• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- InstCombineCompares.cpp --------------------------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the visitICmp and visitFCmp functions.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/APSInt.h"
16 #include "llvm/ADT/SetVector.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/Analysis/ConstantFolding.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/TargetLibraryInfo.h"
21 #include "llvm/IR/ConstantRange.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/GetElementPtrTypeIterator.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/PatternMatch.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/KnownBits.h"
28 
29 using namespace llvm;
30 using namespace PatternMatch;
31 
32 #define DEBUG_TYPE "instcombine"
33 
34 // How many times is a select replaced by one of its operands?
35 STATISTIC(NumSel, "Number of select opts");
36 
37 
38 /// Compute Result = In1+In2, returning true if the result overflowed for this
39 /// type.
addWithOverflow(APInt & Result,const APInt & In1,const APInt & In2,bool IsSigned=false)40 static bool addWithOverflow(APInt &Result, const APInt &In1,
41                             const APInt &In2, bool IsSigned = false) {
42   bool Overflow;
43   if (IsSigned)
44     Result = In1.sadd_ov(In2, Overflow);
45   else
46     Result = In1.uadd_ov(In2, Overflow);
47 
48   return Overflow;
49 }
50 
51 /// Compute Result = In1-In2, returning true if the result overflowed for this
52 /// type.
subWithOverflow(APInt & Result,const APInt & In1,const APInt & In2,bool IsSigned=false)53 static bool subWithOverflow(APInt &Result, const APInt &In1,
54                             const APInt &In2, bool IsSigned = false) {
55   bool Overflow;
56   if (IsSigned)
57     Result = In1.ssub_ov(In2, Overflow);
58   else
59     Result = In1.usub_ov(In2, Overflow);
60 
61   return Overflow;
62 }
63 
64 /// Given an icmp instruction, return true if any use of this comparison is a
65 /// branch on sign bit comparison.
hasBranchUse(ICmpInst & I)66 static bool hasBranchUse(ICmpInst &I) {
67   for (auto *U : I.users())
68     if (isa<BranchInst>(U))
69       return true;
70   return false;
71 }
72 
73 /// Given an exploded icmp instruction, return true if the comparison only
74 /// checks the sign bit. If it only checks the sign bit, set TrueIfSigned if the
75 /// result of the comparison is true when the input value is signed.
isSignBitCheck(ICmpInst::Predicate Pred,const APInt & RHS,bool & TrueIfSigned)76 static bool isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS,
77                            bool &TrueIfSigned) {
78   switch (Pred) {
79   case ICmpInst::ICMP_SLT:   // True if LHS s< 0
80     TrueIfSigned = true;
81     return RHS.isNullValue();
82   case ICmpInst::ICMP_SLE:   // True if LHS s<= RHS and RHS == -1
83     TrueIfSigned = true;
84     return RHS.isAllOnesValue();
85   case ICmpInst::ICMP_SGT:   // True if LHS s> -1
86     TrueIfSigned = false;
87     return RHS.isAllOnesValue();
88   case ICmpInst::ICMP_UGT:
89     // True if LHS u> RHS and RHS == high-bit-mask - 1
90     TrueIfSigned = true;
91     return RHS.isMaxSignedValue();
92   case ICmpInst::ICMP_UGE:
93     // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
94     TrueIfSigned = true;
95     return RHS.isSignMask();
96   default:
97     return false;
98   }
99 }
100 
101 /// Returns true if the exploded icmp can be expressed as a signed comparison
102 /// to zero and updates the predicate accordingly.
103 /// The signedness of the comparison is preserved.
104 /// TODO: Refactor with decomposeBitTestICmp()?
isSignTest(ICmpInst::Predicate & Pred,const APInt & C)105 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
106   if (!ICmpInst::isSigned(Pred))
107     return false;
108 
109   if (C.isNullValue())
110     return ICmpInst::isRelational(Pred);
111 
112   if (C.isOneValue()) {
113     if (Pred == ICmpInst::ICMP_SLT) {
114       Pred = ICmpInst::ICMP_SLE;
115       return true;
116     }
117   } else if (C.isAllOnesValue()) {
118     if (Pred == ICmpInst::ICMP_SGT) {
119       Pred = ICmpInst::ICMP_SGE;
120       return true;
121     }
122   }
123 
124   return false;
125 }
126 
127 /// Given a signed integer type and a set of known zero and one bits, compute
128 /// the maximum and minimum values that could have the specified known zero and
129 /// known one bits, returning them in Min/Max.
130 /// TODO: Move to method on KnownBits struct?
computeSignedMinMaxValuesFromKnownBits(const KnownBits & Known,APInt & Min,APInt & Max)131 static void computeSignedMinMaxValuesFromKnownBits(const KnownBits &Known,
132                                                    APInt &Min, APInt &Max) {
133   assert(Known.getBitWidth() == Min.getBitWidth() &&
134          Known.getBitWidth() == Max.getBitWidth() &&
135          "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
136   APInt UnknownBits = ~(Known.Zero|Known.One);
137 
138   // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
139   // bit if it is unknown.
140   Min = Known.One;
141   Max = Known.One|UnknownBits;
142 
143   if (UnknownBits.isNegative()) { // Sign bit is unknown
144     Min.setSignBit();
145     Max.clearSignBit();
146   }
147 }
148 
149 /// Given an unsigned integer type and a set of known zero and one bits, compute
150 /// the maximum and minimum values that could have the specified known zero and
151 /// known one bits, returning them in Min/Max.
152 /// TODO: Move to method on KnownBits struct?
computeUnsignedMinMaxValuesFromKnownBits(const KnownBits & Known,APInt & Min,APInt & Max)153 static void computeUnsignedMinMaxValuesFromKnownBits(const KnownBits &Known,
154                                                      APInt &Min, APInt &Max) {
155   assert(Known.getBitWidth() == Min.getBitWidth() &&
156          Known.getBitWidth() == Max.getBitWidth() &&
157          "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
158   APInt UnknownBits = ~(Known.Zero|Known.One);
159 
160   // The minimum value is when the unknown bits are all zeros.
161   Min = Known.One;
162   // The maximum value is when the unknown bits are all ones.
163   Max = Known.One|UnknownBits;
164 }
165 
166 /// This is called when we see this pattern:
167 ///   cmp pred (load (gep GV, ...)), cmpcst
168 /// where GV is a global variable with a constant initializer. Try to simplify
169 /// this into some simple computation that does not need the load. For example
170 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
171 ///
172 /// If AndCst is non-null, then the loaded value is masked with that constant
173 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
foldCmpLoadFromIndexedGlobal(GetElementPtrInst * GEP,GlobalVariable * GV,CmpInst & ICI,ConstantInt * AndCst)174 Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
175                                                         GlobalVariable *GV,
176                                                         CmpInst &ICI,
177                                                         ConstantInt *AndCst) {
178   Constant *Init = GV->getInitializer();
179   if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
180     return nullptr;
181 
182   uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
183   // Don't blow up on huge arrays.
184   if (ArrayElementCount > MaxArraySizeForCombine)
185     return nullptr;
186 
187   // There are many forms of this optimization we can handle, for now, just do
188   // the simple index into a single-dimensional array.
189   //
190   // Require: GEP GV, 0, i {{, constant indices}}
191   if (GEP->getNumOperands() < 3 ||
192       !isa<ConstantInt>(GEP->getOperand(1)) ||
193       !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
194       isa<Constant>(GEP->getOperand(2)))
195     return nullptr;
196 
197   // Check that indices after the variable are constants and in-range for the
198   // type they index.  Collect the indices.  This is typically for arrays of
199   // structs.
200   SmallVector<unsigned, 4> LaterIndices;
201 
202   Type *EltTy = Init->getType()->getArrayElementType();
203   for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
204     ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
205     if (!Idx) return nullptr;  // Variable index.
206 
207     uint64_t IdxVal = Idx->getZExtValue();
208     if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
209 
210     if (StructType *STy = dyn_cast<StructType>(EltTy))
211       EltTy = STy->getElementType(IdxVal);
212     else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
213       if (IdxVal >= ATy->getNumElements()) return nullptr;
214       EltTy = ATy->getElementType();
215     } else {
216       return nullptr; // Unknown type.
217     }
218 
219     LaterIndices.push_back(IdxVal);
220   }
221 
222   enum { Overdefined = -3, Undefined = -2 };
223 
224   // Variables for our state machines.
225 
226   // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
227   // "i == 47 | i == 87", where 47 is the first index the condition is true for,
228   // and 87 is the second (and last) index.  FirstTrueElement is -2 when
229   // undefined, otherwise set to the first true element.  SecondTrueElement is
230   // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
231   int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
232 
233   // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
234   // form "i != 47 & i != 87".  Same state transitions as for true elements.
235   int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
236 
237   /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
238   /// define a state machine that triggers for ranges of values that the index
239   /// is true or false for.  This triggers on things like "abbbbc"[i] == 'b'.
240   /// This is -2 when undefined, -3 when overdefined, and otherwise the last
241   /// index in the range (inclusive).  We use -2 for undefined here because we
242   /// use relative comparisons and don't want 0-1 to match -1.
243   int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
244 
245   // MagicBitvector - This is a magic bitvector where we set a bit if the
246   // comparison is true for element 'i'.  If there are 64 elements or less in
247   // the array, this will fully represent all the comparison results.
248   uint64_t MagicBitvector = 0;
249 
250   // Scan the array and see if one of our patterns matches.
251   Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
252   for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
253     Constant *Elt = Init->getAggregateElement(i);
254     if (!Elt) return nullptr;
255 
256     // If this is indexing an array of structures, get the structure element.
257     if (!LaterIndices.empty())
258       Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
259 
260     // If the element is masked, handle it.
261     if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
262 
263     // Find out if the comparison would be true or false for the i'th element.
264     Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
265                                                   CompareRHS, DL, &TLI);
266     // If the result is undef for this element, ignore it.
267     if (isa<UndefValue>(C)) {
268       // Extend range state machines to cover this element in case there is an
269       // undef in the middle of the range.
270       if (TrueRangeEnd == (int)i-1)
271         TrueRangeEnd = i;
272       if (FalseRangeEnd == (int)i-1)
273         FalseRangeEnd = i;
274       continue;
275     }
276 
277     // If we can't compute the result for any of the elements, we have to give
278     // up evaluating the entire conditional.
279     if (!isa<ConstantInt>(C)) return nullptr;
280 
281     // Otherwise, we know if the comparison is true or false for this element,
282     // update our state machines.
283     bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
284 
285     // State machine for single/double/range index comparison.
286     if (IsTrueForElt) {
287       // Update the TrueElement state machine.
288       if (FirstTrueElement == Undefined)
289         FirstTrueElement = TrueRangeEnd = i;  // First true element.
290       else {
291         // Update double-compare state machine.
292         if (SecondTrueElement == Undefined)
293           SecondTrueElement = i;
294         else
295           SecondTrueElement = Overdefined;
296 
297         // Update range state machine.
298         if (TrueRangeEnd == (int)i-1)
299           TrueRangeEnd = i;
300         else
301           TrueRangeEnd = Overdefined;
302       }
303     } else {
304       // Update the FalseElement state machine.
305       if (FirstFalseElement == Undefined)
306         FirstFalseElement = FalseRangeEnd = i; // First false element.
307       else {
308         // Update double-compare state machine.
309         if (SecondFalseElement == Undefined)
310           SecondFalseElement = i;
311         else
312           SecondFalseElement = Overdefined;
313 
314         // Update range state machine.
315         if (FalseRangeEnd == (int)i-1)
316           FalseRangeEnd = i;
317         else
318           FalseRangeEnd = Overdefined;
319       }
320     }
321 
322     // If this element is in range, update our magic bitvector.
323     if (i < 64 && IsTrueForElt)
324       MagicBitvector |= 1ULL << i;
325 
326     // If all of our states become overdefined, bail out early.  Since the
327     // predicate is expensive, only check it every 8 elements.  This is only
328     // really useful for really huge arrays.
329     if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
330         SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
331         FalseRangeEnd == Overdefined)
332       return nullptr;
333   }
334 
335   // Now that we've scanned the entire array, emit our new comparison(s).  We
336   // order the state machines in complexity of the generated code.
337   Value *Idx = GEP->getOperand(2);
338 
339   // If the index is larger than the pointer size of the target, truncate the
340   // index down like the GEP would do implicitly.  We don't have to do this for
341   // an inbounds GEP because the index can't be out of range.
342   if (!GEP->isInBounds()) {
343     Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
344     unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
345     if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
346       Idx = Builder.CreateTrunc(Idx, IntPtrTy);
347   }
348 
349   // If the comparison is only true for one or two elements, emit direct
350   // comparisons.
351   if (SecondTrueElement != Overdefined) {
352     // None true -> false.
353     if (FirstTrueElement == Undefined)
354       return replaceInstUsesWith(ICI, Builder.getFalse());
355 
356     Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
357 
358     // True for one element -> 'i == 47'.
359     if (SecondTrueElement == Undefined)
360       return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
361 
362     // True for two elements -> 'i == 47 | i == 72'.
363     Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
364     Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
365     Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
366     return BinaryOperator::CreateOr(C1, C2);
367   }
368 
369   // If the comparison is only false for one or two elements, emit direct
370   // comparisons.
371   if (SecondFalseElement != Overdefined) {
372     // None false -> true.
373     if (FirstFalseElement == Undefined)
374       return replaceInstUsesWith(ICI, Builder.getTrue());
375 
376     Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
377 
378     // False for one element -> 'i != 47'.
379     if (SecondFalseElement == Undefined)
380       return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
381 
382     // False for two elements -> 'i != 47 & i != 72'.
383     Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
384     Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
385     Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
386     return BinaryOperator::CreateAnd(C1, C2);
387   }
388 
389   // If the comparison can be replaced with a range comparison for the elements
390   // where it is true, emit the range check.
391   if (TrueRangeEnd != Overdefined) {
392     assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
393 
394     // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
395     if (FirstTrueElement) {
396       Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
397       Idx = Builder.CreateAdd(Idx, Offs);
398     }
399 
400     Value *End = ConstantInt::get(Idx->getType(),
401                                   TrueRangeEnd-FirstTrueElement+1);
402     return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
403   }
404 
405   // False range check.
406   if (FalseRangeEnd != Overdefined) {
407     assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
408     // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
409     if (FirstFalseElement) {
410       Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
411       Idx = Builder.CreateAdd(Idx, Offs);
412     }
413 
414     Value *End = ConstantInt::get(Idx->getType(),
415                                   FalseRangeEnd-FirstFalseElement);
416     return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
417   }
418 
419   // If a magic bitvector captures the entire comparison state
420   // of this load, replace it with computation that does:
421   //   ((magic_cst >> i) & 1) != 0
422   {
423     Type *Ty = nullptr;
424 
425     // Look for an appropriate type:
426     // - The type of Idx if the magic fits
427     // - The smallest fitting legal type
428     if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
429       Ty = Idx->getType();
430     else
431       Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
432 
433     if (Ty) {
434       Value *V = Builder.CreateIntCast(Idx, Ty, false);
435       V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
436       V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
437       return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
438     }
439   }
440 
441   return nullptr;
442 }
443 
444 /// Return a value that can be used to compare the *offset* implied by a GEP to
445 /// zero. For example, if we have &A[i], we want to return 'i' for
446 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
447 /// are involved. The above expression would also be legal to codegen as
448 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
449 /// This latter form is less amenable to optimization though, and we are allowed
450 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
451 ///
452 /// If we can't emit an optimized form for this expression, this returns null.
453 ///
evaluateGEPOffsetExpression(User * GEP,InstCombiner & IC,const DataLayout & DL)454 static Value *evaluateGEPOffsetExpression(User *GEP, InstCombiner &IC,
455                                           const DataLayout &DL) {
456   gep_type_iterator GTI = gep_type_begin(GEP);
457 
458   // Check to see if this gep only has a single variable index.  If so, and if
459   // any constant indices are a multiple of its scale, then we can compute this
460   // in terms of the scale of the variable index.  For example, if the GEP
461   // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
462   // because the expression will cross zero at the same point.
463   unsigned i, e = GEP->getNumOperands();
464   int64_t Offset = 0;
465   for (i = 1; i != e; ++i, ++GTI) {
466     if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
467       // Compute the aggregate offset of constant indices.
468       if (CI->isZero()) continue;
469 
470       // Handle a struct index, which adds its field offset to the pointer.
471       if (StructType *STy = GTI.getStructTypeOrNull()) {
472         Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
473       } else {
474         uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
475         Offset += Size*CI->getSExtValue();
476       }
477     } else {
478       // Found our variable index.
479       break;
480     }
481   }
482 
483   // If there are no variable indices, we must have a constant offset, just
484   // evaluate it the general way.
485   if (i == e) return nullptr;
486 
487   Value *VariableIdx = GEP->getOperand(i);
488   // Determine the scale factor of the variable element.  For example, this is
489   // 4 if the variable index is into an array of i32.
490   uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
491 
492   // Verify that there are no other variable indices.  If so, emit the hard way.
493   for (++i, ++GTI; i != e; ++i, ++GTI) {
494     ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
495     if (!CI) return nullptr;
496 
497     // Compute the aggregate offset of constant indices.
498     if (CI->isZero()) continue;
499 
500     // Handle a struct index, which adds its field offset to the pointer.
501     if (StructType *STy = GTI.getStructTypeOrNull()) {
502       Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
503     } else {
504       uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
505       Offset += Size*CI->getSExtValue();
506     }
507   }
508 
509   // Okay, we know we have a single variable index, which must be a
510   // pointer/array/vector index.  If there is no offset, life is simple, return
511   // the index.
512   Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
513   unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
514   if (Offset == 0) {
515     // Cast to intptrty in case a truncation occurs.  If an extension is needed,
516     // we don't need to bother extending: the extension won't affect where the
517     // computation crosses zero.
518     if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
519       VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
520     }
521     return VariableIdx;
522   }
523 
524   // Otherwise, there is an index.  The computation we will do will be modulo
525   // the pointer size, so get it.
526   uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
527 
528   Offset &= PtrSizeMask;
529   VariableScale &= PtrSizeMask;
530 
531   // To do this transformation, any constant index must be a multiple of the
532   // variable scale factor.  For example, we can evaluate "12 + 4*i" as "3 + i",
533   // but we can't evaluate "10 + 3*i" in terms of i.  Check that the offset is a
534   // multiple of the variable scale.
535   int64_t NewOffs = Offset / (int64_t)VariableScale;
536   if (Offset != NewOffs*(int64_t)VariableScale)
537     return nullptr;
538 
539   // Okay, we can do this evaluation.  Start by converting the index to intptr.
540   if (VariableIdx->getType() != IntPtrTy)
541     VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
542                                             true /*Signed*/);
543   Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
544   return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
545 }
546 
547 /// Returns true if we can rewrite Start as a GEP with pointer Base
548 /// and some integer offset. The nodes that need to be re-written
549 /// for this transformation will be added to Explored.
canRewriteGEPAsOffset(Value * Start,Value * Base,const DataLayout & DL,SetVector<Value * > & Explored)550 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
551                                   const DataLayout &DL,
552                                   SetVector<Value *> &Explored) {
553   SmallVector<Value *, 16> WorkList(1, Start);
554   Explored.insert(Base);
555 
556   // The following traversal gives us an order which can be used
557   // when doing the final transformation. Since in the final
558   // transformation we create the PHI replacement instructions first,
559   // we don't have to get them in any particular order.
560   //
561   // However, for other instructions we will have to traverse the
562   // operands of an instruction first, which means that we have to
563   // do a post-order traversal.
564   while (!WorkList.empty()) {
565     SetVector<PHINode *> PHIs;
566 
567     while (!WorkList.empty()) {
568       if (Explored.size() >= 100)
569         return false;
570 
571       Value *V = WorkList.back();
572 
573       if (Explored.count(V) != 0) {
574         WorkList.pop_back();
575         continue;
576       }
577 
578       if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
579           !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
580         // We've found some value that we can't explore which is different from
581         // the base. Therefore we can't do this transformation.
582         return false;
583 
584       if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
585         auto *CI = dyn_cast<CastInst>(V);
586         if (!CI->isNoopCast(DL))
587           return false;
588 
589         if (Explored.count(CI->getOperand(0)) == 0)
590           WorkList.push_back(CI->getOperand(0));
591       }
592 
593       if (auto *GEP = dyn_cast<GEPOperator>(V)) {
594         // We're limiting the GEP to having one index. This will preserve
595         // the original pointer type. We could handle more cases in the
596         // future.
597         if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
598             GEP->getType() != Start->getType())
599           return false;
600 
601         if (Explored.count(GEP->getOperand(0)) == 0)
602           WorkList.push_back(GEP->getOperand(0));
603       }
604 
605       if (WorkList.back() == V) {
606         WorkList.pop_back();
607         // We've finished visiting this node, mark it as such.
608         Explored.insert(V);
609       }
610 
611       if (auto *PN = dyn_cast<PHINode>(V)) {
612         // We cannot transform PHIs on unsplittable basic blocks.
613         if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
614           return false;
615         Explored.insert(PN);
616         PHIs.insert(PN);
617       }
618     }
619 
620     // Explore the PHI nodes further.
621     for (auto *PN : PHIs)
622       for (Value *Op : PN->incoming_values())
623         if (Explored.count(Op) == 0)
624           WorkList.push_back(Op);
625   }
626 
627   // Make sure that we can do this. Since we can't insert GEPs in a basic
628   // block before a PHI node, we can't easily do this transformation if
629   // we have PHI node users of transformed instructions.
630   for (Value *Val : Explored) {
631     for (Value *Use : Val->uses()) {
632 
633       auto *PHI = dyn_cast<PHINode>(Use);
634       auto *Inst = dyn_cast<Instruction>(Val);
635 
636       if (Inst == Base || Inst == PHI || !Inst || !PHI ||
637           Explored.count(PHI) == 0)
638         continue;
639 
640       if (PHI->getParent() == Inst->getParent())
641         return false;
642     }
643   }
644   return true;
645 }
646 
647 // Sets the appropriate insert point on Builder where we can add
648 // a replacement Instruction for V (if that is possible).
setInsertionPoint(IRBuilder<> & Builder,Value * V,bool Before=true)649 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
650                               bool Before = true) {
651   if (auto *PHI = dyn_cast<PHINode>(V)) {
652     Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
653     return;
654   }
655   if (auto *I = dyn_cast<Instruction>(V)) {
656     if (!Before)
657       I = &*std::next(I->getIterator());
658     Builder.SetInsertPoint(I);
659     return;
660   }
661   if (auto *A = dyn_cast<Argument>(V)) {
662     // Set the insertion point in the entry block.
663     BasicBlock &Entry = A->getParent()->getEntryBlock();
664     Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
665     return;
666   }
667   // Otherwise, this is a constant and we don't need to set a new
668   // insertion point.
669   assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
670 }
671 
672 /// Returns a re-written value of Start as an indexed GEP using Base as a
673 /// pointer.
rewriteGEPAsOffset(Value * Start,Value * Base,const DataLayout & DL,SetVector<Value * > & Explored)674 static Value *rewriteGEPAsOffset(Value *Start, Value *Base,
675                                  const DataLayout &DL,
676                                  SetVector<Value *> &Explored) {
677   // Perform all the substitutions. This is a bit tricky because we can
678   // have cycles in our use-def chains.
679   // 1. Create the PHI nodes without any incoming values.
680   // 2. Create all the other values.
681   // 3. Add the edges for the PHI nodes.
682   // 4. Emit GEPs to get the original pointers.
683   // 5. Remove the original instructions.
684   Type *IndexType = IntegerType::get(
685       Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
686 
687   DenseMap<Value *, Value *> NewInsts;
688   NewInsts[Base] = ConstantInt::getNullValue(IndexType);
689 
690   // Create the new PHI nodes, without adding any incoming values.
691   for (Value *Val : Explored) {
692     if (Val == Base)
693       continue;
694     // Create empty phi nodes. This avoids cyclic dependencies when creating
695     // the remaining instructions.
696     if (auto *PHI = dyn_cast<PHINode>(Val))
697       NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
698                                       PHI->getName() + ".idx", PHI);
699   }
700   IRBuilder<> Builder(Base->getContext());
701 
702   // Create all the other instructions.
