LLVM  9.0.0svn
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 /// Given an exploded icmp instruction, return true if the comparison only
73 /// checks the sign bit. If it only checks the sign bit, set TrueIfSigned if the
74 /// result of the comparison is true when the input value is signed.
75 static bool isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS,
76  bool &TrueIfSigned) {
77  switch (Pred) {
78  case ICmpInst::ICMP_SLT: // True if LHS s< 0
79  TrueIfSigned = true;
80  return RHS.isNullValue();
81  case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
82  TrueIfSigned = true;
83  return RHS.isAllOnesValue();
84  case ICmpInst::ICMP_SGT: // True if LHS s> -1
85  TrueIfSigned = false;
86  return RHS.isAllOnesValue();
87  case ICmpInst::ICMP_UGT:
88  // True if LHS u> RHS and RHS == high-bit-mask - 1
89  TrueIfSigned = true;
90  return RHS.isMaxSignedValue();
91  case ICmpInst::ICMP_UGE:
92  // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
93  TrueIfSigned = true;
94  return RHS.isSignMask();
95  default:
96  return false;
97  }
98 }
99 
100 /// Returns true if the exploded icmp can be expressed as a signed comparison
101 /// to zero and updates the predicate accordingly.
102 /// The signedness of the comparison is preserved.
103 /// TODO: Refactor with decomposeBitTestICmp()?
104 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
105  if (!ICmpInst::isSigned(Pred))
106  return false;
107 
108  if (C.isNullValue())
109  return ICmpInst::isRelational(Pred);
110 
111  if (C.isOneValue()) {
112  if (Pred == ICmpInst::ICMP_SLT) {
113  Pred = ICmpInst::ICMP_SLE;
114  return true;
115  }
116  } else if (C.isAllOnesValue()) {
117  if (Pred == ICmpInst::ICMP_SGT) {
118  Pred = ICmpInst::ICMP_SGE;
119  return true;
120  }
121  }
122 
123  return false;
124 }
125 
126 /// Given a signed integer type and a set of known zero and one bits, compute
127 /// the maximum and minimum values that could have the specified known zero and
128 /// known one bits, returning them in Min/Max.
129 /// TODO: Move to method on KnownBits struct?
131  APInt &Min, APInt &Max) {
132  assert(Known.getBitWidth() == Min.getBitWidth() &&
133  Known.getBitWidth() == Max.getBitWidth() &&
134  "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
135  APInt UnknownBits = ~(Known.Zero|Known.One);
136 
137  // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
138  // bit if it is unknown.
139  Min = Known.One;
140  Max = Known.One|UnknownBits;
141 
142  if (UnknownBits.isNegative()) { // Sign bit is unknown
143  Min.setSignBit();
144  Max.clearSignBit();
145  }
146 }
147 
148 /// Given an unsigned integer type and a set of known zero and one bits, compute
149 /// the maximum and minimum values that could have the specified known zero and
150 /// known one bits, returning them in Min/Max.
151 /// TODO: Move to method on KnownBits struct?
153  APInt &Min, APInt &Max) {
154  assert(Known.getBitWidth() == Min.getBitWidth() &&
155  Known.getBitWidth() == Max.getBitWidth() &&
156  "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
157  APInt UnknownBits = ~(Known.Zero|Known.One);
158 
159  // The minimum value is when the unknown bits are all zeros.
160  Min = Known.One;
161  // The maximum value is when the unknown bits are all ones.
162  Max = Known.One|UnknownBits;
163 }
164 
165 /// This is called when we see this pattern:
166 /// cmp pred (load (gep GV, ...)), cmpcst
167 /// where GV is a global variable with a constant initializer. Try to simplify
168 /// this into some simple computation that does not need the load. For example
169 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
170 ///
171 /// If AndCst is non-null, then the loaded value is masked with that constant
172 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
173 Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
174  GlobalVariable *GV,
175  CmpInst &ICI,
176  ConstantInt *AndCst) {
177  Constant *Init = GV->getInitializer();
178  if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
179  return nullptr;
180 
181  uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
182  // Don't blow up on huge arrays.
183  if (ArrayElementCount > MaxArraySizeForCombine)
184  return nullptr;
185 
186  // There are many forms of this optimization we can handle, for now, just do
187  // the simple index into a single-dimensional array.
188  //
189  // Require: GEP GV, 0, i {{, constant indices}}
190  if (GEP->getNumOperands() < 3 ||
191  !isa<ConstantInt>(GEP->getOperand(1)) ||
192  !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
193  isa<Constant>(GEP->getOperand(2)))
194  return nullptr;
195 
196  // Check that indices after the variable are constants and in-range for the
197  // type they index. Collect the indices. This is typically for arrays of
198  // structs.
199  SmallVector<unsigned, 4> LaterIndices;
200 
201  Type *EltTy = Init->getType()->getArrayElementType();
202  for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
203  ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
204  if (!Idx) return nullptr; // Variable index.
205 
206  uint64_t IdxVal = Idx->getZExtValue();
207  if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
208 
209  if (StructType *STy = dyn_cast<StructType>(EltTy))
210  EltTy = STy->getElementType(IdxVal);
211  else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
212  if (IdxVal >= ATy->getNumElements()) return nullptr;
213  EltTy = ATy->getElementType();
214  } else {
215  return nullptr; // Unknown type.
216  }
217 
218  LaterIndices.push_back(IdxVal);
219  }
220 
221  enum { Overdefined = -3, Undefined = -2 };
222 
223  // Variables for our state machines.
224 
225  // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
226  // "i == 47 | i == 87", where 47 is the first index the condition is true for,
227  // and 87 is the second (and last) index. FirstTrueElement is -2 when
228  // undefined, otherwise set to the first true element. SecondTrueElement is
229  // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
230  int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
231 
232  // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
233  // form "i != 47 & i != 87". Same state transitions as for true elements.
234  int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
235 
236  /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
237  /// define a state machine that triggers for ranges of values that the index
238  /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
239  /// This is -2 when undefined, -3 when overdefined, and otherwise the last
240  /// index in the range (inclusive). We use -2 for undefined here because we
241  /// use relative comparisons and don't want 0-1 to match -1.
242  int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
243 
244  // MagicBitvector - This is a magic bitvector where we set a bit if the
245  // comparison is true for element 'i'. If there are 64 elements or less in
246  // the array, this will fully represent all the comparison results.
247  uint64_t MagicBitvector = 0;
248 
249  // Scan the array and see if one of our patterns matches.
250  Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
251  for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
252  Constant *Elt = Init->getAggregateElement(i);
253  if (!Elt) return nullptr;
254 
255  // If this is indexing an array of structures, get the structure element.
256  if (!LaterIndices.empty())
257  Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
258 
259  // If the element is masked, handle it.
260  if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
261 
262  // Find out if the comparison would be true or false for the i'th element.
264  CompareRHS, DL, &TLI);
265  // If the result is undef for this element, ignore it.
266  if (isa<UndefValue>(C)) {
267  // Extend range state machines to cover this element in case there is an
268  // undef in the middle of the range.
269  if (TrueRangeEnd == (int)i-1)
270  TrueRangeEnd = i;
271  if (FalseRangeEnd == (int)i-1)
272  FalseRangeEnd = i;
273  continue;
274  }
275 
276  // If we can't compute the result for any of the elements, we have to give
277  // up evaluating the entire conditional.
278  if (!isa<ConstantInt>(C)) return nullptr;
279 
280  // Otherwise, we know if the comparison is true or false for this element,
281  // update our state machines.
282  bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
283 
284  // State machine for single/double/range index comparison.
285  if (IsTrueForElt) {
286  // Update the TrueElement state machine.
287  if (FirstTrueElement == Undefined)
288  FirstTrueElement = TrueRangeEnd = i; // First true element.
289  else {
290  // Update double-compare state machine.
291  if (SecondTrueElement == Undefined)
292  SecondTrueElement = i;
293  else
294  SecondTrueElement = Overdefined;
295 
296  // Update range state machine.
297  if (TrueRangeEnd == (int)i-1)
298  TrueRangeEnd = i;
299  else
300  TrueRangeEnd = Overdefined;
301  }
302  } else {
303  // Update the FalseElement state machine.
304  if (FirstFalseElement == Undefined)
305  FirstFalseElement = FalseRangeEnd = i; // First false element.
306  else {
307  // Update double-compare state machine.
308  if (SecondFalseElement == Undefined)
309  SecondFalseElement = i;
310  else
311  SecondFalseElement = Overdefined;
312 
313  // Update range state machine.
314  if (FalseRangeEnd == (int)i-1)
315  FalseRangeEnd = i;
316  else
317  FalseRangeEnd = Overdefined;
318  }
319  }
320 
321  // If this element is in range, update our magic bitvector.
322  if (i < 64 && IsTrueForElt)
323  MagicBitvector |= 1ULL << i;
324 
325  // If all of our states become overdefined, bail out early. Since the
326  // predicate is expensive, only check it every 8 elements. This is only
327  // really useful for really huge arrays.
328  if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
329  SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
330  FalseRangeEnd == Overdefined)
331  return nullptr;
332  }
333 
334  // Now that we've scanned the entire array, emit our new comparison(s). We
335  // order the state machines in complexity of the generated code.
336  Value *Idx = GEP->getOperand(2);
337 
338  // If the index is larger than the pointer size of the target, truncate the
339  // index down like the GEP would do implicitly. We don't have to do this for
340  // an inbounds GEP because the index can't be out of range.
341  if (!GEP->isInBounds()) {
342  Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
343  unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
344  if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
345  Idx = Builder.CreateTrunc(Idx, IntPtrTy);
346  }
347 
348  // If the comparison is only true for one or two elements, emit direct
349  // comparisons.
350  if (SecondTrueElement != Overdefined) {
351  // None true -> false.
352  if (FirstTrueElement == Undefined)
353  return replaceInstUsesWith(ICI, Builder.getFalse());
354 
355  Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
356 
357  // True for one element -> 'i == 47'.
358  if (SecondTrueElement == Undefined)
359  return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
360 
361  // True for two elements -> 'i == 47 | i == 72'.
362  Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
363  Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
364  Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
365  return BinaryOperator::CreateOr(C1, C2);
366  }
367 
368  // If the comparison is only false for one or two elements, emit direct
369  // comparisons.
370  if (SecondFalseElement != Overdefined) {
371  // None false -> true.
372  if (FirstFalseElement == Undefined)
373  return replaceInstUsesWith(ICI, Builder.getTrue());
374 
375  Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
376 
377  // False for one element -> 'i != 47'.
378  if (SecondFalseElement == Undefined)
379  return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
380 
381  // False for two elements -> 'i != 47 & i != 72'.
382  Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
383  Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
384  Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
385  return BinaryOperator::CreateAnd(C1, C2);
386  }
387 
388  // If the comparison can be replaced with a range comparison for the elements
389  // where it is true, emit the range check.
390  if (TrueRangeEnd != Overdefined) {
391  assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
392 
393  // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
394  if (FirstTrueElement) {
395  Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
396  Idx = Builder.CreateAdd(Idx, Offs);
397  }
398 
399  Value *End = ConstantInt::get(Idx->getType(),
400  TrueRangeEnd-FirstTrueElement+1);
401  return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
402  }
403 
404  // False range check.
405  if (FalseRangeEnd != Overdefined) {
406  assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
407  // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
408  if (FirstFalseElement) {
409  Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
410  Idx = Builder.CreateAdd(Idx, Offs);
411  }
412 
413  Value *End = ConstantInt::get(Idx->getType(),
414  FalseRangeEnd-FirstFalseElement);
415  return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
416  }
417 
418  // If a magic bitvector captures the entire comparison state
419  // of this load, replace it with computation that does:
420  // ((magic_cst >> i) & 1) != 0
421  {
422  Type *Ty = nullptr;
423 
424  // Look for an appropriate type:
425  // - The type of Idx if the magic fits
426  // - The smallest fitting legal type
427  if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
428  Ty = Idx->getType();
429  else
430  Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
431 
432  if (Ty) {
433  Value *V = Builder.CreateIntCast(Idx, Ty, false);
434  V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
435  V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
436  return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
437  }
438  }
439 
440  return nullptr;
441 }
442 
443 /// Return a value that can be used to compare the *offset* implied by a GEP to
444 /// zero. For example, if we have &A[i], we want to return 'i' for
445 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
446 /// are involved. The above expression would also be legal to codegen as
447 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
448 /// This latter form is less amenable to optimization though, and we are allowed
449 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
450 ///
451 /// If we can't emit an optimized form for this expression, this returns null.
452 ///
454  const DataLayout &DL) {
456 
457  // Check to see if this gep only has a single variable index. If so, and if
458  // any constant indices are a multiple of its scale, then we can compute this
459  // in terms of the scale of the variable index. For example, if the GEP
460  // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
461  // because the expression will cross zero at the same point.
462  unsigned i, e = GEP->getNumOperands();
463  int64_t Offset = 0;
464  for (i = 1; i != e; ++i, ++GTI) {
465  if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
466  // Compute the aggregate offset of constant indices.
467  if (CI->isZero()) continue;
468 
469  // Handle a struct index, which adds its field offset to the pointer.
470  if (StructType *STy = GTI.getStructTypeOrNull()) {
471  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
472  } else {
473  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
474  Offset += Size*CI->getSExtValue();
475  }
476  } else {
477  // Found our variable index.
478  break;
479  }
480  }
481 
482  // If there are no variable indices, we must have a constant offset, just
483  // evaluate it the general way.
484  if (i == e) return nullptr;
485 
486  Value *VariableIdx = GEP->getOperand(i);
487  // Determine the scale factor of the variable element. For example, this is
488  // 4 if the variable index is into an array of i32.
489  uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
490 
491  // Verify that there are no other variable indices. If so, emit the hard way.
492  for (++i, ++GTI; i != e; ++i, ++GTI) {
493  ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
494  if (!CI) return nullptr;
495 
496  // Compute the aggregate offset of constant indices.
497  if (CI->isZero()) continue;
498 
499  // Handle a struct index, which adds its field offset to the pointer.
500  if (StructType *STy = GTI.getStructTypeOrNull()) {
501  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
502  } else {
503  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
504  Offset += Size*CI->getSExtValue();
505  }
506  }
507 
508  // Okay, we know we have a single variable index, which must be a
509  // pointer/array/vector index. If there is no offset, life is simple, return
510  // the index.
511  Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
512  unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
513  if (Offset == 0) {
514  // Cast to intptrty in case a truncation occurs. If an extension is needed,
515  // we don't need to bother extending: the extension won't affect where the
516  // computation crosses zero.
517  if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
518  VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
519  }
520  return VariableIdx;
521  }
522 
523  // Otherwise, there is an index. The computation we will do will be modulo
524  // the pointer size.
525  Offset = SignExtend64(Offset, IntPtrWidth);
526  VariableScale = SignExtend64(VariableScale, IntPtrWidth);
527 
528  // To do this transformation, any constant index must be a multiple of the
529  // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
530  // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
531  // multiple of the variable scale.
532  int64_t NewOffs = Offset / (int64_t)VariableScale;
533  if (Offset != NewOffs*(int64_t)VariableScale)
534  return nullptr;
535 
536  // Okay, we can do this evaluation. Start by converting the index to intptr.
537  if (VariableIdx->getType() != IntPtrTy)
538  VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
539  true /*Signed*/);
540  Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
541  return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
542 }
543 
544 /// Returns true if we can rewrite Start as a GEP with pointer Base
545 /// and some integer offset. The nodes that need to be re-written
546 /// for this transformation will be added to Explored.
547 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
548  const DataLayout &DL,
549  SetVector<Value *> &Explored) {
550  SmallVector<Value *, 16> WorkList(1, Start);
551  Explored.insert(Base);
552 
553  // The following traversal gives us an order which can be used
554  // when doing the final transformation. Since in the final
555  // transformation we create the PHI replacement instructions first,
556  // we don't have to get them in any particular order.
557  //
558  // However, for other instructions we will have to traverse the
559  // operands of an instruction first, which means that we have to
560  // do a post-order traversal.
561  while (!WorkList.empty()) {
563 
564  while (!WorkList.empty()) {
565  if (Explored.size() >= 100)
566  return false;
567 
568  Value *V = WorkList.back();
569 
570  if (Explored.count(V) != 0) {
571  WorkList.pop_back();
572  continue;
573  }
574 
575  if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
576  !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
577  // We've found some value that we can't explore which is different from
578  // the base. Therefore we can't do this transformation.
579  return false;
580 
581  if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
582  auto *CI = dyn_cast<CastInst>(V);
583  if (!CI->isNoopCast(DL))
584  return false;
585 
586  if (Explored.count(CI->getOperand(0)) == 0)
587  WorkList.push_back(CI->getOperand(0));
588  }
589 
590  if (auto *GEP = dyn_cast<GEPOperator>(V)) {
591  // We're limiting the GEP to having one index. This will preserve
592  // the original pointer type. We could handle more cases in the
593  // future.
594  if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
595  GEP->getType() != Start->getType())
596  return false;
597 
598  if (Explored.count(GEP->getOperand(0)) == 0)
599  WorkList.push_back(GEP->getOperand(0));
600  }
601 
602  if (WorkList.back() == V) {
603  WorkList.pop_back();
604  // We've finished visiting this node, mark it as such.
605  Explored.insert(V);
606  }
607 
608  if (auto *PN = dyn_cast<PHINode>(V)) {
609  // We cannot transform PHIs on unsplittable basic blocks.
610  if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
611  return false;
612  Explored.insert(PN);
613  PHIs.insert(PN);
614  }
615  }
616 
617  // Explore the PHI nodes further.
618  for (auto *PN : PHIs)
619  for (Value *Op : PN->incoming_values())
620  if (Explored.count(Op) == 0)
621  WorkList.push_back(Op);
622  }
623 
624  // Make sure that we can do this. Since we can't insert GEPs in a basic
625  // block before a PHI node, we can't easily do this transformation if
626  // we have PHI node users of transformed instructions.
627  for (Value *Val : Explored) {
628  for (Value *Use : Val->uses()) {
629 
630  auto *PHI = dyn_cast<PHINode>(Use);
631  auto *Inst = dyn_cast<Instruction>(Val);
632 
633  if (Inst == Base || Inst == PHI || !Inst || !PHI ||
634  Explored.count(PHI) == 0)
635  continue;
636 
637  if (PHI->getParent() == Inst->getParent())
638  return false;
639  }
640  }
641  return true;
642 }
643 
644 // Sets the appropriate insert point on Builder where we can add
645 // a replacement Instruction for V (if that is possible).
646 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
647  bool Before = true) {
648  if (auto *PHI = dyn_cast<PHINode>(V)) {
649  Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
650  return;
651  }
652  if (auto *I = dyn_cast<Instruction>(V)) {
653  if (!Before)
654  I = &*std::next(I->getIterator());
655  Builder.SetInsertPoint(I);
656  return;
657  }
658  if (auto *A = dyn_cast<Argument>(V)) {
659  // Set the insertion point in the entry block.
660  BasicBlock &Entry = A->getParent()->getEntryBlock();
661  Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
662  return;
663  }
664  // Otherwise, this is a constant and we don't need to set a new
665  // insertion point.
666  assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
667 }
668 
669 /// Returns a re-written value of Start as an indexed GEP using Base as a
670 /// pointer.
672  const DataLayout &DL,
673  SetVector<Value *> &Explored) {
674  // Perform all the substitutions. This is a bit tricky because we can
675  // have cycles in our use-def chains.
676  // 1. Create the PHI nodes without any incoming values.
677  // 2. Create all the other values.
678  // 3. Add the edges for the PHI nodes.
679  // 4. Emit GEPs to get the original pointers.
680  // 5. Remove the original instructions.
681  Type *IndexType = IntegerType::get(
682  Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
683 
685  NewInsts[Base] = ConstantInt::getNullValue(IndexType);
686 
687  // Create the new PHI nodes, without adding any incoming values.
688  for (Value *Val : Explored) {
689  if (Val == Base)
690  continue;
691  // Create empty phi nodes. This avoids cyclic dependencies when creating
692  // the remaining instructions.
693  if (auto *PHI = dyn_cast<PHINode>(Val))
694  NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
695  PHI->getName() + ".idx", PHI);
696  }
697  IRBuilder<> Builder(Base->getContext());
698 
699  // Create all the other instructions.
