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  NewInsts[CI] = NewInsts[CI->getOperand(0)];
707  continue;
708  }
709  if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
710  Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
711  : GEP->getOperand(1);
712  setInsertionPoint(Builder, GEP);
713  // Indices might need to be sign extended. GEPs will magically do
714  // this, but we need to do it ourselves here.
715  if (Index->getType()->getScalarSizeInBits() !=
716  NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
717  Index = Builder.CreateSExtOrTrunc(
718  Index, NewInsts[GEP->getOperand(0)]->getType(),
719  GEP->getOperand(0)->getName() + ".sext");
720  }
721 
722  auto *Op = NewInsts[GEP->getOperand(0)];
723  if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
724  NewInsts[GEP] = Index;
725  else
726  NewInsts[GEP] = Builder.CreateNSWAdd(
727  Op, Index, GEP->getOperand(0)->getName() + ".add");
728  continue;
729  }
730  if (isa<PHINode>(Val))
731  continue;
732 
733  llvm_unreachable("Unexpected instruction type");
734  }
735 
736  // Add the incoming values to the PHI nodes.
737  for (Value *Val : Explored) {
738  if (Val == Base)
739  continue;
740  // All the instructions have been created, we can now add edges to the
741  // phi nodes.
742  if (auto *PHI = dyn_cast<PHINode>(Val)) {
743  PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
744  for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
745  Value *NewIncoming = PHI->getIncomingValue(I);
746 
747  if (NewInsts.find(NewIncoming) != NewInsts.end())
748  NewIncoming = NewInsts[NewIncoming];
749 
750  NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
751  }
752  }
753  }
754 
755  for (Value *Val : Explored) {
756  if (Val == Base)
757  continue;
758 
759  // Depending on the type, for external users we have to emit
760  // a GEP or a GEP + ptrtoint.
761  setInsertionPoint(Builder, Val, false);
762 
763  // If required, create an inttoptr instruction for Base.
764  Value *NewBase = Base;
765  if (!Base->getType()->isPointerTy())
766  NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
767  Start->getName() + "to.ptr");
768 
769  Value *GEP = Builder.CreateInBoundsGEP(
770  Start->getType()->getPointerElementType(), NewBase,
771  makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
772 
773  if (!Val->getType()->isPointerTy()) {
774  Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
775  Val->getName() + ".conv");
776  GEP = Cast;
777  }
778  Val->replaceAllUsesWith(GEP);
779  }
780 
781  return NewInsts[Start];
782 }
783 
784 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
785 /// the input Value as a constant indexed GEP. Returns a pair containing
786 /// the GEPs Pointer and Index.
787 static std::pair<Value *, Value *>
789  Type *IndexType = IntegerType::get(V->getContext(),
791 
793  while (true) {
794  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
795  // We accept only inbouds GEPs here to exclude the possibility of
796  // overflow.
797  if (!GEP->isInBounds())
798  break;
799  if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
800  GEP->getType() == V->getType()) {
801  V = GEP->getOperand(0);
802  Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
803  Index = ConstantExpr::getAdd(
804  Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
805  continue;
806  }
807  break;
808  }
809  if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
810  if (!CI->isNoopCast(DL))
811  break;
812  V = CI->getOperand(0);
813  continue;
814  }
815  if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
816  if (!CI->isNoopCast(DL))
817  break;
818  V = CI->getOperand(0);
819  continue;
820  }
821  break;
822  }
823  return {V, Index};
824 }
825 
826 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
827 /// We can look through PHIs, GEPs and casts in order to determine a common base
828 /// between GEPLHS and RHS.
830  ICmpInst::Predicate Cond,
831  const DataLayout &DL) {
832  if (!GEPLHS->hasAllConstantIndices())
833  return nullptr;
834 
835  // Make sure the pointers have the same type.
836  if (GEPLHS->getType() != RHS->getType())
837  return nullptr;
838 
839  Value *PtrBase, *Index;
840  std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
841 
842  // The set of nodes that will take part in this transformation.
843  SetVector<Value *> Nodes;
844 
845  if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
846  return nullptr;
847 
848  // We know we can re-write this as
849  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
850  // Since we've only looked through inbouds GEPs we know that we
851  // can't have overflow on either side. We can therefore re-write
852  // this as:
853  // OFFSET1 cmp OFFSET2
854  Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
855 
856  // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
857  // GEP having PtrBase as the pointer base, and has returned in NewRHS the
858  // offset. Since Index is the offset of LHS to the base pointer, we will now
859  // compare the offsets instead of comparing the pointers.
860  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
861 }
862 
863 /// Fold comparisons between a GEP instruction and something else. At this point
864 /// we know that the GEP is on the LHS of the comparison.
865 Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
866  ICmpInst::Predicate Cond,
867  Instruction &I) {
868  // Don't transform signed compares of GEPs into index compares. Even if the
869  // GEP is inbounds, the final add of the base pointer can have signed overflow
870  // and would change the result of the icmp.
871  // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
872  // the maximum signed value for the pointer type.
873  if (ICmpInst::isSigned(Cond))
874  return nullptr;
875 
876  // Look through bitcasts and addrspacecasts. We do not however want to remove
877  // 0 GEPs.
878  if (!isa<GetElementPtrInst>(RHS))
879  RHS = RHS->stripPointerCasts();
880 
881  Value *PtrBase = GEPLHS->getOperand(0);
882  if (PtrBase == RHS && GEPLHS->isInBounds()) {
883  // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
884  // This transformation (ignoring the base and scales) is valid because we
885  // know pointers can't overflow since the gep is inbounds. See if we can
886  // output an optimized form.
887  Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
888 
889  // If not, synthesize the offset the hard way.
890  if (!Offset)
891  Offset = EmitGEPOffset(GEPLHS);
892  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
893  Constant::getNullValue(Offset->getType()));
894  } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
895  // If the base pointers are different, but the indices are the same, just
896  // compare the base pointer.
897  if (PtrBase != GEPRHS->getOperand(0)) {
898  bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
899  IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
900  GEPRHS->getOperand(0)->getType();
901  if (IndicesTheSame)
902  for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
903  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
904  IndicesTheSame = false;
905  break;
906  }
907 
908  // If all indices are the same, just compare the base pointers.
909  Type *BaseType = GEPLHS->getOperand(0)->getType();
910  if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
911  return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
912 
913  // If we're comparing GEPs with two base pointers that only differ in type
914  // and both GEPs have only constant indices or just one use, then fold
915  // the compare with the adjusted indices.
916  if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
917  (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
918  (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
919  PtrBase->stripPointerCasts() ==
920  GEPRHS->getOperand(0)->stripPointerCasts()) {
921  Value *LOffset = EmitGEPOffset(GEPLHS);
922  Value *ROffset = EmitGEPOffset(GEPRHS);
923 
924  // If we looked through an addrspacecast between different sized address
925  // spaces, the LHS and RHS pointers are different sized
926  // integers. Truncate to the smaller one.
927  Type *LHSIndexTy = LOffset->getType();
928  Type *RHSIndexTy = ROffset->getType();
929  if (LHSIndexTy != RHSIndexTy) {
930  if (LHSIndexTy->getPrimitiveSizeInBits() <
931  RHSIndexTy->getPrimitiveSizeInBits()) {
932  ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
933  } else
934  LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
935  }
936 
937  Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
938  LOffset, ROffset);
939  return replaceInstUsesWith(I, Cmp);
940  }
941 
942  // Otherwise, the base pointers are different and the indices are
943  // different. Try convert this to an indexed compare by looking through
944  // PHIs/casts.
945  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
946  }
947 
948  // If one of the GEPs has all zero indices, recurse.
949  if (GEPLHS->hasAllZeroIndices())
950  return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
952 
953  // If the other GEP has all zero indices, recurse.
954  if (GEPRHS->hasAllZeroIndices())
955  return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
956 
957  bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
958  if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
959  // If the GEPs only differ by one index, compare it.
960  unsigned NumDifferences = 0; // Keep track of # differences.
961  unsigned DiffOperand = 0; // The operand that differs.
962  for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
963  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
964  if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
965  GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
966  // Irreconcilable differences.
967  NumDifferences = 2;
968  break;
969  } else {
970  if (NumDifferences++) break;
971  DiffOperand = i;
972  }
973  }
974 
975  if (NumDifferences == 0) // SAME GEP?
976  return replaceInstUsesWith(I, // No comparison is needed here.
978 
979  else if (NumDifferences == 1 && GEPsInBounds) {
980  Value *LHSV = GEPLHS->getOperand(DiffOperand);
981  Value *RHSV = GEPRHS->getOperand(DiffOperand);
982  // Make sure we do a signed comparison here.
983  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
984  }
985  }
986 
987  // Only lower this if the icmp is the only user of the GEP or if we expect
988  // the result to fold to a constant!
989  if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
990  (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
991  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
992  Value *L = EmitGEPOffset(GEPLHS);
993  Value *R = EmitGEPOffset(GEPRHS);
994  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
995  }
996  }
997 
998  // Try convert this to an indexed compare by looking through PHIs/casts as a
999  // last resort.
1000  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1001 }
1002 
1003 Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI,
1004  const AllocaInst *Alloca,
1005  const Value *Other) {
1006  assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1007 
1008  // It would be tempting to fold away comparisons between allocas and any
1009  // pointer not based on that alloca (e.g. an argument). However, even
1010  // though such pointers cannot alias, they can still compare equal.
1011  //
1012  // But LLVM doesn't specify where allocas get their memory, so if the alloca
1013  // doesn't escape we can argue that it's impossible to guess its value, and we
1014  // can therefore act as if any such guesses are wrong.
1015  //
1016  // The code below checks that the alloca doesn't escape, and that it's only
1017  // used in a comparison once (the current instruction). The
1018  // single-comparison-use condition ensures that we're trivially folding all
1019  // comparisons against the alloca consistently, and avoids the risk of
1020  // erroneously folding a comparison of the pointer with itself.
1021 
1022  unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1023 
1025  for (const Use &U : Alloca->uses()) {
1026  if (Worklist.size() >= MaxIter)
1027  return nullptr;
1028  Worklist.push_back(&U);
1029  }
1030 
1031  unsigned NumCmps = 0;
1032  while (!Worklist.empty()) {
1033  assert(Worklist.size() <= MaxIter);
1034  const Use *U = Worklist.pop_back_val();
1035  const Value *V = U->getUser();
1036  --MaxIter;
1037 
1038  if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1039  isa<SelectInst>(V)) {
1040  // Track the uses.
1041  } else if (isa<LoadInst>(V)) {
1042  // Loading from the pointer doesn't escape it.
1043  continue;
1044  } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1045  // Storing *to* the pointer is fine, but storing the pointer escapes it.
1046  if (SI->getValueOperand() == U->get())
1047  return nullptr;
1048  continue;
1049  } else if (isa<ICmpInst>(V)) {
1050  if (NumCmps++)
1051  return nullptr; // Found more than one cmp.
1052  continue;
1053  } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1054  switch (Intrin->getIntrinsicID()) {
1055  // These intrinsics don't escape or compare the pointer. Memset is safe
1056  // because we don't allow ptrtoint. Memcpy and memmove are safe because
1057  // we don't allow stores, so src cannot point to V.
1058  case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1059  case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1060  continue;
1061  default:
1062  return nullptr;
1063  }
1064  } else {
1065  return nullptr;
1066  }
1067  for (const Use &U : V->uses()) {
1068  if (Worklist.size() >= MaxIter)
1069  return nullptr;
1070  Worklist.push_back(&U);
1071  }
1072  }
1073 
1074  Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1075  return replaceInstUsesWith(
1076  ICI,
1078 }
1079 
1080 /// Fold "icmp pred (X+C), X".
1081 Instruction *InstCombiner::foldICmpAddOpConst(Value *X, const APInt &C,
1082  ICmpInst::Predicate Pred) {
1083  // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1084  // so the values can never be equal. Similarly for all other "or equals"
1085  // operators.
1086  assert(!!C && "C should not be zero!");
1087 
1088  // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1089  // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1090  // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1091  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1092  Constant *R = ConstantInt::get(X->getType(),
1093  APInt::getMaxValue(C.getBitWidth()) - C);
1094  return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1095  }
1096 
1097  // (X+1) >u X --> X <u (0-1) --> X != 255
1098  // (X+2) >u X --> X <u (0-2) --> X <u 254
1099  // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1100  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1101  return new ICmpInst(ICmpInst::ICMP_ULT, X,
1102  ConstantInt::get(X->getType(), -C));
1103 
1105 
1106  // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1107  // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1108  // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1109  // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1110  // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1111  // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1112  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1113  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1114  ConstantInt::get(X->getType(), SMax - C));
1115 
1116  // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1117  // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1118  // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1119  // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1120  // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1121  // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1122 
1123  assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1124  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1125  ConstantInt::get(X->getType(), SMax - (C - 1)));
1126 }
1127 
1128 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1129 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1130 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1131 Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A,
1132  const APInt &AP1,
1133  const APInt &AP2) {
1134  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1135 
1136  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1137  if (I.getPredicate() == I.ICMP_NE)
1138  Pred = CmpInst::getInversePredicate(Pred);
1139  return new ICmpInst(Pred, LHS, RHS);
1140  };
1141 
1142  // Don't bother doing any work for cases which InstSimplify handles.
1143  if (AP2.isNullValue())
1144  return nullptr;
1145 
1146  bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1147  if (IsAShr) {
1148  if (AP2.isAllOnesValue())
1149  return nullptr;
1150  if (AP2.isNegative() != AP1.isNegative())
1151  return nullptr;
1152  if (AP2.sgt(AP1))
1153  return nullptr;
1154  }
1155 
1156  if (!AP1)
1157  // 'A' must be large enough to shift out the highest set bit.
1158  return getICmp(I.ICMP_UGT, A,
1159  ConstantInt::get(A->getType(), AP2.logBase2()));
1160 
1161  if (AP1 == AP2)
1162  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1163 
1164  int Shift;
1165  if (IsAShr && AP1.isNegative())
1166  Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1167  else
1168  Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1169 
1170  if (Shift > 0) {
1171  if (IsAShr && AP1 == AP2.ashr(Shift)) {
1172  // There are multiple solutions if we are comparing against -1 and the LHS
1173  // of the ashr is not a power of two.
1174  if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1175  return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1176  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1177  } else if (AP1 == AP2.lshr(Shift)) {
1178  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1179  }
1180  }
1181 
1182  // Shifting const2 will never be equal to const1.
1183  // FIXME: This should always be handled by InstSimplify?
1184  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1185  return replaceInstUsesWith(I, TorF);
1186 }
1187 
1188 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1189 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1190 Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A,
1191  const APInt &AP1,
1192  const APInt &AP2) {
1193  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1194 
1195  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1196  if (I.getPredicate() == I.ICMP_NE)
1197  Pred = CmpInst::getInversePredicate(Pred);
1198  return new ICmpInst(Pred, LHS, RHS);
1199  };
1200 
1201  // Don't bother doing any work for cases which InstSimplify handles.
1202  if (AP2.isNullValue())
1203  return nullptr;
1204 
1205  unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1206 
1207  if (!AP1 && AP2TrailingZeros != 0)
1208  return getICmp(
1209  I.ICMP_UGE, A,
1210  ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1211 
1212  if (AP1 == AP2)
1213  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1214 
1215  // Get the distance between the lowest bits that are set.
1216  int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1217 
1218  if (Shift > 0 && AP2.shl(Shift) == AP1)
1219  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1220 
1221  // Shifting const2 will never be equal to const1.
1222  // FIXME: This should always be handled by InstSimplify?
1223  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1224  return replaceInstUsesWith(I, TorF);
1225 }
1226 
1227 /// The caller has matched a pattern of the form:
1228 /// I = icmp ugt (add (add A, B), CI2), CI1
1229 /// If this is of the form:
1230 /// sum = a + b
1231 /// if (sum+128 >u 255)
1232 /// Then replace it with llvm.sadd.with.overflow.i8.
1233 ///
1235  ConstantInt *CI2, ConstantInt *CI1,
1236  InstCombiner &IC) {
1237  // The transformation we're trying to do here is to transform this into an
1238  // llvm.sadd.with.overflow. To do this, we have to replace the original add
1239  // with a narrower add, and discard the add-with-constant that is part of the
1240  // range check (if we can't eliminate it, this isn't profitable).
1241 
1242  // In order to eliminate the add-with-constant, the compare can be its only
1243  // use.
1244  Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1245  if (!AddWithCst->hasOneUse())
1246  return nullptr;
1247 
1248  // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1249  if (!CI2->getValue().isPowerOf2())
1250  return nullptr;
1251  unsigned NewWidth = CI2->getValue().countTrailingZeros();
1252  if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1253  return nullptr;
1254 
1255  // The width of the new add formed is 1 more than the bias.
1256  ++NewWidth;
1257 
1258  // Check to see that CI1 is an all-ones value with NewWidth bits.
1259  if (CI1->getBitWidth() == NewWidth ||
1260  CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1261  return nullptr;
1262 
1263  // This is only really a signed overflow check if the inputs have been
1264  // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1265  // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1266  unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1267  if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1268  IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1269  return nullptr;
1270 
1271  // In order to replace the original add with a narrower
1272  // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1273  // and truncates that discard the high bits of the add. Verify that this is
1274  // the case.
1275  Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1276  for (User *U : OrigAdd->users()) {
1277  if (U == AddWithCst)
1278  continue;
1279 
1280  // Only accept truncates for now. We would really like a nice recursive
1281  // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1282  // chain to see which bits of a value are actually demanded. If the
1283  // original add had another add which was then immediately truncated, we
1284  // could still do the transformation.
1285  TruncInst *TI = dyn_cast<TruncInst>(U);
1286  if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1287  return nullptr;
1288  }
1289 
1290  // If the pattern matches, truncate the inputs to the narrower type and
1291  // use the sadd_with_overflow intrinsic to efficiently compute both the
1292  // result and the overflow bit.
1293  Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1295  I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1296 
1297  InstCombiner::BuilderTy &Builder = IC.Builder;
1298 
1299  // Put the new code above the original add, in case there are any uses of the
1300  // add between the add and the compare.
