return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1921  }
1922 
1923  // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1924  // a shorter form that has more potential to be folded even further.
1925  Value *X1, *X2, *X3, *X4;
1926  if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1927  match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1928  // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1929  // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1930  Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1931  Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1932  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1933  return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1934  }
1935 
1936  return nullptr;
1937 }
1938 
1939 /// Fold icmp (mul X, Y), C.
1940 Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp,
1941  BinaryOperator *Mul,
1942  const APInt &C) {
1943  const APInt *MulC;
1944  if (!match(Mul->getOperand(1), m_APInt(MulC)))
1945  return nullptr;
1946 
1947  // If this is a test of the sign bit and the multiply is sign-preserving with
1948  // a constant operand, use the multiply LHS operand instead.
1949  ICmpInst::Predicate Pred = Cmp.getPredicate();
1950  if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1951  if (MulC->isNegative())
1952  Pred = ICmpInst::getSwappedPredicate(Pred);
1953  return new ICmpInst(Pred, Mul->getOperand(0),
1955  }
1956 
1957  return nullptr;
1958 }
1959 
1960 /// Fold icmp (shl 1, Y), C.
1962  const APInt &C) {
1963  Value *Y;
1964  if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1965  return nullptr;
1966 
1967  Type *ShiftType = Shl->getType();
1968  unsigned TypeBits = C.getBitWidth();
1969  bool CIsPowerOf2 = C.isPowerOf2();
1970  ICmpInst::Predicate Pred = Cmp.getPredicate();
1971  if (Cmp.isUnsigned()) {
1972  // (1 << Y) pred C -> Y pred Log2(C)
1973  if (!CIsPowerOf2) {
1974  // (1 << Y) < 30 -> Y <= 4
1975  // (1 << Y) <= 30 -> Y <= 4
1976  // (1 << Y) >= 30 -> Y > 4
1977  // (1 << Y) > 30 -> Y > 4
1978  if (Pred == ICmpInst::ICMP_ULT)
1979  Pred = ICmpInst::ICMP_ULE;
1980  else if (Pred == ICmpInst::ICMP_UGE)
1981  Pred = ICmpInst::ICMP_UGT;
1982  }
1983 
1984  // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
1985  // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
1986  unsigned CLog2 = C.logBase2();
1987  if (CLog2 == TypeBits - 1) {
1988  if (Pred == ICmpInst::ICMP_UGE)
1989  Pred = ICmpInst::ICMP_EQ;
1990  else if (Pred == ICmpInst::ICMP_ULT)
1991  Pred = ICmpInst::ICMP_NE;
1992  }
1993  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
1994  } else if (Cmp.isSigned()) {
1995  Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
1996  if (C.isAllOnesValue()) {
1997  // (1 << Y) <= -1 -> Y == 31
1998  if (Pred == ICmpInst::ICMP_SLE)
1999  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2000 
2001  // (1 << Y) > -1 -> Y != 31
2002  if (Pred == ICmpInst::ICMP_SGT)
2003  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2004  } else if (!C) {
2005  // (1 << Y) < 0 -> Y == 31
2006  // (1 << Y) <= 0 -> Y == 31
2007  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
2008  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2009 
2010  // (1 << Y) >= 0 -> Y != 31
2011  // (1 << Y) > 0 -> Y != 31
2012  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
2013  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2014  }
2015  } else if (Cmp.isEquality() && CIsPowerOf2) {
2016  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
2017  }
2018 
2019  return nullptr;
2020 }
2021 
2022 /// Fold icmp (shl X, Y), C.
2023 Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp,
2024  BinaryOperator *Shl,
2025  const APInt &C) {
2026  const APInt *ShiftVal;
2027  if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
2028  return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
2029 
2030  const APInt *ShiftAmt;
2031  if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
2032  return foldICmpShlOne(Cmp, Shl, C);
2033 
2034  // Check that the shift amount is in range. If not, don't perform undefined
2035  // shifts. When the shift is visited, it will be simplified.
2036  unsigned TypeBits = C.getBitWidth();
2037  if (ShiftAmt->uge(TypeBits))
2038  return nullptr;
2039 
2040  ICmpInst::Predicate Pred = Cmp.getPredicate();
2041  Value *X = Shl->getOperand(0);
2042  Type *ShType = Shl->getType();
2043 
2044  // NSW guarantees that we are only shifting out sign bits from the high bits,
2045  // so we can ASHR the compare constant without needing a mask and eliminate
2046  // the shift.
2047  if (Shl->hasNoSignedWrap()) {
2048  if (Pred == ICmpInst::ICMP_SGT) {
2049  // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2050  APInt ShiftedC = C.ashr(*ShiftAmt);
2051  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2052  }
2053  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2054  C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
2055  APInt ShiftedC = C.ashr(*ShiftAmt);
2056  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2057  }
2058  if (Pred == ICmpInst::ICMP_SLT) {
2059  // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2060  // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2061  // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2062  // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2063  assert(!C.isMinSignedValue() && "Unexpected icmp slt");
2064  APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2065  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2066  }
2067  // If this is a signed comparison to 0 and the shift is sign preserving,
2068  // use the shift LHS operand instead; isSignTest may change 'Pred', so only
2069  // do that if we're sure to not continue on in this function.
2070  if (isSignTest(Pred, C))
2071  return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
2072  }
2073 
2074  // NUW guarantees that we are only shifting out zero bits from the high bits,
2075  // so we can LSHR the compare constant without needing a mask and eliminate
2076  // the shift.
2077  if (Shl->hasNoUnsignedWrap()) {
2078  if (Pred == ICmpInst::ICMP_UGT) {
2079  // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2080  APInt ShiftedC = C.lshr(*ShiftAmt);
2081  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2082  }
2083  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2084  C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2085  APInt ShiftedC = C.lshr(*ShiftAmt);
2086  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2087  }
2088  if (Pred == ICmpInst::ICMP_ULT) {
2089  // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2090  // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2091  // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2092  // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2093  assert(C.ugt(0) && "ult 0 should have been eliminated");
2094  APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2095  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2096  }
2097  }
2098 
2099  if (Cmp.isEquality() && Shl->hasOneUse()) {
2100  // Strength-reduce the shift into an 'and'.
2102  ShType,
2103  APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2104  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2105  Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2106  return new ICmpInst(Pred, And, LShrC);
2107  }
2108 
2109  // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2110  bool TrueIfSigned = false;
2111  if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2112  // (X << 31) <s 0 --> (X & 1) != 0
2114  ShType,
2115  APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2116  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2117  return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2118  And, Constant::getNullValue(ShType));
2119  }
2120 
2121  // Simplify 'shl' inequality test into 'and' equality test.
2122  if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2123  // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2124  if ((C + 1).isPowerOf2() &&
2125  (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2126  Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2127  return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2129  And, Constant::getNullValue(ShType));
2130  }
2131  // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2132  if (C.isPowerOf2() &&
2133  (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2134  Value *And =
2135  Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2136  return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2138  And, Constant::getNullValue(ShType));
2139  }
2140  }
2141 
2142  // Transform (icmp pred iM (shl iM %v, N), C)
2143  // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2144  // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2145  // This enables us to get rid of the shift in favor of a trunc that may be
2146  // free on the target. It has the additional benefit of comparing to a
2147  // smaller constant that may be more target-friendly.
2148  unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2149  if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2150  DL.isLegalInteger(TypeBits - Amt)) {
2151  Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2152  if (ShType->isVectorTy())
2153  TruncTy = VectorType::get(TruncTy, ShType->getVectorNumElements());
2154  Constant *NewC =
2155  ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2156  return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2157  }
2158 
2159  return nullptr;
2160 }
2161 
2162 /// Fold icmp ({al}shr X, Y), C.
2163 Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp,
2164  BinaryOperator *Shr,
2165  const APInt &C) {
2166  // An exact shr only shifts out zero bits, so:
2167  // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2168  Value *X = Shr->getOperand(0);
2169  CmpInst::Predicate Pred = Cmp.getPredicate();
2170  if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2171  C.isNullValue())
2172  return new ICmpInst(Pred, X, Cmp.getOperand(1));
2173 
2174  const APInt *ShiftVal;
2175  if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2176  return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2177 
2178  const APInt *ShiftAmt;
2179  if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2180  return nullptr;
2181 
2182  // Check that the shift amount is in range. If not, don't perform undefined
2183  // shifts. When the shift is visited it will be simplified.
2184  unsigned TypeBits = C.getBitWidth();
2185  unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2186  if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2187  return nullptr;
2188 
2189  bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2190  bool IsExact = Shr->isExact();
2191  Type *ShrTy = Shr->getType();
2192  // TODO: If we could guarantee that InstSimplify would handle all of the
2193  // constant-value-based preconditions in the folds below, then we could assert
2194  // those conditions rather than checking them. This is difficult because of
2195  // undef/poison (PR34838).
2196  if (IsAShr) {
2197  if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2198  // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2199  // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2200  APInt ShiftedC = C.shl(ShAmtVal);
2201  if (ShiftedC.ashr(ShAmtVal) == C)
2202  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2203  }
2204  if (Pred == CmpInst::ICMP_SGT) {
2205  // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2206  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2207  if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2208  (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2209  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2210  }
2211  } else {
2212  if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2213  // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2214  // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2215  APInt ShiftedC = C.shl(ShAmtVal);
2216  if (ShiftedC.lshr(ShAmtVal) == C)
2217  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2218  }
2219  if (Pred == CmpInst::ICMP_UGT) {
2220  // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2221  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2222  if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2223  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2224  }
2225  }
2226 
2227  if (!Cmp.isEquality())
2228  return nullptr;
2229 
2230  // Handle equality comparisons of shift-by-constant.
2231 
2232  // If the comparison constant changes with the shift, the comparison cannot
2233  // succeed (bits of the comparison constant cannot match the shifted value).
2234  // This should be known by InstSimplify and already be folded to true/false.
2235  assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2236  (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2237  "Expected icmp+shr simplify did not occur.");
2238 
2239  // If the bits shifted out are known zero, compare the unshifted value:
2240  // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2241  if (Shr->isExact())
2242  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2243 
2244  if (Shr->hasOneUse()) {
2245  // Canonicalize the shift into an 'and':
2246  // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2247  APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2248  Constant *Mask = ConstantInt::get(ShrTy, Val);
2249  Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2250  return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2251  }
2252 
2253  return nullptr;
2254 }
2255 
2256 Instruction *InstCombiner::foldICmpSRemConstant(ICmpInst &Cmp,
2257  BinaryOperator *SRem,
2258  const APInt &C) {
2259  // Match an 'is positive' or 'is negative' comparison of remainder by a
2260  // constant power-of-2 value:
2261  // (X % pow2C) sgt/slt 0
2262  const ICmpInst::Predicate Pred = Cmp.getPredicate();
2263  if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT)
2264  return nullptr;
2265 
2266  // TODO: The one-use check is standard because we do not typically want to
2267  // create longer instruction sequences, but this might be a special-case
2268  // because srem is not good for analysis or codegen.
2269  if (!SRem->hasOneUse())
2270  return nullptr;
2271 
2272  const APInt *DivisorC;
2273  if (!C.isNullValue() || !match(SRem->getOperand(1), m_Power2(DivisorC)))
2274  return nullptr;
2275 
2276  // Mask off the sign bit and the modulo bits (low-bits).
2277  Type *Ty = SRem->getType();
2278  APInt SignMask = APInt::getSignMask(Ty->getScalarSizeInBits());
2279  Constant *MaskC = ConstantInt::get(Ty, SignMask | (*DivisorC - 1));
2280  Value *And = Builder.CreateAnd(SRem->getOperand(0), MaskC);
2281 
2282  // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2283  // bit is set. Example:
2284  // (i8 X % 32) s> 0 --> (X & 159) s> 0
2285  if (Pred == ICmpInst::ICMP_SGT)
2287 
2288  // For 'is negative?' check that the sign-bit is set and at least 1 masked
2289  // bit is set. Example:
2290  // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2291  return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, SignMask));
2292 }
2293 
2294 /// Fold icmp (udiv X, Y), C.
2295 Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp,
2296  BinaryOperator *UDiv,
2297  const APInt &C) {
2298  const APInt *C2;
2299  if (!match(UDiv->getOperand(0), m_APInt(C2)))
2300  return nullptr;
2301 
2302  assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2303 
2304  // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2305  Value *Y = UDiv->getOperand(1);
2306  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2307  assert(!C.isMaxValue() &&
2308  "icmp ugt X, UINT_MAX should have been simplified already.");
2309  return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2310  ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2311  }
2312 
2313  // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2314  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2315  assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2316  return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2317  ConstantInt::get(Y->getType(), C2->udiv(C)));
2318  }
2319 
2320  return nullptr;
2321 }
2322 
2323 /// Fold icmp ({su}div X, Y), C.
2324 Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp,
2325  BinaryOperator *Div,
2326  const APInt &C) {
2327  // Fold: icmp pred ([us]div X, C2), C -> range test
2328  // Fold this div into the comparison, producing a range check.
2329  // Determine, based on the divide type, what the range is being
2330  // checked. If there is an overflow on the low or high side, remember
2331  // it, otherwise compute the range [low, hi) bounding the new value.
2332  // See: InsertRangeTest above for the kinds of replacements possible.