700  for (Value *Val : Explored) {
701 
702  if (NewInsts.find(Val) != NewInsts.end())
703  continue;
704 
705  if (auto *CI = dyn_cast<CastInst>(Val)) {
706  // Don't get rid of the intermediate variable here; the store can grow
707  // the map which will invalidate the reference to the input value.
708  Value *V = NewInsts[CI->getOperand(0)];
709  NewInsts[CI] = V;
710  continue;
711  }
712  if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
713  Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
714  : GEP->getOperand(1);
715  setInsertionPoint(Builder, GEP);
716  // Indices might need to be sign extended. GEPs will magically do
717  // this, but we need to do it ourselves here.
718  if (Index->getType()->getScalarSizeInBits() !=
719  NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
720  Index = Builder.CreateSExtOrTrunc(
721  Index, NewInsts[GEP->getOperand(0)]->getType(),
722  GEP->getOperand(0)->getName() + ".sext");
723  }
724 
725  auto *Op = NewInsts[GEP->getOperand(0)];
726  if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
727  NewInsts[GEP] = Index;
728  else
729  NewInsts[GEP] = Builder.CreateNSWAdd(
730  Op, Index, GEP->getOperand(0)->getName() + ".add");
731  continue;
732  }
733  if (isa<PHINode>(Val))
734  continue;
735 
736  llvm_unreachable("Unexpected instruction type");
737  }
738 
739  // Add the incoming values to the PHI nodes.
740  for (Value *Val : Explored) {
741  if (Val == Base)
742  continue;
743  // All the instructions have been created, we can now add edges to the
744  // phi nodes.
745  if (auto *PHI = dyn_cast<PHINode>(Val)) {
746  PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
747  for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
748  Value *NewIncoming = PHI->getIncomingValue(I);
749 
750  if (NewInsts.find(NewIncoming) != NewInsts.end())
751  NewIncoming = NewInsts[NewIncoming];
752 
753  NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
754  }
755  }
756  }
757 
758  for (Value *Val : Explored) {
759  if (Val == Base)
760  continue;
761 
762  // Depending on the type, for external users we have to emit
763  // a GEP or a GEP + ptrtoint.
764  setInsertionPoint(Builder, Val, false);
765 
766  // If required, create an inttoptr instruction for Base.
767  Value *NewBase = Base;
768  if (!Base->getType()->isPointerTy())
769  NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
770  Start->getName() + "to.ptr");
771 
772  Value *GEP = Builder.CreateInBoundsGEP(
773  Start->getType()->getPointerElementType(), NewBase,
774  makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
775 
776  if (!Val->getType()->isPointerTy()) {
777  Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
778  Val->getName() + ".conv");
779  GEP = Cast;
780  }
781  Val->replaceAllUsesWith(GEP);
782  }
783 
784  return NewInsts[Start];
785 }
786 
787 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
788 /// the input Value as a constant indexed GEP. Returns a pair containing
789 /// the GEPs Pointer and Index.
790 static std::pair<Value *, Value *>
792  Type *IndexType = IntegerType::get(V->getContext(),
794 
796  while (true) {
797  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
798  // We accept only inbouds GEPs here to exclude the possibility of
799  // overflow.
800  if (!GEP->isInBounds())
801  break;
802  if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
803  GEP->getType() == V->getType()) {
804  V = GEP->getOperand(0);
805  Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
806  Index = ConstantExpr::getAdd(
807  Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
808  continue;
809  }
810  break;
811  }
812  if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
813  if (!CI->isNoopCast(DL))
814  break;
815  V = CI->getOperand(0);
816  continue;
817  }
818  if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
819  if (!CI->isNoopCast(DL))
820  break;
821  V = CI->getOperand(0);
822  continue;
823  }
824  break;
825  }
826  return {V, Index};
827 }
828 
829 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
830 /// We can look through PHIs, GEPs and casts in order to determine a common base
831 /// between GEPLHS and RHS.
833  ICmpInst::Predicate Cond,
834  const DataLayout &DL) {
835  if (!GEPLHS->hasAllConstantIndices())
836  return nullptr;
837 
838  // Make sure the pointers have the same type.
839  if (GEPLHS->getType() != RHS->getType())
840  return nullptr;
841 
842  Value *PtrBase, *Index;
843  std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
844 
845  // The set of nodes that will take part in this transformation.
846  SetVector<Value *> Nodes;
847 
848  if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
849  return nullptr;
850 
851  // We know we can re-write this as
852  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
853  // Since we've only looked through inbouds GEPs we know that we
854  // can't have overflow on either side. We can therefore re-write
855  // this as:
856  // OFFSET1 cmp OFFSET2
857  Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
858 
859  // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
860  // GEP having PtrBase as the pointer base, and has returned in NewRHS the
861  // offset. Since Index is the offset of LHS to the base pointer, we will now
862  // compare the offsets instead of comparing the pointers.
863  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
864 }
865 
866 /// Fold comparisons between a GEP instruction and something else. At this point
867 /// we know that the GEP is on the LHS of the comparison.
868 Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
869  ICmpInst::Predicate Cond,
870  Instruction &I) {
871  // Don't transform signed compares of GEPs into index compares. Even if the
872  // GEP is inbounds, the final add of the base pointer can have signed overflow
873  // and would change the result of the icmp.
874  // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
875  // the maximum signed value for the pointer type.
876  if (ICmpInst::isSigned(Cond))
877  return nullptr;
878 
879  // Look through bitcasts and addrspacecasts. We do not however want to remove
880  // 0 GEPs.
881  if (!isa<GetElementPtrInst>(RHS))
882  RHS = RHS->stripPointerCasts();
883 
884  Value *PtrBase = GEPLHS->getOperand(0);
885  if (PtrBase == RHS && GEPLHS->isInBounds()) {
886  // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
887  // This transformation (ignoring the base and scales) is valid because we
888  // know pointers can't overflow since the gep is inbounds. See if we can
889  // output an optimized form.
890  Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
891 
892  // If not, synthesize the offset the hard way.
893  if (!Offset)
894  Offset = EmitGEPOffset(GEPLHS);
895  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
896  Constant::getNullValue(Offset->getType()));
897  } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
898  // If the base pointers are different, but the indices are the same, just
899  // compare the base pointer.
900  if (PtrBase != GEPRHS->getOperand(0)) {
901  bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
902  IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
903  GEPRHS->getOperand(0)->getType();
904  if (IndicesTheSame)
905  for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
906  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
907  IndicesTheSame = false;
908  break;
909  }
910 
911  // If all indices are the same, just compare the base pointers.
912  Type *BaseType = GEPLHS->getOperand(0)->getType();
913  if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
914  return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
915 
916  // If we're comparing GEPs with two base pointers that only differ in type
917  // and both GEPs have only constant indices or just one use, then fold
918  // the compare with the adjusted indices.
919  if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
920  (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
921  (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
922  PtrBase->stripPointerCasts() ==
923  GEPRHS->getOperand(0)->stripPointerCasts()) {
924  Value *LOffset = EmitGEPOffset(GEPLHS);
925  Value *ROffset = EmitGEPOffset(GEPRHS);
926 
927  // If we looked through an addrspacecast between different sized address
928  // spaces, the LHS and RHS pointers are different sized
929  // integers. Truncate to the smaller one.
930  Type *LHSIndexTy = LOffset->getType();
931  Type *RHSIndexTy = ROffset->getType();
932  if (LHSIndexTy != RHSIndexTy) {
933  if (LHSIndexTy->getPrimitiveSizeInBits() <
934  RHSIndexTy->getPrimitiveSizeInBits()) {
935  ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
936  } else
937  LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
938  }
939 
940  Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
941  LOffset, ROffset);
942  return replaceInstUsesWith(I, Cmp);
943  }
944 
945  // Otherwise, the base pointers are different and the indices are
946  // different. Try convert this to an indexed compare by looking through
947  // PHIs/casts.
948  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
949  }
950 
951  // If one of the GEPs has all zero indices, recurse.
952  if (GEPLHS->hasAllZeroIndices())
953  return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
955 
956  // If the other GEP has all zero indices, recurse.
957  if (GEPRHS->hasAllZeroIndices())
958  return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
959 
960  bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
961  if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
962  // If the GEPs only differ by one index, compare it.
963  unsigned NumDifferences = 0; // Keep track of # differences.
964  unsigned DiffOperand = 0; // The operand that differs.
965  for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
966  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
967  if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
968  GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
969  // Irreconcilable differences.
970  NumDifferences = 2;
971  break;
972  } else {
973  if (NumDifferences++) break;
974  DiffOperand = i;
975  }
976  }
977 
978  if (NumDifferences == 0) // SAME GEP?
979  return replaceInstUsesWith(I, // No comparison is needed here.
981 
982  else if (NumDifferences == 1 && GEPsInBounds) {
983  Value *LHSV = GEPLHS->getOperand(DiffOperand);
984  Value *RHSV = GEPRHS->getOperand(DiffOperand);
985  // Make sure we do a signed comparison here.
986  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
987  }
988  }
989 
990  // Only lower this if the icmp is the only user of the GEP or if we expect
991  // the result to fold to a constant!
992  if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
993  (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
994  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
995  Value *L = EmitGEPOffset(GEPLHS);
996  Value *R = EmitGEPOffset(GEPRHS);
997  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
998  }
999  }
1000 
1001  // Try convert this to an indexed compare by looking through PHIs/casts as a
1002  // last resort.
1003  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1004 }
1005 
1006 Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI,
1007  const AllocaInst *Alloca,
1008  const Value *Other) {
1009  assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1010 
1011  // It would be tempting to fold away comparisons between allocas and any
1012  // pointer not based on that alloca (e.g. an argument). However, even
1013  // though such pointers cannot alias, they can still compare equal.
1014  //
1015  // But LLVM doesn't specify where allocas get their memory, so if the alloca
1016  // doesn't escape we can argue that it's impossible to guess its value, and we
1017  // can therefore act as if any such guesses are wrong.
1018  //
1019  // The code below checks that the alloca doesn't escape, and that it's only
1020  // used in a comparison once (the current instruction). The
1021  // single-comparison-use condition ensures that we're trivially folding all
1022  // comparisons against the alloca consistently, and avoids the risk of
1023  // erroneously folding a comparison of the pointer with itself.
1024 
1025  unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1026 
1028  for (const Use &U : Alloca->uses()) {
1029  if (Worklist.size() >= MaxIter)
1030  return nullptr;
1031  Worklist.push_back(&U);
1032  }
1033 
1034  unsigned NumCmps = 0;
1035  while (!Worklist.empty()) {
1036  assert(Worklist.size() <= MaxIter);
1037  const Use *U = Worklist.pop_back_val();
1038  const Value *V = U->getUser();
1039  --MaxIter;
1040 
1041  if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1042  isa<SelectInst>(V)) {
1043  // Track the uses.
1044  } else if (isa<LoadInst>(V)) {
1045  // Loading from the pointer doesn't escape it.
1046  continue;
1047  } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1048  // Storing *to* the pointer is fine, but storing the pointer escapes it.
1049  if (SI->getValueOperand() == U->get())
1050  return nullptr;
1051  continue;
1052  } else if (isa<ICmpInst>(V)) {
1053  if (NumCmps++)
1054  return nullptr; // Found more than one cmp.
1055  continue;
1056  } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1057  switch (Intrin->getIntrinsicID()) {
1058  // These intrinsics don't escape or compare the pointer. Memset is safe
1059  // because we don't allow ptrtoint. Memcpy and memmove are safe because
1060  // we don't allow stores, so src cannot point to V.
1061  case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1062  case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1063  continue;
1064  default:
1065  return nullptr;
1066  }
1067  } else {
1068  return nullptr;
1069  }
1070  for (const Use &U : V->uses()) {
1071  if (Worklist.size() >= MaxIter)
1072  return nullptr;
1073  Worklist.push_back(&U);
1074  }
1075  }
1076 
1077  Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1078  return replaceInstUsesWith(
1079  ICI,
1081 }
1082 
1083 /// Fold "icmp pred (X+C), X".
1084 Instruction *InstCombiner::foldICmpAddOpConst(Value *X, const APInt &C,
1085  ICmpInst::Predicate Pred) {
1086  // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1087  // so the values can never be equal. Similarly for all other "or equals"
1088  // operators.
1089  assert(!!C && "C should not be zero!");
1090 
1091  // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1092  // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1093  // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1094  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1095  Constant *R = ConstantInt::get(X->getType(),
1096  APInt::getMaxValue(C.getBitWidth()) - C);
1097  return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1098  }
1099 
1100  // (X+1) >u X --> X <u (0-1) --> X != 255
1101  // (X+2) >u X --> X <u (0-2) --> X <u 254
1102  // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1103  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1104  return new ICmpInst(ICmpInst::ICMP_ULT, X,
1105  ConstantInt::get(X->getType(), -C));
1106 
1108 
1109  // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1110  // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1111  // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1112  // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1113  // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1114  // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1115  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1116  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1117  ConstantInt::get(X->getType(), SMax - C));
1118 
1119  // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1120  // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1121  // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1122  // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1123  // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1124  // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1125 
1126  assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1127  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1128  ConstantInt::get(X->getType(), SMax - (C - 1)));
1129 }
1130 
1131 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1132 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1133 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1134 Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A,
1135  const APInt &AP1,
1136  const APInt &AP2) {
1137  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1138 
1139  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1140  if (I.getPredicate() == I.ICMP_NE)
1141  Pred = CmpInst::getInversePredicate(Pred);
1142  return new ICmpInst(Pred, LHS, RHS);
1143  };
1144 
1145  // Don't bother doing any work for cases which InstSimplify handles.
1146  if (AP2.isNullValue())
1147  return nullptr;
1148 
1149  bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1150  if (IsAShr) {
1151  if (AP2.isAllOnesValue())
1152  return nullptr;
1153  if (AP2.isNegative() != AP1.isNegative())
1154  return nullptr;
1155  if (AP2.sgt(AP1))
1156  return nullptr;
1157  }
1158 
1159  if (!AP1)
1160  // 'A' must be large enough to shift out the highest set bit.
1161  return getICmp(I.ICMP_UGT, A,
1162  ConstantInt::get(A->getType(), AP2.logBase2()));
1163 
1164  if (AP1 == AP2)
1165  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1166 
1167  int Shift;
1168  if (IsAShr && AP1.isNegative())
1169  Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1170  else
1171  Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1172 
1173  if (Shift > 0) {
1174  if (IsAShr && AP1 == AP2.ashr(Shift)) {
1175  // There are multiple solutions if we are comparing against -1 and the LHS
1176  // of the ashr is not a power of two.
1177  if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1178  return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1179  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1180  } else if (AP1 == AP2.lshr(Shift)) {
1181  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1182  }
1183  }
1184 
1185  // Shifting const2 will never be equal to const1.
1186  // FIXME: This should always be handled by InstSimplify?
1187  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1188  return replaceInstUsesWith(I, TorF);
1189 }
1190 
1191 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1192 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1193 Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A,
1194  const APInt &AP1,
1195  const APInt &AP2) {
1196  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1197 
1198  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1199  if (I.getPredicate() == I.ICMP_NE)
1200  Pred = CmpInst::getInversePredicate(Pred);
1201  return new ICmpInst(Pred, LHS, RHS);
1202  };
1203 
1204  // Don't bother doing any work for cases which InstSimplify handles.
1205  if (AP2.isNullValue())
1206  return nullptr;
1207 
1208  unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1209 
1210  if (!AP1 && AP2TrailingZeros != 0)
1211  return getICmp(
1212  I.ICMP_UGE, A,
1213  ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1214 
1215  if (AP1 == AP2)
1216  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1217 
1218  // Get the distance between the lowest bits that are set.
1219  int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1220 
1221  if (Shift > 0 && AP2.shl(Shift) == AP1)
1222  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1223 
1224  // Shifting const2 will never be equal to const1.
1225  // FIXME: This should always be handled by InstSimplify?
1226  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1227  return replaceInstUsesWith(I, TorF);
1228 }
1229 
1230 /// The caller has matched a pattern of the form:
1231 /// I = icmp ugt (add (add A, B), CI2), CI1
1232 /// If this is of the form:
1233 /// sum = a + b
1234 /// if (sum+128 >u 255)
1235 /// Then replace it with llvm.sadd.with.overflow.i8.
1236 ///
1238  ConstantInt *CI2, ConstantInt *CI1,
1239  InstCombiner &IC) {
1240  // The transformation we're trying to do here is to transform this into an
1241  // llvm.sadd.with.overflow. To do this, we have to replace the original add
1242  // with a narrower add, and discard the add-with-constant that is part of the
1243  // range check (if we can't eliminate it, this isn't profitable).
1244 
1245  // In order to eliminate the add-with-constant, the compare can be its only
1246  // use.
1247  Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1248  if (!AddWithCst->hasOneUse())
1249  return nullptr;
1250 
1251  // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1252  if (!CI2->getValue().isPowerOf2())
1253  return nullptr;
1254  unsigned NewWidth = CI2->getValue().countTrailingZeros();
1255  if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1256  return nullptr;
1257 
1258  // The width of the new add formed is 1 more than the bias.
1259  ++NewWidth;
1260 
1261  // Check to see that CI1 is an all-ones value with NewWidth bits.
1262  if (CI1->getBitWidth() == NewWidth ||
1263  CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1264  return nullptr;
1265 
1266  // This is only really a signed overflow check if the inputs have been
1267  // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1268  // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1269  unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1270  if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1271  IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1272  return nullptr;
1273 
1274  // In order to replace the original add with a narrower
1275  // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1276  // and truncates that discard the high bits of the add. Verify that this is
1277  // the case.
1278  Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1279  for (User *U : OrigAdd->users()) {
1280  if (U == AddWithCst)
1281  continue;
1282 
1283  // Only accept truncates for now. We would really like a nice recursive
1284  // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1285  // chain to see which bits of a value are actually demanded. If the
1286  // original add had another add which was then immediately truncated, we
1287  // could still do the transformation.
1288  TruncInst *TI = dyn_cast<TruncInst>(U);
1289  if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1290  return nullptr;
1291  }
1292 
1293  // If the pattern matches, truncate the inputs to the narrower type and
1294  // use the sadd_with_overflow intrinsic to efficiently compute both the
1295  // result and the overflow bit.
1296  Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1298  I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1299 
1300  InstCombiner::BuilderTy &Builder = IC.Builder;
1301 
1302  // Put the new code above the original add, in case there are any uses of the
1303  // add between the add and the compare.
1304  Builder.SetInsertPoint(OrigAdd);
1305 
1306  Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1307  Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1308  CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1309  Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1310  Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1311 
1312  // The inner add was the result of the narrow add, zero extended to the
1313  // wider type. Replace it with the result computed by the intrinsic.
1314  IC.replaceInstUsesWith(*OrigAdd, ZExt);
1315 
1316  // The original icmp gets replaced with the overflow value.
1317  return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1318 }
1319 
1320 // Handle (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1321 Instruction *InstCombiner::foldICmpWithZero(ICmpInst &Cmp) {
1322  CmpInst::Predicate Pred = Cmp.getPredicate();
1323  Value *X = Cmp.getOperand(0);
1324 
1325  if (match(Cmp.getOperand(1), m_Zero()) && Pred == ICmpInst::ICMP_SGT) {
1326  Value *A, *B;
1327  SelectPatternResult SPR = matchSelectPattern(X, A, B);
1328  if (SPR.Flavor == SPF_SMIN) {
1329  if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1330  return new ICmpInst(Pred, B, Cmp.getOperand(1));
1331  if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1332  return new ICmpInst(Pred, A, Cmp.getOperand(1));
1333  }
1334  }
1335  return nullptr;
1336 }
1337 
1338 /// Fold icmp Pred X, C.
1339 /// TODO: This code structure does not make sense. The saturating add fold
1340 /// should be moved to some other helper and extended as noted below (it is also
1341 /// possible that code has been made unnecessary - do we canonicalize IR to
1342 /// overflow/saturating intrinsics or not?).
1343 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
1344  // Match the following pattern, which is a common idiom when writing
1345  // overflow-safe integer arithmetic functions. The source performs an addition
1346  // in wider type and explicitly checks for overflow using comparisons against
1347  // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1348  //
1349  // TODO: This could probably be generalized to handle other overflow-safe
1350  // operations if we worked out the formulas to compute the appropriate magic
1351  // constants.
1352  //
1353  // sum = a + b
1354  // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1355  CmpInst::Predicate Pred = Cmp.getPredicate();
1356  Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1357  Value *A, *B;
1358  ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1359  if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1360  match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1361  if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1362  return Res;
1363 
1364  return nullptr;
1365 }
1366 
1367 /// Canonicalize icmp instructions based on dominating conditions.
