LLVM API Documentation

InstCombineCompares.cpp
Go to the documentation of this file.
00001 //===- InstCombineCompares.cpp --------------------------------------------===//
00002 //
00003 //                     The LLVM Compiler Infrastructure
00004 //
00005 // This file is distributed under the University of Illinois Open Source
00006 // License. See LICENSE.TXT for details.
00007 //
00008 //===----------------------------------------------------------------------===//
00009 //
00010 // This file implements the visitICmp and visitFCmp functions.
00011 //
00012 //===----------------------------------------------------------------------===//
00013 
00014 #include "InstCombine.h"
00015 #include "llvm/Analysis/ConstantFolding.h"
00016 #include "llvm/Analysis/InstructionSimplify.h"
00017 #include "llvm/Analysis/MemoryBuiltins.h"
00018 #include "llvm/IR/ConstantRange.h"
00019 #include "llvm/IR/DataLayout.h"
00020 #include "llvm/IR/GetElementPtrTypeIterator.h"
00021 #include "llvm/IR/IntrinsicInst.h"
00022 #include "llvm/IR/PatternMatch.h"
00023 #include "llvm/Target/TargetLibraryInfo.h"
00024 using namespace llvm;
00025 using namespace PatternMatch;
00026 
00027 #define DEBUG_TYPE "instcombine"
00028 
00029 static ConstantInt *getOne(Constant *C) {
00030   return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
00031 }
00032 
00033 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
00034   return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
00035 }
00036 
00037 static bool HasAddOverflow(ConstantInt *Result,
00038                            ConstantInt *In1, ConstantInt *In2,
00039                            bool IsSigned) {
00040   if (!IsSigned)
00041     return Result->getValue().ult(In1->getValue());
00042 
00043   if (In2->isNegative())
00044     return Result->getValue().sgt(In1->getValue());
00045   return Result->getValue().slt(In1->getValue());
00046 }
00047 
00048 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
00049 /// overflowed for this type.
00050 static bool AddWithOverflow(Constant *&Result, Constant *In1,
00051                             Constant *In2, bool IsSigned = false) {
00052   Result = ConstantExpr::getAdd(In1, In2);
00053 
00054   if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
00055     for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
00056       Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
00057       if (HasAddOverflow(ExtractElement(Result, Idx),
00058                          ExtractElement(In1, Idx),
00059                          ExtractElement(In2, Idx),
00060                          IsSigned))
00061         return true;
00062     }
00063     return false;
00064   }
00065 
00066   return HasAddOverflow(cast<ConstantInt>(Result),
00067                         cast<ConstantInt>(In1), cast<ConstantInt>(In2),
00068                         IsSigned);
00069 }
00070 
00071 static bool HasSubOverflow(ConstantInt *Result,
00072                            ConstantInt *In1, ConstantInt *In2,
00073                            bool IsSigned) {
00074   if (!IsSigned)
00075     return Result->getValue().ugt(In1->getValue());
00076 
00077   if (In2->isNegative())
00078     return Result->getValue().slt(In1->getValue());
00079 
00080   return Result->getValue().sgt(In1->getValue());
00081 }
00082 
00083 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
00084 /// overflowed for this type.
00085 static bool SubWithOverflow(Constant *&Result, Constant *In1,
00086                             Constant *In2, bool IsSigned = false) {
00087   Result = ConstantExpr::getSub(In1, In2);
00088 
00089   if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
00090     for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
00091       Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
00092       if (HasSubOverflow(ExtractElement(Result, Idx),
00093                          ExtractElement(In1, Idx),
00094                          ExtractElement(In2, Idx),
00095                          IsSigned))
00096         return true;
00097     }
00098     return false;
00099   }
00100 
00101   return HasSubOverflow(cast<ConstantInt>(Result),
00102                         cast<ConstantInt>(In1), cast<ConstantInt>(In2),
00103                         IsSigned);
00104 }
00105 
00106 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
00107 /// comparison only checks the sign bit.  If it only checks the sign bit, set
00108 /// TrueIfSigned if the result of the comparison is true when the input value is
00109 /// signed.
00110 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
00111                            bool &TrueIfSigned) {
00112   switch (pred) {
00113   case ICmpInst::ICMP_SLT:   // True if LHS s< 0
00114     TrueIfSigned = true;
00115     return RHS->isZero();
00116   case ICmpInst::ICMP_SLE:   // True if LHS s<= RHS and RHS == -1
00117     TrueIfSigned = true;
00118     return RHS->isAllOnesValue();
00119   case ICmpInst::ICMP_SGT:   // True if LHS s> -1
00120     TrueIfSigned = false;
00121     return RHS->isAllOnesValue();
00122   case ICmpInst::ICMP_UGT:
00123     // True if LHS u> RHS and RHS == high-bit-mask - 1
00124     TrueIfSigned = true;
00125     return RHS->isMaxValue(true);
00126   case ICmpInst::ICMP_UGE:
00127     // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
00128     TrueIfSigned = true;
00129     return RHS->getValue().isSignBit();
00130   default:
00131     return false;
00132   }
00133 }
00134 
00135 /// Returns true if the exploded icmp can be expressed as a signed comparison
00136 /// to zero and updates the predicate accordingly.
00137 /// The signedness of the comparison is preserved.
00138 static bool isSignTest(ICmpInst::Predicate &pred, const ConstantInt *RHS) {
00139   if (!ICmpInst::isSigned(pred))
00140     return false;
00141 
00142   if (RHS->isZero())
00143     return ICmpInst::isRelational(pred);
00144 
00145   if (RHS->isOne()) {
00146     if (pred == ICmpInst::ICMP_SLT) {
00147       pred = ICmpInst::ICMP_SLE;
00148       return true;
00149     }
00150   } else if (RHS->isAllOnesValue()) {
00151     if (pred == ICmpInst::ICMP_SGT) {
00152       pred = ICmpInst::ICMP_SGE;
00153       return true;
00154     }
00155   }
00156 
00157   return false;
00158 }
00159 
00160 // isHighOnes - Return true if the constant is of the form 1+0+.
00161 // This is the same as lowones(~X).
00162 static bool isHighOnes(const ConstantInt *CI) {
00163   return (~CI->getValue() + 1).isPowerOf2();
00164 }
00165 
00166 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
00167 /// set of known zero and one bits, compute the maximum and minimum values that
00168 /// could have the specified known zero and known one bits, returning them in
00169 /// min/max.
00170 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
00171                                                    const APInt& KnownOne,
00172                                                    APInt& Min, APInt& Max) {
00173   assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
00174          KnownZero.getBitWidth() == Min.getBitWidth() &&
00175          KnownZero.getBitWidth() == Max.getBitWidth() &&
00176          "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
00177   APInt UnknownBits = ~(KnownZero|KnownOne);
00178 
00179   // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
00180   // bit if it is unknown.
00181   Min = KnownOne;
00182   Max = KnownOne|UnknownBits;
00183 
00184   if (UnknownBits.isNegative()) { // Sign bit is unknown
00185     Min.setBit(Min.getBitWidth()-1);
00186     Max.clearBit(Max.getBitWidth()-1);
00187   }
00188 }
00189 
00190 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
00191 // a set of known zero and one bits, compute the maximum and minimum values that
00192 // could have the specified known zero and known one bits, returning them in
00193 // min/max.
00194 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
00195                                                      const APInt &KnownOne,
00196                                                      APInt &Min, APInt &Max) {
00197   assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
00198          KnownZero.getBitWidth() == Min.getBitWidth() &&
00199          KnownZero.getBitWidth() == Max.getBitWidth() &&
00200          "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
00201   APInt UnknownBits = ~(KnownZero|KnownOne);
00202 
00203   // The minimum value is when the unknown bits are all zeros.
00204   Min = KnownOne;
00205   // The maximum value is when the unknown bits are all ones.
00206   Max = KnownOne|UnknownBits;
00207 }
00208 
00209 
00210 
00211 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
00212 ///   cmp pred (load (gep GV, ...)), cmpcst
00213 /// where GV is a global variable with a constant initializer.  Try to simplify
00214 /// this into some simple computation that does not need the load.  For example
00215 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
00216 ///
00217 /// If AndCst is non-null, then the loaded value is masked with that constant
00218 /// before doing the comparison.  This handles cases like "A[i]&4 == 0".
00219 Instruction *InstCombiner::
00220 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
00221                              CmpInst &ICI, ConstantInt *AndCst) {
00222   // We need TD information to know the pointer size unless this is inbounds.
00223   if (!GEP->isInBounds() && !DL)
00224     return nullptr;
00225 
00226   Constant *Init = GV->getInitializer();
00227   if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
00228     return nullptr;
00229 
00230   uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
00231   if (ArrayElementCount > 1024) return nullptr; // Don't blow up on huge arrays.
00232 
00233   // There are many forms of this optimization we can handle, for now, just do
00234   // the simple index into a single-dimensional array.
00235   //
00236   // Require: GEP GV, 0, i {{, constant indices}}
00237   if (GEP->getNumOperands() < 3 ||
00238       !isa<ConstantInt>(GEP->getOperand(1)) ||
00239       !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
00240       isa<Constant>(GEP->getOperand(2)))
00241     return nullptr;
00242 
00243   // Check that indices after the variable are constants and in-range for the
00244   // type they index.  Collect the indices.  This is typically for arrays of
00245   // structs.