703   for (Value *Val : Explored) {
704 
705     if (NewInsts.find(Val) != NewInsts.end())
706       continue;
707 
708     if (auto *CI = dyn_cast<CastInst>(Val)) {
709       NewInsts[CI] = NewInsts[CI->getOperand(0)];
710       continue;
711     }
712     if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
713       Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
714                                                   : GEP->getOperand(1);
715       setInsertionPoint(Builder, GEP);
716       // Indices might need to be sign extended. GEPs will magically do
717       // this, but we need to do it ourselves here.
718       if (Index->getType()->getScalarSizeInBits() !=
719           NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
720         Index = Builder.CreateSExtOrTrunc(
721             Index, NewInsts[GEP->getOperand(0)]->getType(),
722             GEP->getOperand(0)->getName() + ".sext");
723       }
724 
725       auto *Op = NewInsts[GEP->getOperand(0)];
726       if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
727         NewInsts[GEP] = Index;
728       else
729         NewInsts[GEP] = Builder.CreateNSWAdd(
730             Op, Index, GEP->getOperand(0)->getName() + ".add");
731       continue;
732     }
733     if (isa<PHINode>(Val))
734       continue;
735 
736     llvm_unreachable("Unexpected instruction type");
737   }
738 
739   // Add the incoming values to the PHI nodes.
740   for (Value *Val : Explored) {
741     if (Val == Base)
742       continue;
743     // All the instructions have been created, we can now add edges to the
744     // phi nodes.
745     if (auto *PHI = dyn_cast<PHINode>(Val)) {
746       PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
747       for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
748         Value *NewIncoming = PHI->getIncomingValue(I);
749 
750         if (NewInsts.find(NewIncoming) != NewInsts.end())
751           NewIncoming = NewInsts[NewIncoming];
752 
753         NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
754       }
755     }
756   }
757 
758   for (Value *Val : Explored) {
759     if (Val == Base)
760       continue;
761 
762     // Depending on the type, for external users we have to emit
763     // a GEP or a GEP + ptrtoint.
764     setInsertionPoint(Builder, Val, false);
765 
766     // If required, create an inttoptr instruction for Base.
767     Value *NewBase = Base;
768     if (!Base->getType()->isPointerTy())
769       NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
770                                                Start->getName() + "to.ptr");
771 
772     Value *GEP = Builder.CreateInBoundsGEP(
773         Start->getType()->getPointerElementType(), NewBase,
774         makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
775 
776     if (!Val->getType()->isPointerTy()) {
777       Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
778                                               Val->getName() + ".conv");
779       GEP = Cast;
780     }
781     Val->replaceAllUsesWith(GEP);
782   }
783 
784   return NewInsts[Start];
785 }
786 
787 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
788 /// the input Value as a constant indexed GEP. Returns a pair containing
789 /// the GEPs Pointer and Index.
790 static std::pair<Value *, Value *>
getAsConstantIndexedAddress(Value * V,const DataLayout & DL)791 getAsConstantIndexedAddress(Value *V, const DataLayout &DL) {
792   Type *IndexType = IntegerType::get(V->getContext(),
793                                      DL.getIndexTypeSizeInBits(V->getType()));
794 
795   Constant *Index = ConstantInt::getNullValue(IndexType);
796   while (true) {
797     if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
798       // We accept only inbouds GEPs here to exclude the possibility of
799       // overflow.
800       if (!GEP->isInBounds())
801         break;
802       if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
803           GEP->getType() == V->getType()) {
804         V = GEP->getOperand(0);
805         Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
806         Index = ConstantExpr::getAdd(
807             Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
808         continue;
809       }
810       break;
811     }
812     if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
813       if (!CI->isNoopCast(DL))
814         break;
815       V = CI->getOperand(0);
816       continue;
817     }
818     if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
819       if (!CI->isNoopCast(DL))
820         break;
821       V = CI->getOperand(0);
822       continue;
823     }
824     break;
825   }
826   return {V, Index};
827 }
828 
829 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
830 /// We can look through PHIs, GEPs and casts in order to determine a common base
831 /// between GEPLHS and RHS.
transformToIndexedCompare(GEPOperator * GEPLHS,Value * RHS,ICmpInst::Predicate Cond,const DataLayout & DL)832 static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
833                                               ICmpInst::Predicate Cond,
834                                               const DataLayout &DL) {
835   if (!GEPLHS->hasAllConstantIndices())
836     return nullptr;
837 
838   // Make sure the pointers have the same type.
839   if (GEPLHS->getType() != RHS->getType())
840     return nullptr;
841 
842   Value *PtrBase, *Index;
843   std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
844 
845   // The set of nodes that will take part in this transformation.
846   SetVector<Value *> Nodes;
847 
848   if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
849     return nullptr;
850 
851   // We know we can re-write this as
852   //  ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
853   // Since we've only looked through inbouds GEPs we know that we
854   // can't have overflow on either side. We can therefore re-write
855   // this as:
856   //   OFFSET1 cmp OFFSET2
857   Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
858 
859   // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
860   // GEP having PtrBase as the pointer base, and has returned in NewRHS the
861   // offset. Since Index is the offset of LHS to the base pointer, we will now
862   // compare the offsets instead of comparing the pointers.
863   return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
864 }
865 
866 /// Fold comparisons between a GEP instruction and something else. At this point
867 /// we know that the GEP is on the LHS of the comparison.
foldGEPICmp(GEPOperator * GEPLHS,Value * RHS,ICmpInst::Predicate Cond,Instruction & I)868 Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
869                                        ICmpInst::Predicate Cond,
870                                        Instruction &I) {
871   // Don't transform signed compares of GEPs into index compares. Even if the
872   // GEP is inbounds, the final add of the base pointer can have signed overflow
873   // and would change the result of the icmp.
874   // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
875   // the maximum signed value for the pointer type.
876   if (ICmpInst::isSigned(Cond))
877     return nullptr;
878 
879   // Look through bitcasts and addrspacecasts. We do not however want to remove
880   // 0 GEPs.
881   if (!isa<GetElementPtrInst>(RHS))
882     RHS = RHS->stripPointerCasts();
883 
884   Value *PtrBase = GEPLHS->getOperand(0);
885   if (PtrBase == RHS && GEPLHS->isInBounds()) {
886     // ((gep Ptr, OFFSET) cmp Ptr)   ---> (OFFSET cmp 0).
887     // This transformation (ignoring the base and scales) is valid because we
888     // know pointers can't overflow since the gep is inbounds.  See if we can
889     // output an optimized form.
890     Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
891 
892     // If not, synthesize the offset the hard way.
893     if (!Offset)
894       Offset = EmitGEPOffset(GEPLHS);
895     return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
896                         Constant::getNullValue(Offset->getType()));
897   } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
898     // If the base pointers are different, but the indices are the same, just
899     // compare the base pointer.
900     if (PtrBase != GEPRHS->getOperand(0)) {
901       bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
902       IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
903                         GEPRHS->getOperand(0)->getType();
904       if (IndicesTheSame)
905         for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
906           if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
907             IndicesTheSame = false;
908             break;
909           }
910 
911       // If all indices are the same, just compare the base pointers.
912       if (IndicesTheSame)
913         return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
914 
915       // If we're comparing GEPs with two base pointers that only differ in type
916       // and both GEPs have only constant indices or just one use, then fold
917       // the compare with the adjusted indices.
918       if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
919           (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
920           (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
921           PtrBase->stripPointerCasts() ==
922               GEPRHS->getOperand(0)->stripPointerCasts()) {
923         Value *LOffset = EmitGEPOffset(GEPLHS);
924         Value *ROffset = EmitGEPOffset(GEPRHS);
925 
926         // If we looked through an addrspacecast between different sized address
927         // spaces, the LHS and RHS pointers are different sized
928         // integers. Truncate to the smaller one.
929         Type *LHSIndexTy = LOffset->getType();
930         Type *RHSIndexTy = ROffset->getType();
931         if (LHSIndexTy != RHSIndexTy) {
932           if (LHSIndexTy->getPrimitiveSizeInBits() <
933               RHSIndexTy->getPrimitiveSizeInBits()) {
934             ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
935           } else
936             LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
937         }
938 
939         Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
940                                         LOffset, ROffset);
941         return replaceInstUsesWith(I, Cmp);
942       }
943 
944       // Otherwise, the base pointers are different and the indices are
945       // different. Try convert this to an indexed compare by looking through
946       // PHIs/casts.
947       return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
948     }
949 
950     // If one of the GEPs has all zero indices, recurse.
951     if (GEPLHS->hasAllZeroIndices())
952       return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
953                          ICmpInst::getSwappedPredicate(Cond), I);
954 
955     // If the other GEP has all zero indices, recurse.
956     if (GEPRHS->hasAllZeroIndices())
957       return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
958 
959     bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
960     if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
961       // If the GEPs only differ by one index, compare it.
962       unsigned NumDifferences = 0;  // Keep track of # differences.
963       unsigned DiffOperand = 0;     // The operand that differs.
964       for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
965         if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
966           if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
967                    GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
968             // Irreconcilable differences.
969             NumDifferences = 2;
970             break;
971           } else {
972             if (NumDifferences++) break;
973             DiffOperand = i;
974           }
975         }
976 
977       if (NumDifferences == 0)   // SAME GEP?
978         return replaceInstUsesWith(I, // No comparison is needed here.
979                              Builder.getInt1(ICmpInst::isTrueWhenEqual(Cond)));
980 
981       else if (NumDifferences == 1 && GEPsInBounds) {
982         Value *LHSV = GEPLHS->getOperand(DiffOperand);
983         Value *RHSV = GEPRHS->getOperand(DiffOperand);
984         // Make sure we do a signed comparison here.
985         return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
986       }
987     }
988 
989     // Only lower this if the icmp is the only user of the GEP or if we expect
990     // the result to fold to a constant!
991     if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
992         (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
993       // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)  --->  (OFFSET1 cmp OFFSET2)
994       Value *L = EmitGEPOffset(GEPLHS);
995       Value *R = EmitGEPOffset(GEPRHS);
996       return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
997     }
998   }
999 
1000   // Try convert this to an indexed compare by looking through PHIs/casts as a
1001   // last resort.
1002   return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1003 }
1004 
foldAllocaCmp(ICmpInst & ICI,const AllocaInst * Alloca,const Value * Other)1005 Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI,
1006                                          const AllocaInst *Alloca,
1007                                          const Value *Other) {
1008   assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1009 
1010   // It would be tempting to fold away comparisons between allocas and any
1011   // pointer not based on that alloca (e.g. an argument). However, even
1012   // though such pointers cannot alias, they can still compare equal.
1013   //
1014   // But LLVM doesn't specify where allocas get their memory, so if the alloca
1015   // doesn't escape we can argue that it's impossible to guess its value, and we
1016   // can therefore act as if any such guesses are wrong.
1017   //
1018   // The code below checks that the alloca doesn't escape, and that it's only
1019   // used in a comparison once (the current instruction). The
1020   // single-comparison-use condition ensures that we're trivially folding all
1021   // comparisons against the alloca consistently, and avoids the risk of
1022   // erroneously folding a comparison of the pointer with itself.
1023 
1024   unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1025 
1026   SmallVector<const Use *, 32> Worklist;
1027   for (const Use &U : Alloca->uses()) {
1028     if (Worklist.size() >= MaxIter)
1029       return nullptr;
1030     Worklist.push_back(&U);
1031   }
1032 
1033   unsigned NumCmps = 0;
1034   while (!Worklist.empty()) {
1035     assert(Worklist.size() <= MaxIter);
1036     const Use *U = Worklist.pop_back_val();
1037     const Value *V = U->getUser();
1038     --MaxIter;
1039 
1040     if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1041         isa<SelectInst>(V)) {
1042       // Track the uses.
1043     } else if (isa<LoadInst>(V)) {
1044       // Loading from the pointer doesn't escape it.
1045       continue;
1046     } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1047       // Storing *to* the pointer is fine, but storing the pointer escapes it.
1048       if (SI->getValueOperand() == U->get())
1049         return nullptr;
1050       continue;
1051     } else if (isa<ICmpInst>(V)) {
1052       if (NumCmps++)
1053         return nullptr; // Found more than one cmp.
1054       continue;
1055     } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1056       switch (Intrin->getIntrinsicID()) {
1057         // These intrinsics don't escape or compare the pointer. Memset is safe
1058         // because we don't allow ptrtoint. Memcpy and memmove are safe because
1059         // we don't allow stores, so src cannot point to V.
1060         case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1061         case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1062           continue;
1063         default:
1064           return nullptr;
1065       }
1066     } else {
1067       return nullptr;
1068     }
1069     for (const Use &U : V->uses()) {
1070       if (Worklist.size() >= MaxIter)
1071         return nullptr;
1072       Worklist.push_back(&U);
1073     }
1074   }
1075 
1076   Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1077   return replaceInstUsesWith(
1078       ICI,
1079       ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate())));
1080 }
1081 
1082 /// Fold "icmp pred (X+CI), X".
foldICmpAddOpConst(Value * X,ConstantInt * CI,ICmpInst::Predicate Pred)1083 Instruction *InstCombiner::foldICmpAddOpConst(Value *X, ConstantInt *CI,
1084                                               ICmpInst::Predicate Pred) {
1085   // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1086   // so the values can never be equal.  Similarly for all other "or equals"
1087   // operators.
1088 
1089   // (X+1) <u X        --> X >u (MAXUINT-1)        --> X == 255
1090   // (X+2) <u X        --> X >u (MAXUINT-2)        --> X > 253
1091   // (X+MAXUINT) <u X  --> X >u (MAXUINT-MAXUINT)  --> X != 0
1092   if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1093     Value *R =
1094       ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
1095     return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1096   }
1097 
1098   // (X+1) >u X        --> X <u (0-1)        --> X != 255
1099   // (X+2) >u X        --> X <u (0-2)        --> X <u 254
1100   // (X+MAXUINT) >u X  --> X <u (0-MAXUINT)  --> X <u 1  --> X == 0
1101   if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1102     return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
1103 
1104   unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
1105   ConstantInt *SMax = ConstantInt::get(X->getContext(),
1106                                        APInt::getSignedMaxValue(BitWidth));
1107 
1108   // (X+ 1) <s X       --> X >s (MAXSINT-1)          --> X == 127
1109   // (X+ 2) <s X       --> X >s (MAXSINT-2)          --> X >s 125
1110   // (X+MAXSINT) <s X  --> X >s (MAXSINT-MAXSINT)    --> X >s 0
1111   // (X+MINSINT) <s X  --> X >s (MAXSINT-MINSINT)    --> X >s -1
1112   // (X+ -2) <s X      --> X >s (MAXSINT- -2)        --> X >s 126
1113   // (X+ -1) <s X      --> X >s (MAXSINT- -1)        --> X != 127
1114   if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1115     return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
1116 
1117   // (X+ 1) >s X       --> X <s (MAXSINT-(1-1))       --> X != 127
1118   // (X+ 2) >s X       --> X <s (MAXSINT-(2-1))       --> X <s 126
1119   // (X+MAXSINT) >s X  --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1120   // (X+MINSINT) >s X  --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1121   // (X+ -2) >s X      --> X <s (MAXSINT-(-2-1))      --> X <s -126
1122   // (X+ -1) >s X      --> X <s (MAXSINT-(-1-1))      --> X == -128
1123 
1124   assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1125   Constant *C = Builder.getInt(CI->getValue() - 1);
1126   return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
1127 }
1128 
1129 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1130 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1131 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
foldICmpShrConstConst(ICmpInst & I,Value * A,const APInt & AP1,const APInt & AP2)1132 Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A,
1133                                                  const APInt &AP1,
1134                                                  const APInt &AP2) {
1135   assert(I.isEquality() && "Cannot fold icmp gt/lt");
1136 
1137   auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1138     if (I.getPredicate() == I.ICMP_NE)
1139       Pred = CmpInst::getInversePredicate(Pred);
1140     return new ICmpInst(Pred, LHS, RHS);
1141   };
1142 
1143   // Don't bother doing any work for cases which InstSimplify handles.
1144   if (AP2.isNullValue())
1145     return nullptr;
1146 
1147   bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1148   if (IsAShr) {
1149     if (AP2.isAllOnesValue())
1150       return nullptr;
1151     if (AP2.isNegative() != AP1.isNegative())
1152       return nullptr;
1153     if (AP2.sgt(AP1))
1154       return nullptr;
1155   }
1156 
1157   if (!AP1)
1158     // 'A' must be large enough to shift out the highest set bit.
1159     return getICmp(I.ICMP_UGT, A,
1160                    ConstantInt::get(A->getType(), AP2.logBase2()));
1161 
1162   if (AP1 == AP2)
1163     return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1164 
1165   int Shift;
1166   if (IsAShr && AP1.isNegative())
1167     Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1168   else
1169     Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1170 
1171   if (Shift > 0) {
1172     if (IsAShr && AP1 == AP2.ashr(Shift)) {
1173       // There are multiple solutions if we are comparing against -1 and the LHS
1174       // of the ashr is not a power of two.
1175       if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1176         return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1177       return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1178     } else if (AP1 == AP2.lshr(Shift)) {
1179       return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1180     }
1181   }
1182 
1183   // Shifting const2 will never be equal to const1.
1184   // FIXME: This should always be handled by InstSimplify?
1185   auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1186   return replaceInstUsesWith(I, TorF);
1187 }
1188 
1189 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1190 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
foldICmpShlConstConst(ICmpInst & I,Value * A,const APInt & AP1,const APInt & AP2)1191 Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A,
1192                                                  const APInt &AP1,
1193                                                  const APInt &AP2) {
1194   assert(I.isEquality() && "Cannot fold icmp gt/lt");
1195 
1196   auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1197     if (I.getPredicate() == I.ICMP_NE)
1198       Pred = CmpInst::getInversePredicate(Pred);
1199     return new ICmpInst(Pred, LHS, RHS);
1200   };
1201 
1202   // Don't bother doing any work for cases which InstSimplify handles.
1203   if (AP2.isNullValue())
1204     return nullptr;
1205 
1206   unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1207 
1208   if (!AP1 && AP2TrailingZeros != 0)
1209     return getICmp(
1210         I.ICMP_UGE, A,
1211         ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1212 
1213   if (AP1 == AP2)
1214     return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1215 
1216   // Get the distance between the lowest bits that are set.
1217   int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1218 
1219   if (Shift > 0 && AP2.shl(Shift) == AP1)
1220     return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1221 
1222   // Shifting const2 will never be equal to const1.
1223   // FIXME: This should always be handled by InstSimplify?
1224   auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1225   return replaceInstUsesWith(I, TorF);
1226 }
1227 
1228 /// The caller has matched a pattern of the form:
1229 ///   I = icmp ugt (add (add A, B), CI2), CI1
1230 /// If this is of the form:
1231 ///   sum = a + b
1232 ///   if (sum+128 >u 255)
1233 /// Then replace it with llvm.sadd.with.overflow.i8.
1234 ///
processUGT_ADDCST_ADD(ICmpInst & I,Value * A,Value * B,ConstantInt * CI2,ConstantInt * CI1,InstCombiner & IC)1235 static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1236                                           ConstantInt *CI2, ConstantInt *CI1,
1237                                           InstCombiner &IC) {
1238   // The transformation we're trying to do here is to transform this into an
1239   // llvm.sadd.with.overflow.  To do this, we have to replace the original add
1240   // with a narrower add, and discard the add-with-constant that is part of the
1241   // range check (if we can't eliminate it, this isn't profitable).
1242 
1243   // In order to eliminate the add-with-constant, the compare can be its only
1244   // use.
1245   Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1246   if (!AddWithCst->hasOneUse())
1247     return nullptr;
1248 
1249   // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1250   if (!CI2->getValue().isPowerOf2())
1251     return nullptr;
1252   unsigned NewWidth = CI2->getValue().countTrailingZeros();
1253   if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1254     return nullptr;
1255 
1256   // The width of the new add formed is 1 more than the bias.
1257   ++NewWidth;
1258 
1259   // Check to see that CI1 is an all-ones value with NewWidth bits.
1260   if (CI1->getBitWidth() == NewWidth ||
1261       CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1262     return nullptr;
1263 
1264   // This is only really a signed overflow check if the inputs have been
1265   // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1266   // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1267   unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1268   if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1269       IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1270     return nullptr;
1271 
1272   // In order to replace the original add with a narrower
1273   // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1274   // and truncates that discard the high bits of the add.  Verify that this is
1275   // the case.
1276   Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1277   for (User *U : OrigAdd->users()) {
1278     if (U == AddWithCst)
1279       continue;
1280 
1281     // Only accept truncates for now.  We would really like a nice recursive
1282     // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1283     // chain to see which bits of a value are actually demanded.  If the
1284     // original add had another add which was then immediately truncated, we
1285     // could still do the transformation.
1286     TruncInst *TI = dyn_cast<TruncInst>(U);
1287     if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1288       return nullptr;
1289   }
1290 
1291   // If the pattern matches, truncate the inputs to the narrower type and
1292   // use the sadd_with_overflow intrinsic to efficiently compute both the
1293   // result and the overflow bit.
1294   Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1295   Value *F = Intrinsic::getDeclaration(I.getModule(),
1296                                        Intrinsic::sadd_with_overflow, NewType);
1297 
1298   InstCombiner::BuilderTy &Builder = IC.Builder;
1299 
1300   // Put the new code above the original add, in case there are any uses of the
1301   // add between the add and the compare.
1302   Builder.SetInsertPoint(OrigAdd);
1303 
1304   Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1305   Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1306   CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1307   Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1308   Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1309 
1310   // The inner add was the result of the narrow add, zero extended to the
1311   // wider type.  Replace it with the result computed by the intrinsic.
1312   IC.replaceInstUsesWith(*OrigAdd, ZExt);
1313 
1314   // The original icmp gets replaced with the overflow value.
1315   return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1316 }
1317 
1318 // Handle (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
foldICmpWithZero(ICmpInst & Cmp)1319 Instruction *InstCombiner::foldICmpWithZero(ICmpInst &Cmp) {
1320   CmpInst::Predicate Pred = Cmp.getPredicate();
1321   Value *X = Cmp.getOperand(0);
1322 
1323   if (match(Cmp.getOperand(1), m_Zero()) && Pred == ICmpInst::ICMP_SGT) {
1324     Value *A, *B;
1325     SelectPatternResult SPR = matchSelectPattern(X, A, B);
1326     if (SPR.Flavor == SPF_SMIN) {
1327       if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1328         return new ICmpInst(Pred, B, Cmp.getOperand(1));
1329       if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1330         return new ICmpInst(Pred, A, Cmp.getOperand(1));
1331     }
1332   }
1333   return nullptr;
1334 }
1335 
1336 // Fold icmp Pred X, C.
foldICmpWithConstant(ICmpInst & Cmp)1337 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
1338   CmpInst::Predicate Pred = Cmp.getPredicate();
1339   Value *X = Cmp.getOperand(0);
1340 
1341   const APInt *C;
1342   if (!match(Cmp.getOperand(1), m_APInt(C)))
1343     return nullptr;
1344 
1345   Value *A = nullptr, *B = nullptr;
1346 
1347   // Match the following pattern, which is a common idiom when writing
1348   // overflow-safe integer arithmetic functions. The source performs an addition
1349   // in wider type and explicitly checks for overflow using comparisons against
1350   // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1351   //
1352   // TODO: This could probably be generalized to handle other overflow-safe
1353   // operations if we worked out the formulas to compute the appropriate magic
1354   // constants.
1355   //
1356   // sum = a + b
1357   // if (sum+128 >u 255)  ...  -> llvm.sadd.with.overflow.i8
1358   {
1359     ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1360     if (Pred == ICmpInst::ICMP_UGT &&
1361         match(X, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1362       if (Instruction *Res = processUGT_ADDCST_ADD(
1363               Cmp, A, B, CI2, cast<ConstantInt>(Cmp.getOperand(1)), *this))
1364         return Res;
1365   }
1366 
1367   // FIXME: Use m_APInt to allow folds for splat constants.
1368   ConstantInt *CI = dyn_cast<ConstantInt>(Cmp.getOperand(1));
1369   if (!CI)
1370     return nullptr;
1371 
1372   // Canonicalize icmp instructions based on dominating conditions.