2333  const APInt *C2;
2334  if (!match(Div->getOperand(1), m_APInt(C2)))
2335  return nullptr;
2336 
2337  // FIXME: If the operand types don't match the type of the divide
2338  // then don't attempt this transform. The code below doesn't have the
2339  // logic to deal with a signed divide and an unsigned compare (and
2340  // vice versa). This is because (x /s C2) <s C produces different
2341  // results than (x /s C2) <u C or (x /u C2) <s C or even
2342  // (x /u C2) <u C. Simply casting the operands and result won't
2343  // work. :( The if statement below tests that condition and bails
2344  // if it finds it.
2345  bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2346  if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2347  return nullptr;
2348 
2349  // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2350  // INT_MIN will also fail if the divisor is 1. Although folds of all these
2351  // division-by-constant cases should be present, we can not assert that they
2352  // have happened before we reach this icmp instruction.
2353  if (C2->isNullValue() || C2->isOneValue() ||
2354  (DivIsSigned && C2->isAllOnesValue()))
2355  return nullptr;
2356 
2357  // Compute Prod = C * C2. We are essentially solving an equation of
2358  // form X / C2 = C. We solve for X by multiplying C2 and C.
2359  // By solving for X, we can turn this into a range check instead of computing
2360  // a divide.
2361  APInt Prod = C * *C2;
2362 
2363  // Determine if the product overflows by seeing if the product is not equal to
2364  // the divide. Make sure we do the same kind of divide as in the LHS
2365  // instruction that we're folding.
2366  bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2367 
2368  ICmpInst::Predicate Pred = Cmp.getPredicate();
2369 
2370  // If the division is known to be exact, then there is no remainder from the
2371  // divide, so the covered range size is unit, otherwise it is the divisor.
2372  APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2373 
2374  // Figure out the interval that is being checked. For example, a comparison
2375  // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2376  // Compute this interval based on the constants involved and the signedness of
2377  // the compare/divide. This computes a half-open interval, keeping track of
2378  // whether either value in the interval overflows. After analysis each
2379  // overflow variable is set to 0 if it's corresponding bound variable is valid
2380  // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2381  int LoOverflow = 0, HiOverflow = 0;
2382  APInt LoBound, HiBound;
2383 
2384  if (!DivIsSigned) { // udiv
2385  // e.g. X/5 op 3 --> [15, 20)
2386  LoBound = Prod;
2387  HiOverflow = LoOverflow = ProdOV;
2388  if (!HiOverflow) {
2389  // If this is not an exact divide, then many values in the range collapse
2390  // to the same result value.
2391  HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2392  }
2393  } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2394  if (C.isNullValue()) { // (X / pos) op 0
2395  // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2396  LoBound = -(RangeSize - 1);
2397  HiBound = RangeSize;
2398  } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2399  LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2400  HiOverflow = LoOverflow = ProdOV;
2401  if (!HiOverflow)
2402  HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2403  } else { // (X / pos) op neg
2404  // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2405  HiBound = Prod + 1;
2406  LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2407  if (!LoOverflow) {
2408  APInt DivNeg = -RangeSize;
2409  LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2410  }
2411  }
2412  } else if (C2->isNegative()) { // Divisor is < 0.
2413  if (Div->isExact())
2414  RangeSize.negate();
2415  if (C.isNullValue()) { // (X / neg) op 0
2416  // e.g. X/-5 op 0 --> [-4, 5)
2417  LoBound = RangeSize + 1;
2418  HiBound = -RangeSize;
2419  if (HiBound == *C2) { // -INTMIN = INTMIN
2420  HiOverflow = 1; // [INTMIN+1, overflow)
2421  HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2422  }
2423  } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2424  // e.g. X/-5 op 3 --> [-19, -14)
2425  HiBound = Prod + 1;
2426  HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2427  if (!LoOverflow)
2428  LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2429  } else { // (X / neg) op neg
2430  LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2431  LoOverflow = HiOverflow = ProdOV;
2432  if (!HiOverflow)
2433  HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2434  }
2435 
2436  // Dividing by a negative swaps the condition. LT <-> GT
2437  Pred = ICmpInst::getSwappedPredicate(Pred);
2438  }
2439 
2440  Value *X = Div->getOperand(0);
2441  switch (Pred) {
2442  default: llvm_unreachable("Unhandled icmp opcode!");
2443  case ICmpInst::ICMP_EQ:
2444  if (LoOverflow && HiOverflow)
2445  return replaceInstUsesWith(Cmp, Builder.getFalse());
2446  if (HiOverflow)
2447  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2448  ICmpInst::ICMP_UGE, X,
2449  ConstantInt::get(Div->getType(), LoBound));
2450  if (LoOverflow)
2451  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2452  ICmpInst::ICMP_ULT, X,
2453  ConstantInt::get(Div->getType(), HiBound));
2454  return replaceInstUsesWith(
2455  Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2456  case ICmpInst::ICMP_NE:
2457  if (LoOverflow && HiOverflow)
2458  return replaceInstUsesWith(Cmp, Builder.getTrue());
2459  if (HiOverflow)
2460  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2461  ICmpInst::ICMP_ULT, X,
2462  ConstantInt::get(Div->getType(), LoBound));
2463  if (LoOverflow)
2464  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2465  ICmpInst::ICMP_UGE, X,
2466  ConstantInt::get(Div->getType(), HiBound));
2467  return replaceInstUsesWith(Cmp,
2468  insertRangeTest(X, LoBound, HiBound,
2469  DivIsSigned, false));
2470  case ICmpInst::ICMP_ULT:
2471  case ICmpInst::ICMP_SLT:
2472  if (LoOverflow == +1) // Low bound is greater than input range.
2473  return replaceInstUsesWith(Cmp, Builder.getTrue());
2474  if (LoOverflow == -1) // Low bound is less than input range.
2475  return replaceInstUsesWith(Cmp, Builder.getFalse());
2476  return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2477  case ICmpInst::ICMP_UGT:
2478  case ICmpInst::ICMP_SGT:
2479  if (HiOverflow == +1) // High bound greater than input range.
2480  return replaceInstUsesWith(Cmp, Builder.getFalse());
2481  if (HiOverflow == -1) // High bound less than input range.
2482  return replaceInstUsesWith(Cmp, Builder.getTrue());
2483  if (Pred == ICmpInst::ICMP_UGT)
2484  return new ICmpInst(ICmpInst::ICMP_UGE, X,
2485  ConstantInt::get(Div->getType(), HiBound));
2486  return new ICmpInst(ICmpInst::ICMP_SGE, X,
2487  ConstantInt::get(Div->getType(), HiBound));
2488  }
2489 
2490  return nullptr;
2491 }
2492 
2493 /// Fold icmp (sub X, Y), C.
2494 Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp,
2495  BinaryOperator *Sub,
2496  const APInt &C) {
2497  Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2498  ICmpInst::Predicate Pred = Cmp.getPredicate();
2499  const APInt *C2;
2500  APInt SubResult;
2501 
2502  // icmp eq/ne (sub C, Y), C -> icmp eq/ne Y, 0
2503  if (match(X, m_APInt(C2)) && *C2 == C && Cmp.isEquality())
2504  return new ICmpInst(Cmp.getPredicate(), Y,
2505  ConstantInt::get(Y->getType(), 0));
2506 
2507  // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2508  if (match(X, m_APInt(C2)) &&
2509  ((Cmp.isUnsigned() && Sub->hasNoUnsignedWrap()) ||
2510  (Cmp.isSigned() && Sub->hasNoSignedWrap())) &&
2511  !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2512  return new ICmpInst(Cmp.getSwappedPredicate(), Y,
2513  ConstantInt::get(Y->getType(), SubResult));
2514 
2515  // The following transforms are only worth it if the only user of the subtract
2516  // is the icmp.
2517  if (!Sub->hasOneUse())
2518  return nullptr;
2519 
2520  if (Sub->hasNoSignedWrap()) {
2521  // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2522  if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2523  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2524 
2525  // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2526  if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2527  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2528 
2529  // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2530  if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2531  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2532 
2533  // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2534  if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2535  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2536  }
2537 
2538  if (!match(X, m_APInt(C2)))
2539  return nullptr;
2540 
2541  // C2 - Y <u C -> (Y | (C - 1)) == C2
2542  // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2543  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2544  (*C2 & (C - 1)) == (C - 1))
2545  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2546 
2547  // C2 - Y >u C -> (Y | C) != C2
2548  // iff C2 & C == C and C + 1 is a power of 2
2549  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2550  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2551 
2552  return nullptr;
2553 }
2554 
2555 /// Fold icmp (add X, Y), C.
2556 Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp,
2558  const APInt &C) {
2559  Value *Y = Add->getOperand(1);
2560  const APInt *C2;
2561  if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2562  return nullptr;
2563 
2564  // Fold icmp pred (add X, C2), C.
2565  Value *X = Add->getOperand(0);
2566  Type *Ty = Add->getType();
2567  CmpInst::Predicate Pred = Cmp.getPredicate();
2568 
2569  if (!Add->hasOneUse())
2570  return nullptr;
2571 
2572  // If the add does not wrap, we can always adjust the compare by subtracting
2573  // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2574  // are canonicalized to SGT/SLT/UGT/ULT.
2575  if ((Add->hasNoSignedWrap() &&
2576  (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2577  (Add->hasNoUnsignedWrap() &&
2578  (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2579  bool Overflow;
2580  APInt NewC =
2581  Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2582  // If there is overflow, the result must be true or false.
2583  // TODO: Can we assert there is no overflow because InstSimplify always
2584  // handles those cases?
2585  if (!Overflow)
2586  // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2587  return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2588  }
2589 
2590  auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2591  const APInt &Upper = CR.getUpper();
2592  const APInt &Lower = CR.getLower();
2593  if (Cmp.isSigned()) {
2594  if (Lower.isSignMask())
2595  return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2596  if (Upper.isSignMask())
2597  return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2598  } else {
2599  if (Lower.isMinValue())
2600  return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2601  if (Upper.isMinValue())
2602  return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2603  }
2604 
2605  // X+C <u C2 -> (X & -C2) == C
2606  // iff C & (C2-1) == 0
2607  // C2 is a power of 2
2608  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2609  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2610  ConstantExpr::getNeg(cast<Constant>(Y)));
2611 
2612  // X+C >u C2 -> (X & ~C2) != C
2613  // iff C & C2 == 0
2614  // C2+1 is a power of 2
2615  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2616  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2617  ConstantExpr::getNeg(cast<Constant>(Y)));
2618 
2619  return nullptr;
2620 }
2621 
2622 bool InstCombiner::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2623  Value *&RHS, ConstantInt *&Less,
2624  ConstantInt *&Equal,
2625  ConstantInt *&Greater) {
2626  // TODO: Generalize this to work with other comparison idioms or ensure
2627  // they get canonicalized into this form.
2628 
2629  // select i1 (a == b),
2630  // i32 Equal,
2631  // i32 (select i1 (a < b), i32 Less, i32 Greater)
2632  // where Equal, Less and Greater are placeholders for any three constants.
2633  ICmpInst::Predicate PredA;
2634  if (!match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) ||
2635  !ICmpInst::isEquality(PredA))
2636  return false;
2637  Value *EqualVal = SI->getTrueValue();
2638  Value *UnequalVal = SI->getFalseValue();
2639  // We still can get non-canonical predicate here, so canonicalize.
2640  if (PredA == ICmpInst::ICMP_NE)
2641  std::swap(EqualVal, UnequalVal);
2642  if (!match(EqualVal, m_ConstantInt(Equal)))
2643  return false;
2644  ICmpInst::Predicate PredB;
2645  Value *LHS2, *RHS2;
2646  if (!match(UnequalVal, m_Select(m_ICmp(PredB, m_Value(LHS2), m_Value(RHS2)),
2647  m_ConstantInt(Less), m_ConstantInt(Greater))))
2648  return false;
2649  // We can get predicate mismatch here, so canonicalize if possible:
2650  // First, ensure that 'LHS' match.
2651  if (LHS2 != LHS) {
2652  // x sgt y <--> y slt x
2653  std::swap(LHS2, RHS2);
2654  PredB = ICmpInst::getSwappedPredicate(PredB);
2655  }
2656  if (LHS2 != LHS)
2657  return false;
2658  // We also need to canonicalize 'RHS'.
2659  if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(RHS2)) {
2660  // x sgt C-1 <--> x sge C <--> not(x slt C)
2661  auto FlippedStrictness =
2662  getFlippedStrictnessPredicateAndConstant(PredB, cast<Constant>(RHS2));
2663  if (!FlippedStrictness)
2664  return false;
2665  assert(FlippedStrictness->first == ICmpInst::ICMP_SGE && "Sanity check");
2666  RHS2 = FlippedStrictness->second;
2667  // And kind-of perform the result swap.
2668  std::swap(Less, Greater);
2669  PredB = ICmpInst::ICMP_SLT;
2670  }
2671  return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
2672 }
2673 
2674 Instruction *InstCombiner::foldICmpSelectConstant(ICmpInst &Cmp,
2675  SelectInst *Select,
2676  ConstantInt *C) {
2677 
2678  assert(C && "Cmp RHS should be a constant int!");
2679  // If we're testing a constant value against the result of a three way
2680  // comparison, the result can be expressed directly in terms of the
2681  // original values being compared. Note: We could possibly be more
2682  // aggressive here and remove the hasOneUse test. The original select is
2683  // really likely to simplify or sink when we remove a test of the result.