1368 Instruction *InstCombiner::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1369  // This is a cheap/incomplete check for dominance - just match a single
1370  // predecessor with a conditional branch.
1371  BasicBlock *CmpBB = Cmp.getParent();
1372  BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1373  if (!DomBB)
1374  return nullptr;
1375 
1376  Value *DomCond;
1377  BasicBlock *TrueBB, *FalseBB;
1378  if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1379  return nullptr;
1380 
1381  assert((TrueBB == CmpBB || FalseBB == CmpBB) &&
1382  "Predecessor block does not point to successor?");
1383 
1384  // The branch should get simplified. Don't bother simplifying this condition.
1385  if (TrueBB == FalseBB)
1386  return nullptr;
1387 
1388  // Try to simplify this compare to T/F based on the dominating condition.
1389  Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1390  if (Imp)
1391  return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1392 
1393  CmpInst::Predicate Pred = Cmp.getPredicate();
1394  Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1395  ICmpInst::Predicate DomPred;
1396  const APInt *C, *DomC;
1397  if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1398  match(Y, m_APInt(C))) {
1399  // We have 2 compares of a variable with constants. Calculate the constant
1400  // ranges of those compares to see if we can transform the 2nd compare:
1401  // DomBB:
1402  // DomCond = icmp DomPred X, DomC
1403  // br DomCond, CmpBB, FalseBB
1404  // CmpBB:
1405  // Cmp = icmp Pred X, C
1407  ConstantRange DominatingCR =
1408  (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1410  CmpInst::getInversePredicate(DomPred), *DomC);
1411  ConstantRange Intersection = DominatingCR.intersectWith(CR);
1412  ConstantRange Difference = DominatingCR.difference(CR);
1413  if (Intersection.isEmptySet())
1414  return replaceInstUsesWith(Cmp, Builder.getFalse());
1415  if (Difference.isEmptySet())
1416  return replaceInstUsesWith(Cmp, Builder.getTrue());
1417 
1418  // Canonicalizing a sign bit comparison that gets used in a branch,
1419  // pessimizes codegen by generating branch on zero instruction instead
1420  // of a test and branch. So we avoid canonicalizing in such situations
1421  // because test and branch instruction has better branch displacement
1422  // than compare and branch instruction.
1423  bool UnusedBit;
1424  bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1425  if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1426  return nullptr;
1427 
1428  if (const APInt *EqC = Intersection.getSingleElement())
1429  return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1430  if (const APInt *NeC = Difference.getSingleElement())
1431  return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1432  }
1433 
1434  return nullptr;
1435 }
1436 
1437 /// Fold icmp (trunc X, Y), C.
1438 Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
1439  TruncInst *Trunc,
1440  const APInt &C) {
1441  ICmpInst::Predicate Pred = Cmp.getPredicate();
1442  Value *X = Trunc->getOperand(0);
1443  if (C.isOneValue() && C.getBitWidth() > 1) {
1444  // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1445  Value *V = nullptr;
1446  if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1447  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1448  ConstantInt::get(V->getType(), 1));
1449  }
1450 
1451  if (Cmp.isEquality() && Trunc->hasOneUse()) {
1452  // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1453  // of the high bits truncated out of x are known.
1454  unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1455  SrcBits = X->getType()->getScalarSizeInBits();
1456  KnownBits Known = computeKnownBits(X, 0, &Cmp);
1457 
1458  // If all the high bits are known, we can do this xform.
1459  if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1460  // Pull in the high bits from known-ones set.
1461  APInt NewRHS = C.zext(SrcBits);
1462  NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1463  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1464  }
1465  }
1466 
1467  return nullptr;
1468 }
1469 
1470 /// Fold icmp (xor X, Y), C.
1471 Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp,
1472  BinaryOperator *Xor,
1473  const APInt &C) {
1474  Value *X = Xor->getOperand(0);
1475  Value *Y = Xor->getOperand(1);
1476  const APInt *XorC;
1477  if (!match(Y, m_APInt(XorC)))
1478  return nullptr;
1479 
1480  // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1481  // fold the xor.
1482  ICmpInst::Predicate Pred = Cmp.getPredicate();
1483  bool TrueIfSigned = false;
1484  if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1485 
1486  // If the sign bit of the XorCst is not set, there is no change to
1487  // the operation, just stop using the Xor.
1488  if (!XorC->isNegative()) {
1489  Cmp.setOperand(0, X);
1490  Worklist.Add(Xor);
1491  return &Cmp;
1492  }
1493 
1494  // Emit the opposite comparison.
1495  if (TrueIfSigned)
1496  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1498  else
1499  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1501  }
1502 
1503  if (Xor->hasOneUse()) {
1504  // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1505  if (!Cmp.isEquality() && XorC->isSignMask()) {
1506  Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1507  : Cmp.getSignedPredicate();
1508  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1509  }
1510 
1511  // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1512  if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1513  Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1514  : Cmp.getSignedPredicate();
1515  Pred = Cmp.getSwappedPredicate(Pred);
1516  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1517  }
1518  }
1519 
1520  // Mask constant magic can eliminate an 'xor' with unsigned compares.
1521  if (Pred == ICmpInst::ICMP_UGT) {
1522  // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1523  if (*XorC == ~C && (C + 1).isPowerOf2())
1524  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1525  // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1526  if (*XorC == C && (C + 1).isPowerOf2())
1527  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1528  }
1529  if (Pred == ICmpInst::ICMP_ULT) {
1530  // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1531  if (*XorC == -C && C.isPowerOf2())
1532  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1533  ConstantInt::get(X->getType(), ~C));
1534  // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1535  if (*XorC == C && (-C).isPowerOf2())
1536  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1537  ConstantInt::get(X->getType(), ~C));
1538  }
1539  return nullptr;
1540 }
1541 
1542 /// Fold icmp (and (sh X, Y), C2), C1.
1543 Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
1544  const APInt &C1, const APInt &C2) {
1545  BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1546  if (!Shift || !Shift->isShift())
1547  return nullptr;
1548 
1549  // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1550  // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1551  // code produced by the clang front-end, for bitfield access.
1552  // This seemingly simple opportunity to fold away a shift turns out to be
1553  // rather complicated. See PR17827 for details.
1554  unsigned ShiftOpcode = Shift->getOpcode();
1555  bool IsShl = ShiftOpcode == Instruction::Shl;
1556  const APInt *C3;
1557  if (match(Shift->getOperand(1), m_APInt(C3))) {
1558  bool CanFold = false;
1559  if (ShiftOpcode == Instruction::Shl) {
1560  // For a left shift, we can fold if the comparison is not signed. We can
1561  // also fold a signed comparison if the mask value and comparison value
1562  // are not negative. These constraints may not be obvious, but we can
1563  // prove that they are correct using an SMT solver.
1564  if (!Cmp.isSigned() || (!C2.isNegative() && !C1.isNegative()))
1565  CanFold = true;
1566  } else {
1567  bool IsAshr = ShiftOpcode == Instruction::AShr;
1568  // For a logical right shift, we can fold if the comparison is not signed.
1569  // We can also fold a signed comparison if the shifted mask value and the
1570  // shifted comparison value are not negative. These constraints may not be
1571  // obvious, but we can prove that they are correct using an SMT solver.
1572  // For an arithmetic shift right we can do the same, if we ensure
1573  // the And doesn't use any bits being shifted in. Normally these would
1574  // be turned into lshr by SimplifyDemandedBits, but not if there is an
1575  // additional user.
1576  if (!IsAshr || (C2.shl(*C3).lshr(*C3) == C2)) {
1577  if (!Cmp.isSigned() ||
1578  (!C2.shl(*C3).isNegative() && !C1.shl(*C3).isNegative()))
1579  CanFold = true;
1580  }
1581  }
1582 
1583  if (CanFold) {
1584  APInt NewCst = IsShl ? C1.lshr(*C3) : C1.shl(*C3);
1585  APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3);
1586  // Check to see if we are shifting out any of the bits being compared.
1587  if (SameAsC1 != C1) {
1588  // If we shifted bits out, the fold is not going to work out. As a
1589  // special case, check to see if this means that the result is always
1590  // true or false now.
1591  if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1592  return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1593  if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1594  return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1595  } else {
1596  Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst));
1597  APInt NewAndCst = IsShl ? C2.lshr(*C3) : C2.shl(*C3);
1598  And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst));
1599  And->setOperand(0, Shift->getOperand(0));
1600  Worklist.Add(Shift); // Shift is dead.
1601  return &Cmp;
1602  }
1603  }
1604  }
1605 
1606  // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1607  // preferable because it allows the C2 << Y expression to be hoisted out of a
1608  // loop if Y is invariant and X is not.
1609  if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1610  !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1611  // Compute C2 << Y.
1612  Value *NewShift =
1613  IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1614  : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1615 
1616  // Compute X & (C2 << Y).
1617  Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1618  Cmp.setOperand(0, NewAnd);
1619  return &Cmp;
1620  }
1621 
1622  return nullptr;
1623 }
1624 
1625 /// Fold icmp (and X, C2), C1.
1626 Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp,
1627  BinaryOperator *And,
1628  const APInt &C1) {
1629  bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1630 
1631  // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1632  // TODO: We canonicalize to the longer form for scalars because we have
1633  // better analysis/folds for icmp, and codegen may be better with icmp.
1634  if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isNullValue() &&
1635  match(And->getOperand(1), m_One()))
1636  return new TruncInst(And->getOperand(0), Cmp.getType());
1637 
1638  const APInt *C2;
1639  Value *X;
1640  if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1641  return nullptr;
1642 
1643  // Don't perform the following transforms if the AND has multiple uses
1644  if (!And->hasOneUse())
1645  return nullptr;
1646 
1647  if (Cmp.isEquality() && C1.isNullValue()) {
1648  // Restrict this fold to single-use 'and' (PR10267).
1649  // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1650  if (C2->isSignMask()) {
1651  Constant *Zero = Constant::getNullValue(X->getType());
1652  auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1653  return new ICmpInst(NewPred, X, Zero);
1654  }
1655  }
1656 
1657  // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1658  // the input width without changing the value produced, eliminate the cast:
1659  //
1660  // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1661  //
1662  // We can do this transformation if the constants do not have their sign bits
1663  // set or if it is an equality comparison. Extending a relational comparison
1664  // when we're checking the sign bit would not work.
1665  Value *W;
1666  if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1667  (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1668  // TODO: Is this a good transform for vectors? Wider types may reduce
1669  // throughput. Should this transform be limited (even for scalars) by using
1670  // shouldChangeType()?
1671  if (!Cmp.getType()->isVectorTy()) {
1672  Type *WideType = W->getType();
1673  unsigned WideScalarBits = WideType->getScalarSizeInBits();
1674  Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1675  Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1676  Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1677  return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1678  }
1679  }
1680 
1681  if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1682  return I;
1683 
1684  // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1685  // (icmp pred (and A, (or (shl 1, B), 1), 0))
1686  //
1687  // iff pred isn't signed
1688  if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1689  match(And->getOperand(1), m_One())) {
1690  Constant *One = cast<Constant>(And->getOperand(1));
1691  Value *Or = And->getOperand(0);
1692  Value *A, *B, *LShr;
1693  if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1694  match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1695  unsigned UsesRemoved = 0;
1696  if (And->hasOneUse())
1697  ++UsesRemoved;
1698  if (Or->hasOneUse())
1699  ++UsesRemoved;
1700  if (LShr->hasOneUse())
1701  ++UsesRemoved;
1702 
1703  // Compute A & ((1 << B) | 1)
1704  Value *NewOr = nullptr;
1705  if (auto *C = dyn_cast<Constant>(B)) {
1706  if (UsesRemoved >= 1)
1707  NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1708  } else {
1709  if (UsesRemoved >= 3)
1710  NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1711  /*HasNUW=*/true),
1712  One, Or->getName());
1713  }
1714  if (NewOr) {
1715  Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1716  Cmp.setOperand(0, NewAnd);
1717  return &Cmp;
1718  }
1719  }
1720  }
1721 
1722  return nullptr;
1723 }
1724 
1725 /// Fold icmp (and X, Y), C.
1726 Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp,
1727  BinaryOperator *And,
1728  const APInt &C) {
1729  if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1730  return I;
1731 
1732  // TODO: These all require that Y is constant too, so refactor with the above.
1733 
1734  // Try to optimize things like "A[i] & 42 == 0" to index computations.
1735  Value *X = And->getOperand(0);
1736  Value *Y = And->getOperand(1);
1737  if (auto *LI = dyn_cast<LoadInst>(X))
1738  if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1739  if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1740  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1741  !LI->isVolatile() && isa<ConstantInt>(Y)) {
1742  ConstantInt *C2 = cast<ConstantInt>(Y);
1743  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1744  return Res;
1745  }
1746 
1747  if (!Cmp.isEquality())
1748  return nullptr;
1749 
1750  // X & -C == -C -> X > u ~C
1751  // X & -C != -C -> X <= u ~C
1752  // iff C is a power of 2
1753  if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1754  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1756  return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1757  }
1758 
1759  // (X & C2) == 0 -> (trunc X) >= 0
1760  // (X & C2) != 0 -> (trunc X) < 0
1761  // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1762  const APInt *C2;
1763  if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1764  int32_t ExactLogBase2 = C2->exactLogBase2();
1765  if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1766  Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1767  if (And->getType()->isVectorTy())
1768  NTy = VectorType::get(NTy, And->getType()->getVectorNumElements());
1769  Value *Trunc = Builder.CreateTrunc(X, NTy);
1770  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1772  return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1773  }
1774  }
1775 
1776  return nullptr;
1777 }
1778 
1779 /// Fold icmp (or X, Y), C.
1780 Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
1781  const APInt &C) {
1782  ICmpInst::Predicate Pred = Cmp.getPredicate();
1783  if (C.isOneValue()) {
1784  // icmp slt signum(V) 1 --> icmp slt V, 1
1785  Value *V = nullptr;
1786  if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1787  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1788  ConstantInt::get(V->getType(), 1));
1789  }
1790 
1791  Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1792  if (Cmp.isEquality() && Cmp.getOperand(1) == OrOp1) {
1793  // X | C == C --> X <=u C
1794  // X | C != C --> X >u C
1795  // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1796  if ((C + 1).isPowerOf2()) {
1798  return new ICmpInst(Pred, OrOp0, OrOp1);
1799  }
1800  // More general: are all bits outside of a mask constant set or not set?
1801  // X | C == C --> (X & ~C) == 0
1802  // X | C != C --> (X & ~C) != 0
1803  if (Or->hasOneUse()) {
1804  Value *A = Builder.CreateAnd(OrOp0, ~C);
1805  return new ICmpInst(Pred, A, ConstantInt::getNullValue(OrOp0->getType()));
1806  }
1807  }
1808 
1809  if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1810  return nullptr;
1811 
1812  Value *P, *Q;
1813  if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1814  // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1815  // -> and (icmp eq P, null), (icmp eq Q, null).
1816  Value *CmpP =
1817  Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1818  Value *CmpQ =
1819  Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1820  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1821  return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1822  }
1823 
1824  // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1825  // a shorter form that has more potential to be folded even further.
1826  Value *X1, *X2, *X3, *X4;
1827  if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1828  match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1829  // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1830  // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1831  Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1832  Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1833  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1834  return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1835  }
1836 
1837  return nullptr;
1838 }
1839 
1840 /// Fold icmp (mul X, Y), C.
1841 Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp,
1842  BinaryOperator *Mul,
1843  const APInt &C) {
1844  const APInt *MulC;
1845  if (!match(Mul->getOperand(1), m_APInt(MulC)))
1846  return nullptr;
1847 
1848  // If this is a test of the sign bit and the multiply is sign-preserving with
1849  // a constant operand, use the multiply LHS operand instead.
1850  ICmpInst::Predicate Pred = Cmp.getPredicate();
1851  if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1852  if (MulC->isNegative())
1853  Pred = ICmpInst::getSwappedPredicate(Pred);
1854  return new ICmpInst(Pred, Mul->getOperand(0),
1856  }
1857 
1858  return nullptr;
1859 }
1860 
1861 /// Fold icmp (shl 1, Y), C.
1863  const APInt &C) {
1864  Value *Y;
1865  if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1866  return nullptr;
1867 
1868  Type *ShiftType = Shl->getType();
1869  unsigned TypeBits = C.getBitWidth();
1870  bool CIsPowerOf2 = C.isPowerOf2();
1871  ICmpInst::Predicate Pred = Cmp.getPredicate();
1872  if (Cmp.isUnsigned()) {
1873  // (1 << Y) pred C -> Y pred Log2(C)
1874  if (!CIsPowerOf2) {
1875  // (1 << Y) < 30 -> Y <= 4
1876  // (1 << Y) <= 30 -> Y <= 4
1877  // (1 << Y) >= 30 -> Y > 4
1878  // (1 << Y) > 30 -> Y > 4
1879  if (Pred == ICmpInst::ICMP_ULT)
1880  Pred = ICmpInst::ICMP_ULE;
1881  else if (Pred == ICmpInst::ICMP_UGE)
1882  Pred = ICmpInst::ICMP_UGT;
1883  }
1884 
1885  // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
1886  // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
1887  unsigned CLog2 = C.logBase2();
1888  if (CLog2 == TypeBits - 1) {
1889  if (Pred == ICmpInst::ICMP_UGE)
1890  Pred = ICmpInst::ICMP_EQ;
1891  else if (Pred == ICmpInst::ICMP_ULT)
1892  Pred = ICmpInst::ICMP_NE;
1893  }
1894  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
1895  } else if (Cmp.isSigned()) {
1896  Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
1897  if (C.isAllOnesValue()) {
1898  // (1 << Y) <= -1 -> Y == 31
1899  if (Pred == ICmpInst::ICMP_SLE)
1900  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1901 
1902  // (1 << Y) > -1 -> Y != 31
1903  if (Pred == ICmpInst::ICMP_SGT)
1904  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1905  } else if (!C) {
1906  // (1 << Y) < 0 -> Y == 31
1907  // (1 << Y) <= 0 -> Y == 31
1908  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1909  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1910 
1911  // (1 << Y) >= 0 -> Y != 31
1912  // (1 << Y) > 0 -> Y != 31
1913  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1914  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1915  }
1916  } else if (Cmp.isEquality() && CIsPowerOf2) {
1917  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
1918  }
1919 
1920  return nullptr;
1921 }
1922 
1923 /// Fold icmp (shl X, Y), C.
1924 Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp,
1925  BinaryOperator *Shl,
1926  const APInt &C) {
1927  const APInt *ShiftVal;
1928  if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
1929  return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
1930 
1931  const APInt *ShiftAmt;
1932  if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
1933  return foldICmpShlOne(Cmp, Shl, C);
1934 
1935  // Check that the shift amount is in range. If not, don't perform undefined
1936  // shifts. When the shift is visited, it will be simplified.
1937  unsigned TypeBits = C.getBitWidth();
1938  if (ShiftAmt->uge(TypeBits))
1939  return nullptr;
1940 
1941  ICmpInst::Predicate Pred = Cmp.getPredicate();
1942  Value *X = Shl->getOperand(0);
1943  Type *ShType = Shl->getType();
1944 
1945  // NSW guarantees that we are only shifting out sign bits from the high bits,
1946  // so we can ASHR the compare constant without needing a mask and eliminate
1947  // the shift.
1948  if (Shl->hasNoSignedWrap()) {
1949  if (Pred == ICmpInst::ICMP_SGT) {
1950  // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
1951  APInt ShiftedC = C.ashr(*ShiftAmt);
1952  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1953  }
1954  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
1955  C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
1956  APInt ShiftedC = C.ashr(*ShiftAmt);
1957  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1958  }
1959  if (Pred == ICmpInst::ICMP_SLT) {
1960  // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
1961  // (X << S) <=s C is equiv to X <=s (C >> S) for all C
1962  // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
1963  // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
1964  assert(!C.isMinSignedValue() && "Unexpected icmp slt");
1965  APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
1966  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1967  }
1968  // If this is a signed comparison to 0 and the shift is sign preserving,
1969  // use the shift LHS operand instead; isSignTest may change 'Pred', so only
1970  // do that if we're sure to not continue on in this function.
1971  if (isSignTest(Pred, C))
1972  return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
1973  }
1974 
1975  // NUW guarantees that we are only shifting out zero bits from the high bits,
1976  // so we can LSHR the compare constant without needing a mask and eliminate
1977  // the shift.