1301  Builder.SetInsertPoint(OrigAdd);
1302 
1303  Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1304  Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1305  CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1306  Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1307  Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1308 
1309  // The inner add was the result of the narrow add, zero extended to the
1310  // wider type. Replace it with the result computed by the intrinsic.
1311  IC.replaceInstUsesWith(*OrigAdd, ZExt);
1312 
1313  // The original icmp gets replaced with the overflow value.
1314  return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1315 }
1316 
1317 // Handle (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1318 Instruction *InstCombiner::foldICmpWithZero(ICmpInst &Cmp) {
1319  CmpInst::Predicate Pred = Cmp.getPredicate();
1320  Value *X = Cmp.getOperand(0);
1321 
1322  if (match(Cmp.getOperand(1), m_Zero()) && Pred == ICmpInst::ICMP_SGT) {
1323  Value *A, *B;
1324  SelectPatternResult SPR = matchSelectPattern(X, A, B);
1325  if (SPR.Flavor == SPF_SMIN) {
1326  if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1327  return new ICmpInst(Pred, B, Cmp.getOperand(1));
1328  if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1329  return new ICmpInst(Pred, A, Cmp.getOperand(1));
1330  }
1331  }
1332  return nullptr;
1333 }
1334 
1335 /// Fold icmp Pred X, C.
1336 /// TODO: This code structure does not make sense. The saturating add fold
1337 /// should be moved to some other helper and extended as noted below (it is also
1338 /// possible that code has been made unnecessary - do we canonicalize IR to
1339 /// overflow/saturating intrinsics or not?).
1340 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
1341  // Match the following pattern, which is a common idiom when writing
1342  // overflow-safe integer arithmetic functions. The source performs an addition
1343  // in wider type and explicitly checks for overflow using comparisons against
1344  // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1345  //
1346  // TODO: This could probably be generalized to handle other overflow-safe
1347  // operations if we worked out the formulas to compute the appropriate magic
1348  // constants.
1349  //
1350  // sum = a + b
1351  // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1352  CmpInst::Predicate Pred = Cmp.getPredicate();
1353  Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1354  Value *A, *B;
1355  ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1356  if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1357  match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1358  if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1359  return Res;
1360 
1361  return nullptr;
1362 }
1363 
1364 /// Canonicalize icmp instructions based on dominating conditions.
1365 Instruction *InstCombiner::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1366  // This is a cheap/incomplete check for dominance - just match a single
1367  // predecessor with a conditional branch.
1368  BasicBlock *CmpBB = Cmp.getParent();
1369  BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1370  if (!DomBB)
1371  return nullptr;
1372 
1373  Value *DomCond;
1374  BasicBlock *TrueBB, *FalseBB;
1375  if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1376  return nullptr;
1377 
1378  assert((TrueBB == CmpBB || FalseBB == CmpBB) &&
1379  "Predecessor block does not point to successor?");
1380 
1381  // The branch should get simplified. Don't bother simplifying this condition.
1382  if (TrueBB == FalseBB)
1383  return nullptr;
1384 
1385  // Try to simplify this compare to T/F based on the dominating condition.
1386  Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1387  if (Imp)
1388  return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1389 
1390  CmpInst::Predicate Pred = Cmp.getPredicate();
1391  Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1392  ICmpInst::Predicate DomPred;
1393  const APInt *C, *DomC;
1394  if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1395  match(Y, m_APInt(C))) {
1396  // We have 2 compares of a variable with constants. Calculate the constant
1397  // ranges of those compares to see if we can transform the 2nd compare:
1398  // DomBB:
1399  // DomCond = icmp DomPred X, DomC
1400  // br DomCond, CmpBB, FalseBB
1401  // CmpBB:
1402  // Cmp = icmp Pred X, C
1404  ConstantRange DominatingCR =
1405  (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1407  CmpInst::getInversePredicate(DomPred), *DomC);
1408  ConstantRange Intersection = DominatingCR.intersectWith(CR);
1409  ConstantRange Difference = DominatingCR.difference(CR);
1410  if (Intersection.isEmptySet())
1411  return replaceInstUsesWith(Cmp, Builder.getFalse());
1412  if (Difference.isEmptySet())
1413  return replaceInstUsesWith(Cmp, Builder.getTrue());
1414 
1415  // Canonicalizing a sign bit comparison that gets used in a branch,
1416  // pessimizes codegen by generating branch on zero instruction instead
1417  // of a test and branch. So we avoid canonicalizing in such situations
1418  // because test and branch instruction has better branch displacement
1419  // than compare and branch instruction.
1420  bool UnusedBit;
1421  bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1422  if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1423  return nullptr;
1424 
1425  if (const APInt *EqC = Intersection.getSingleElement())
1426  return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1427  if (const APInt *NeC = Difference.getSingleElement())
1428  return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1429  }
1430 
1431  return nullptr;
1432 }
1433 
1434 /// Fold icmp (trunc X, Y), C.
1435 Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
1436  TruncInst *Trunc,
1437  const APInt &C) {
1438  ICmpInst::Predicate Pred = Cmp.getPredicate();
1439  Value *X = Trunc->getOperand(0);
1440  if (C.isOneValue() && C.getBitWidth() > 1) {
1441  // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1442  Value *V = nullptr;
1443  if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1444  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1445  ConstantInt::get(V->getType(), 1));
1446  }
1447 
1448  if (Cmp.isEquality() && Trunc->hasOneUse()) {
1449  // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1450  // of the high bits truncated out of x are known.
1451  unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1452  SrcBits = X->getType()->getScalarSizeInBits();
1453  KnownBits Known = computeKnownBits(X, 0, &Cmp);
1454 
1455  // If all the high bits are known, we can do this xform.
1456  if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1457  // Pull in the high bits from known-ones set.
1458  APInt NewRHS = C.zext(SrcBits);
1459  NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1460  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1461  }
1462  }
1463 
1464  return nullptr;
1465 }
1466 
1467 /// Fold icmp (xor X, Y), C.
1468 Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp,
1469  BinaryOperator *Xor,
1470  const APInt &C) {
1471  Value *X = Xor->getOperand(0);
1472  Value *Y = Xor->getOperand(1);
1473  const APInt *XorC;
1474  if (!match(Y, m_APInt(XorC)))
1475  return nullptr;
1476 
1477  // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1478  // fold the xor.
1479  ICmpInst::Predicate Pred = Cmp.getPredicate();
1480  bool TrueIfSigned = false;
1481  if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1482 
1483  // If the sign bit of the XorCst is not set, there is no change to
1484  // the operation, just stop using the Xor.
1485  if (!XorC->isNegative()) {
1486  Cmp.setOperand(0, X);
1487  Worklist.Add(Xor);
1488  return &Cmp;
1489  }
1490 
1491  // Emit the opposite comparison.
1492  if (TrueIfSigned)
1493  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1495  else
1496  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1498  }
1499 
1500  if (Xor->hasOneUse()) {
1501  // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1502  if (!Cmp.isEquality() && XorC->isSignMask()) {
1503  Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1504  : Cmp.getSignedPredicate();
1505  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1506  }
1507 
1508  // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1509  if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1510  Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1511  : Cmp.getSignedPredicate();
1512  Pred = Cmp.getSwappedPredicate(Pred);
1513  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1514  }
1515  }
1516 
1517  // Mask constant magic can eliminate an 'xor' with unsigned compares.
1518  if (Pred == ICmpInst::ICMP_UGT) {
1519  // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1520  if (*XorC == ~C && (C + 1).isPowerOf2())
1521  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
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_UGT, X, Y);
1525  }
1526  if (Pred == ICmpInst::ICMP_ULT) {
1527  // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1528  if (*XorC == -C && C.isPowerOf2())
1529  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1530  ConstantInt::get(X->getType(), ~C));
1531  // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1532  if (*XorC == C && (-C).isPowerOf2())
1533  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1534  ConstantInt::get(X->getType(), ~C));
1535  }
1536  return nullptr;
1537 }
1538 
1539 /// Fold icmp (and (sh X, Y), C2), C1.
1540 Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
1541  const APInt &C1, const APInt &C2) {
1542  BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1543  if (!Shift || !Shift->isShift())
1544  return nullptr;
1545 
1546  // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1547  // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1548  // code produced by the clang front-end, for bitfield access.
1549  // This seemingly simple opportunity to fold away a shift turns out to be
1550  // rather complicated. See PR17827 for details.
1551  unsigned ShiftOpcode = Shift->getOpcode();
1552  bool IsShl = ShiftOpcode == Instruction::Shl;
1553  const APInt *C3;
1554  if (match(Shift->getOperand(1), m_APInt(C3))) {
1555  bool CanFold = false;
1556  if (ShiftOpcode == Instruction::Shl) {
1557  // For a left shift, we can fold if the comparison is not signed. We can
1558  // also fold a signed comparison if the mask value and comparison value
1559  // are not negative. These constraints may not be obvious, but we can
1560  // prove that they are correct using an SMT solver.
1561  if (!Cmp.isSigned() || (!C2.isNegative() && !C1.isNegative()))
1562  CanFold = true;
1563  } else {
1564  bool IsAshr = ShiftOpcode == Instruction::AShr;
1565  // For a logical right shift, we can fold if the comparison is not signed.
1566  // We can also fold a signed comparison if the shifted mask value and the
1567  // shifted comparison value are not negative. These constraints may not be
1568  // obvious, but we can prove that they are correct using an SMT solver.
1569  // For an arithmetic shift right we can do the same, if we ensure
1570  // the And doesn't use any bits being shifted in. Normally these would
1571  // be turned into lshr by SimplifyDemandedBits, but not if there is an
1572  // additional user.
1573  if (!IsAshr || (C2.shl(*C3).lshr(*C3) == C2)) {
1574  if (!Cmp.isSigned() ||
1575  (!C2.shl(*C3).isNegative() && !C1.shl(*C3).isNegative()))
1576  CanFold = true;
1577  }
1578  }
1579 
1580  if (CanFold) {
1581  APInt NewCst = IsShl ? C1.lshr(*C3) : C1.shl(*C3);
1582  APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3);
1583  // Check to see if we are shifting out any of the bits being compared.
1584  if (SameAsC1 != C1) {
1585  // If we shifted bits out, the fold is not going to work out. As a
1586  // special case, check to see if this means that the result is always
1587  // true or false now.
1588  if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1589  return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1590  if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1591  return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1592  } else {
1593  Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst));
1594  APInt NewAndCst = IsShl ? C2.lshr(*C3) : C2.shl(*C3);
1595  And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst));
1596  And->setOperand(0, Shift->getOperand(0));
1597  Worklist.Add(Shift); // Shift is dead.
1598  return &Cmp;
1599  }
1600  }
1601  }
1602 
1603  // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1604  // preferable because it allows the C2 << Y expression to be hoisted out of a
1605  // loop if Y is invariant and X is not.
1606  if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1607  !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1608  // Compute C2 << Y.
1609  Value *NewShift =
1610  IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1611  : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1612 
1613  // Compute X & (C2 << Y).
1614  Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1615  Cmp.setOperand(0, NewAnd);
1616  return &Cmp;
1617  }
1618 
1619  return nullptr;
1620 }
1621 
1622 /// Fold icmp (and X, C2), C1.
1623 Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp,
1624  BinaryOperator *And,
1625  const APInt &C1) {
1626  // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1627  // TODO: We canonicalize to the longer form for scalars because we have
1628  // better analysis/folds for icmp, and codegen may be better with icmp.
1629  if (Cmp.getPredicate() == CmpInst::ICMP_NE && Cmp.getType()->isVectorTy() &&
1630  C1.isNullValue() && match(And->getOperand(1), m_One()))
1631  return new TruncInst(And->getOperand(0), Cmp.getType());
1632 
1633  const APInt *C2;
1634  if (!match(And->getOperand(1), m_APInt(C2)))
1635  return nullptr;
1636 
1637  if (!And->hasOneUse())
1638  return nullptr;
1639 
1640  // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1641  // the input width without changing the value produced, eliminate the cast:
1642  //
1643  // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1644  //
1645  // We can do this transformation if the constants do not have their sign bits
1646  // set or if it is an equality comparison. Extending a relational comparison
1647  // when we're checking the sign bit would not work.
1648  Value *W;
1649  if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1650  (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1651  // TODO: Is this a good transform for vectors? Wider types may reduce
1652  // throughput. Should this transform be limited (even for scalars) by using
1653  // shouldChangeType()?
1654  if (!Cmp.getType()->isVectorTy()) {
1655  Type *WideType = W->getType();
1656  unsigned WideScalarBits = WideType->getScalarSizeInBits();
1657  Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1658  Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1659  Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1660  return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1661  }
1662  }
1663 
1664  if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1665  return I;
1666 
1667  // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1668  // (icmp pred (and A, (or (shl 1, B), 1), 0))
1669  //
1670  // iff pred isn't signed
1671  if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1672  match(And->getOperand(1), m_One())) {
1673  Constant *One = cast<Constant>(And->getOperand(1));
1674  Value *Or = And->getOperand(0);
1675  Value *A, *B, *LShr;
1676  if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1677  match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1678  unsigned UsesRemoved = 0;
1679  if (And->hasOneUse())
1680  ++UsesRemoved;
1681  if (Or->hasOneUse())
1682  ++UsesRemoved;
1683  if (LShr->hasOneUse())
1684  ++UsesRemoved;
1685 
1686  // Compute A & ((1 << B) | 1)
1687  Value *NewOr = nullptr;
1688  if (auto *C = dyn_cast<Constant>(B)) {
1689  if (UsesRemoved >= 1)
1690  NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1691  } else {
1692  if (UsesRemoved >= 3)
1693  NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1694  /*HasNUW=*/true),
1695  One, Or->getName());
1696  }
1697  if (NewOr) {
1698  Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1699  Cmp.setOperand(0, NewAnd);
1700  return &Cmp;
1701  }
1702  }
1703  }
1704 
1705  return nullptr;
1706 }
1707 
1708 /// Fold icmp (and X, Y), C.
1709 Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp,
1710  BinaryOperator *And,
1711  const APInt &C) {
1712  if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1713  return I;
1714 
1715  // TODO: These all require that Y is constant too, so refactor with the above.
1716 
1717  // Try to optimize things like "A[i] & 42 == 0" to index computations.
1718  Value *X = And->getOperand(0);
1719  Value *Y = And->getOperand(1);
1720  if (auto *LI = dyn_cast<LoadInst>(X))
1721  if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1722  if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1723  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1724  !LI->isVolatile() && isa<ConstantInt>(Y)) {
1725  ConstantInt *C2 = cast<ConstantInt>(Y);
1726  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1727  return Res;
1728  }
1729 
1730  if (!Cmp.isEquality())
1731  return nullptr;
1732 
1733  // X & -C == -C -> X > u ~C
1734  // X & -C != -C -> X <= u ~C
1735  // iff C is a power of 2
1736  if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1737  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1739  return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1740  }
1741 
1742  // (X & C2) == 0 -> (trunc X) >= 0
1743  // (X & C2) != 0 -> (trunc X) < 0
1744  // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1745  const APInt *C2;
1746  if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1747  int32_t ExactLogBase2 = C2->exactLogBase2();
1748  if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1749  Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1750  if (And->getType()->isVectorTy())
1751  NTy = VectorType::get(NTy, And->getType()->getVectorNumElements());
1752  Value *Trunc = Builder.CreateTrunc(X, NTy);
1753  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1755  return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1756  }
1757  }
1758 
1759  return nullptr;
1760 }
1761 
1762 /// Fold icmp (or X, Y), C.
1763 Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
1764  const APInt &C) {
1765  ICmpInst::Predicate Pred = Cmp.getPredicate();
1766  if (C.isOneValue()) {
1767  // icmp slt signum(V) 1 --> icmp slt V, 1
1768  Value *V = nullptr;
1769  if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1770  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1771  ConstantInt::get(V->getType(), 1));
1772  }
1773 
1774  Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1775  if (Cmp.isEquality() && Cmp.getOperand(1) == OrOp1) {
1776  // X | C == C --> X <=u C
1777  // X | C != C --> X >u C
1778  // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1779  if ((C + 1).isPowerOf2()) {
1781  return new ICmpInst(Pred, OrOp0, OrOp1);
1782  }
1783  // More general: are all bits outside of a mask constant set or not set?
1784  // X | C == C --> (X & ~C) == 0
1785  // X | C != C --> (X & ~C) != 0
1786  if (Or->hasOneUse()) {
1787  Value *A = Builder.CreateAnd(OrOp0, ~C);
1788  return new ICmpInst(Pred, A, ConstantInt::getNullValue(OrOp0->getType()));
1789  }
1790  }
1791 
1792  if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1793  return nullptr;
1794 
1795  Value *P, *Q;
1796  if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1797  // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1798  // -> and (icmp eq P, null), (icmp eq Q, null).
1799  Value *CmpP =
1800  Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1801  Value *CmpQ =
1802  Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1803  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1804  return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1805  }
1806 
1807  // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1808  // a shorter form that has more potential to be folded even further.
1809  Value *X1, *X2, *X3, *X4;
1810  if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1811  match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1812  // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1813  // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1814  Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1815  Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1816  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1817  return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1818  }
1819 
1820  return nullptr;
1821 }
1822 
1823 /// Fold icmp (mul X, Y), C.
1824 Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp,
1825  BinaryOperator *Mul,
1826  const APInt &C) {
1827  const APInt *MulC;
1828  if (!match(Mul->getOperand(1), m_APInt(MulC)))
1829  return nullptr;
1830 
1831  // If this is a test of the sign bit and the multiply is sign-preserving with
1832  // a constant operand, use the multiply LHS operand instead.
1833  ICmpInst::Predicate Pred = Cmp.getPredicate();
1834  if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1835  if (MulC->isNegative())
1836  Pred = ICmpInst::getSwappedPredicate(Pred);
1837  return new ICmpInst(Pred, Mul->getOperand(0),
1839  }
1840 
1841  return nullptr;
1842 }
1843 
1844 /// Fold icmp (shl 1, Y), C.