2684  Value *OrigLHS, *OrigRHS;
2685  ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2686  if (Cmp.hasOneUse() &&
2687  matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2688  C3GreaterThan)) {
2689  assert(C1LessThan && C2Equal && C3GreaterThan);
2690 
2691  bool TrueWhenLessThan =
2692  ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2693  ->isAllOnesValue();
2694  bool TrueWhenEqual =
2695  ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2696  ->isAllOnesValue();
2697  bool TrueWhenGreaterThan =
2698  ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2699  ->isAllOnesValue();
2700 
2701  // This generates the new instruction that will replace the original Cmp
2702  // Instruction. Instead of enumerating the various combinations when
2703  // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2704  // false, we rely on chaining of ORs and future passes of InstCombine to
2705  // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2706 
2707  // When none of the three constants satisfy the predicate for the RHS (C),
2708  // the entire original Cmp can be simplified to a false.
2709  Value *Cond = Builder.getFalse();
2710  if (TrueWhenLessThan)
2711  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2712  OrigLHS, OrigRHS));
2713  if (TrueWhenEqual)
2714  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2715  OrigLHS, OrigRHS));
2716  if (TrueWhenGreaterThan)
2717  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2718  OrigLHS, OrigRHS));
2719 
2720  return replaceInstUsesWith(Cmp, Cond);
2721  }
2722  return nullptr;
2723 }
2724 
2726  InstCombiner::BuilderTy &Builder) {
2727  auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2728  if (!Bitcast)
2729  return nullptr;
2730 
2731  ICmpInst::Predicate Pred = Cmp.getPredicate();
2732  Value *Op1 = Cmp.getOperand(1);
2733  Value *BCSrcOp = Bitcast->getOperand(0);
2734 
2735  // Make sure the bitcast doesn't change the number of vector elements.
2736  if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2737  Bitcast->getDestTy()->getScalarSizeInBits()) {
2738  // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2739  Value *X;
2740  if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2741  // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2742  // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2743  // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2744  // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2745  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2746  Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2747  match(Op1, m_Zero()))
2748  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2749 
2750  // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2751  if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2752  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2753 
2754  // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2755  if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2756  return new ICmpInst(Pred, X,
2758  }
2759 
2760  // Zero-equality checks are preserved through unsigned floating-point casts:
2761  // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2762  // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2763  if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2764  if (Cmp.isEquality() && match(Op1, m_Zero()))
2765  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2766  }
2767 
2768  // Test to see if the operands of the icmp are casted versions of other
2769  // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2770  if (Bitcast->getType()->isPointerTy() &&
2771  (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2772  // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2773  // so eliminate it as well.
2774  if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2775  Op1 = BC2->getOperand(0);
2776 
2777  Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2778  return new ICmpInst(Pred, BCSrcOp, Op1);
2779  }
2780 
2781  // Folding: icmp <pred> iN X, C
2782  // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2783  // and C is a splat of a K-bit pattern
2784  // and SC is a constant vector = <C', C', C', ..., C'>
2785  // Into:
2786  // %E = extractelement <M x iK> %vec, i32 C'
2787  // icmp <pred> iK %E, trunc(C)
2788  const APInt *C;
2789  if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2790  !Bitcast->getType()->isIntegerTy() ||
2791  !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2792  return nullptr;
2793 
2794  Value *Vec;
2795  Constant *Mask;
2796  if (match(BCSrcOp,
2797  m_ShuffleVector(m_Value(Vec), m_Undef(), m_Constant(Mask)))) {
2798  // Check whether every element of Mask is the same constant
2799  if (auto *Elem = dyn_cast_or_null<ConstantInt>(Mask->getSplatValue())) {
2800  auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2801  auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2802  if (C->isSplat(EltTy->getBitWidth())) {
2803  // Fold the icmp based on the value of C
2804  // If C is M copies of an iK sized bit pattern,
2805  // then:
2806  // => %E = extractelement <N x iK> %vec, i32 Elem
2807  // icmp <pred> iK %SplatVal, <pattern>
2808  Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2809  Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2810  return new ICmpInst(Pred, Extract, NewC);
2811  }
2812  }
2813  }
2814  return nullptr;
2815 }
2816 
2817 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2818 /// where X is some kind of instruction.
2819 Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) {
2820  const APInt *C;
2821  if (!match(Cmp.getOperand(1), m_APInt(C)))
2822  return nullptr;
2823 
2824  if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2825  switch (BO->getOpcode()) {
2826  case Instruction::Xor:
2827  if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2828  return I;
2829  break;
2830  case Instruction::And:
2831  if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2832  return I;
2833  break;
2834  case Instruction::Or:
2835  if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2836  return I;
2837  break;
2838  case Instruction::Mul:
2839  if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2840  return I;
2841  break;
2842  case Instruction::Shl:
2843  if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2844  return I;
2845  break;
2846  case Instruction::LShr:
2847  case Instruction::AShr:
2848  if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2849  return I;
2850  break;
2851  case Instruction::SRem:
2852  if (Instruction *I = foldICmpSRemConstant(Cmp, BO, *C))
2853  return I;
2854  break;
2855  case Instruction::UDiv:
2856  if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2857  return I;
2859  case Instruction::SDiv:
2860  if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2861  return I;
2862  break;
2863  case Instruction::Sub:
2864  if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2865  return I;
2866  break;
2867  case Instruction::Add:
2868  if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2869  return I;
2870  break;
2871  default:
2872  break;
2873  }
2874  // TODO: These folds could be refactored to be part of the above calls.
2875  if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2876  return I;
2877  }
2878 
2879  // Match against CmpInst LHS being instructions other than binary operators.
2880 
2881  if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2882  // For now, we only support constant integers while folding the
2883  // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2884  // similar to the cases handled by binary ops above.
2885  if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2886  if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2887  return I;
2888  }
2889 
2890  if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2891  if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2892  return I;
2893  }
2894 
2895  if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
2896  if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
2897  return I;
2898 
2899  return nullptr;
2900 }
2901 
2902 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2903 /// icmp eq/ne BO, C.
2904 Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
2905  BinaryOperator *BO,
2906  const APInt &C) {
2907  // TODO: Some of these folds could work with arbitrary constants, but this
2908  // function is limited to scalar and vector splat constants.
2909  if (!Cmp.isEquality())
2910  return nullptr;
2911 
2912  ICmpInst::Predicate Pred = Cmp.getPredicate();
2913  bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2914  Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2915  Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2916 
2917  switch (BO->getOpcode()) {
2918  case Instruction::SRem:
2919  // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2920  if (C.isNullValue() && BO->hasOneUse()) {
2921  const APInt *BOC;
2922  if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2923  Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
2924  return new ICmpInst(Pred, NewRem,
2926  }
2927  }
2928  break;
2929  case Instruction::Add: {
2930  // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
2931  const APInt *BOC;
2932  if (match(BOp1, m_APInt(BOC))) {
2933  if (BO->hasOneUse()) {
2934  Constant *SubC = ConstantExpr::getSub(RHS, cast<Constant>(BOp1));
2935  return new ICmpInst(Pred, BOp0, SubC);
2936  }
2937  } else if (C.isNullValue()) {
2938  // Replace ((add A, B) != 0) with (A != -B) if A or B is
2939  // efficiently invertible, or if the add has just this one use.
2940  if (Value *NegVal = dyn_castNegVal(BOp1))
2941  return new ICmpInst(Pred, BOp0, NegVal);
2942  if (Value *NegVal = dyn_castNegVal(BOp0))
2943  return new ICmpInst(Pred, NegVal, BOp1);
2944  if (BO->hasOneUse()) {
2945  Value *Neg = Builder.CreateNeg(BOp1);
2946  Neg->takeName(BO);
2947  return new ICmpInst(Pred, BOp0, Neg);
2948  }
2949  }
2950  break;
2951  }
2952  case Instruction::Xor:
2953  if (BO->hasOneUse()) {
2954  if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2955  // For the xor case, we can xor two constants together, eliminating
2956  // the explicit xor.
2957  return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
2958  } else if (C.isNullValue()) {
2959  // Replace ((xor A, B) != 0) with (A != B)
2960  return new ICmpInst(Pred, BOp0, BOp1);
2961  }
2962  }
2963  break;
2964  case Instruction::Sub:
2965  if (BO->hasOneUse()) {
2966  const APInt *BOC;
2967  if (match(BOp0, m_APInt(BOC))) {
2968  // Replace ((sub BOC, B) != C) with (B != BOC-C).
2969  Constant *SubC = ConstantExpr::getSub(cast<Constant>(BOp0), RHS);
2970  return new ICmpInst(Pred, BOp1, SubC);
2971  } else if (C.isNullValue()) {
2972  // Replace ((sub A, B) != 0) with (A != B).
2973  return new ICmpInst(Pred, BOp0, BOp1);
2974  }
2975  }
2976  break;
2977  case Instruction::Or: {
2978  const APInt *BOC;
2979  if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
2980  // Comparing if all bits outside of a constant mask are set?
2981  // Replace (X | C) == -1 with (X & ~C) == ~C.
2982  // This removes the -1 constant.
2983  Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
2984  Value *And = Builder.CreateAnd(BOp0, NotBOC);
2985  return new ICmpInst(Pred, And, NotBOC);
2986  }
2987  break;
2988  }
2989  case Instruction::And: {
2990  const APInt *BOC;
2991  if (match(BOp1, m_APInt(BOC))) {
2992  // If we have ((X & C) == C), turn it into ((X & C) != 0).
2993  if (C == *BOC && C.isPowerOf2())
2994  return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
2995  BO, Constant::getNullValue(RHS->getType()));
2996  }
2997  break;
2998  }
2999  case Instruction::Mul:
3000  if (C.isNullValue() && BO->hasNoSignedWrap()) {
3001  const APInt *BOC;
3002  if (match(BOp1, m_APInt(BOC)) && !BOC->isNullValue()) {
3003  // The trivial case (mul X, 0) is handled by InstSimplify.
3004  // General case : (mul X, C) != 0 iff X != 0
3005  // (mul X, C) == 0 iff X == 0
3006  return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType()));
3007  }
3008  }
3009  break;
3010  case Instruction::UDiv:
3011  if (C.isNullValue()) {
3012  // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3013  auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3014  return new ICmpInst(NewPred, BOp1, BOp0);
3015  }
3016  break;
3017  default:
3018  break;
3019  }
3020  return nullptr;
3021 }
3022 
3023 /// Fold an equality icmp with LLVM intrinsic and constant operand.
3024 Instruction *InstCombiner::foldICmpEqIntrinsicWithConstant(ICmpInst &Cmp,
3025  IntrinsicInst *II,
3026  const APInt &C) {
3027  Type *Ty = II->getType();
3028  unsigned BitWidth = C.getBitWidth();
3029  switch (II->getIntrinsicID()) {
3030  case Intrinsic::bswap:
3031  Worklist.Add(II);
3032  Cmp.setOperand(0, II->getArgOperand(0));
3033  Cmp.setOperand(1, ConstantInt::get(Ty, C.byteSwap()));
3034  return &Cmp;
3035 
3036  case Intrinsic::ctlz:
3037  case Intrinsic::cttz: {
3038  // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3039  if (C == BitWidth) {
3040  Worklist.Add(II);
3041  Cmp.setOperand(0, II->getArgOperand(0));
3043  return &Cmp;
3044  }
3045 
3046  // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3047  // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3048  // Limit to one use to ensure we don't increase instruction count.
3049  unsigned Num = C.getLimitedValue(BitWidth);
3050  if (Num != BitWidth && II->hasOneUse()) {
3051  bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3052  APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
3053  : APInt::getHighBitsSet(BitWidth, Num + 1);
3054  APInt Mask2 = IsTrailing
3055  ? APInt::getOneBitSet(BitWidth, Num)
3056  : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3057  Cmp.setOperand(0, Builder.CreateAnd(II->getArgOperand(0), Mask1));
3058  Cmp.setOperand(1, ConstantInt::get(Ty, Mask2));
3059  Worklist.Add(II);
3060  return &Cmp;
3061  }
3062  break;
3063  }
3064 
3065  case Intrinsic::ctpop: {
3066  // popcount(A) == 0 -> A == 0 and likewise for !=
3067  // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3068  bool IsZero = C.isNullValue();
3069  if (IsZero || C == BitWidth) {
3070  Worklist.Add(II);
3071  Cmp.setOperand(0, II->getArgOperand(0));
3072  auto *NewOp =
3074  Cmp.setOperand(1, NewOp);
3075  return &Cmp;
3076  }
3077  break;
3078  }
3079 
3080  case Intrinsic::uadd_sat: {
3081  // uadd.sat(a, b) == 0 -> (a | b) == 0
3082  if (C.isNullValue()) {
3083  Value *Or = Builder.CreateOr(II->getArgOperand(0), II->getArgOperand(1));
3084  return replaceInstUsesWith(Cmp, Builder.CreateICmp(
3085  Cmp.getPredicate(), Or, Constant::getNullValue(Ty)));
3086 
3087  }
3088  break;
3089  }
3090 
3091  case Intrinsic::usub_sat: {
3092  // usub.sat(a, b) == 0 -> a <= b
3093  if (C.isNullValue()) {
3096  return ICmpInst::Create(Instruction::ICmp, NewPred,
3097  II->getArgOperand(0), II->getArgOperand(1));
3098  }
3099  break;
3100  }
3101  default:
3102  break;
3103  }
3104 
3105  return nullptr;
3106 }
3107 
3108 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
3109 Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
3110  IntrinsicInst *II,
3111  const APInt &C) {
3112  if (Cmp.isEquality())
3113  return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
3114 
3115  Type *Ty = II->getType();
3116  unsigned BitWidth = C.getBitWidth();
3117  switch (II->getIntrinsicID()) {
3118  case Intrinsic::ctlz: {
3119  // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
3120  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3121  unsigned Num = C.getLimitedValue();
3122  APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3123  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
3124  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3125  }
3126 
3127  // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
3128  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
3129  C.uge(1) && C.ule(BitWidth)) {
3130  unsigned Num = C.getLimitedValue();
3131  APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
3132  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
3133  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3134  }
3135  break;
3136  }
3137  case Intrinsic::cttz: {
3138  // Limit to one use to ensure we don't increase instruction count.