1978  if (Shl->hasNoUnsignedWrap()) {
1979  if (Pred == ICmpInst::ICMP_UGT) {
1980  // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
1981  APInt ShiftedC = C.lshr(*ShiftAmt);
1982  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1983  }
1984  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
1985  C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
1986  APInt ShiftedC = C.lshr(*ShiftAmt);
1987  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1988  }
1989  if (Pred == ICmpInst::ICMP_ULT) {
1990  // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
1991  // (X << S) <=u C is equiv to X <=u (C >> S) for all C
1992  // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
1993  // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
1994  assert(C.ugt(0) && "ult 0 should have been eliminated");
1995  APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
1996  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1997  }
1998  }
1999 
2000  if (Cmp.isEquality() && Shl->hasOneUse()) {
2001  // Strength-reduce the shift into an 'and'.
2003  ShType,
2004  APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2005  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2006  Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2007  return new ICmpInst(Pred, And, LShrC);
2008  }
2009 
2010  // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2011  bool TrueIfSigned = false;
2012  if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2013  // (X << 31) <s 0 --> (X & 1) != 0
2015  ShType,
2016  APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2017  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2018  return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2019  And, Constant::getNullValue(ShType));
2020  }
2021 
2022  // Transform (icmp pred iM (shl iM %v, N), C)
2023  // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2024  // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2025  // This enables us to get rid of the shift in favor of a trunc that may be
2026  // free on the target. It has the additional benefit of comparing to a
2027  // smaller constant that may be more target-friendly.
2028  unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2029  if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2030  DL.isLegalInteger(TypeBits - Amt)) {
2031  Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2032  if (ShType->isVectorTy())
2033  TruncTy = VectorType::get(TruncTy, ShType->getVectorNumElements());
2034  Constant *NewC =
2035  ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2036  return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2037  }
2038 
2039  return nullptr;
2040 }
2041 
2042 /// Fold icmp ({al}shr X, Y), C.
2043 Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp,
2044  BinaryOperator *Shr,
2045  const APInt &C) {
2046  // An exact shr only shifts out zero bits, so:
2047  // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2048  Value *X = Shr->getOperand(0);
2049  CmpInst::Predicate Pred = Cmp.getPredicate();
2050  if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2051  C.isNullValue())
2052  return new ICmpInst(Pred, X, Cmp.getOperand(1));
2053 
2054  const APInt *ShiftVal;
2055  if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2056  return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2057 
2058  const APInt *ShiftAmt;
2059  if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2060  return nullptr;
2061 
2062  // Check that the shift amount is in range. If not, don't perform undefined
2063  // shifts. When the shift is visited it will be simplified.
2064  unsigned TypeBits = C.getBitWidth();
2065  unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2066  if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2067  return nullptr;
2068 
2069  bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2070  bool IsExact = Shr->isExact();
2071  Type *ShrTy = Shr->getType();
2072  // TODO: If we could guarantee that InstSimplify would handle all of the
2073  // constant-value-based preconditions in the folds below, then we could assert
2074  // those conditions rather than checking them. This is difficult because of
2075  // undef/poison (PR34838).
2076  if (IsAShr) {
2077  if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2078  // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2079  // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2080  APInt ShiftedC = C.shl(ShAmtVal);
2081  if (ShiftedC.ashr(ShAmtVal) == C)
2082  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2083  }
2084  if (Pred == CmpInst::ICMP_SGT) {
2085  // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2086  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2087  if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2088  (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2089  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2090  }
2091  } else {
2092  if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2093  // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2094  // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2095  APInt ShiftedC = C.shl(ShAmtVal);
2096  if (ShiftedC.lshr(ShAmtVal) == C)
2097  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2098  }
2099  if (Pred == CmpInst::ICMP_UGT) {
2100  // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2101  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2102  if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2103  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2104  }
2105  }
2106 
2107  if (!Cmp.isEquality())
2108  return nullptr;
2109 
2110  // Handle equality comparisons of shift-by-constant.
2111 
2112  // If the comparison constant changes with the shift, the comparison cannot
2113  // succeed (bits of the comparison constant cannot match the shifted value).
2114  // This should be known by InstSimplify and already be folded to true/false.
2115  assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2116  (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2117  "Expected icmp+shr simplify did not occur.");
2118 
2119  // If the bits shifted out are known zero, compare the unshifted value:
2120  // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2121  if (Shr->isExact())
2122  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2123 
2124  if (Shr->hasOneUse()) {
2125  // Canonicalize the shift into an 'and':
2126  // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2127  APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2128  Constant *Mask = ConstantInt::get(ShrTy, Val);
2129  Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2130  return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2131  }
2132 
2133  return nullptr;
2134 }
2135 
2136 /// Fold icmp (udiv X, Y), C.
2137 Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp,
2138  BinaryOperator *UDiv,
2139  const APInt &C) {
2140  const APInt *C2;
2141  if (!match(UDiv->getOperand(0), m_APInt(C2)))
2142  return nullptr;
2143 
2144  assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2145 
2146  // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2147  Value *Y = UDiv->getOperand(1);
2148  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2149  assert(!C.isMaxValue() &&
2150  "icmp ugt X, UINT_MAX should have been simplified already.");
2151  return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2152  ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2153  }
2154 
2155  // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2156  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2157  assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2158  return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2159  ConstantInt::get(Y->getType(), C2->udiv(C)));
2160  }
2161 
2162  return nullptr;
2163 }
2164 
2165 /// Fold icmp ({su}div X, Y), C.
2166 Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp,
2167  BinaryOperator *Div,
2168  const APInt &C) {
2169  // Fold: icmp pred ([us]div X, C2), C -> range test
2170  // Fold this div into the comparison, producing a range check.
2171  // Determine, based on the divide type, what the range is being
2172  // checked. If there is an overflow on the low or high side, remember
2173  // it, otherwise compute the range [low, hi) bounding the new value.
2174  // See: InsertRangeTest above for the kinds of replacements possible.
2175  const APInt *C2;
2176  if (!match(Div->getOperand(1), m_APInt(C2)))
2177  return nullptr;
2178 
2179  // FIXME: If the operand types don't match the type of the divide
2180  // then don't attempt this transform. The code below doesn't have the
2181  // logic to deal with a signed divide and an unsigned compare (and
2182  // vice versa). This is because (x /s C2) <s C produces different
2183  // results than (x /s C2) <u C or (x /u C2) <s C or even
2184  // (x /u C2) <u C. Simply casting the operands and result won't
2185  // work. :( The if statement below tests that condition and bails
2186  // if it finds it.
2187  bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2188  if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2189  return nullptr;
2190 
2191  // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2192  // INT_MIN will also fail if the divisor is 1. Although folds of all these
2193  // division-by-constant cases should be present, we can not assert that they
2194  // have happened before we reach this icmp instruction.
2195  if (C2->isNullValue() || C2->isOneValue() ||
2196  (DivIsSigned && C2->isAllOnesValue()))
2197  return nullptr;
2198 
2199  // Compute Prod = C * C2. We are essentially solving an equation of
2200  // form X / C2 = C. We solve for X by multiplying C2 and C.
2201  // By solving for X, we can turn this into a range check instead of computing
2202  // a divide.
2203  APInt Prod = C * *C2;
2204 
2205  // Determine if the product overflows by seeing if the product is not equal to
2206  // the divide. Make sure we do the same kind of divide as in the LHS
2207  // instruction that we're folding.
2208  bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2209 
2210  ICmpInst::Predicate Pred = Cmp.getPredicate();
2211 
2212  // If the division is known to be exact, then there is no remainder from the
2213  // divide, so the covered range size is unit, otherwise it is the divisor.
2214  APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2215 
2216  // Figure out the interval that is being checked. For example, a comparison
2217  // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2218  // Compute this interval based on the constants involved and the signedness of
2219  // the compare/divide. This computes a half-open interval, keeping track of
2220  // whether either value in the interval overflows. After analysis each
2221  // overflow variable is set to 0 if it's corresponding bound variable is valid
2222  // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2223  int LoOverflow = 0, HiOverflow = 0;
2224  APInt LoBound, HiBound;
2225 
2226  if (!DivIsSigned) { // udiv
2227  // e.g. X/5 op 3 --> [15, 20)
2228  LoBound = Prod;
2229  HiOverflow = LoOverflow = ProdOV;
2230  if (!HiOverflow) {
2231  // If this is not an exact divide, then many values in the range collapse
2232  // to the same result value.
2233  HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2234  }
2235  } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2236  if (C.isNullValue()) { // (X / pos) op 0
2237  // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2238  LoBound = -(RangeSize - 1);
2239  HiBound = RangeSize;
2240  } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2241  LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2242  HiOverflow = LoOverflow = ProdOV;
2243  if (!HiOverflow)
2244  HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2245  } else { // (X / pos) op neg
2246  // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2247  HiBound = Prod + 1;
2248  LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2249  if (!LoOverflow) {
2250  APInt DivNeg = -RangeSize;
2251  LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2252  }
2253  }
2254  } else if (C2->isNegative()) { // Divisor is < 0.
2255  if (Div->isExact())
2256  RangeSize.negate();
2257  if (C.isNullValue()) { // (X / neg) op 0
2258  // e.g. X/-5 op 0 --> [-4, 5)
2259  LoBound = RangeSize + 1;
2260  HiBound = -RangeSize;
2261  if (HiBound == *C2) { // -INTMIN = INTMIN
2262  HiOverflow = 1; // [INTMIN+1, overflow)
2263  HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2264  }
2265  } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2266  // e.g. X/-5 op 3 --> [-19, -14)
2267  HiBound = Prod + 1;
2268  HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2269  if (!LoOverflow)
2270  LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2271  } else { // (X / neg) op neg
2272  LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2273  LoOverflow = HiOverflow = ProdOV;
2274  if (!HiOverflow)
2275  HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2276  }
2277 
2278  // Dividing by a negative swaps the condition. LT <-> GT
2279  Pred = ICmpInst::getSwappedPredicate(Pred);
2280  }
2281 
2282  Value *X = Div->getOperand(0);
2283  switch (Pred) {
2284  default: llvm_unreachable("Unhandled icmp opcode!");
2285  case ICmpInst::ICMP_EQ:
2286  if (LoOverflow && HiOverflow)
2287  return replaceInstUsesWith(Cmp, Builder.getFalse());
2288  if (HiOverflow)
2289  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2290  ICmpInst::ICMP_UGE, X,
2291  ConstantInt::get(Div->getType(), LoBound));
2292  if (LoOverflow)
2293  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2294  ICmpInst::ICMP_ULT, X,
2295  ConstantInt::get(Div->getType(), HiBound));
2296  return replaceInstUsesWith(
2297  Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2298  case ICmpInst::ICMP_NE:
2299  if (LoOverflow && HiOverflow)
2300  return replaceInstUsesWith(Cmp, Builder.getTrue());
2301  if (HiOverflow)
2302  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2303  ICmpInst::ICMP_ULT, X,
2304  ConstantInt::get(Div->getType(), LoBound));
2305  if (LoOverflow)
2306  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2307  ICmpInst::ICMP_UGE, X,
2308  ConstantInt::get(Div->getType(), HiBound));
2309  return replaceInstUsesWith(Cmp,
2310  insertRangeTest(X, LoBound, HiBound,
2311  DivIsSigned, false));
2312  case ICmpInst::ICMP_ULT:
2313  case ICmpInst::ICMP_SLT:
2314  if (LoOverflow == +1) // Low bound is greater than input range.
2315  return replaceInstUsesWith(Cmp, Builder.getTrue());
2316  if (LoOverflow == -1) // Low bound is less than input range.
2317  return replaceInstUsesWith(Cmp, Builder.getFalse());
2318  return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2319  case ICmpInst::ICMP_UGT:
2320  case ICmpInst::ICMP_SGT:
2321  if (HiOverflow == +1) // High bound greater than input range.
2322  return replaceInstUsesWith(Cmp, Builder.getFalse());
2323  if (HiOverflow == -1) // High bound less than input range.
2324  return replaceInstUsesWith(Cmp, Builder.getTrue());
2325  if (Pred == ICmpInst::ICMP_UGT)
2326  return new ICmpInst(ICmpInst::ICMP_UGE, X,
2327  ConstantInt::get(Div->getType(), HiBound));
2328  return new ICmpInst(ICmpInst::ICMP_SGE, X,
2329  ConstantInt::get(Div->getType(), HiBound));
2330  }
2331 
2332  return nullptr;
2333 }
2334 
2335 /// Fold icmp (sub X, Y), C.
2336 Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp,
2337  BinaryOperator *Sub,
2338  const APInt &C) {
2339  Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2340  ICmpInst::Predicate Pred = Cmp.getPredicate();
2341  const APInt *C2;
2342  APInt SubResult;
2343 
2344  // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2345  if (match(X, m_APInt(C2)) &&
2346  ((Cmp.isUnsigned() && Sub->hasNoUnsignedWrap()) ||
2347  (Cmp.isSigned() && Sub->hasNoSignedWrap())) &&
2348  !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2349  return new ICmpInst(Cmp.getSwappedPredicate(), Y,
2350  ConstantInt::get(Y->getType(), SubResult));
2351 
2352  // The following transforms are only worth it if the only user of the subtract
2353  // is the icmp.
2354  if (!Sub->hasOneUse())
2355  return nullptr;
2356 
2357  if (Sub->hasNoSignedWrap()) {
2358  // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2359  if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2360  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2361 
2362  // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2363  if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2364  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2365 
2366  // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2367  if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2368  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2369 
2370  // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2371  if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2372  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2373  }
2374 
2375  if (!match(X, m_APInt(C2)))
2376  return nullptr;
2377 
2378  // C2 - Y <u C -> (Y | (C - 1)) == C2
2379  // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2380  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2381  (*C2 & (C - 1)) == (C - 1))
2382  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2383 
2384  // C2 - Y >u C -> (Y | C) != C2
2385  // iff C2 & C == C and C + 1 is a power of 2
2386  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2387  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2388 
2389  return nullptr;
2390 }
2391 
2392 /// Fold icmp (add X, Y), C.
2393 Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp,
2395  const APInt &C) {
2396  Value *Y = Add->getOperand(1);
2397  const APInt *C2;
2398  if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2399  return nullptr;
2400 
2401  // Fold icmp pred (add X, C2), C.
2402  Value *X = Add->getOperand(0);
2403  Type *Ty = Add->getType();
2404  CmpInst::Predicate Pred = Cmp.getPredicate();
2405 
2406  if (!Add->hasOneUse())
2407  return nullptr;
2408 
2409  // If the add does not wrap, we can always adjust the compare by subtracting
2410  // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2411  // are canonicalized to SGT/SLT/UGT/ULT.
2412  if ((Add->hasNoSignedWrap() &&
2413  (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2414  (Add->hasNoUnsignedWrap() &&
2415  (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2416  bool Overflow;
2417  APInt NewC =
2418  Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2419  // If there is overflow, the result must be true or false.
2420  // TODO: Can we assert there is no overflow because InstSimplify always
2421  // handles those cases?
2422  if (!Overflow)
2423  // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2424  return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2425  }
2426 
2427  auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2428  const APInt &Upper = CR.getUpper();
2429  const APInt &Lower = CR.getLower();
2430  if (Cmp.isSigned()) {
2431  if (Lower.isSignMask())
2432  return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2433  if (Upper.isSignMask())
2434  return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2435  } else {
2436  if (Lower.isMinValue())
2437  return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2438  if (Upper.isMinValue())
2439  return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2440  }
2441 
2442  // X+C <u C2 -> (X & -C2) == C
2443  // iff C & (C2-1) == 0
2444  // C2 is a power of 2
2445  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2446  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2447  ConstantExpr::getNeg(cast<Constant>(Y)));
2448 
2449  // X+C >u C2 -> (X & ~C2) != C
2450  // iff C & C2 == 0
2451  // C2+1 is a power of 2
2452  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2453  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2454  ConstantExpr::getNeg(cast<Constant>(Y)));
2455 
2456  return nullptr;
2457 }
2458 
2459 bool InstCombiner::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2460  Value *&RHS, ConstantInt *&Less,
2461  ConstantInt *&Equal,
2462  ConstantInt *&Greater) {
2463  // TODO: Generalize this to work with other comparison idioms or ensure
2464  // they get canonicalized into this form.
2465 
2466  // select i1 (a == b), i32 Equal, i32 (select i1 (a < b), i32 Less, i32
2467  // Greater), where Equal, Less and Greater are placeholders for any three
2468  // constants.
2469  ICmpInst::Predicate PredA, PredB;
2470  if (match(SI->getTrueValue(), m_ConstantInt(Equal)) &&
2471  match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) &&
2472  PredA == ICmpInst::ICMP_EQ &&
2473  match(SI->getFalseValue(),
2474  m_Select(m_ICmp(PredB, m_Specific(LHS), m_Specific(RHS)),
2475  m_ConstantInt(Less), m_ConstantInt(Greater))) &&
2476  PredB == ICmpInst::ICMP_SLT) {
2477  return true;
2478  }
2479  return false;
2480 }
2481 
2482 Instruction *InstCombiner::foldICmpSelectConstant(ICmpInst &Cmp,
2483  SelectInst *Select,
2484  ConstantInt *C) {
2485 
2486  assert(C && "Cmp RHS should be a constant int!");
2487  // If we're testing a constant value against the result of a three way
2488  // comparison, the result can be expressed directly in terms of the
2489  // original values being compared. Note: We could possibly be more
2490  // aggressive here and remove the hasOneUse test. The original select is
2491  // really likely to simplify or sink when we remove a test of the result.
2492  Value *OrigLHS, *OrigRHS;
2493  ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2494  if (Cmp.hasOneUse() &&
2495  matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2496  C3GreaterThan)) {
2497  assert(C1LessThan && C2Equal && C3GreaterThan);
2498 
2499  bool TrueWhenLessThan =
2500  ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2501  ->isAllOnesValue();
2502  bool TrueWhenEqual =
2503  ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2504  ->isAllOnesValue();
2505  bool TrueWhenGreaterThan =
2506  ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2507  ->isAllOnesValue();
2508 
2509  // This generates the new instruction that will replace the original Cmp
2510  // Instruction. Instead of enumerating the various combinations when
2511  // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2512  // false, we rely on chaining of ORs and future passes of InstCombine to
2513  // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2514 
2515  // When none of the three constants satisfy the predicate for the RHS (C),
2516  // the entire original Cmp can be simplified to a false.
2517  Value *Cond = Builder.getFalse();
2518  if (TrueWhenLessThan)
2519  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2520  OrigLHS, OrigRHS));
2521  if (TrueWhenEqual)
2522  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2523  OrigLHS, OrigRHS));
2524  if (TrueWhenGreaterThan)
2525  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2526  OrigLHS, OrigRHS));
2527 
2528  return replaceInstUsesWith(Cmp, Cond);
2529  }
2530  return nullptr;
2531 }
2532 
2534  InstCombiner::BuilderTy &Builder) {
2535  auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2536  if (!Bitcast)
2537  return nullptr;
2538 
2539  ICmpInst::Predicate Pred = Cmp.getPredicate();
2540  Value *Op1 = Cmp.getOperand(1);
2541  Value *BCSrcOp = Bitcast->getOperand(0);
2542 
2543  // Make sure the bitcast doesn't change the number of vector elements.
2544  if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2545  Bitcast->getDestTy()->getScalarSizeInBits()) {
2546  // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2547  Value *X;
2548  if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2549  // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2550  // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2551  // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2552  // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2553  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2554  Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2555  match(Op1, m_Zero()))
2556  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2557 
2558  // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2559  if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2560  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2561 
2562  // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2563  if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2564  return new ICmpInst(Pred, X,
2566  }
2567 
2568  // Zero-equality checks are preserved through unsigned floating-point casts:
2569  // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2570  // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2571  if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2572  if (Cmp.isEquality() && match(Op1, m_Zero()))
2573  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2574  }
2575 
2576  // Test to see if the operands of the icmp are casted versions of other
2577  // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2578  if (Bitcast->getType()->isPointerTy() &&
2579  (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2580  // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2581  // so eliminate it as well.