00246   SmallVector<unsigned, 4> LaterIndices;
00247 
00248   Type *EltTy = Init->getType()->getArrayElementType();
00249   for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
00250     ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
00251     if (!Idx) return nullptr;  // Variable index.
00252 
00253     uint64_t IdxVal = Idx->getZExtValue();
00254     if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
00255 
00256     if (StructType *STy = dyn_cast<StructType>(EltTy))
00257       EltTy = STy->getElementType(IdxVal);
00258     else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
00259       if (IdxVal >= ATy->getNumElements()) return nullptr;
00260       EltTy = ATy->getElementType();
00261     } else {
00262       return nullptr; // Unknown type.
00263     }
00264 
00265     LaterIndices.push_back(IdxVal);
00266   }
00267 
00268   enum { Overdefined = -3, Undefined = -2 };
00269 
00270   // Variables for our state machines.
00271 
00272   // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
00273   // "i == 47 | i == 87", where 47 is the first index the condition is true for,
00274   // and 87 is the second (and last) index.  FirstTrueElement is -2 when
00275   // undefined, otherwise set to the first true element.  SecondTrueElement is
00276   // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
00277   int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
00278 
00279   // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
00280   // form "i != 47 & i != 87".  Same state transitions as for true elements.
00281   int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
00282 
00283   /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
00284   /// define a state machine that triggers for ranges of values that the index
00285   /// is true or false for.  This triggers on things like "abbbbc"[i] == 'b'.
00286   /// This is -2 when undefined, -3 when overdefined, and otherwise the last
00287   /// index in the range (inclusive).  We use -2 for undefined here because we
00288   /// use relative comparisons and don't want 0-1 to match -1.
00289   int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
00290 
00291   // MagicBitvector - This is a magic bitvector where we set a bit if the
00292   // comparison is true for element 'i'.  If there are 64 elements or less in
00293   // the array, this will fully represent all the comparison results.
00294   uint64_t MagicBitvector = 0;
00295 
00296 
00297   // Scan the array and see if one of our patterns matches.
00298   Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
00299   for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
00300     Constant *Elt = Init->getAggregateElement(i);
00301     if (!Elt) return nullptr;
00302 
00303     // If this is indexing an array of structures, get the structure element.
00304     if (!LaterIndices.empty())
00305       Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
00306 
00307     // If the element is masked, handle it.
00308     if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
00309 
00310     // Find out if the comparison would be true or false for the i'th element.
00311     Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
00312                                                   CompareRHS, DL, TLI);
00313     // If the result is undef for this element, ignore it.
00314     if (isa<UndefValue>(C)) {
00315       // Extend range state machines to cover this element in case there is an
00316       // undef in the middle of the range.
00317       if (TrueRangeEnd == (int)i-1)
00318         TrueRangeEnd = i;
00319       if (FalseRangeEnd == (int)i-1)
00320         FalseRangeEnd = i;
00321       continue;
00322     }
00323 
00324     // If we can't compute the result for any of the elements, we have to give
00325     // up evaluating the entire conditional.
00326     if (!isa<ConstantInt>(C)) return nullptr;
00327 
00328     // Otherwise, we know if the comparison is true or false for this element,
00329     // update our state machines.
00330     bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
00331 
00332     // State machine for single/double/range index comparison.
00333     if (IsTrueForElt) {
00334       // Update the TrueElement state machine.
00335       if (FirstTrueElement == Undefined)
00336         FirstTrueElement = TrueRangeEnd = i;  // First true element.
00337       else {
00338         // Update double-compare state machine.
00339         if (SecondTrueElement == Undefined)
00340           SecondTrueElement = i;
00341         else
00342           SecondTrueElement = Overdefined;
00343 
00344         // Update range state machine.
00345         if (TrueRangeEnd == (int)i-1)
00346           TrueRangeEnd = i;
00347         else
00348           TrueRangeEnd = Overdefined;
00349       }
00350     } else {
00351       // Update the FalseElement state machine.
00352       if (FirstFalseElement == Undefined)
00353         FirstFalseElement = FalseRangeEnd = i; // First false element.
00354       else {
00355         // Update double-compare state machine.
00356         if (SecondFalseElement == Undefined)
00357           SecondFalseElement = i;
00358         else
00359           SecondFalseElement = Overdefined;
00360 
00361         // Update range state machine.
00362         if (FalseRangeEnd == (int)i-1)
00363           FalseRangeEnd = i;
00364         else
00365           FalseRangeEnd = Overdefined;
00366       }
00367     }
00368 
00369 
00370     // If this element is in range, update our magic bitvector.
00371     if (i < 64 && IsTrueForElt)
00372       MagicBitvector |= 1ULL << i;
00373 
00374     // If all of our states become overdefined, bail out early.  Since the
00375     // predicate is expensive, only check it every 8 elements.  This is only
00376     // really useful for really huge arrays.
00377     if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
00378         SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
00379         FalseRangeEnd == Overdefined)
00380       return nullptr;
00381   }
00382 
00383   // Now that we've scanned the entire array, emit our new comparison(s).  We
00384   // order the state machines in complexity of the generated code.
00385   Value *Idx = GEP->getOperand(2);
00386 
00387   // If the index is larger than the pointer size of the target, truncate the
00388   // index down like the GEP would do implicitly.  We don't have to do this for
00389   // an inbounds GEP because the index can't be out of range.
00390   if (!GEP->isInBounds()) {
00391     Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
00392     unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
00393     if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
00394       Idx = Builder->CreateTrunc(Idx, IntPtrTy);
00395   }
00396 
00397   // If the comparison is only true for one or two elements, emit direct
00398   // comparisons.
00399   if (SecondTrueElement != Overdefined) {
00400     // None true -> false.
00401     if (FirstTrueElement == Undefined)
00402       return ReplaceInstUsesWith(ICI, Builder->getFalse());
00403 
00404     Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
00405 
00406     // True for one element -> 'i == 47'.
00407     if (SecondTrueElement == Undefined)
00408       return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
00409 
00410     // True for two elements -> 'i == 47 | i == 72'.
00411     Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
00412     Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
00413     Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
00414     return BinaryOperator::CreateOr(C1, C2);
00415   }
00416 
00417   // If the comparison is only false for one or two elements, emit direct
00418   // comparisons.
00419   if (SecondFalseElement != Overdefined) {
00420     // None false -> true.
00421     if (FirstFalseElement == Undefined)
00422       return ReplaceInstUsesWith(ICI, Builder->getTrue());
00423 
00424     Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
00425 
00426     // False for one element -> 'i != 47'.
00427     if (SecondFalseElement == Undefined)
00428       return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
00429 
00430     // False for two elements -> 'i != 47 & i != 72'.
00431     Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
00432     Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
00433     Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
00434     return BinaryOperator::CreateAnd(C1, C2);
00435   }
00436 
00437   // If the comparison can be replaced with a range comparison for the elements
00438   // where it is true, emit the range check.
00439   if (TrueRangeEnd != Overdefined) {
00440     assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
00441 
00442     // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
00443     if (FirstTrueElement) {
00444       Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
00445       Idx = Builder->CreateAdd(Idx, Offs);
00446     }
00447 
00448     Value *End = ConstantInt::get(Idx->getType(),
00449                                   TrueRangeEnd-FirstTrueElement+1);
00450     return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
00451   }
00452 
00453   // False range check.
00454   if (FalseRangeEnd != Overdefined) {
00455     assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
00456     // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
00457     if (FirstFalseElement) {
00458       Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
00459       Idx = Builder->CreateAdd(Idx, Offs);
00460     }
00461 
00462     Value *End = ConstantInt::get(Idx->getType(),
00463                                   FalseRangeEnd-FirstFalseElement);
00464     return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
00465   }
00466 
00467 
00468   // If a magic bitvector captures the entire comparison state
00469   // of this load, replace it with computation that does:
00470   //   ((magic_cst >> i) & 1) != 0
00471   {
00472     Type *Ty = nullptr;
00473 
00474     // Look for an appropriate type:
00475     // - The type of Idx if the magic fits
00476     // - The smallest fitting legal type if we have a DataLayout
00477     // - Default to i32
00478     if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
00479       Ty = Idx->getType();
00480     else if (DL)
00481       Ty = DL->getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
00482     else if (ArrayElementCount <= 32)
00483       Ty = Type::getInt32Ty(Init->getContext());
00484 
00485     if (Ty) {
00486       Value *V = Builder->CreateIntCast(Idx, Ty, false);
00487       V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
00488       V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
00489       return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
00490     }
00491   }
00492 
00493   return nullptr;
00494 }
00495 
00496 
00497 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
00498 /// the *offset* implied by a GEP to zero.  For example, if we have &A[i], we
00499 /// want to return 'i' for "icmp ne i, 0".  Note that, in general, indices can
00500 /// be complex, and scales are involved.  The above expression would also be
00501 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
00502 /// This later form is less amenable to optimization though, and we are allowed
00503 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
00504 ///
00505 /// If we can't emit an optimized form for this expression, this returns null.
00506 ///
00507 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
00508   const DataLayout &DL = *IC.getDataLayout();
00509   gep_type_iterator GTI = gep_type_begin(GEP);
00510 
00511   // Check to see if this gep only has a single variable index.  If so, and if
00512   // any constant indices are a multiple of its scale, then we can compute this
00513   // in terms of the scale of the variable index.  For example, if the GEP
00514   // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
00515   // because the expression will cross zero at the same point.