1373   BasicBlock *Parent = Cmp.getParent();
1374   BasicBlock *Dom = Parent->getSinglePredecessor();
1375   auto *BI = Dom ? dyn_cast<BranchInst>(Dom->getTerminator()) : nullptr;
1376   ICmpInst::Predicate Pred2;
1377   BasicBlock *TrueBB, *FalseBB;
1378   ConstantInt *CI2;
1379   if (BI && match(BI, m_Br(m_ICmp(Pred2, m_Specific(X), m_ConstantInt(CI2)),
1380                            TrueBB, FalseBB)) &&
1381       TrueBB != FalseBB) {
1382     ConstantRange CR =
1383         ConstantRange::makeAllowedICmpRegion(Pred, CI->getValue());
1384     ConstantRange DominatingCR =
1385         (Parent == TrueBB)
1386             ? ConstantRange::makeExactICmpRegion(Pred2, CI2->getValue())
1387             : ConstantRange::makeExactICmpRegion(
1388                   CmpInst::getInversePredicate(Pred2), CI2->getValue());
1389     ConstantRange Intersection = DominatingCR.intersectWith(CR);
1390     ConstantRange Difference = DominatingCR.difference(CR);
1391     if (Intersection.isEmptySet())
1392       return replaceInstUsesWith(Cmp, Builder.getFalse());
1393     if (Difference.isEmptySet())
1394       return replaceInstUsesWith(Cmp, Builder.getTrue());
1395 
1396     // If this is a normal comparison, it demands all bits. If it is a sign
1397     // bit comparison, it only demands the sign bit.
1398     bool UnusedBit;
1399     bool IsSignBit = isSignBitCheck(Pred, CI->getValue(), UnusedBit);
1400 
1401     // Canonicalizing a sign bit comparison that gets used in a branch,
1402     // pessimizes codegen by generating branch on zero instruction instead
1403     // of a test and branch. So we avoid canonicalizing in such situations
1404     // because test and branch instruction has better branch displacement
1405     // than compare and branch instruction.
1406     if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1407       return nullptr;
1408 
1409     if (auto *AI = Intersection.getSingleElement())
1410       return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*AI));
1411     if (auto *AD = Difference.getSingleElement())
1412       return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*AD));
1413   }
1414 
1415   return nullptr;
1416 }
1417 
1418 /// Fold icmp (trunc X, Y), C.
foldICmpTruncConstant(ICmpInst & Cmp,TruncInst * Trunc,const APInt & C)1419 Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
1420                                                  TruncInst *Trunc,
1421                                                  const APInt &C) {
1422   ICmpInst::Predicate Pred = Cmp.getPredicate();
1423   Value *X = Trunc->getOperand(0);
1424   if (C.isOneValue() && C.getBitWidth() > 1) {
1425     // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1426     Value *V = nullptr;
1427     if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1428       return new ICmpInst(ICmpInst::ICMP_SLT, V,
1429                           ConstantInt::get(V->getType(), 1));
1430   }
1431 
1432   if (Cmp.isEquality() && Trunc->hasOneUse()) {
1433     // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1434     // of the high bits truncated out of x are known.
1435     unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1436              SrcBits = X->getType()->getScalarSizeInBits();
1437     KnownBits Known = computeKnownBits(X, 0, &Cmp);
1438 
1439     // If all the high bits are known, we can do this xform.
1440     if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1441       // Pull in the high bits from known-ones set.
1442       APInt NewRHS = C.zext(SrcBits);
1443       NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1444       return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1445     }
1446   }
1447 
1448   return nullptr;
1449 }
1450 
1451 /// Fold icmp (xor X, Y), C.
foldICmpXorConstant(ICmpInst & Cmp,BinaryOperator * Xor,const APInt & C)1452 Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp,
1453                                                BinaryOperator *Xor,
1454                                                const APInt &C) {
1455   Value *X = Xor->getOperand(0);
1456   Value *Y = Xor->getOperand(1);
1457   const APInt *XorC;
1458   if (!match(Y, m_APInt(XorC)))
1459     return nullptr;
1460 
1461   // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1462   // fold the xor.
1463   ICmpInst::Predicate Pred = Cmp.getPredicate();
1464   bool TrueIfSigned = false;
1465   if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1466 
1467     // If the sign bit of the XorCst is not set, there is no change to
1468     // the operation, just stop using the Xor.
1469     if (!XorC->isNegative()) {
1470       Cmp.setOperand(0, X);
1471       Worklist.Add(Xor);
1472       return &Cmp;
1473     }
1474 
1475     // Emit the opposite comparison.
1476     if (TrueIfSigned)
1477       return new ICmpInst(ICmpInst::ICMP_SGT, X,
1478                           ConstantInt::getAllOnesValue(X->getType()));
1479     else
1480       return new ICmpInst(ICmpInst::ICMP_SLT, X,
1481                           ConstantInt::getNullValue(X->getType()));
1482   }
1483 
1484   if (Xor->hasOneUse()) {
1485     // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1486     if (!Cmp.isEquality() && XorC->isSignMask()) {
1487       Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1488                             : Cmp.getSignedPredicate();
1489       return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1490     }
1491 
1492     // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1493     if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1494       Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1495                             : Cmp.getSignedPredicate();
1496       Pred = Cmp.getSwappedPredicate(Pred);
1497       return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1498     }
1499   }
1500 
1501   // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1502   //   iff -C is a power of 2
1503   if (Pred == ICmpInst::ICMP_UGT && *XorC == ~C && (C + 1).isPowerOf2())
1504     return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1505 
1506   // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1507   //   iff -C is a power of 2
1508   if (Pred == ICmpInst::ICMP_ULT && *XorC == -C && C.isPowerOf2())
1509     return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
1510 
1511   return nullptr;
1512 }
1513 
1514 /// Fold icmp (and (sh X, Y), C2), C1.
foldICmpAndShift(ICmpInst & Cmp,BinaryOperator * And,const APInt & C1,const APInt & C2)1515 Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
1516                                             const APInt &C1, const APInt &C2) {
1517   BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1518   if (!Shift || !Shift->isShift())
1519     return nullptr;
1520 
1521   // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1522   // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1523   // code produced by the clang front-end, for bitfield access.
1524   // This seemingly simple opportunity to fold away a shift turns out to be
1525   // rather complicated. See PR17827 for details.
1526   unsigned ShiftOpcode = Shift->getOpcode();
1527   bool IsShl = ShiftOpcode == Instruction::Shl;
1528   const APInt *C3;
1529   if (match(Shift->getOperand(1), m_APInt(C3))) {
1530     bool CanFold = false;
1531     if (ShiftOpcode == Instruction::Shl) {
1532       // For a left shift, we can fold if the comparison is not signed. We can
1533       // also fold a signed comparison if the mask value and comparison value
1534       // are not negative. These constraints may not be obvious, but we can
1535       // prove that they are correct using an SMT solver.
1536       if (!Cmp.isSigned() || (!C2.isNegative() && !C1.isNegative()))
1537         CanFold = true;
1538     } else {
1539       bool IsAshr = ShiftOpcode == Instruction::AShr;
1540       // For a logical right shift, we can fold if the comparison is not signed.
1541       // We can also fold a signed comparison if the shifted mask value and the
1542       // shifted comparison value are not negative. These constraints may not be
1543       // obvious, but we can prove that they are correct using an SMT solver.
1544       // For an arithmetic shift right we can do the same, if we ensure
1545       // the And doesn't use any bits being shifted in. Normally these would
1546       // be turned into lshr by SimplifyDemandedBits, but not if there is an
1547       // additional user.
1548       if (!IsAshr || (C2.shl(*C3).lshr(*C3) == C2)) {
1549         if (!Cmp.isSigned() ||
1550             (!C2.shl(*C3).isNegative() && !C1.shl(*C3).isNegative()))
1551           CanFold = true;
1552       }
1553     }
1554 
1555     if (CanFold) {
1556       APInt NewCst = IsShl ? C1.lshr(*C3) : C1.shl(*C3);
1557       APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3);
1558       // Check to see if we are shifting out any of the bits being compared.
1559       if (SameAsC1 != C1) {
1560         // If we shifted bits out, the fold is not going to work out. As a
1561         // special case, check to see if this means that the result is always
1562         // true or false now.
1563         if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1564           return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1565         if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1566           return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1567       } else {
1568         Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst));
1569         APInt NewAndCst = IsShl ? C2.lshr(*C3) : C2.shl(*C3);
1570         And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst));
1571         And->setOperand(0, Shift->getOperand(0));
1572         Worklist.Add(Shift); // Shift is dead.
1573         return &Cmp;
1574       }
1575     }
1576   }
1577 
1578   // Turn ((X >> Y) & C2) == 0  into  (X & (C2 << Y)) == 0.  The latter is
1579   // preferable because it allows the C2 << Y expression to be hoisted out of a
1580   // loop if Y is invariant and X is not.
1581   if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1582       !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1583     // Compute C2 << Y.
1584     Value *NewShift =
1585         IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1586               : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1587 
1588     // Compute X & (C2 << Y).
1589     Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1590     Cmp.setOperand(0, NewAnd);
1591     return &Cmp;
1592   }
1593 
1594   return nullptr;
1595 }
1596 
1597 /// Fold icmp (and X, C2), C1.
foldICmpAndConstConst(ICmpInst & Cmp,BinaryOperator * And,const APInt & C1)1598 Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp,
1599                                                  BinaryOperator *And,
1600                                                  const APInt &C1) {
1601   const APInt *C2;
1602   if (!match(And->getOperand(1), m_APInt(C2)))
1603     return nullptr;
1604 
1605   if (!And->hasOneUse())
1606     return nullptr;
1607 
1608   // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1609   // the input width without changing the value produced, eliminate the cast:
1610   //
1611   // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1612   //
1613   // We can do this transformation if the constants do not have their sign bits
1614   // set or if it is an equality comparison. Extending a relational comparison
1615   // when we're checking the sign bit would not work.
1616   Value *W;
1617   if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1618       (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1619     // TODO: Is this a good transform for vectors? Wider types may reduce
1620     // throughput. Should this transform be limited (even for scalars) by using
1621     // shouldChangeType()?
1622     if (!Cmp.getType()->isVectorTy()) {
1623       Type *WideType = W->getType();
1624       unsigned WideScalarBits = WideType->getScalarSizeInBits();
1625       Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1626       Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1627       Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1628       return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1629     }
1630   }
1631 
1632   if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1633     return I;
1634 
1635   // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1636   // (icmp pred (and A, (or (shl 1, B), 1), 0))
1637   //
1638   // iff pred isn't signed
1639   if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1640       match(And->getOperand(1), m_One())) {
1641     Constant *One = cast<Constant>(And->getOperand(1));
1642     Value *Or = And->getOperand(0);
1643     Value *A, *B, *LShr;
1644     if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1645         match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1646       unsigned UsesRemoved = 0;
1647       if (And->hasOneUse())
1648         ++UsesRemoved;
1649       if (Or->hasOneUse())
1650         ++UsesRemoved;
1651       if (LShr->hasOneUse())
1652         ++UsesRemoved;
1653 
1654       // Compute A & ((1 << B) | 1)
1655       Value *NewOr = nullptr;
1656       if (auto *C = dyn_cast<Constant>(B)) {
1657         if (UsesRemoved >= 1)
1658           NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1659       } else {
1660         if (UsesRemoved >= 3)
1661           NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1662                                                      /*HasNUW=*/true),
1663                                    One, Or->getName());
1664       }
1665       if (NewOr) {
1666         Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1667         Cmp.setOperand(0, NewAnd);
1668         return &Cmp;
1669       }
1670     }
1671   }
1672 
1673   return nullptr;
1674 }
1675 
1676 /// Fold icmp (and X, Y), C.
foldICmpAndConstant(ICmpInst & Cmp,BinaryOperator * And,const APInt & C)1677 Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp,
1678                                                BinaryOperator *And,
1679                                                const APInt &C) {
1680   if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1681     return I;
1682 
1683   // TODO: These all require that Y is constant too, so refactor with the above.
1684 
1685   // Try to optimize things like "A[i] & 42 == 0" to index computations.
1686   Value *X = And->getOperand(0);
1687   Value *Y = And->getOperand(1);
1688   if (auto *LI = dyn_cast<LoadInst>(X))
1689     if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1690       if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1691         if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1692             !LI->isVolatile() && isa<ConstantInt>(Y)) {
1693           ConstantInt *C2 = cast<ConstantInt>(Y);
1694           if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1695             return Res;
1696         }
1697 
1698   if (!Cmp.isEquality())
1699     return nullptr;
1700 
1701   // X & -C == -C -> X >  u ~C
1702   // X & -C != -C -> X <= u ~C
1703   //   iff C is a power of 2
1704   if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1705     auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1706                                                           : CmpInst::ICMP_ULE;
1707     return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1708   }
1709 
1710   // (X & C2) == 0 -> (trunc X) >= 0
1711   // (X & C2) != 0 -> (trunc X) <  0
1712   //   iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1713   const APInt *C2;
1714   if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1715     int32_t ExactLogBase2 = C2->exactLogBase2();
1716     if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1717       Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1718       if (And->getType()->isVectorTy())
1719         NTy = VectorType::get(NTy, And->getType()->getVectorNumElements());
1720       Value *Trunc = Builder.CreateTrunc(X, NTy);
1721       auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1722                                                             : CmpInst::ICMP_SLT;
1723       return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1724     }
1725   }
1726 
1727   return nullptr;
1728 }
1729 
1730 /// Fold icmp (or X, Y), C.
foldICmpOrConstant(ICmpInst & Cmp,BinaryOperator * Or,const APInt & C)1731 Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
1732                                               const APInt &C) {
1733   ICmpInst::Predicate Pred = Cmp.getPredicate();
1734   if (C.isOneValue()) {
1735     // icmp slt signum(V) 1 --> icmp slt V, 1
1736     Value *V = nullptr;
1737     if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1738       return new ICmpInst(ICmpInst::ICMP_SLT, V,
1739                           ConstantInt::get(V->getType(), 1));
1740   }
1741 
1742   // X | C == C --> X <=u C
1743   // X | C != C --> X  >u C
1744   //   iff C+1 is a power of 2 (C is a bitmask of the low bits)
1745   if (Cmp.isEquality() && Cmp.getOperand(1) == Or->getOperand(1) &&
1746       (C + 1).isPowerOf2()) {
1747     Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
1748     return new ICmpInst(Pred, Or->getOperand(0), Or->getOperand(1));
1749   }
1750 
1751   if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1752     return nullptr;
1753 
1754   Value *P, *Q;
1755   if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1756     // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1757     // -> and (icmp eq P, null), (icmp eq Q, null).
1758     Value *CmpP =
1759         Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1760     Value *CmpQ =
1761         Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1762     auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1763     return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1764   }
1765 
1766   // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1767   // a shorter form that has more potential to be folded even further.
1768   Value *X1, *X2, *X3, *X4;
1769   if (match(Or->getOperand(0), m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1770       match(Or->getOperand(1), m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1771     // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1772     // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1773     Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1774     Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1775     auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1776     return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1777   }
1778 
1779   return nullptr;
1780 }
1781 
1782 /// Fold icmp (mul X, Y), C.
foldICmpMulConstant(ICmpInst & Cmp,BinaryOperator * Mul,const APInt & C)1783 Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp,
1784                                                BinaryOperator *Mul,
1785                                                const APInt &C) {
1786   const APInt *MulC;
1787   if (!match(Mul->getOperand(1), m_APInt(MulC)))
1788     return nullptr;
1789 
1790   // If this is a test of the sign bit and the multiply is sign-preserving with
1791   // a constant operand, use the multiply LHS operand instead.
1792   ICmpInst::Predicate Pred = Cmp.getPredicate();
1793   if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1794     if (MulC->isNegative())
1795       Pred = ICmpInst::getSwappedPredicate(Pred);
1796     return new ICmpInst(Pred, Mul->getOperand(0),
1797                         Constant::getNullValue(Mul->getType()));
1798   }
1799 
1800   return nullptr;
1801 }
1802 
1803 /// Fold icmp (shl 1, Y), C.
foldICmpShlOne(ICmpInst & Cmp,Instruction * Shl,const APInt & C)1804 static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl,
1805                                    const APInt &C) {
1806   Value *Y;
1807   if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1808     return nullptr;
1809 
1810   Type *ShiftType = Shl->getType();
1811   unsigned TypeBits = C.getBitWidth();
1812   bool CIsPowerOf2 = C.isPowerOf2();
1813   ICmpInst::Predicate Pred = Cmp.getPredicate();
1814   if (Cmp.isUnsigned()) {
1815     // (1 << Y) pred C -> Y pred Log2(C)
1816     if (!CIsPowerOf2) {
1817       // (1 << Y) <  30 -> Y <= 4
1818       // (1 << Y) <= 30 -> Y <= 4
1819       // (1 << Y) >= 30 -> Y >  4
1820       // (1 << Y) >  30 -> Y >  4
1821       if (Pred == ICmpInst::ICMP_ULT)
1822         Pred = ICmpInst::ICMP_ULE;
1823       else if (Pred == ICmpInst::ICMP_UGE)
1824         Pred = ICmpInst::ICMP_UGT;
1825     }
1826 
1827     // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
1828     // (1 << Y) <  2147483648 -> Y <  31 -> Y != 31
1829     unsigned CLog2 = C.logBase2();
1830     if (CLog2 == TypeBits - 1) {
1831       if (Pred == ICmpInst::ICMP_UGE)
1832         Pred = ICmpInst::ICMP_EQ;
1833       else if (Pred == ICmpInst::ICMP_ULT)
1834         Pred = ICmpInst::ICMP_NE;
1835     }
1836     return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
1837   } else if (Cmp.isSigned()) {
1838     Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
1839     if (C.isAllOnesValue()) {
1840       // (1 << Y) <= -1 -> Y == 31
1841       if (Pred == ICmpInst::ICMP_SLE)
1842         return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1843 
1844       // (1 << Y) >  -1 -> Y != 31
1845       if (Pred == ICmpInst::ICMP_SGT)
1846         return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1847     } else if (!C) {
1848       // (1 << Y) <  0 -> Y == 31
1849       // (1 << Y) <= 0 -> Y == 31
1850       if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1851         return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1852 
1853       // (1 << Y) >= 0 -> Y != 31
1854       // (1 << Y) >  0 -> Y != 31
1855       if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1856         return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1857     }
1858   } else if (Cmp.isEquality() && CIsPowerOf2) {
1859     return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
1860   }
1861 
1862   return nullptr;
1863 }
1864 
1865 /// Fold icmp (shl X, Y), C.
foldICmpShlConstant(ICmpInst & Cmp,BinaryOperator * Shl,const APInt & C)1866 Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp,
1867                                                BinaryOperator *Shl,
1868                                                const APInt &C) {
1869   const APInt *ShiftVal;
1870   if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
1871     return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
1872 
1873   const APInt *ShiftAmt;
1874   if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
1875     return foldICmpShlOne(Cmp, Shl, C);
1876 
1877   // Check that the shift amount is in range. If not, don't perform undefined
1878   // shifts. When the shift is visited, it will be simplified.
1879   unsigned TypeBits = C.getBitWidth();
1880   if (ShiftAmt->uge(TypeBits))
1881     return nullptr;
1882 
1883   ICmpInst::Predicate Pred = Cmp.getPredicate();
1884   Value *X = Shl->getOperand(0);
1885   Type *ShType = Shl->getType();
1886 
1887   // NSW guarantees that we are only shifting out sign bits from the high bits,
1888   // so we can ASHR the compare constant without needing a mask and eliminate
1889   // the shift.
1890   if (Shl->hasNoSignedWrap()) {
1891     if (Pred == ICmpInst::ICMP_SGT) {
1892       // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
1893       APInt ShiftedC = C.ashr(*ShiftAmt);
1894       return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1895     }
1896     if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
1897         C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
1898       APInt ShiftedC = C.ashr(*ShiftAmt);
1899       return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1900     }
1901     if (Pred == ICmpInst::ICMP_SLT) {
1902       // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
1903       // (X << S) <=s C is equiv to X <=s (C >> S) for all C
1904       // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
1905       // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
1906       assert(!C.isMinSignedValue() && "Unexpected icmp slt");
1907       APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
1908       return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1909     }
1910     // If this is a signed comparison to 0 and the shift is sign preserving,
1911     // use the shift LHS operand instead; isSignTest may change 'Pred', so only
1912     // do that if we're sure to not continue on in this function.
1913     if (isSignTest(Pred, C))
1914       return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
1915   }
1916 
1917   // NUW guarantees that we are only shifting out zero bits from the high bits,
1918   // so we can LSHR the compare constant without needing a mask and eliminate
1919   // the shift.
1920   if (Shl->hasNoUnsignedWrap()) {
1921     if (Pred == ICmpInst::ICMP_UGT) {
1922       // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
1923       APInt ShiftedC = C.lshr(*ShiftAmt);
1924       return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1925     }
1926     if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
1927         C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
1928       APInt ShiftedC = C.lshr(*ShiftAmt);
1929       return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1930     }
1931     if (Pred == ICmpInst::ICMP_ULT) {
1932       // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
1933       // (X << S) <=u C is equiv to X <=u (C >> S) for all C
1934       // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
1935       // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
1936       assert(C.ugt(0) && "ult 0 should have been eliminated");
1937       APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
1938       return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1939     }
1940   }
1941 
1942   if (Cmp.isEquality() && Shl->hasOneUse()) {
1943     // Strength-reduce the shift into an 'and'.
1944     Constant *Mask = ConstantInt::get(
1945         ShType,
1946         APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
1947     Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
1948     Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
1949     return new ICmpInst(Pred, And, LShrC);
1950   }
1951 
1952   // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1953   bool TrueIfSigned = false;
1954   if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
1955     // (X << 31) <s 0  --> (X & 1) != 0
1956     Constant *Mask = ConstantInt::get(
1957         ShType,
1958         APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
1959     Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
1960     return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1961                         And, Constant::getNullValue(ShType));
1962   }
1963 
1964   // Transform (icmp pred iM (shl iM %v, N), C)
1965   // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
1966   // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
1967   // This enables us to get rid of the shift in favor of a trunc that may be
1968   // free on the target. It has the additional benefit of comparing to a
1969   // smaller constant that may be more target-friendly.
1970   unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
1971   if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
1972       DL.isLegalInteger(TypeBits - Amt)) {
1973     Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
1974     if (ShType->isVectorTy())
1975       TruncTy = VectorType::get(TruncTy, ShType->getVectorNumElements());
1976     Constant *NewC =
1977         ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
1978     return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
1979   }
1980 
1981   return nullptr;
1982 }
1983 
1984 /// Fold icmp ({al}shr X, Y), C.
foldICmpShrConstant(ICmpInst & Cmp,BinaryOperator * Shr,const APInt & C)1985 Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp,
1986                                                BinaryOperator *Shr,
1987                                                const APInt &C) {
1988   // An exact shr only shifts out zero bits, so:
1989   // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
1990   Value *X = Shr->getOperand(0);
1991   CmpInst::Predicate Pred = Cmp.getPredicate();
1992   if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
1993       C.isNullValue())
1994     return new ICmpInst(Pred, X, Cmp.getOperand(1));
1995 
1996   const APInt *ShiftVal;
1997   if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
1998     return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
1999 
2000   const APInt *ShiftAmt;
2001   if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2002     return nullptr;
2003 
2004   // Check that the shift amount is in range. If not, don't perform undefined
2005   // shifts. When the shift is visited it will be simplified.
2006   unsigned TypeBits = C.getBitWidth();
2007   unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2008   if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2009     return nullptr;
2010 
2011   bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2012   bool IsExact = Shr->isExact();
2013   Type *ShrTy = Shr->getType();
2014   // TODO: If we could guarantee that InstSimplify would handle all of the
2015   // constant-value-based preconditions in the folds below, then we could assert
2016   // those conditions rather than checking them. This is difficult because of
2017   // undef/poison (PR34838).
2018   if (IsAShr) {
2019     if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2020       // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2021       // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2022       APInt ShiftedC = C.shl(ShAmtVal);
2023       if (ShiftedC.ashr(ShAmtVal) == C)
2024         return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2025     }
2026     if (Pred == CmpInst::ICMP_SGT) {
2027       // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2028       APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2029       if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2030           (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2031         return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2032     }
2033   } else {
2034     if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2035       // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2036       // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2037       APInt ShiftedC = C.shl(ShAmtVal);
2038       if (ShiftedC.lshr(ShAmtVal) == C)
2039         return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2040     }
2041     if (Pred == CmpInst::ICMP_UGT) {
2042       // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2043       APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2044       if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2045         return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2046     }
2047   }
2048 
2049   if (!Cmp.isEquality())
2050     return nullptr;
2051 
2052   // Handle equality comparisons of shift-by-constant.