1846  const APInt &C) {
1847  Value *Y;
1848  if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1849  return nullptr;
1850 
1851  Type *ShiftType = Shl->getType();
1852  unsigned TypeBits = C.getBitWidth();
1853  bool CIsPowerOf2 = C.isPowerOf2();
1854  ICmpInst::Predicate Pred = Cmp.getPredicate();
1855  if (Cmp.isUnsigned()) {
1856  // (1 << Y) pred C -> Y pred Log2(C)
1857  if (!CIsPowerOf2) {
1858  // (1 << Y) < 30 -> Y <= 4
1859  // (1 << Y) <= 30 -> Y <= 4
1860  // (1 << Y) >= 30 -> Y > 4
1861  // (1 << Y) > 30 -> Y > 4
1862  if (Pred == ICmpInst::ICMP_ULT)
1863  Pred = ICmpInst::ICMP_ULE;
1864  else if (Pred == ICmpInst::ICMP_UGE)
1865  Pred = ICmpInst::ICMP_UGT;
1866  }
1867 
1868  // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
1869  // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
1870  unsigned CLog2 = C.logBase2();
1871  if (CLog2 == TypeBits - 1) {
1872  if (Pred == ICmpInst::ICMP_UGE)
1873  Pred = ICmpInst::ICMP_EQ;
1874  else if (Pred == ICmpInst::ICMP_ULT)
1875  Pred = ICmpInst::ICMP_NE;
1876  }
1877  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
1878  } else if (Cmp.isSigned()) {
1879  Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
1880  if (C.isAllOnesValue()) {
1881  // (1 << Y) <= -1 -> Y == 31
1882  if (Pred == ICmpInst::ICMP_SLE)
1883  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1884 
1885  // (1 << Y) > -1 -> Y != 31
1886  if (Pred == ICmpInst::ICMP_SGT)
1887  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1888  } else if (!C) {
1889  // (1 << Y) < 0 -> Y == 31
1890  // (1 << Y) <= 0 -> Y == 31
1891  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1892  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1893 
1894  // (1 << Y) >= 0 -> Y != 31
1895  // (1 << Y) > 0 -> Y != 31
1896  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1897  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1898  }
1899  } else if (Cmp.isEquality() && CIsPowerOf2) {
1900  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
1901  }
1902 
1903  return nullptr;
1904 }
1905 
1906 /// Fold icmp (shl X, Y), C.
1907 Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp,
1908  BinaryOperator *Shl,
1909  const APInt &C) {
1910  const APInt *ShiftVal;
1911  if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
1912  return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
1913 
1914  const APInt *ShiftAmt;
1915  if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
1916  return foldICmpShlOne(Cmp, Shl, C);
1917 
1918  // Check that the shift amount is in range. If not, don't perform undefined
1919  // shifts. When the shift is visited, it will be simplified.
1920  unsigned TypeBits = C.getBitWidth();
1921  if (ShiftAmt->uge(TypeBits))
1922  return nullptr;
1923 
1924  ICmpInst::Predicate Pred = Cmp.getPredicate();
1925  Value *X = Shl->getOperand(0);
1926  Type *ShType = Shl->getType();
1927 
1928  // NSW guarantees that we are only shifting out sign bits from the high bits,
1929  // so we can ASHR the compare constant without needing a mask and eliminate
1930  // the shift.
1931  if (Shl->hasNoSignedWrap()) {
1932  if (Pred == ICmpInst::ICMP_SGT) {
1933  // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
1934  APInt ShiftedC = C.ashr(*ShiftAmt);
1935  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1936  }
1937  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
1938  C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
1939  APInt ShiftedC = C.ashr(*ShiftAmt);
1940  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1941  }
1942  if (Pred == ICmpInst::ICMP_SLT) {
1943  // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
1944  // (X << S) <=s C is equiv to X <=s (C >> S) for all C
1945  // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
1946  // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
1947  assert(!C.isMinSignedValue() && "Unexpected icmp slt");
1948  APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
1949  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1950  }
1951  // If this is a signed comparison to 0 and the shift is sign preserving,
1952  // use the shift LHS operand instead; isSignTest may change 'Pred', so only
1953  // do that if we're sure to not continue on in this function.
1954  if (isSignTest(Pred, C))
1955  return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
1956  }
1957 
1958  // NUW guarantees that we are only shifting out zero bits from the high bits,
1959  // so we can LSHR the compare constant without needing a mask and eliminate
1960  // the shift.
1961  if (Shl->hasNoUnsignedWrap()) {
1962  if (Pred == ICmpInst::ICMP_UGT) {
1963  // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
1964  APInt ShiftedC = C.lshr(*ShiftAmt);
1965  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1966  }
1967  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
1968  C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
1969  APInt ShiftedC = C.lshr(*ShiftAmt);
1970  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1971  }
1972  if (Pred == ICmpInst::ICMP_ULT) {
1973  // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
1974  // (X << S) <=u C is equiv to X <=u (C >> S) for all C
1975  // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
1976  // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
1977  assert(C.ugt(0) && "ult 0 should have been eliminated");
1978  APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
1979  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1980  }
1981  }
1982 
1983  if (Cmp.isEquality() && Shl->hasOneUse()) {
1984  // Strength-reduce the shift into an 'and'.
1986  ShType,
1987  APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
1988  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
1989  Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
1990  return new ICmpInst(Pred, And, LShrC);
1991  }
1992 
1993  // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1994  bool TrueIfSigned = false;
1995  if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
1996  // (X << 31) <s 0 --> (X & 1) != 0
1998  ShType,
1999  APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2000  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2001  return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2002  And, Constant::getNullValue(ShType));
2003  }
2004 
2005  // Transform (icmp pred iM (shl iM %v, N), C)
2006  // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2007  // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2008  // This enables us to get rid of the shift in favor of a trunc that may be
2009  // free on the target. It has the additional benefit of comparing to a
2010  // smaller constant that may be more target-friendly.
2011  unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2012  if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2013  DL.isLegalInteger(TypeBits - Amt)) {
2014  Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2015  if (ShType->isVectorTy())
2016  TruncTy = VectorType::get(TruncTy, ShType->getVectorNumElements());
2017  Constant *NewC =
2018  ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2019  return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2020  }
2021 
2022  return nullptr;
2023 }
2024 
2025 /// Fold icmp ({al}shr X, Y), C.
2026 Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp,
2027  BinaryOperator *Shr,
2028  const APInt &C) {
2029  // An exact shr only shifts out zero bits, so:
2030  // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2031  Value *X = Shr->getOperand(0);
2032  CmpInst::Predicate Pred = Cmp.getPredicate();
2033  if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2034  C.isNullValue())
2035  return new ICmpInst(Pred, X, Cmp.getOperand(1));
2036 
2037  const APInt *ShiftVal;
2038  if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2039  return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2040 
2041  const APInt *ShiftAmt;
2042  if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2043  return nullptr;
2044 
2045  // Check that the shift amount is in range. If not, don't perform undefined
2046  // shifts. When the shift is visited it will be simplified.
2047  unsigned TypeBits = C.getBitWidth();
2048  unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2049  if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2050  return nullptr;
2051 
2052  bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2053  bool IsExact = Shr->isExact();
2054  Type *ShrTy = Shr->getType();
2055  // TODO: If we could guarantee that InstSimplify would handle all of the
2056  // constant-value-based preconditions in the folds below, then we could assert
2057  // those conditions rather than checking them. This is difficult because of
2058  // undef/poison (PR34838).
2059  if (IsAShr) {
2060  if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2061  // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2062  // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2063  APInt ShiftedC = C.shl(ShAmtVal);
2064  if (ShiftedC.ashr(ShAmtVal) == C)
2065  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2066  }
2067  if (Pred == CmpInst::ICMP_SGT) {
2068  // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2069  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2070  if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2071  (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2072  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2073  }
2074  } else {
2075  if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2076  // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2077  // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2078  APInt ShiftedC = C.shl(ShAmtVal);
2079  if (ShiftedC.lshr(ShAmtVal) == C)
2080  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2081  }
2082  if (Pred == CmpInst::ICMP_UGT) {
2083  // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2084  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2085  if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2086  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2087  }
2088  }
2089 
2090  if (!Cmp.isEquality())
2091  return nullptr;
2092 
2093  // Handle equality comparisons of shift-by-constant.
2094 
2095  // If the comparison constant changes with the shift, the comparison cannot
2096  // succeed (bits of the comparison constant cannot match the shifted value).
2097  // This should be known by InstSimplify and already be folded to true/false.
2098  assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2099  (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2100  "Expected icmp+shr simplify did not occur.");
2101 
2102  // If the bits shifted out are known zero, compare the unshifted value:
2103  // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2104  if (Shr->isExact())
2105  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2106 
2107  if (Shr->hasOneUse()) {
2108  // Canonicalize the shift into an 'and':
2109  // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2110  APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2111  Constant *Mask = ConstantInt::get(ShrTy, Val);
2112  Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2113  return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2114  }
2115 
2116  return nullptr;
2117 }
2118 
2119 /// Fold icmp (udiv X, Y), C.
2120 Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp,
2121  BinaryOperator *UDiv,
2122  const APInt &C) {
2123  const APInt *C2;
2124  if (!match(UDiv->getOperand(0), m_APInt(C2)))
2125  return nullptr;
2126 
2127  assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2128 
2129  // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2130  Value *Y = UDiv->getOperand(1);
2131  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2132  assert(!C.isMaxValue() &&
2133  "icmp ugt X, UINT_MAX should have been simplified already.");
2134  return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2135  ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2136  }
2137 
2138  // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2139  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2140  assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2141  return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2142  ConstantInt::get(Y->getType(), C2->udiv(C)));
2143  }
2144 
2145  return nullptr;
2146 }
2147 
2148 /// Fold icmp ({su}div X, Y), C.
2149 Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp,
2150  BinaryOperator *Div,
2151  const APInt &C) {
2152  // Fold: icmp pred ([us]div X, C2), C -> range test
2153  // Fold this div into the comparison, producing a range check.
2154  // Determine, based on the divide type, what the range is being
2155  // checked. If there is an overflow on the low or high side, remember
2156  // it, otherwise compute the range [low, hi) bounding the new value.
2157  // See: InsertRangeTest above for the kinds of replacements possible.
2158  const APInt *C2;
2159  if (!match(Div->getOperand(1), m_APInt(C2)))
2160  return nullptr;
2161 
2162  // FIXME: If the operand types don't match the type of the divide
2163  // then don't attempt this transform. The code below doesn't have the
2164  // logic to deal with a signed divide and an unsigned compare (and
2165  // vice versa). This is because (x /s C2) <s C produces different
2166  // results than (x /s C2) <u C or (x /u C2) <s C or even
2167  // (x /u C2) <u C. Simply casting the operands and result won't
2168  // work. :( The if statement below tests that condition and bails
2169  // if it finds it.
2170  bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2171  if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2172  return nullptr;
2173 
2174  // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2175  // INT_MIN will also fail if the divisor is 1. Although folds of all these
2176  // division-by-constant cases should be present, we can not assert that they
2177  // have happened before we reach this icmp instruction.
2178  if (C2->isNullValue() || C2->isOneValue() ||
2179  (DivIsSigned && C2->isAllOnesValue()))
2180  return nullptr;
2181 
2182  // Compute Prod = C * C2. We are essentially solving an equation of
2183  // form X / C2 = C. We solve for X by multiplying C2 and C.
2184  // By solving for X, we can turn this into a range check instead of computing
2185  // a divide.
2186  APInt Prod = C * *C2;
2187 
2188  // Determine if the product overflows by seeing if the product is not equal to
2189  // the divide. Make sure we do the same kind of divide as in the LHS
2190  // instruction that we're folding.
2191  bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2192 
2193  ICmpInst::Predicate Pred = Cmp.getPredicate();
2194 
2195  // If the division is known to be exact, then there is no remainder from the
2196  // divide, so the covered range size is unit, otherwise it is the divisor.
2197  APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2198 
2199  // Figure out the interval that is being checked. For example, a comparison
2200  // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2201  // Compute this interval based on the constants involved and the signedness of
2202  // the compare/divide. This computes a half-open interval, keeping track of
2203  // whether either value in the interval overflows. After analysis each
2204  // overflow variable is set to 0 if it's corresponding bound variable is valid
2205  // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2206  int LoOverflow = 0, HiOverflow = 0;
2207  APInt LoBound, HiBound;
2208 
2209  if (!DivIsSigned) { // udiv
2210  // e.g. X/5 op 3 --> [15, 20)
2211  LoBound = Prod;
2212  HiOverflow = LoOverflow = ProdOV;
2213  if (!HiOverflow) {
2214  // If this is not an exact divide, then many values in the range collapse
2215  // to the same result value.
2216  HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2217  }
2218  } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2219  if (C.isNullValue()) { // (X / pos) op 0
2220  // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2221  LoBound = -(RangeSize - 1);
2222  HiBound = RangeSize;
2223  } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2224  LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2225  HiOverflow = LoOverflow = ProdOV;
2226  if (!HiOverflow)
2227  HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2228  } else { // (X / pos) op neg
2229  // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2230  HiBound = Prod + 1;
2231  LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2232  if (!LoOverflow) {
2233  APInt DivNeg = -RangeSize;
2234  LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2235  }
2236  }
2237  } else if (C2->isNegative()) { // Divisor is < 0.
2238  if (Div->isExact())
2239  RangeSize.negate();
2240  if (C.isNullValue()) { // (X / neg) op 0
2241  // e.g. X/-5 op 0 --> [-4, 5)
2242  LoBound = RangeSize + 1;
2243  HiBound = -RangeSize;
2244  if (HiBound == *C2) { // -INTMIN = INTMIN
2245  HiOverflow = 1; // [INTMIN+1, overflow)
2246  HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2247  }
2248  } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2249  // e.g. X/-5 op 3 --> [-19, -14)
2250  HiBound = Prod + 1;
2251  HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2252  if (!LoOverflow)
2253  LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2254  } else { // (X / neg) op neg
2255  LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2256  LoOverflow = HiOverflow = ProdOV;
2257  if (!HiOverflow)
2258  HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2259  }
2260 
2261  // Dividing by a negative swaps the condition. LT <-> GT
2262  Pred = ICmpInst::getSwappedPredicate(Pred);
2263  }
2264 
2265  Value *X = Div->getOperand(0);
2266  switch (Pred) {
2267  default: llvm_unreachable("Unhandled icmp opcode!");
2268  case ICmpInst::ICMP_EQ:
2269  if (LoOverflow && HiOverflow)
2270  return replaceInstUsesWith(Cmp, Builder.getFalse());
2271  if (HiOverflow)
2272  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2273  ICmpInst::ICMP_UGE, X,
2274  ConstantInt::get(Div->getType(), LoBound));
2275  if (LoOverflow)
2276  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2277  ICmpInst::ICMP_ULT, X,
2278  ConstantInt::get(Div->getType(), HiBound));
2279  return replaceInstUsesWith(
2280  Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2281  case ICmpInst::ICMP_NE:
2282  if (LoOverflow && HiOverflow)
2283  return replaceInstUsesWith(Cmp, Builder.getTrue());
2284  if (HiOverflow)
2285  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2286  ICmpInst::ICMP_ULT, X,
2287  ConstantInt::get(Div->getType(), LoBound));
2288  if (LoOverflow)
2289  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2290  ICmpInst::ICMP_UGE, X,
2291  ConstantInt::get(Div->getType(), HiBound));
2292  return replaceInstUsesWith(Cmp,
2293  insertRangeTest(X, LoBound, HiBound,
2294  DivIsSigned, false));
2295  case ICmpInst::ICMP_ULT:
2296  case ICmpInst::ICMP_SLT:
2297  if (LoOverflow == +1) // Low bound is greater than input range.
2298  return replaceInstUsesWith(Cmp, Builder.getTrue());
2299  if (LoOverflow == -1) // Low bound is less than input range.
2300  return replaceInstUsesWith(Cmp, Builder.getFalse());
2301  return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2302  case ICmpInst::ICMP_UGT:
2303  case ICmpInst::ICMP_SGT:
2304  if (HiOverflow == +1) // High bound greater than input range.
2305  return replaceInstUsesWith(Cmp, Builder.getFalse());
2306  if (HiOverflow == -1) // High bound less than input range.
2307  return replaceInstUsesWith(Cmp, Builder.getTrue());
2308  if (Pred == ICmpInst::ICMP_UGT)
2309  return new ICmpInst(ICmpInst::ICMP_UGE, X,
2310  ConstantInt::get(Div->getType(), HiBound));
2311  return new ICmpInst(ICmpInst::ICMP_SGE, X,
2312  ConstantInt::get(Div->getType(), HiBound));
2313  }
2314 
2315  return nullptr;
2316 }
2317 
2318 /// Fold icmp (sub X, Y), C.
2319 Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp,
2320  BinaryOperator *Sub,
2321  const APInt &C) {
2322  Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2323  ICmpInst::Predicate Pred = Cmp.getPredicate();
2324 
2325  // The following transforms are only worth it if the only user of the subtract
2326  // is the icmp.
2327  if (!Sub->hasOneUse())
2328  return nullptr;
2329 
2330  if (Sub->hasNoSignedWrap()) {
2331  // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2332  if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2333  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2334 
2335  // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2336  if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2337  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2338 
2339  // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2340  if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2341  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2342 
2343  // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2344  if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2345  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2346  }
2347 
2348  const APInt *C2;
2349  if (!match(X, m_APInt(C2)))
2350  return nullptr;
2351 
2352  // C2 - Y <u C -> (Y | (C - 1)) == C2
2353  // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2354  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2355  (*C2 & (C - 1)) == (C - 1))
2356  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2357 
2358  // C2 - Y >u C -> (Y | C) != C2
2359  // iff C2 & C == C and C + 1 is a power of 2
2360  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2361  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2362 
2363  return nullptr;
2364 }
2365 
2366 /// Fold icmp (add X, Y), C.