00516   unsigned i, e = GEP->getNumOperands();
00517   int64_t Offset = 0;
00518   for (i = 1; i != e; ++i, ++GTI) {
00519     if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
00520       // Compute the aggregate offset of constant indices.
00521       if (CI->isZero()) continue;
00522 
00523       // Handle a struct index, which adds its field offset to the pointer.
00524       if (StructType *STy = dyn_cast<StructType>(*GTI)) {
00525         Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
00526       } else {
00527         uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
00528         Offset += Size*CI->getSExtValue();
00529       }
00530     } else {
00531       // Found our variable index.
00532       break;
00533     }
00534   }
00535 
00536   // If there are no variable indices, we must have a constant offset, just
00537   // evaluate it the general way.
00538   if (i == e) return nullptr;
00539 
00540   Value *VariableIdx = GEP->getOperand(i);
00541   // Determine the scale factor of the variable element.  For example, this is
00542   // 4 if the variable index is into an array of i32.
00543   uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
00544 
00545   // Verify that there are no other variable indices.  If so, emit the hard way.
00546   for (++i, ++GTI; i != e; ++i, ++GTI) {
00547     ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
00548     if (!CI) return nullptr;
00549 
00550     // Compute the aggregate offset of constant indices.
00551     if (CI->isZero()) continue;
00552 
00553     // Handle a struct index, which adds its field offset to the pointer.
00554     if (StructType *STy = dyn_cast<StructType>(*GTI)) {
00555       Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
00556     } else {
00557       uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
00558       Offset += Size*CI->getSExtValue();
00559     }
00560   }
00561 
00562 
00563 
00564   // Okay, we know we have a single variable index, which must be a
00565   // pointer/array/vector index.  If there is no offset, life is simple, return
00566   // the index.
00567   Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
00568   unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
00569   if (Offset == 0) {
00570     // Cast to intptrty in case a truncation occurs.  If an extension is needed,
00571     // we don't need to bother extending: the extension won't affect where the
00572     // computation crosses zero.
00573     if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
00574       VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
00575     }
00576     return VariableIdx;
00577   }
00578 
00579   // Otherwise, there is an index.  The computation we will do will be modulo
00580   // the pointer size, so get it.
00581   uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
00582 
00583   Offset &= PtrSizeMask;
00584   VariableScale &= PtrSizeMask;
00585 
00586   // To do this transformation, any constant index must be a multiple of the
00587   // variable scale factor.  For example, we can evaluate "12 + 4*i" as "3 + i",
00588   // but we can't evaluate "10 + 3*i" in terms of i.  Check that the offset is a
00589   // multiple of the variable scale.
00590   int64_t NewOffs = Offset / (int64_t)VariableScale;
00591   if (Offset != NewOffs*(int64_t)VariableScale)
00592     return nullptr;
00593 
00594   // Okay, we can do this evaluation.  Start by converting the index to intptr.
00595   if (VariableIdx->getType() != IntPtrTy)
00596     VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
00597                                             true /*Signed*/);
00598   Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
00599   return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
00600 }
00601 
00602 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
00603 /// else.  At this point we know that the GEP is on the LHS of the comparison.
00604 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
00605                                        ICmpInst::Predicate Cond,
00606                                        Instruction &I) {
00607   // Don't transform signed compares of GEPs into index compares. Even if the
00608   // GEP is inbounds, the final add of the base pointer can have signed overflow
00609   // and would change the result of the icmp.
00610   // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
00611   // the maximum signed value for the pointer type.
00612   if (ICmpInst::isSigned(Cond))
00613     return nullptr;
00614 
00615   // Look through bitcasts and addrspacecasts. We do not however want to remove
00616   // 0 GEPs.
00617   if (!isa<GetElementPtrInst>(RHS))
00618     RHS = RHS->stripPointerCasts();
00619 
00620   Value *PtrBase = GEPLHS->getOperand(0);
00621   if (DL && PtrBase == RHS && GEPLHS->isInBounds()) {
00622     // ((gep Ptr, OFFSET) cmp Ptr)   ---> (OFFSET cmp 0).
00623     // This transformation (ignoring the base and scales) is valid because we
00624     // know pointers can't overflow since the gep is inbounds.  See if we can
00625     // output an optimized form.
00626     Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
00627 
00628     // If not, synthesize the offset the hard way.
00629     if (!Offset)
00630       Offset = EmitGEPOffset(GEPLHS);
00631     return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
00632                         Constant::getNullValue(Offset->getType()));
00633   } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
00634     // If the base pointers are different, but the indices are the same, just
00635     // compare the base pointer.
00636     if (PtrBase != GEPRHS->getOperand(0)) {
00637       bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
00638       IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
00639                         GEPRHS->getOperand(0)->getType();
00640       if (IndicesTheSame)
00641         for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
00642           if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
00643             IndicesTheSame = false;
00644             break;
00645           }
00646 
00647       // If all indices are the same, just compare the base pointers.
00648       if (IndicesTheSame)
00649         return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
00650 
00651       // If we're comparing GEPs with two base pointers that only differ in type
00652       // and both GEPs have only constant indices or just one use, then fold
00653       // the compare with the adjusted indices.
00654       if (DL && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
00655           (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
00656           (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
00657           PtrBase->stripPointerCasts() ==
00658             GEPRHS->getOperand(0)->stripPointerCasts()) {
00659         Value *LOffset = EmitGEPOffset(GEPLHS);
00660         Value *ROffset = EmitGEPOffset(GEPRHS);
00661 
00662         // If we looked through an addrspacecast between different sized address
00663         // spaces, the LHS and RHS pointers are different sized
00664         // integers. Truncate to the smaller one.
00665         Type *LHSIndexTy = LOffset->getType();
00666         Type *RHSIndexTy = ROffset->getType();
00667         if (LHSIndexTy != RHSIndexTy) {
00668           if (LHSIndexTy->getPrimitiveSizeInBits() <
00669               RHSIndexTy->getPrimitiveSizeInBits()) {
00670             ROffset = Builder->CreateTrunc(ROffset, LHSIndexTy);
00671           } else
00672             LOffset = Builder->CreateTrunc(LOffset, RHSIndexTy);
00673         }
00674 
00675         Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
00676                                          LOffset, ROffset);
00677         return ReplaceInstUsesWith(I, Cmp);
00678       }
00679 
00680       // Otherwise, the base pointers are different and the indices are
00681       // different, bail out.
00682       return nullptr;
00683     }
00684 
00685     // If one of the GEPs has all zero indices, recurse.
00686     if (GEPLHS->hasAllZeroIndices())
00687       return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
00688                          ICmpInst::getSwappedPredicate(Cond), I);
00689 
00690     // If the other GEP has all zero indices, recurse.
00691     if (GEPRHS->hasAllZeroIndices())
00692       return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
00693 
00694     bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
00695     if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
00696       // If the GEPs only differ by one index, compare it.
00697       unsigned NumDifferences = 0;  // Keep track of # differences.
00698       unsigned DiffOperand = 0;     // The operand that differs.
00699       for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
00700         if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
00701           if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
00702                    GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
00703             // Irreconcilable differences.
00704             NumDifferences = 2;
00705             break;
00706           } else {
00707             if (NumDifferences++) break;
00708             DiffOperand = i;
00709           }
00710         }
00711 
00712       if (NumDifferences == 0)   // SAME GEP?
00713         return ReplaceInstUsesWith(I, // No comparison is needed here.
00714                              Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
00715 
00716       else if (NumDifferences == 1 && GEPsInBounds) {
00717         Value *LHSV = GEPLHS->getOperand(DiffOperand);
00718         Value *RHSV = GEPRHS->getOperand(DiffOperand);
00719         // Make sure we do a signed comparison here.
00720         return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
00721       }
00722     }
00723 
00724     // Only lower this if the icmp is the only user of the GEP or if we expect
00725     // the result to fold to a constant!
00726     if (DL &&
00727         GEPsInBounds &&
00728         (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
00729         (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
00730       // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)  --->  (OFFSET1 cmp OFFSET2)
00731       Value *L = EmitGEPOffset(GEPLHS);
00732       Value *R = EmitGEPOffset(GEPRHS);
00733       return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
00734     }
00735   }
00736   return nullptr;
00737 }
00738 
00739 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
00740 Instruction *InstCombiner::FoldICmpAddOpCst(Instruction &ICI,
00741                                             Value *X, ConstantInt *CI,
00742                                             ICmpInst::Predicate Pred) {
00743   // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
00744   // so the values can never be equal.  Similarly for all other "or equals"
00745   // operators.