2053 
2054   // If the comparison constant changes with the shift, the comparison cannot
2055   // succeed (bits of the comparison constant cannot match the shifted value).
2056   // This should be known by InstSimplify and already be folded to true/false.
2057   assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2058           (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2059          "Expected icmp+shr simplify did not occur.");
2060 
2061   // If the bits shifted out are known zero, compare the unshifted value:
2062   //  (X & 4) >> 1 == 2  --> (X & 4) == 4.
2063   if (Shr->isExact())
2064     return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2065 
2066   if (Shr->hasOneUse()) {
2067     // Canonicalize the shift into an 'and':
2068     // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2069     APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2070     Constant *Mask = ConstantInt::get(ShrTy, Val);
2071     Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2072     return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2073   }
2074 
2075   return nullptr;
2076 }
2077 
2078 /// Fold icmp (udiv X, Y), C.
foldICmpUDivConstant(ICmpInst & Cmp,BinaryOperator * UDiv,const APInt & C)2079 Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp,
2080                                                 BinaryOperator *UDiv,
2081                                                 const APInt &C) {
2082   const APInt *C2;
2083   if (!match(UDiv->getOperand(0), m_APInt(C2)))
2084     return nullptr;
2085 
2086   assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2087 
2088   // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2089   Value *Y = UDiv->getOperand(1);
2090   if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2091     assert(!C.isMaxValue() &&
2092            "icmp ugt X, UINT_MAX should have been simplified already.");
2093     return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2094                         ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2095   }
2096 
2097   // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2098   if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2099     assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2100     return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2101                         ConstantInt::get(Y->getType(), C2->udiv(C)));
2102   }
2103 
2104   return nullptr;
2105 }
2106 
2107 /// Fold icmp ({su}div X, Y), C.
foldICmpDivConstant(ICmpInst & Cmp,BinaryOperator * Div,const APInt & C)2108 Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp,
2109                                                BinaryOperator *Div,
2110                                                const APInt &C) {
2111   // Fold: icmp pred ([us]div X, C2), C -> range test
2112   // Fold this div into the comparison, producing a range check.
2113   // Determine, based on the divide type, what the range is being
2114   // checked.  If there is an overflow on the low or high side, remember
2115   // it, otherwise compute the range [low, hi) bounding the new value.
2116   // See: InsertRangeTest above for the kinds of replacements possible.
2117   const APInt *C2;
2118   if (!match(Div->getOperand(1), m_APInt(C2)))
2119     return nullptr;
2120 
2121   // FIXME: If the operand types don't match the type of the divide
2122   // then don't attempt this transform. The code below doesn't have the
2123   // logic to deal with a signed divide and an unsigned compare (and
2124   // vice versa). This is because (x /s C2) <s C  produces different
2125   // results than (x /s C2) <u C or (x /u C2) <s C or even
2126   // (x /u C2) <u C.  Simply casting the operands and result won't
2127   // work. :(  The if statement below tests that condition and bails
2128   // if it finds it.
2129   bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2130   if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2131     return nullptr;
2132 
2133   // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2134   // INT_MIN will also fail if the divisor is 1. Although folds of all these
2135   // division-by-constant cases should be present, we can not assert that they
2136   // have happened before we reach this icmp instruction.
2137   if (C2->isNullValue() || C2->isOneValue() ||
2138       (DivIsSigned && C2->isAllOnesValue()))
2139     return nullptr;
2140 
2141   // Compute Prod = C * C2. We are essentially solving an equation of
2142   // form X / C2 = C. We solve for X by multiplying C2 and C.
2143   // By solving for X, we can turn this into a range check instead of computing
2144   // a divide.
2145   APInt Prod = C * *C2;
2146 
2147   // Determine if the product overflows by seeing if the product is not equal to
2148   // the divide. Make sure we do the same kind of divide as in the LHS
2149   // instruction that we're folding.
2150   bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2151 
2152   ICmpInst::Predicate Pred = Cmp.getPredicate();
2153 
2154   // If the division is known to be exact, then there is no remainder from the
2155   // divide, so the covered range size is unit, otherwise it is the divisor.
2156   APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2157 
2158   // Figure out the interval that is being checked.  For example, a comparison
2159   // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2160   // Compute this interval based on the constants involved and the signedness of
2161   // the compare/divide.  This computes a half-open interval, keeping track of
2162   // whether either value in the interval overflows.  After analysis each
2163   // overflow variable is set to 0 if it's corresponding bound variable is valid
2164   // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2165   int LoOverflow = 0, HiOverflow = 0;
2166   APInt LoBound, HiBound;
2167 
2168   if (!DivIsSigned) {  // udiv
2169     // e.g. X/5 op 3  --> [15, 20)
2170     LoBound = Prod;
2171     HiOverflow = LoOverflow = ProdOV;
2172     if (!HiOverflow) {
2173       // If this is not an exact divide, then many values in the range collapse
2174       // to the same result value.
2175       HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2176     }
2177   } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2178     if (C.isNullValue()) {       // (X / pos) op 0
2179       // Can't overflow.  e.g.  X/2 op 0 --> [-1, 2)
2180       LoBound = -(RangeSize - 1);
2181       HiBound = RangeSize;
2182     } else if (C.isStrictlyPositive()) {   // (X / pos) op pos
2183       LoBound = Prod;     // e.g.   X/5 op 3 --> [15, 20)
2184       HiOverflow = LoOverflow = ProdOV;
2185       if (!HiOverflow)
2186         HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2187     } else {                       // (X / pos) op neg
2188       // e.g. X/5 op -3  --> [-15-4, -15+1) --> [-19, -14)
2189       HiBound = Prod + 1;
2190       LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2191       if (!LoOverflow) {
2192         APInt DivNeg = -RangeSize;
2193         LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2194       }
2195     }
2196   } else if (C2->isNegative()) { // Divisor is < 0.
2197     if (Div->isExact())
2198       RangeSize.negate();
2199     if (C.isNullValue()) { // (X / neg) op 0
2200       // e.g. X/-5 op 0  --> [-4, 5)
2201       LoBound = RangeSize + 1;
2202       HiBound = -RangeSize;
2203       if (HiBound == *C2) {        // -INTMIN = INTMIN
2204         HiOverflow = 1;            // [INTMIN+1, overflow)
2205         HiBound = APInt();         // e.g. X/INTMIN = 0 --> X > INTMIN
2206       }
2207     } else if (C.isStrictlyPositive()) {   // (X / neg) op pos
2208       // e.g. X/-5 op 3  --> [-19, -14)
2209       HiBound = Prod + 1;
2210       HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2211       if (!LoOverflow)
2212         LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2213     } else {                       // (X / neg) op neg
2214       LoBound = Prod;       // e.g. X/-5 op -3  --> [15, 20)
2215       LoOverflow = HiOverflow = ProdOV;
2216       if (!HiOverflow)
2217         HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2218     }
2219 
2220     // Dividing by a negative swaps the condition.  LT <-> GT
2221     Pred = ICmpInst::getSwappedPredicate(Pred);
2222   }
2223 
2224   Value *X = Div->getOperand(0);
2225   switch (Pred) {
2226     default: llvm_unreachable("Unhandled icmp opcode!");
2227     case ICmpInst::ICMP_EQ:
2228       if (LoOverflow && HiOverflow)
2229         return replaceInstUsesWith(Cmp, Builder.getFalse());
2230       if (HiOverflow)
2231         return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2232                             ICmpInst::ICMP_UGE, X,
2233                             ConstantInt::get(Div->getType(), LoBound));
2234       if (LoOverflow)
2235         return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2236                             ICmpInst::ICMP_ULT, X,
2237                             ConstantInt::get(Div->getType(), HiBound));
2238       return replaceInstUsesWith(
2239           Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2240     case ICmpInst::ICMP_NE:
2241       if (LoOverflow && HiOverflow)
2242         return replaceInstUsesWith(Cmp, Builder.getTrue());
2243       if (HiOverflow)
2244         return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2245                             ICmpInst::ICMP_ULT, X,
2246                             ConstantInt::get(Div->getType(), LoBound));
2247       if (LoOverflow)
2248         return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2249                             ICmpInst::ICMP_UGE, X,
2250                             ConstantInt::get(Div->getType(), HiBound));
2251       return replaceInstUsesWith(Cmp,
2252                                  insertRangeTest(X, LoBound, HiBound,
2253                                                  DivIsSigned, false));
2254     case ICmpInst::ICMP_ULT:
2255     case ICmpInst::ICMP_SLT:
2256       if (LoOverflow == +1)   // Low bound is greater than input range.
2257         return replaceInstUsesWith(Cmp, Builder.getTrue());
2258       if (LoOverflow == -1)   // Low bound is less than input range.
2259         return replaceInstUsesWith(Cmp, Builder.getFalse());
2260       return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2261     case ICmpInst::ICMP_UGT:
2262     case ICmpInst::ICMP_SGT:
2263       if (HiOverflow == +1)       // High bound greater than input range.
2264         return replaceInstUsesWith(Cmp, Builder.getFalse());
2265       if (HiOverflow == -1)       // High bound less than input range.
2266         return replaceInstUsesWith(Cmp, Builder.getTrue());
2267       if (Pred == ICmpInst::ICMP_UGT)
2268         return new ICmpInst(ICmpInst::ICMP_UGE, X,
2269                             ConstantInt::get(Div->getType(), HiBound));
2270       return new ICmpInst(ICmpInst::ICMP_SGE, X,
2271                           ConstantInt::get(Div->getType(), HiBound));
2272   }
2273 
2274   return nullptr;
2275 }
2276 
2277 /// Fold icmp (sub X, Y), C.
foldICmpSubConstant(ICmpInst & Cmp,BinaryOperator * Sub,const APInt & C)2278 Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp,
2279                                                BinaryOperator *Sub,
2280                                                const APInt &C) {
2281   Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2282   ICmpInst::Predicate Pred = Cmp.getPredicate();
2283 
2284   // The following transforms are only worth it if the only user of the subtract
2285   // is the icmp.
2286   if (!Sub->hasOneUse())
2287     return nullptr;
2288 
2289   if (Sub->hasNoSignedWrap()) {
2290     // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2291     if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2292       return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2293 
2294     // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2295     if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2296       return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2297 
2298     // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2299     if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2300       return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2301 
2302     // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2303     if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2304       return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2305   }
2306 
2307   const APInt *C2;
2308   if (!match(X, m_APInt(C2)))
2309     return nullptr;
2310 
2311   // C2 - Y <u C -> (Y | (C - 1)) == C2
2312   //   iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2313   if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2314       (*C2 & (C - 1)) == (C - 1))
2315     return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2316 
2317   // C2 - Y >u C -> (Y | C) != C2
2318   //   iff C2 & C == C and C + 1 is a power of 2
2319   if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2320     return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2321 
2322   return nullptr;
2323 }
2324 
2325 /// Fold icmp (add X, Y), C.
foldICmpAddConstant(ICmpInst & Cmp,BinaryOperator * Add,const APInt & C)2326 Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp,
2327                                                BinaryOperator *Add,
2328                                                const APInt &C) {
2329   Value *Y = Add->getOperand(1);
2330   const APInt *C2;
2331   if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2332     return nullptr;
2333 
2334   // Fold icmp pred (add X, C2), C.
2335   Value *X = Add->getOperand(0);
2336   Type *Ty = Add->getType();
2337   CmpInst::Predicate Pred = Cmp.getPredicate();
2338 
2339   // If the add does not wrap, we can always adjust the compare by subtracting
2340   // the constants. Equality comparisons are handled elsewhere. SGE/SLE are
2341   // canonicalized to SGT/SLT.
2342   if (Add->hasNoSignedWrap() &&
2343       (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) {
2344     bool Overflow;
2345     APInt NewC = C.ssub_ov(*C2, Overflow);
2346     // If there is overflow, the result must be true or false.
2347     // TODO: Can we assert there is no overflow because InstSimplify always
2348     // handles those cases?
2349     if (!Overflow)
2350       // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2351       return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2352   }
2353 
2354   auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2355   const APInt &Upper = CR.getUpper();
2356   const APInt &Lower = CR.getLower();
2357   if (Cmp.isSigned()) {
2358     if (Lower.isSignMask())
2359       return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2360     if (Upper.isSignMask())
2361       return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2362   } else {
2363     if (Lower.isMinValue())
2364       return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2365     if (Upper.isMinValue())
2366       return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2367   }
2368 
2369   if (!Add->hasOneUse())
2370     return nullptr;
2371 
2372   // X+C <u C2 -> (X & -C2) == C
2373   //   iff C & (C2-1) == 0
2374   //       C2 is a power of 2
2375   if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2376     return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2377                         ConstantExpr::getNeg(cast<Constant>(Y)));
2378 
2379   // X+C >u C2 -> (X & ~C2) != C
2380   //   iff C & C2 == 0
2381   //       C2+1 is a power of 2
2382   if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2383     return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2384                         ConstantExpr::getNeg(cast<Constant>(Y)));
2385 
2386   return nullptr;
2387 }
2388 
matchThreeWayIntCompare(SelectInst * SI,Value * & LHS,Value * & RHS,ConstantInt * & Less,ConstantInt * & Equal,ConstantInt * & Greater)2389 bool InstCombiner::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2390                                            Value *&RHS, ConstantInt *&Less,
2391                                            ConstantInt *&Equal,
2392                                            ConstantInt *&Greater) {
2393   // TODO: Generalize this to work with other comparison idioms or ensure
2394   // they get canonicalized into this form.
2395 
2396   // select i1 (a == b), i32 Equal, i32 (select i1 (a < b), i32 Less, i32
2397   // Greater), where Equal, Less and Greater are placeholders for any three
2398   // constants.
2399   ICmpInst::Predicate PredA, PredB;
2400   if (match(SI->getTrueValue(), m_ConstantInt(Equal)) &&
2401       match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) &&
2402       PredA == ICmpInst::ICMP_EQ &&
2403       match(SI->getFalseValue(),
2404             m_Select(m_ICmp(PredB, m_Specific(LHS), m_Specific(RHS)),
2405                      m_ConstantInt(Less), m_ConstantInt(Greater))) &&
2406       PredB == ICmpInst::ICMP_SLT) {
2407     return true;
2408   }
2409   return false;
2410 }
2411 
foldICmpSelectConstant(ICmpInst & Cmp,SelectInst * Select,ConstantInt * C)2412 Instruction *InstCombiner::foldICmpSelectConstant(ICmpInst &Cmp,
2413                                                   SelectInst *Select,
2414                                                   ConstantInt *C) {
2415 
2416   assert(C && "Cmp RHS should be a constant int!");
2417   // If we're testing a constant value against the result of a three way
2418   // comparison, the result can be expressed directly in terms of the
2419   // original values being compared.  Note: We could possibly be more
2420   // aggressive here and remove the hasOneUse test. The original select is
2421   // really likely to simplify or sink when we remove a test of the result.
2422   Value *OrigLHS, *OrigRHS;
2423   ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2424   if (Cmp.hasOneUse() &&
2425       matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2426                               C3GreaterThan)) {
2427     assert(C1LessThan && C2Equal && C3GreaterThan);
2428 
2429     bool TrueWhenLessThan =
2430         ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2431             ->isAllOnesValue();
2432     bool TrueWhenEqual =
2433         ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2434             ->isAllOnesValue();
2435     bool TrueWhenGreaterThan =
2436         ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2437             ->isAllOnesValue();
2438 
2439     // This generates the new instruction that will replace the original Cmp
2440     // Instruction. Instead of enumerating the various combinations when
2441     // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2442     // false, we rely on chaining of ORs and future passes of InstCombine to
2443     // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2444 
2445     // When none of the three constants satisfy the predicate for the RHS (C),
2446     // the entire original Cmp can be simplified to a false.
2447     Value *Cond = Builder.getFalse();
2448     if (TrueWhenLessThan)
2449       Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT, OrigLHS, OrigRHS));
2450     if (TrueWhenEqual)
2451       Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ, OrigLHS, OrigRHS));
2452     if (TrueWhenGreaterThan)
2453       Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT, OrigLHS, OrigRHS));
2454 
2455     return replaceInstUsesWith(Cmp, Cond);
2456   }
2457   return nullptr;
2458 }
2459 
foldICmpBitCastConstant(ICmpInst & Cmp,BitCastInst * Bitcast,const APInt & C)2460 Instruction *InstCombiner::foldICmpBitCastConstant(ICmpInst &Cmp,
2461                                                    BitCastInst *Bitcast,
2462                                                    const APInt &C) {
2463   // Folding: icmp <pred> iN X, C
2464   //  where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2465   //    and C is a splat of a K-bit pattern
2466   //    and SC is a constant vector = <C', C', C', ..., C'>
2467   // Into:
2468   //   %E = extractelement <M x iK> %vec, i32 C'
2469   //   icmp <pred> iK %E, trunc(C)
2470   if (!Bitcast->getType()->isIntegerTy() ||
2471       !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2472     return nullptr;
2473 
2474   Value *BCIOp = Bitcast->getOperand(0);
2475   Value *Vec = nullptr;     // 1st vector arg of the shufflevector
2476   Constant *Mask = nullptr; // Mask arg of the shufflevector
2477   if (match(BCIOp,
2478             m_ShuffleVector(m_Value(Vec), m_Undef(), m_Constant(Mask)))) {
2479     // Check whether every element of Mask is the same constant
2480     if (auto *Elem = dyn_cast_or_null<ConstantInt>(Mask->getSplatValue())) {
2481       auto *VecTy = cast<VectorType>(BCIOp->getType());
2482       auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2483       auto Pred = Cmp.getPredicate();
2484       if (C.isSplat(EltTy->getBitWidth())) {
2485         // Fold the icmp based on the value of C
2486         // If C is M copies of an iK sized bit pattern,
2487         // then:
2488         //   =>  %E = extractelement <N x iK> %vec, i32 Elem
2489         //       icmp <pred> iK %SplatVal, <pattern>
2490         Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2491         Value *NewC = ConstantInt::get(EltTy, C.trunc(EltTy->getBitWidth()));
2492         return new ICmpInst(Pred, Extract, NewC);
2493       }
2494     }
2495   }
2496   return nullptr;
2497 }
2498 
2499 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2500 /// where X is some kind of instruction.
foldICmpInstWithConstant(ICmpInst & Cmp)2501 Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) {
2502   const APInt *C;
2503   if (!match(Cmp.getOperand(1), m_APInt(C)))
2504     return nullptr;
2505 
2506   if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2507     switch (BO->getOpcode()) {
2508     case Instruction::Xor:
2509       if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2510         return I;
2511       break;
2512     case Instruction::And:
2513       if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2514         return I;
2515       break;
2516     case Instruction::Or:
2517       if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2518         return I;
2519       break;
2520     case Instruction::Mul:
2521       if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2522         return I;
2523       break;
2524     case Instruction::Shl:
2525       if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2526         return I;
2527       break;
2528     case Instruction::LShr:
2529     case Instruction::AShr:
2530       if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2531         return I;
2532       break;
2533     case Instruction::UDiv:
2534       if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2535         return I;
2536       LLVM_FALLTHROUGH;
2537     case Instruction::SDiv:
2538       if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2539         return I;
2540       break;
2541     case Instruction::Sub:
2542       if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2543         return I;
2544       break;
2545     case Instruction::Add:
2546       if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2547         return I;
2548       break;
2549     default:
2550       break;
2551     }
2552     // TODO: These folds could be refactored to be part of the above calls.
2553     if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2554       return I;
2555   }
2556 
2557   // Match against CmpInst LHS being instructions other than binary operators.
2558 
2559   if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2560     // For now, we only support constant integers while folding the
2561     // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2562     // similar to the cases handled by binary ops above.
2563     if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2564       if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2565         return I;
2566   }
2567 
2568   if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2569     if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2570       return I;
2571   }
2572 
2573   if (auto *BCI = dyn_cast<BitCastInst>(Cmp.getOperand(0))) {
2574     if (Instruction *I = foldICmpBitCastConstant(Cmp, BCI, *C))
2575       return I;
2576   }
2577 
2578   if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, *C))
2579     return I;
2580 
2581   return nullptr;
2582 }
2583 
2584 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2585 /// icmp eq/ne BO, C.
foldICmpBinOpEqualityWithConstant(ICmpInst & Cmp,BinaryOperator * BO,const APInt & C)2586 Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
2587                                                              BinaryOperator *BO,
2588                                                              const APInt &C) {
2589   // TODO: Some of these folds could work with arbitrary constants, but this
2590   // function is limited to scalar and vector splat constants.
2591   if (!Cmp.isEquality())
2592     return nullptr;
2593 
2594   ICmpInst::Predicate Pred = Cmp.getPredicate();
2595   bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2596   Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2597   Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2598 
2599   switch (BO->getOpcode()) {
2600   case Instruction::SRem:
2601     // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2602     if (C.isNullValue() && BO->hasOneUse()) {
2603       const APInt *BOC;
2604       if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2605         Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
2606         return new ICmpInst(Pred, NewRem,
2607                             Constant::getNullValue(BO->getType()));
2608       }
2609     }
2610     break;
2611   case Instruction::Add: {
2612     // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
2613     const APInt *BOC;
2614     if (match(BOp1, m_APInt(BOC))) {
2615       if (BO->hasOneUse()) {
2616         Constant *SubC = ConstantExpr::getSub(RHS, cast<Constant>(BOp1));
2617         return new ICmpInst(Pred, BOp0, SubC);
2618       }
2619     } else if (C.isNullValue()) {
2620       // Replace ((add A, B) != 0) with (A != -B) if A or B is
2621       // efficiently invertible, or if the add has just this one use.
2622       if (Value *NegVal = dyn_castNegVal(BOp1))
2623         return new ICmpInst(Pred, BOp0, NegVal);
2624       if (Value *NegVal = dyn_castNegVal(BOp0))
2625         return new ICmpInst(Pred, NegVal, BOp1);
2626       if (BO->hasOneUse()) {
2627         Value *Neg = Builder.CreateNeg(BOp1);
2628         Neg->takeName(BO);
2629         return new ICmpInst(Pred, BOp0, Neg);
2630       }
2631     }
2632     break;
2633   }
2634   case Instruction::Xor:
2635     if (BO->hasOneUse()) {
2636       if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2637         // For the xor case, we can xor two constants together, eliminating
2638         // the explicit xor.
2639         return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
2640       } else if (C.isNullValue()) {
2641         // Replace ((xor A, B) != 0) with (A != B)
2642         return new ICmpInst(Pred, BOp0, BOp1);
2643       }
2644     }
2645     break;
2646   case Instruction::Sub:
2647     if (BO->hasOneUse()) {
2648       const APInt *BOC;
2649       if (match(BOp0, m_APInt(BOC))) {
2650         // Replace ((sub BOC, B) != C) with (B != BOC-C).
2651         Constant *SubC = ConstantExpr::getSub(cast<Constant>(BOp0), RHS);
2652         return new ICmpInst(Pred, BOp1, SubC);
2653       } else if (C.isNullValue()) {
2654         // Replace ((sub A, B) != 0) with (A != B).
2655         return new ICmpInst(Pred, BOp0, BOp1);
2656       }
2657     }
2658     break;
2659   case Instruction::Or: {
2660     const APInt *BOC;
2661     if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
2662       // Comparing if all bits outside of a constant mask are set?
2663       // Replace (X | C) == -1 with (X & ~C) == ~C.
2664       // This removes the -1 constant.
2665       Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
2666       Value *And = Builder.CreateAnd(BOp0, NotBOC);
2667       return new ICmpInst(Pred, And, NotBOC);
2668     }
2669     break;
2670   }
2671   case Instruction::And: {
2672     const APInt *BOC;
2673     if (match(BOp1, m_APInt(BOC))) {
2674       // If we have ((X & C) == C), turn it into ((X & C) != 0).
2675       if (C == *BOC && C.isPowerOf2())
2676         return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
2677                             BO, Constant::getNullValue(RHS->getType()));
2678 
2679       // Don't perform the following transforms if the AND has multiple uses
2680       if (!BO->hasOneUse())
2681         break;
2682 
2683       // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
2684       if (BOC->isSignMask()) {
2685         Constant *Zero = Constant::getNullValue(BOp0->getType());
2686         auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
2687         return new ICmpInst(NewPred, BOp0, Zero);
2688       }
2689 
2690       // ((X & ~7) == 0) --> X < 8
2691       if (C.isNullValue() && (~(*BOC) + 1).isPowerOf2()) {
2692         Constant *NegBOC = ConstantExpr::getNeg(cast<Constant>(BOp1));
2693         auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
2694         return new ICmpInst(NewPred, BOp0, NegBOC);
2695       }
2696     }
2697     break;
2698   }
2699   case Instruction::Mul:
2700     if (C.isNullValue() && BO->hasNoSignedWrap()) {
2701       const APInt *BOC;
2702       if (match(BOp1, m_APInt(BOC)) && !BOC->isNullValue()) {
2703         // The trivial case (mul X, 0) is handled by InstSimplify.