2367 Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp,
2369  const APInt &C) {
2370  Value *Y = Add->getOperand(1);
2371  const APInt *C2;
2372  if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2373  return nullptr;
2374 
2375  // Fold icmp pred (add X, C2), C.
2376  Value *X = Add->getOperand(0);
2377  Type *Ty = Add->getType();
2378  CmpInst::Predicate Pred = Cmp.getPredicate();
2379 
2380  if (!Add->hasOneUse())
2381  return nullptr;
2382 
2383  // If the add does not wrap, we can always adjust the compare by subtracting
2384  // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2385  // are canonicalized to SGT/SLT/UGT/ULT.
2386  if ((Add->hasNoSignedWrap() &&
2387  (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2388  (Add->hasNoUnsignedWrap() &&
2389  (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2390  bool Overflow;
2391  APInt NewC =
2392  Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2393  // If there is overflow, the result must be true or false.
2394  // TODO: Can we assert there is no overflow because InstSimplify always
2395  // handles those cases?
2396  if (!Overflow)
2397  // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2398  return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2399  }
2400 
2401  auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2402  const APInt &Upper = CR.getUpper();
2403  const APInt &Lower = CR.getLower();
2404  if (Cmp.isSigned()) {
2405  if (Lower.isSignMask())
2406  return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2407  if (Upper.isSignMask())
2408  return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2409  } else {
2410  if (Lower.isMinValue())
2411  return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2412  if (Upper.isMinValue())
2413  return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2414  }
2415 
2416  // X+C <u C2 -> (X & -C2) == C
2417  // iff C & (C2-1) == 0
2418  // C2 is a power of 2
2419  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2420  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2421  ConstantExpr::getNeg(cast<Constant>(Y)));
2422 
2423  // X+C >u C2 -> (X & ~C2) != C
2424  // iff C & C2 == 0
2425  // C2+1 is a power of 2
2426  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2427  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2428  ConstantExpr::getNeg(cast<Constant>(Y)));
2429 
2430  return nullptr;
2431 }
2432 
2433 bool InstCombiner::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2434  Value *&RHS, ConstantInt *&Less,
2435  ConstantInt *&Equal,
2436  ConstantInt *&Greater) {
2437  // TODO: Generalize this to work with other comparison idioms or ensure
2438  // they get canonicalized into this form.
2439 
2440  // select i1 (a == b), i32 Equal, i32 (select i1 (a < b), i32 Less, i32
2441  // Greater), where Equal, Less and Greater are placeholders for any three
2442  // constants.
2443  ICmpInst::Predicate PredA, PredB;
2444  if (match(SI->getTrueValue(), m_ConstantInt(Equal)) &&
2445  match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) &&
2446  PredA == ICmpInst::ICMP_EQ &&
2447  match(SI->getFalseValue(),
2448  m_Select(m_ICmp(PredB, m_Specific(LHS), m_Specific(RHS)),
2449  m_ConstantInt(Less), m_ConstantInt(Greater))) &&
2450  PredB == ICmpInst::ICMP_SLT) {
2451  return true;
2452  }
2453  return false;
2454 }
2455 
2456 Instruction *InstCombiner::foldICmpSelectConstant(ICmpInst &Cmp,
2457  SelectInst *Select,
2458  ConstantInt *C) {
2459 
2460  assert(C && "Cmp RHS should be a constant int!");
2461  // If we're testing a constant value against the result of a three way
2462  // comparison, the result can be expressed directly in terms of the
2463  // original values being compared. Note: We could possibly be more
2464  // aggressive here and remove the hasOneUse test. The original select is
2465  // really likely to simplify or sink when we remove a test of the result.
2466  Value *OrigLHS, *OrigRHS;
2467  ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2468  if (Cmp.hasOneUse() &&
2469  matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2470  C3GreaterThan)) {
2471  assert(C1LessThan && C2Equal && C3GreaterThan);
2472 
2473  bool TrueWhenLessThan =
2474  ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2475  ->isAllOnesValue();
2476  bool TrueWhenEqual =
2477  ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2478  ->isAllOnesValue();
2479  bool TrueWhenGreaterThan =
2480  ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2481  ->isAllOnesValue();
2482 
2483  // This generates the new instruction that will replace the original Cmp
2484  // Instruction. Instead of enumerating the various combinations when
2485  // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2486  // false, we rely on chaining of ORs and future passes of InstCombine to
2487  // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2488 
2489  // When none of the three constants satisfy the predicate for the RHS (C),
2490  // the entire original Cmp can be simplified to a false.
2491  Value *Cond = Builder.getFalse();
2492  if (TrueWhenLessThan)
2493  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2494  OrigLHS, OrigRHS));
2495  if (TrueWhenEqual)
2496  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2497  OrigLHS, OrigRHS));
2498  if (TrueWhenGreaterThan)
2499  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2500  OrigLHS, OrigRHS));
2501 
2502  return replaceInstUsesWith(Cmp, Cond);
2503  }
2504  return nullptr;
2505 }
2506 
2508  InstCombiner::BuilderTy &Builder) {
2509  auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2510  if (!Bitcast)
2511  return nullptr;
2512 
2513  ICmpInst::Predicate Pred = Cmp.getPredicate();
2514  Value *Op1 = Cmp.getOperand(1);
2515  Value *BCSrcOp = Bitcast->getOperand(0);
2516 
2517  // Make sure the bitcast doesn't change the number of vector elements.
2518  if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2519  Bitcast->getDestTy()->getScalarSizeInBits()) {
2520  // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2521  Value *X;
2522  if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2523  // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2524  // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2525  // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2526  // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2527  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2528  Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2529  match(Op1, m_Zero()))
2530  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2531 
2532  // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2533  if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2534  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2535 
2536  // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2537  if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2538  return new ICmpInst(Pred, X,
2540  }
2541 
2542  // Zero-equality checks are preserved through unsigned floating-point casts:
2543  // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2544  // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2545  if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2546  if (Cmp.isEquality() && match(Op1, m_Zero()))
2547  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2548  }
2549 
2550  // Test to see if the operands of the icmp are casted versions of other
2551  // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2552  if (Bitcast->getType()->isPointerTy() &&
2553  (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2554  // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2555  // so eliminate it as well.
2556  if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2557  Op1 = BC2->getOperand(0);
2558 
2559  Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2560  return new ICmpInst(Pred, BCSrcOp, Op1);
2561  }
2562 
2563  // Folding: icmp <pred> iN X, C
2564  // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2565  // and C is a splat of a K-bit pattern
2566  // and SC is a constant vector = <C', C', C', ..., C'>
2567  // Into:
2568  // %E = extractelement <M x iK> %vec, i32 C'
2569  // icmp <pred> iK %E, trunc(C)
2570  const APInt *C;
2571  if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2572  !Bitcast->getType()->isIntegerTy() ||
2573  !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2574  return nullptr;
2575 
2576  Value *Vec;
2577  Constant *Mask;
2578  if (match(BCSrcOp,
2579  m_ShuffleVector(m_Value(Vec), m_Undef(), m_Constant(Mask)))) {
2580  // Check whether every element of Mask is the same constant
2581  if (auto *Elem = dyn_cast_or_null<ConstantInt>(Mask->getSplatValue())) {
2582  auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2583  auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2584  if (C->isSplat(EltTy->getBitWidth())) {
2585  // Fold the icmp based on the value of C
2586  // If C is M copies of an iK sized bit pattern,
2587  // then:
2588  // => %E = extractelement <N x iK> %vec, i32 Elem
2589  // icmp <pred> iK %SplatVal, <pattern>
2590  Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2591  Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2592  return new ICmpInst(Pred, Extract, NewC);
2593  }
2594  }
2595  }
2596  return nullptr;
2597 }
2598 
2599 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2600 /// where X is some kind of instruction.
2601 Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) {
2602  const APInt *C;
2603  if (!match(Cmp.getOperand(1), m_APInt(C)))
2604  return nullptr;
2605 
2606  if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2607  switch (BO->getOpcode()) {
2608  case Instruction::Xor:
2609  if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2610  return I;
2611  break;
2612  case Instruction::And:
2613  if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2614  return I;
2615  break;
2616  case Instruction::Or:
2617  if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2618  return I;
2619  break;
2620  case Instruction::Mul:
2621  if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2622  return I;
2623  break;
2624  case Instruction::Shl:
2625  if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2626  return I;
2627  break;
2628  case Instruction::LShr:
2629  case Instruction::AShr:
2630  if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2631  return I;
2632  break;
2633  case Instruction::UDiv:
2634  if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2635  return I;
2637  case Instruction::SDiv:
2638  if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2639  return I;
2640  break;
2641  case Instruction::Sub:
2642  if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2643  return I;
2644  break;
2645  case Instruction::Add:
2646  if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2647  return I;
2648  break;
2649  default:
2650  break;
2651  }
2652  // TODO: These folds could be refactored to be part of the above calls.
2653  if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2654  return I;
2655  }
2656 
2657  // Match against CmpInst LHS being instructions other than binary operators.
2658 
2659  if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2660  // For now, we only support constant integers while folding the
2661  // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2662  // similar to the cases handled by binary ops above.
2663  if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2664  if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2665  return I;
2666  }
2667 
2668  if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2669  if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2670  return I;
2671  }
2672 
2673  if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
2674  if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
2675  return I;
2676 
2677  return nullptr;
2678 }
2679 
2680 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2681 /// icmp eq/ne BO, C.
2682 Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
2683  BinaryOperator *BO,
2684  const APInt &C) {
2685  // TODO: Some of these folds could work with arbitrary constants, but this
2686  // function is limited to scalar and vector splat constants.
2687  if (!Cmp.isEquality())
2688  return nullptr;
2689 
2690  ICmpInst::Predicate Pred = Cmp.getPredicate();
2691  bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2692  Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2693  Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2694 
2695  switch (BO->getOpcode()) {
2696  case Instruction::SRem:
2697  // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2698  if (C.isNullValue() && BO->hasOneUse()) {
2699  const APInt *BOC;
2700  if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2701  Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
2702  return new ICmpInst(Pred, NewRem,
2704  }
2705  }
2706  break;
2707  case Instruction::Add: {
2708  // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
2709  const APInt *BOC;
2710  if (match(BOp1, m_APInt(BOC))) {
2711  if (BO->hasOneUse()) {
2712  Constant *SubC = ConstantExpr::getSub(RHS, cast<Constant>(BOp1));
2713  return new ICmpInst(Pred, BOp0, SubC);
2714  }
2715  } else if (C.isNullValue()) {
2716  // Replace ((add A, B) != 0) with (A != -B) if A or B is
2717  // efficiently invertible, or if the add has just this one use.
2718  if (Value *NegVal = dyn_castNegVal(BOp1))
2719  return new ICmpInst(Pred, BOp0, NegVal);
2720  if (Value *NegVal = dyn_castNegVal(BOp0))
2721  return new ICmpInst(Pred, NegVal, BOp1);
2722  if (BO->hasOneUse()) {
2723  Value *Neg = Builder.CreateNeg(BOp1);
2724  Neg->takeName(BO);
2725  return new ICmpInst(Pred, BOp0, Neg);
2726  }
2727  }
2728  break;
2729  }
2730  case Instruction::Xor:
2731  if (BO->hasOneUse()) {
2732  if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2733  // For the xor case, we can xor two constants together, eliminating
2734  // the explicit xor.
2735  return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
2736  } else if (C.isNullValue()) {
2737  // Replace ((xor A, B) != 0) with (A != B)
2738  return new ICmpInst(Pred, BOp0, BOp1);
2739  }
2740  }
2741  break;
2742  case Instruction::Sub:
2743  if (BO->hasOneUse()) {
2744  const APInt *BOC;
2745  if (match(BOp0, m_APInt(BOC))) {
2746  // Replace ((sub BOC, B) != C) with (B != BOC-C).
2747  Constant *SubC = ConstantExpr::getSub(cast<Constant>(BOp0), RHS);
2748  return new ICmpInst(Pred, BOp1, SubC);
2749  } else if (C.isNullValue()) {
2750  // Replace ((sub A, B) != 0) with (A != B).
2751  return new ICmpInst(Pred, BOp0, BOp1);
2752  }
2753  }
2754  break;
2755  case Instruction::Or: {
2756  const APInt *BOC;
2757  if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
2758  // Comparing if all bits outside of a constant mask are set?
2759  // Replace (X | C) == -1 with (X & ~C) == ~C.
2760  // This removes the -1 constant.
2761  Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
2762  Value *And = Builder.CreateAnd(BOp0, NotBOC);
2763  return new ICmpInst(Pred, And, NotBOC);
2764  }
2765  break;
2766  }
2767  case Instruction::And: {
2768  const APInt *BOC;
2769  if (match(BOp1, m_APInt(BOC))) {
2770  // If we have ((X & C) == C), turn it into ((X & C) != 0).
2771  if (C == *BOC && C.isPowerOf2())
2772  return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
2773  BO, Constant::getNullValue(RHS->getType()));
2774 
2775  // Don't perform the following transforms if the AND has multiple uses
2776  if (!BO->hasOneUse())
2777  break;
2778 
2779  // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
2780  if (BOC->isSignMask()) {
2781  Constant *Zero = Constant::getNullValue(BOp0->getType());
2782  auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
2783  return new ICmpInst(NewPred, BOp0, Zero);
2784  }
2785 
2786  // ((X & ~7) == 0) --> X < 8
2787  if (C.isNullValue() && (~(*BOC) + 1).isPowerOf2()) {
2788  Constant *NegBOC = ConstantExpr::getNeg(cast<Constant>(BOp1));
2789  auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
2790  return new ICmpInst(NewPred, BOp0, NegBOC);
2791  }
2792  }
2793  break;
2794  }
2795  case Instruction::Mul:
2796  if (C.isNullValue() && BO->hasNoSignedWrap()) {
2797  const APInt *BOC;
2798  if (match(BOp1, m_APInt(BOC)) && !BOC->isNullValue()) {
2799  // The trivial case (mul X, 0) is handled by InstSimplify.
2800  // General case : (mul X, C) != 0 iff X != 0
2801  // (mul X, C) == 0 iff X == 0
2802  return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType()));
2803  }
2804  }
2805  break;
2806  case Instruction::UDiv:
2807  if (C.isNullValue()) {
2808  // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
2809  auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
2810  return new ICmpInst(NewPred, BOp1, BOp0);
2811  }
2812  break;
2813  default:
2814  break;
2815  }
2816  return nullptr;
2817 }
2818 
2819 /// Fold an equality icmp with LLVM intrinsic and constant operand.
2820 Instruction *InstCombiner::foldICmpEqIntrinsicWithConstant(ICmpInst &Cmp,
2821  IntrinsicInst *II,
2822  const APInt &C) {
2823  Type *Ty = II->getType();
2824  unsigned BitWidth = C.getBitWidth();
2825  switch (II->getIntrinsicID()) {
2826  case Intrinsic::bswap:
2827  Worklist.Add(II);
2828  Cmp.setOperand(0, II->getArgOperand(0));
2829  Cmp.setOperand(1, ConstantInt::get(Ty, C.byteSwap()));
2830  return &Cmp;
2831 
2832  case Intrinsic::ctlz:
2833  case Intrinsic::cttz: {
2834  // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
2835  if (C == BitWidth) {
2836  Worklist.Add(II);
2837  Cmp.setOperand(0, II->getArgOperand(0));
2839  return &Cmp;
2840  }
2841 
2842  // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
2843  // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
2844  // Limit to one use to ensure we don't increase instruction count.
2845  unsigned Num = C.getLimitedValue(BitWidth);
2846  if (Num != BitWidth && II->hasOneUse()) {
2847  bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
2848  APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
2849  : APInt::getHighBitsSet(BitWidth, Num + 1);
2850  APInt Mask2 = IsTrailing
2851  ? APInt::getOneBitSet(BitWidth, Num)
2852  : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
2853  Cmp.setOperand(0, Builder.CreateAnd(II->getArgOperand(0), Mask1));
2854  Cmp.setOperand(1, ConstantInt::get(Ty, Mask2));
2855  Worklist.Add(II);
2856  return &Cmp;
2857  }
2858  break;
2859  }
2860 
2861  case Intrinsic::ctpop: {
2862  // popcount(A) == 0 -> A == 0 and likewise for !=
2863  // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
2864  bool IsZero = C.isNullValue();
2865  if (IsZero || C == BitWidth) {
2866  Worklist.Add(II);
2867  Cmp.setOperand(0, II->getArgOperand(0));
2868  auto *NewOp =
2870  Cmp.setOperand(1, NewOp);
2871  return &Cmp;
2872  }
2873  break;
2874  }
2875  default:
2876  break;
2877  }
2878 
2879  return nullptr;
2880 }
2881 
2882 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
2883 Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
2884  IntrinsicInst *II,
2885  const APInt &C) {
2886  if (Cmp.isEquality())
2887  return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
2888 
2889  Type *Ty = II->getType();
2890  unsigned BitWidth = C.getBitWidth();
2891  switch (II->getIntrinsicID()) {
2892  case Intrinsic::ctlz: {
2893  // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
2894  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
2895  unsigned Num = C.getLimitedValue();
2896  APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
2897  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
2898  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
2899  }
2900 
2901  // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
2902  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
2903  C.uge(1) && C.ule(BitWidth)) {
2904  unsigned Num = C.getLimitedValue();
2905  APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
2906  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
2907  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
2908  }
2909  break;
2910  }
2911  case Intrinsic::cttz: {
2912  // Limit to one use to ensure we don't increase instruction count.
2913  if (!II->hasOneUse())
2914  return nullptr;
2915 
2916  // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
2917  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
2918  APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
2919  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
2920  Builder.CreateAnd(II->getArgOperand(0), Mask),
2922  }
2923 
2924  // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
2925  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
2926  C.uge(1) && C.ule(BitWidth)) {
2928  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
2929  Builder.CreateAnd(II->getArgOperand(0), Mask),
2931  }
2932  break;
2933  }
2934  default:
2935  break;
2936  }
2937 
2938  return nullptr;
2939 }
2940 
2941 /// Handle icmp with constant (but not simple integer constant) RHS.