3139  if (!II->hasOneUse())
3140  return nullptr;
3141 
3142  // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
3143  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3144  APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
3145  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
3146  Builder.CreateAnd(II->getArgOperand(0), Mask),
3148  }
3149 
3150  // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
3151  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
3152  C.uge(1) && C.ule(BitWidth)) {
3154  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
3155  Builder.CreateAnd(II->getArgOperand(0), Mask),
3157  }
3158  break;
3159  }
3160  default:
3161  break;
3162  }
3163 
3164  return nullptr;
3165 }
3166 
3167 /// Handle icmp with constant (but not simple integer constant) RHS.
3168 Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) {
3169  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3170  Constant *RHSC = dyn_cast<Constant>(Op1);
3171  Instruction *LHSI = dyn_cast<Instruction>(Op0);
3172  if (!RHSC || !LHSI)
3173  return nullptr;
3174 
3175  switch (LHSI->getOpcode()) {
3176  case Instruction::GetElementPtr:
3177  // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3178  if (RHSC->isNullValue() &&
3179  cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3180  return new ICmpInst(
3181  I.getPredicate(), LHSI->getOperand(0),
3182  Constant::getNullValue(LHSI->getOperand(0)->getType()));
3183  break;
3184  case Instruction::PHI:
3185  // Only fold icmp into the PHI if the phi and icmp are in the same
3186  // block. If in the same block, we're encouraging jump threading. If
3187  // not, we are just pessimizing the code by making an i1 phi.
3188  if (LHSI->getParent() == I.getParent())
3189  if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
3190  return NV;
3191  break;
3192  case Instruction::Select: {
3193  // If either operand of the select is a constant, we can fold the
3194  // comparison into the select arms, which will cause one to be
3195  // constant folded and the select turned into a bitwise or.
3196  Value *Op1 = nullptr, *Op2 = nullptr;
3197  ConstantInt *CI = nullptr;
3198  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3199  Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3200  CI = dyn_cast<ConstantInt>(Op1);
3201  }
3202  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3203  Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3204  CI = dyn_cast<ConstantInt>(Op2);
3205  }
3206 
3207  // We only want to perform this transformation if it will not lead to
3208  // additional code. This is true if either both sides of the select
3209  // fold to a constant (in which case the icmp is replaced with a select
3210  // which will usually simplify) or this is the only user of the
3211  // select (in which case we are trading a select+icmp for a simpler
3212  // select+icmp) or all uses of the select can be replaced based on
3213  // dominance information ("Global cases").
3214  bool Transform = false;
3215  if (Op1 && Op2)
3216  Transform = true;
3217  else if (Op1 || Op2) {
3218  // Local case
3219  if (LHSI->hasOneUse())
3220  Transform = true;
3221  // Global cases
3222  else if (CI && !CI->isZero())
3223  // When Op1 is constant try replacing select with second operand.
3224  // Otherwise Op2 is constant and try replacing select with first
3225  // operand.
3226  Transform =
3227  replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3228  }
3229  if (Transform) {
3230  if (!Op1)
3231  Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3232  I.getName());
3233  if (!Op2)
3234  Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3235  I.getName());
3236  return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3237  }
3238  break;
3239  }
3240  case Instruction::IntToPtr:
3241  // icmp pred inttoptr(X), null -> icmp pred X, 0
3242  if (RHSC->isNullValue() &&
3243  DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3244  return new ICmpInst(
3245  I.getPredicate(), LHSI->getOperand(0),
3246  Constant::getNullValue(LHSI->getOperand(0)->getType()));
3247  break;
3248 
3249  case Instruction::Load:
3250  // Try to optimize things like "A[i] > 4" to index computations.
3251  if (GetElementPtrInst *GEP =
3252  dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3253  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3254  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3255  !cast<LoadInst>(LHSI)->isVolatile())
3256  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3257  return Res;
3258  }
3259  break;
3260  }
3261 
3262  return nullptr;
3263 }
3264 
3265 /// Some comparisons can be simplified.
3266 /// In this case, we are looking for comparisons that look like
3267 /// a check for a lossy truncation.
3268 /// Folds:
3269 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3270 /// Where Mask is some pattern that produces all-ones in low bits:
3271 /// (-1 >> y)
3272 /// ((-1 << y) >> y) <- non-canonical, has extra uses
3273 /// ~(-1 << y)
3274 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
3275 /// The Mask can be a constant, too.
3276 /// For some predicates, the operands are commutative.
3277 /// For others, x can only be on a specific side.
3279  InstCombiner::BuilderTy &Builder) {
3280  ICmpInst::Predicate SrcPred;
3281  Value *X, *M, *Y;
3282  auto m_VariableMask = m_CombineOr(
3284  m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3286  m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
3287  auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3288  if (!match(&I, m_c_ICmp(SrcPred,
3289  m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
3290  m_Deferred(X))))
3291  return nullptr;
3292 
3293  ICmpInst::Predicate DstPred;
3294  switch (SrcPred) {
3295  case ICmpInst::Predicate::ICMP_EQ:
3296  // x & (-1 >> y) == x -> x u<= (-1 >> y)
3297  DstPred = ICmpInst::Predicate::ICMP_ULE;
3298  break;
3299  case ICmpInst::Predicate::ICMP_NE:
3300  // x & (-1 >> y) != x -> x u> (-1 >> y)
3301  DstPred = ICmpInst::Predicate::ICMP_UGT;
3302  break;
3303  case ICmpInst::Predicate::ICMP_UGT:
3304  // x u> x & (-1 >> y) -> x u> (-1 >> y)
3305  assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
3306  DstPred = ICmpInst::Predicate::ICMP_UGT;
3307  break;
3308  case ICmpInst::Predicate::ICMP_UGE:
3309  // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3310  assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
3311  DstPred = ICmpInst::Predicate::ICMP_ULE;
3312  break;
3313  case ICmpInst::Predicate::ICMP_ULT:
3314  // x & (-1 >> y) u< x -> x u> (-1 >> y)
3315  assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
3316  DstPred = ICmpInst::Predicate::ICMP_UGT;
3317  break;
3318  case ICmpInst::Predicate::ICMP_ULE:
3319  // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3320  assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
3321  DstPred = ICmpInst::Predicate::ICMP_ULE;
3322  break;
3323  case ICmpInst::Predicate::ICMP_SGT:
3324  // x s> x & (-1 >> y) -> x s> (-1 >> y)
3325  if (X != I.getOperand(0)) // X must be on LHS of comparison!
3326  return nullptr; // Ignore the other case.
3327  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3328  return nullptr;
3329  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3330  return nullptr;
3331  DstPred = ICmpInst::Predicate::ICMP_SGT;
3332  break;
3333  case ICmpInst::Predicate::ICMP_SGE:
3334  // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3335  if (X != I.getOperand(1)) // X must be on RHS of comparison!
3336  return nullptr; // Ignore the other case.
3337  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3338  return nullptr;
3339  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3340  return nullptr;
3341  DstPred = ICmpInst::Predicate::ICMP_SLE;
3342  break;
3343  case ICmpInst::Predicate::ICMP_SLT:
3344  // x & (-1 >> y) s< x -> x s> (-1 >> y)
3345  if (X != I.getOperand(1)) // X must be on RHS of comparison!
3346  return nullptr; // Ignore the other case.
3347  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3348  return nullptr;
3349  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3350  return nullptr;
3351  DstPred = ICmpInst::Predicate::ICMP_SGT;
3352  break;
3353  case ICmpInst::Predicate::ICMP_SLE:
3354  // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3355  if (X != I.getOperand(0)) // X must be on LHS of comparison!
3356  return nullptr; // Ignore the other case.
3357  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3358  return nullptr;
3359  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3360  return nullptr;
3361  DstPred = ICmpInst::Predicate::ICMP_SLE;
3362  break;
3363  default:
3364  llvm_unreachable("All possible folds are handled.");
3365  }
3366 
3367  return Builder.CreateICmp(DstPred, X, M);
3368 }
3369 
3370 /// Some comparisons can be simplified.
3371 /// In this case, we are looking for comparisons that look like
3372 /// a check for a lossy signed truncation.
3373 /// Folds: (MaskedBits is a constant.)
3374 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3375 /// Into:
3376 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3377 /// Where KeptBits = bitwidth(%x) - MaskedBits
3378 static Value *
3380  InstCombiner::BuilderTy &Builder) {
3381  ICmpInst::Predicate SrcPred;
3382  Value *X;
3383  const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3384  // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3385  if (!match(&I, m_c_ICmp(SrcPred,
3386  m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3387  m_APInt(C1))),
3388  m_Deferred(X))))
3389  return nullptr;
3390 
3391  // Potential handling of non-splats: for each element:
3392  // * if both are undef, replace with constant 0.
3393  // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3394  // * if both are not undef, and are different, bailout.
3395  // * else, only one is undef, then pick the non-undef one.
3396 
3397  // The shift amount must be equal.
3398  if (*C0 != *C1)
3399  return nullptr;
3400  const APInt &MaskedBits = *C0;
3401  assert(MaskedBits != 0 && "shift by zero should be folded away already.");
3402 
3403  ICmpInst::Predicate DstPred;
3404  switch (SrcPred) {
3405  case ICmpInst::Predicate::ICMP_EQ:
3406  // ((%x << MaskedBits) a>> MaskedBits) == %x
3407  // =>
3408  // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3409  DstPred = ICmpInst::Predicate::ICMP_ULT;
3410  break;
3411  case ICmpInst::Predicate::ICMP_NE:
3412  // ((%x << MaskedBits) a>> MaskedBits) != %x
3413  // =>
3414  // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3415  DstPred = ICmpInst::Predicate::ICMP_UGE;
3416  break;
3417  // FIXME: are more folds possible?
3418  default:
3419  return nullptr;
3420  }
3421 
3422  auto *XType = X->getType();
3423  const unsigned XBitWidth = XType->getScalarSizeInBits();
3424  const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3425  assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3426 
3427  // KeptBits = bitwidth(%x) - MaskedBits
3428  const APInt KeptBits = BitWidth - MaskedBits;
3429  assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3430  // ICmpCst = (1 << KeptBits)
3431  const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3432  assert(ICmpCst.isPowerOf2());
3433  // AddCst = (1 << (KeptBits-1))
3434  const APInt AddCst = ICmpCst.lshr(1);
3435  assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3436 
3437  // T0 = add %x, AddCst
3438  Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3439  // T1 = T0 DstPred ICmpCst
3440  Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3441 
3442  return T1;
3443 }
3444 
3445 // Given pattern:
3446 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3447 // we should move shifts to the same hand of 'and', i.e. rewrite as
3448 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3449 // We are only interested in opposite logical shifts here.
3450 // One of the shifts can be truncated.
3451 // If we can, we want to end up creating 'lshr' shift.
3452 static Value *
3454  InstCombiner::BuilderTy &Builder) {
3455  if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
3456  !I.getOperand(0)->hasOneUse())
3457  return nullptr;
3458 
3459  auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
3460 
3461  // Look for an 'and' of two logical shifts, one of which may be truncated.
3462  // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
3463  Instruction *XShift, *MaybeTruncation, *YShift;
3464  if (!match(
3465  I.getOperand(0),
3466  m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
3468  m_AnyLogicalShift, m_Instruction(YShift))),
3469  m_Instruction(MaybeTruncation)))))
3470  return nullptr;
3471 
3472  // We potentially looked past 'trunc', but only when matching YShift,
3473  // therefore YShift must have the widest type.
3474  Instruction *WidestShift = YShift;
3475  // Therefore XShift must have the shallowest type.
3476  // Or they both have identical types if there was no truncation.
3477  Instruction *NarrowestShift = XShift;
3478 
3479  Type *WidestTy = WidestShift->getType();
3480  assert(NarrowestShift->getType() == I.getOperand(0)->getType() &&
3481  "We did not look past any shifts while matching XShift though.");
3482  bool HadTrunc = WidestTy != I.getOperand(0)->getType();
3483 
3484  // If YShift is a 'lshr', swap the shifts around.