00746 
00747   // (X+1) <u X        --> X >u (MAXUINT-1)        --> X == 255
00748   // (X+2) <u X        --> X >u (MAXUINT-2)        --> X > 253
00749   // (X+MAXUINT) <u X  --> X >u (MAXUINT-MAXUINT)  --> X != 0
00750   if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
00751     Value *R =
00752       ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
00753     return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
00754   }
00755 
00756   // (X+1) >u X        --> X <u (0-1)        --> X != 255
00757   // (X+2) >u X        --> X <u (0-2)        --> X <u 254
00758   // (X+MAXUINT) >u X  --> X <u (0-MAXUINT)  --> X <u 1  --> X == 0
00759   if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
00760     return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
00761 
00762   unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
00763   ConstantInt *SMax = ConstantInt::get(X->getContext(),
00764                                        APInt::getSignedMaxValue(BitWidth));
00765 
00766   // (X+ 1) <s X       --> X >s (MAXSINT-1)          --> X == 127
00767   // (X+ 2) <s X       --> X >s (MAXSINT-2)          --> X >s 125
00768   // (X+MAXSINT) <s X  --> X >s (MAXSINT-MAXSINT)    --> X >s 0
00769   // (X+MINSINT) <s X  --> X >s (MAXSINT-MINSINT)    --> X >s -1
00770   // (X+ -2) <s X      --> X >s (MAXSINT- -2)        --> X >s 126
00771   // (X+ -1) <s X      --> X >s (MAXSINT- -1)        --> X != 127
00772   if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
00773     return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
00774 
00775   // (X+ 1) >s X       --> X <s (MAXSINT-(1-1))       --> X != 127
00776   // (X+ 2) >s X       --> X <s (MAXSINT-(2-1))       --> X <s 126
00777   // (X+MAXSINT) >s X  --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
00778   // (X+MINSINT) >s X  --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
00779   // (X+ -2) >s X      --> X <s (MAXSINT-(-2-1))      --> X <s -126
00780   // (X+ -1) >s X      --> X <s (MAXSINT-(-1-1))      --> X == -128
00781 
00782   assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
00783   Constant *C = Builder->getInt(CI->getValue()-1);
00784   return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
00785 }
00786 
00787 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
00788 /// and CmpRHS are both known to be integer constants.
00789 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
00790                                           ConstantInt *DivRHS) {
00791   ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
00792   const APInt &CmpRHSV = CmpRHS->getValue();
00793 
00794   // FIXME: If the operand types don't match the type of the divide
00795   // then don't attempt this transform. The code below doesn't have the
00796   // logic to deal with a signed divide and an unsigned compare (and
00797   // vice versa). This is because (x /s C1) <s C2  produces different
00798   // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
00799   // (x /u C1) <u C2.  Simply casting the operands and result won't
00800   // work. :(  The if statement below tests that condition and bails
00801   // if it finds it.
00802   bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
00803   if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
00804     return nullptr;
00805   if (DivRHS->isZero())
00806     return nullptr; // The ProdOV computation fails on divide by zero.
00807   if (DivIsSigned && DivRHS->isAllOnesValue())
00808     return nullptr; // The overflow computation also screws up here
00809   if (DivRHS->isOne()) {
00810     // This eliminates some funny cases with INT_MIN.
00811     ICI.setOperand(0, DivI->getOperand(0));   // X/1 == X.
00812     return &ICI;
00813   }
00814 
00815   // Compute Prod = CI * DivRHS. We are essentially solving an equation
00816   // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
00817   // C2 (CI). By solving for X we can turn this into a range check
00818   // instead of computing a divide.
00819   Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
00820 
00821   // Determine if the product overflows by seeing if the product is
00822   // not equal to the divide. Make sure we do the same kind of divide
00823   // as in the LHS instruction that we're folding.
00824   bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
00825                  ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
00826 
00827   // Get the ICmp opcode
00828   ICmpInst::Predicate Pred = ICI.getPredicate();
00829 
00830   /// If the division is known to be exact, then there is no remainder from the
00831   /// divide, so the covered range size is unit, otherwise it is the divisor.
00832   ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
00833 
00834   // Figure out the interval that is being checked.  For example, a comparison
00835   // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
00836   // Compute this interval based on the constants involved and the signedness of
00837   // the compare/divide.  This computes a half-open interval, keeping track of
00838   // whether either value in the interval overflows.  After analysis each
00839   // overflow variable is set to 0 if it's corresponding bound variable is valid
00840   // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
00841   int LoOverflow = 0, HiOverflow = 0;
00842   Constant *LoBound = nullptr, *HiBound = nullptr;
00843 
00844   if (!DivIsSigned) {  // udiv
00845     // e.g. X/5 op 3  --> [15, 20)
00846     LoBound = Prod;
00847     HiOverflow = LoOverflow = ProdOV;
00848     if (!HiOverflow) {
00849       // If this is not an exact divide, then many values in the range collapse
00850       // to the same result value.
00851       HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
00852     }
00853 
00854   } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
00855     if (CmpRHSV == 0) {       // (X / pos) op 0
00856       // Can't overflow.  e.g.  X/2 op 0 --> [-1, 2)
00857       LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
00858       HiBound = RangeSize;
00859     } else if (CmpRHSV.isStrictlyPositive()) {   // (X / pos) op pos
00860       LoBound = Prod;     // e.g.   X/5 op 3 --> [15, 20)
00861       HiOverflow = LoOverflow = ProdOV;
00862       if (!HiOverflow)
00863         HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
00864     } else {                       // (X / pos) op neg
00865       // e.g. X/5 op -3  --> [-15-4, -15+1) --> [-19, -14)
00866       HiBound = AddOne(Prod);
00867       LoOverflow = HiOverflow = ProdOV ? -1 : 0;
00868       if (!LoOverflow) {
00869         ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
00870         LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
00871       }
00872     }
00873   } else if (DivRHS->isNegative()) { // Divisor is < 0.
00874     if (DivI->isExact())
00875       RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
00876     if (CmpRHSV == 0) {       // (X / neg) op 0
00877       // e.g. X/-5 op 0  --> [-4, 5)
00878       LoBound = AddOne(RangeSize);
00879       HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
00880       if (HiBound == DivRHS) {     // -INTMIN = INTMIN
00881         HiOverflow = 1;            // [INTMIN+1, overflow)
00882         HiBound = nullptr;         // e.g. X/INTMIN = 0 --> X > INTMIN
00883       }
00884     } else if (CmpRHSV.isStrictlyPositive()) {   // (X / neg) op pos
00885       // e.g. X/-5 op 3  --> [-19, -14)
00886       HiBound = AddOne(Prod);
00887       HiOverflow = LoOverflow = ProdOV ? -1 : 0;
00888       if (!LoOverflow)
00889         LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
00890     } else {                       // (X / neg) op neg
00891       LoBound = Prod;       // e.g. X/-5 op -3  --> [15, 20)
00892       LoOverflow = HiOverflow = ProdOV;
00893       if (!HiOverflow)
00894         HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
00895     }
00896 
00897     // Dividing by a negative swaps the condition.  LT <-> GT
00898     Pred = ICmpInst::getSwappedPredicate(Pred);
00899   }
00900 
00901   Value *X = DivI->getOperand(0);
00902   switch (Pred) {
00903   default: llvm_unreachable("Unhandled icmp opcode!");
00904   case ICmpInst::ICMP_EQ:
00905     if (LoOverflow && HiOverflow)
00906       return ReplaceInstUsesWith(ICI, Builder->getFalse());
00907     if (HiOverflow)
00908       return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
00909                           ICmpInst::ICMP_UGE, X, LoBound);
00910     if (LoOverflow)
00911       return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
00912                           ICmpInst::ICMP_ULT, X, HiBound);
00913     return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
00914                                                     DivIsSigned, true));
00915   case ICmpInst::ICMP_NE:
00916     if (LoOverflow && HiOverflow)
00917       return ReplaceInstUsesWith(ICI, Builder->getTrue());
00918     if (HiOverflow)
00919       return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
00920                           ICmpInst::ICMP_ULT, X, LoBound);
00921     if (LoOverflow)
00922       return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
00923                           ICmpInst::ICMP_UGE, X, HiBound);
00924     return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
00925                                                     DivIsSigned, false));
00926   case ICmpInst::ICMP_ULT:
00927   case ICmpInst::ICMP_SLT:
00928     if (LoOverflow == +1)   // Low bound is greater than input range.
00929       return ReplaceInstUsesWith(ICI, Builder->getTrue());
00930     if (LoOverflow == -1)   // Low bound is less than input range.
00931       return ReplaceInstUsesWith(ICI, Builder->getFalse());
00932     return new ICmpInst(Pred, X, LoBound);
00933   case ICmpInst::ICMP_UGT:
00934   case ICmpInst::ICMP_SGT:
00935     if (HiOverflow == +1)       // High bound greater than input range.
00936       return ReplaceInstUsesWith(ICI, Builder->getFalse());
00937     if (HiOverflow == -1)       // High bound less than input range.
00938       return ReplaceInstUsesWith(ICI, Builder->getTrue());
00939     if (Pred == ICmpInst::ICMP_UGT)
00940       return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
00941     return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
00942   }
00943 }
00944 
00945 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
00946 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
00947                                           ConstantInt *ShAmt) {
00948   const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
00949 
00950   // Check that the shift amount is in range.  If not, don't perform
00951   // undefined shifts.  When the shift is visited it will be
00952   // simplified.
00953   uint32_t TypeBits = CmpRHSV.getBitWidth();
00954   uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
00955   if (ShAmtVal >= TypeBits || ShAmtVal == 0)
00956     return nullptr;
00957 
00958   if (!ICI.isEquality()) {
00959     // If we have an unsigned comparison and an ashr, we can't simplify this.
00960     // Similarly for signed comparisons with lshr.
00961     if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
00962       return nullptr;
00963 
00964     // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
00965     // by a power of 2.  Since we already have logic to simplify these,
00966     // transform to div and then simplify the resultant comparison.