2942 Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) {
2943  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2944  Constant *RHSC = dyn_cast<Constant>(Op1);
2945  Instruction *LHSI = dyn_cast<Instruction>(Op0);
2946  if (!RHSC || !LHSI)
2947  return nullptr;
2948 
2949  switch (LHSI->getOpcode()) {
2950  case Instruction::GetElementPtr:
2951  // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2952  if (RHSC->isNullValue() &&
2953  cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2954  return new ICmpInst(
2955  I.getPredicate(), LHSI->getOperand(0),
2956  Constant::getNullValue(LHSI->getOperand(0)->getType()));
2957  break;
2958  case Instruction::PHI:
2959  // Only fold icmp into the PHI if the phi and icmp are in the same
2960  // block. If in the same block, we're encouraging jump threading. If
2961  // not, we are just pessimizing the code by making an i1 phi.
2962  if (LHSI->getParent() == I.getParent())
2963  if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
2964  return NV;
2965  break;
2966  case Instruction::Select: {
2967  // If either operand of the select is a constant, we can fold the
2968  // comparison into the select arms, which will cause one to be
2969  // constant folded and the select turned into a bitwise or.
2970  Value *Op1 = nullptr, *Op2 = nullptr;
2971  ConstantInt *CI = nullptr;
2972  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2973  Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2974  CI = dyn_cast<ConstantInt>(Op1);
2975  }
2976  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2977  Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2978  CI = dyn_cast<ConstantInt>(Op2);
2979  }
2980 
2981  // We only want to perform this transformation if it will not lead to
2982  // additional code. This is true if either both sides of the select
2983  // fold to a constant (in which case the icmp is replaced with a select
2984  // which will usually simplify) or this is the only user of the
2985  // select (in which case we are trading a select+icmp for a simpler
2986  // select+icmp) or all uses of the select can be replaced based on
2987  // dominance information ("Global cases").
2988  bool Transform = false;
2989  if (Op1 && Op2)
2990  Transform = true;
2991  else if (Op1 || Op2) {
2992  // Local case
2993  if (LHSI->hasOneUse())
2994  Transform = true;
2995  // Global cases
2996  else if (CI && !CI->isZero())
2997  // When Op1 is constant try replacing select with second operand.
2998  // Otherwise Op2 is constant and try replacing select with first
2999  // operand.
3000  Transform =
3001  replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3002  }
3003  if (Transform) {
3004  if (!Op1)
3005  Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3006  I.getName());
3007  if (!Op2)
3008  Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3009  I.getName());
3010  return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3011  }
3012  break;
3013  }
3014  case Instruction::IntToPtr:
3015  // icmp pred inttoptr(X), null -> icmp pred X, 0
3016  if (RHSC->isNullValue() &&
3017  DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3018  return new ICmpInst(
3019  I.getPredicate(), LHSI->getOperand(0),
3020  Constant::getNullValue(LHSI->getOperand(0)->getType()));
3021  break;
3022 
3023  case Instruction::Load:
3024  // Try to optimize things like "A[i] > 4" to index computations.
3025  if (GetElementPtrInst *GEP =
3026  dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3027  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3028  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3029  !cast<LoadInst>(LHSI)->isVolatile())
3030  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3031  return Res;
3032  }
3033  break;
3034  }
3035 
3036  return nullptr;
3037 }
3038 
3039 /// Some comparisons can be simplified.
3040 /// In this case, we are looking for comparisons that look like
3041 /// a check for a lossy truncation.
3042 /// Folds:
3043 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3044 /// Where Mask is some pattern that produces all-ones in low bits:
3045 /// (-1 >> y)
3046 /// ((-1 << y) >> y) <- non-canonical, has extra uses
3047 /// ~(-1 << y)
3048 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
3049 /// The Mask can be a constant, too.
3050 /// For some predicates, the operands are commutative.
3051 /// For others, x can only be on a specific side.
3053  InstCombiner::BuilderTy &Builder) {
3054  ICmpInst::Predicate SrcPred;
3055  Value *X, *M, *Y;
3056  auto m_VariableMask = m_CombineOr(
3058  m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3060  m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
3061  auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3062  if (!match(&I, m_c_ICmp(SrcPred,
3063  m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
3064  m_Deferred(X))))
3065  return nullptr;
3066 
3067  ICmpInst::Predicate DstPred;
3068  switch (SrcPred) {
3069  case ICmpInst::Predicate::ICMP_EQ:
3070  // x & (-1 >> y) == x -> x u<= (-1 >> y)
3071  DstPred = ICmpInst::Predicate::ICMP_ULE;
3072  break;
3073  case ICmpInst::Predicate::ICMP_NE:
3074  // x & (-1 >> y) != x -> x u> (-1 >> y)
3075  DstPred = ICmpInst::Predicate::ICMP_UGT;
3076  break;
3077  case ICmpInst::Predicate::ICMP_UGT:
3078  // x u> x & (-1 >> y) -> x u> (-1 >> y)
3079  assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
3080  DstPred = ICmpInst::Predicate::ICMP_UGT;
3081  break;
3082  case ICmpInst::Predicate::ICMP_UGE:
3083  // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3084  assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
3085  DstPred = ICmpInst::Predicate::ICMP_ULE;
3086  break;
3087  case ICmpInst::Predicate::ICMP_ULT:
3088  // x & (-1 >> y) u< x -> x u> (-1 >> y)
3089  assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
3090  DstPred = ICmpInst::Predicate::ICMP_UGT;
3091  break;
3092  case ICmpInst::Predicate::ICMP_ULE:
3093  // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3094  assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
3095  DstPred = ICmpInst::Predicate::ICMP_ULE;
3096  break;
3097  case ICmpInst::Predicate::ICMP_SGT:
3098  // x s> x & (-1 >> y) -> x s> (-1 >> y)
3099  if (X != I.getOperand(0)) // X must be on LHS of comparison!
3100  return nullptr; // Ignore the other case.
3101  DstPred = ICmpInst::Predicate::ICMP_SGT;
3102  break;
3103  case ICmpInst::Predicate::ICMP_SGE:
3104  // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3105  if (X != I.getOperand(1)) // X must be on RHS of comparison!
3106  return nullptr; // Ignore the other case.
3107  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3108  return nullptr;
3109  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3110  return nullptr;
3111  DstPred = ICmpInst::Predicate::ICMP_SLE;
3112  break;
3113  case ICmpInst::Predicate::ICMP_SLT:
3114  // x & (-1 >> y) s< x -> x s> (-1 >> y)
3115  if (X != I.getOperand(1)) // X must be on RHS of comparison!
3116  return nullptr; // Ignore the other case.
3117  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3118  return nullptr;
3119  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3120  return nullptr;
3121  DstPred = ICmpInst::Predicate::ICMP_SGT;
3122  break;
3123  case ICmpInst::Predicate::ICMP_SLE:
3124  // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3125  if (X != I.getOperand(0)) // X must be on LHS of comparison!
3126  return nullptr; // Ignore the other case.
3127  DstPred = ICmpInst::Predicate::ICMP_SLE;
3128  break;
3129  default:
3130  llvm_unreachable("All possible folds are handled.");
3131  }
3132 
3133  return Builder.CreateICmp(DstPred, X, M);
3134 }
3135 
3136 /// Some comparisons can be simplified.
3137 /// In this case, we are looking for comparisons that look like
3138 /// a check for a lossy signed truncation.
3139 /// Folds: (MaskedBits is a constant.)
3140 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3141 /// Into:
3142 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3143 /// Where KeptBits = bitwidth(%x) - MaskedBits
3144 static Value *
3146  InstCombiner::BuilderTy &Builder) {
3147  ICmpInst::Predicate SrcPred;
3148  Value *X;
3149  const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3150  // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3151  if (!match(&I, m_c_ICmp(SrcPred,
3152  m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3153  m_APInt(C1))),
3154  m_Deferred(X))))
3155  return nullptr;
3156 
3157  // Potential handling of non-splats: for each element:
3158  // * if both are undef, replace with constant 0.
3159  // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3160  // * if both are not undef, and are different, bailout.
3161  // * else, only one is undef, then pick the non-undef one.
3162 
3163  // The shift amount must be equal.
3164  if (*C0 != *C1)
3165  return nullptr;
3166  const APInt &MaskedBits = *C0;
3167  assert(MaskedBits != 0 && "shift by zero should be folded away already.");
3168 
3169  ICmpInst::Predicate DstPred;
3170  switch (SrcPred) {
3171  case ICmpInst::Predicate::ICMP_EQ:
3172  // ((%x << MaskedBits) a>> MaskedBits) == %x
3173  // =>
3174  // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3175  DstPred = ICmpInst::Predicate::ICMP_ULT;
3176  break;
3177  case ICmpInst::Predicate::ICMP_NE:
3178  // ((%x << MaskedBits) a>> MaskedBits) != %x
3179  // =>
3180  // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3181  DstPred = ICmpInst::Predicate::ICMP_UGE;
3182  break;
3183  // FIXME: are more folds possible?
3184  default:
3185  return nullptr;
3186  }
3187 
3188  auto *XType = X->getType();
3189  const unsigned XBitWidth = XType->getScalarSizeInBits();
3190  const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3191  assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3192 
3193  // KeptBits = bitwidth(%x) - MaskedBits
3194  const APInt KeptBits = BitWidth - MaskedBits;
3195  assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3196  // ICmpCst = (1 << KeptBits)
3197  const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3198  assert(ICmpCst.isPowerOf2());
3199  // AddCst = (1 << (KeptBits-1))
3200  const APInt AddCst = ICmpCst.lshr(1);
3201  assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3202 
3203  // T0 = add %x, AddCst
3204  Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3205  // T1 = T0 DstPred ICmpCst
3206  Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3207 
3208  return T1;
3209 }
3210 
3211 /// Try to fold icmp (binop), X or icmp X, (binop).
3212 /// TODO: A large part of this logic is duplicated in InstSimplify's
3213 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3214 /// duplication.
3215 Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I) {
3216  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3217 
3218  // Special logic for binary operators.
3219  BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3220  BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3221  if (!BO0 && !BO1)
3222  return nullptr;
3223 
3224  const CmpInst::Predicate Pred = I.getPredicate();
3225  Value *X;
3226 
3227  // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3228  // (Op1 + X) <u Op1 --> ~Op1 <u X
3229  // Op0 >u (Op0 + X) --> X >u ~Op0
3230  if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3231  Pred == ICmpInst::ICMP_ULT)
3232  return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3233  if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
3234  Pred == ICmpInst::ICMP_UGT)
3235  return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3236 
3237  bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3238  if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3239  NoOp0WrapProblem =
3240  ICmpInst::isEquality(Pred) ||
3241  (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3242  (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3243  if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3244  NoOp1WrapProblem =
3245  ICmpInst::isEquality(Pred) ||
3246  (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3247  (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3248 
3249  // Analyze the case when either Op0 or Op1 is an add instruction.
3250  // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3251  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3252  if (BO0 && BO0->getOpcode() == Instruction::Add) {
3253  A = BO0->getOperand(0);
3254  B = BO0->getOperand(1);
3255  }
3256  if (BO1 && BO1->getOpcode() == Instruction::Add) {
3257  C = BO1->getOperand(0);
3258  D = BO1->getOperand(1);
3259  }
3260 
3261  // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3262  if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3263  return new ICmpInst(Pred, A == Op1 ? B : A,
3264  Constant::getNullValue(Op1->getType()));
3265 
3266  // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3267  if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3268  return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3269  C == Op0 ? D : C);
3270 
3271  // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
3272  if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3273  NoOp1WrapProblem &&
3274  // Try not to increase register pressure.
3275  BO0->hasOneUse() && BO1->hasOneUse()) {
3276  // Determine Y and Z in the form icmp (X+Y), (X+Z).
3277  Value *Y, *Z;
3278  if (A == C) {
3279  // C + B == C + D -> B == D
3280  Y = B;
3281  Z = D;
3282  } else if (A == D) {
3283  // D + B == C + D -> B == C
3284  Y = B;
3285  Z = C;
3286  } else if (B == C) {
3287  // A + C == C + D -> A == D
3288  Y = A;
3289  Z = D;
3290  } else {
3291  assert(B == D);
3292  // A + D == C + D -> A == C
3293  Y = A;
3294  Z = C;
3295  }
3296  return new ICmpInst(Pred, Y, Z);
3297  }
3298 
3299  // icmp slt (X + -1), Y -> icmp sle X, Y
3300  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3301  match(B, m_AllOnes()))
3302  return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3303 
3304  // icmp sge (X + -1), Y -> icmp sgt X, Y
3305  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3306  match(B, m_AllOnes()))
3307  return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3308 
3309  // icmp sle (X + 1), Y -> icmp slt X, Y
3310  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
3311  return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3312 
3313  // icmp sgt (X + 1), Y -> icmp sge X, Y
3314  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
3315  return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3316 
3317  // icmp sgt X, (Y + -1) -> icmp sge X, Y
3318  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
3319  match(D, m_AllOnes()))
3320  return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3321 
3322  // icmp sle X, (Y + -1) -> icmp slt X, Y
3323  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
3324  match(D, m_AllOnes()))
3325  return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3326 
3327  // icmp sge X, (Y + 1) -> icmp sgt X, Y
3328  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
3329  return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3330 
3331  // icmp slt X, (Y + 1) -> icmp sle X, Y
3332  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
3333  return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3334 
3335  // TODO: The subtraction-related identities shown below also hold, but
3336  // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3337  // wouldn't happen even if they were implemented.
3338  //
3339  // icmp ult (X - 1), Y -> icmp ule X, Y
3340  // icmp uge (X - 1), Y -> icmp ugt X, Y
3341  // icmp ugt X, (Y - 1) -> icmp uge X, Y
3342  // icmp ule X, (Y - 1) -> icmp ult X, Y
3343 
3344  // icmp ule (X + 1), Y -> icmp ult X, Y
3345  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
3346  return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3347 
3348  // icmp ugt (X + 1), Y -> icmp uge X, Y
3349  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
3350  return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3351 
3352  // icmp uge X, (Y + 1) -> icmp ugt X, Y
3353  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
3354  return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3355 
3356  // icmp ult X, (Y + 1) -> icmp ule X, Y
3357  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
3358  return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3359 
3360  // if C1 has greater magnitude than C2:
3361  // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3362  // s.t. C3 = C1 - C2
3363  //
3364  // if C2 has greater magnitude than C1:
3365  // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3366  // s.t. C3 = C2 - C1
3367  if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3368  (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3369  if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3370  if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3371  const APInt &AP1 = C1->getValue();
3372  const APInt &AP2 = C2->getValue();
3373  if (AP1.isNegative() == AP2.isNegative()) {
3374  APInt AP1Abs = C1->getValue().abs();
3375  APInt AP2Abs = C2->getValue().abs();
3376  if (AP1Abs.uge(AP2Abs)) {
3377  ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3378  Value *NewAdd = Builder.CreateNSWAdd(A, C3);
3379  return new ICmpInst(Pred, NewAdd, C);
3380  } else {
3381  ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3382  Value *NewAdd = Builder.CreateNSWAdd(C, C3);
3383  return new ICmpInst(Pred, A, NewAdd);
3384  }
3385  }
3386  }
3387 
3388  // Analyze the case when either Op0 or Op1 is a sub instruction.
3389  // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3390  A = nullptr;
3391  B = nullptr;
3392  C = nullptr;
3393  D = nullptr;
3394  if (BO0 && BO0->getOpcode() == Instruction::Sub) {
3395  A = BO0->getOperand(0);
3396  B = BO0->getOperand(1);
3397  }
3398  if (BO1 && BO1->getOpcode() == Instruction::Sub) {
3399  C = BO1->getOperand(0);
3400  D = BO1->getOperand(1);
3401  }
3402 
3403  // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3404  if (A == Op1 && NoOp0WrapProblem)
3405  return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3406  // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3407  if (C == Op0 && NoOp1WrapProblem)
3408  return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3409 
3410  // (A - B) >u A --> A <u B
3411  if (A == Op1 && Pred == ICmpInst::ICMP_UGT)
3412  return new ICmpInst(ICmpInst::ICMP_ULT, A, B);
3413  // C <u (C - D) --> C <u D
3414  if (C == Op0 && Pred == ICmpInst::ICMP_ULT)
3415  return new ICmpInst(ICmpInst::ICMP_ULT, C, D);
3416 
3417  // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3418  if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3419  // Try not to increase register pressure.
3420  BO0->hasOneUse() && BO1->hasOneUse())
3421  return new ICmpInst(Pred, A, C);
3422  // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3423  if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3424  // Try not to increase register pressure.
3425  BO0->hasOneUse() && BO1->hasOneUse())
3426  return new ICmpInst(Pred, D, B);
3427 
3428  // icmp (0-X) < cst --> x > -cst
3429  if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3430  Value *X;
3431  if (match(BO0, m_Neg(m_Value(X))))
3432  if (Constant *RHSC = dyn_cast<Constant>(Op1))
3433  if (RHSC->isNotMinSignedValue())
3434  return new ICmpInst(I.getSwappedPredicate(), X,
3435  ConstantExpr::getNeg(RHSC));
3436  }
3437 
3438  BinaryOperator *SRem = nullptr;
3439  // icmp (srem X, Y), Y
3440  if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
3441  SRem = BO0;
3442  // icmp Y, (srem X, Y)
3443  else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3444  Op0 == BO1->getOperand(1))
3445  SRem = BO1;
3446  if (SRem) {
3447  // We don't check hasOneUse to avoid increasing register pressure because
3448  // the value we use is the same value this instruction was already using.