3485  if (match(YShift, m_LShr(m_Value(), m_Value())))
3486  std::swap(XShift, YShift);
3487 
3488  // The shifts must be in opposite directions.
3489  auto XShiftOpcode = XShift->getOpcode();
3490  if (XShiftOpcode == YShift->getOpcode())
3491  return nullptr; // Do not care about same-direction shifts here.
3492 
3493  Value *X, *XShAmt, *Y, *YShAmt;
3494  match(XShift, m_BinOp(m_Value(X), m_ZExtOrSelf(m_Value(XShAmt))));
3495  match(YShift, m_BinOp(m_Value(Y), m_ZExtOrSelf(m_Value(YShAmt))));
3496 
3497  // If one of the values being shifted is a constant, then we will end with
3498  // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
3499  // however, we will need to ensure that we won't increase instruction count.
3500  if (!isa<Constant>(X) && !isa<Constant>(Y)) {
3501  // At least one of the hands of the 'and' should be one-use shift.
3502  if (!match(I.getOperand(0),
3503  m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
3504  return nullptr;
3505  if (HadTrunc) {
3506  // Due to the 'trunc', we will need to widen X. For that either the old
3507  // 'trunc' or the shift amt in the non-truncated shift should be one-use.
3508  if (!MaybeTruncation->hasOneUse() &&
3509  !NarrowestShift->getOperand(1)->hasOneUse())
3510  return nullptr;
3511  }
3512  }
3513 
3514  // We have two shift amounts from two different shifts. The types of those
3515  // shift amounts may not match. If that's the case let's bailout now.
3516  if (XShAmt->getType() != YShAmt->getType())
3517  return nullptr;
3518 
3519  // Can we fold (XShAmt+YShAmt) ?
3520  auto *NewShAmt = dyn_cast_or_null<Constant>(
3521  SimplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,
3522  /*isNUW=*/false, SQ.getWithInstruction(&I)));
3523  if (!NewShAmt)
3524  return nullptr;
3525  NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, WidestTy);
3526  unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
3527 
3528  // Is the new shift amount smaller than the bit width?
3529  // FIXME: could also rely on ConstantRange.
3530  if (!match(NewShAmt,
3531  m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
3532  APInt(WidestBitWidth, WidestBitWidth))))
3533  return nullptr;
3534 
3535  // An extra legality check is needed if we had trunc-of-lshr.
3536  if (HadTrunc && match(WidestShift, m_LShr(m_Value(), m_Value()))) {
3537  auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
3538  WidestShift]() {
3539  // It isn't obvious whether it's worth it to analyze non-constants here.
3540  // Also, let's basically give up on non-splat cases, pessimizing vectors.
3541  // If *any* of these preconditions matches we can perform the fold.
3542  Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
3543  ? NewShAmt->getSplatValue()
3544  : NewShAmt;
3545  // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
3546  if (NewShAmtSplat &&
3547  (NewShAmtSplat->isNullValue() ||
3548  NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
3549  return true;
3550  // We consider *min* leading zeros so a single outlier
3551  // blocks the transform as opposed to allowing it.
3552  if (auto *C = dyn_cast<Constant>(NarrowestShift->getOperand(0))) {
3553  KnownBits Known = computeKnownBits(C, SQ.DL);
3554  unsigned MinLeadZero = Known.countMinLeadingZeros();
3555  // If the value being shifted has at most lowest bit set we can fold.
3556  unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3557  if (MaxActiveBits <= 1)
3558  return true;
3559  // Precondition: NewShAmt u<= countLeadingZeros(C)
3560  if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(MinLeadZero))
3561  return true;
3562  }
3563  if (auto *C = dyn_cast<Constant>(WidestShift->getOperand(0))) {
3564  KnownBits Known = computeKnownBits(C, SQ.DL);
3565  unsigned MinLeadZero = Known.countMinLeadingZeros();
3566  // If the value being shifted has at most lowest bit set we can fold.
3567  unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3568  if (MaxActiveBits <= 1)
3569  return true;
3570  // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
3571  if (NewShAmtSplat) {
3572  APInt AdjNewShAmt =
3573  (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
3574  if (AdjNewShAmt.ule(MinLeadZero))
3575  return true;
3576  }
3577  }
3578  return false; // Can't tell if it's ok.
3579  };
3580  if (!CanFold())
3581  return nullptr;
3582  }
3583 
3584  // All good, we can do this fold.
3585  X = Builder.CreateZExt(X, WidestTy);
3586  Y = Builder.CreateZExt(Y, WidestTy);
3587  // The shift is the same that was for X.
3588  Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
3589  ? Builder.CreateLShr(X, NewShAmt)
3590  : Builder.CreateShl(X, NewShAmt);
3591  Value *T1 = Builder.CreateAnd(T0, Y);
3592  return Builder.CreateICmp(I.getPredicate(), T1,
3593  Constant::getNullValue(WidestTy));
3594 }
3595 
3596 /// Fold
3597 /// (-1 u/ x) u< y
3598 /// ((x * y) u/ x) != y
3599 /// to
3600 /// @llvm.umul.with.overflow(x, y) plus extraction of overflow bit
3601 /// Note that the comparison is commutative, while inverted (u>=, ==) predicate
3602 /// will mean that we are looking for the opposite answer.
3603 Value *InstCombiner::foldUnsignedMultiplicationOverflowCheck(ICmpInst &I) {
3604  ICmpInst::Predicate Pred;
3605  Value *X, *Y;
3606  Instruction *Mul;
3607  bool NeedNegation;
3608  // Look for: (-1 u/ x) u</u>= y
3609  if (!I.isEquality() &&
3610  match(&I, m_c_ICmp(Pred, m_OneUse(m_UDiv(m_AllOnes(), m_Value(X))),
3611  m_Value(Y)))) {
3612  Mul = nullptr;
3613  // Canonicalize as-if y was on RHS.
3614  if (I.getOperand(1) != Y)
3615  Pred = I.getSwappedPredicate();
3616 
3617  // Are we checking that overflow does not happen, or does happen?
3618  switch (Pred) {
3619  case ICmpInst::Predicate::ICMP_ULT:
3620  NeedNegation = false;
3621  break; // OK
3622  case ICmpInst::Predicate::ICMP_UGE:
3623  NeedNegation = true;
3624  break; // OK
3625  default:
3626  return nullptr; // Wrong predicate.
3627  }
3628  } else // Look for: ((x * y) u/ x) !=/== y
3629  if (I.isEquality() &&
3630  match(&I, m_c_ICmp(Pred, m_Value(Y),
3632  m_Value(X)),
3633  m_Instruction(Mul)),
3634  m_Deferred(X)))))) {
3635  NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
3636  } else
3637  return nullptr;
3638 
3639  BuilderTy::InsertPointGuard Guard(Builder);
3640  // If the pattern included (x * y), we'll want to insert new instructions
3641  // right before that original multiplication so that we can replace it.
3642  bool MulHadOtherUses = Mul && !Mul->hasOneUse();
3643  if (MulHadOtherUses)
3644  Builder.SetInsertPoint(Mul);
3645 
3647  I.getModule(), Intrinsic::umul_with_overflow, X->getType());
3648  CallInst *Call = Builder.CreateCall(F, {X, Y}, "umul");
3649 
3650  // If the multiplication was used elsewhere, to ensure that we don't leave
3651  // "duplicate" instructions, replace uses of that original multiplication
3652  // with the multiplication result from the with.overflow intrinsic.
3653  if (MulHadOtherUses)
3654  replaceInstUsesWith(*Mul, Builder.CreateExtractValue(Call, 0, "umul.val"));
3655 
3656  Value *Res = Builder.CreateExtractValue(Call, 1, "umul.ov");
3657  if (NeedNegation) // This technically increases instruction count.
3658  Res = Builder.CreateNot(Res, "umul.not.ov");
3659 
3660  return Res;
3661 }
3662 
3663 /// Try to fold icmp (binop), X or icmp X, (binop).
3664 /// TODO: A large part of this logic is duplicated in InstSimplify's
3665 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3666 /// duplication.
3667 Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I, const SimplifyQuery &SQ) {
3668  const SimplifyQuery Q = SQ.getWithInstruction(&I);
3669  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3670 
3671  // Special logic for binary operators.
3672  BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3673  BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3674  if (!BO0 && !BO1)
3675  return nullptr;
3676 
3677  const CmpInst::Predicate Pred = I.getPredicate();
3678  Value *X;
3679 
3680  // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3681  // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
3682  if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3683  (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3684  return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3685  // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
3686  if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
3687  (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3688  return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3689 
3690  bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3691  if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3692  NoOp0WrapProblem =
3693  ICmpInst::isEquality(Pred) ||
3694  (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3695  (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3696  if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3697  NoOp1WrapProblem =
3698  ICmpInst::isEquality(Pred) ||
3699  (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3700  (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3701 
3702  // Analyze the case when either Op0 or Op1 is an add instruction.
3703  // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3704  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3705  if (BO0 && BO0->getOpcode() == Instruction::Add) {
3706  A = BO0->getOperand(0);
3707  B = BO0->getOperand(1);
3708  }
3709  if (BO1 && BO1->getOpcode() == Instruction::Add) {
3710  C = BO1->getOperand(0);
3711  D = BO1->getOperand(1);
3712  }
3713 
3714  // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
3715  // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
3716  if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3717  return new ICmpInst(Pred, A == Op1 ? B : A,
3718  Constant::getNullValue(Op1->getType()));
3719 
3720  // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
3721  // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
3722  if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3723  return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3724  C == Op0 ? D : C);
3725 
3726  // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
3727  if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3728  NoOp1WrapProblem) {
3729  // Determine Y and Z in the form icmp (X+Y), (X+Z).
3730  Value *Y, *Z;
3731  if (A == C) {
3732  // C + B == C + D -> B == D
3733  Y = B;
3734  Z = D;
3735  } else if (A == D) {
3736  // D + B == C + D -> B == C
3737  Y = B;
3738  Z = C;
3739  } else if (B == C) {
3740  // A + C == C + D -> A == D
3741  Y = A;
3742  Z = D;
3743  } else {
3744  assert(B == D);
3745  // A + D == C + D -> A == C
3746  Y = A;
3747  Z = C;
3748  }
3749  return new ICmpInst(Pred, Y, Z);
3750  }
3751 
3752  // icmp slt (A + -1), Op1 -> icmp sle A, Op1
3753  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3754  match(B, m_AllOnes()))
3755  return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3756 
3757  // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
3758  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3759  match(B, m_AllOnes()))
3760  return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3761 
3762  // icmp sle (A + 1), Op1 -> icmp slt A, Op1
3763  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
3764  return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3765 
3766  // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
3767  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
3768  return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3769 
3770  // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
3771  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
3772  match(D, m_AllOnes()))
3773  return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3774 
3775  // icmp sle Op0, (C + -1) -> icmp slt Op0, C
3776  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
3777  match(D, m_AllOnes()))
3778  return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3779 
3780  // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
3781  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
3782  return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3783 
3784  // icmp slt Op0, (C + 1) -> icmp sle Op0, C
3785  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
3786  return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3787 
3788  // TODO: The subtraction-related identities shown below also hold, but
3789  // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3790  // wouldn't happen even if they were implemented.
3791  //
3792  // icmp ult (A - 1), Op1 -> icmp ule A, Op1
3793  // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
3794  // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
3795  // icmp ule Op0, (C - 1) -> icmp ult Op0, C
3796 
3797  // icmp ule (A + 1), Op0 -> icmp ult A, Op1
3798  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
3799  return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3800 
3801  // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
3802  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
3803  return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3804 
3805  // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
3806  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
3807  return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3808 
3809  // icmp ult Op0, (C + 1) -> icmp ule Op0, C
3810  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
3811  return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3812 
3813  // if C1 has greater magnitude than C2:
3814  // icmp (A + C1), (C + C2) -> icmp (A + C3), C
3815  // s.t. C3 = C1 - C2
3816  //
3817  // if C2 has greater magnitude than C1:
3818  // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
3819  // s.t. C3 = C2 - C1
3820  if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3821  (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3822  if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3823  if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3824  const APInt &AP1 = C1->getValue();
3825  const APInt &AP2 = C2->getValue();
3826  if (AP1.isNegative() == AP2.isNegative()) {
3827  APInt AP1Abs = C1->getValue().abs();
3828  APInt AP2Abs = C2->getValue().abs();
3829  if (AP1Abs.uge(AP2Abs)) {
3830  ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3831  Value *NewAdd = Builder.CreateNSWAdd(A, C3);
3832  return new ICmpInst(Pred, NewAdd, C);
3833  } else {
3834  ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3835  Value *NewAdd = Builder.CreateNSWAdd(C, C3);
3836  return new ICmpInst(Pred, A, NewAdd);
3837  }
3838  }
3839  }
3840 
3841  // Analyze the case when either Op0 or Op1 is a sub instruction.