2582  if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2583  Op1 = BC2->getOperand(0);
2584 
2585  Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2586  return new ICmpInst(Pred, BCSrcOp, Op1);
2587  }
2588 
2589  // Folding: icmp <pred> iN X, C
2590  // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2591  // and C is a splat of a K-bit pattern
2592  // and SC is a constant vector = <C', C', C', ..., C'>
2593  // Into:
2594  // %E = extractelement <M x iK> %vec, i32 C'
2595  // icmp <pred> iK %E, trunc(C)
2596  const APInt *C;
2597  if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2598  !Bitcast->getType()->isIntegerTy() ||
2599  !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2600  return nullptr;
2601 
2602  Value *Vec;
2603  Constant *Mask;
2604  if (match(BCSrcOp,
2605  m_ShuffleVector(m_Value(Vec), m_Undef(), m_Constant(Mask)))) {
2606  // Check whether every element of Mask is the same constant
2607  if (auto *Elem = dyn_cast_or_null<ConstantInt>(Mask->getSplatValue())) {
2608  auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2609  auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2610  if (C->isSplat(EltTy->getBitWidth())) {
2611  // Fold the icmp based on the value of C
2612  // If C is M copies of an iK sized bit pattern,
2613  // then:
2614  // => %E = extractelement <N x iK> %vec, i32 Elem
2615  // icmp <pred> iK %SplatVal, <pattern>
2616  Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2617  Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2618  return new ICmpInst(Pred, Extract, NewC);
2619  }
2620  }
2621  }
2622  return nullptr;
2623 }
2624 
2625 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2626 /// where X is some kind of instruction.
2627 Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) {
2628  const APInt *C;
2629  if (!match(Cmp.getOperand(1), m_APInt(C)))
2630  return nullptr;
2631 
2632  if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2633  switch (BO->getOpcode()) {
2634  case Instruction::Xor:
2635  if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2636  return I;
2637  break;
2638  case Instruction::And:
2639  if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2640  return I;
2641  break;
2642  case Instruction::Or:
2643  if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2644  return I;
2645  break;
2646  case Instruction::Mul:
2647  if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2648  return I;
2649  break;
2650  case Instruction::Shl:
2651  if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2652  return I;
2653  break;
2654  case Instruction::LShr:
2655  case Instruction::AShr:
2656  if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2657  return I;
2658  break;
2659  case Instruction::UDiv:
2660  if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2661  return I;
2663  case Instruction::SDiv:
2664  if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2665  return I;
2666  break;
2667  case Instruction::Sub:
2668  if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2669  return I;
2670  break;
2671  case Instruction::Add:
2672  if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2673  return I;
2674  break;
2675  default:
2676  break;
2677  }
2678  // TODO: These folds could be refactored to be part of the above calls.
2679  if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2680  return I;
2681  }
2682 
2683  // Match against CmpInst LHS being instructions other than binary operators.
2684 
2685  if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2686  // For now, we only support constant integers while folding the
2687  // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2688  // similar to the cases handled by binary ops above.
2689  if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2690  if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2691  return I;
2692  }
2693 
2694  if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2695  if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2696  return I;
2697  }
2698 
2699  if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
2700  if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
2701  return I;
2702 
2703  return nullptr;
2704 }
2705 
2706 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2707 /// icmp eq/ne BO, C.
2708 Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
2709  BinaryOperator *BO,
2710  const APInt &C) {
2711  // TODO: Some of these folds could work with arbitrary constants, but this
2712  // function is limited to scalar and vector splat constants.
2713  if (!Cmp.isEquality())
2714  return nullptr;
2715 
2716  ICmpInst::Predicate Pred = Cmp.getPredicate();
2717  bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2718  Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2719  Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2720 
2721  switch (BO->getOpcode()) {
2722  case Instruction::SRem:
2723  // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2724  if (C.isNullValue() && BO->hasOneUse()) {
2725  const APInt *BOC;
2726  if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2727  Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
2728  return new ICmpInst(Pred, NewRem,
2730  }
2731  }
2732  break;
2733  case Instruction::Add: {
2734  // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
2735  const APInt *BOC;
2736  if (match(BOp1, m_APInt(BOC))) {
2737  if (BO->hasOneUse()) {
2738  Constant *SubC = ConstantExpr::getSub(RHS, cast<Constant>(BOp1));
2739  return new ICmpInst(Pred, BOp0, SubC);
2740  }
2741  } else if (C.isNullValue()) {
2742  // Replace ((add A, B) != 0) with (A != -B) if A or B is
2743  // efficiently invertible, or if the add has just this one use.
2744  if (Value *NegVal = dyn_castNegVal(BOp1))
2745  return new ICmpInst(Pred, BOp0, NegVal);
2746  if (Value *NegVal = dyn_castNegVal(BOp0))
2747  return new ICmpInst(Pred, NegVal, BOp1);
2748  if (BO->hasOneUse()) {
2749  Value *Neg = Builder.CreateNeg(BOp1);
2750  Neg->takeName(BO);
2751  return new ICmpInst(Pred, BOp0, Neg);
2752  }
2753  }
2754  break;
2755  }
2756  case Instruction::Xor:
2757  if (BO->hasOneUse()) {
2758  if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2759  // For the xor case, we can xor two constants together, eliminating
2760  // the explicit xor.
2761  return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
2762  } else if (C.isNullValue()) {
2763  // Replace ((xor A, B) != 0) with (A != B)
2764  return new ICmpInst(Pred, BOp0, BOp1);
2765  }
2766  }
2767  break;
2768  case Instruction::Sub:
2769  if (BO->hasOneUse()) {
2770  const APInt *BOC;
2771  if (match(BOp0, m_APInt(BOC))) {
2772  // Replace ((sub BOC, B) != C) with (B != BOC-C).
2773  Constant *SubC = ConstantExpr::getSub(cast<Constant>(BOp0), RHS);
2774  return new ICmpInst(Pred, BOp1, SubC);
2775  } else if (C.isNullValue()) {
2776  // Replace ((sub A, B) != 0) with (A != B).
2777  return new ICmpInst(Pred, BOp0, BOp1);
2778  }
2779  }
2780  break;
2781  case Instruction::Or: {
2782  const APInt *BOC;
2783  if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
2784  // Comparing if all bits outside of a constant mask are set?
2785  // Replace (X | C) == -1 with (X & ~C) == ~C.
2786  // This removes the -1 constant.
2787  Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
2788  Value *And = Builder.CreateAnd(BOp0, NotBOC);
2789  return new ICmpInst(Pred, And, NotBOC);
2790  }
2791  break;
2792  }
2793  case Instruction::And: {
2794  const APInt *BOC;
2795  if (match(BOp1, m_APInt(BOC))) {
2796  // If we have ((X & C) == C), turn it into ((X & C) != 0).
2797  if (C == *BOC && C.isPowerOf2())
2798  return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
2799  BO, Constant::getNullValue(RHS->getType()));
2800 
2801  // Don't perform the following transforms if the AND has multiple uses
2802  if (!BO->hasOneUse())
2803  break;
2804 
2805  // ((X & ~7) == 0) --> X < 8
2806  if (C.isNullValue() && (~(*BOC) + 1).isPowerOf2()) {
2807  Constant *NegBOC = ConstantExpr::getNeg(cast<Constant>(BOp1));
2808  auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
2809  return new ICmpInst(NewPred, BOp0, NegBOC);
2810  }
2811  }
2812  break;
2813  }
2814  case Instruction::Mul:
2815  if (C.isNullValue() && BO->hasNoSignedWrap()) {
2816  const APInt *BOC;
2817  if (match(BOp1, m_APInt(BOC)) && !BOC->isNullValue()) {
2818  // The trivial case (mul X, 0) is handled by InstSimplify.
2819  // General case : (mul X, C) != 0 iff X != 0
2820  // (mul X, C) == 0 iff X == 0
2821  return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType()));
2822  }
2823  }
2824  break;
2825  case Instruction::UDiv:
2826  if (C.isNullValue()) {
2827  // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
2828  auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
2829  return new ICmpInst(NewPred, BOp1, BOp0);
2830  }
2831  break;
2832  default:
2833  break;
2834  }
2835  return nullptr;
2836 }
2837 
2838 /// Fold an equality icmp with LLVM intrinsic and constant operand.
2839 Instruction *InstCombiner::foldICmpEqIntrinsicWithConstant(ICmpInst &Cmp,
2840  IntrinsicInst *II,
2841  const APInt &C) {
2842  Type *Ty = II->getType();
2843  unsigned BitWidth = C.getBitWidth();
2844  switch (II->getIntrinsicID()) {
2845  case Intrinsic::bswap:
2846  Worklist.Add(II);
2847  Cmp.setOperand(0, II->getArgOperand(0));
2848  Cmp.setOperand(1, ConstantInt::get(Ty, C.byteSwap()));
2849  return &Cmp;
2850 
2851  case Intrinsic::ctlz:
2852  case Intrinsic::cttz: {
2853  // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
2854  if (C == BitWidth) {
2855  Worklist.Add(II);
2856  Cmp.setOperand(0, II->getArgOperand(0));
2858  return &Cmp;
2859  }
2860 
2861  // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
2862  // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
2863  // Limit to one use to ensure we don't increase instruction count.
2864  unsigned Num = C.getLimitedValue(BitWidth);
2865  if (Num != BitWidth && II->hasOneUse()) {
2866  bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
2867  APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
2868  : APInt::getHighBitsSet(BitWidth, Num + 1);
2869  APInt Mask2 = IsTrailing
2870  ? APInt::getOneBitSet(BitWidth, Num)
2871  : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
2872  Cmp.setOperand(0, Builder.CreateAnd(II->getArgOperand(0), Mask1));
2873  Cmp.setOperand(1, ConstantInt::get(Ty, Mask2));
2874  Worklist.Add(II);
2875  return &Cmp;
2876  }
2877  break;
2878  }
2879 
2880  case Intrinsic::ctpop: {
2881  // popcount(A) == 0 -> A == 0 and likewise for !=
2882  // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
2883  bool IsZero = C.isNullValue();
2884  if (IsZero || C == BitWidth) {
2885  Worklist.Add(II);
2886  Cmp.setOperand(0, II->getArgOperand(0));
2887  auto *NewOp =
2889  Cmp.setOperand(1, NewOp);
2890  return &Cmp;
2891  }
2892  break;
2893  }
2894  default:
2895  break;
2896  }
2897 
2898  return nullptr;
2899 }
2900 
2901 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
2902 Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
2903  IntrinsicInst *II,
2904  const APInt &C) {
2905  if (Cmp.isEquality())
2906  return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
2907 
2908  Type *Ty = II->getType();
2909  unsigned BitWidth = C.getBitWidth();
2910  switch (II->getIntrinsicID()) {
2911  case Intrinsic::ctlz: {
2912  // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
2913  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
2914  unsigned Num = C.getLimitedValue();
2915  APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
2916  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
2917  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
2918  }
2919 
2920  // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
2921  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
2922  C.uge(1) && C.ule(BitWidth)) {
2923  unsigned Num = C.getLimitedValue();
2924  APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
2925  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
2926  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
2927  }
2928  break;
2929  }
2930  case Intrinsic::cttz: {
2931  // Limit to one use to ensure we don't increase instruction count.
2932  if (!II->hasOneUse())
2933  return nullptr;
2934 
2935  // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
2936  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
2937  APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
2938  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
2939  Builder.CreateAnd(II->getArgOperand(0), Mask),
2941  }
2942 
2943  // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
2944  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
2945  C.uge(1) && C.ule(BitWidth)) {
2947  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
2948  Builder.CreateAnd(II->getArgOperand(0), Mask),
2950  }
2951  break;
2952  }
2953  default:
2954  break;
2955  }
2956 
2957  return nullptr;
2958 }
2959 
2960 /// Handle icmp with constant (but not simple integer constant) RHS.
2961 Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) {
2962  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2963  Constant *RHSC = dyn_cast<Constant>(Op1);
2964  Instruction *LHSI = dyn_cast<Instruction>(Op0);
2965  if (!RHSC || !LHSI)
2966  return nullptr;
2967 
2968  switch (LHSI->getOpcode()) {
2969  case Instruction::GetElementPtr:
2970  // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2971  if (RHSC->isNullValue() &&
2972  cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2973  return new ICmpInst(
2974  I.getPredicate(), LHSI->getOperand(0),
2975  Constant::getNullValue(LHSI->getOperand(0)->getType()));
2976  break;
2977  case Instruction::PHI:
2978  // Only fold icmp into the PHI if the phi and icmp are in the same
2979  // block. If in the same block, we're encouraging jump threading. If
2980  // not, we are just pessimizing the code by making an i1 phi.
2981  if (LHSI->getParent() == I.getParent())
2982  if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
2983  return NV;
2984  break;
2985  case Instruction::Select: {
2986  // If either operand of the select is a constant, we can fold the
2987  // comparison into the select arms, which will cause one to be
2988  // constant folded and the select turned into a bitwise or.
2989  Value *Op1 = nullptr, *Op2 = nullptr;
2990  ConstantInt *CI = nullptr;
2991  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2992  Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2993  CI = dyn_cast<ConstantInt>(Op1);
2994  }
2995  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2996  Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2997  CI = dyn_cast<ConstantInt>(Op2);
2998  }
2999 
3000  // We only want to perform this transformation if it will not lead to
3001  // additional code. This is true if either both sides of the select
3002  // fold to a constant (in which case the icmp is replaced with a select
3003  // which will usually simplify) or this is the only user of the
3004  // select (in which case we are trading a select+icmp for a simpler
3005  // select+icmp) or all uses of the select can be replaced based on
3006  // dominance information ("Global cases").
3007  bool Transform = false;
3008  if (Op1 && Op2)
3009  Transform = true;
3010  else if (Op1 || Op2) {
3011  // Local case
3012  if (LHSI->hasOneUse())
3013  Transform = true;
3014  // Global cases
3015  else if (CI && !CI->isZero())
3016  // When Op1 is constant try replacing select with second operand.
3017  // Otherwise Op2 is constant and try replacing select with first
3018  // operand.
3019  Transform =
3020  replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3021  }
3022  if (Transform) {
3023  if (!Op1)
3024  Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3025  I.getName());
3026  if (!Op2)
3027  Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3028  I.getName());
3029  return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3030  }
3031  break;
3032  }
3033  case Instruction::IntToPtr:
3034  // icmp pred inttoptr(X), null -> icmp pred X, 0
3035  if (RHSC->isNullValue() &&
3036  DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3037  return new ICmpInst(
3038  I.getPredicate(), LHSI->getOperand(0),
3039  Constant::getNullValue(LHSI->getOperand(0)->getType()));
3040  break;
3041 
3042  case Instruction::Load:
3043  // Try to optimize things like "A[i] > 4" to index computations.
3044  if (GetElementPtrInst *GEP =
3045  dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3046  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3047  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3048  !cast<LoadInst>(LHSI)->isVolatile())
3049  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3050  return Res;
3051  }
3052  break;
3053  }
3054 
3055  return nullptr;
3056 }
3057 
3058 /// Some comparisons can be simplified.
3059 /// In this case, we are looking for comparisons that look like
3060 /// a check for a lossy truncation.
3061 /// Folds:
3062 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3063 /// Where Mask is some pattern that produces all-ones in low bits:
3064 /// (-1 >> y)
3065 /// ((-1 << y) >> y) <- non-canonical, has extra uses
3066 /// ~(-1 << y)
3067 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
3068 /// The Mask can be a constant, too.
3069 /// For some predicates, the operands are commutative.
3070 /// For others, x can only be on a specific side.
3072  InstCombiner::BuilderTy &Builder) {
3073  ICmpInst::Predicate SrcPred;
3074  Value *X, *M, *Y;
3075  auto m_VariableMask = m_CombineOr(
3077  m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3079  m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
3080  auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3081  if (!match(&I, m_c_ICmp(SrcPred,
3082  m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
3083  m_Deferred(X))))
3084  return nullptr;
3085 
3086  ICmpInst::Predicate DstPred;
3087  switch (SrcPred) {
3088  case ICmpInst::Predicate::ICMP_EQ:
3089  // x & (-1 >> y) == x -> x u<= (-1 >> y)
3090  DstPred = ICmpInst::Predicate::ICMP_ULE;
3091  break;
3092  case ICmpInst::Predicate::ICMP_NE:
3093  // x & (-1 >> y) != x -> x u> (-1 >> y)
3094  DstPred = ICmpInst::Predicate::ICMP_UGT;
3095  break;
3096  case ICmpInst::Predicate::ICMP_UGT:
3097  // x u> x & (-1 >> y) -> x u> (-1 >> y)
3098  assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
3099  DstPred = ICmpInst::Predicate::ICMP_UGT;
3100  break;
3101  case ICmpInst::Predicate::ICMP_UGE:
3102  // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3103  assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
3104  DstPred = ICmpInst::Predicate::ICMP_ULE;
3105  break;
3106  case ICmpInst::Predicate::ICMP_ULT:
3107  // x & (-1 >> y) u< x -> x u> (-1 >> y)
3108  assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
3109  DstPred = ICmpInst::Predicate::ICMP_UGT;
3110  break;
3111  case ICmpInst::Predicate::ICMP_ULE:
3112  // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3113  assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
3114  DstPred = ICmpInst::Predicate::ICMP_ULE;
3115  break;
3116  case ICmpInst::Predicate::ICMP_SGT:
3117  // x s> x & (-1 >> y) -> x s> (-1 >> y)
3118  if (X != I.getOperand(0)) // X must be on LHS of comparison!
3119  return nullptr; // Ignore the other case.
3120  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3121  return nullptr;
3122  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3123  return nullptr;
3124  DstPred = ICmpInst::Predicate::ICMP_SGT;
3125  break;
3126  case ICmpInst::Predicate::ICMP_SGE:
3127  // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3128  if (X != I.getOperand(1)) // X must be on RHS of comparison!
3129  return nullptr; // Ignore the other case.
3130  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3131  return nullptr;
3132  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3133  return nullptr;
3134  DstPred = ICmpInst::Predicate::ICMP_SLE;
3135  break;
3136  case ICmpInst::Predicate::ICMP_SLT:
3137  // x & (-1 >> y) s< x -> x s> (-1 >> y)
3138  if (X != I.getOperand(1)) // X must be on RHS of comparison!
3139  return nullptr; // Ignore the other case.
3140  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3141  return nullptr;
3142  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3143  return nullptr;
3144  DstPred = ICmpInst::Predicate::ICMP_SGT;
3145  break;
3146  case ICmpInst::Predicate::ICMP_SLE:
3147  // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3148  if (X != I.getOperand(0)) // X must be on LHS of comparison!
3149  return nullptr; // Ignore the other case.
3150  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3151  return nullptr;
3152  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3153  return nullptr;
3154  DstPred = ICmpInst::Predicate::ICMP_SLE;
3155  break;
3156  default:
3157  llvm_unreachable("All possible folds are handled.");
3158  }
3159 
3160  return Builder.CreateICmp(DstPred, X, M);
3161 }
3162 
3163 /// Some comparisons can be simplified.
3164 /// In this case, we are looking for comparisons that look like
3165 /// a check for a lossy signed truncation.
3166 /// Folds: (MaskedBits is a constant.)
3167 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3168 /// Into:
3169 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3170 /// Where KeptBits = bitwidth(%x) - MaskedBits
3171 static Value *
3173  InstCombiner::BuilderTy &Builder) {
3174  ICmpInst::Predicate SrcPred;
3175  Value *X;
3176  const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3177  // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3178  if (!match(&I, m_c_ICmp(SrcPred,
3179  m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3180  m_APInt(C1))),
3181  m_Deferred(X))))
3182  return nullptr;
3183 
3184  // Potential handling of non-splats: for each element:
3185  // * if both are undef, replace with constant 0.
3186  // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3187  // * if both are not undef, and are different, bailout.
3188  // * else, only one is undef, then pick the non-undef one.
3189 
3190  // The shift amount must be equal.
3191  if (*C0 != *C1)
3192  return nullptr;
3193  const APInt &MaskedBits = *C0;
3194  assert(MaskedBits != 0 && "shift by zero should be folded away already.");
3195 
3196  ICmpInst::Predicate DstPred;
3197  switch (SrcPred) {
3198  case ICmpInst::Predicate::ICMP_EQ:
3199  // ((%x << MaskedBits) a>> MaskedBits) == %x
3200  // =>
3201  // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3202  DstPred = ICmpInst::Predicate::ICMP_ULT;
3203  break;
3204  case ICmpInst::Predicate::ICMP_NE:
3205  // ((%x << MaskedBits) a>> MaskedBits) != %x
3206  // =>
3207  // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3208  DstPred = ICmpInst::Predicate::ICMP_UGE;
3209  break;
3210  // FIXME: are more folds possible?