00967     if (Shr->getOpcode() == Instruction::AShr &&
00968         (!Shr->isExact() || ShAmtVal == TypeBits - 1))
00969       return nullptr;
00970 
00971     // Revisit the shift (to delete it).
00972     Worklist.Add(Shr);
00973 
00974     Constant *DivCst =
00975       ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
00976 
00977     Value *Tmp =
00978       Shr->getOpcode() == Instruction::AShr ?
00979       Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
00980       Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
00981 
00982     ICI.setOperand(0, Tmp);
00983 
00984     // If the builder folded the binop, just return it.
00985     BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
00986     if (!TheDiv)
00987       return &ICI;
00988 
00989     // Otherwise, fold this div/compare.
00990     assert(TheDiv->getOpcode() == Instruction::SDiv ||
00991            TheDiv->getOpcode() == Instruction::UDiv);
00992 
00993     Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
00994     assert(Res && "This div/cst should have folded!");
00995     return Res;
00996   }
00997 
00998 
00999   // If we are comparing against bits always shifted out, the
01000   // comparison cannot succeed.
01001   APInt Comp = CmpRHSV << ShAmtVal;
01002   ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
01003   if (Shr->getOpcode() == Instruction::LShr)
01004     Comp = Comp.lshr(ShAmtVal);
01005   else
01006     Comp = Comp.ashr(ShAmtVal);
01007 
01008   if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
01009     bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
01010     Constant *Cst = Builder->getInt1(IsICMP_NE);
01011     return ReplaceInstUsesWith(ICI, Cst);
01012   }
01013 
01014   // Otherwise, check to see if the bits shifted out are known to be zero.
01015   // If so, we can compare against the unshifted value:
01016   //  (X & 4) >> 1 == 2  --> (X & 4) == 4.
01017   if (Shr->hasOneUse() && Shr->isExact())
01018     return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
01019 
01020   if (Shr->hasOneUse()) {
01021     // Otherwise strength reduce the shift into an and.
01022     APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
01023     Constant *Mask = Builder->getInt(Val);
01024 
01025     Value *And = Builder->CreateAnd(Shr->getOperand(0),
01026                                     Mask, Shr->getName()+".mask");
01027     return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
01028   }
01029   return nullptr;
01030 }
01031 
01032 /// FoldICmpCstShrCst - Handle "(icmp eq/ne (ashr/lshr const2, A), const1)" ->
01033 /// (icmp eq/ne A, Log2(const2/const1)) ->
01034 /// (icmp eq/ne A, Log2(const2) - Log2(const1)).
01035 Instruction *InstCombiner::FoldICmpCstShrCst(ICmpInst &I, Value *Op, Value *A,
01036                                              ConstantInt *CI1,
01037                                              ConstantInt *CI2) {
01038   assert(I.isEquality() && "Cannot fold icmp gt/lt");
01039 
01040   auto getConstant = [&I, this](bool IsTrue) {
01041     if (I.getPredicate() == I.ICMP_NE)
01042       IsTrue = !IsTrue;
01043     return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
01044   };
01045 
01046   auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
01047     if (I.getPredicate() == I.ICMP_NE)
01048       Pred = CmpInst::getInversePredicate(Pred);
01049     return new ICmpInst(Pred, LHS, RHS);
01050   };
01051 
01052   APInt AP1 = CI1->getValue();
01053   APInt AP2 = CI2->getValue();
01054 
01055   assert(AP2 != 0 && "Handled in InstSimplify");
01056   assert(!AP2.isAllOnesValue() && "Handled in InstSimplify");
01057 
01058   if (!AP1)
01059     // 'A' must be large enough to shift out the highest set bit.
01060     return getICmp(I.ICMP_UGT, A,
01061                    ConstantInt::get(A->getType(), AP2.logBase2()));
01062 
01063   if (AP1 == AP2)
01064     return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
01065 
01066   bool IsAShr = isa<AShrOperator>(Op);
01067   // If we are dealing with an arithmetic shift, both constants should agree in
01068   // sign.  InstSimplify's SimplifyICmpInst range analysis is supposed to catch
01069   // the cases when they disagree.
01070   assert((!IsAShr || (AP1.isNegative() == AP2.isNegative() && AP1.sgt(AP2))) &&
01071          "Handled in InstSimplify");
01072 
01073   // Get the distance between the highest bit that's set.
01074   int Shift;
01075   // Both the constants are negative, take their positive to calculate log.
01076   if (IsAShr && AP1.isNegative())
01077     // Get the ones' complement of AP2 and AP1 when computing the distance.
01078     Shift = (~AP2).logBase2() - (~AP1).logBase2();
01079   else
01080     Shift = AP2.logBase2() - AP1.logBase2();
01081 
01082   if (Shift > 0) {
01083     if (IsAShr ? AP1 == AP2.ashr(Shift) : AP1 == AP2.lshr(Shift))
01084       return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
01085   }
01086   // Shifting const2 will never be equal to const1.
01087   return getConstant(false);
01088 }
01089 
01090 /// FoldICmpCstShlCst - Handle "(icmp eq/ne (shl const2, A), const1)" ->
01091 /// (icmp eq/ne A, TrailingZeros(const1) - TrailingZeros(const2)).
01092 Instruction *InstCombiner::FoldICmpCstShlCst(ICmpInst &I, Value *Op, Value *A,
01093                                              ConstantInt *CI1,
01094                                              ConstantInt *CI2) {
01095   assert(I.isEquality() && "Cannot fold icmp gt/lt");
01096 
01097   auto getConstant = [&I, this](bool IsTrue) {
01098     if (I.getPredicate() == I.ICMP_NE)
01099       IsTrue = !IsTrue;
01100     return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
01101   };
01102 
01103   auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
01104     if (I.getPredicate() == I.ICMP_NE)
01105       Pred = CmpInst::getInversePredicate(Pred);
01106     return new ICmpInst(Pred, LHS, RHS);
01107   };
01108 
01109   APInt AP1 = CI1->getValue();
01110   APInt AP2 = CI2->getValue();
01111 
01112   assert(AP2 != 0 && "Handled in InstSimplify");
01113 
01114   unsigned AP2TrailingZeros = AP2.countTrailingZeros();
01115 
01116   if (!AP1 && AP2TrailingZeros != 0)
01117     return getICmp(I.ICMP_UGE, A,
01118                    ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
01119 
01120   if (AP1 == AP2)
01121     return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
01122 
01123   // Get the distance between the lowest bits that are set.
01124   int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
01125 
01126   if (Shift > 0 && AP2.shl(Shift) == AP1)
01127     return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
01128 
01129   // Shifting const2 will never be equal to const1.
01130   return getConstant(false);
01131 }
01132 
01133 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
01134 ///
01135 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
01136                                                           Instruction *LHSI,
01137                                                           ConstantInt *RHS) {
01138   const APInt &RHSV = RHS->getValue();
01139 
01140   switch (LHSI->getOpcode()) {
01141   case Instruction::Trunc:
01142     if (ICI.isEquality() && LHSI->hasOneUse()) {
01143       // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
01144       // of the high bits truncated out of x are known.
01145       unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
01146              SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
01147       APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
01148       computeKnownBits(LHSI->getOperand(0), KnownZero, KnownOne, 0, &ICI);
01149 
01150       // If all the high bits are known, we can do this xform.
01151       if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
01152         // Pull in the high bits from known-ones set.
01153         APInt NewRHS = RHS->getValue().zext(SrcBits);
01154         NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
01155         return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
01156                             Builder->getInt(NewRHS));
01157       }
01158     }
01159     break;
01160 
01161   case Instruction::Xor:         // (icmp pred (xor X, XorCst), CI)
01162     if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
01163       // If this is a comparison that tests the signbit (X < 0) or (x > -1),
01164       // fold the xor.
01165       if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
01166           (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
01167         Value *CompareVal = LHSI->getOperand(0);
01168 
01169         // If the sign bit of the XorCst is not set, there is no change to
01170         // the operation, just stop using the Xor.
01171         if (!XorCst->isNegative()) {
01172           ICI.setOperand(0, CompareVal);
01173           Worklist.Add(LHSI);
01174           return &ICI;
01175         }
01176 
01177         // Was the old condition true if the operand is positive?
01178         bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
01179 
01180         // If so, the new one isn't.
01181         isTrueIfPositive ^= true;
01182 
01183         if (isTrueIfPositive)
01184           return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
01185                               SubOne(RHS));
01186         else
01187           return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
01188                               AddOne(RHS));
01189       }
01190 
01191       if (LHSI->hasOneUse()) {
01192         // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
01193         if (!ICI.isEquality() && XorCst->getValue().isSignBit()) {
01194           const APInt &SignBit = XorCst->getValue();
01195           ICmpInst::Predicate Pred = ICI.isSigned()
01196                                          ? ICI.getUnsignedPredicate()
01197                                          : ICI.getSignedPredicate();
01198           return new ICmpInst(Pred, LHSI->getOperand(0),
01199                               Builder->getInt(RHSV ^ SignBit));
01200         }
01201 
01202         // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
01203         if (!ICI.isEquality() && XorCst->isMaxValue(true)) {
01204           const APInt &NotSignBit = XorCst->getValue();
01205           ICmpInst::Predicate Pred = ICI.isSigned()
01206                                          ? ICI.getUnsignedPredicate()
01207                                          : ICI.getSignedPredicate();
01208           Pred = ICI.getSwappedPredicate(Pred);
01209           return new ICmpInst(Pred, LHSI->getOperand(0),
01210                               Builder->getInt(RHSV ^ NotSignBit));
01211         }
01212       }
01213 
01214       // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
01215       //   iff -C is a power of 2
01216       if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
01217           XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
01218         return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst);
01219 
01220       // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
01221       //   iff -C is a power of 2
01222       if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
01223           XorCst->getValue() == -RHSV && RHSV.isPowerOf2())
01224         return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst);
01225     }
01226     break;
01227   case Instruction::And:         // (icmp pred (and X, AndCst), RHS)
01228     if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
01229         LHSI->getOperand(0)->hasOneUse()) {
01230       ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1));
01231 
01232       // If the LHS is an AND of a truncating cast, we can widen the
01233       // and/compare to be the input width without changing the value
01234       // produced, eliminating a cast.