3842  // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3843  A = nullptr;
3844  B = nullptr;
3845  C = nullptr;
3846  D = nullptr;
3847  if (BO0 && BO0->getOpcode() == Instruction::Sub) {
3848  A = BO0->getOperand(0);
3849  B = BO0->getOperand(1);
3850  }
3851  if (BO1 && BO1->getOpcode() == Instruction::Sub) {
3852  C = BO1->getOperand(0);
3853  D = BO1->getOperand(1);
3854  }
3855 
3856  // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
3857  if (A == Op1 && NoOp0WrapProblem)
3858  return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3859  // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
3860  if (C == Op0 && NoOp1WrapProblem)
3861  return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3862 
3863  // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
3864  // (A - B) u>/u<= A --> B u>/u<= A
3865  if (A == Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3866  return new ICmpInst(Pred, B, A);
3867  // C u</u>= (C - D) --> C u</u>= D
3868  if (C == Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3869  return new ICmpInst(Pred, C, D);
3870  // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
3871  if (A == Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
3872  isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
3873  return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), B, A);
3874  // C u<=/u> (C - D) --> C u</u>= D iff B != 0
3875  if (C == Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
3876  isKnownNonZero(D, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
3877  return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), C, D);
3878 
3879  // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
3880  if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
3881  return new ICmpInst(Pred, A, C);
3882 
3883  // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
3884  if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
3885  return new ICmpInst(Pred, D, B);
3886 
3887  // icmp (0-X) < cst --> x > -cst
3888  if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3889  Value *X;
3890  if (match(BO0, m_Neg(m_Value(X))))
3891  if (Constant *RHSC = dyn_cast<Constant>(Op1))
3892  if (RHSC->isNotMinSignedValue())
3893  return new ICmpInst(I.getSwappedPredicate(), X,
3894  ConstantExpr::getNeg(RHSC));
3895  }
3896 
3897  BinaryOperator *SRem = nullptr;
3898  // icmp (srem X, Y), Y
3899  if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
3900  SRem = BO0;
3901  // icmp Y, (srem X, Y)
3902  else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3903  Op0 == BO1->getOperand(1))
3904  SRem = BO1;
3905  if (SRem) {
3906  // We don't check hasOneUse to avoid increasing register pressure because
3907  // the value we use is the same value this instruction was already using.
3908  switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3909  default:
3910  break;
3911  case ICmpInst::ICMP_EQ:
3912  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3913  case ICmpInst::ICMP_NE:
3914  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3915  case ICmpInst::ICMP_SGT:
3916  case ICmpInst::ICMP_SGE:
3917  return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3919  case ICmpInst::ICMP_SLT:
3920  case ICmpInst::ICMP_SLE:
3921  return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3922  Constant::getNullValue(SRem->getType()));
3923  }
3924  }
3925 
3926  if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
3927  BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
3928  switch (BO0->getOpcode()) {
3929  default:
3930  break;
3931  case Instruction::Add:
3932  case Instruction::Sub:
3933  case Instruction::Xor: {
3934  if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3935  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3936 
3937  const APInt *C;
3938  if (match(BO0->getOperand(1), m_APInt(C))) {
3939  // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
3940  if (C->isSignMask()) {
3941  ICmpInst::Predicate NewPred =
3943  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3944  }
3945 
3946  // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
3947  if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
3948  ICmpInst::Predicate NewPred =
3950  NewPred = I.getSwappedPredicate(NewPred);
3951  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3952  }
3953  }
3954  break;
3955  }
3956  case Instruction::Mul: {
3957  if (!I.isEquality())
3958  break;
3959 
3960  const APInt *C;
3961  if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
3962  !C->isOneValue()) {
3963  // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
3964  // Mask = -1 >> count-trailing-zeros(C).
3965  if (unsigned TZs = C->countTrailingZeros()) {
3967  BO0->getType(),
3968  APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
3969  Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
3970  Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
3971  return new ICmpInst(Pred, And1, And2);
3972  }
3973  // If there are no trailing zeros in the multiplier, just eliminate
3974  // the multiplies (no masking is needed):
3975  // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
3976  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3977  }
3978  break;
3979  }
3980  case Instruction::UDiv:
3981  case Instruction::LShr:
3982  if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
3983  break;
3984  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3985 
3986  case Instruction::SDiv:
3987  if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
3988  break;
3989  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3990 
3991  case Instruction::AShr:
3992  if (!BO0->isExact() || !BO1->isExact())
3993  break;
3994  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3995 
3996  case Instruction::Shl: {
3997  bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3998  bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3999  if (!NUW && !NSW)
4000  break;
4001  if (!NSW && I.isSigned())
4002  break;
4003  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4004  }
4005  }
4006  }
4007 
4008  if (BO0) {
4009  // Transform A & (L - 1) `ult` L --> L != 0
4010  auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
4011  auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
4012 
4013  if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
4014  auto *Zero = Constant::getNullValue(BO0->getType());
4015  return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
4016  }
4017  }
4018 
4019  if (Value *V = foldUnsignedMultiplicationOverflowCheck(I))
4020  return replaceInstUsesWith(I, V);
4021 
4022  if (Value *V = foldICmpWithLowBitMaskedVal(I, Builder))
4023  return replaceInstUsesWith(I, V);
4024 
4025  if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
4026  return replaceInstUsesWith(I, V);
4027 
4028  if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder))
4029  return replaceInstUsesWith(I, V);
4030 
4031  return nullptr;
4032 }
4033 
4034 /// Fold icmp Pred min|max(X, Y), X.
4036  ICmpInst::Predicate Pred = Cmp.getPredicate();
4037  Value *Op0 = Cmp.getOperand(0);
4038  Value *X = Cmp.getOperand(1);
4039 
4040  // Canonicalize minimum or maximum operand to LHS of the icmp.
4041  if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
4042  match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
4043  match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
4044  match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
4045  std::swap(Op0, X);
4046  Pred = Cmp.getSwappedPredicate();
4047  }
4048 
4049  Value *Y;
4050  if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
4051  // smin(X, Y) == X --> X s<= Y
4052  // smin(X, Y) s>= X --> X s<= Y
4053  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
4054  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
4055 
4056  // smin(X, Y) != X --> X s> Y
4057  // smin(X, Y) s< X --> X s> Y
4058  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
4059  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
4060 
4061  // These cases should be handled in InstSimplify:
4062  // smin(X, Y) s<= X --> true
4063  // smin(X, Y) s> X --> false
4064  return nullptr;
4065  }
4066 
4067  if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
4068  // smax(X, Y) == X --> X s>= Y
4069  // smax(X, Y) s<= X --> X s>= Y
4070  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
4071  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
4072 
4073  // smax(X, Y) != X --> X s< Y
4074  // smax(X, Y) s> X --> X s< Y
4075  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
4076  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
4077 
4078  // These cases should be handled in InstSimplify:
4079  // smax(X, Y) s>= X --> true
4080  // smax(X, Y) s< X --> false
4081  return nullptr;
4082  }
4083 
4084  if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
4085  // umin(X, Y) == X --> X u<= Y
4086  // umin(X, Y) u>= X --> X u<= Y
4087  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
4088  return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
4089 
4090  // umin(X, Y) != X --> X u> Y
4091  // umin(X, Y) u< X --> X u> Y
4092  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
4093  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
4094 
4095  // These cases should be handled in InstSimplify:
4096  // umin(X, Y) u<= X --> true
4097  // umin(X, Y) u> X --> false
4098  return nullptr;
4099  }
4100 
4101  if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
4102  // umax(X, Y) == X --> X u>= Y
4103  // umax(X, Y) u<= X --> X u>= Y
4104  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
4105  return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
4106 
4107  // umax(X, Y) != X --> X u< Y
4108  // umax(X, Y) u> X --> X u< Y
4109  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
4110  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
4111 
4112  // These cases should be handled in InstSimplify:
4113  // umax(X, Y) u>= X --> true
4114  // umax(X, Y) u< X --> false
4115  return nullptr;
4116  }
4117 
4118  return nullptr;
4119 }
4120 
4121 Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
4122  if (!I.isEquality())
4123  return nullptr;
4124 
4125  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4126  const CmpInst::Predicate Pred = I.getPredicate();
4127  Value *A, *B, *C, *D;
4128  if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4129  if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
4130  Value *OtherVal = A == Op1 ? B : A;
4131  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4132  }
4133 
4134  if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
4135  // A^c1 == C^c2 --> A == C^(c1^c2)
4136  ConstantInt *C1, *C2;
4137  if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
4138  Op1->hasOneUse()) {
4139  Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
4140  Value *Xor = Builder.CreateXor(C, NC);
4141  return new ICmpInst(Pred, A, Xor);
4142  }
4143 
4144  // A^B == A^D -> B == D
4145  if (A == C)
4146  return new ICmpInst(Pred, B, D);
4147  if (A == D)
4148  return new ICmpInst(Pred, B, C);
4149  if (B == C)
4150  return new ICmpInst(Pred, A, D);
4151  if (B == D)
4152  return new ICmpInst(Pred, A, C);
4153  }
4154  }
4155 
4156  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
4157  // A == (A^B) -> B == 0
4158  Value *OtherVal = A == Op0 ? B : A;
4159  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4160  }
4161 
4162  // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4163  if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
4164  match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
4165  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
4166 
4167  if (A == C) {
4168  X = B;
4169  Y = D;
4170  Z = A;
4171  } else if (A == D) {
4172  X = B;
4173  Y = C;
4174  Z = A;
4175  } else if (B == C) {
4176  X = A;
4177  Y = D;
4178  Z = B;
4179  } else if (B == D) {
4180  X = A;
4181  Y = C;
4182  Z = B;
4183  }
4184 
4185  if (X) { // Build (X^Y) & Z
4186  Op1 = Builder.CreateXor(X, Y);
4187  Op1 = Builder.CreateAnd(Op1, Z);
4188  I.setOperand(0, Op1);
4189  I.setOperand(1, Constant::getNullValue(Op1->getType()));
4190  return &I;
4191  }
4192  }
4193 
4194  // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
4195  // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
4196  ConstantInt *Cst1;
4197  if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
4198  match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
4199  (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
4200  match(Op1, m_ZExt(m_Value(A))))) {
4201  APInt Pow2 = Cst1->getValue() + 1;
4202  if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
4203  Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
4204  return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
4205  }
4206 
4207  // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
4208  // For lshr and ashr pairs.
4209  if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4210  match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
4211  (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4212  match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
4213  unsigned TypeBits = Cst1->getBitWidth();
4214  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4215  if (ShAmt < TypeBits && ShAmt != 0) {
4216  ICmpInst::Predicate NewPred =
4218  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4219  APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
4220  return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
4221  }
4222  }
4223 
4224  // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
4225  if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
4226  match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
4227  unsigned TypeBits = Cst1->getBitWidth();
4228  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4229  if (ShAmt < TypeBits && ShAmt != 0) {
4230  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4231  APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
4232  Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
4233  I.getName() + ".mask");
4234  return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
4235  }
4236  }
4237 
4238  // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
4239  // "icmp (and X, mask), cst"
4240  uint64_t ShAmt = 0;
4241  if (Op0->hasOneUse() &&
4242  match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
4243  match(Op1, m_ConstantInt(Cst1)) &&
4244  // Only do this when A has multiple uses. This is most important to do
4245  // when it exposes other optimizations.
4246  !A->hasOneUse()) {
4247  unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
4248 
4249  if (ShAmt < ASize) {
4250  APInt MaskV =
4252  MaskV <<= ShAmt;
4253 
4254  APInt CmpV = Cst1->getValue().zext(ASize);
4255  CmpV <<= ShAmt;
4256 
4257  Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
4258  return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
4259  }
4260  }
4261 
4262  // If both operands are byte-swapped or bit-reversed, just compare the
4263  // original values.
4264  // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
4265  // and handle more intrinsics.
4266  if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
4267  (match(Op0, m_BitReverse(m_Value(A))) &&
4268  match(Op1, m_BitReverse(m_Value(B)))))
4269  return new ICmpInst(Pred, A, B);
4270 
4271  // Canonicalize checking for a power-of-2-or-zero value:
4272  // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
4273  // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
4274  if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
4275  m_Deferred(A)))) ||
4276  !match(Op1, m_ZeroInt()))
4277  A = nullptr;
4278 
4279  // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
4280  // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
4281  if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
4282  A = Op1;
4283  else if (match(Op1,
4284  m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
4285  A = Op0;
4286 
4287  if (A) {
4288  Type *Ty = A->getType();
4289  CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
4290  return Pred == ICmpInst::ICMP_EQ
4291  ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, ConstantInt::get(Ty, 2))
4292  : new ICmpInst(ICmpInst::ICMP_UGT, CtPop, ConstantInt::get(Ty, 1));
4293  }
4294 
4295  return nullptr;
4296 }
4297 
4299  InstCombiner::BuilderTy &Builder) {
4300  assert(isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0");
4301  auto *CastOp0 = cast<CastInst>(ICmp.getOperand(0));
4302  Value *X;
4303  if (!match(CastOp0, m_ZExtOrSExt(m_Value(X))))
4304  return nullptr;
4305 
4306  bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt;
4307  bool IsSignedCmp = ICmp.isSigned();
4308  if (auto *CastOp1 = dyn_cast<CastInst>(ICmp.getOperand(1))) {
4309  // If the signedness of the two casts doesn't agree (i.e. one is a sext
4310  // and the other is a zext), then we can't handle this.
4311  // TODO: This is too strict. We can handle some predicates (equality?).
4312  if (CastOp0->getOpcode() != CastOp1->getOpcode())
4313  return nullptr;
4314 
4315  // Not an extension from the same type?