3449  switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3450  default:
3451  break;
3452  case ICmpInst::ICMP_EQ:
3453  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3454  case ICmpInst::ICMP_NE:
3455  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3456  case ICmpInst::ICMP_SGT:
3457  case ICmpInst::ICMP_SGE:
3458  return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3460  case ICmpInst::ICMP_SLT:
3461  case ICmpInst::ICMP_SLE:
3462  return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3463  Constant::getNullValue(SRem->getType()));
3464  }
3465  }
3466 
3467  if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
3468  BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
3469  switch (BO0->getOpcode()) {
3470  default:
3471  break;
3472  case Instruction::Add:
3473  case Instruction::Sub:
3474  case Instruction::Xor: {
3475  if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3476  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3477 
3478  const APInt *C;
3479  if (match(BO0->getOperand(1), m_APInt(C))) {
3480  // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
3481  if (C->isSignMask()) {
3482  ICmpInst::Predicate NewPred =
3484  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3485  }
3486 
3487  // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
3488  if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
3489  ICmpInst::Predicate NewPred =
3491  NewPred = I.getSwappedPredicate(NewPred);
3492  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3493  }
3494  }
3495  break;
3496  }
3497  case Instruction::Mul: {
3498  if (!I.isEquality())
3499  break;
3500 
3501  const APInt *C;
3502  if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
3503  !C->isOneValue()) {
3504  // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
3505  // Mask = -1 >> count-trailing-zeros(C).
3506  if (unsigned TZs = C->countTrailingZeros()) {
3508  BO0->getType(),
3509  APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
3510  Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
3511  Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
3512  return new ICmpInst(Pred, And1, And2);
3513  }
3514  // If there are no trailing zeros in the multiplier, just eliminate
3515  // the multiplies (no masking is needed):
3516  // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
3517  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3518  }
3519  break;
3520  }
3521  case Instruction::UDiv:
3522  case Instruction::LShr:
3523  if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
3524  break;
3525  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3526 
3527  case Instruction::SDiv:
3528  if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
3529  break;
3530  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3531 
3532  case Instruction::AShr:
3533  if (!BO0->isExact() || !BO1->isExact())
3534  break;
3535  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3536 
3537  case Instruction::Shl: {
3538  bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3539  bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3540  if (!NUW && !NSW)
3541  break;
3542  if (!NSW && I.isSigned())
3543  break;
3544  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3545  }
3546  }
3547  }
3548 
3549  if (BO0) {
3550  // Transform A & (L - 1) `ult` L --> L != 0
3551  auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
3552  auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
3553 
3554  if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
3555  auto *Zero = Constant::getNullValue(BO0->getType());
3556  return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
3557  }
3558  }
3559 
3560  if (Value *V = foldICmpWithLowBitMaskedVal(I, Builder))
3561  return replaceInstUsesWith(I, V);
3562 
3563  if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
3564  return replaceInstUsesWith(I, V);
3565 
3566  return nullptr;
3567 }
3568 
3569 /// Fold icmp Pred min|max(X, Y), X.
3571  ICmpInst::Predicate Pred = Cmp.getPredicate();
3572  Value *Op0 = Cmp.getOperand(0);
3573  Value *X = Cmp.getOperand(1);
3574 
3575  // Canonicalize minimum or maximum operand to LHS of the icmp.
3576  if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
3577  match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
3578  match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
3579  match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
3580  std::swap(Op0, X);
3581  Pred = Cmp.getSwappedPredicate();
3582  }
3583 
3584  Value *Y;
3585  if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
3586  // smin(X, Y) == X --> X s<= Y
3587  // smin(X, Y) s>= X --> X s<= Y
3588  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
3589  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
3590 
3591  // smin(X, Y) != X --> X s> Y
3592  // smin(X, Y) s< X --> X s> Y
3593  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
3594  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
3595 
3596  // These cases should be handled in InstSimplify:
3597  // smin(X, Y) s<= X --> true
3598  // smin(X, Y) s> X --> false
3599  return nullptr;
3600  }
3601 
3602  if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
3603  // smax(X, Y) == X --> X s>= Y
3604  // smax(X, Y) s<= X --> X s>= Y
3605  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
3606  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
3607 
3608  // smax(X, Y) != X --> X s< Y
3609  // smax(X, Y) s> X --> X s< Y
3610  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
3611  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
3612 
3613  // These cases should be handled in InstSimplify:
3614  // smax(X, Y) s>= X --> true
3615  // smax(X, Y) s< X --> false
3616  return nullptr;
3617  }
3618 
3619  if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
3620  // umin(X, Y) == X --> X u<= Y
3621  // umin(X, Y) u>= X --> X u<= Y
3622  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
3623  return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
3624 
3625  // umin(X, Y) != X --> X u> Y
3626  // umin(X, Y) u< X --> X u> Y
3627  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
3628  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
3629 
3630  // These cases should be handled in InstSimplify:
3631  // umin(X, Y) u<= X --> true
3632  // umin(X, Y) u> X --> false
3633  return nullptr;
3634  }
3635 
3636  if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
3637  // umax(X, Y) == X --> X u>= Y
3638  // umax(X, Y) u<= X --> X u>= Y
3639  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
3640  return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
3641 
3642  // umax(X, Y) != X --> X u< Y
3643  // umax(X, Y) u> X --> X u< Y
3644  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
3645  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
3646 
3647  // These cases should be handled in InstSimplify:
3648  // umax(X, Y) u>= X --> true
3649  // umax(X, Y) u< X --> false
3650  return nullptr;
3651  }
3652 
3653  return nullptr;
3654 }
3655 
3656 Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
3657  if (!I.isEquality())
3658  return nullptr;
3659 
3660  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3661  const CmpInst::Predicate Pred = I.getPredicate();
3662  Value *A, *B, *C, *D;
3663  if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3664  if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3665  Value *OtherVal = A == Op1 ? B : A;
3666  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3667  }
3668 
3669  if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3670  // A^c1 == C^c2 --> A == C^(c1^c2)
3671  ConstantInt *C1, *C2;
3672  if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
3673  Op1->hasOneUse()) {
3674  Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
3675  Value *Xor = Builder.CreateXor(C, NC);
3676  return new ICmpInst(Pred, A, Xor);
3677  }
3678 
3679  // A^B == A^D -> B == D
3680  if (A == C)
3681  return new ICmpInst(Pred, B, D);
3682  if (A == D)
3683  return new ICmpInst(Pred, B, C);
3684  if (B == C)
3685  return new ICmpInst(Pred, A, D);
3686  if (B == D)
3687  return new ICmpInst(Pred, A, C);
3688  }
3689  }
3690 
3691  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
3692  // A == (A^B) -> B == 0
3693  Value *OtherVal = A == Op0 ? B : A;
3694  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3695  }
3696 
3697  // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3698  if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3699  match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3700  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3701 
3702  if (A == C) {
3703  X = B;
3704  Y = D;
3705  Z = A;
3706  } else if (A == D) {
3707  X = B;
3708  Y = C;
3709  Z = A;
3710  } else if (B == C) {
3711  X = A;
3712  Y = D;
3713  Z = B;
3714  } else if (B == D) {
3715  X = A;
3716  Y = C;
3717  Z = B;
3718  }
3719 
3720  if (X) { // Build (X^Y) & Z
3721  Op1 = Builder.CreateXor(X, Y);
3722  Op1 = Builder.CreateAnd(Op1, Z);
3723  I.setOperand(0, Op1);
3724  I.setOperand(1, Constant::getNullValue(Op1->getType()));
3725  return &I;
3726  }
3727  }
3728 
3729  // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3730  // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3731  ConstantInt *Cst1;
3732  if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
3733  match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3734  (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3735  match(Op1, m_ZExt(m_Value(A))))) {
3736  APInt Pow2 = Cst1->getValue() + 1;
3737  if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3738  Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3739  return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
3740  }
3741 
3742  // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3743  // For lshr and ashr pairs.
3744  if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3745  match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3746  (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3747  match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3748  unsigned TypeBits = Cst1->getBitWidth();
3749  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3750  if (ShAmt < TypeBits && ShAmt != 0) {
3751  ICmpInst::Predicate NewPred =
3753  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3754  APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3755  return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
3756  }
3757  }
3758 
3759  // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
3760  if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
3761  match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
3762  unsigned TypeBits = Cst1->getBitWidth();
3763  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3764  if (ShAmt < TypeBits && ShAmt != 0) {
3765  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3766  APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
3767  Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
3768  I.getName() + ".mask");
3769  return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
3770  }
3771  }
3772 
3773  // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3774  // "icmp (and X, mask), cst"
3775  uint64_t ShAmt = 0;
3776  if (Op0->hasOneUse() &&
3777  match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
3778  match(Op1, m_ConstantInt(Cst1)) &&
3779  // Only do this when A has multiple uses. This is most important to do
3780  // when it exposes other optimizations.
3781  !A->hasOneUse()) {
3782  unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3783 
3784  if (ShAmt < ASize) {
3785  APInt MaskV =
3787  MaskV <<= ShAmt;
3788 
3789  APInt CmpV = Cst1->getValue().zext(ASize);
3790  CmpV <<= ShAmt;
3791 
3792  Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
3793  return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
3794  }
3795  }
3796 
3797  // If both operands are byte-swapped or bit-reversed, just compare the
3798  // original values.
3799  // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
3800  // and handle more intrinsics.
3801  if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
3802  (match(Op0, m_BitReverse(m_Value(A))) &&
3803  match(Op1, m_BitReverse(m_Value(B)))))
3804  return new ICmpInst(Pred, A, B);
3805 
3806  return nullptr;
3807 }
3808 
3809 /// Handle icmp (cast x to y), (cast/cst). We only handle extending casts so
3810 /// far.
3811 Instruction *InstCombiner::foldICmpWithCastAndCast(ICmpInst &ICmp) {
3812  const CastInst *LHSCI = cast<CastInst>(ICmp.getOperand(0));
3813  Value *LHSCIOp = LHSCI->getOperand(0);
3814  Type *SrcTy = LHSCIOp->getType();
3815  Type *DestTy = LHSCI->getType();
3816  Value *RHSCIOp;
3817 
3818  // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
3819  // integer type is the same size as the pointer type.
3820  const auto& CompatibleSizes = [&](Type* SrcTy, Type* DestTy) -> bool {
3821  if (isa<VectorType>(SrcTy)) {
3822  SrcTy = cast<VectorType>(SrcTy)->getElementType();
3823  DestTy = cast<VectorType>(DestTy)->getElementType();
3824  }
3825  return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
3826  };
3827  if (LHSCI->getOpcode() == Instruction::PtrToInt &&
3828  CompatibleSizes(SrcTy, DestTy)) {
3829  Value *RHSOp = nullptr;
3830  if (auto *RHSC = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
3831  Value *RHSCIOp = RHSC->getOperand(0);
3832  if (RHSCIOp->getType()->getPointerAddressSpace() ==
3833  LHSCIOp->getType()->getPointerAddressSpace()) {
3834  RHSOp = RHSC->getOperand(0);
3835  // If the pointer types don't match, insert a bitcast.
3836  if (LHSCIOp->getType() != RHSOp->getType())
3837  RHSOp = Builder.CreateBitCast(RHSOp, LHSCIOp->getType());
3838  }
3839  } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
3840  RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
3841  }
3842 
3843  if (RHSOp)
3844  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSOp);
3845  }
3846 
3847  // The code below only handles extension cast instructions, so far.
3848  // Enforce this.
3849  if (LHSCI->getOpcode() != Instruction::ZExt &&
3850  LHSCI->getOpcode() != Instruction::SExt)
3851  return nullptr;
3852 
3853  bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
3854  bool isSignedCmp = ICmp.isSigned();
3855 
3856  if (auto *CI = dyn_cast<CastInst>(ICmp.getOperand(1))) {
3857  // Not an extension from the same type?
3858  RHSCIOp = CI->getOperand(0);
3859  if (RHSCIOp->getType() != LHSCIOp->getType())
3860  return nullptr;
3861 
3862  // If the signedness of the two casts doesn't agree (i.e. one is a sext
3863  // and the other is a zext), then we can't handle this.
3864  if (CI->getOpcode() != LHSCI->getOpcode())
3865  return nullptr;
3866 
3867  // Deal with equality cases early.
3868  if (ICmp.isEquality())
3869  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3870 
3871  // A signed comparison of sign extended values simplifies into a
3872  // signed comparison.
3873  if (isSignedCmp && isSignedExt)
3874  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3875 
3876  // The other three cases all fold into an unsigned comparison.
3877  return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
3878  }
3879 
3880  // If we aren't dealing with a constant on the RHS, exit early.
3881  auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
3882  if (!C)
3883  return nullptr;
3884 
3885  // Compute the constant that would happen if we truncated to SrcTy then
3886  // re-extended to DestTy.
3887  Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
3888  Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
3889 
3890  // If the re-extended constant didn't change...
3891  if (Res2 == C) {
3892  // Deal with equality cases early.
3893  if (ICmp.isEquality())
3894  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3895 
3896  // A signed comparison of sign extended values simplifies into a
3897  // signed comparison.
3898  if (isSignedExt && isSignedCmp)
3899  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3900 
3901  // The other three cases all fold into an unsigned comparison.
3902  return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, Res1);
3903  }
3904 
3905  // The re-extended constant changed, partly changed (in the case of a vector),
3906  // or could not be determined to be equal (in the case of a constant
3907  // expression), so the constant cannot be represented in the shorter type.
3908  // Consequently, we cannot emit a simple comparison.
3909  // All the cases that fold to true or false will have already been handled
3910  // by SimplifyICmpInst, so only deal with the tricky case.
3911 
3912  if (isSignedCmp || !isSignedExt || !isa<ConstantInt>(C))
3913  return nullptr;
3914 
3915  // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
3916  // should have been folded away previously and not enter in here.
3917 
3918  // We're performing an unsigned comp with a sign extended value.
3919  // This is true if the input is >= 0. [aka >s -1]
3920  Constant *NegOne = Constant::getAllOnesValue(SrcTy);
3921  Value *Result = Builder.CreateICmpSGT(LHSCIOp, NegOne, ICmp.getName());
3922 
3923  // Finally, return the value computed.
3924  if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
3925  return replaceInstUsesWith(ICmp, Result);
3926 
3927  assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
3928  return BinaryOperator::CreateNot(Result);
3929 }
3930 
3931 bool InstCombiner::OptimizeOverflowCheck(OverflowCheckFlavor OCF, Value *LHS,
3932  Value *RHS, Instruction &OrigI,
3933  Value *&Result, Constant *&Overflow) {
3934  if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
3935  std::swap(LHS, RHS);
3936 
3937  auto SetResult = [&](Value *OpResult, Constant *OverflowVal, bool ReuseName) {
3938  Result = OpResult;
3939  Overflow = OverflowVal;
3940  if (ReuseName)
3941  Result->takeName(&OrigI);
3942  return true;
3943  };
3944 
3945  // If the overflow check was an add followed by a compare, the insertion point
3946  // may be pointing to the compare. We want to insert the new instructions
3947  // before the add in case there are uses of the add between the add and the
3948  // compare.
3949  Builder.SetInsertPoint(&OrigI);
3950 
3951  switch (OCF) {
3952  case OCF_INVALID:
3953  llvm_unreachable("bad overflow check kind!");
3954 
3955  case OCF_UNSIGNED_ADD: {
3956  OverflowResult OR = computeOverflowForUnsignedAdd(LHS, RHS, &OrigI);
3958  return SetResult(Builder.CreateNUWAdd(LHS, RHS), Builder.getFalse(),
3959  true);
3960 
3962  return SetResult(Builder.CreateAdd(LHS, RHS), Builder.getTrue(), true);
3963 
3964  // Fall through uadd into sadd
3966  }
3967  case OCF_SIGNED_ADD: {
3968  // X + 0 -> {X, false}
3969  if (match(RHS, m_Zero()))
3970  return SetResult(LHS, Builder.getFalse(), false);
3971 
3972  // We can strength reduce this signed add into a regular add if we can prove
3973  // that it will never overflow.
3974  if (OCF == OCF_SIGNED_ADD)
3975  if (willNotOverflowSignedAdd(LHS, RHS, OrigI))
3976  return SetResult(Builder.CreateNSWAdd(LHS, RHS), Builder.getFalse(),
3977  true);
3978  break;
3979  }
3980 
3981  case OCF_UNSIGNED_SUB:
3982  case OCF_SIGNED_SUB: {
3983  // X - 0 -> {X, false}
3984  if (match(RHS, m_Zero()))
3985  return SetResult(LHS, Builder.getFalse(), false);
3986 
3987  if (OCF == OCF_SIGNED_SUB) {
3988  if (willNotOverflowSignedSub(LHS, RHS, OrigI))
3989  return SetResult(Builder.CreateNSWSub(LHS, RHS), Builder.getFalse(),
3990  true);
3991  } else {
3992  if (willNotOverflowUnsignedSub(LHS, RHS, OrigI))
3993  return SetResult(Builder.CreateNUWSub(LHS, RHS), Builder.getFalse(),
3994  true);
3995  }
3996  break;
3997  }
3998 
3999  case OCF_UNSIGNED_MUL: {
4000  OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, &OrigI);
4002  return SetResult(Builder.CreateNUWMul(LHS, RHS), Builder.getFalse(),
4003  true);
4005  return SetResult(Builder.CreateMul(LHS, RHS), Builder.getTrue(), true);
4007  }
4008  case OCF_SIGNED_MUL:
4009  // X * undef -> undef
4010  if (isa<UndefValue>(RHS))
4011  return SetResult(RHS, UndefValue::get(Builder.getInt1Ty()), false);
4012 
4013  // X * 0 -> {0, false}
4014  if (match(RHS, m_Zero()))
4015  return SetResult(RHS, Builder.getFalse(), false);
4016 
4017  // X * 1 -> {X, false}
4018  if (match(RHS, m_One()))
4019  return SetResult(LHS, Builder.getFalse(), false);
4020 
4021  if (OCF == OCF_SIGNED_MUL)
4022  if (willNotOverflowSignedMul(LHS, RHS, OrigI))
4023  return SetResult(Builder.CreateNSWMul(LHS, RHS), Builder.getFalse(),
4024  true);
4025  break;
4026  }
4027 
4028  return false;
4029 }
4030 
4031 /// Recognize and process idiom involving test for multiplication
4032 /// overflow.
4033 ///
4034 /// The caller has matched a pattern of the form:
4035 /// I = cmp u (mul(zext A, zext B), V
4036 /// The function checks if this is a test for overflow and if so replaces
4037 /// multiplication with call to 'mul.with.overflow' intrinsic.