3211  default:
3212  return nullptr;
3213  }
3214 
3215  auto *XType = X->getType();
3216  const unsigned XBitWidth = XType->getScalarSizeInBits();
3217  const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3218  assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3219 
3220  // KeptBits = bitwidth(%x) - MaskedBits
3221  const APInt KeptBits = BitWidth - MaskedBits;
3222  assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3223  // ICmpCst = (1 << KeptBits)
3224  const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3225  assert(ICmpCst.isPowerOf2());
3226  // AddCst = (1 << (KeptBits-1))
3227  const APInt AddCst = ICmpCst.lshr(1);
3228  assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3229 
3230  // T0 = add %x, AddCst
3231  Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3232  // T1 = T0 DstPred ICmpCst
3233  Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3234 
3235  return T1;
3236 }
3237 
3238 /// Try to fold icmp (binop), X or icmp X, (binop).
3239 /// TODO: A large part of this logic is duplicated in InstSimplify's
3240 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3241 /// duplication.
3242 Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I) {
3243  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3244 
3245  // Special logic for binary operators.
3246  BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3247  BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3248  if (!BO0 && !BO1)
3249  return nullptr;
3250 
3251  const CmpInst::Predicate Pred = I.getPredicate();
3252  Value *X;
3253 
3254  // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3255  // (Op1 + X) <u Op1 --> ~Op1 <u X
3256  // Op0 >u (Op0 + X) --> X >u ~Op0
3257  if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3258  Pred == ICmpInst::ICMP_ULT)
3259  return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3260  if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
3261  Pred == ICmpInst::ICMP_UGT)
3262  return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3263 
3264  bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3265  if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3266  NoOp0WrapProblem =
3267  ICmpInst::isEquality(Pred) ||
3268  (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3269  (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3270  if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3271  NoOp1WrapProblem =
3272  ICmpInst::isEquality(Pred) ||
3273  (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3274  (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3275 
3276  // Analyze the case when either Op0 or Op1 is an add instruction.
3277  // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3278  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3279  if (BO0 && BO0->getOpcode() == Instruction::Add) {
3280  A = BO0->getOperand(0);
3281  B = BO0->getOperand(1);
3282  }
3283  if (BO1 && BO1->getOpcode() == Instruction::Add) {
3284  C = BO1->getOperand(0);
3285  D = BO1->getOperand(1);
3286  }
3287 
3288  // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3289  if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3290  return new ICmpInst(Pred, A == Op1 ? B : A,
3291  Constant::getNullValue(Op1->getType()));
3292 
3293  // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3294  if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3295  return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3296  C == Op0 ? D : C);
3297 
3298  // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
3299  if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3300  NoOp1WrapProblem &&
3301  // Try not to increase register pressure.
3302  BO0->hasOneUse() && BO1->hasOneUse()) {
3303  // Determine Y and Z in the form icmp (X+Y), (X+Z).
3304  Value *Y, *Z;
3305  if (A == C) {
3306  // C + B == C + D -> B == D
3307  Y = B;
3308  Z = D;
3309  } else if (A == D) {
3310  // D + B == C + D -> B == C
3311  Y = B;
3312  Z = C;
3313  } else if (B == C) {
3314  // A + C == C + D -> A == D
3315  Y = A;
3316  Z = D;
3317  } else {
3318  assert(B == D);
3319  // A + D == C + D -> A == C
3320  Y = A;
3321  Z = C;
3322  }
3323  return new ICmpInst(Pred, Y, Z);
3324  }
3325 
3326  // icmp slt (X + -1), Y -> icmp sle X, Y
3327  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3328  match(B, m_AllOnes()))
3329  return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3330 
3331  // icmp sge (X + -1), Y -> icmp sgt X, Y
3332  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3333  match(B, m_AllOnes()))
3334  return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3335 
3336  // icmp sle (X + 1), Y -> icmp slt X, Y
3337  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
3338  return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3339 
3340  // icmp sgt (X + 1), Y -> icmp sge X, Y
3341  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
3342  return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3343 
3344  // icmp sgt X, (Y + -1) -> icmp sge X, Y
3345  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
3346  match(D, m_AllOnes()))
3347  return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3348 
3349  // icmp sle X, (Y + -1) -> icmp slt X, Y
3350  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
3351  match(D, m_AllOnes()))
3352  return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3353 
3354  // icmp sge X, (Y + 1) -> icmp sgt X, Y
3355  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
3356  return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3357 
3358  // icmp slt X, (Y + 1) -> icmp sle X, Y
3359  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
3360  return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3361 
3362  // TODO: The subtraction-related identities shown below also hold, but
3363  // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3364  // wouldn't happen even if they were implemented.
3365  //
3366  // icmp ult (X - 1), Y -> icmp ule X, Y
3367  // icmp uge (X - 1), Y -> icmp ugt X, Y
3368  // icmp ugt X, (Y - 1) -> icmp uge X, Y
3369  // icmp ule X, (Y - 1) -> icmp ult X, Y
3370 
3371  // icmp ule (X + 1), Y -> icmp ult X, Y
3372  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
3373  return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3374 
3375  // icmp ugt (X + 1), Y -> icmp uge X, Y
3376  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
3377  return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3378 
3379  // icmp uge X, (Y + 1) -> icmp ugt X, Y
3380  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
3381  return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3382 
3383  // icmp ult X, (Y + 1) -> icmp ule X, Y
3384  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
3385  return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3386 
3387  // if C1 has greater magnitude than C2:
3388  // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3389  // s.t. C3 = C1 - C2
3390  //
3391  // if C2 has greater magnitude than C1:
3392  // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3393  // s.t. C3 = C2 - C1
3394  if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3395  (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3396  if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3397  if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3398  const APInt &AP1 = C1->getValue();
3399  const APInt &AP2 = C2->getValue();
3400  if (AP1.isNegative() == AP2.isNegative()) {
3401  APInt AP1Abs = C1->getValue().abs();
3402  APInt AP2Abs = C2->getValue().abs();
3403  if (AP1Abs.uge(AP2Abs)) {
3404  ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3405  Value *NewAdd = Builder.CreateNSWAdd(A, C3);
3406  return new ICmpInst(Pred, NewAdd, C);
3407  } else {
3408  ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3409  Value *NewAdd = Builder.CreateNSWAdd(C, C3);
3410  return new ICmpInst(Pred, A, NewAdd);
3411  }
3412  }
3413  }
3414 
3415  // Analyze the case when either Op0 or Op1 is a sub instruction.
3416  // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3417  A = nullptr;
3418  B = nullptr;
3419  C = nullptr;
3420  D = nullptr;
3421  if (BO0 && BO0->getOpcode() == Instruction::Sub) {
3422  A = BO0->getOperand(0);
3423  B = BO0->getOperand(1);
3424  }
3425  if (BO1 && BO1->getOpcode() == Instruction::Sub) {
3426  C = BO1->getOperand(0);
3427  D = BO1->getOperand(1);
3428  }
3429 
3430  // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3431  if (A == Op1 && NoOp0WrapProblem)
3432  return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3433  // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3434  if (C == Op0 && NoOp1WrapProblem)
3435  return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3436 
3437  // (A - B) >u A --> A <u B
3438  if (A == Op1 && Pred == ICmpInst::ICMP_UGT)
3439  return new ICmpInst(ICmpInst::ICMP_ULT, A, B);
3440  // C <u (C - D) --> C <u D
3441  if (C == Op0 && Pred == ICmpInst::ICMP_ULT)
3442  return new ICmpInst(ICmpInst::ICMP_ULT, C, D);
3443 
3444  // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3445  if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3446  // Try not to increase register pressure.
3447  BO0->hasOneUse() && BO1->hasOneUse())
3448  return new ICmpInst(Pred, A, C);
3449  // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3450  if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3451  // Try not to increase register pressure.
3452  BO0->hasOneUse() && BO1->hasOneUse())
3453  return new ICmpInst(Pred, D, B);
3454 
3455  // icmp (0-X) < cst --> x > -cst
3456  if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3457  Value *X;
3458  if (match(BO0, m_Neg(m_Value(X))))
3459  if (Constant *RHSC = dyn_cast<Constant>(Op1))
3460  if (RHSC->isNotMinSignedValue())
3461  return new ICmpInst(I.getSwappedPredicate(), X,
3462  ConstantExpr::getNeg(RHSC));
3463  }
3464 
3465  BinaryOperator *SRem = nullptr;
3466  // icmp (srem X, Y), Y
3467  if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
3468  SRem = BO0;
3469  // icmp Y, (srem X, Y)
3470  else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3471  Op0 == BO1->getOperand(1))
3472  SRem = BO1;
3473  if (SRem) {
3474  // We don't check hasOneUse to avoid increasing register pressure because
3475  // the value we use is the same value this instruction was already using.
3476  switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3477  default:
3478  break;
3479  case ICmpInst::ICMP_EQ:
3480  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3481  case ICmpInst::ICMP_NE:
3482  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3483  case ICmpInst::ICMP_SGT:
3484  case ICmpInst::ICMP_SGE:
3485  return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3487  case ICmpInst::ICMP_SLT:
3488  case ICmpInst::ICMP_SLE:
3489  return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3490  Constant::getNullValue(SRem->getType()));
3491  }
3492  }
3493 
3494  if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
3495  BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
3496  switch (BO0->getOpcode()) {
3497  default:
3498  break;
3499  case Instruction::Add:
3500  case Instruction::Sub:
3501  case Instruction::Xor: {
3502  if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3503  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3504 
3505  const APInt *C;
3506  if (match(BO0->getOperand(1), m_APInt(C))) {
3507  // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
3508  if (C->isSignMask()) {
3509  ICmpInst::Predicate NewPred =
3511  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3512  }
3513 
3514  // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
3515  if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
3516  ICmpInst::Predicate NewPred =
3518  NewPred = I.getSwappedPredicate(NewPred);
3519  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3520  }
3521  }
3522  break;
3523  }
3524  case Instruction::Mul: {
3525  if (!I.isEquality())
3526  break;
3527 
3528  const APInt *C;
3529  if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
3530  !C->isOneValue()) {
3531  // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
3532  // Mask = -1 >> count-trailing-zeros(C).
3533  if (unsigned TZs = C->countTrailingZeros()) {
3535  BO0->getType(),
3536  APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
3537  Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
3538  Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
3539  return new ICmpInst(Pred, And1, And2);
3540  }
3541  // If there are no trailing zeros in the multiplier, just eliminate
3542  // the multiplies (no masking is needed):
3543  // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
3544  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3545  }
3546  break;
3547  }
3548  case Instruction::UDiv:
3549  case Instruction::LShr:
3550  if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
3551  break;
3552  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3553 
3554  case Instruction::SDiv:
3555  if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
3556  break;
3557  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3558 
3559  case Instruction::AShr:
3560  if (!BO0->isExact() || !BO1->isExact())
3561  break;
3562  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3563 
3564  case Instruction::Shl: {
3565  bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3566  bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3567  if (!NUW && !NSW)
3568  break;
3569  if (!NSW && I.isSigned())
3570  break;
3571  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3572  }
3573  }
3574  }
3575 
3576  if (BO0) {
3577  // Transform A & (L - 1) `ult` L --> L != 0
3578  auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
3579  auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
3580 
3581  if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
3582  auto *Zero = Constant::getNullValue(BO0->getType());
3583  return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
3584  }
3585  }
3586 
3587  if (Value *V = foldICmpWithLowBitMaskedVal(I, Builder))
3588  return replaceInstUsesWith(I, V);
3589 
3590  if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
3591  return replaceInstUsesWith(I, V);
3592 
3593  return nullptr;
3594 }
3595 
3596 /// Fold icmp Pred min|max(X, Y), X.
3598  ICmpInst::Predicate Pred = Cmp.getPredicate();
3599  Value *Op0 = Cmp.getOperand(0);
3600  Value *X = Cmp.getOperand(1);
3601 
3602  // Canonicalize minimum or maximum operand to LHS of the icmp.
3603  if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
3604  match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
3605  match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
3606  match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
3607  std::swap(Op0, X);
3608  Pred = Cmp.getSwappedPredicate();
3609  }
3610 
3611  Value *Y;
3612  if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
3613  // smin(X, Y) == X --> X s<= Y
3614  // smin(X, Y) s>= X --> X s<= Y
3615  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
3616  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
3617 
3618  // smin(X, Y) != X --> X s> Y
3619  // smin(X, Y) s< X --> X s> Y
3620  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
3621  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
3622 
3623  // These cases should be handled in InstSimplify:
3624  // smin(X, Y) s<= X --> true
3625  // smin(X, Y) s> X --> false
3626  return nullptr;
3627  }
3628 
3629  if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
3630  // smax(X, Y) == X --> X s>= Y
3631  // smax(X, Y) s<= X --> X s>= Y
3632  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
3633  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
3634 
3635  // smax(X, Y) != X --> X s< Y
3636  // smax(X, Y) s> X --> X s< Y
3637  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
3638  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
3639 
3640  // These cases should be handled in InstSimplify:
3641  // smax(X, Y) s>= X --> true
3642  // smax(X, Y) s< X --> false
3643  return nullptr;
3644  }
3645 
3646  if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
3647  // umin(X, Y) == X --> X u<= Y
3648  // umin(X, Y) u>= X --> X u<= Y
3649  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
3650  return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
3651 
3652  // umin(X, Y) != X --> X u> Y
3653  // umin(X, Y) u< X --> X u> Y
3654  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
3655  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
3656 
3657  // These cases should be handled in InstSimplify:
3658  // umin(X, Y) u<= X --> true
3659  // umin(X, Y) u> X --> false
3660  return nullptr;
3661  }
3662 
3663  if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
3664  // umax(X, Y) == X --> X u>= Y
3665  // umax(X, Y) u<= X --> X u>= Y
3666  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
3667  return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
3668 
3669  // umax(X, Y) != X --> X u< Y
3670  // umax(X, Y) u> X --> X u< Y
3671  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
3672  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
3673 
3674  // These cases should be handled in InstSimplify:
3675  // umax(X, Y) u>= X --> true
3676  // umax(X, Y) u< X --> false
3677  return nullptr;
3678  }
3679 
3680  return nullptr;
3681 }
3682 
3683 Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
3684  if (!I.isEquality())
3685  return nullptr;
3686 
3687  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3688  const CmpInst::Predicate Pred = I.getPredicate();
3689  Value *A, *B, *C, *D;
3690  if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3691  if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3692  Value *OtherVal = A == Op1 ? B : A;
3693  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3694  }
3695 
3696  if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3697  // A^c1 == C^c2 --> A == C^(c1^c2)
3698  ConstantInt *C1, *C2;
3699  if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
3700  Op1->hasOneUse()) {
3701  Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
3702  Value *Xor = Builder.CreateXor(C, NC);
3703  return new ICmpInst(Pred, A, Xor);
3704  }
3705 
3706  // A^B == A^D -> B == D
3707  if (A == C)
3708  return new ICmpInst(Pred, B, D);
3709  if (A == D)
3710  return new ICmpInst(Pred, B, C);
3711  if (B == C)
3712  return new ICmpInst(Pred, A, D);
3713  if (B == D)
3714  return new ICmpInst(Pred, A, C);
3715  }
3716  }
3717 
3718  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
3719  // A == (A^B) -> B == 0
3720  Value *OtherVal = A == Op0 ? B : A;
3721  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3722  }
3723 
3724  // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3725  if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3726  match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3727  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3728 
3729  if (A == C) {
3730  X = B;
3731  Y = D;
3732  Z = A;
3733  } else if (A == D) {
3734  X = B;
3735  Y = C;
3736  Z = A;
3737  } else if (B == C) {
3738  X = A;
3739  Y = D;
3740  Z = B;
3741  } else if (B == D) {
3742  X = A;
3743  Y = C;
3744  Z = B;
3745  }
3746 
3747  if (X) { // Build (X^Y) & Z
3748  Op1 = Builder.CreateXor(X, Y);
3749  Op1 = Builder.CreateAnd(Op1, Z);
3750  I.setOperand(0, Op1);
3751  I.setOperand(1, Constant::getNullValue(Op1->getType()));
3752  return &I;
3753  }
3754  }
3755 
3756  // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3757  // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3758  ConstantInt *Cst1;
3759  if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
3760  match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3761  (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3762  match(Op1, m_ZExt(m_Value(A))))) {
3763  APInt Pow2 = Cst1->getValue() + 1;
3764  if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3765  Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3766  return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
3767  }
3768 
3769  // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3770  // For lshr and ashr pairs.
3771  if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3772  match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3773  (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3774  match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3775  unsigned TypeBits = Cst1->getBitWidth();
3776  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3777  if (ShAmt < TypeBits && ShAmt != 0) {
3778  ICmpInst::Predicate NewPred =
3780  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3781  APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3782  return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
3783  }
3784  }
3785 
3786  // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
3787  if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
3788  match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
3789  unsigned TypeBits = Cst1->getBitWidth();
3790  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3791  if (ShAmt < TypeBits && ShAmt != 0) {
3792  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3793  APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
3794  Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
3795  I.getName() + ".mask");
3796  return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
3797  }
3798  }
3799 
3800  // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3801  // "icmp (and X, mask), cst"
3802  uint64_t ShAmt = 0;
3803  if (Op0->hasOneUse() &&
3804  match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
3805  match(Op1, m_ConstantInt(Cst1)) &&
3806  // Only do this when A has multiple uses. This is most important to do
3807  // when it exposes other optimizations.
3808  !A->hasOneUse()) {
3809  unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3810 
3811  if (ShAmt < ASize) {
3812  APInt MaskV =
3814  MaskV <<= ShAmt;
3815 
3816  APInt CmpV = Cst1->getValue().zext(ASize);
3817  CmpV <<= ShAmt;
3818 
3819  Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
3820  return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
3821  }
3822  }
3823 
3824  // If both operands are byte-swapped or bit-reversed, just compare the
3825  // original values.
3826  // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
3827  // and handle more intrinsics.
3828  if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
3829  (match(Op0, m_BitReverse(m_Value(A))) &&
3830  match(Op1, m_BitReverse(m_Value(B)))))
3831  return new ICmpInst(Pred, A, B);
3832 
3833  return nullptr;
3834 }
3835 
3836 /// Handle icmp (cast x to y), (cast/cst). We only handle extending casts so
3837 /// far.
3838 Instruction *InstCombiner::foldICmpWithCastAndCast(ICmpInst &ICmp) {
3839  const CastInst *LHSCI = cast<CastInst>(ICmp.getOperand(0));
3840  Value *LHSCIOp = LHSCI->getOperand(0);
3841  Type *SrcTy = LHSCIOp->getType();
3842  Type *DestTy = LHSCI->getType();
3843 
3844  // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
3845  // integer type is the same size as the pointer type.
3846  const auto& CompatibleSizes = [&](Type* SrcTy, Type* DestTy) -> bool {
3847  if (isa<VectorType>(SrcTy)) {
3848  SrcTy = cast<VectorType>(SrcTy)->getElementType();
3849  DestTy = cast<VectorType>(DestTy)->getElementType();
3850  }
3851  return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
3852  };
3853  if (LHSCI->getOpcode() == Instruction::PtrToInt &&
3854  CompatibleSizes(SrcTy, DestTy)) {
3855  Value *RHSOp = nullptr;
3856  if (auto *RHSC = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
3857  Value *RHSCIOp = RHSC->getOperand(0);
3858  if (RHSCIOp->getType()->getPointerAddressSpace() ==
3859  LHSCIOp->getType()->getPointerAddressSpace()) {
3860  RHSOp = RHSC->getOperand(0);
3861  // If the pointer types don't match, insert a bitcast.
3862  if (LHSCIOp->getType() != RHSOp->getType())
3863  RHSOp = Builder.CreateBitCast(RHSOp, LHSCIOp->getType());
3864  }
3865  } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
3866  RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
3867  }
3868 
3869  if (RHSOp)
3870  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSOp);
3871  }
3872 
3873  // The code below only handles extension cast instructions, so far.
3874  // Enforce this.
3875  if (LHSCI->getOpcode() != Instruction::ZExt &&
3876  LHSCI->getOpcode() != Instruction::SExt)
3877  return nullptr;
3878 
3879  bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
3880  bool isSignedCmp = ICmp.isSigned();
3881 
3882  if (auto *CI = dyn_cast<CastInst>(ICmp.getOperand(1))) {
3883  // Not an extension from the same type?