01235       if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
01236         // We can do this transformation if either the AND constant does not
01237         // have its sign bit set or if it is an equality comparison.
01238         // Extending a relational comparison when we're checking the sign
01239         // bit would not work.
01240         if (ICI.isEquality() ||
01241             (!AndCst->isNegative() && RHSV.isNonNegative())) {
01242           Value *NewAnd =
01243             Builder->CreateAnd(Cast->getOperand(0),
01244                                ConstantExpr::getZExt(AndCst, Cast->getSrcTy()));
01245           NewAnd->takeName(LHSI);
01246           return new ICmpInst(ICI.getPredicate(), NewAnd,
01247                               ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
01248         }
01249       }
01250 
01251       // If the LHS is an AND of a zext, and we have an equality compare, we can
01252       // shrink the and/compare to the smaller type, eliminating the cast.
01253       if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
01254         IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
01255         // Make sure we don't compare the upper bits, SimplifyDemandedBits
01256         // should fold the icmp to true/false in that case.
01257         if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
01258           Value *NewAnd =
01259             Builder->CreateAnd(Cast->getOperand(0),
01260                                ConstantExpr::getTrunc(AndCst, Ty));
01261           NewAnd->takeName(LHSI);
01262           return new ICmpInst(ICI.getPredicate(), NewAnd,
01263                               ConstantExpr::getTrunc(RHS, Ty));
01264         }
01265       }
01266 
01267       // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
01268       // could exist), turn it into (X & (C2 << C1)) != (C3 << C1).  This
01269       // happens a LOT in code produced by the C front-end, for bitfield
01270       // access.
01271       BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
01272       if (Shift && !Shift->isShift())
01273         Shift = nullptr;
01274 
01275       ConstantInt *ShAmt;
01276       ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : nullptr;
01277 
01278       // This seemingly simple opportunity to fold away a shift turns out to
01279       // be rather complicated. See PR17827
01280       // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details.
01281       if (ShAmt) {
01282         bool CanFold = false;
01283         unsigned ShiftOpcode = Shift->getOpcode();
01284         if (ShiftOpcode == Instruction::AShr) {
01285           // There may be some constraints that make this possible,
01286           // but nothing simple has been discovered yet.
01287           CanFold = false;
01288         } else if (ShiftOpcode == Instruction::Shl) {
01289           // For a left shift, we can fold if the comparison is not signed.
01290           // We can also fold a signed comparison if the mask value and
01291           // comparison value are not negative. These constraints may not be
01292           // obvious, but we can prove that they are correct using an SMT
01293           // solver.
01294           if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative()))
01295             CanFold = true;
01296         } else if (ShiftOpcode == Instruction::LShr) {
01297           // For a logical right shift, we can fold if the comparison is not
01298           // signed. We can also fold a signed comparison if the shifted mask
01299           // value and the shifted comparison value are not negative.
01300           // These constraints may not be obvious, but we can prove that they
01301           // are correct using an SMT solver.
01302           if (!ICI.isSigned())
01303             CanFold = true;
01304           else {
01305             ConstantInt *ShiftedAndCst =
01306               cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt));
01307             ConstantInt *ShiftedRHSCst =
01308               cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt));
01309             
01310             if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative())
01311               CanFold = true;
01312           }
01313         }
01314 
01315         if (CanFold) {
01316           Constant *NewCst;
01317           if (ShiftOpcode == Instruction::Shl)
01318             NewCst = ConstantExpr::getLShr(RHS, ShAmt);
01319           else
01320             NewCst = ConstantExpr::getShl(RHS, ShAmt);
01321 
01322           // Check to see if we are shifting out any of the bits being
01323           // compared.
01324           if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) {
01325             // If we shifted bits out, the fold is not going to work out.
01326             // As a special case, check to see if this means that the
01327             // result is always true or false now.
01328             if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
01329               return ReplaceInstUsesWith(ICI, Builder->getFalse());
01330             if (ICI.getPredicate() == ICmpInst::ICMP_NE)
01331               return ReplaceInstUsesWith(ICI, Builder->getTrue());
01332           } else {
01333             ICI.setOperand(1, NewCst);
01334             Constant *NewAndCst;
01335             if (ShiftOpcode == Instruction::Shl)
01336               NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt);
01337             else
01338               NewAndCst = ConstantExpr::getShl(AndCst, ShAmt);
01339             LHSI->setOperand(1, NewAndCst);
01340             LHSI->setOperand(0, Shift->getOperand(0));
01341             Worklist.Add(Shift); // Shift is dead.
01342             return &ICI;
01343           }
01344         }
01345       }
01346 
01347       // Turn ((X >> Y) & C) == 0  into  (X & (C << Y)) == 0.  The later is
01348       // preferable because it allows the C<<Y expression to be hoisted out
01349       // of a loop if Y is invariant and X is not.
01350       if (Shift && Shift->hasOneUse() && RHSV == 0 &&
01351           ICI.isEquality() && !Shift->isArithmeticShift() &&
01352           !isa<Constant>(Shift->getOperand(0))) {
01353         // Compute C << Y.
01354         Value *NS;
01355         if (Shift->getOpcode() == Instruction::LShr) {
01356           NS = Builder->CreateShl(AndCst, Shift->getOperand(1));
01357         } else {
01358           // Insert a logical shift.
01359           NS = Builder->CreateLShr(AndCst, Shift->getOperand(1));
01360         }
01361 
01362         // Compute X & (C << Y).
01363         Value *NewAnd =
01364           Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
01365 
01366         ICI.setOperand(0, NewAnd);
01367         return &ICI;
01368       }
01369 
01370       // (icmp pred (and (or (lshr X, Y), X), 1), 0) -->
01371       //    (icmp pred (and X, (or (shl 1, Y), 1), 0))
01372       //
01373       // iff pred isn't signed
01374       {
01375         Value *X, *Y, *LShr;
01376         if (!ICI.isSigned() && RHSV == 0) {
01377           if (match(LHSI->getOperand(1), m_One())) {
01378             Constant *One = cast<Constant>(LHSI->getOperand(1));
01379             Value *Or = LHSI->getOperand(0);
01380             if (match(Or, m_Or(m_Value(LShr), m_Value(X))) &&
01381                 match(LShr, m_LShr(m_Specific(X), m_Value(Y)))) {
01382               unsigned UsesRemoved = 0;
01383               if (LHSI->hasOneUse())
01384                 ++UsesRemoved;
01385               if (Or->hasOneUse())
01386                 ++UsesRemoved;
01387               if (LShr->hasOneUse())
01388                 ++UsesRemoved;
01389               Value *NewOr = nullptr;
01390               // Compute X & ((1 << Y) | 1)
01391               if (auto *C = dyn_cast<Constant>(Y)) {
01392                 if (UsesRemoved >= 1)
01393                   NewOr =
01394                       ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
01395               } else {
01396                 if (UsesRemoved >= 3)
01397                   NewOr = Builder->CreateOr(Builder->CreateShl(One, Y,
01398                                                                LShr->getName(),
01399                                                                /*HasNUW=*/true),
01400                                             One, Or->getName());
01401               }
01402               if (NewOr) {
01403                 Value *NewAnd = Builder->CreateAnd(X, NewOr, LHSI->getName());
01404                 ICI.setOperand(0, NewAnd);
01405                 return &ICI;
01406               }
01407             }
01408           }
01409         }
01410       }
01411 
01412       // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any
01413       // bit set in (X & AndCst) will produce a result greater than RHSV.
01414       if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
01415         unsigned NTZ = AndCst->getValue().countTrailingZeros();
01416         if ((NTZ < AndCst->getBitWidth()) &&
01417             APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV))
01418           return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
01419                               Constant::getNullValue(RHS->getType()));
01420       }
01421     }
01422 
01423     // Try to optimize things like "A[i]&42 == 0" to index computations.
01424     if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
01425       if (GetElementPtrInst *GEP =
01426           dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
01427         if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
01428           if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
01429               !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
01430             ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
01431             if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
01432               return Res;
01433           }
01434     }
01435 
01436     // X & -C == -C -> X >  u ~C
01437     // X & -C != -C -> X <= u ~C
01438     //   iff C is a power of 2
01439     if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
01440       return new ICmpInst(
01441           ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
01442                                                   : ICmpInst::ICMP_ULE,
01443           LHSI->getOperand(0), SubOne(RHS));
01444     break;
01445 
01446   case Instruction::Or: {
01447     if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
01448       break;
01449     Value *P, *Q;
01450     if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
01451       // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
01452       // -> and (icmp eq P, null), (icmp eq Q, null).