4038 ///
4039 /// \param I Compare instruction.
4040 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
4041 /// the compare instruction. Must be of integer type.
4042 /// \param OtherVal The other argument of compare instruction.
4043 /// \returns Instruction which must replace the compare instruction, NULL if no
4044 /// replacement required.
4046  Value *OtherVal, InstCombiner &IC) {
4047  // Don't bother doing this transformation for pointers, don't do it for
4048  // vectors.
4049  if (!isa<IntegerType>(MulVal->getType()))
4050  return nullptr;
4051 
4052  assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
4053  assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
4054  auto *MulInstr = dyn_cast<Instruction>(MulVal);
4055  if (!MulInstr)
4056  return nullptr;
4057  assert(MulInstr->getOpcode() == Instruction::Mul);
4058 
4059  auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
4060  *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
4061  assert(LHS->getOpcode() == Instruction::ZExt);
4062  assert(RHS->getOpcode() == Instruction::ZExt);
4063  Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
4064 
4065  // Calculate type and width of the result produced by mul.with.overflow.
4066  Type *TyA = A->getType(), *TyB = B->getType();
4067  unsigned WidthA = TyA->getPrimitiveSizeInBits(),
4068  WidthB = TyB->getPrimitiveSizeInBits();
4069  unsigned MulWidth;
4070  Type *MulType;
4071  if (WidthB > WidthA) {
4072  MulWidth = WidthB;
4073  MulType = TyB;
4074  } else {
4075  MulWidth = WidthA;
4076  MulType = TyA;
4077  }
4078 
4079  // In order to replace the original mul with a narrower mul.with.overflow,
4080  // all uses must ignore upper bits of the product. The number of used low
4081  // bits must be not greater than the width of mul.with.overflow.
4082  if (MulVal->hasNUsesOrMore(2))
4083  for (User *U : MulVal->users()) {
4084  if (U == &I)
4085  continue;
4086  if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4087  // Check if truncation ignores bits above MulWidth.
4088  unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
4089  if (TruncWidth > MulWidth)
4090  return nullptr;
4091  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4092  // Check if AND ignores bits above MulWidth.
4093  if (BO->getOpcode() != Instruction::And)
4094  return nullptr;
4095  if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4096  const APInt &CVal = CI->getValue();
4097  if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
4098  return nullptr;
4099  } else {
4100  // In this case we could have the operand of the binary operation
4101  // being defined in another block, and performing the replacement
4102  // could break the dominance relation.
4103  return nullptr;
4104  }
4105  } else {
4106  // Other uses prohibit this transformation.
4107  return nullptr;
4108  }
4109  }
4110 
4111  // Recognize patterns
4112  switch (I.getPredicate()) {
4113  case ICmpInst::ICMP_EQ:
4114  case ICmpInst::ICMP_NE:
4115  // Recognize pattern:
4116  // mulval = mul(zext A, zext B)
4117  // cmp eq/neq mulval, zext trunc mulval
4118  if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
4119  if (Zext->hasOneUse()) {
4120  Value *ZextArg = Zext->getOperand(0);
4121  if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
4122  if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
4123  break; //Recognized
4124  }
4125 
4126  // Recognize pattern:
4127  // mulval = mul(zext A, zext B)
4128  // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
4129  ConstantInt *CI;
4130  Value *ValToMask;
4131  if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
4132  if (ValToMask != MulVal)
4133  return nullptr;
4134  const APInt &CVal = CI->getValue() + 1;
4135  if (CVal.isPowerOf2()) {
4136  unsigned MaskWidth = CVal.logBase2();
4137  if (MaskWidth == MulWidth)
4138  break; // Recognized
4139  }
4140  }
4141  return nullptr;
4142 
4143  case ICmpInst::ICMP_UGT:
4144  // Recognize pattern:
4145  // mulval = mul(zext A, zext B)
4146  // cmp ugt mulval, max
4147  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4148  APInt MaxVal = APInt::getMaxValue(MulWidth);
4149  MaxVal = MaxVal.zext(CI->getBitWidth());
4150  if (MaxVal.eq(CI->getValue()))
4151  break; // Recognized
4152  }
4153  return nullptr;
4154 
4155  case ICmpInst::ICMP_UGE:
4156  // Recognize pattern:
4157  // mulval = mul(zext A, zext B)
4158  // cmp uge mulval, max+1
4159  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4160  APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4161  if (MaxVal.eq(CI->getValue()))
4162  break; // Recognized
4163  }
4164  return nullptr;
4165 
4166  case ICmpInst::ICMP_ULE:
4167  // Recognize pattern:
4168  // mulval = mul(zext A, zext B)
4169  // cmp ule mulval, max
4170  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4171  APInt MaxVal = APInt::getMaxValue(MulWidth);
4172  MaxVal = MaxVal.zext(CI->getBitWidth());
4173  if (MaxVal.eq(CI->getValue()))
4174  break; // Recognized
4175  }
4176  return nullptr;
4177 
4178  case ICmpInst::ICMP_ULT:
4179  // Recognize pattern:
4180  // mulval = mul(zext A, zext B)
4181  // cmp ule mulval, max + 1
4182  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4183  APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4184  if (MaxVal.eq(CI->getValue()))
4185  break; // Recognized
4186  }
4187  return nullptr;
4188 
4189  default:
4190  return nullptr;
4191  }
4192 
4193  InstCombiner::BuilderTy &Builder = IC.Builder;
4194  Builder.SetInsertPoint(MulInstr);
4195 
4196  // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4197  Value *MulA = A, *MulB = B;
4198  if (WidthA < MulWidth)
4199  MulA = Builder.CreateZExt(A, MulType);
4200  if (WidthB < MulWidth)
4201  MulB = Builder.CreateZExt(B, MulType);
4203  I.getModule(), Intrinsic::umul_with_overflow, MulType);
4204  CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4205  IC.Worklist.Add(MulInstr);
4206 
4207  // If there are uses of mul result other than the comparison, we know that
4208  // they are truncation or binary AND. Change them to use result of
4209  // mul.with.overflow and adjust properly mask/size.
4210  if (MulVal->hasNUsesOrMore(2)) {
4211  Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4212  for (auto UI = MulVal->user_begin(), UE = MulVal->user_end(); UI != UE;) {
4213  User *U = *UI++;
4214  if (U == &I || U == OtherVal)
4215  continue;
4216  if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4217  if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
4218  IC.replaceInstUsesWith(*TI, Mul);
4219  else
4220  TI->setOperand(0, Mul);
4221  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4222  assert(BO->getOpcode() == Instruction::And);
4223  // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
4224  ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
4225  APInt ShortMask = CI->getValue().trunc(MulWidth);
4226  Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
4227  Instruction *Zext =
4228  cast<Instruction>(Builder.CreateZExt(ShortAnd, BO->getType()));
4229  IC.Worklist.Add(Zext);
4230  IC.replaceInstUsesWith(*BO, Zext);
4231  } else {
4232  llvm_unreachable("Unexpected Binary operation");
4233  }
4234  IC.Worklist.Add(cast<Instruction>(U));
4235  }
4236  }
4237  if (isa<Instruction>(OtherVal))
4238  IC.Worklist.Add(cast<Instruction>(OtherVal));
4239 
4240  // The original icmp gets replaced with the overflow value, maybe inverted
4241  // depending on predicate.
4242  bool Inverse = false;
4243  switch (I.getPredicate()) {
4244  case ICmpInst::ICMP_NE:
4245  break;
4246  case ICmpInst::ICMP_EQ:
4247  Inverse = true;
4248  break;
4249  case ICmpInst::ICMP_UGT:
4250  case ICmpInst::ICMP_UGE:
4251  if (I.getOperand(0) == MulVal)
4252  break;
4253  Inverse = true;
4254  break;
4255  case ICmpInst::ICMP_ULT:
4256  case ICmpInst::ICMP_ULE:
4257  if (I.getOperand(1) == MulVal)
4258  break;
4259  Inverse = true;
4260  break;
4261  default:
4262  llvm_unreachable("Unexpected predicate");
4263  }
4264  if (Inverse) {
4265  Value *Res = Builder.CreateExtractValue(Call, 1);
4266  return BinaryOperator::CreateNot(Res);
4267  }
4268 
4269  return ExtractValueInst::Create(Call, 1);
4270 }
4271 
4272 /// When performing a comparison against a constant, it is possible that not all
4273 /// the bits in the LHS are demanded. This helper method computes the mask that
4274 /// IS demanded.
4275 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
4276  const APInt *RHS;
4277  if (!match(I.getOperand(1), m_APInt(RHS)))
4278  return APInt::getAllOnesValue(BitWidth);
4279 
4280  // If this is a normal comparison, it demands all bits. If it is a sign bit
4281  // comparison, it only demands the sign bit.
4282  bool UnusedBit;
4283  if (isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
4284  return APInt::getSignMask(BitWidth);
4285 
4286  switch (I.getPredicate()) {
4287  // For a UGT comparison, we don't care about any bits that
4288  // correspond to the trailing ones of the comparand. The value of these
4289  // bits doesn't impact the outcome of the comparison, because any value
4290  // greater than the RHS must differ in a bit higher than these due to carry.
4291  case ICmpInst::ICMP_UGT:
4292  return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
4293 
4294  // Similarly, for a ULT comparison, we don't care about the trailing zeros.
4295  // Any value less than the RHS must differ in a higher bit because of carries.
4296  case ICmpInst::ICMP_ULT:
4297  return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
4298 
4299  default:
4300  return APInt::getAllOnesValue(BitWidth);
4301  }
4302 }
4303 
4304 /// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
4305 /// should be swapped.
4306 /// The decision is based on how many times these two operands are reused
4307 /// as subtract operands and their positions in those instructions.
4308 /// The rationale is that several architectures use the same instruction for
4309 /// both subtract and cmp. Thus, it is better if the order of those operands
4310 /// match.
4311 /// \return true if Op0 and Op1 should be swapped.
4312 static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
4313  // Filter out pointer values as those cannot appear directly in subtract.
4314  // FIXME: we may want to go through inttoptrs or bitcasts.
4315  if (Op0->getType()->isPointerTy())
4316  return false;
4317  // If a subtract already has the same operands as a compare, swapping would be
4318  // bad. If a subtract has the same operands as a compare but in reverse order,
4319  // then swapping is good.
4320  int GoodToSwap = 0;
4321  for (const User *U : Op0->users()) {
4322  if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
4323  GoodToSwap++;
4324  else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
4325  GoodToSwap--;
4326  }
4327  return GoodToSwap > 0;
4328 }
4329 
4330 /// Check that one use is in the same block as the definition and all
4331 /// other uses are in blocks dominated by a given block.
4332 ///
4333 /// \param DI Definition
4334 /// \param UI Use
4335 /// \param DB Block that must dominate all uses of \p DI outside
4336 /// the parent block
4337 /// \return true when \p UI is the only use of \p DI in the parent block
4338 /// and all other uses of \p DI are in blocks dominated by \p DB.
4339 ///
4341  const Instruction *UI,
4342  const BasicBlock *DB) const {
4343  assert(DI && UI && "Instruction not defined\n");
4344  // Ignore incomplete definitions.
4345  if (!DI->getParent())
4346  return false;
4347  // DI and UI must be in the same block.
4348  if (DI->getParent() != UI->getParent())
4349  return false;
4350  // Protect from self-referencing blocks.
4351  if (DI->getParent() == DB)
4352  return false;
4353  for (const User *U : DI->users()) {
4354  auto *Usr = cast<Instruction>(U);
4355  if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4356  return false;
4357  }
4358  return true;
4359 }
4360 
4361 /// Return true when the instruction sequence within a block is select-cmp-br.
4362 static bool isChainSelectCmpBranch(const SelectInst *SI) {
4363  const BasicBlock *BB = SI->getParent();
4364  if (!BB)
4365  return false;
4366  auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4367  if (!BI || BI->getNumSuccessors() != 2)
4368  return false;
4369  auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4370  if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4371  return false;
4372  return true;
4373 }
4374 
4375 /// True when a select result is replaced by one of its operands
4376 /// in select-icmp sequence. This will eventually result in the elimination
4377 /// of the select.
4378 ///
4379 /// \param SI Select instruction
4380 /// \param Icmp Compare instruction
4381 /// \param SIOpd Operand that replaces the select
4382 ///
4383 /// Notes:
4384 /// - The replacement is global and requires dominator information
4385 /// - The caller is responsible for the actual replacement
4386 ///
4387 /// Example:
4388 ///
4389 /// entry:
4390 /// %4 = select i1 %3, %C* %0, %C* null
4391 /// %5 = icmp eq %C* %4, null
4392 /// br i1 %5, label %9, label %7
4393 /// ...
4394 /// ; <label>:7 ; preds = %entry
4395 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4396 /// ...
4397 ///
4398 /// can be transformed to
4399 ///
4400 /// %5 = icmp eq %C* %0, null
4401 /// %6 = select i1 %3, i1 %5, i1 true
4402 /// br i1 %6, label %9, label %7
4403 /// ...
4404 /// ; <label>:7 ; preds = %entry
4405 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
4406 ///
4407 /// Similar when the first operand of the select is a constant or/and
4408 /// the compare is for not equal rather than equal.
4409 ///
4410 /// NOTE: The function is only called when the select and compare constants
4411 /// are equal, the optimization can work only for EQ predicates. This is not a
4412 /// major restriction since a NE compare should be 'normalized' to an equal
4413 /// compare, which usually happens in the combiner and test case
4414 /// select-cmp-br.ll checks for it.
4416  const ICmpInst *Icmp,
4417  const unsigned SIOpd) {
4418  assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
4419  if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4420  BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4421  // The check for the single predecessor is not the best that can be
4422  // done. But it protects efficiently against cases like when SI's
4423  // home block has two successors, Succ and Succ1, and Succ1 predecessor
4424  // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4425  // replaced can be reached on either path. So the uniqueness check
4426  // guarantees that the path all uses of SI (outside SI's parent) are on
4427  // is disjoint from all other paths out of SI. But that information
4428  // is more expensive to compute, and the trade-off here is in favor
4429  // of compile-time. It should also be noticed that we check for a single
4430  // predecessor and not only uniqueness. This to handle the situation when
4431  // Succ and Succ1 points to the same basic block.
4432  if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
4433  NumSel++;
4434  SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
4435  return true;
4436  }
4437  }
4438  return false;
4439 }
4440 
4441 /// Try to fold the comparison based on range information we can get by checking
4442 /// whether bits are known to be zero or one in the inputs.
4443 Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) {
4444  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4445  Type *Ty = Op0->getType();
4446  ICmpInst::Predicate Pred = I.getPredicate();
4447 
4448  // Get scalar or pointer size.
4449  unsigned BitWidth = Ty->isIntOrIntVectorTy()
4450  ? Ty->getScalarSizeInBits()
4451  : DL.getIndexTypeSizeInBits(Ty->getScalarType());
4452 
4453  if (!BitWidth)
4454  return nullptr;
4455 
4456  KnownBits Op0Known(BitWidth);
4457  KnownBits Op1Known(BitWidth);
4458 
4459  if (SimplifyDemandedBits(&I, 0,
4460  getDemandedBitsLHSMask(I, BitWidth),
4461  Op0Known, 0))
4462  return &I;
4463 
4464  if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
4465  Op1Known, 0))
4466  return &I;
4467 
4468  // Given the known and unknown bits, compute a range that the LHS could be
4469  // in. Compute the Min, Max and RHS values based on the known bits. For the
4470  // EQ and NE we use unsigned values.
4471  APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
4472  APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
4473  if (I.isSigned()) {
4474  computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4475  computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4476  } else {
4477  computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4478  computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4479  }
4480 
4481  // If Min and Max are known to be the same, then SimplifyDemandedBits figured
4482  // out that the LHS or RHS is a constant. Constant fold this now, so that
4483  // code below can assume that Min != Max.
4484  if (!isa<Constant>(Op0) && Op0Min == Op0Max)
4485  return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
4486  if (!isa<Constant>(Op1) && Op1Min == Op1Max)
4487  return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
4488 
4489  // Based on the range information we know about the LHS, see if we can
4490  // simplify this comparison. For example, (x&4) < 8 is always true.
4491  switch (Pred) {
4492  default:
4493  llvm_unreachable("Unknown icmp opcode!");
4494  case ICmpInst::ICMP_EQ:
4495  case ICmpInst::ICMP_NE: {
4496  if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
4497  return Pred == CmpInst::ICMP_EQ
4498  ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
4499  : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4500  }
4501 
4502  // If all bits are known zero except for one, then we know at most one bit
4503  // is set. If the comparison is against zero, then this is a check to see if
4504  // *that* bit is set.
4505  APInt Op0KnownZeroInverted = ~Op0Known.Zero;
4506  if (Op1Known.isZero()) {
4507  // If the LHS is an AND with the same constant, look through it.
4508  Value *LHS = nullptr;
4509  const APInt *LHSC;
4510  if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
4511  *LHSC != Op0KnownZeroInverted)
4512  LHS = Op0;
4513 
4514  Value *X;
4515  if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
4516  APInt ValToCheck = Op0KnownZeroInverted;
4517  Type *XTy = X->getType();
4518  if (ValToCheck.isPowerOf2()) {
4519  // ((1 << X) & 8) == 0 -> X != 3
4520  // ((1 << X) & 8) != 0 -> X == 3
4521  auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4522  auto NewPred = ICmpInst::getInversePredicate(Pred);
4523  return new ICmpInst(NewPred, X, CmpC);
4524  } else if ((++ValToCheck).isPowerOf2()) {
4525  // ((1 << X) & 7) == 0 -> X >= 3
4526  // ((1 << X) & 7) != 0 -> X < 3
4527  auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4528  auto NewPred =
4530  return new ICmpInst(NewPred, X, CmpC);
4531  }
4532  }
4533 
4534  // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
4535  const APInt *CI;
4536  if (Op0KnownZeroInverted.isOneValue() &&
4537  match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
4538  // ((8 >>u X) & 1) == 0 -> X != 3
4539  // ((8 >>u X) & 1) != 0 -> X == 3
4540  unsigned CmpVal = CI->countTrailingZeros();
4541  auto NewPred = ICmpInst::getInversePredicate(Pred);
4542  return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
4543  }
4544  }
4545  break;
4546  }
4547  case ICmpInst::ICMP_ULT: {
4548  if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
4549  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4550  if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
4551  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4552  if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
4553  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4554 
4555  const APInt *CmpC;
4556  if (match(Op1, m_APInt(CmpC))) {
4557  // A <u C -> A == C-1 if min(A)+1 == C
4558  if (*CmpC == Op0Min + 1)
4559  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4560  ConstantInt::get(Op1->getType(), *CmpC - 1));
4561  // X <u C --> X == 0, if the number of zero bits in the bottom of X
4562  // exceeds the log2 of C.