2704         // General case : (mul X, C) != 0 iff X != 0
2705         //                (mul X, C) == 0 iff X == 0
2706         return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType()));
2707       }
2708     }
2709     break;
2710   case Instruction::UDiv:
2711     if (C.isNullValue()) {
2712       // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
2713       auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
2714       return new ICmpInst(NewPred, BOp1, BOp0);
2715     }
2716     break;
2717   default:
2718     break;
2719   }
2720   return nullptr;
2721 }
2722 
2723 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
foldICmpIntrinsicWithConstant(ICmpInst & Cmp,const APInt & C)2724 Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
2725                                                          const APInt &C) {
2726   IntrinsicInst *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0));
2727   if (!II || !Cmp.isEquality())
2728     return nullptr;
2729 
2730   // Handle icmp {eq|ne} <intrinsic>, Constant.
2731   Type *Ty = II->getType();
2732   switch (II->getIntrinsicID()) {
2733   case Intrinsic::bswap:
2734     Worklist.Add(II);
2735     Cmp.setOperand(0, II->getArgOperand(0));
2736     Cmp.setOperand(1, ConstantInt::get(Ty, C.byteSwap()));
2737     return &Cmp;
2738 
2739   case Intrinsic::ctlz:
2740   case Intrinsic::cttz:
2741     // ctz(A) == bitwidth(A)  ->  A == 0 and likewise for !=
2742     if (C == C.getBitWidth()) {
2743       Worklist.Add(II);
2744       Cmp.setOperand(0, II->getArgOperand(0));
2745       Cmp.setOperand(1, ConstantInt::getNullValue(Ty));
2746       return &Cmp;
2747     }
2748     break;
2749 
2750   case Intrinsic::ctpop: {
2751     // popcount(A) == 0  ->  A == 0 and likewise for !=
2752     // popcount(A) == bitwidth(A)  ->  A == -1 and likewise for !=
2753     bool IsZero = C.isNullValue();
2754     if (IsZero || C == C.getBitWidth()) {
2755       Worklist.Add(II);
2756       Cmp.setOperand(0, II->getArgOperand(0));
2757       auto *NewOp =
2758           IsZero ? Constant::getNullValue(Ty) : Constant::getAllOnesValue(Ty);
2759       Cmp.setOperand(1, NewOp);
2760       return &Cmp;
2761     }
2762     break;
2763   }
2764   default:
2765     break;
2766   }
2767 
2768   return nullptr;
2769 }
2770 
2771 /// Handle icmp with constant (but not simple integer constant) RHS.
foldICmpInstWithConstantNotInt(ICmpInst & I)2772 Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) {
2773   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2774   Constant *RHSC = dyn_cast<Constant>(Op1);
2775   Instruction *LHSI = dyn_cast<Instruction>(Op0);
2776   if (!RHSC || !LHSI)
2777     return nullptr;
2778 
2779   switch (LHSI->getOpcode()) {
2780   case Instruction::GetElementPtr:
2781     // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2782     if (RHSC->isNullValue() &&
2783         cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2784       return new ICmpInst(
2785           I.getPredicate(), LHSI->getOperand(0),
2786           Constant::getNullValue(LHSI->getOperand(0)->getType()));
2787     break;
2788   case Instruction::PHI:
2789     // Only fold icmp into the PHI if the phi and icmp are in the same
2790     // block.  If in the same block, we're encouraging jump threading.  If
2791     // not, we are just pessimizing the code by making an i1 phi.
2792     if (LHSI->getParent() == I.getParent())
2793       if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
2794         return NV;
2795     break;
2796   case Instruction::Select: {
2797     // If either operand of the select is a constant, we can fold the
2798     // comparison into the select arms, which will cause one to be
2799     // constant folded and the select turned into a bitwise or.
2800     Value *Op1 = nullptr, *Op2 = nullptr;
2801     ConstantInt *CI = nullptr;
2802     if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2803       Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2804       CI = dyn_cast<ConstantInt>(Op1);
2805     }
2806     if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2807       Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2808       CI = dyn_cast<ConstantInt>(Op2);
2809     }
2810 
2811     // We only want to perform this transformation if it will not lead to
2812     // additional code. This is true if either both sides of the select
2813     // fold to a constant (in which case the icmp is replaced with a select
2814     // which will usually simplify) or this is the only user of the
2815     // select (in which case we are trading a select+icmp for a simpler
2816     // select+icmp) or all uses of the select can be replaced based on
2817     // dominance information ("Global cases").
2818     bool Transform = false;
2819     if (Op1 && Op2)
2820       Transform = true;
2821     else if (Op1 || Op2) {
2822       // Local case
2823       if (LHSI->hasOneUse())
2824         Transform = true;
2825       // Global cases
2826       else if (CI && !CI->isZero())
2827         // When Op1 is constant try replacing select with second operand.
2828         // Otherwise Op2 is constant and try replacing select with first
2829         // operand.
2830         Transform =
2831             replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
2832     }
2833     if (Transform) {
2834       if (!Op1)
2835         Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
2836                                  I.getName());
2837       if (!Op2)
2838         Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
2839                                  I.getName());
2840       return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2841     }
2842     break;
2843   }
2844   case Instruction::IntToPtr:
2845     // icmp pred inttoptr(X), null -> icmp pred X, 0
2846     if (RHSC->isNullValue() &&
2847         DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
2848       return new ICmpInst(
2849           I.getPredicate(), LHSI->getOperand(0),
2850           Constant::getNullValue(LHSI->getOperand(0)->getType()));
2851     break;
2852 
2853   case Instruction::Load:
2854     // Try to optimize things like "A[i] > 4" to index computations.
2855     if (GetElementPtrInst *GEP =
2856             dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2857       if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2858         if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2859             !cast<LoadInst>(LHSI)->isVolatile())
2860           if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
2861             return Res;
2862     }
2863     break;
2864   }
2865 
2866   return nullptr;
2867 }
2868 
2869 /// Some comparisons can be simplified.
2870 /// In this case, we are looking for comparisons that look like
2871 /// a check for a lossy truncation.
2872 /// Folds:
2873 ///   x & (-1 >> y) SrcPred x    to    x DstPred (-1 >> y)
2874 /// The Mask can be a constant, too.
2875 /// For some predicates, the operands are commutative.
2876 /// For others, x can only be on a specific side.
foldICmpWithLowBitMaskedVal(ICmpInst & I,InstCombiner::BuilderTy & Builder)2877 static Value *foldICmpWithLowBitMaskedVal(ICmpInst &I,
2878                                           InstCombiner::BuilderTy &Builder) {
2879   ICmpInst::Predicate SrcPred;
2880   Value *X, *M;
2881   auto m_Mask = m_CombineOr(m_LShr(m_AllOnes(), m_Value()), m_LowBitMask());
2882   if (!match(&I, m_c_ICmp(SrcPred,
2883                           m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
2884                           m_Deferred(X))))
2885     return nullptr;
2886 
2887   ICmpInst::Predicate DstPred;
2888   switch (SrcPred) {
2889   case ICmpInst::Predicate::ICMP_EQ:
2890     //  x & (-1 >> y) == x    ->    x u<= (-1 >> y)
2891     DstPred = ICmpInst::Predicate::ICMP_ULE;
2892     break;
2893   case ICmpInst::Predicate::ICMP_NE:
2894     //  x & (-1 >> y) != x    ->    x u> (-1 >> y)
2895     DstPred = ICmpInst::Predicate::ICMP_UGT;
2896     break;
2897   case ICmpInst::Predicate::ICMP_UGT:
2898     //  x u> x & (-1 >> y)    ->    x u> (-1 >> y)
2899     assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
2900     DstPred = ICmpInst::Predicate::ICMP_UGT;
2901     break;
2902   case ICmpInst::Predicate::ICMP_UGE:
2903     //  x & (-1 >> y) u>= x    ->    x u<= (-1 >> y)
2904     assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
2905     DstPred = ICmpInst::Predicate::ICMP_ULE;
2906     break;
2907   case ICmpInst::Predicate::ICMP_ULT:
2908     //  x & (-1 >> y) u< x    ->    x u> (-1 >> y)
2909     assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
2910     DstPred = ICmpInst::Predicate::ICMP_UGT;
2911     break;
2912   case ICmpInst::Predicate::ICMP_ULE:
2913     //  x u<= x & (-1 >> y)    ->    x u<= (-1 >> y)
2914     assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
2915     DstPred = ICmpInst::Predicate::ICMP_ULE;
2916     break;
2917   case ICmpInst::Predicate::ICMP_SGT:
2918     //  x s> x & (-1 >> y)    ->    x s> (-1 >> y)
2919     if (X != I.getOperand(0)) // X must be on LHS of comparison!
2920       return nullptr;         // Ignore the other case.
2921     DstPred = ICmpInst::Predicate::ICMP_SGT;
2922     break;
2923   case ICmpInst::Predicate::ICMP_SGE:
2924     //  x & (-1 >> y) s>= x    ->    x s<= (-1 >> y)
2925     if (X != I.getOperand(1)) // X must be on RHS of comparison!
2926       return nullptr;         // Ignore the other case.
2927     if (!match(M, m_Constant())) // Can not do this fold with non-constant.
2928       return nullptr;
2929     if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
2930       return nullptr;
2931     DstPred = ICmpInst::Predicate::ICMP_SLE;
2932     break;
2933   case ICmpInst::Predicate::ICMP_SLT:
2934     //  x & (-1 >> y) s< x    ->    x s> (-1 >> y)
2935     if (X != I.getOperand(1)) // X must be on RHS of comparison!
2936       return nullptr;         // Ignore the other case.
2937     if (!match(M, m_Constant())) // Can not do this fold with non-constant.
2938       return nullptr;
2939     if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
2940       return nullptr;
2941     DstPred = ICmpInst::Predicate::ICMP_SGT;
2942     break;
2943   case ICmpInst::Predicate::ICMP_SLE:
2944     //  x s<= x & (-1 >> y)    ->    x s<= (-1 >> y)
2945     if (X != I.getOperand(0)) // X must be on LHS of comparison!
2946       return nullptr;         // Ignore the other case.
2947     DstPred = ICmpInst::Predicate::ICMP_SLE;
2948     break;
2949   default:
2950     llvm_unreachable("All possible folds are handled.");
2951   }
2952 
2953   return Builder.CreateICmp(DstPred, X, M);
2954 }
2955 
2956 /// Some comparisons can be simplified.
2957 /// In this case, we are looking for comparisons that look like
2958 /// a check for a lossy signed truncation.
2959 /// Folds:   (MaskedBits is a constant.)
2960 ///   ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
2961 /// Into:
2962 ///   (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
2963 /// Where  KeptBits = bitwidth(%x) - MaskedBits
2964 static Value *
foldICmpWithTruncSignExtendedVal(ICmpInst & I,InstCombiner::BuilderTy & Builder)2965 foldICmpWithTruncSignExtendedVal(ICmpInst &I,
2966                                  InstCombiner::BuilderTy &Builder) {
2967   ICmpInst::Predicate SrcPred;
2968   Value *X;
2969   const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
2970   // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
2971   if (!match(&I, m_c_ICmp(SrcPred,
2972                           m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
2973                                           m_APInt(C1))),
2974                           m_Deferred(X))))
2975     return nullptr;
2976 
2977   // Potential handling of non-splats: for each element:
2978   //  * if both are undef, replace with constant 0.
2979   //    Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
2980   //  * if both are not undef, and are different, bailout.
2981   //  * else, only one is undef, then pick the non-undef one.
2982 
2983   // The shift amount must be equal.
2984   if (*C0 != *C1)
2985     return nullptr;
2986   const APInt &MaskedBits = *C0;
2987   assert(MaskedBits != 0 && "shift by zero should be folded away already.");
2988 
2989   ICmpInst::Predicate DstPred;
2990   switch (SrcPred) {
2991   case ICmpInst::Predicate::ICMP_EQ:
2992     // ((%x << MaskedBits) a>> MaskedBits) == %x
2993     //   =>
2994     // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
2995     DstPred = ICmpInst::Predicate::ICMP_ULT;
2996     break;
2997   case ICmpInst::Predicate::ICMP_NE:
2998     // ((%x << MaskedBits) a>> MaskedBits) != %x
2999     //   =>
3000     // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3001     DstPred = ICmpInst::Predicate::ICMP_UGE;
3002     break;
3003   // FIXME: are more folds possible?
3004   default:
3005     return nullptr;
3006   }
3007 
3008   auto *XType = X->getType();
3009   const unsigned XBitWidth = XType->getScalarSizeInBits();
3010   const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3011   assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3012 
3013   // KeptBits = bitwidth(%x) - MaskedBits
3014   const APInt KeptBits = BitWidth - MaskedBits;
3015   assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3016   // ICmpCst = (1 << KeptBits)
3017   const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3018   assert(ICmpCst.isPowerOf2());
3019   // AddCst = (1 << (KeptBits-1))
3020   const APInt AddCst = ICmpCst.lshr(1);
3021   assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3022 
3023   // T0 = add %x, AddCst
3024   Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3025   // T1 = T0 DstPred ICmpCst
3026   Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3027 
3028   return T1;
3029 }
3030 
3031 /// Try to fold icmp (binop), X or icmp X, (binop).
3032 /// TODO: A large part of this logic is duplicated in InstSimplify's
3033 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3034 /// duplication.
foldICmpBinOp(ICmpInst & I)3035 Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I) {
3036   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3037 
3038   // Special logic for binary operators.
3039   BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3040   BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3041   if (!BO0 && !BO1)
3042     return nullptr;
3043 
3044   const CmpInst::Predicate Pred = I.getPredicate();
3045   bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3046   if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3047     NoOp0WrapProblem =
3048         ICmpInst::isEquality(Pred) ||
3049         (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3050         (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3051   if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3052     NoOp1WrapProblem =
3053         ICmpInst::isEquality(Pred) ||
3054         (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3055         (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3056 
3057   // Analyze the case when either Op0 or Op1 is an add instruction.
3058   // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3059   Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3060   if (BO0 && BO0->getOpcode() == Instruction::Add) {
3061     A = BO0->getOperand(0);
3062     B = BO0->getOperand(1);
3063   }
3064   if (BO1 && BO1->getOpcode() == Instruction::Add) {
3065     C = BO1->getOperand(0);
3066     D = BO1->getOperand(1);
3067   }
3068 
3069   // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3070   if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3071     return new ICmpInst(Pred, A == Op1 ? B : A,
3072                         Constant::getNullValue(Op1->getType()));
3073 
3074   // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3075   if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3076     return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3077                         C == Op0 ? D : C);
3078 
3079   // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
3080   if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3081       NoOp1WrapProblem &&
3082       // Try not to increase register pressure.
3083       BO0->hasOneUse() && BO1->hasOneUse()) {
3084     // Determine Y and Z in the form icmp (X+Y), (X+Z).
3085     Value *Y, *Z;
3086     if (A == C) {
3087       // C + B == C + D  ->  B == D
3088       Y = B;
3089       Z = D;
3090     } else if (A == D) {
3091       // D + B == C + D  ->  B == C
3092       Y = B;
3093       Z = C;
3094     } else if (B == C) {
3095       // A + C == C + D  ->  A == D
3096       Y = A;
3097       Z = D;
3098     } else {
3099       assert(B == D);
3100       // A + D == C + D  ->  A == C
3101       Y = A;
3102       Z = C;
3103     }
3104     return new ICmpInst(Pred, Y, Z);
3105   }
3106 
3107   // icmp slt (X + -1), Y -> icmp sle X, Y
3108   if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3109       match(B, m_AllOnes()))
3110     return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3111 
3112   // icmp sge (X + -1), Y -> icmp sgt X, Y
3113   if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3114       match(B, m_AllOnes()))
3115     return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3116 
3117   // icmp sle (X + 1), Y -> icmp slt X, Y
3118   if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
3119     return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3120 
3121   // icmp sgt (X + 1), Y -> icmp sge X, Y
3122   if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
3123     return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3124 
3125   // icmp sgt X, (Y + -1) -> icmp sge X, Y
3126   if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
3127       match(D, m_AllOnes()))
3128     return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3129 
3130   // icmp sle X, (Y + -1) -> icmp slt X, Y
3131   if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
3132       match(D, m_AllOnes()))
3133     return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3134 
3135   // icmp sge X, (Y + 1) -> icmp sgt X, Y
3136   if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
3137     return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3138 
3139   // icmp slt X, (Y + 1) -> icmp sle X, Y
3140   if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
3141     return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3142 
3143   // TODO: The subtraction-related identities shown below also hold, but
3144   // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3145   // wouldn't happen even if they were implemented.
3146   //
3147   // icmp ult (X - 1), Y -> icmp ule X, Y
3148   // icmp uge (X - 1), Y -> icmp ugt X, Y
3149   // icmp ugt X, (Y - 1) -> icmp uge X, Y
3150   // icmp ule X, (Y - 1) -> icmp ult X, Y
3151 
3152   // icmp ule (X + 1), Y -> icmp ult X, Y
3153   if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
3154     return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3155 
3156   // icmp ugt (X + 1), Y -> icmp uge X, Y
3157   if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
3158     return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3159 
3160   // icmp uge X, (Y + 1) -> icmp ugt X, Y
3161   if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
3162     return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3163 
3164   // icmp ult X, (Y + 1) -> icmp ule X, Y
3165   if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
3166     return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3167 
3168   // if C1 has greater magnitude than C2:
3169   //  icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3170   //  s.t. C3 = C1 - C2
3171   //
3172   // if C2 has greater magnitude than C1:
3173   //  icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3174   //  s.t. C3 = C2 - C1
3175   if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3176       (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3177     if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3178       if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3179         const APInt &AP1 = C1->getValue();
3180         const APInt &AP2 = C2->getValue();
3181         if (AP1.isNegative() == AP2.isNegative()) {
3182           APInt AP1Abs = C1->getValue().abs();
3183           APInt AP2Abs = C2->getValue().abs();
3184           if (AP1Abs.uge(AP2Abs)) {
3185             ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3186             Value *NewAdd = Builder.CreateNSWAdd(A, C3);
3187             return new ICmpInst(Pred, NewAdd, C);
3188           } else {
3189             ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3190             Value *NewAdd = Builder.CreateNSWAdd(C, C3);
3191             return new ICmpInst(Pred, A, NewAdd);
3192           }
3193         }
3194       }
3195 
3196   // Analyze the case when either Op0 or Op1 is a sub instruction.
3197   // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3198   A = nullptr;
3199   B = nullptr;
3200   C = nullptr;
3201   D = nullptr;
3202   if (BO0 && BO0->getOpcode() == Instruction::Sub) {
3203     A = BO0->getOperand(0);
3204     B = BO0->getOperand(1);
3205   }
3206   if (BO1 && BO1->getOpcode() == Instruction::Sub) {
3207     C = BO1->getOperand(0);
3208     D = BO1->getOperand(1);
3209   }
3210 
3211   // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3212   if (A == Op1 && NoOp0WrapProblem)
3213     return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3214   // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3215   if (C == Op0 && NoOp1WrapProblem)
3216     return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3217 
3218   // (A - B) >u A --> A <u B
3219   if (A == Op1 && Pred == ICmpInst::ICMP_UGT)
3220     return new ICmpInst(ICmpInst::ICMP_ULT, A, B);
3221   // C <u (C - D) --> C <u D
3222   if (C == Op0 && Pred == ICmpInst::ICMP_ULT)
3223     return new ICmpInst(ICmpInst::ICMP_ULT, C, D);
3224 
3225   // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3226   if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3227       // Try not to increase register pressure.
3228       BO0->hasOneUse() && BO1->hasOneUse())
3229     return new ICmpInst(Pred, A, C);
3230   // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3231   if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3232       // Try not to increase register pressure.
3233       BO0->hasOneUse() && BO1->hasOneUse())
3234     return new ICmpInst(Pred, D, B);
3235 
3236   // icmp (0-X) < cst --> x > -cst
3237   if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3238     Value *X;
3239     if (match(BO0, m_Neg(m_Value(X))))
3240       if (Constant *RHSC = dyn_cast<Constant>(Op1))
3241         if (RHSC->isNotMinSignedValue())
3242           return new ICmpInst(I.getSwappedPredicate(), X,
3243                               ConstantExpr::getNeg(RHSC));
3244   }
3245 
3246   BinaryOperator *SRem = nullptr;
3247   // icmp (srem X, Y), Y
3248   if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
3249     SRem = BO0;
3250   // icmp Y, (srem X, Y)
3251   else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3252            Op0 == BO1->getOperand(1))
3253     SRem = BO1;
3254   if (SRem) {
3255     // We don't check hasOneUse to avoid increasing register pressure because
3256     // the value we use is the same value this instruction was already using.
3257     switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3258     default:
3259       break;
3260     case ICmpInst::ICMP_EQ:
3261       return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3262     case ICmpInst::ICMP_NE:
3263       return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3264     case ICmpInst::ICMP_SGT:
3265     case ICmpInst::ICMP_SGE:
3266       return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3267                           Constant::getAllOnesValue(SRem->getType()));
3268     case ICmpInst::ICMP_SLT:
3269     case ICmpInst::ICMP_SLE:
3270       return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3271                           Constant::getNullValue(SRem->getType()));
3272     }
3273   }
3274 
3275   if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
3276       BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
3277     switch (BO0->getOpcode()) {
3278     default:
3279       break;
3280     case Instruction::Add:
3281     case Instruction::Sub:
3282     case Instruction::Xor: {
3283       if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3284         return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3285 
3286       const APInt *C;
3287       if (match(BO0->getOperand(1), m_APInt(C))) {
3288         // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
3289         if (C->isSignMask()) {
3290           ICmpInst::Predicate NewPred =
3291               I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
3292           return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3293         }
3294 
3295         // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
3296         if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
3297           ICmpInst::Predicate NewPred =
3298               I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
3299           NewPred = I.getSwappedPredicate(NewPred);
3300           return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3301         }
3302       }
3303       break;
3304     }
3305     case Instruction::Mul: {
3306       if (!I.isEquality())
3307         break;
3308 
3309       const APInt *C;
3310       if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
3311           !C->isOneValue()) {
3312         // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
3313         // Mask = -1 >> count-trailing-zeros(C).
3314         if (unsigned TZs = C->countTrailingZeros()) {
3315           Constant *Mask = ConstantInt::get(
3316               BO0->getType(),
3317               APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
3318           Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
3319           Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
3320           return new ICmpInst(Pred, And1, And2);
3321         }
3322         // If there are no trailing zeros in the multiplier, just eliminate
3323         // the multiplies (no masking is needed):
3324         // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
3325         return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3326       }
3327       break;
3328     }
3329     case Instruction::UDiv:
3330     case Instruction::LShr:
3331       if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
3332         break;
3333       return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3334 
3335     case Instruction::SDiv:
3336       if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
3337         break;
3338       return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3339 
3340     case Instruction::AShr:
3341       if (!BO0->isExact() || !BO1->isExact())
3342         break;
3343       return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3344 
3345     case Instruction::Shl: {
3346       bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3347       bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3348       if (!NUW && !NSW)
3349         break;
3350       if (!NSW && I.isSigned())
3351         break;
3352       return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3353     }
3354     }
3355   }
3356 
3357   if (BO0) {
3358     // Transform  A & (L - 1) `ult` L --> L != 0
3359     auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
3360     auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
3361 
3362     if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
3363       auto *Zero = Constant::getNullValue(BO0->getType());
3364       return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
3365     }
3366   }
3367 
3368   if (Value *V = foldICmpWithLowBitMaskedVal(I, Builder))
3369     return replaceInstUsesWith(I, V);
3370 
3371   if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
3372     return replaceInstUsesWith(I, V);
3373 
3374   return nullptr;
3375 }
3376 
3377 /// Fold icmp Pred min|max(X, Y), X.
foldICmpWithMinMax(ICmpInst & Cmp)3378 static Instruction *foldICmpWithMinMax(ICmpInst &Cmp) {
3379   ICmpInst::Predicate Pred = Cmp.getPredicate();
3380   Value *Op0 = Cmp.getOperand(0);
3381   Value *X = Cmp.getOperand(1);
3382 
3383   // Canonicalize minimum or maximum operand to LHS of the icmp.