01453       Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
01454                                         Constant::getNullValue(P->getType()));
01455       Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
01456                                         Constant::getNullValue(Q->getType()));
01457       Instruction *Op;
01458       if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
01459         Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
01460       else
01461         Op = BinaryOperator::CreateOr(ICIP, ICIQ);
01462       return Op;
01463     }
01464     break;
01465   }
01466 
01467   case Instruction::Mul: {       // (icmp pred (mul X, Val), CI)
01468     ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
01469     if (!Val) break;
01470 
01471     // If this is a signed comparison to 0 and the mul is sign preserving,
01472     // use the mul LHS operand instead.
01473     ICmpInst::Predicate pred = ICI.getPredicate();
01474     if (isSignTest(pred, RHS) && !Val->isZero() &&
01475         cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
01476       return new ICmpInst(Val->isNegative() ?
01477                           ICmpInst::getSwappedPredicate(pred) : pred,
01478                           LHSI->getOperand(0),
01479                           Constant::getNullValue(RHS->getType()));
01480 
01481     break;
01482   }
01483 
01484   case Instruction::Shl: {       // (icmp pred (shl X, ShAmt), CI)
01485     uint32_t TypeBits = RHSV.getBitWidth();
01486     ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
01487     if (!ShAmt) {
01488       Value *X;
01489       // (1 << X) pred P2 -> X pred Log2(P2)
01490       if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
01491         bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
01492         ICmpInst::Predicate Pred = ICI.getPredicate();
01493         if (ICI.isUnsigned()) {
01494           if (!RHSVIsPowerOf2) {
01495             // (1 << X) <  30 -> X <= 4
01496             // (1 << X) <= 30 -> X <= 4
01497             // (1 << X) >= 30 -> X >  4
01498             // (1 << X) >  30 -> X >  4
01499             if (Pred == ICmpInst::ICMP_ULT)
01500               Pred = ICmpInst::ICMP_ULE;
01501             else if (Pred == ICmpInst::ICMP_UGE)
01502               Pred = ICmpInst::ICMP_UGT;
01503           }
01504           unsigned RHSLog2 = RHSV.logBase2();
01505 
01506           // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
01507           // (1 << X) <  2147483648 -> X <  31 -> X != 31
01508           if (RHSLog2 == TypeBits-1) {
01509             if (Pred == ICmpInst::ICMP_UGE)
01510               Pred = ICmpInst::ICMP_EQ;
01511             else if (Pred == ICmpInst::ICMP_ULT)
01512               Pred = ICmpInst::ICMP_NE;
01513           }
01514 
01515           return new ICmpInst(Pred, X,
01516                               ConstantInt::get(RHS->getType(), RHSLog2));
01517         } else if (ICI.isSigned()) {
01518           if (RHSV.isAllOnesValue()) {
01519             // (1 << X) <= -1 -> X == 31
01520             if (Pred == ICmpInst::ICMP_SLE)
01521               return new ICmpInst(ICmpInst::ICMP_EQ, X,
01522                                   ConstantInt::get(RHS->getType(), TypeBits-1));
01523 
01524             // (1 << X) >  -1 -> X != 31
01525             if (Pred == ICmpInst::ICMP_SGT)
01526               return new ICmpInst(ICmpInst::ICMP_NE, X,
01527                                   ConstantInt::get(RHS->getType(), TypeBits-1));
01528           } else if (!RHSV) {
01529             // (1 << X) <  0 -> X == 31
01530             // (1 << X) <= 0 -> X == 31
01531             if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
01532               return new ICmpInst(ICmpInst::ICMP_EQ, X,
01533                                   ConstantInt::get(RHS->getType(), TypeBits-1));
01534 
01535             // (1 << X) >= 0 -> X != 31
01536             // (1 << X) >  0 -> X != 31
01537             if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
01538               return new ICmpInst(ICmpInst::ICMP_NE, X,
01539                                   ConstantInt::get(RHS->getType(), TypeBits-1));
01540           }
01541         } else if (ICI.isEquality()) {
01542           if (RHSVIsPowerOf2)
01543             return new ICmpInst(
01544                 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
01545         }
01546       }
01547       break;
01548     }
01549 
01550     // Check that the shift amount is in range.  If not, don't perform
01551     // undefined shifts.  When the shift is visited it will be
01552     // simplified.
01553     if (ShAmt->uge(TypeBits))
01554       break;
01555 
01556     if (ICI.isEquality()) {
01557       // If we are comparing against bits always shifted out, the
01558       // comparison cannot succeed.
01559       Constant *Comp =
01560         ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
01561                                                                  ShAmt);
01562       if (Comp != RHS) {// Comparing against a bit that we know is zero.
01563         bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
01564         Constant *Cst = Builder->getInt1(IsICMP_NE);
01565         return ReplaceInstUsesWith(ICI, Cst);
01566       }
01567 
01568       // If the shift is NUW, then it is just shifting out zeros, no need for an
01569       // AND.
01570       if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
01571         return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
01572                             ConstantExpr::getLShr(RHS, ShAmt));
01573 
01574       // If the shift is NSW and we compare to 0, then it is just shifting out
01575       // sign bits, no need for an AND either.
01576       if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
01577         return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
01578                             ConstantExpr::getLShr(RHS, ShAmt));
01579 
01580       if (LHSI->hasOneUse()) {
01581         // Otherwise strength reduce the shift into an and.
01582         uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
01583         Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
01584                                                           TypeBits - ShAmtVal));
01585 
01586         Value *And =
01587           Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
01588         return new ICmpInst(ICI.getPredicate(), And,
01589                             ConstantExpr::getLShr(RHS, ShAmt));
01590       }
01591     }
01592 
01593     // If this is a signed comparison to 0 and the shift is sign preserving,
01594     // use the shift LHS operand instead.
01595     ICmpInst::Predicate pred = ICI.getPredicate();
01596     if (isSignTest(pred, RHS) &&
01597         cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
01598       return new ICmpInst(pred,
01599                           LHSI->getOperand(0),
01600                           Constant::getNullValue(RHS->getType()));
01601 
01602     // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
01603     bool TrueIfSigned = false;
01604     if (LHSI->hasOneUse() &&
01605         isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
01606       // (X << 31) <s 0  --> (X&1) != 0
01607       Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
01608                                         APInt::getOneBitSet(TypeBits,
01609                                             TypeBits-ShAmt->getZExtValue()-1));
01610       Value *And =
01611         Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
01612       return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
01613                           And, Constant::getNullValue(And->getType()));
01614     }
01615 
01616     // Transform (icmp pred iM (shl iM %v, N), CI)
01617     // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
01618     // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
01619     // This enables to get rid of the shift in favor of a trunc which can be
01620     // free on the target. It has the additional benefit of comparing to a
01621     // smaller constant, which will be target friendly.
01622     unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
01623     if (LHSI->hasOneUse() &&
01624         Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
01625       Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
01626       Constant *NCI = ConstantExpr::getTrunc(
01627                         ConstantExpr::getAShr(RHS,
01628                           ConstantInt::get(RHS->getType(), Amt)),
01629                         NTy);
01630       return new ICmpInst(ICI.getPredicate(),
01631                           Builder->CreateTrunc(LHSI->getOperand(0), NTy),
01632                           NCI);
01633     }
01634 
01635     break;
01636   }
01637 
01638   case Instruction::LShr:         // (icmp pred (shr X, ShAmt), CI)
01639   case Instruction::AShr: {
01640     // Handle equality comparisons of shift-by-constant.
01641     BinaryOperator *BO = cast<BinaryOperator>(LHSI);
01642     if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
01643       if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
01644         return Res;
01645     }
01646 
01647     // Handle exact shr's.
01648     if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
01649       if (RHSV.isMinValue())
01650         return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
01651     }
01652     break;
01653   }
01654 
01655   case Instruction::SDiv:
01656   case Instruction::UDiv:
01657     // Fold: icmp pred ([us]div X, C1), C2 -> range test
01658     // Fold this div into the comparison, producing a range check.
01659     // Determine, based on the divide type, what the range is being
01660     // checked.  If there is an overflow on the low or high side, remember
01661     // it, otherwise compute the range [low, hi) bounding the new value.
01662     // See: InsertRangeTest above for the kinds of replacements possible.