3884  Value *RHSCIOp = CI->getOperand(0);
3885  if (RHSCIOp->getType() != LHSCIOp->getType())
3886  return nullptr;
3887 
3888  // If the signedness of the two casts doesn't agree (i.e. one is a sext
3889  // and the other is a zext), then we can't handle this.
3890  if (CI->getOpcode() != LHSCI->getOpcode())
3891  return nullptr;
3892 
3893  // Deal with equality cases early.
3894  if (ICmp.isEquality())
3895  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3896 
3897  // A signed comparison of sign extended values simplifies into a
3898  // signed comparison.
3899  if (isSignedCmp && isSignedExt)
3900  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3901 
3902  // The other three cases all fold into an unsigned comparison.
3903  return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
3904  }
3905 
3906  // If we aren't dealing with a constant on the RHS, exit early.
3907  auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
3908  if (!C)
3909  return nullptr;
3910 
3911  // Compute the constant that would happen if we truncated to SrcTy then
3912  // re-extended to DestTy.
3913  Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
3914  Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
3915 
3916  // If the re-extended constant didn't change...
3917  if (Res2 == C) {
3918  // Deal with equality cases early.
3919  if (ICmp.isEquality())
3920  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3921 
3922  // A signed comparison of sign extended values simplifies into a
3923  // signed comparison.
3924  if (isSignedExt && isSignedCmp)
3925  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3926 
3927  // The other three cases all fold into an unsigned comparison.
3928  return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, Res1);
3929  }
3930 
3931  // The re-extended constant changed, partly changed (in the case of a vector),
3932  // or could not be determined to be equal (in the case of a constant
3933  // expression), so the constant cannot be represented in the shorter type.
3934  // Consequently, we cannot emit a simple comparison.
3935  // All the cases that fold to true or false will have already been handled
3936  // by SimplifyICmpInst, so only deal with the tricky case.
3937 
3938  if (isSignedCmp || !isSignedExt || !isa<ConstantInt>(C))
3939  return nullptr;
3940 
3941  // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
3942  // should have been folded away previously and not enter in here.
3943 
3944  // We're performing an unsigned comp with a sign extended value.
3945  // This is true if the input is >= 0. [aka >s -1]
3946  Constant *NegOne = Constant::getAllOnesValue(SrcTy);
3947  Value *Result = Builder.CreateICmpSGT(LHSCIOp, NegOne, ICmp.getName());
3948 
3949  // Finally, return the value computed.
3950  if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
3951  return replaceInstUsesWith(ICmp, Result);
3952 
3953  assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
3954  return BinaryOperator::CreateNot(Result);
3955 }
3956 
3957 static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS) {
3958  switch (BinaryOp) {
3959  default:
3960  llvm_unreachable("Unsupported binary op");
3961  case Instruction::Add:
3962  case Instruction::Sub:
3963  return match(RHS, m_Zero());
3964  case Instruction::Mul:
3965  return match(RHS, m_One());
3966  }
3967 }
3968 
3970  Instruction::BinaryOps BinaryOp, bool IsSigned,
3971  Value *LHS, Value *RHS, Instruction *CxtI) const {
3972  switch (BinaryOp) {
3973  default:
3974  llvm_unreachable("Unsupported binary op");
3975  case Instruction::Add:
3976  if (IsSigned)
3977  return computeOverflowForSignedAdd(LHS, RHS, CxtI);
3978  else
3979  return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
3980  case Instruction::Sub:
3981  if (IsSigned)
3982  return computeOverflowForSignedSub(LHS, RHS, CxtI);
3983  else
3984  return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
3985  case Instruction::Mul:
3986  if (IsSigned)
3987  return computeOverflowForSignedMul(LHS, RHS, CxtI);
3988  else
3989  return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
3990  }
3991 }
3992 
3993 bool InstCombiner::OptimizeOverflowCheck(
3994  Instruction::BinaryOps BinaryOp, bool IsSigned, Value *LHS, Value *RHS,
3995  Instruction &OrigI, Value *&Result, Constant *&Overflow) {
3996  if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
3997  std::swap(LHS, RHS);
3998 
3999  // If the overflow check was an add followed by a compare, the insertion point
4000  // may be pointing to the compare. We want to insert the new instructions
4001  // before the add in case there are uses of the add between the add and the
4002  // compare.
4003  Builder.SetInsertPoint(&OrigI);
4004 
4005  if (isNeutralValue(BinaryOp, RHS)) {
4006  Result = LHS;
4007  Overflow = Builder.getFalse();
4008  return true;
4009  }
4010 
4011  switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, &OrigI)) {
4013  return false;
4016  Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4017  Result->takeName(&OrigI);
4018  Overflow = Builder.getTrue();
4019  return true;
4021  Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4022  Result->takeName(&OrigI);
4023  Overflow = Builder.getFalse();
4024  if (auto *Inst = dyn_cast<Instruction>(Result)) {
4025  if (IsSigned)
4026  Inst->setHasNoSignedWrap();
4027  else
4028  Inst->setHasNoUnsignedWrap();
4029  }
4030  return true;
4031  }
4032 
4033  llvm_unreachable("Unexpected overflow result");
4034 }
4035 
4036 /// Recognize and process idiom involving test for multiplication
4037 /// overflow.
4038 ///
4039 /// The caller has matched a pattern of the form:
4040 /// I = cmp u (mul(zext A, zext B), V
4041 /// The function checks if this is a test for overflow and if so replaces
4042 /// multiplication with call to 'mul.with.overflow' intrinsic.
4043 ///
4044 /// \param I Compare instruction.
4045 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
4046 /// the compare instruction. Must be of integer type.
4047 /// \param OtherVal The other argument of compare instruction.
4048 /// \returns Instruction which must replace the compare instruction, NULL if no
4049 /// replacement required.
4051  Value *OtherVal, InstCombiner &IC) {
4052  // Don't bother doing this transformation for pointers, don't do it for
4053  // vectors.
4054  if (!isa<IntegerType>(MulVal->getType()))
4055  return nullptr;
4056 
4057  assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
4058  assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
4059  auto *MulInstr = dyn_cast<Instruction>(MulVal);
4060  if (!MulInstr)
4061  return nullptr;
4062  assert(MulInstr->getOpcode() == Instruction::Mul);
4063 
4064  auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
4065  *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
4066  assert(LHS->getOpcode() == Instruction::ZExt);
4067  assert(RHS->getOpcode() == Instruction::ZExt);
4068  Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
4069 
4070  // Calculate type and width of the result produced by mul.with.overflow.
4071  Type *TyA = A->getType(), *TyB = B->getType();
4072  unsigned WidthA = TyA->getPrimitiveSizeInBits(),
4073  WidthB = TyB->getPrimitiveSizeInBits();
4074  unsigned MulWidth;
4075  Type *MulType;
4076  if (WidthB > WidthA) {
4077  MulWidth = WidthB;
4078  MulType = TyB;
4079  } else {
4080  MulWidth = WidthA;
4081  MulType = TyA;
4082  }
4083 
4084  // In order to replace the original mul with a narrower mul.with.overflow,
4085  // all uses must ignore upper bits of the product. The number of used low
4086  // bits must be not greater than the width of mul.with.overflow.
4087  if (MulVal->hasNUsesOrMore(2))
4088  for (User *U : MulVal->users()) {
4089  if (U == &I)
4090  continue;
4091  if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4092  // Check if truncation ignores bits above MulWidth.
4093  unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
4094  if (TruncWidth > MulWidth)
4095  return nullptr;
4096  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4097  // Check if AND ignores bits above MulWidth.
4098  if (BO->getOpcode() != Instruction::And)
4099  return nullptr;
4100  if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4101  const APInt &CVal = CI->getValue();
4102  if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
4103  return nullptr;
4104  } else {
4105  // In this case we could have the operand of the binary operation
4106  // being defined in another block, and performing the replacement
4107  // could break the dominance relation.
4108  return nullptr;
4109  }
4110  } else {
4111  // Other uses prohibit this transformation.
4112  return nullptr;
4113  }
4114  }
4115 
4116  // Recognize patterns
4117  switch (I.getPredicate()) {
4118  case ICmpInst::ICMP_EQ:
4119  case ICmpInst::ICMP_NE:
4120  // Recognize pattern:
4121  // mulval = mul(zext A, zext B)
4122  // cmp eq/neq mulval, zext trunc mulval
4123  if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
4124  if (Zext->hasOneUse()) {
4125  Value *ZextArg = Zext->getOperand(0);
4126  if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
4127  if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
4128  break; //Recognized
4129  }
4130 
4131  // Recognize pattern:
4132  // mulval = mul(zext A, zext B)
4133  // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
4134  ConstantInt *CI;
4135  Value *ValToMask;
4136  if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
4137  if (ValToMask != MulVal)
4138  return nullptr;
4139  const APInt &CVal = CI->getValue() + 1;
4140  if (CVal.isPowerOf2()) {
4141  unsigned MaskWidth = CVal.logBase2();
4142  if (MaskWidth == MulWidth)
4143  break; // Recognized
4144  }
4145  }
4146  return nullptr;
4147 
4148  case ICmpInst::ICMP_UGT:
4149  // Recognize pattern:
4150  // mulval = mul(zext A, zext B)
4151  // cmp ugt mulval, max
4152  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4153  APInt MaxVal = APInt::getMaxValue(MulWidth);
4154  MaxVal = MaxVal.zext(CI->getBitWidth());
4155  if (MaxVal.eq(CI->getValue()))
4156  break; // Recognized
4157  }
4158  return nullptr;
4159 
4160  case ICmpInst::ICMP_UGE:
4161  // Recognize pattern:
4162  // mulval = mul(zext A, zext B)
4163  // cmp uge mulval, max+1
4164  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4165  APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4166  if (MaxVal.eq(CI->getValue()))
4167  break; // Recognized
4168  }
4169  return nullptr;
4170 
4171  case ICmpInst::ICMP_ULE:
4172  // Recognize pattern:
4173  // mulval = mul(zext A, zext B)
4174  // cmp ule mulval, max
4175  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4176  APInt MaxVal = APInt::getMaxValue(MulWidth);
4177  MaxVal = MaxVal.zext(CI->getBitWidth());
4178  if (MaxVal.eq(CI->getValue()))
4179  break; // Recognized
4180  }
4181  return nullptr;
4182 
4183  case ICmpInst::ICMP_ULT:
4184  // Recognize pattern:
4185  // mulval = mul(zext A, zext B)
4186  // cmp ule mulval, max + 1
4187  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4188  APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4189  if (MaxVal.eq(CI->getValue()))
4190  break; // Recognized
4191  }
4192  return nullptr;
4193 
4194  default:
4195  return nullptr;
4196  }
4197 
4198  InstCombiner::BuilderTy &Builder = IC.Builder;
4199  Builder.SetInsertPoint(MulInstr);
4200 
4201  // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4202  Value *MulA = A, *MulB = B;
4203  if (WidthA < MulWidth)
4204  MulA = Builder.CreateZExt(A, MulType);
4205  if (WidthB < MulWidth)
4206  MulB = Builder.CreateZExt(B, MulType);
4208  I.getModule(), Intrinsic::umul_with_overflow, MulType);
4209  CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4210  IC.Worklist.Add(MulInstr);
4211 
4212  // If there are uses of mul result other than the comparison, we know that
4213  // they are truncation or binary AND. Change them to use result of
4214  // mul.with.overflow and adjust properly mask/size.
4215  if (MulVal->hasNUsesOrMore(2)) {
4216  Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4217  for (auto UI = MulVal->user_begin(), UE = MulVal->user_end(); UI != UE;) {
4218  User *U = *UI++;
4219  if (U == &I || U == OtherVal)
4220  continue;
4221  if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4222  if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
4223  IC.replaceInstUsesWith(*TI, Mul);
4224  else
4225  TI->setOperand(0, Mul);
4226  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4227  assert(BO->getOpcode() == Instruction::And);
4228  // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
4229  ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
4230  APInt ShortMask = CI->getValue().trunc(MulWidth);
4231  Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
4232  Instruction *Zext =
4233  cast<Instruction>(Builder.CreateZExt(ShortAnd, BO->getType()));
4234  IC.Worklist.Add(Zext);
4235  IC.replaceInstUsesWith(*BO, Zext);
4236  } else {
4237  llvm_unreachable("Unexpected Binary operation");
4238  }
4239  IC.Worklist.Add(cast<Instruction>(U));
4240  }
4241  }
4242  if (isa<Instruction>(OtherVal))
4243  IC.Worklist.Add(cast<Instruction>(OtherVal));
4244 
4245  // The original icmp gets replaced with the overflow value, maybe inverted
4246  // depending on predicate.
4247  bool Inverse = false;
4248  switch (I.getPredicate()) {
4249  case ICmpInst::ICMP_NE:
4250  break;
4251  case ICmpInst::ICMP_EQ:
4252  Inverse = true;
4253  break;
4254  case ICmpInst::ICMP_UGT:
4255  case ICmpInst::ICMP_UGE:
4256  if (I.getOperand(0) == MulVal)
4257  break;
4258  Inverse = true;
4259  break;
4260  case ICmpInst::ICMP_ULT:
4261  case ICmpInst::ICMP_ULE:
4262  if (I.getOperand(1) == MulVal)
4263  break;
4264  Inverse = true;
4265  break;
4266  default:
4267  llvm_unreachable("Unexpected predicate");
4268  }
4269  if (Inverse) {
4270  Value *Res = Builder.CreateExtractValue(Call, 1);
4271  return BinaryOperator::CreateNot(Res);
4272  }
4273 
4274  return ExtractValueInst::Create(Call, 1);
4275 }
4276 
4277 /// When performing a comparison against a constant, it is possible that not all
4278 /// the bits in the LHS are demanded. This helper method computes the mask that
4279 /// IS demanded.
4280 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
4281  const APInt *RHS;
4282  if (!match(I.getOperand(1), m_APInt(RHS)))
4283  return APInt::getAllOnesValue(BitWidth);
4284 
4285  // If this is a normal comparison, it demands all bits. If it is a sign bit
4286  // comparison, it only demands the sign bit.
4287  bool UnusedBit;
4288  if (isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
4289  return APInt::getSignMask(BitWidth);
4290 
4291  switch (I.getPredicate()) {
4292  // For a UGT comparison, we don't care about any bits that
4293  // correspond to the trailing ones of the comparand. The value of these
4294  // bits doesn't impact the outcome of the comparison, because any value
4295  // greater than the RHS must differ in a bit higher than these due to carry.
4296  case ICmpInst::ICMP_UGT:
4297  return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
4298 
4299  // Similarly, for a ULT comparison, we don't care about the trailing zeros.
4300  // Any value less than the RHS must differ in a higher bit because of carries.
4301  case ICmpInst::ICMP_ULT:
4302  return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
4303 
4304  default:
4305  return APInt::getAllOnesValue(BitWidth);
4306  }
4307 }
4308 
4309 /// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
4310 /// should be swapped.
4311 /// The decision is based on how many times these two operands are reused
4312 /// as subtract operands and their positions in those instructions.
4313 /// The rationale is that several architectures use the same instruction for
4314 /// both subtract and cmp. Thus, it is better if the order of those operands
4315 /// match.
4316 /// \return true if Op0 and Op1 should be swapped.
4317 static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
4318  // Filter out pointer values as those cannot appear directly in subtract.
4319  // FIXME: we may want to go through inttoptrs or bitcasts.
4320  if (Op0->getType()->isPointerTy())
4321  return false;
4322  // If a subtract already has the same operands as a compare, swapping would be
4323  // bad. If a subtract has the same operands as a compare but in reverse order,
4324  // then swapping is good.
4325  int GoodToSwap = 0;
4326  for (const User *U : Op0->users()) {
4327  if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
4328  GoodToSwap++;
4329  else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
4330  GoodToSwap--;
4331  }
4332  return GoodToSwap > 0;
4333 }
4334 
4335 /// Check that one use is in the same block as the definition and all
4336 /// other uses are in blocks dominated by a given block.
4337 ///
4338 /// \param DI Definition
4339 /// \param UI Use
4340 /// \param DB Block that must dominate all uses of \p DI outside
4341 /// the parent block
4342 /// \return true when \p UI is the only use of \p DI in the parent block
4343 /// and all other uses of \p DI are in blocks dominated by \p DB.
4344 ///
4346  const Instruction *UI,
4347  const BasicBlock *DB) const {
4348  assert(DI && UI && "Instruction not defined\n");
4349  // Ignore incomplete definitions.
4350  if (!DI->getParent())
4351  return false;
4352  // DI and UI must be in the same block.
4353  if (DI->getParent() != UI->getParent())
4354  return false;
4355  // Protect from self-referencing blocks.
4356  if (DI->getParent() == DB)
4357  return false;
4358  for (const User *U : DI->users()) {
4359  auto *Usr = cast<Instruction>(U);
4360  if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4361  return false;
4362  }
4363  return true;
4364 }
4365 
4366 /// Return true when the instruction sequence within a block is select-cmp-br.
4367 static bool isChainSelectCmpBranch(const SelectInst *SI) {
4368  const BasicBlock *BB = SI->getParent();
4369  if (!BB)
4370  return false;
4371  auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4372  if (!BI || BI->getNumSuccessors() != 2)
4373  return false;
4374  auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4375  if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4376  return false;
4377  return true;
4378 }
4379 
4380 /// True when a select result is replaced by one of its operands
4381 /// in select-icmp sequence. This will eventually result in the elimination
4382 /// of the select.
4383 ///
4384 /// \param SI Select instruction
4385 /// \param Icmp Compare instruction
4386 /// \param SIOpd Operand that replaces the select
4387 ///
4388 /// Notes:
4389 /// - The replacement is global and requires dominator information
4390 /// - The caller is responsible for the actual replacement
4391 ///
4392 /// Example:
4393 ///
4394 /// entry:
4395 /// %4 = select i1 %3, %C* %0, %C* null
4396 /// %5 = icmp eq %C* %4, null
4397 /// br i1 %5, label %9, label %7
4398 /// ...
4399 /// ; <label>:7 ; preds = %entry
4400 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4401 /// ...
4402 ///
4403 /// can be transformed to
4404 ///
4405 /// %5 = icmp eq %C* %0, null
4406 /// %6 = select i1 %3, i1 %5, i1 true
4407 /// br i1 %6, label %9, label %7
4408 /// ...
4409 /// ; <label>:7 ; preds = %entry
4410 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
4411 ///
4412 /// Similar when the first operand of the select is a constant or/and
4413 /// the compare is for not equal rather than equal.
4414 ///
4415 /// NOTE: The function is only called when the select and compare constants
4416 /// are equal, the optimization can work only for EQ predicates. This is not a
4417 /// major restriction since a NE compare should be 'normalized' to an equal
4418 /// compare, which usually happens in the combiner and test case
4419 /// select-cmp-br.ll checks for it.
4421  const ICmpInst *Icmp,
4422  const unsigned SIOpd) {
4423  assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
4424  if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4425  BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4426  // The check for the single predecessor is not the best that can be
4427  // done. But it protects efficiently against cases like when SI's
4428  // home block has two successors, Succ and Succ1, and Succ1 predecessor
4429  // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4430  // replaced can be reached on either path. So the uniqueness check
4431  // guarantees that the path all uses of SI (outside SI's parent) are on
4432  // is disjoint from all other paths out of SI. But that information
4433  // is more expensive to compute, and the trade-off here is in favor
4434  // of compile-time. It should also be noticed that we check for a single
4435  // predecessor and not only uniqueness. This to handle the situation when
4436  // Succ and Succ1 points to the same basic block.
4437  if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
4438  NumSel++;
4439  SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
4440  return true;
4441  }
4442  }
4443  return false;
4444 }
4445 
4446 /// Try to fold the comparison based on range information we can get by checking
4447 /// whether bits are known to be zero or one in the inputs.
4448 Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) {
4449  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4450  Type *Ty = Op0->getType();
4451  ICmpInst::Predicate Pred = I.getPredicate();
4452 
4453  // Get scalar or pointer size.