3384   if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
3385       match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
3386       match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
3387       match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
3388     std::swap(Op0, X);
3389     Pred = Cmp.getSwappedPredicate();
3390   }
3391 
3392   Value *Y;
3393   if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
3394     // smin(X, Y)  == X --> X s<= Y
3395     // smin(X, Y) s>= X --> X s<= Y
3396     if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
3397       return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
3398 
3399     // smin(X, Y) != X --> X s> Y
3400     // smin(X, Y) s< X --> X s> Y
3401     if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
3402       return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
3403 
3404     // These cases should be handled in InstSimplify:
3405     // smin(X, Y) s<= X --> true
3406     // smin(X, Y) s> X --> false
3407     return nullptr;
3408   }
3409 
3410   if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
3411     // smax(X, Y)  == X --> X s>= Y
3412     // smax(X, Y) s<= X --> X s>= Y
3413     if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
3414       return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
3415 
3416     // smax(X, Y) != X --> X s< Y
3417     // smax(X, Y) s> X --> X s< Y
3418     if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
3419       return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
3420 
3421     // These cases should be handled in InstSimplify:
3422     // smax(X, Y) s>= X --> true
3423     // smax(X, Y) s< X --> false
3424     return nullptr;
3425   }
3426 
3427   if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
3428     // umin(X, Y)  == X --> X u<= Y
3429     // umin(X, Y) u>= X --> X u<= Y
3430     if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
3431       return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
3432 
3433     // umin(X, Y) != X --> X u> Y
3434     // umin(X, Y) u< X --> X u> Y
3435     if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
3436       return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
3437 
3438     // These cases should be handled in InstSimplify:
3439     // umin(X, Y) u<= X --> true
3440     // umin(X, Y) u> X --> false
3441     return nullptr;
3442   }
3443 
3444   if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
3445     // umax(X, Y)  == X --> X u>= Y
3446     // umax(X, Y) u<= X --> X u>= Y
3447     if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
3448       return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
3449 
3450     // umax(X, Y) != X --> X u< Y
3451     // umax(X, Y) u> X --> X u< Y
3452     if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
3453       return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
3454 
3455     // These cases should be handled in InstSimplify:
3456     // umax(X, Y) u>= X --> true
3457     // umax(X, Y) u< X --> false
3458     return nullptr;
3459   }
3460 
3461   return nullptr;
3462 }
3463 
foldICmpEquality(ICmpInst & I)3464 Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
3465   if (!I.isEquality())
3466     return nullptr;
3467 
3468   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3469   const CmpInst::Predicate Pred = I.getPredicate();
3470   Value *A, *B, *C, *D;
3471   if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3472     if (A == Op1 || B == Op1) { // (A^B) == A  ->  B == 0
3473       Value *OtherVal = A == Op1 ? B : A;
3474       return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3475     }
3476 
3477     if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3478       // A^c1 == C^c2 --> A == C^(c1^c2)
3479       ConstantInt *C1, *C2;
3480       if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
3481           Op1->hasOneUse()) {
3482         Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
3483         Value *Xor = Builder.CreateXor(C, NC);
3484         return new ICmpInst(Pred, A, Xor);
3485       }
3486 
3487       // A^B == A^D -> B == D
3488       if (A == C)
3489         return new ICmpInst(Pred, B, D);
3490       if (A == D)
3491         return new ICmpInst(Pred, B, C);
3492       if (B == C)
3493         return new ICmpInst(Pred, A, D);
3494       if (B == D)
3495         return new ICmpInst(Pred, A, C);
3496     }
3497   }
3498 
3499   if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
3500     // A == (A^B)  ->  B == 0
3501     Value *OtherVal = A == Op0 ? B : A;
3502     return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3503   }
3504 
3505   // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3506   if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3507       match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3508     Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3509 
3510     if (A == C) {
3511       X = B;
3512       Y = D;
3513       Z = A;
3514     } else if (A == D) {
3515       X = B;
3516       Y = C;
3517       Z = A;
3518     } else if (B == C) {
3519       X = A;
3520       Y = D;
3521       Z = B;
3522     } else if (B == D) {
3523       X = A;
3524       Y = C;
3525       Z = B;
3526     }
3527 
3528     if (X) { // Build (X^Y) & Z
3529       Op1 = Builder.CreateXor(X, Y);
3530       Op1 = Builder.CreateAnd(Op1, Z);
3531       I.setOperand(0, Op1);
3532       I.setOperand(1, Constant::getNullValue(Op1->getType()));
3533       return &I;
3534     }
3535   }
3536 
3537   // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3538   // and       (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3539   ConstantInt *Cst1;
3540   if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
3541        match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3542       (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3543        match(Op1, m_ZExt(m_Value(A))))) {
3544     APInt Pow2 = Cst1->getValue() + 1;
3545     if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3546         Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3547       return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
3548   }
3549 
3550   // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3551   // For lshr and ashr pairs.
3552   if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3553        match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3554       (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3555        match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3556     unsigned TypeBits = Cst1->getBitWidth();
3557     unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3558     if (ShAmt < TypeBits && ShAmt != 0) {
3559       ICmpInst::Predicate NewPred =
3560           Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
3561       Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3562       APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3563       return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
3564     }
3565   }
3566 
3567   // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
3568   if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
3569       match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
3570     unsigned TypeBits = Cst1->getBitWidth();
3571     unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3572     if (ShAmt < TypeBits && ShAmt != 0) {
3573       Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3574       APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
3575       Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
3576                                       I.getName() + ".mask");
3577       return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
3578     }
3579   }
3580 
3581   // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3582   // "icmp (and X, mask), cst"
3583   uint64_t ShAmt = 0;
3584   if (Op0->hasOneUse() &&
3585       match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
3586       match(Op1, m_ConstantInt(Cst1)) &&
3587       // Only do this when A has multiple uses.  This is most important to do
3588       // when it exposes other optimizations.
3589       !A->hasOneUse()) {
3590     unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3591 
3592     if (ShAmt < ASize) {
3593       APInt MaskV =
3594           APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
3595       MaskV <<= ShAmt;
3596 
3597       APInt CmpV = Cst1->getValue().zext(ASize);
3598       CmpV <<= ShAmt;
3599 
3600       Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
3601       return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
3602     }
3603   }
3604 
3605   // If both operands are byte-swapped or bit-reversed, just compare the
3606   // original values.
3607   // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
3608   // and handle more intrinsics.
3609   if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
3610       (match(Op0, m_BitReverse(m_Value(A))) &&
3611        match(Op1, m_BitReverse(m_Value(B)))))
3612     return new ICmpInst(Pred, A, B);
3613 
3614   return nullptr;
3615 }
3616 
3617 /// Handle icmp (cast x to y), (cast/cst). We only handle extending casts so
3618 /// far.
foldICmpWithCastAndCast(ICmpInst & ICmp)3619 Instruction *InstCombiner::foldICmpWithCastAndCast(ICmpInst &ICmp) {
3620   const CastInst *LHSCI = cast<CastInst>(ICmp.getOperand(0));
3621   Value *LHSCIOp        = LHSCI->getOperand(0);
3622   Type *SrcTy     = LHSCIOp->getType();
3623   Type *DestTy    = LHSCI->getType();
3624   Value *RHSCIOp;
3625 
3626   // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
3627   // integer type is the same size as the pointer type.
3628   const auto& CompatibleSizes = [&](Type* SrcTy, Type* DestTy) -> bool {
3629     if (isa<VectorType>(SrcTy)) {
3630       SrcTy = cast<VectorType>(SrcTy)->getElementType();
3631       DestTy = cast<VectorType>(DestTy)->getElementType();
3632     }
3633     return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
3634   };
3635   if (LHSCI->getOpcode() == Instruction::PtrToInt &&
3636       CompatibleSizes(SrcTy, DestTy)) {
3637     Value *RHSOp = nullptr;
3638     if (auto *RHSC = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
3639       Value *RHSCIOp = RHSC->getOperand(0);
3640       if (RHSCIOp->getType()->getPointerAddressSpace() ==
3641           LHSCIOp->getType()->getPointerAddressSpace()) {
3642         RHSOp = RHSC->getOperand(0);
3643         // If the pointer types don't match, insert a bitcast.
3644         if (LHSCIOp->getType() != RHSOp->getType())
3645           RHSOp = Builder.CreateBitCast(RHSOp, LHSCIOp->getType());
3646       }
3647     } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
3648       RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
3649     }
3650 
3651     if (RHSOp)
3652       return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSOp);
3653   }
3654 
3655   // The code below only handles extension cast instructions, so far.
3656   // Enforce this.
3657   if (LHSCI->getOpcode() != Instruction::ZExt &&
3658       LHSCI->getOpcode() != Instruction::SExt)
3659     return nullptr;
3660 
3661   bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
3662   bool isSignedCmp = ICmp.isSigned();
3663 
3664   if (auto *CI = dyn_cast<CastInst>(ICmp.getOperand(1))) {
3665     // Not an extension from the same type?
3666     RHSCIOp = CI->getOperand(0);
3667     if (RHSCIOp->getType() != LHSCIOp->getType())
3668       return nullptr;
3669 
3670     // If the signedness of the two casts doesn't agree (i.e. one is a sext
3671     // and the other is a zext), then we can't handle this.
3672     if (CI->getOpcode() != LHSCI->getOpcode())
3673       return nullptr;
3674 
3675     // Deal with equality cases early.
3676     if (ICmp.isEquality())
3677       return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3678 
3679     // A signed comparison of sign extended values simplifies into a
3680     // signed comparison.
3681     if (isSignedCmp && isSignedExt)
3682       return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3683 
3684     // The other three cases all fold into an unsigned comparison.
3685     return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
3686   }
3687 
3688   // If we aren't dealing with a constant on the RHS, exit early.
3689   auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
3690   if (!C)
3691     return nullptr;
3692 
3693   // Compute the constant that would happen if we truncated to SrcTy then
3694   // re-extended to DestTy.
3695   Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
3696   Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
3697 
3698   // If the re-extended constant didn't change...
3699   if (Res2 == C) {
3700     // Deal with equality cases early.
3701     if (ICmp.isEquality())
3702       return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3703 
3704     // A signed comparison of sign extended values simplifies into a
3705     // signed comparison.
3706     if (isSignedExt && isSignedCmp)
3707       return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3708 
3709     // The other three cases all fold into an unsigned comparison.
3710     return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, Res1);
3711   }
3712 
3713   // The re-extended constant changed, partly changed (in the case of a vector),
3714   // or could not be determined to be equal (in the case of a constant
3715   // expression), so the constant cannot be represented in the shorter type.
3716   // Consequently, we cannot emit a simple comparison.
3717   // All the cases that fold to true or false will have already been handled
3718   // by SimplifyICmpInst, so only deal with the tricky case.
3719 
3720   if (isSignedCmp || !isSignedExt || !isa<ConstantInt>(C))
3721     return nullptr;
3722 
3723   // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
3724   // should have been folded away previously and not enter in here.
3725 
3726   // We're performing an unsigned comp with a sign extended value.
3727   // This is true if the input is >= 0. [aka >s -1]
3728   Constant *NegOne = Constant::getAllOnesValue(SrcTy);
3729   Value *Result = Builder.CreateICmpSGT(LHSCIOp, NegOne, ICmp.getName());
3730 
3731   // Finally, return the value computed.
3732   if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
3733     return replaceInstUsesWith(ICmp, Result);
3734 
3735   assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
3736   return BinaryOperator::CreateNot(Result);
3737 }
3738 
OptimizeOverflowCheck(OverflowCheckFlavor OCF,Value * LHS,Value * RHS,Instruction & OrigI,Value * & Result,Constant * & Overflow)3739 bool InstCombiner::OptimizeOverflowCheck(OverflowCheckFlavor OCF, Value *LHS,
3740                                          Value *RHS, Instruction &OrigI,
3741                                          Value *&Result, Constant *&Overflow) {
3742   if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
3743     std::swap(LHS, RHS);
3744 
3745   auto SetResult = [&](Value *OpResult, Constant *OverflowVal, bool ReuseName) {
3746     Result = OpResult;
3747     Overflow = OverflowVal;
3748     if (ReuseName)
3749       Result->takeName(&OrigI);
3750     return true;
3751   };
3752 
3753   // If the overflow check was an add followed by a compare, the insertion point
3754   // may be pointing to the compare.  We want to insert the new instructions
3755   // before the add in case there are uses of the add between the add and the
3756   // compare.
3757   Builder.SetInsertPoint(&OrigI);
3758 
3759   switch (OCF) {
3760   case OCF_INVALID:
3761     llvm_unreachable("bad overflow check kind!");
3762 
3763   case OCF_UNSIGNED_ADD: {
3764     OverflowResult OR = computeOverflowForUnsignedAdd(LHS, RHS, &OrigI);
3765     if (OR == OverflowResult::NeverOverflows)
3766       return SetResult(Builder.CreateNUWAdd(LHS, RHS), Builder.getFalse(),
3767                        true);
3768 
3769     if (OR == OverflowResult::AlwaysOverflows)
3770       return SetResult(Builder.CreateAdd(LHS, RHS), Builder.getTrue(), true);
3771 
3772     // Fall through uadd into sadd
3773     LLVM_FALLTHROUGH;
3774   }
3775   case OCF_SIGNED_ADD: {
3776     // X + 0 -> {X, false}
3777     if (match(RHS, m_Zero()))
3778       return SetResult(LHS, Builder.getFalse(), false);
3779 
3780     // We can strength reduce this signed add into a regular add if we can prove
3781     // that it will never overflow.
3782     if (OCF == OCF_SIGNED_ADD)
3783       if (willNotOverflowSignedAdd(LHS, RHS, OrigI))
3784         return SetResult(Builder.CreateNSWAdd(LHS, RHS), Builder.getFalse(),
3785                          true);
3786     break;
3787   }
3788 
3789   case OCF_UNSIGNED_SUB:
3790   case OCF_SIGNED_SUB: {
3791     // X - 0 -> {X, false}
3792     if (match(RHS, m_Zero()))
3793       return SetResult(LHS, Builder.getFalse(), false);
3794 
3795     if (OCF == OCF_SIGNED_SUB) {
3796       if (willNotOverflowSignedSub(LHS, RHS, OrigI))
3797         return SetResult(Builder.CreateNSWSub(LHS, RHS), Builder.getFalse(),
3798                          true);
3799     } else {
3800       if (willNotOverflowUnsignedSub(LHS, RHS, OrigI))
3801         return SetResult(Builder.CreateNUWSub(LHS, RHS), Builder.getFalse(),
3802                          true);
3803     }
3804     break;
3805   }
3806 
3807   case OCF_UNSIGNED_MUL: {
3808     OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, &OrigI);
3809     if (OR == OverflowResult::NeverOverflows)
3810       return SetResult(Builder.CreateNUWMul(LHS, RHS), Builder.getFalse(),
3811                        true);
3812     if (OR == OverflowResult::AlwaysOverflows)
3813       return SetResult(Builder.CreateMul(LHS, RHS), Builder.getTrue(), true);
3814     LLVM_FALLTHROUGH;
3815   }
3816   case OCF_SIGNED_MUL:
3817     // X * undef -> undef
3818     if (isa<UndefValue>(RHS))
3819       return SetResult(RHS, UndefValue::get(Builder.getInt1Ty()), false);
3820 
3821     // X * 0 -> {0, false}
3822     if (match(RHS, m_Zero()))
3823       return SetResult(RHS, Builder.getFalse(), false);
3824 
3825     // X * 1 -> {X, false}
3826     if (match(RHS, m_One()))
3827       return SetResult(LHS, Builder.getFalse(), false);
3828 
3829     if (OCF == OCF_SIGNED_MUL)
3830       if (willNotOverflowSignedMul(LHS, RHS, OrigI))
3831         return SetResult(Builder.CreateNSWMul(LHS, RHS), Builder.getFalse(),
3832                          true);
3833     break;
3834   }
3835 
3836   return false;
3837 }
3838 
3839 /// Recognize and process idiom involving test for multiplication
3840 /// overflow.
3841 ///
3842 /// The caller has matched a pattern of the form:
3843 ///   I = cmp u (mul(zext A, zext B), V
3844 /// The function checks if this is a test for overflow and if so replaces
3845 /// multiplication with call to 'mul.with.overflow' intrinsic.
3846 ///
3847 /// \param I Compare instruction.
3848 /// \param MulVal Result of 'mult' instruction.  It is one of the arguments of
3849 ///               the compare instruction.  Must be of integer type.
3850 /// \param OtherVal The other argument of compare instruction.
3851 /// \returns Instruction which must replace the compare instruction, NULL if no
3852 ///          replacement required.
processUMulZExtIdiom(ICmpInst & I,Value * MulVal,Value * OtherVal,InstCombiner & IC)3853 static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
3854                                          Value *OtherVal, InstCombiner &IC) {
3855   // Don't bother doing this transformation for pointers, don't do it for
3856   // vectors.
3857   if (!isa<IntegerType>(MulVal->getType()))
3858     return nullptr;
3859 
3860   assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
3861   assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
3862   auto *MulInstr = dyn_cast<Instruction>(MulVal);
3863   if (!MulInstr)
3864     return nullptr;
3865   assert(MulInstr->getOpcode() == Instruction::Mul);
3866 
3867   auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
3868        *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
3869   assert(LHS->getOpcode() == Instruction::ZExt);
3870   assert(RHS->getOpcode() == Instruction::ZExt);
3871   Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
3872 
3873   // Calculate type and width of the result produced by mul.with.overflow.
3874   Type *TyA = A->getType(), *TyB = B->getType();
3875   unsigned WidthA = TyA->getPrimitiveSizeInBits(),
3876            WidthB = TyB->getPrimitiveSizeInBits();
3877   unsigned MulWidth;
3878   Type *MulType;
3879   if (WidthB > WidthA) {
3880     MulWidth = WidthB;
3881     MulType = TyB;
3882   } else {
3883     MulWidth = WidthA;
3884     MulType = TyA;
3885   }
3886 
3887   // In order to replace the original mul with a narrower mul.with.overflow,
3888   // all uses must ignore upper bits of the product.  The number of used low
3889   // bits must be not greater than the width of mul.with.overflow.
3890   if (MulVal->hasNUsesOrMore(2))
3891     for (User *U : MulVal->users()) {
3892       if (U == &I)
3893         continue;
3894       if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
3895         // Check if truncation ignores bits above MulWidth.
3896         unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
3897         if (TruncWidth > MulWidth)
3898           return nullptr;
3899       } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
3900         // Check if AND ignores bits above MulWidth.
3901         if (BO->getOpcode() != Instruction::And)
3902           return nullptr;
3903         if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3904           const APInt &CVal = CI->getValue();
3905           if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
3906             return nullptr;
3907         } else {
3908           // In this case we could have the operand of the binary operation
3909           // being defined in another block, and performing the replacement
3910           // could break the dominance relation.
3911           return nullptr;
3912         }
3913       } else {
3914         // Other uses prohibit this transformation.
3915         return nullptr;
3916       }
3917     }
3918 
3919   // Recognize patterns
3920   switch (I.getPredicate()) {
3921   case ICmpInst::ICMP_EQ:
3922   case ICmpInst::ICMP_NE:
3923     // Recognize pattern:
3924     //   mulval = mul(zext A, zext B)
3925     //   cmp eq/neq mulval, zext trunc mulval
3926     if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
3927       if (Zext->hasOneUse()) {
3928         Value *ZextArg = Zext->getOperand(0);
3929         if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
3930           if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
3931             break; //Recognized
3932       }
3933 
3934     // Recognize pattern:
3935     //   mulval = mul(zext A, zext B)
3936     //   cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
3937     ConstantInt *CI;
3938     Value *ValToMask;
3939     if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
3940       if (ValToMask != MulVal)
3941         return nullptr;
3942       const APInt &CVal = CI->getValue() + 1;
3943       if (CVal.isPowerOf2()) {
3944         unsigned MaskWidth = CVal.logBase2();
3945         if (MaskWidth == MulWidth)
3946           break; // Recognized
3947       }
3948     }
3949     return nullptr;
3950 
3951   case ICmpInst::ICMP_UGT:
3952     // Recognize pattern:
3953     //   mulval = mul(zext A, zext B)
3954     //   cmp ugt mulval, max
3955     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3956       APInt MaxVal = APInt::getMaxValue(MulWidth);
3957       MaxVal = MaxVal.zext(CI->getBitWidth());
3958       if (MaxVal.eq(CI->getValue()))
3959         break; // Recognized
3960     }
3961     return nullptr;
3962 
3963   case ICmpInst::ICMP_UGE:
3964     // Recognize pattern:
3965     //   mulval = mul(zext A, zext B)
3966     //   cmp uge mulval, max+1
3967     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3968       APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
3969       if (MaxVal.eq(CI->getValue()))
3970         break; // Recognized
3971     }
3972     return nullptr;
3973 
3974   case ICmpInst::ICMP_ULE:
3975     // Recognize pattern:
3976     //   mulval = mul(zext A, zext B)
3977     //   cmp ule mulval, max
3978     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3979       APInt MaxVal = APInt::getMaxValue(MulWidth);
3980       MaxVal = MaxVal.zext(CI->getBitWidth());
3981       if (MaxVal.eq(CI->getValue()))
3982         break; // Recognized
3983     }
3984     return nullptr;
3985 
3986   case ICmpInst::ICMP_ULT:
3987     // Recognize pattern:
3988     //   mulval = mul(zext A, zext B)
3989     //   cmp ule mulval, max + 1
3990     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3991       APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
3992       if (MaxVal.eq(CI->getValue()))
3993         break; // Recognized
3994     }
3995     return nullptr;
3996 
3997   default:
3998     return nullptr;
3999   }
4000 
4001   InstCombiner::BuilderTy &Builder = IC.Builder;
4002   Builder.SetInsertPoint(MulInstr);
4003 
4004   // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4005   Value *MulA = A, *MulB = B;
4006   if (WidthA < MulWidth)
4007     MulA = Builder.CreateZExt(A, MulType);
4008   if (WidthB < MulWidth)
4009     MulB = Builder.CreateZExt(B, MulType);
4010   Value *F = Intrinsic::getDeclaration(I.getModule(),
4011                                        Intrinsic::umul_with_overflow, MulType);
4012   CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4013   IC.Worklist.Add(MulInstr);
4014 
4015   // If there are uses of mul result other than the comparison, we know that
4016   // they are truncation or binary AND. Change them to use result of
4017   // mul.with.overflow and adjust properly mask/size.
4018   if (MulVal->hasNUsesOrMore(2)) {
4019     Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4020     for (auto UI = MulVal->user_begin(), UE = MulVal->user_end(); UI != UE;) {
4021       User *U = *UI++;
4022       if (U == &I || U == OtherVal)
4023         continue;
4024       if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4025         if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
4026           IC.replaceInstUsesWith(*TI, Mul);
4027         else
4028           TI->setOperand(0, Mul);
4029       } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4030         assert(BO->getOpcode() == Instruction::And);
4031         // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
4032         ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
4033         APInt ShortMask = CI->getValue().trunc(MulWidth);
4034         Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
4035         Instruction *Zext =
4036             cast<Instruction>(Builder.CreateZExt(ShortAnd, BO->getType()));
4037         IC.Worklist.Add(Zext);
4038         IC.replaceInstUsesWith(*BO, Zext);
4039       } else {
4040         llvm_unreachable("Unexpected Binary operation");
4041       }
4042       IC.Worklist.Add(cast<Instruction>(U));
4043     }
4044   }
4045   if (isa<Instruction>(OtherVal))
4046     IC.Worklist.Add(cast<Instruction>(OtherVal));
4047 
4048   // The original icmp gets replaced with the overflow value, maybe inverted
4049   // depending on predicate.