4563  if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
4564  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4565  Constant::getNullValue(Op1->getType()));
4566  }
4567  break;
4568  }
4569  case ICmpInst::ICMP_UGT: {
4570  if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
4571  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4572  if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
4573  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4574  if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
4575  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4576 
4577  const APInt *CmpC;
4578  if (match(Op1, m_APInt(CmpC))) {
4579  // A >u C -> A == C+1 if max(a)-1 == C
4580  if (*CmpC == Op0Max - 1)
4581  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4582  ConstantInt::get(Op1->getType(), *CmpC + 1));
4583  // X >u C --> X != 0, if the number of zero bits in the bottom of X
4584  // exceeds the log2 of C.
4585  if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
4586  return new ICmpInst(ICmpInst::ICMP_NE, Op0,
4587  Constant::getNullValue(Op1->getType()));
4588  }
4589  break;
4590  }
4591  case ICmpInst::ICMP_SLT: {
4592  if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
4593  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4594  if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
4595  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4596  if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
4597  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4598  const APInt *CmpC;
4599  if (match(Op1, m_APInt(CmpC))) {
4600  if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
4601  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4602  ConstantInt::get(Op1->getType(), *CmpC - 1));
4603  }
4604  break;
4605  }
4606  case ICmpInst::ICMP_SGT: {
4607  if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
4608  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4609  if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
4610  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4611  if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
4612  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4613  const APInt *CmpC;
4614  if (match(Op1, m_APInt(CmpC))) {
4615  if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
4616  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4617  ConstantInt::get(Op1->getType(), *CmpC + 1));
4618  }
4619  break;
4620  }
4621  case ICmpInst::ICMP_SGE:
4622  assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
4623  if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
4624  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4625  if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
4626  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4627  if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
4628  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4629  break;
4630  case ICmpInst::ICMP_SLE:
4631  assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
4632  if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
4633  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4634  if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
4635  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4636  if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
4637  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4638  break;
4639  case ICmpInst::ICMP_UGE:
4640  assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
4641  if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
4642  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4643  if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
4644  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4645  if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
4646  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4647  break;
4648  case ICmpInst::ICMP_ULE:
4649  assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
4650  if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
4651  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4652  if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
4653  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4654  if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
4655  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4656  break;
4657  }
4658 
4659  // Turn a signed comparison into an unsigned one if both operands are known to
4660  // have the same sign.
4661  if (I.isSigned() &&
4662  ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
4663  (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
4664  return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
4665 
4666  return nullptr;
4667 }
4668 
4669 /// If we have an icmp le or icmp ge instruction with a constant operand, turn
4670 /// it into the appropriate icmp lt or icmp gt instruction. This transform
4671 /// allows them to be folded in visitICmpInst.
4673  ICmpInst::Predicate Pred = I.getPredicate();
4674  if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGE &&
4675  Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_UGE)
4676  return nullptr;
4677 
4678  Value *Op0 = I.getOperand(0);
4679  Value *Op1 = I.getOperand(1);
4680  auto *Op1C = dyn_cast<Constant>(Op1);
4681  if (!Op1C)
4682  return nullptr;
4683 
4684  // Check if the constant operand can be safely incremented/decremented without
4685  // overflowing/underflowing. For scalars, SimplifyICmpInst has already handled
4686  // the edge cases for us, so we just assert on them. For vectors, we must
4687  // handle the edge cases.
4688  Type *Op1Type = Op1->getType();
4689  bool IsSigned = I.isSigned();
4690  bool IsLE = (Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_ULE);
4691  auto *CI = dyn_cast<ConstantInt>(Op1C);
4692  if (CI) {
4693  // A <= MAX -> TRUE ; A >= MIN -> TRUE
4694  assert(IsLE ? !CI->isMaxValue(IsSigned) : !CI->isMinValue(IsSigned));
4695  } else if (Op1Type->isVectorTy()) {
4696  // TODO? If the edge cases for vectors were guaranteed to be handled as they
4697  // are for scalar, we could remove the min/max checks. However, to do that,
4698  // we would have to use insertelement/shufflevector to replace edge values.
4699  unsigned NumElts = Op1Type->getVectorNumElements();
4700  for (unsigned i = 0; i != NumElts; ++i) {
4701  Constant *Elt = Op1C->getAggregateElement(i);
4702  if (!Elt)
4703  return nullptr;
4704 
4705  if (isa<UndefValue>(Elt))
4706  continue;
4707 
4708  // Bail out if we can't determine if this constant is min/max or if we
4709  // know that this constant is min/max.
4710  auto *CI = dyn_cast<ConstantInt>(Elt);
4711  if (!CI || (IsLE ? CI->isMaxValue(IsSigned) : CI->isMinValue(IsSigned)))
4712  return nullptr;
4713  }
4714  } else {
4715  // ConstantExpr?
4716  return nullptr;
4717  }
4718 
4719  // Increment or decrement the constant and set the new comparison predicate:
4720  // ULE -> ULT ; UGE -> UGT ; SLE -> SLT ; SGE -> SGT
4721  Constant *OneOrNegOne = ConstantInt::get(Op1Type, IsLE ? 1 : -1, true);
4723  NewPred = IsSigned ? ICmpInst::getSignedPredicate(NewPred) : NewPred;
4724  return new ICmpInst(NewPred, Op0, ConstantExpr::getAdd(Op1C, OneOrNegOne));
4725 }
4726 
4727 /// Integer compare with boolean values can always be turned into bitwise ops.
4729  InstCombiner::BuilderTy &Builder) {
4730  Value *A = I.getOperand(0), *B = I.getOperand(1);
4731  assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only");
4732 
4733  // A boolean compared to true/false can be simplified to Op0/true/false in
4734  // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
4735  // Cases not handled by InstSimplify are always 'not' of Op0.
4736  if (match(B, m_Zero())) {
4737  switch (I.getPredicate()) {
4738  case CmpInst::ICMP_EQ: // A == 0 -> !A
4739  case CmpInst::ICMP_ULE: // A <=u 0 -> !A
4740  case CmpInst::ICMP_SGE: // A >=s 0 -> !A
4741  return BinaryOperator::CreateNot(A);
4742  default:
4743  llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4744  }
4745  } else if (match(B, m_One())) {
4746  switch (I.getPredicate()) {
4747  case CmpInst::ICMP_NE: // A != 1 -> !A
4748  case CmpInst::ICMP_ULT: // A <u 1 -> !A
4749  case CmpInst::ICMP_SGT: // A >s -1 -> !A
4750  return BinaryOperator::CreateNot(A);
4751  default:
4752  llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4753  }
4754  }
4755 
4756  switch (I.getPredicate()) {
4757  default:
4758  llvm_unreachable("Invalid icmp instruction!");
4759  case ICmpInst::ICMP_EQ:
4760  // icmp eq i1 A, B -> ~(A ^ B)
4761  return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
4762 
4763  case ICmpInst::ICMP_NE:
4764  // icmp ne i1 A, B -> A ^ B
4765  return BinaryOperator::CreateXor(A, B);
4766 
4767  case ICmpInst::ICMP_UGT:
4768  // icmp ugt -> icmp ult
4769  std::swap(A, B);
4771  case ICmpInst::ICMP_ULT:
4772  // icmp ult i1 A, B -> ~A & B
4773  return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
4774 
4775  case ICmpInst::ICMP_SGT:
4776  // icmp sgt -> icmp slt
4777  std::swap(A, B);
4779  case ICmpInst::ICMP_SLT:
4780  // icmp slt i1 A, B -> A & ~B
4781  return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
4782 
4783  case ICmpInst::ICMP_UGE:
4784  // icmp uge -> icmp ule
4785  std::swap(A, B);
4787  case ICmpInst::ICMP_ULE:
4788  // icmp ule i1 A, B -> ~A | B
4789  return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
4790 
4791  case ICmpInst::ICMP_SGE:
4792  // icmp sge -> icmp sle
4793  std::swap(A, B);
4795  case ICmpInst::ICMP_SLE:
4796  // icmp sle i1 A, B -> A | ~B
4797  return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
4798  }
4799 }
4800 
4801 // Transform pattern like:
4802 // (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
4803 // (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
4804 // Into:
4805 // (X l>> Y) != 0
4806 // (X l>> Y) == 0
4808  InstCombiner::BuilderTy &Builder) {
4809  ICmpInst::Predicate Pred, NewPred;
4810  Value *X, *Y;
4811  if (match(&Cmp,
4812  m_c_ICmp(Pred, m_OneUse(m_Shl(m_One(), m_Value(Y))), m_Value(X)))) {
4813  // We want X to be the icmp's second operand, so swap predicate if it isn't.
4814  if (Cmp.getOperand(0) == X)
4815  Pred = Cmp.getSwappedPredicate();
4816 
4817  switch (Pred) {
4818  case ICmpInst::ICMP_ULE:
4819  NewPred = ICmpInst::ICMP_NE;
4820  break;
4821  case ICmpInst::ICMP_UGT:
4822  NewPred = ICmpInst::ICMP_EQ;
4823  break;
4824  default:
4825  return nullptr;
4826  }
4827  } else if (match(&Cmp, m_c_ICmp(Pred,
4829  m_Not(m_Shl(m_AllOnes(), m_Value(Y))),
4830  m_Add(m_Shl(m_One(), m_Value(Y)),
4831  m_AllOnes()))),
4832  m_Value(X)))) {
4833  // The variant with 'add' is not canonical, (the variant with 'not' is)
4834  // we only get it because it has extra uses, and can't be canonicalized,
4835 
4836  // We want X to be the icmp's second operand, so swap predicate if it isn't.
4837  if (Cmp.getOperand(0) == X)
4838  Pred = Cmp.getSwappedPredicate();
4839 
4840  switch (Pred) {
4841  case ICmpInst::ICMP_ULT:
4842  NewPred = ICmpInst::ICMP_NE;
4843  break;
4844  case ICmpInst::ICMP_UGE:
4845  NewPred = ICmpInst::ICMP_EQ;
4846  break;
4847  default:
4848  return nullptr;
4849  }
4850  } else
4851  return nullptr;
4852 
4853  Value *NewX = Builder.CreateLShr(X, Y, X->getName() + ".highbits");
4854  Constant *Zero = Constant::getNullValue(NewX->getType());
4855  return CmpInst::Create(Instruction::ICmp, NewPred, NewX, Zero);
4856 }
4857 
4859  InstCombiner::BuilderTy &Builder) {
4860  // If both arguments of the cmp are shuffles that use the same mask and
4861  // shuffle within a single vector, move the shuffle after the cmp.
4862  Value *LHS = Cmp.getOperand(0), *RHS = Cmp.getOperand(1);
4863  Value *V1, *V2;
4864  Constant *M;
4865  if (match(LHS, m_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(M))) &&
4866  match(RHS, m_ShuffleVector(m_Value(V2), m_Undef(), m_Specific(M))) &&
4867  V1->getType() == V2->getType() &&
4868  (LHS->hasOneUse() || RHS->hasOneUse())) {
4869  // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
4871  Value *NewCmp = isa<ICmpInst>(Cmp) ? Builder.CreateICmp(P, V1, V2)
4872  : Builder.CreateFCmp(P, V1, V2);
4873  return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()), M);
4874  }
4875  return nullptr;
4876 }
4877 
4879  bool Changed = false;
4880  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4881  unsigned Op0Cplxity = getComplexity(Op0);
4882  unsigned Op1Cplxity = getComplexity(Op1);
4883 
4884  /// Orders the operands of the compare so that they are listed from most
4885  /// complex to least complex. This puts constants before unary operators,
4886  /// before binary operators.
4887  if (Op0Cplxity < Op1Cplxity ||
4888  (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
4889  I.swapOperands();
4890  std::swap(Op0, Op1);
4891  Changed = true;
4892  }
4893 
4894  if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1,
4895  SQ.getWithInstruction(&I)))
4896  return replaceInstUsesWith(I, V);
4897 
4898  // Comparing -val or val with non-zero is the same as just comparing val
4899  // ie, abs(val) != 0 -> val != 0
4900  if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
4901  Value *Cond, *SelectTrue, *SelectFalse;
4902  if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
4903  m_Value(SelectFalse)))) {
4904  if (Value *V = dyn_castNegVal(SelectTrue)) {
4905  if (V == SelectFalse)
4906  return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4907  }
4908  else if (Value *V = dyn_castNegVal(SelectFalse)) {
4909  if (V == SelectTrue)
4910  return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4911  }
4912  }
4913  }
4914 
4915  if (Op0->getType()->isIntOrIntVectorTy(1))
4916  if (Instruction *Res = canonicalizeICmpBool(I, Builder))
4917  return Res;
4918 
4919  if (ICmpInst *NewICmp = canonicalizeCmpWithConstant(I))
4920  return NewICmp;
4921 
4922  if (Instruction *Res = foldICmpWithConstant(I))
4923  return Res;
4924 
4925  if (Instruction *Res = foldICmpWithDominatingICmp(I))
4926  return Res;
4927 
4928  if (Instruction *Res = foldICmpUsingKnownBits(I))
4929  return Res;
4930 
4931  // Test if the ICmpInst instruction is used exclusively by a select as
4932  // part of a minimum or maximum operation. If so, refrain from doing
4933  // any other folding. This helps out other analyses which understand
4934  // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
4935  // and CodeGen. And in this case, at least one of the comparison
4936  // operands has at least one user besides the compare (the select),
4937  // which would often largely negate the benefit of folding anyway.
4938  //
4939  // Do the same for the other patterns recognized by matchSelectPattern.
4940  if (I.hasOneUse())
4941  if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
4942  Value *A, *B;
4943  SelectPatternResult SPR = matchSelectPattern(SI, A, B);
4944  if (SPR.Flavor != SPF_UNKNOWN)
4945  return nullptr;
4946  }
4947 
4948  // Do this after checking for min/max to prevent infinite looping.
4949  if (Instruction *Res = foldICmpWithZero(I))
4950  return Res;
4951 
4952  // FIXME: We only do this after checking for min/max to prevent infinite
4953  // looping caused by a reverse canonicalization of these patterns for min/max.
4954  // FIXME: The organization of folds is a mess. These would naturally go into
4955  // canonicalizeCmpWithConstant(), but we can't move all of the above folds
4956  // down here after the min/max restriction.
4957  ICmpInst::Predicate Pred = I.getPredicate();
4958  const APInt *C;
4959  if (match(Op1, m_APInt(C))) {
4960  // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
4961  if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
4962  Constant *Zero = Constant::getNullValue(Op0->getType());
4963  return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
4964  }
4965 
4966  // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
4967  if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
4968  Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
4969  return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
4970  }
4971  }
4972 
4973  if (Instruction *Res = foldICmpInstWithConstant(I))
4974  return Res;
4975 
4976  if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
4977  return Res;
4978 
4979  // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4980  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
4981  if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I))
4982  return NI;
4983  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
4984  if (Instruction *NI = foldGEPICmp(GEP, Op0,
4986  return NI;
4987 
4988  // Try to optimize equality comparisons against alloca-based pointers.
4989  if (Op0->getType()->isPointerTy() && I.isEquality()) {
4990  assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?");
4991  if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op0, DL)))
4992  if (Instruction *New = foldAllocaCmp(I, Alloca, Op1))
4993  return New;
4994  if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op1, DL)))
4995  if (Instruction *New = foldAllocaCmp(I, Alloca, Op0))
4996  return New;
4997  }
4998 
4999  if (Instruction *Res = foldICmpBitCast(I, Builder))
5000  return Res;
5001 
5002  if (isa<CastInst>(Op0)) {
5003  // Handle the special case of: icmp (cast bool to X), <cst>
5004  // This comes up when you have code like
5005  // int X = A < B;
5006  // if (X) ...
5007  // For generality, we handle any zero-extension of any operand comparison
5008  // with a constant or another cast from the same type.
5009  if (isa<Constant>(Op1) || isa<CastInst>(Op1))
5010  if (Instruction *R = foldICmpWithCastAndCast(I))
5011  return R;
5012  }
5013 
5014  if (Instruction *Res = foldICmpBinOp(I))
5015  return Res;
5016 
5017  if (Instruction *Res = foldICmpWithMinMax(I))
5018  return Res;
5019 
5020  {
5021  Value *A, *B;
5022  // Transform (A & ~B) == 0 --> (A & B) != 0
5023  // and (A & ~B) != 0 --> (A & B) == 0
5024  // if A is a power of 2.
5025  if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
5026  match(Op1, m_Zero()) &&
5027  isKnownToBeAPowerOfTwo(A, false, 0, &I) && I.isEquality())
5028  return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(A, B),
5029  Op1);
5030 
5031  // ~X < ~Y --> Y < X
5032  // ~X < C --> X > ~C
5033  if (match(Op0, m_Not(m_Value(A)))) {
5034  if (match(Op1, m_Not(m_Value(B))))
5035  return new ICmpInst(I.getPredicate(), B, A);
5036 
5037  const APInt *C;
5038  if (match(Op1, m_APInt(C)))
5039  return new ICmpInst(I.getSwappedPredicate(), A,
5040  ConstantInt::get(Op1->getType(), ~(*C)));
5041  }
5042