4316  Value *Y = CastOp1->getOperand(0);
4317  Type *XTy = X->getType(), *YTy = Y->getType();
4318  if (XTy != YTy) {
4319  // One of the casts must have one use because we are creating a new cast.
4320  if (!CastOp0->hasOneUse() && !CastOp1->hasOneUse())
4321  return nullptr;
4322  // Extend the narrower operand to the type of the wider operand.
4323  if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits())
4324  X = Builder.CreateCast(CastOp0->getOpcode(), X, YTy);
4325  else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits())
4326  Y = Builder.CreateCast(CastOp0->getOpcode(), Y, XTy);
4327  else
4328  return nullptr;
4329  }
4330 
4331  // (zext X) == (zext Y) --> X == Y
4332  // (sext X) == (sext Y) --> X == Y
4333  if (ICmp.isEquality())
4334  return new ICmpInst(ICmp.getPredicate(), X, Y);
4335 
4336  // A signed comparison of sign extended values simplifies into a
4337  // signed comparison.
4338  if (IsSignedCmp && IsSignedExt)
4339  return new ICmpInst(ICmp.getPredicate(), X, Y);
4340 
4341  // The other three cases all fold into an unsigned comparison.
4342  return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y);
4343  }
4344 
4345  // Below here, we are only folding a compare with constant.
4346  auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
4347  if (!C)
4348  return nullptr;
4349 
4350  // Compute the constant that would happen if we truncated to SrcTy then
4351  // re-extended to DestTy.
4352  Type *SrcTy = CastOp0->getSrcTy();
4353  Type *DestTy = CastOp0->getDestTy();
4354  Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
4355  Constant *Res2 = ConstantExpr::getCast(CastOp0->getOpcode(), Res1, DestTy);
4356 
4357  // If the re-extended constant didn't change...
4358  if (Res2 == C) {
4359  if (ICmp.isEquality())
4360  return new ICmpInst(ICmp.getPredicate(), X, Res1);
4361 
4362  // A signed comparison of sign extended values simplifies into a
4363  // signed comparison.
4364  if (IsSignedExt && IsSignedCmp)
4365  return new ICmpInst(ICmp.getPredicate(), X, Res1);
4366 
4367  // The other three cases all fold into an unsigned comparison.
4368  return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res1);
4369  }
4370 
4371  // The re-extended constant changed, partly changed (in the case of a vector),
4372  // or could not be determined to be equal (in the case of a constant
4373  // expression), so the constant cannot be represented in the shorter type.
4374  // All the cases that fold to true or false will have already been handled
4375  // by SimplifyICmpInst, so only deal with the tricky case.
4376  if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(C))
4377  return nullptr;
4378 
4379  // Is source op positive?
4380  // icmp ult (sext X), C --> icmp sgt X, -1
4381  if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
4382  return new ICmpInst(CmpInst::ICMP_SGT, X, Constant::getAllOnesValue(SrcTy));
4383 
4384  // Is source op negative?
4385  // icmp ugt (sext X), C --> icmp slt X, 0
4386  assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
4387  return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(SrcTy));
4388 }
4389 
4390 /// Handle icmp (cast x), (cast or constant).
4391 Instruction *InstCombiner::foldICmpWithCastOp(ICmpInst &ICmp) {
4392  auto *CastOp0 = dyn_cast<CastInst>(ICmp.getOperand(0));
4393  if (!CastOp0)
4394  return nullptr;
4395  if (!isa<Constant>(ICmp.getOperand(1)) && !isa<CastInst>(ICmp.getOperand(1)))
4396  return nullptr;
4397 
4398  Value *Op0Src = CastOp0->getOperand(0);
4399  Type *SrcTy = CastOp0->getSrcTy();
4400  Type *DestTy = CastOp0->getDestTy();
4401 
4402  // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
4403  // integer type is the same size as the pointer type.
4404  auto CompatibleSizes = [&](Type *SrcTy, Type *DestTy) {
4405  if (isa<VectorType>(SrcTy)) {
4406  SrcTy = cast<VectorType>(SrcTy)->getElementType();
4407  DestTy = cast<VectorType>(DestTy)->getElementType();
4408  }
4409  return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
4410  };
4411  if (CastOp0->getOpcode() == Instruction::PtrToInt &&
4412  CompatibleSizes(SrcTy, DestTy)) {
4413  Value *NewOp1 = nullptr;
4414  if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
4415  Value *PtrSrc = PtrToIntOp1->getOperand(0);
4416  if (PtrSrc->getType()->getPointerAddressSpace() ==
4417  Op0Src->getType()->getPointerAddressSpace()) {
4418  NewOp1 = PtrToIntOp1->getOperand(0);
4419  // If the pointer types don't match, insert a bitcast.
4420  if (Op0Src->getType() != NewOp1->getType())
4421  NewOp1 = Builder.CreateBitCast(NewOp1, Op0Src->getType());
4422  }
4423  } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
4424  NewOp1 = ConstantExpr::getIntToPtr(RHSC, SrcTy);
4425  }
4426 
4427  if (NewOp1)
4428  return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
4429  }
4430 
4431  return foldICmpWithZextOrSext(ICmp, Builder);
4432 }
4433 
4434 static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS) {
4435  switch (BinaryOp) {
4436  default:
4437  llvm_unreachable("Unsupported binary op");
4438  case Instruction::Add:
4439  case Instruction::Sub:
4440  return match(RHS, m_Zero());
4441  case Instruction::Mul:
4442  return match(RHS, m_One());
4443  }
4444 }
4445 
4447  Instruction::BinaryOps BinaryOp, bool IsSigned,
4448  Value *LHS, Value *RHS, Instruction *CxtI) const {
4449  switch (BinaryOp) {
4450  default:
4451  llvm_unreachable("Unsupported binary op");
4452  case Instruction::Add:
4453  if (IsSigned)
4454  return computeOverflowForSignedAdd(LHS, RHS, CxtI);
4455  else
4456  return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
4457  case Instruction::Sub:
4458  if (IsSigned)
4459  return computeOverflowForSignedSub(LHS, RHS, CxtI);
4460  else
4461  return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
4462  case Instruction::Mul:
4463  if (IsSigned)
4464  return computeOverflowForSignedMul(LHS, RHS, CxtI);
4465  else
4466  return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
4467  }
4468 }
4469 
4470 bool InstCombiner::OptimizeOverflowCheck(
4471  Instruction::BinaryOps BinaryOp, bool IsSigned, Value *LHS, Value *RHS,
4472  Instruction &OrigI, Value *&Result, Constant *&Overflow) {
4473  if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
4474  std::swap(LHS, RHS);
4475 
4476  // If the overflow check was an add followed by a compare, the insertion point
4477  // may be pointing to the compare. We want to insert the new instructions
4478  // before the add in case there are uses of the add between the add and the
4479  // compare.
4480  Builder.SetInsertPoint(&OrigI);
4481 
4482  if (isNeutralValue(BinaryOp, RHS)) {
4483  Result = LHS;
4484  Overflow = Builder.getFalse();
4485  return true;
4486  }
4487 
4488  switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, &OrigI)) {
4490  return false;
4493  Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4494  Result->takeName(&OrigI);
4495  Overflow = Builder.getTrue();
4496  return true;
4498  Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4499  Result->takeName(&OrigI);
4500  Overflow = Builder.getFalse();
4501  if (auto *Inst = dyn_cast<Instruction>(Result)) {
4502  if (IsSigned)
4503  Inst->setHasNoSignedWrap();
4504  else
4505  Inst->setHasNoUnsignedWrap();
4506  }
4507  return true;
4508  }
4509 
4510  llvm_unreachable("Unexpected overflow result");
4511 }
4512 
4513 /// Recognize and process idiom involving test for multiplication
4514 /// overflow.
4515 ///
4516 /// The caller has matched a pattern of the form:
4517 /// I = cmp u (mul(zext A, zext B), V
4518 /// The function checks if this is a test for overflow and if so replaces
4519 /// multiplication with call to 'mul.with.overflow' intrinsic.
4520 ///
4521 /// \param I Compare instruction.
4522 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
4523 /// the compare instruction. Must be of integer type.
4524 /// \param OtherVal The other argument of compare instruction.
4525 /// \returns Instruction which must replace the compare instruction, NULL if no
4526 /// replacement required.
4528  Value *OtherVal, InstCombiner &IC) {
4529  // Don't bother doing this transformation for pointers, don't do it for
4530  // vectors.
4531  if (!isa<IntegerType>(MulVal->getType()))
4532  return nullptr;
4533 
4534  assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
4535  assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
4536  auto *MulInstr = dyn_cast<Instruction>(MulVal);
4537  if (!MulInstr)
4538  return nullptr;
4539  assert(MulInstr->getOpcode() == Instruction::Mul);
4540 
4541  auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
4542  *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
4543  assert(LHS->getOpcode() == Instruction::ZExt);
4544  assert(RHS->getOpcode() == Instruction::ZExt);
4545  Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
4546 
4547  // Calculate type and width of the result produced by mul.with.overflow.
4548  Type *TyA = A->getType(), *TyB = B->getType();
4549  unsigned WidthA = TyA->getPrimitiveSizeInBits(),
4550  WidthB = TyB->getPrimitiveSizeInBits();
4551  unsigned MulWidth;
4552  Type *MulType;
4553  if (WidthB > WidthA) {
4554  MulWidth = WidthB;
4555  MulType = TyB;
4556  } else {
4557  MulWidth = WidthA;
4558  MulType = TyA;
4559  }
4560 
4561  // In order to replace the original mul with a narrower mul.with.overflow,
4562  // all uses must ignore upper bits of the product. The number of used low
4563  // bits must be not greater than the width of mul.with.overflow.
4564  if (MulVal->hasNUsesOrMore(2))
4565  for (User *U : MulVal->users()) {
4566  if (U == &I)
4567  continue;
4568  if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4569  // Check if truncation ignores bits above MulWidth.
4570  unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
4571  if (TruncWidth > MulWidth)
4572  return nullptr;
4573  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4574  // Check if AND ignores bits above MulWidth.
4575  if (BO->getOpcode() != Instruction::And)
4576  return nullptr;
4577  if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4578  const APInt &CVal = CI->getValue();
4579  if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
4580  return nullptr;
4581  } else {
4582  // In this case we could have the operand of the binary operation
4583  // being defined in another block, and performing the replacement
4584  // could break the dominance relation.
4585  return nullptr;
4586  }
4587  } else {
4588  // Other uses prohibit this transformation.
4589  return nullptr;
4590  }
4591  }
4592 
4593  // Recognize patterns
4594  switch (I.getPredicate()) {
4595  case ICmpInst::ICMP_EQ:
4596  case ICmpInst::ICMP_NE:
4597  // Recognize pattern:
4598  // mulval = mul(zext A, zext B)
4599  // cmp eq/neq mulval, zext trunc mulval
4600  if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
4601  if (Zext->hasOneUse()) {
4602  Value *ZextArg = Zext->getOperand(0);
4603  if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
4604  if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
4605  break; //Recognized
4606  }
4607 
4608  // Recognize pattern:
4609  // mulval = mul(zext A, zext B)
4610  // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
4611  ConstantInt *CI;
4612  Value *ValToMask;
4613  if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
4614  if (ValToMask != MulVal)
4615  return nullptr;
4616  const APInt &CVal = CI->getValue() + 1;
4617  if (CVal.isPowerOf2()) {
4618  unsigned MaskWidth = CVal.logBase2();
4619  if (MaskWidth == MulWidth)
4620  break; // Recognized
4621  }
4622  }
4623  return nullptr;
4624 
4625  case ICmpInst::ICMP_UGT:
4626  // Recognize pattern:
4627  // mulval = mul(zext A, zext B)
4628  // cmp ugt mulval, max
4629  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4630  APInt MaxVal = APInt::getMaxValue(MulWidth);
4631  MaxVal = MaxVal.zext(CI->getBitWidth());
4632  if (MaxVal.eq(CI->getValue()))
4633  break; // Recognized
4634  }
4635  return nullptr;
4636 
4637  case ICmpInst::ICMP_UGE:
4638  // Recognize pattern:
4639  // mulval = mul(zext A, zext B)
4640  // cmp uge mulval, max+1
4641  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4642  APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4643  if (MaxVal.eq(CI->getValue()))
4644  break; // Recognized
4645  }
4646  return nullptr;
4647 
4648  case ICmpInst::ICMP_ULE:
4649  // Recognize pattern:
4650  // mulval = mul(zext A, zext B)
4651  // cmp ule mulval, max
4652  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4653  APInt MaxVal = APInt::getMaxValue(MulWidth);
4654  MaxVal = MaxVal.zext(CI->getBitWidth());
4655  if (MaxVal.eq(CI->getValue()))
4656  break; // Recognized
4657  }
4658  return nullptr;
4659 
4660  case ICmpInst::ICMP_ULT:
4661  // Recognize pattern:
4662  // mulval = mul(zext A, zext B)
4663  // cmp ule mulval, max + 1
4664  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4665  APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4666  if (MaxVal.eq(CI->getValue()))
4667  break; // Recognized
4668  }
4669  return nullptr;
4670 
4671  default:
4672  return nullptr;
4673  }
4674 
4675  InstCombiner::BuilderTy &Builder = IC.Builder;
4676  Builder.SetInsertPoint(MulInstr);
4677 
4678  // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4679  Value *MulA = A, *MulB = B;
4680  if (WidthA < MulWidth)
4681  MulA = Builder.CreateZExt(A, MulType);
4682  if (WidthB < MulWidth)
4683  MulB = Builder.CreateZExt(B, MulType);
4685  I.getModule(), Intrinsic::umul_with_overflow, MulType);
4686  CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4687  IC.Worklist.Add(MulInstr);
4688 
4689  // If there are uses of mul result other than the comparison, we know that
4690  // they are truncation or binary AND. Change them to use result of
4691  // mul.with.overflow and adjust properly mask/size.