4050   bool Inverse = false;
4051   switch (I.getPredicate()) {
4052   case ICmpInst::ICMP_NE:
4053     break;
4054   case ICmpInst::ICMP_EQ:
4055     Inverse = true;
4056     break;
4057   case ICmpInst::ICMP_UGT:
4058   case ICmpInst::ICMP_UGE:
4059     if (I.getOperand(0) == MulVal)
4060       break;
4061     Inverse = true;
4062     break;
4063   case ICmpInst::ICMP_ULT:
4064   case ICmpInst::ICMP_ULE:
4065     if (I.getOperand(1) == MulVal)
4066       break;
4067     Inverse = true;
4068     break;
4069   default:
4070     llvm_unreachable("Unexpected predicate");
4071   }
4072   if (Inverse) {
4073     Value *Res = Builder.CreateExtractValue(Call, 1);
4074     return BinaryOperator::CreateNot(Res);
4075   }
4076 
4077   return ExtractValueInst::Create(Call, 1);
4078 }
4079 
4080 /// When performing a comparison against a constant, it is possible that not all
4081 /// the bits in the LHS are demanded. This helper method computes the mask that
4082 /// IS demanded.
getDemandedBitsLHSMask(ICmpInst & I,unsigned BitWidth)4083 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
4084   const APInt *RHS;
4085   if (!match(I.getOperand(1), m_APInt(RHS)))
4086     return APInt::getAllOnesValue(BitWidth);
4087 
4088   // If this is a normal comparison, it demands all bits. If it is a sign bit
4089   // comparison, it only demands the sign bit.
4090   bool UnusedBit;
4091   if (isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
4092     return APInt::getSignMask(BitWidth);
4093 
4094   switch (I.getPredicate()) {
4095   // For a UGT comparison, we don't care about any bits that
4096   // correspond to the trailing ones of the comparand.  The value of these
4097   // bits doesn't impact the outcome of the comparison, because any value
4098   // greater than the RHS must differ in a bit higher than these due to carry.
4099   case ICmpInst::ICMP_UGT:
4100     return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
4101 
4102   // Similarly, for a ULT comparison, we don't care about the trailing zeros.
4103   // Any value less than the RHS must differ in a higher bit because of carries.
4104   case ICmpInst::ICMP_ULT:
4105     return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
4106 
4107   default:
4108     return APInt::getAllOnesValue(BitWidth);
4109   }
4110 }
4111 
4112 /// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
4113 /// should be swapped.
4114 /// The decision is based on how many times these two operands are reused
4115 /// as subtract operands and their positions in those instructions.
4116 /// The rationale is that several architectures use the same instruction for
4117 /// both subtract and cmp. Thus, it is better if the order of those operands
4118 /// match.
4119 /// \return true if Op0 and Op1 should be swapped.
swapMayExposeCSEOpportunities(const Value * Op0,const Value * Op1)4120 static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
4121   // Filter out pointer values as those cannot appear directly in subtract.
4122   // FIXME: we may want to go through inttoptrs or bitcasts.
4123   if (Op0->getType()->isPointerTy())
4124     return false;
4125   // If a subtract already has the same operands as a compare, swapping would be
4126   // bad. If a subtract has the same operands as a compare but in reverse order,
4127   // then swapping is good.
4128   int GoodToSwap = 0;
4129   for (const User *U : Op0->users()) {
4130     if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
4131       GoodToSwap++;
4132     else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
4133       GoodToSwap--;
4134   }
4135   return GoodToSwap > 0;
4136 }
4137 
4138 /// Check that one use is in the same block as the definition and all
4139 /// other uses are in blocks dominated by a given block.
4140 ///
4141 /// \param DI Definition
4142 /// \param UI Use
4143 /// \param DB Block that must dominate all uses of \p DI outside
4144 ///           the parent block
4145 /// \return true when \p UI is the only use of \p DI in the parent block
4146 /// and all other uses of \p DI are in blocks dominated by \p DB.
4147 ///
dominatesAllUses(const Instruction * DI,const Instruction * UI,const BasicBlock * DB) const4148 bool InstCombiner::dominatesAllUses(const Instruction *DI,
4149                                     const Instruction *UI,
4150                                     const BasicBlock *DB) const {
4151   assert(DI && UI && "Instruction not defined\n");
4152   // Ignore incomplete definitions.
4153   if (!DI->getParent())
4154     return false;
4155   // DI and UI must be in the same block.
4156   if (DI->getParent() != UI->getParent())
4157     return false;
4158   // Protect from self-referencing blocks.
4159   if (DI->getParent() == DB)
4160     return false;
4161   for (const User *U : DI->users()) {
4162     auto *Usr = cast<Instruction>(U);
4163     if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4164       return false;
4165   }
4166   return true;
4167 }
4168 
4169 /// Return true when the instruction sequence within a block is select-cmp-br.
isChainSelectCmpBranch(const SelectInst * SI)4170 static bool isChainSelectCmpBranch(const SelectInst *SI) {
4171   const BasicBlock *BB = SI->getParent();
4172   if (!BB)
4173     return false;
4174   auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4175   if (!BI || BI->getNumSuccessors() != 2)
4176     return false;
4177   auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4178   if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4179     return false;
4180   return true;
4181 }
4182 
4183 /// True when a select result is replaced by one of its operands
4184 /// in select-icmp sequence. This will eventually result in the elimination
4185 /// of the select.
4186 ///
4187 /// \param SI    Select instruction
4188 /// \param Icmp  Compare instruction
4189 /// \param SIOpd Operand that replaces the select
4190 ///
4191 /// Notes:
4192 /// - The replacement is global and requires dominator information
4193 /// - The caller is responsible for the actual replacement
4194 ///
4195 /// Example:
4196 ///
4197 /// entry:
4198 ///  %4 = select i1 %3, %C* %0, %C* null
4199 ///  %5 = icmp eq %C* %4, null
4200 ///  br i1 %5, label %9, label %7
4201 ///  ...
4202 ///  ; <label>:7                                       ; preds = %entry
4203 ///  %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4204 ///  ...
4205 ///
4206 /// can be transformed to
4207 ///
4208 ///  %5 = icmp eq %C* %0, null
4209 ///  %6 = select i1 %3, i1 %5, i1 true
4210 ///  br i1 %6, label %9, label %7
4211 ///  ...
4212 ///  ; <label>:7                                       ; preds = %entry
4213 ///  %8 = getelementptr inbounds %C* %0, i64 0, i32 0  // replace by %0!
4214 ///
4215 /// Similar when the first operand of the select is a constant or/and
4216 /// the compare is for not equal rather than equal.
4217 ///
4218 /// NOTE: The function is only called when the select and compare constants
4219 /// are equal, the optimization can work only for EQ predicates. This is not a
4220 /// major restriction since a NE compare should be 'normalized' to an equal
4221 /// compare, which usually happens in the combiner and test case
4222 /// select-cmp-br.ll checks for it.
replacedSelectWithOperand(SelectInst * SI,const ICmpInst * Icmp,const unsigned SIOpd)4223 bool InstCombiner::replacedSelectWithOperand(SelectInst *SI,
4224                                              const ICmpInst *Icmp,
4225                                              const unsigned SIOpd) {
4226   assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
4227   if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4228     BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4229     // The check for the single predecessor is not the best that can be
4230     // done. But it protects efficiently against cases like when SI's
4231     // home block has two successors, Succ and Succ1, and Succ1 predecessor
4232     // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4233     // replaced can be reached on either path. So the uniqueness check
4234     // guarantees that the path all uses of SI (outside SI's parent) are on
4235     // is disjoint from all other paths out of SI. But that information
4236     // is more expensive to compute, and the trade-off here is in favor
4237     // of compile-time. It should also be noticed that we check for a single
4238     // predecessor and not only uniqueness. This to handle the situation when
4239     // Succ and Succ1 points to the same basic block.
4240     if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
4241       NumSel++;
4242       SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
4243       return true;
4244     }
4245   }
4246   return false;
4247 }
4248 
4249 /// Try to fold the comparison based on range information we can get by checking
4250 /// whether bits are known to be zero or one in the inputs.
foldICmpUsingKnownBits(ICmpInst & I)4251 Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) {
4252   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4253   Type *Ty = Op0->getType();
4254   ICmpInst::Predicate Pred = I.getPredicate();
4255 
4256   // Get scalar or pointer size.
4257   unsigned BitWidth = Ty->isIntOrIntVectorTy()
4258                           ? Ty->getScalarSizeInBits()
4259                           : DL.getIndexTypeSizeInBits(Ty->getScalarType());
4260 
4261   if (!BitWidth)
4262     return nullptr;
4263 
4264   KnownBits Op0Known(BitWidth);
4265   KnownBits Op1Known(BitWidth);
4266 
4267   if (SimplifyDemandedBits(&I, 0,
4268                            getDemandedBitsLHSMask(I, BitWidth),
4269                            Op0Known, 0))
4270     return &I;
4271 
4272   if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
4273                            Op1Known, 0))
4274     return &I;
4275 
4276   // Given the known and unknown bits, compute a range that the LHS could be
4277   // in.  Compute the Min, Max and RHS values based on the known bits. For the
4278   // EQ and NE we use unsigned values.
4279   APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
4280   APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
4281   if (I.isSigned()) {
4282     computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4283     computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4284   } else {
4285     computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4286     computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4287   }
4288 
4289   // If Min and Max are known to be the same, then SimplifyDemandedBits figured
4290   // out that the LHS or RHS is a constant. Constant fold this now, so that
4291   // code below can assume that Min != Max.
4292   if (!isa<Constant>(Op0) && Op0Min == Op0Max)
4293     return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
4294   if (!isa<Constant>(Op1) && Op1Min == Op1Max)
4295     return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
4296 
4297   // Based on the range information we know about the LHS, see if we can
4298   // simplify this comparison.  For example, (x&4) < 8 is always true.
4299   switch (Pred) {
4300   default:
4301     llvm_unreachable("Unknown icmp opcode!");
4302   case ICmpInst::ICMP_EQ:
4303   case ICmpInst::ICMP_NE: {
4304     if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
4305       return Pred == CmpInst::ICMP_EQ
4306                  ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
4307                  : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4308     }
4309 
4310     // If all bits are known zero except for one, then we know at most one bit
4311     // is set. If the comparison is against zero, then this is a check to see if
4312     // *that* bit is set.
4313     APInt Op0KnownZeroInverted = ~Op0Known.Zero;
4314     if (Op1Known.isZero()) {
4315       // If the LHS is an AND with the same constant, look through it.
4316       Value *LHS = nullptr;
4317       const APInt *LHSC;
4318       if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
4319           *LHSC != Op0KnownZeroInverted)
4320         LHS = Op0;
4321 
4322       Value *X;
4323       if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
4324         APInt ValToCheck = Op0KnownZeroInverted;
4325         Type *XTy = X->getType();
4326         if (ValToCheck.isPowerOf2()) {
4327           // ((1 << X) & 8) == 0 -> X != 3
4328           // ((1 << X) & 8) != 0 -> X == 3
4329           auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4330           auto NewPred = ICmpInst::getInversePredicate(Pred);
4331           return new ICmpInst(NewPred, X, CmpC);
4332         } else if ((++ValToCheck).isPowerOf2()) {
4333           // ((1 << X) & 7) == 0 -> X >= 3
4334           // ((1 << X) & 7) != 0 -> X  < 3
4335           auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4336           auto NewPred =
4337               Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
4338           return new ICmpInst(NewPred, X, CmpC);
4339         }
4340       }
4341 
4342       // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
4343       const APInt *CI;
4344       if (Op0KnownZeroInverted.isOneValue() &&
4345           match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
4346         // ((8 >>u X) & 1) == 0 -> X != 3
4347         // ((8 >>u X) & 1) != 0 -> X == 3
4348         unsigned CmpVal = CI->countTrailingZeros();
4349         auto NewPred = ICmpInst::getInversePredicate(Pred);
4350         return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
4351       }
4352     }
4353     break;
4354   }
4355   case ICmpInst::ICMP_ULT: {
4356     if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
4357       return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4358     if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
4359       return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4360     if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
4361       return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4362 
4363     const APInt *CmpC;
4364     if (match(Op1, m_APInt(CmpC))) {
4365       // A <u C -> A == C-1 if min(A)+1 == C
4366       if (*CmpC == Op0Min + 1)
4367         return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4368                             ConstantInt::get(Op1->getType(), *CmpC - 1));
4369       // X <u C --> X == 0, if the number of zero bits in the bottom of X
4370       // exceeds the log2 of C.
4371       if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
4372         return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4373                             Constant::getNullValue(Op1->getType()));
4374     }
4375     break;
4376   }
4377   case ICmpInst::ICMP_UGT: {
4378     if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
4379       return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4380     if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
4381       return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4382     if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
4383       return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4384 
4385     const APInt *CmpC;
4386     if (match(Op1, m_APInt(CmpC))) {
4387       // A >u C -> A == C+1 if max(a)-1 == C
4388       if (*CmpC == Op0Max - 1)
4389         return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4390                             ConstantInt::get(Op1->getType(), *CmpC + 1));
4391       // X >u C --> X != 0, if the number of zero bits in the bottom of X
4392       // exceeds the log2 of C.
4393       if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
4394         return new ICmpInst(ICmpInst::ICMP_NE, Op0,
4395                             Constant::getNullValue(Op1->getType()));
4396     }
4397     break;
4398   }
4399   case ICmpInst::ICMP_SLT: {
4400     if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
4401       return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4402     if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
4403       return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4404     if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
4405       return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4406     const APInt *CmpC;
4407     if (match(Op1, m_APInt(CmpC))) {
4408       if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
4409         return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4410                             ConstantInt::get(Op1->getType(), *CmpC - 1));
4411     }
4412     break;
4413   }
4414   case ICmpInst::ICMP_SGT: {
4415     if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
4416       return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4417     if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
4418       return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4419     if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
4420       return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4421     const APInt *CmpC;
4422     if (match(Op1, m_APInt(CmpC))) {
4423       if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
4424         return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4425                             ConstantInt::get(Op1->getType(), *CmpC + 1));
4426     }
4427     break;
4428   }
4429   case ICmpInst::ICMP_SGE:
4430     assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
4431     if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
4432       return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4433     if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
4434       return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4435     if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
4436       return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4437     break;
4438   case ICmpInst::ICMP_SLE:
4439     assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
4440     if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
4441       return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4442     if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
4443       return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4444     if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
4445       return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4446     break;
4447   case ICmpInst::ICMP_UGE:
4448     assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
4449     if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
4450       return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4451     if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
4452       return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4453     if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
4454       return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4455     break;
4456   case ICmpInst::ICMP_ULE:
4457     assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
4458     if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
4459       return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4460     if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
4461       return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4462     if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
4463       return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4464     break;
4465   }
4466 
4467   // Turn a signed comparison into an unsigned one if both operands are known to
4468   // have the same sign.
4469   if (I.isSigned() &&
4470       ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
4471        (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
4472     return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
4473 
4474   return nullptr;
4475 }
4476 
4477 /// If we have an icmp le or icmp ge instruction with a constant operand, turn
4478 /// it into the appropriate icmp lt or icmp gt instruction. This transform
4479 /// allows them to be folded in visitICmpInst.
canonicalizeCmpWithConstant(ICmpInst & I)4480 static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
4481   ICmpInst::Predicate Pred = I.getPredicate();
4482   if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGE &&
4483       Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_UGE)
4484     return nullptr;
4485 
4486   Value *Op0 = I.getOperand(0);
4487   Value *Op1 = I.getOperand(1);
4488   auto *Op1C = dyn_cast<Constant>(Op1);
4489   if (!Op1C)
4490     return nullptr;
4491 
4492   // Check if the constant operand can be safely incremented/decremented without
4493   // overflowing/underflowing. For scalars, SimplifyICmpInst has already handled
4494   // the edge cases for us, so we just assert on them. For vectors, we must
4495   // handle the edge cases.
4496   Type *Op1Type = Op1->getType();
4497   bool IsSigned = I.isSigned();
4498   bool IsLE = (Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_ULE);
4499   auto *CI = dyn_cast<ConstantInt>(Op1C);
4500   if (CI) {
4501     // A <= MAX -> TRUE ; A >= MIN -> TRUE
4502     assert(IsLE ? !CI->isMaxValue(IsSigned) : !CI->isMinValue(IsSigned));
4503   } else if (Op1Type->isVectorTy()) {
4504     // TODO? If the edge cases for vectors were guaranteed to be handled as they
4505     // are for scalar, we could remove the min/max checks. However, to do that,
4506     // we would have to use insertelement/shufflevector to replace edge values.
4507     unsigned NumElts = Op1Type->getVectorNumElements();
4508     for (unsigned i = 0; i != NumElts; ++i) {
4509       Constant *Elt = Op1C->getAggregateElement(i);
4510       if (!Elt)
4511         return nullptr;
4512 
4513       if (isa<UndefValue>(Elt))
4514         continue;
4515 
4516       // Bail out if we can't determine if this constant is min/max or if we
4517       // know that this constant is min/max.
4518       auto *CI = dyn_cast<ConstantInt>(Elt);
4519       if (!CI || (IsLE ? CI->isMaxValue(IsSigned) : CI->isMinValue(IsSigned)))
4520         return nullptr;
4521     }
4522   } else {
4523     // ConstantExpr?
4524     return nullptr;
4525   }
4526 
4527   // Increment or decrement the constant and set the new comparison predicate:
4528   // ULE -> ULT ; UGE -> UGT ; SLE -> SLT ; SGE -> SGT
4529   Constant *OneOrNegOne = ConstantInt::get(Op1Type, IsLE ? 1 : -1, true);
4530   CmpInst::Predicate NewPred = IsLE ? ICmpInst::ICMP_ULT: ICmpInst::ICMP_UGT;
4531   NewPred = IsSigned ? ICmpInst::getSignedPredicate(NewPred) : NewPred;
4532   return new ICmpInst(NewPred, Op0, ConstantExpr::getAdd(Op1C, OneOrNegOne));
4533 }
4534 
4535 /// Integer compare with boolean values can always be turned into bitwise ops.
canonicalizeICmpBool(ICmpInst & I,InstCombiner::BuilderTy & Builder)4536 static Instruction *canonicalizeICmpBool(ICmpInst &I,
4537                                          InstCombiner::BuilderTy &Builder) {
4538   Value *A = I.getOperand(0), *B = I.getOperand(1);
4539   assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only");
4540 
4541   // A boolean compared to true/false can be simplified to Op0/true/false in
4542   // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
4543   // Cases not handled by InstSimplify are always 'not' of Op0.
4544   if (match(B, m_Zero())) {
4545     switch (I.getPredicate()) {
4546       case CmpInst::ICMP_EQ:  // A ==   0 -> !A
4547       case CmpInst::ICMP_ULE: // A <=u  0 -> !A
4548       case CmpInst::ICMP_SGE: // A >=s  0 -> !A
4549         return BinaryOperator::CreateNot(A);
4550       default:
4551         llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4552     }
4553   } else if (match(B, m_One())) {
4554     switch (I.getPredicate()) {
4555       case CmpInst::ICMP_NE:  // A !=  1 -> !A
4556       case CmpInst::ICMP_ULT: // A <u  1 -> !A
4557       case CmpInst::ICMP_SGT: // A >s -1 -> !A
4558         return BinaryOperator::CreateNot(A);
4559       default:
4560         llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4561     }
4562   }
4563 
4564   switch (I.getPredicate()) {
4565   default:
4566     llvm_unreachable("Invalid icmp instruction!");
4567   case ICmpInst::ICMP_EQ:
4568     // icmp eq i1 A, B -> ~(A ^ B)
4569     return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
4570 
4571   case ICmpInst::ICMP_NE:
4572     // icmp ne i1 A, B -> A ^ B
4573     return BinaryOperator::CreateXor(A, B);
4574 
4575   case ICmpInst::ICMP_UGT:
4576     // icmp ugt -> icmp ult
4577     std::swap(A, B);
4578     LLVM_FALLTHROUGH;
4579   case ICmpInst::ICMP_ULT:
4580     // icmp ult i1 A, B -> ~A & B
4581     return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
4582 
4583   case ICmpInst::ICMP_SGT:
4584     // icmp sgt -> icmp slt
4585     std::swap(A, B);
4586     LLVM_FALLTHROUGH;
4587   case ICmpInst::ICMP_SLT:
4588     // icmp slt i1 A, B -> A & ~B
4589     return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
4590 
4591   case ICmpInst::ICMP_UGE:
4592     // icmp uge -> icmp ule
4593     std::swap(A, B);
4594     LLVM_FALLTHROUGH;
4595   case ICmpInst::ICMP_ULE:
4596     // icmp ule i1 A, B -> ~A | B
4597     return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
4598 
4599   case ICmpInst::ICMP_SGE:
4600     // icmp sge -> icmp sle
4601     std::swap(A, B);
4602     LLVM_FALLTHROUGH;
4603   case ICmpInst::ICMP_SLE:
4604     // icmp sle i1 A, B -> A | ~B
4605     return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
4606   }
4607 }
4608 
visitICmpInst(ICmpInst & I)4609 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4610   bool Changed = false;
4611   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4612   unsigned Op0Cplxity = getComplexity(Op0);
4613   unsigned Op1Cplxity = getComplexity(Op1);
4614 
4615   /// Orders the operands of the compare so that they are listed from most
4616   /// complex to least complex.  This puts constants before unary operators,
4617   /// before binary operators.
4618   if (Op0Cplxity < Op1Cplxity ||
4619       (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
4620     I.swapOperands();
4621     std::swap(Op0, Op1);
4622     Changed = true;
4623   }
4624 
4625   if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1,
4626                                   SQ.getWithInstruction(&I)))
4627     return replaceInstUsesWith(I, V);
4628 
4629   // Comparing -val or val with non-zero is the same as just comparing val
4630   // ie, abs(val) != 0 -> val != 0
4631   if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
4632     Value *Cond, *SelectTrue, *SelectFalse;
4633     if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
4634                             m_Value(SelectFalse)))) {
4635       if (Value *V = dyn_castNegVal(SelectTrue)) {
4636         if (V == SelectFalse)
4637           return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4638       }
4639       else if (Value *V = dyn_castNegVal(SelectFalse)) {
4640         if (V == SelectTrue)
4641           return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4642       }
4643     }
4644   }
4645 
4646   if (Op0->getType()->isIntOrIntVectorTy(1))
4647     if (Instruction *Res = canonicalizeICmpBool(I, Builder))
4648       return Res;
4649 
4650   if (ICmpInst *NewICmp = canonicalizeCmpWithConstant(I))
4651     return NewICmp;
4652 
4653   if (Instruction *Res = foldICmpWithConstant(I))
4654     return Res;
4655 
4656   if (Instruction *Res = foldICmpUsingKnownBits(I))
4657     return Res;
4658 
4659   // Test if the ICmpInst instruction is used exclusively by a select as
4660   // part of a minimum or maximum operation. If so, refrain from doing
4661   // any other folding. This helps out other analyses which understand
4662   // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
4663   // and CodeGen. And in this case, at least one of the comparison
4664   // operands has at least one user besides the compare (the select),
4665   // which would often largely negate the benefit of folding anyway.
4666   //
4667   // Do the same for the other patterns recognized by matchSelectPattern.
4668   if (I.hasOneUse())
4669     if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
4670       Value *A, *B;
4671       SelectPatternResult SPR = matchSelectPattern(SI, A, B);
4672       if (SPR.Flavor != SPF_UNKNOWN)
4673         return nullptr;
4674     }
4675 
4676   // Do this after checking for min/max to prevent infinite looping.
4677   if (Instruction *Res = foldICmpWithZero(I))
4678     return Res;
4679 
4680   // FIXME: We only do this after checking for min/max to prevent infinite
4681   // looping caused by a reverse canonicalization of these patterns for min/max.
4682   // FIXME: The organization of folds is a mess. These would naturally go into
4683   // canonicalizeCmpWithConstant(), but we can't move all of the above folds
4684   // down here after the min/max restriction.
4685   ICmpInst::Predicate Pred = I.getPredicate();
4686   const APInt *C;
4687   if (match(Op1, m_APInt(C))) {
4688     // For i32: x >u 2147483647 -> x <s 0  -> true if sign bit set
4689     if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
4690       Constant *Zero = Constant::getNullValue(Op0->getType());
4691       return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
4692     }
4693 
4694     // For i32: x <u 2147483648 -> x >s -1  -> true if sign bit clear
4695     if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
4696       Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
4697       return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
4698     }
4699   }
4700 
4701   if (Instruction *Res = foldICmpInstWithConstant(I))
4702     return Res;
4703 
4704   if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
4705     return Res;
4706 
4707   // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4708   if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
4709     if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I))
4710       return NI;
4711   if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
4712     if (Instruction *NI = foldGEPICmp(GEP, Op0,
4713                            ICmpInst::getSwappedPredicate(I.getPredicate()), I))
4714       return NI;
4715 
4716   // Try to optimize equality comparisons against alloca-based pointers.