4454  unsigned BitWidth = Ty->isIntOrIntVectorTy()
4455  ? Ty->getScalarSizeInBits()
4456  : DL.getIndexTypeSizeInBits(Ty->getScalarType());
4457 
4458  if (!BitWidth)
4459  return nullptr;
4460 
4461  KnownBits Op0Known(BitWidth);
4462  KnownBits Op1Known(BitWidth);
4463 
4464  if (SimplifyDemandedBits(&I, 0,
4465  getDemandedBitsLHSMask(I, BitWidth),
4466  Op0Known, 0))
4467  return &I;
4468 
4469  if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
4470  Op1Known, 0))
4471  return &I;
4472 
4473  // Given the known and unknown bits, compute a range that the LHS could be
4474  // in. Compute the Min, Max and RHS values based on the known bits. For the
4475  // EQ and NE we use unsigned values.
4476  APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
4477  APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
4478  if (I.isSigned()) {
4479  computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4480  computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4481  } else {
4482  computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4483  computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4484  }
4485 
4486  // If Min and Max are known to be the same, then SimplifyDemandedBits figured
4487  // out that the LHS or RHS is a constant. Constant fold this now, so that
4488  // code below can assume that Min != Max.
4489  if (!isa<Constant>(Op0) && Op0Min == Op0Max)
4490  return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
4491  if (!isa<Constant>(Op1) && Op1Min == Op1Max)
4492  return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
4493 
4494  // Based on the range information we know about the LHS, see if we can
4495  // simplify this comparison. For example, (x&4) < 8 is always true.
4496  switch (Pred) {
4497  default:
4498  llvm_unreachable("Unknown icmp opcode!");
4499  case ICmpInst::ICMP_EQ:
4500  case ICmpInst::ICMP_NE: {
4501  if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
4502  return Pred == CmpInst::ICMP_EQ
4503  ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
4504  : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4505  }
4506 
4507  // If all bits are known zero except for one, then we know at most one bit
4508  // is set. If the comparison is against zero, then this is a check to see if
4509  // *that* bit is set.
4510  APInt Op0KnownZeroInverted = ~Op0Known.Zero;
4511  if (Op1Known.isZero()) {
4512  // If the LHS is an AND with the same constant, look through it.
4513  Value *LHS = nullptr;
4514  const APInt *LHSC;
4515  if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
4516  *LHSC != Op0KnownZeroInverted)
4517  LHS = Op0;
4518 
4519  Value *X;
4520  if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
4521  APInt ValToCheck = Op0KnownZeroInverted;
4522  Type *XTy = X->getType();
4523  if (ValToCheck.isPowerOf2()) {
4524  // ((1 << X) & 8) == 0 -> X != 3
4525  // ((1 << X) & 8) != 0 -> X == 3
4526  auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4527  auto NewPred = ICmpInst::getInversePredicate(Pred);
4528  return new ICmpInst(NewPred, X, CmpC);
4529  } else if ((++ValToCheck).isPowerOf2()) {
4530  // ((1 << X) & 7) == 0 -> X >= 3
4531  // ((1 << X) & 7) != 0 -> X < 3
4532  auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4533  auto NewPred =
4535  return new ICmpInst(NewPred, X, CmpC);
4536  }
4537  }
4538 
4539  // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
4540  const APInt *CI;
4541  if (Op0KnownZeroInverted.isOneValue() &&
4542  match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
4543  // ((8 >>u X) & 1) == 0 -> X != 3
4544  // ((8 >>u X) & 1) != 0 -> X == 3
4545  unsigned CmpVal = CI->countTrailingZeros();
4546  auto NewPred = ICmpInst::getInversePredicate(Pred);
4547  return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
4548  }
4549  }
4550  break;
4551  }
4552  case ICmpInst::ICMP_ULT: {
4553  if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
4554  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4555  if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
4556  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4557  if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
4558  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4559 
4560  const APInt *CmpC;
4561  if (match(Op1, m_APInt(CmpC))) {
4562  // A <u C -> A == C-1 if min(A)+1 == C
4563  if (*CmpC == Op0Min + 1)
4564  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4565  ConstantInt::get(Op1->getType(), *CmpC - 1));
4566  // X <u C --> X == 0, if the number of zero bits in the bottom of X
4567  // exceeds the log2 of C.
4568  if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
4569  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4570  Constant::getNullValue(Op1->getType()));
4571  }
4572  break;
4573  }
4574  case ICmpInst::ICMP_UGT: {
4575  if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
4576  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4577  if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
4578  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4579  if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
4580  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4581 
4582  const APInt *CmpC;
4583  if (match(Op1, m_APInt(CmpC))) {
4584  // A >u C -> A == C+1 if max(a)-1 == C
4585  if (*CmpC == Op0Max - 1)
4586  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4587  ConstantInt::get(Op1->getType(), *CmpC + 1));
4588  // X >u C --> X != 0, if the number of zero bits in the bottom of X
4589  // exceeds the log2 of C.
4590  if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
4591  return new ICmpInst(ICmpInst::ICMP_NE, Op0,
4592  Constant::getNullValue(Op1->getType()));
4593  }
4594  break;
4595  }
4596  case ICmpInst::ICMP_SLT: {
4597  if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
4598  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4599  if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
4600  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4601  if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
4602  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4603  const APInt *CmpC;
4604  if (match(Op1, m_APInt(CmpC))) {
4605  if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
4606  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4607  ConstantInt::get(Op1->getType(), *CmpC - 1));
4608  }
4609  break;
4610  }
4611  case ICmpInst::ICMP_SGT: {
4612  if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
4613  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4614  if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
4615  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4616  if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
4617  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4618  const APInt *CmpC;
4619  if (match(Op1, m_APInt(CmpC))) {
4620  if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
4621  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4622  ConstantInt::get(Op1->getType(), *CmpC + 1));
4623  }
4624  break;
4625  }
4626  case ICmpInst::ICMP_SGE:
4627  assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
4628  if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
4629  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4630  if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
4631  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4632  if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
4633  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4634  break;
4635  case ICmpInst::ICMP_SLE:
4636  assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
4637  if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
4638  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4639  if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
4640  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4641  if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
4642  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4643  break;
4644  case ICmpInst::ICMP_UGE:
4645  assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
4646  if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
4647  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4648  if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
4649  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4650  if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
4651  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4652  break;
4653  case ICmpInst::ICMP_ULE:
4654  assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
4655  if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
4656  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4657  if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
4658  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4659  if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
4660  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4661  break;
4662  }
4663 
4664  // Turn a signed comparison into an unsigned one if both operands are known to
4665  // have the same sign.
4666  if (I.isSigned() &&
4667  ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
4668  (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
4669  return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
4670 
4671  return nullptr;
4672 }
4673 
4674 /// If we have an icmp le or icmp ge instruction with a constant operand, turn
4675 /// it into the appropriate icmp lt or icmp gt instruction. This transform
4676 /// allows them to be folded in visitICmpInst.
4678  ICmpInst::Predicate Pred = I.getPredicate();
4679  if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGE &&
4680  Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_UGE)
4681  return nullptr;
4682 
4683  Value *Op0 = I.getOperand(0);
4684  Value *Op1 = I.getOperand(1);
4685  auto *Op1C = dyn_cast<Constant>(Op1);
4686  if (!Op1C)
4687  return nullptr;
4688 
4689  // Check if the constant operand can be safely incremented/decremented without
4690  // overflowing/underflowing. For scalars, SimplifyICmpInst has already handled
4691  // the edge cases for us, so we just assert on them. For vectors, we must
4692  // handle the edge cases.
4693  Type *Op1Type = Op1->getType();
4694  bool IsSigned = I.isSigned();
4695  bool IsLE = (Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_ULE);
4696  auto *CI = dyn_cast<ConstantInt>(Op1C);
4697  if (CI) {
4698  // A <= MAX -> TRUE ; A >= MIN -> TRUE
4699  assert(IsLE ? !CI->isMaxValue(IsSigned) : !CI->isMinValue(IsSigned));
4700  } else if (Op1Type->isVectorTy()) {
4701  // TODO? If the edge cases for vectors were guaranteed to be handled as they
4702  // are for scalar, we could remove the min/max checks. However, to do that,
4703  // we would have to use insertelement/shufflevector to replace edge values.
4704  unsigned NumElts = Op1Type->getVectorNumElements();
4705  for (unsigned i = 0; i != NumElts; ++i) {
4706  Constant *Elt = Op1C->getAggregateElement(i);
4707  if (!Elt)
4708  return nullptr;
4709 
4710  if (isa<UndefValue>(Elt))
4711  continue;
4712 
4713  // Bail out if we can't determine if this constant is min/max or if we
4714  // know that this constant is min/max.
4715  auto *CI = dyn_cast<ConstantInt>(Elt);
4716  if (!CI || (IsLE ? CI->isMaxValue(IsSigned) : CI->isMinValue(IsSigned)))
4717  return nullptr;
4718  }
4719  } else {
4720  // ConstantExpr?
4721  return nullptr;
4722  }
4723 
4724  // Increment or decrement the constant and set the new comparison predicate:
4725  // ULE -> ULT ; UGE -> UGT ; SLE -> SLT ; SGE -> SGT
4726  Constant *OneOrNegOne = ConstantInt::get(Op1Type, IsLE ? 1 : -1, true);
4728  NewPred = IsSigned ? ICmpInst::getSignedPredicate(NewPred) : NewPred;
4729  return new ICmpInst(NewPred, Op0, ConstantExpr::getAdd(Op1C, OneOrNegOne));
4730 }
4731 
4732 /// Integer compare with boolean values can always be turned into bitwise ops.
4734  InstCombiner::BuilderTy &Builder) {
4735  Value *A = I.getOperand(0), *B = I.getOperand(1);
4736  assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only");
4737 
4738  // A boolean compared to true/false can be simplified to Op0/true/false in
4739  // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
4740  // Cases not handled by InstSimplify are always 'not' of Op0.
4741  if (match(B, m_Zero())) {
4742  switch (I.getPredicate()) {
4743  case CmpInst::ICMP_EQ: // A == 0 -> !A
4744  case CmpInst::ICMP_ULE: // A <=u 0 -> !A
4745  case CmpInst::ICMP_SGE: // A >=s 0 -> !A
4746  return BinaryOperator::CreateNot(A);
4747  default:
4748  llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4749  }
4750  } else if (match(B, m_One())) {
4751  switch (I.getPredicate()) {
4752  case CmpInst::ICMP_NE: // A != 1 -> !A
4753  case CmpInst::ICMP_ULT: // A <u 1 -> !A
4754  case CmpInst::ICMP_SGT: // A >s -1 -> !A
4755  return BinaryOperator::CreateNot(A);
4756  default:
4757  llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4758  }
4759  }
4760 
4761  switch (I.getPredicate()) {
4762  default:
4763  llvm_unreachable("Invalid icmp instruction!");
4764  case ICmpInst::ICMP_EQ:
4765  // icmp eq i1 A, B -> ~(A ^ B)
4766  return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
4767 
4768  case ICmpInst::ICMP_NE:
4769  // icmp ne i1 A, B -> A ^ B
4770  return BinaryOperator::CreateXor(A, B);
4771 
4772  case ICmpInst::ICMP_UGT:
4773  // icmp ugt -> icmp ult
4774  std::swap(A, B);
4776  case ICmpInst::ICMP_ULT:
4777  // icmp ult i1 A, B -> ~A & B
4778  return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
4779 
4780  case ICmpInst::ICMP_SGT:
4781  // icmp sgt -> icmp slt
4782  std::swap(A, B);
4784  case ICmpInst::ICMP_SLT:
4785  // icmp slt i1 A, B -> A & ~B
4786  return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
4787 
4788  case ICmpInst::ICMP_UGE:
4789  // icmp uge -> icmp ule
4790  std::swap(A, B);
4792  case ICmpInst::ICMP_ULE:
4793  // icmp ule i1 A, B -> ~A | B
4794  return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
4795 
4796  case ICmpInst::ICMP_SGE:
4797  // icmp sge -> icmp sle
4798  std::swap(A, B);
4800  case ICmpInst::ICMP_SLE:
4801  // icmp sle i1 A, B -> A | ~B
4802  return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
4803  }
4804 }
4805 
4806 // Transform pattern like:
4807 // (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
4808 // (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
4809 // Into:
4810 // (X l>> Y) != 0
4811 // (X l>> Y) == 0
4813  InstCombiner::BuilderTy &Builder) {
4814  ICmpInst::Predicate Pred, NewPred;
4815  Value *X, *Y;
4816  if (match(&Cmp,
4817  m_c_ICmp(Pred, m_OneUse(m_Shl(m_One(), m_Value(Y))), m_Value(X)))) {
4818  // We want X to be the icmp's second operand, so swap predicate if it isn't.
4819  if (Cmp.getOperand(0) == X)
4820  Pred = Cmp.getSwappedPredicate();
4821 
4822  switch (Pred) {
4823  case ICmpInst::ICMP_ULE:
4824  NewPred = ICmpInst::ICMP_NE;
4825  break;
4826  case ICmpInst::ICMP_UGT:
4827  NewPred = ICmpInst::ICMP_EQ;
4828  break;
4829  default:
4830  return nullptr;
4831  }
4832  } else if (match(&Cmp, m_c_ICmp(Pred,
4834  m_Not(m_Shl(m_AllOnes(), m_Value(Y))),
4835  m_Add(m_Shl(m_One(), m_Value(Y)),
4836  m_AllOnes()))),
4837  m_Value(X)))) {
4838  // The variant with 'add' is not canonical, (the variant with 'not' is)
4839  // we only get it because it has extra uses, and can't be canonicalized,
4840 
4841  // We want X to be the icmp's second operand, so swap predicate if it isn't.
4842  if (Cmp.getOperand(0) == X)
4843  Pred = Cmp.getSwappedPredicate();
4844 
4845  switch (Pred) {
4846  case ICmpInst::ICMP_ULT:
4847  NewPred = ICmpInst::ICMP_NE;
4848  break;
4849  case ICmpInst::ICMP_UGE:
4850  NewPred = ICmpInst::ICMP_EQ;
4851  break;
4852  default:
4853  return nullptr;
4854  }
4855  } else
4856  return nullptr;
4857 
4858  Value *NewX = Builder.CreateLShr(X, Y, X->getName() + ".highbits");
4859  Constant *Zero = Constant::getNullValue(NewX->getType());
4860  return CmpInst::Create(Instruction::ICmp, NewPred, NewX, Zero);
4861 }
4862 
4864  InstCombiner::BuilderTy &Builder) {
4865  // If both arguments of the cmp are shuffles that use the same mask and
4866  // shuffle within a single vector, move the shuffle after the cmp.
4867  Value *LHS = Cmp.getOperand(0), *RHS = Cmp.getOperand(1);
4868  Value *V1, *V2;
4869  Constant *M;
4870  if (match(LHS, m_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(M))) &&
4871  match(RHS, m_ShuffleVector(m_Value(V2), m_Undef(), m_Specific(M))) &&
4872  V1->getType() == V2->getType() &&
4873  (LHS->hasOneUse() || RHS->hasOneUse())) {
4874  // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
4876  Value *NewCmp = isa<ICmpInst>(Cmp) ? Builder.CreateICmp(P, V1, V2)
4877  : Builder.CreateFCmp(P, V1, V2);
4878  return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()), M);
4879  }
4880  return nullptr;
4881 }
4882 
4884  bool Changed = false;
4885  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4886  unsigned Op0Cplxity = getComplexity(Op0);
4887  unsigned Op1Cplxity = getComplexity(Op1);
4888 
4889  /// Orders the operands of the compare so that they are listed from most
4890  /// complex to least complex. This puts constants before unary operators,
4891  /// before binary operators.
4892  if (Op0Cplxity < Op1Cplxity ||
4893  (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
4894  I.swapOperands();
4895  std::swap(Op0, Op1);
4896  Changed = true;
4897  }
4898 
4899  if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1,
4900  SQ.getWithInstruction(&I)))
4901  return replaceInstUsesWith(I, V);
4902 
4903  // Comparing -val or val with non-zero is the same as just comparing val
4904  // ie, abs(val) != 0 -> val != 0
4905  if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
4906  Value *Cond, *SelectTrue, *SelectFalse;
4907  if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
4908  m_Value(SelectFalse)))) {
4909  if (Value *V = dyn_castNegVal(SelectTrue)) {
4910  if (V == SelectFalse)
4911  return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4912  }
4913  else if (Value *V = dyn_castNegVal(SelectFalse)) {
4914  if (V == SelectTrue)
4915  return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4916  }
4917  }
4918  }
4919 
4920  if (Op0->getType()->isIntOrIntVectorTy(1))
4921  if (Instruction *Res = canonicalizeICmpBool(I, Builder))
4922  return Res;
4923 
4924  if (ICmpInst *NewICmp = canonicalizeCmpWithConstant(I))
4925  return NewICmp;
4926 
4927  if (Instruction *Res = foldICmpWithConstant(I))
4928  return Res;
4929 
4930  if (Instruction *Res = foldICmpWithDominatingICmp(I))
4931  return Res;
4932 
4933  if (Instruction *Res = foldICmpUsingKnownBits(I))
4934  return Res;
4935 
4936  // Test if the ICmpInst instruction is used exclusively by a select as
4937  // part of a minimum or maximum operation. If so, refrain from doing
4938  // any other folding. This helps out other analyses which understand
4939  // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
4940  // and CodeGen. And in this case, at least one of the comparison
4941  // operands has at least one user besides the compare (the select),
4942  // which would often largely negate the benefit of folding anyway.
4943  //
4944  // Do the same for the other patterns recognized by matchSelectPattern.
4945  if (I.hasOneUse())
4946  if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
4947  Value *A, *B;
4948  SelectPatternResult SPR = matchSelectPattern(SI, A, B);
4949  if (SPR.Flavor != SPF_UNKNOWN)
4950  return nullptr;
4951  }
4952 
4953  // Do this after checking for min/max to prevent infinite looping.
4954  if (Instruction *Res = foldICmpWithZero(I))
4955  return Res;
4956 
4957  // FIXME: We only do this after checking for min/max to prevent infinite
4958  // looping caused by a reverse canonicalization of these patterns for min/max.
4959  // FIXME: The organization of folds is a mess. These would naturally go into
4960  // canonicalizeCmpWithConstant(), but we can't move all of the above folds
4961  // down here after the min/max restriction.
4962  ICmpInst::Predicate Pred = I.getPredicate();
4963  const APInt *C;
4964  if (match(Op1, m_APInt(C))) {
4965  // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
4966  if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
4967  Constant *Zero = Constant::getNullValue(Op0->getType());
4968  return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
4969  }
4970 
4971  // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
4972  if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
4973  Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
4974  return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
4975  }
4976  }
4977 
4978  if (Instruction *Res = foldICmpInstWithConstant(I))
4979  return Res;
4980 
4981  if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
4982  return Res;
4983 
4984  // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4985  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
4986  if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I))
4987  return NI;
4988  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
4989  if (Instruction *NI = foldGEPICmp(GEP, Op0,
4991  return NI;
4992 
4993  // Try to optimize equality comparisons against alloca-based pointers.
4994  if (Op0->getType()->isPointerTy() && I.isEquality()) {
4995  assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?");
4996  if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op0, DL)))
4997  if (Instruction *New = foldAllocaCmp(I, Alloca, Op1))
4998  return New;
4999  if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op1, DL)))
5000  if (Instruction *New = foldAllocaCmp(I, Alloca, Op0))
5001  return New;
5002  }
5003 
5004  if (Instruction *Res = foldICmpBitCast(I, Builder))
5005  return Res;
5006 
5007  if (isa<CastInst>(Op0)) {
5008  // Handle the special case of: icmp (cast bool to X), <cst>
5009  // This comes up when you have code like
5010  // int X = A < B;
5011  // if (X) ...
5012  // For generality, we handle any zero-extension of any operand comparison
5013  // with a constant or another cast from the same type.
5014  if (isa<Constant>(Op1) || isa<CastInst>(Op1))
5015  if (Instruction *R = foldICmpWithCastAndCast(I))
5016  return R;
5017  }
5018 
5019  if (Instruction *Res = foldICmpBinOp(I))
5020  return Res;
5021 
5022  if (Instruction *Res = foldICmpWithMinMax(I))
5023  return Res;
5024 
5025  {
5026  Value *A, *B;
5027  // Transform (A & ~B) == 0 --> (A & B) != 0
5028  // and (A & ~B) != 0 --> (A & B) == 0
5029  // if A is a power of 2.
5030  if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
5031  match(Op1, m_Zero()) &&
5032  isKnownToBeAPowerOfTwo(A, false, 0, &I) && I.isEquality())
5033  return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(A, B),
5034  Op1);