4692  if (MulVal->hasNUsesOrMore(2)) {
4693  Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4694  for (auto UI = MulVal->user_begin(), UE = MulVal->user_end(); UI != UE;) {
4695  User *U = *UI++;
4696  if (U == &I || U == OtherVal)
4697  continue;
4698  if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4699  if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
4700  IC.replaceInstUsesWith(*TI, Mul);
4701  else
4702  TI->setOperand(0, Mul);
4703  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4704  assert(BO->getOpcode() == Instruction::And);
4705  // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
4706  ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
4707  APInt ShortMask = CI->getValue().trunc(MulWidth);
4708  Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
4709  Instruction *Zext =
4710  cast<Instruction>(Builder.CreateZExt(ShortAnd, BO->getType()));
4711  IC.Worklist.Add(Zext);
4712  IC.replaceInstUsesWith(*BO, Zext);
4713  } else {
4714  llvm_unreachable("Unexpected Binary operation");
4715  }
4716  IC.Worklist.Add(cast<Instruction>(U));
4717  }
4718  }
4719  if (isa<Instruction>(OtherVal))
4720  IC.Worklist.Add(cast<Instruction>(OtherVal));
4721 
4722  // The original icmp gets replaced with the overflow value, maybe inverted
4723  // depending on predicate.
4724  bool Inverse = false;
4725  switch (I.getPredicate()) {
4726  case ICmpInst::ICMP_NE:
4727  break;
4728  case ICmpInst::ICMP_EQ:
4729  Inverse = true;
4730  break;
4731  case ICmpInst::ICMP_UGT:
4732  case ICmpInst::ICMP_UGE:
4733  if (I.getOperand(0) == MulVal)
4734  break;
4735  Inverse = true;
4736  break;
4737  case ICmpInst::ICMP_ULT:
4738  case ICmpInst::ICMP_ULE:
4739  if (I.getOperand(1) == MulVal)
4740  break;
4741  Inverse = true;
4742  break;
4743  default:
4744  llvm_unreachable("Unexpected predicate");
4745  }
4746  if (Inverse) {
4747  Value *Res = Builder.CreateExtractValue(Call, 1);
4748  return BinaryOperator::CreateNot(Res);
4749  }
4750 
4751  return ExtractValueInst::Create(Call, 1);
4752 }
4753 
4754 /// When performing a comparison against a constant, it is possible that not all
4755 /// the bits in the LHS are demanded. This helper method computes the mask that
4756 /// IS demanded.
4757 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
4758  const APInt *RHS;
4759  if (!match(I.getOperand(1), m_APInt(RHS)))
4760  return APInt::getAllOnesValue(BitWidth);
4761 
4762  // If this is a normal comparison, it demands all bits. If it is a sign bit
4763  // comparison, it only demands the sign bit.
4764  bool UnusedBit;
4765  if (isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
4766  return APInt::getSignMask(BitWidth);
4767 
4768  switch (I.getPredicate()) {
4769  // For a UGT comparison, we don't care about any bits that
4770  // correspond to the trailing ones of the comparand. The value of these
4771  // bits doesn't impact the outcome of the comparison, because any value
4772  // greater than the RHS must differ in a bit higher than these due to carry.
4773  case ICmpInst::ICMP_UGT:
4774  return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
4775 
4776  // Similarly, for a ULT comparison, we don't care about the trailing zeros.
4777  // Any value less than the RHS must differ in a higher bit because of carries.
4778  case ICmpInst::ICMP_ULT:
4779  return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
4780 
4781  default:
4782  return APInt::getAllOnesValue(BitWidth);
4783  }
4784 }
4785 
4786 /// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
4787 /// should be swapped.
4788 /// The decision is based on how many times these two operands are reused
4789 /// as subtract operands and their positions in those instructions.
4790 /// The rationale is that several architectures use the same instruction for
4791 /// both subtract and cmp. Thus, it is better if the order of those operands
4792 /// match.
4793 /// \return true if Op0 and Op1 should be swapped.
4794 static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
4795  // Filter out pointer values as those cannot appear directly in subtract.
4796  // FIXME: we may want to go through inttoptrs or bitcasts.
4797  if (Op0->getType()->isPointerTy())
4798  return false;
4799  // If a subtract already has the same operands as a compare, swapping would be
4800  // bad. If a subtract has the same operands as a compare but in reverse order,
4801  // then swapping is good.
4802  int GoodToSwap = 0;
4803  for (const User *U : Op0->users()) {
4804  if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
4805  GoodToSwap++;
4806  else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
4807  GoodToSwap--;
4808  }
4809  return GoodToSwap > 0;
4810 }
4811 
4812 /// Check that one use is in the same block as the definition and all
4813 /// other uses are in blocks dominated by a given block.
4814 ///
4815 /// \param DI Definition
4816 /// \param UI Use
4817 /// \param DB Block that must dominate all uses of \p DI outside
4818 /// the parent block
4819 /// \return true when \p UI is the only use of \p DI in the parent block
4820 /// and all other uses of \p DI are in blocks dominated by \p DB.
4821 ///
4823  const Instruction *UI,
4824  const BasicBlock *DB) const {
4825  assert(DI && UI && "Instruction not defined\n");
4826  // Ignore incomplete definitions.
4827  if (!DI->getParent())
4828  return false;
4829  // DI and UI must be in the same block.
4830  if (DI->getParent() != UI->getParent())
4831  return false;
4832  // Protect from self-referencing blocks.
4833  if (DI->getParent() == DB)
4834  return false;
4835  for (const User *U : DI->users()) {
4836  auto *Usr = cast<Instruction>(U);
4837  if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4838  return false;
4839  }
4840  return true;
4841 }
4842 
4843 /// Return true when the instruction sequence within a block is select-cmp-br.
4844 static bool isChainSelectCmpBranch(const SelectInst *SI) {
4845  const BasicBlock *BB = SI->getParent();
4846  if (!BB)
4847  return false;
4848  auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4849  if (!BI || BI->getNumSuccessors() != 2)
4850  return false;
4851  auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4852  if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4853  return false;
4854  return true;
4855 }
4856 
4857 /// True when a select result is replaced by one of its operands
4858 /// in select-icmp sequence. This will eventually result in the elimination
4859 /// of the select.
4860 ///
4861 /// \param SI Select instruction
4862 /// \param Icmp Compare instruction
4863 /// \param SIOpd Operand that replaces the select
4864 ///
4865 /// Notes:
4866 /// - The replacement is global and requires dominator information
4867 /// - The caller is responsible for the actual replacement
4868 ///
4869 /// Example:
4870 ///
4871 /// entry:
4872 /// %4 = select i1 %3, %C* %0, %C* null
4873 /// %5 = icmp eq %C* %4, null
4874 /// br i1 %5, label %9, label %7
4875 /// ...
4876 /// ; <label>:7 ; preds = %entry
4877 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4878 /// ...
4879 ///
4880 /// can be transformed to
4881 ///
4882 /// %5 = icmp eq %C* %0, null
4883 /// %6 = select i1 %3, i1 %5, i1 true
4884 /// br i1 %6, label %9, label %7
4885 /// ...
4886 /// ; <label>:7 ; preds = %entry
4887 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
4888 ///
4889 /// Similar when the first operand of the select is a constant or/and
4890 /// the compare is for not equal rather than equal.
4891 ///
4892 /// NOTE: The function is only called when the select and compare constants
4893 /// are equal, the optimization can work only for EQ predicates. This is not a
4894 /// major restriction since a NE compare should be 'normalized' to an equal
4895 /// compare, which usually happens in the combiner and test case
4896 /// select-cmp-br.ll checks for it.
4898  const ICmpInst *Icmp,
4899  const unsigned SIOpd) {
4900  assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
4901  if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4902  BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4903  // The check for the single predecessor is not the best that can be
4904  // done. But it protects efficiently against cases like when SI's
4905  // home block has two successors, Succ and Succ1, and Succ1 predecessor
4906  // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4907  // replaced can be reached on either path. So the uniqueness check
4908  // guarantees that the path all uses of SI (outside SI's parent) are on
4909  // is disjoint from all other paths out of SI. But that information
4910  // is more expensive to compute, and the trade-off here is in favor
4911  // of compile-time. It should also be noticed that we check for a single
4912  // predecessor and not only uniqueness. This to handle the situation when
4913  // Succ and Succ1 points to the same basic block.
4914  if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
4915  NumSel++;
4916  SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
4917  return true;
4918  }
4919  }
4920  return false;
4921 }
4922 
4923 /// Try to fold the comparison based on range information we can get by checking
4924 /// whether bits are known to be zero or one in the inputs.
4925 Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) {
4926  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4927  Type *Ty = Op0->getType();
4928  ICmpInst::Predicate Pred = I.getPredicate();
4929 
4930  // Get scalar or pointer size.
4931  unsigned BitWidth = Ty->isIntOrIntVectorTy()
4932  ? Ty->getScalarSizeInBits()
4933  : DL.getIndexTypeSizeInBits(Ty->getScalarType());
4934 
4935  if (!BitWidth)
4936  return nullptr;
4937 
4938  KnownBits Op0Known(BitWidth);
4939  KnownBits Op1Known(BitWidth);
4940 
4941  if (SimplifyDemandedBits(&I, 0,
4942  getDemandedBitsLHSMask(I, BitWidth),
4943  Op0Known, 0))
4944  return &I;
4945 
4946  if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
4947  Op1Known, 0))
4948  return &I;
4949 
4950  // Given the known and unknown bits, compute a range that the LHS could be
4951  // in. Compute the Min, Max and RHS values based on the known bits. For the
4952  // EQ and NE we use unsigned values.
4953  APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
4954  APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
4955  if (I.isSigned()) {
4956  computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4957  computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4958  } else {
4959  computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4960  computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4961  }
4962 
4963  // If Min and Max are known to be the same, then SimplifyDemandedBits figured
4964  // out that the LHS or RHS is a constant. Constant fold this now, so that
4965  // code below can assume that Min != Max.
4966  if (!isa<Constant>(Op0) && Op0Min == Op0Max)
4967  return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
4968  if (!isa<Constant>(Op1) && Op1Min == Op1Max)
4969  return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
4970 
4971  // Based on the range information we know about the LHS, see if we can
4972  // simplify this comparison. For example, (x&4) < 8 is always true.
4973  switch (Pred) {
4974  default:
4975  llvm_unreachable("Unknown icmp opcode!");
4976  case ICmpInst::ICMP_EQ:
4977  case ICmpInst::ICMP_NE: {
4978  if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
4979  return Pred == CmpInst::ICMP_EQ
4980  ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
4981  : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4982  }
4983 
4984  // If all bits are known zero except for one, then we know at most one bit
4985  // is set. If the comparison is against zero, then this is a check to see if
4986  // *that* bit is set.
4987  APInt Op0KnownZeroInverted = ~Op0Known.Zero;
4988  if (Op1Known.isZero()) {
4989  // If the LHS is an AND with the same constant, look through it.
4990  Value *LHS = nullptr;
4991  const APInt *LHSC;
4992  if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
4993  *LHSC != Op0KnownZeroInverted)
4994  LHS = Op0;
4995 
4996  Value *X;
4997  if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
4998  APInt ValToCheck = Op0KnownZeroInverted;
4999  Type *XTy = X->getType();
5000  if (ValToCheck.isPowerOf2()) {
5001  // ((1 << X) & 8) == 0 -> X != 3
5002  // ((1 << X) & 8) != 0 -> X == 3
5003  auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5004  auto NewPred = ICmpInst::getInversePredicate(Pred);
5005  return new ICmpInst(NewPred, X, CmpC);
5006  } else if ((++ValToCheck).isPowerOf2()) {
5007  // ((1 << X) & 7) == 0 -> X >= 3
5008  // ((1 << X) & 7) != 0 -> X < 3
5009  auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5010  auto NewPred =
5012  return new ICmpInst(NewPred, X, CmpC);
5013  }
5014  }
5015 
5016  // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
5017  const APInt *CI;
5018  if (Op0KnownZeroInverted.isOneValue() &&
5019  match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
5020  // ((8 >>u X) & 1) == 0 -> X != 3
5021  // ((8 >>u X) & 1) != 0 -> X == 3
5022  unsigned CmpVal = CI->countTrailingZeros();
5023  auto NewPred = ICmpInst::getInversePredicate(Pred);
5024  return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
5025  }
5026  }
5027  break;
5028  }
5029  case ICmpInst::ICMP_ULT: {
5030  if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
5031  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));