4717   if (Op0->getType()->isPointerTy() && I.isEquality()) {
4718     assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?");
4719     if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op0, DL)))
4720       if (Instruction *New = foldAllocaCmp(I, Alloca, Op1))
4721         return New;
4722     if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op1, DL)))
4723       if (Instruction *New = foldAllocaCmp(I, Alloca, Op0))
4724         return New;
4725   }
4726 
4727   // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
4728   Value *X;
4729   if (match(Op0, m_BitCast(m_SIToFP(m_Value(X))))) {
4730     // icmp  eq (bitcast (sitofp X)), 0 --> icmp  eq X, 0
4731     // icmp  ne (bitcast (sitofp X)), 0 --> icmp  ne X, 0
4732     // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
4733     // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
4734     if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
4735          Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
4736         match(Op1, m_Zero()))
4737       return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
4738 
4739     // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
4740     if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
4741       return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
4742 
4743     // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
4744     if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
4745       return new ICmpInst(Pred, X, ConstantInt::getAllOnesValue(X->getType()));
4746   }
4747 
4748   // Zero-equality checks are preserved through unsigned floating-point casts:
4749   // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
4750   // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
4751   if (match(Op0, m_BitCast(m_UIToFP(m_Value(X)))))
4752     if (I.isEquality() && match(Op1, m_Zero()))
4753       return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
4754 
4755   // Test to see if the operands of the icmp are casted versions of other
4756   // values.  If the ptr->ptr cast can be stripped off both arguments, we do so
4757   // now.
4758   if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
4759     if (Op0->getType()->isPointerTy() &&
4760         (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
4761       // We keep moving the cast from the left operand over to the right
4762       // operand, where it can often be eliminated completely.
4763       Op0 = CI->getOperand(0);
4764 
4765       // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
4766       // so eliminate it as well.
4767       if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
4768         Op1 = CI2->getOperand(0);
4769 
4770       // If Op1 is a constant, we can fold the cast into the constant.
4771       if (Op0->getType() != Op1->getType()) {
4772         if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
4773           Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
4774         } else {
4775           // Otherwise, cast the RHS right before the icmp
4776           Op1 = Builder.CreateBitCast(Op1, Op0->getType());
4777         }
4778       }
4779       return new ICmpInst(I.getPredicate(), Op0, Op1);
4780     }
4781   }
4782 
4783   if (isa<CastInst>(Op0)) {
4784     // Handle the special case of: icmp (cast bool to X), <cst>
4785     // This comes up when you have code like
4786     //   int X = A < B;
4787     //   if (X) ...
4788     // For generality, we handle any zero-extension of any operand comparison
4789     // with a constant or another cast from the same type.
4790     if (isa<Constant>(Op1) || isa<CastInst>(Op1))
4791       if (Instruction *R = foldICmpWithCastAndCast(I))
4792         return R;
4793   }
4794 
4795   if (Instruction *Res = foldICmpBinOp(I))
4796     return Res;
4797 
4798   if (Instruction *Res = foldICmpWithMinMax(I))
4799     return Res;
4800 
4801   {
4802     Value *A, *B;
4803     // Transform (A & ~B) == 0 --> (A & B) != 0
4804     // and       (A & ~B) != 0 --> (A & B) == 0
4805     // if A is a power of 2.
4806     if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
4807         match(Op1, m_Zero()) &&
4808         isKnownToBeAPowerOfTwo(A, false, 0, &I) && I.isEquality())
4809       return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(A, B),
4810                           Op1);
4811 
4812     // ~X < ~Y --> Y < X
4813     // ~X < C -->  X > ~C
4814     if (match(Op0, m_Not(m_Value(A)))) {
4815       if (match(Op1, m_Not(m_Value(B))))
4816         return new ICmpInst(I.getPredicate(), B, A);
4817 
4818       const APInt *C;
4819       if (match(Op1, m_APInt(C)))
4820         return new ICmpInst(I.getSwappedPredicate(), A,
4821                             ConstantInt::get(Op1->getType(), ~(*C)));
4822     }
4823 
4824     Instruction *AddI = nullptr;
4825     if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
4826                                      m_Instruction(AddI))) &&
4827         isa<IntegerType>(A->getType())) {
4828       Value *Result;
4829       Constant *Overflow;
4830       if (OptimizeOverflowCheck(OCF_UNSIGNED_ADD, A, B, *AddI, Result,
4831                                 Overflow)) {
4832         replaceInstUsesWith(*AddI, Result);
4833         return replaceInstUsesWith(I, Overflow);
4834       }
4835     }
4836 
4837     // (zext a) * (zext b)  --> llvm.umul.with.overflow.
4838     if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
4839       if (Instruction *R = processUMulZExtIdiom(I, Op0, Op1, *this))
4840         return R;
4841     }
4842     if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
4843       if (Instruction *R = processUMulZExtIdiom(I, Op1, Op0, *this))
4844         return R;
4845     }
4846   }
4847 
4848   if (Instruction *Res = foldICmpEquality(I))
4849     return Res;
4850 
4851   // The 'cmpxchg' instruction returns an aggregate containing the old value and
4852   // an i1 which indicates whether or not we successfully did the swap.
4853   //
4854   // Replace comparisons between the old value and the expected value with the
4855   // indicator that 'cmpxchg' returns.
4856   //
4857   // N.B.  This transform is only valid when the 'cmpxchg' is not permitted to
4858   // spuriously fail.  In those cases, the old value may equal the expected
4859   // value but it is possible for the swap to not occur.
4860   if (I.getPredicate() == ICmpInst::ICMP_EQ)
4861     if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
4862       if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
4863         if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
4864             !ACXI->isWeak())
4865           return ExtractValueInst::Create(ACXI, 1);
4866 
4867   {
4868     Value *X; ConstantInt *Cst;
4869     // icmp X+Cst, X
4870     if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
4871       return foldICmpAddOpConst(X, Cst, I.getPredicate());
4872 
4873     // icmp X, X+Cst
4874     if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
4875       return foldICmpAddOpConst(X, Cst, I.getSwappedPredicate());
4876   }
4877 
4878   return Changed ? &I : nullptr;
4879 }
4880 
4881 /// Fold fcmp ([us]itofp x, cst) if possible.
foldFCmpIntToFPConst(FCmpInst & I,Instruction * LHSI,Constant * RHSC)4882 Instruction *InstCombiner::foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI,
4883                                                 Constant *RHSC) {
4884   if (!isa<ConstantFP>(RHSC)) return nullptr;
4885   const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
4886 
4887   // Get the width of the mantissa.  We don't want to hack on conversions that
4888   // might lose information from the integer, e.g. "i64 -> float"
4889   int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
4890   if (MantissaWidth == -1) return nullptr;  // Unknown.
4891 
4892   IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
4893 
4894   bool LHSUnsigned = isa<UIToFPInst>(LHSI);
4895 
4896   if (I.isEquality()) {
4897     FCmpInst::Predicate P = I.getPredicate();
4898     bool IsExact = false;
4899     APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned);
4900     RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
4901 
4902     // If the floating point constant isn't an integer value, we know if we will
4903     // ever compare equal / not equal to it.
4904     if (!IsExact) {
4905       // TODO: Can never be -0.0 and other non-representable values
4906       APFloat RHSRoundInt(RHS);
4907       RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven);
4908       if (RHS.compare(RHSRoundInt) != APFloat::cmpEqual) {
4909         if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
4910           return replaceInstUsesWith(I, Builder.getFalse());
4911 
4912         assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
4913         return replaceInstUsesWith(I, Builder.getTrue());
4914       }
4915     }
4916 
4917     // TODO: If the constant is exactly representable, is it always OK to do
4918     // equality compares as integer?
4919   }
4920 
4921   // Check to see that the input is converted from an integer type that is small
4922   // enough that preserves all bits.  TODO: check here for "known" sign bits.
4923   // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
4924   unsigned InputSize = IntTy->getScalarSizeInBits();
4925 
4926   // Following test does NOT adjust InputSize downwards for signed inputs,
4927   // because the most negative value still requires all the mantissa bits
4928   // to distinguish it from one less than that value.
4929   if ((int)InputSize > MantissaWidth) {
4930     // Conversion would lose accuracy. Check if loss can impact comparison.
4931     int Exp = ilogb(RHS);
4932     if (Exp == APFloat::IEK_Inf) {
4933       int MaxExponent = ilogb(APFloat::getLargest(RHS.getSemantics()));
4934       if (MaxExponent < (int)InputSize - !LHSUnsigned)
4935         // Conversion could create infinity.
4936         return nullptr;
4937     } else {
4938       // Note that if RHS is zero or NaN, then Exp is negative
4939       // and first condition is trivially false.
4940       if (MantissaWidth <= Exp && Exp <= (int)InputSize - !LHSUnsigned)
4941         // Conversion could affect comparison.
4942         return nullptr;
4943     }
4944   }
4945 
4946   // Otherwise, we can potentially simplify the comparison.  We know that it
4947   // will always come through as an integer value and we know the constant is
4948   // not a NAN (it would have been previously simplified).
4949   assert(!RHS.isNaN() && "NaN comparison not already folded!");
4950 
4951   ICmpInst::Predicate Pred;
4952   switch (I.getPredicate()) {
4953   default: llvm_unreachable("Unexpected predicate!");
4954   case FCmpInst::FCMP_UEQ:
4955   case FCmpInst::FCMP_OEQ:
4956     Pred = ICmpInst::ICMP_EQ;
4957     break;
4958   case FCmpInst::FCMP_UGT:
4959   case FCmpInst::FCMP_OGT:
4960     Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
4961     break;
4962   case FCmpInst::FCMP_UGE:
4963   case FCmpInst::FCMP_OGE:
4964     Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
4965     break;
4966   case FCmpInst::FCMP_ULT:
4967   case FCmpInst::FCMP_OLT:
4968     Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
4969     break;
4970   case FCmpInst::FCMP_ULE:
4971   case FCmpInst::FCMP_OLE:
4972     Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
4973     break;
4974   case FCmpInst::FCMP_UNE:
4975   case FCmpInst::FCMP_ONE:
4976     Pred = ICmpInst::ICMP_NE;
4977     break;
4978   case FCmpInst::FCMP_ORD:
4979     return replaceInstUsesWith(I, Builder.getTrue());
4980   case FCmpInst::FCMP_UNO:
4981     return replaceInstUsesWith(I, Builder.getFalse());
4982   }
4983 
4984   // Now we know that the APFloat is a normal number, zero or inf.
4985 
4986   // See if the FP constant is too large for the integer.  For example,
4987   // comparing an i8 to 300.0.
4988   unsigned IntWidth = IntTy->getScalarSizeInBits();
4989 
4990   if (!LHSUnsigned) {
4991     // If the RHS value is > SignedMax, fold the comparison.  This handles +INF
4992     // and large values.
4993     APFloat SMax(RHS.getSemantics());
4994     SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
4995                           APFloat::rmNearestTiesToEven);
4996     if (SMax.compare(RHS) == APFloat::cmpLessThan) {  // smax < 13123.0
4997       if (Pred == ICmpInst::ICMP_NE  || Pred == ICmpInst::ICMP_SLT ||
4998           Pred == ICmpInst::ICMP_SLE)
4999         return replaceInstUsesWith(I, Builder.getTrue());
5000       return replaceInstUsesWith(I, Builder.getFalse());
5001     }
5002   } else {
5003     // If the RHS value is > UnsignedMax, fold the comparison. This handles
5004     // +INF and large values.
5005     APFloat UMax(RHS.getSemantics());
5006     UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5007                           APFloat::rmNearestTiesToEven);
5008     if (UMax.compare(RHS) == APFloat::cmpLessThan) {  // umax < 13123.0
5009       if (Pred == ICmpInst::ICMP_NE  || Pred == ICmpInst::ICMP_ULT ||
5010           Pred == ICmpInst::ICMP_ULE)
5011         return replaceInstUsesWith(I, Builder.getTrue());
5012       return replaceInstUsesWith(I, Builder.getFalse());
5013     }
5014   }
5015 
5016   if (!LHSUnsigned) {
5017     // See if the RHS value is < SignedMin.
5018     APFloat SMin(RHS.getSemantics());
5019     SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5020                           APFloat::rmNearestTiesToEven);
5021     if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5022       if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5023           Pred == ICmpInst::ICMP_SGE)
5024         return replaceInstUsesWith(I, Builder.getTrue());
5025       return replaceInstUsesWith(I, Builder.getFalse());
5026     }
5027   } else {
5028     // See if the RHS value is < UnsignedMin.
5029     APFloat SMin(RHS.getSemantics());
5030     SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
5031                           APFloat::rmNearestTiesToEven);
5032     if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
5033       if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
5034           Pred == ICmpInst::ICMP_UGE)
5035         return replaceInstUsesWith(I, Builder.getTrue());
5036       return replaceInstUsesWith(I, Builder.getFalse());
5037     }
5038   }
5039 
5040   // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5041   // [0, UMAX], but it may still be fractional.  See if it is fractional by
5042   // casting the FP value to the integer value and back, checking for equality.
5043   // Don't do this for zero, because -0.0 is not fractional.
5044   Constant *RHSInt = LHSUnsigned
5045     ? ConstantExpr::getFPToUI(RHSC, IntTy)
5046     : ConstantExpr::getFPToSI(RHSC, IntTy);
5047   if (!RHS.isZero()) {
5048     bool Equal = LHSUnsigned
5049       ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5050       : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5051     if (!Equal) {
5052       // If we had a comparison against a fractional value, we have to adjust
5053       // the compare predicate and sometimes the value.  RHSC is rounded towards
5054       // zero at this point.
5055       switch (Pred) {
5056       default: llvm_unreachable("Unexpected integer comparison!");
5057       case ICmpInst::ICMP_NE:  // (float)int != 4.4   --> true
5058         return replaceInstUsesWith(I, Builder.getTrue());
5059       case ICmpInst::ICMP_EQ:  // (float)int == 4.4   --> false
5060         return replaceInstUsesWith(I, Builder.getFalse());
5061       case ICmpInst::ICMP_ULE:
5062         // (float)int <= 4.4   --> int <= 4
5063         // (float)int <= -4.4  --> false
5064         if (RHS.isNegative())
5065           return replaceInstUsesWith(I, Builder.getFalse());
5066         break;
5067       case ICmpInst::ICMP_SLE:
5068         // (float)int <= 4.4   --> int <= 4
5069         // (float)int <= -4.4  --> int < -4
5070         if (RHS.isNegative())
5071           Pred = ICmpInst::ICMP_SLT;
5072         break;
5073       case ICmpInst::ICMP_ULT:
5074         // (float)int < -4.4   --> false
5075         // (float)int < 4.4    --> int <= 4
5076         if (RHS.isNegative())
5077           return replaceInstUsesWith(I, Builder.getFalse());
5078         Pred = ICmpInst::ICMP_ULE;
5079         break;
5080       case ICmpInst::ICMP_SLT:
5081         // (float)int < -4.4   --> int < -4
5082         // (float)int < 4.4    --> int <= 4
5083         if (!RHS.isNegative())
5084           Pred = ICmpInst::ICMP_SLE;
5085         break;
5086       case ICmpInst::ICMP_UGT:
5087         // (float)int > 4.4    --> int > 4
5088         // (float)int > -4.4   --> true
5089         if (RHS.isNegative())
5090           return replaceInstUsesWith(I, Builder.getTrue());
5091         break;
5092       case ICmpInst::ICMP_SGT:
5093         // (float)int > 4.4    --> int > 4
5094         // (float)int > -4.4   --> int >= -4
5095         if (RHS.isNegative())
5096           Pred = ICmpInst::ICMP_SGE;
5097         break;
5098       case ICmpInst::ICMP_UGE:
5099         // (float)int >= -4.4   --> true
5100         // (float)int >= 4.4    --> int > 4
5101         if (RHS.isNegative())
5102           return replaceInstUsesWith(I, Builder.getTrue());
5103         Pred = ICmpInst::ICMP_UGT;
5104         break;
5105       case ICmpInst::ICMP_SGE:
5106         // (float)int >= -4.4   --> int >= -4
5107         // (float)int >= 4.4    --> int > 4
5108         if (!RHS.isNegative())
5109           Pred = ICmpInst::ICMP_SGT;
5110         break;
5111       }
5112     }
5113   }
5114 
5115   // Lower this FP comparison into an appropriate integer version of the
5116   // comparison.
5117   return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5118 }
5119 
visitFCmpInst(FCmpInst & I)5120 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5121   bool Changed = false;
5122 
5123   /// Orders the operands of the compare so that they are listed from most
5124   /// complex to least complex.  This puts constants before unary operators,
5125   /// before binary operators.
5126   if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
5127     I.swapOperands();
5128     Changed = true;
5129   }
5130 
5131   const CmpInst::Predicate Pred = I.getPredicate();
5132   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5133   if (Value *V = SimplifyFCmpInst(Pred, Op0, Op1, I.getFastMathFlags(),
5134                                   SQ.getWithInstruction(&I)))
5135     return replaceInstUsesWith(I, V);
5136 
5137   // Simplify 'fcmp pred X, X'
5138   if (Op0 == Op1) {
5139     switch (Pred) {
5140       default: break;
5141     case FCmpInst::FCMP_UNO:    // True if unordered: isnan(X) | isnan(Y)
5142     case FCmpInst::FCMP_ULT:    // True if unordered or less than
5143     case FCmpInst::FCMP_UGT:    // True if unordered or greater than
5144     case FCmpInst::FCMP_UNE:    // True if unordered or not equal
5145       // Canonicalize these to be 'fcmp uno %X, 0.0'.
5146       I.setPredicate(FCmpInst::FCMP_UNO);
5147       I.setOperand(1, Constant::getNullValue(Op0->getType()));
5148       return &I;
5149 
5150     case FCmpInst::FCMP_ORD:    // True if ordered (no nans)
5151     case FCmpInst::FCMP_OEQ:    // True if ordered and equal
5152     case FCmpInst::FCMP_OGE:    // True if ordered and greater than or equal
5153     case FCmpInst::FCMP_OLE:    // True if ordered and less than or equal
5154       // Canonicalize these to be 'fcmp ord %X, 0.0'.
5155       I.setPredicate(FCmpInst::FCMP_ORD);
5156       I.setOperand(1, Constant::getNullValue(Op0->getType()));
5157       return &I;
5158     }
5159   }
5160 
5161   // If we're just checking for a NaN (ORD/UNO) and have a non-NaN operand,
5162   // then canonicalize the operand to 0.0.
5163   if (Pred == CmpInst::FCMP_ORD || Pred == CmpInst::FCMP_UNO) {
5164     if (!match(Op0, m_PosZeroFP()) && isKnownNeverNaN(Op0)) {
5165       I.setOperand(0, ConstantFP::getNullValue(Op0->getType()));
5166       return &I;
5167     }
5168     if (!match(Op1, m_PosZeroFP()) && isKnownNeverNaN(Op1)) {
5169       I.setOperand(1, ConstantFP::getNullValue(Op0->getType()));
5170       return &I;
5171     }
5172   }
5173 
5174   // Test if the FCmpInst instruction is used exclusively by a select as
5175   // part of a minimum or maximum operation. If so, refrain from doing
5176   // any other folding. This helps out other analyses which understand
5177   // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5178   // and CodeGen. And in this case, at least one of the comparison
5179   // operands has at least one user besides the compare (the select),
5180   // which would often largely negate the benefit of folding anyway.
5181   if (I.hasOneUse())
5182     if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
5183       Value *A, *B;
5184       SelectPatternResult SPR = matchSelectPattern(SI, A, B);
5185       if (SPR.Flavor != SPF_UNKNOWN)
5186         return nullptr;
5187     }
5188 
5189   // Handle fcmp with constant RHS
5190   if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5191     if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5192       switch (LHSI->getOpcode()) {
5193       case Instruction::FPExt: {
5194         // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
5195         FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
5196         ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
5197         if (!RHSF)
5198           break;
5199 
5200         const fltSemantics *Sem;
5201         // FIXME: This shouldn't be here.
5202         if (LHSExt->getSrcTy()->isHalfTy())
5203           Sem = &APFloat::IEEEhalf();
5204         else if (LHSExt->getSrcTy()->isFloatTy())
5205           Sem = &APFloat::IEEEsingle();
5206         else if (LHSExt->getSrcTy()->isDoubleTy())
5207           Sem = &APFloat::IEEEdouble();
5208         else if (LHSExt->getSrcTy()->isFP128Ty())
5209           Sem = &APFloat::IEEEquad();
5210         else if (LHSExt->getSrcTy()->isX86_FP80Ty())
5211           Sem = &APFloat::x87DoubleExtended();
5212         else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
5213           Sem = &APFloat::PPCDoubleDouble();
5214         else
5215           break;
5216 
5217         bool Lossy;
5218         APFloat F = RHSF->getValueAPF();
5219         F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
5220 
5221         // Avoid lossy conversions and denormals. Zero is a special case
5222         // that's OK to convert.
5223         APFloat Fabs = F;
5224         Fabs.clearSign();
5225         if (!Lossy &&
5226             ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
5227                  APFloat::cmpLessThan) || Fabs.isZero()))
5228 
5229           return new FCmpInst(Pred, LHSExt->getOperand(0),
5230                               ConstantFP::get(RHSC->getContext(), F));
5231         break;
5232       }
5233       case Instruction::PHI:
5234         // Only fold fcmp into the PHI if the phi and fcmp are in the same
5235         // block.  If in the same block, we're encouraging jump threading.  If
5236         // not, we are just pessimizing the code by making an i1 phi.
5237         if (LHSI->getParent() == I.getParent())
5238           if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
5239             return NV;
5240         break;
5241       case Instruction::SIToFP:
5242       case Instruction::UIToFP:
5243         if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC))
5244           return NV;
5245         break;
5246       case Instruction::FSub: {
5247         // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
5248         Value *Op;
5249         if (match(LHSI, m_FNeg(m_Value(Op))))
5250           return new FCmpInst(I.getSwappedPredicate(), Op,
5251                               ConstantExpr::getFNeg(RHSC));
5252         break;
5253       }
5254       case Instruction::Load:
5255         if (GetElementPtrInst *GEP =
5256             dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
5257           if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
5258             if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
5259                 !cast<LoadInst>(LHSI)->isVolatile())
5260               if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
5261                 return Res;
5262         }
5263         break;
5264       case Instruction::Call: {
5265         if (!RHSC->isNullValue())
5266           break;
5267 
5268         CallInst *CI = cast<CallInst>(LHSI);
5269         Intrinsic::ID IID = getIntrinsicForCallSite(CI, &TLI);
5270         if (IID != Intrinsic::fabs)
5271           break;
5272 
5273         // Various optimization for fabs compared with zero.
5274         switch (Pred) {
5275         default:
5276           break;
5277         // fabs(x) < 0 --> false
5278         case FCmpInst::FCMP_OLT:
5279           llvm_unreachable("handled by SimplifyFCmpInst");
5280         // fabs(x) > 0 --> x != 0
5281         case FCmpInst::FCMP_OGT:
5282           return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0), RHSC);
5283         // fabs(x) <= 0 --> x == 0
5284         case FCmpInst::FCMP_OLE:
5285           return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0), RHSC);
5286         // fabs(x) >= 0 --> !isnan(x)
5287         case FCmpInst::FCMP_OGE:
5288           return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0), RHSC);
5289         // fabs(x) == 0 --> x == 0
5290         // fabs(x) != 0 --> x != 0
5291         case FCmpInst::FCMP_OEQ:
5292         case FCmpInst::FCMP_UEQ:
5293         case FCmpInst::FCMP_ONE:
5294         case FCmpInst::FCMP_UNE:
5295           return new FCmpInst(Pred, CI->getArgOperand(0), RHSC);
5296         }
5297       }
5298       }
5299   }
5300 
5301   // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
5302   Value *X, *Y;
5303   if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
5304     return new FCmpInst(I.getSwappedPredicate(), X, Y);
5305 
5306   // fcmp (fpext x), (fpext y) -> fcmp x, y
5307   if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
5308     if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
5309       if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
5310         return new FCmpInst(Pred, LHSExt->getOperand(0), RHSExt->getOperand(0));
5311 
5312   return Changed ? &I : nullptr;
5313 }
5314