01663     if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
01664       if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
01665                                           DivRHS))
01666         return R;
01667     break;
01668 
01669   case Instruction::Sub: {
01670     ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
01671     if (!LHSC) break;
01672     const APInt &LHSV = LHSC->getValue();
01673 
01674     // C1-X <u C2 -> (X|(C2-1)) == C1
01675     //   iff C1 & (C2-1) == C2-1
01676     //       C2 is a power of 2
01677     if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
01678         RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
01679       return new ICmpInst(ICmpInst::ICMP_EQ,
01680                           Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
01681                           LHSC);
01682 
01683     // C1-X >u C2 -> (X|C2) != C1
01684     //   iff C1 & C2 == C2
01685     //       C2+1 is a power of 2
01686     if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
01687         (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
01688       return new ICmpInst(ICmpInst::ICMP_NE,
01689                           Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
01690     break;
01691   }
01692 
01693   case Instruction::Add:
01694     // Fold: icmp pred (add X, C1), C2
01695     if (!ICI.isEquality()) {
01696       ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
01697       if (!LHSC) break;
01698       const APInt &LHSV = LHSC->getValue();
01699 
01700       ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
01701                             .subtract(LHSV);
01702 
01703       if (ICI.isSigned()) {
01704         if (CR.getLower().isSignBit()) {
01705           return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
01706                               Builder->getInt(CR.getUpper()));
01707         } else if (CR.getUpper().isSignBit()) {
01708           return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
01709                               Builder->getInt(CR.getLower()));
01710         }
01711       } else {
01712         if (CR.getLower().isMinValue()) {
01713           return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
01714                               Builder->getInt(CR.getUpper()));
01715         } else if (CR.getUpper().isMinValue()) {
01716           return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
01717                               Builder->getInt(CR.getLower()));
01718         }
01719       }
01720 
01721       // X-C1 <u C2 -> (X & -C2) == C1
01722       //   iff C1 & (C2-1) == 0
01723       //       C2 is a power of 2
01724       if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
01725           RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
01726         return new ICmpInst(ICmpInst::ICMP_EQ,
01727                             Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
01728                             ConstantExpr::getNeg(LHSC));
01729 
01730       // X-C1 >u C2 -> (X & ~C2) != C1
01731       //   iff C1 & C2 == 0
01732       //       C2+1 is a power of 2
01733       if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
01734           (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
01735         return new ICmpInst(ICmpInst::ICMP_NE,
01736                             Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
01737                             ConstantExpr::getNeg(LHSC));
01738     }
01739     break;
01740   }
01741 
01742   // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
01743   if (ICI.isEquality()) {
01744     bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
01745 
01746     // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
01747     // the second operand is a constant, simplify a bit.
01748     if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
01749       switch (BO->getOpcode()) {
01750       case Instruction::SRem:
01751         // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
01752         if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
01753           const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
01754           if (V.sgt(1) && V.isPowerOf2()) {
01755             Value *NewRem =
01756               Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
01757                                   BO->getName());
01758             return new ICmpInst(ICI.getPredicate(), NewRem,
01759                                 Constant::getNullValue(BO->getType()));
01760           }
01761         }
01762         break;
01763       case Instruction::Add:
01764         // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
01765         if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
01766           if (BO->hasOneUse())
01767             return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01768                                 ConstantExpr::getSub(RHS, BOp1C));
01769         } else if (RHSV == 0) {
01770           // Replace ((add A, B) != 0) with (A != -B) if A or B is
01771           // efficiently invertible, or if the add has just this one use.
01772           Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
01773 
01774           if (Value *NegVal = dyn_castNegVal(BOp1))
01775             return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
01776           if (Value *NegVal = dyn_castNegVal(BOp0))
01777             return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
01778           if (BO->hasOneUse()) {
01779             Value *Neg = Builder->CreateNeg(BOp1);
01780             Neg->takeName(BO);
01781             return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
01782           }
01783         }
01784         break;
01785       case Instruction::Xor:
01786         // For the xor case, we can xor two constants together, eliminating
01787         // the explicit xor.
01788         if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
01789           return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01790                               ConstantExpr::getXor(RHS, BOC));
01791         } else if (RHSV == 0) {
01792           // Replace ((xor A, B) != 0) with (A != B)
01793           return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01794                               BO->getOperand(1));
01795         }
01796         break;
01797       case Instruction::Sub:
01798         // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
01799         if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
01800           if (BO->hasOneUse())
01801             return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
01802                                 ConstantExpr::getSub(BOp0C, RHS));
01803         } else if (RHSV == 0) {
01804           // Replace ((sub A, B) != 0) with (A != B)
01805           return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01806                               BO->getOperand(1));
01807         }
01808         break;
01809       case Instruction::Or:
01810         // If bits are being or'd in that are not present in the constant we
01811         // are comparing against, then the comparison could never succeed!
01812         if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
01813           Constant *NotCI = ConstantExpr::getNot(RHS);
01814           if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
01815             return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
01816         }
01817         break;
01818 
01819       case Instruction::And:
01820         if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
01821           // If bits are being compared against that are and'd out, then the
01822           // comparison can never succeed!
01823           if ((RHSV & ~BOC->getValue()) != 0)
01824             return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
01825 
01826           // If we have ((X & C) == C), turn it into ((X & C) != 0).
01827           if (RHS == BOC && RHSV.isPowerOf2())
01828             return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
01829                                 ICmpInst::ICMP_NE, LHSI,
01830                                 Constant::getNullValue(RHS->getType()));
01831 
01832           // Don't perform the following transforms if the AND has multiple uses
01833           if (!BO->hasOneUse())
01834             break;
01835 
01836           // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
01837           if (BOC->getValue().isSignBit()) {
01838             Value *X = BO->getOperand(0);
01839             Constant *Zero = Constant::getNullValue(X->getType());
01840             ICmpInst::Predicate pred = isICMP_NE ?
01841               ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
01842             return new ICmpInst(pred, X, Zero);
01843           }
01844 
01845           // ((X & ~7) == 0) --> X < 8
01846           if (RHSV == 0 && isHighOnes(BOC)) {
01847             Value *X = BO->getOperand(0);
01848             Constant *NegX = ConstantExpr::getNeg(BOC);
01849             ICmpInst::Predicate pred = isICMP_NE ?
01850               ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
01851             return new ICmpInst(pred, X, NegX);
01852           }
01853         }
01854         break;
01855       case Instruction::Mul:
01856         if (RHSV == 0 && BO->hasNoSignedWrap()) {
01857           if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
01858             // The trivial case (mul X, 0) is handled by InstSimplify
01859             // General case : (mul X, C) != 0 iff X != 0
01860             //                (mul X, C) == 0 iff X == 0
01861             if (!BOC->isZero())
01862               return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01863                                   Constant::getNullValue(RHS->getType()));
01864           }
01865         }
01866         break;
01867       default: break;
01868       }
01869     } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
01870       // Handle icmp {eq|ne} <intrinsic>, intcst.
01871       switch (II->getIntrinsicID()) {
01872       case Intrinsic::bswap:
01873         Worklist.Add(II);
01874         ICI.setOperand(0, II->getArgOperand(0));
01875         ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
01876         return &ICI;
01877       case Intrinsic::ctlz:
01878       case Intrinsic::cttz:
01879         // ctz(A) == bitwidth(a)  ->  A == 0 and likewise for !=
01880         if (RHSV == RHS->getType()->getBitWidth()) {
01881           Worklist.Add(II);
01882           ICI.setOperand(0, II->getArgOperand(0));
01883           ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
01884           return &ICI;
01885         }
01886         break;
01887       case Intrinsic::ctpop:
01888         // popcount(A) == 0  ->  A == 0 and likewise for !=
01889         if (RHS->isZero()) {
01890           Worklist.Add(II);
01891           ICI.setOperand(0, II->getArgOperand(0));
01892           ICI.setOperand(1, RHS);
01893           return &ICI;
01894         }
01895         break;
01896       default:
01897         break;
01898       }
01899     }
01900   }
01901   return nullptr;
01902 }
01903 
01904 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
01905 /// We only handle extending casts so far.
01906 ///
01907 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
01908   const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
01909   Value *LHSCIOp        = LHSCI->getOperand(0);
01910   Type *SrcTy     = LHSCIOp->getType();
01911   Type *DestTy    = LHSCI->getType();
01912   Value *RHSCIOp;
01913 
01914   // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
01915   // integer type is the same size as the pointer type.
01916   if (DL && LHSCI->getOpcode() == Instruction::PtrToInt &&
01917       DL->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
01918     Value *RHSOp = nullptr;
01919     if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
01920       RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
01921     } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
01922       RHSOp = RHSC->getOperand(0);
01923       // If the pointer types don't match, insert a bitcast.
01924       if (LHSCIOp->getType() != RHSOp->getType())
01925         RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
01926     }
01927 
01928     if (RHSOp)
01929       return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
01930   }
01931 
01932   // The code below only handles extension cast instructions, so far.
01933   // Enforce this.
01934   if (LHSCI->getOpcode() != Instruction::ZExt &&
01935       LHSCI->getOpcode() != Instruction::SExt)
01936     return nullptr;
01937 
01938   bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
01939   bool isSignedCmp = ICI.isSigned();
01940 
01941   if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
01942     // Not an extension from the same type?
01943     RHSCIOp = CI->getOperand(0);
01944     if (RHSCIOp->getType() != LHSCIOp->getType())
01945       return nullptr;
01946 
01947     // If the signedness of the two casts doesn't agree (i.e. one is a sext
01948     // and the other is a zext), then we can't handle this.
01949     if (CI->getOpcode() != LHSCI->getOpcode())
01950       return nullptr;
01951 
01952     // Deal with equality cases early.
01953     if (ICI.isEquality())
01954       return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
01955 
01956     // A signed comparison of sign extended values simplifies into a
01957     // signed comparison.
01958     if (isSignedCmp && isSignedExt)
01959       return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
01960 
01961     // The other three cases all fold into an unsigned comparison.
01962     return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
01963   }
01964 
01965   // If we aren't dealing with a constant on the RHS, exit early
01966   ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
01967   if (!CI)
01968     return nullptr;
01969 
01970   // Compute the constant that would happen if we truncated to SrcTy then
01971   // reextended to DestTy.
01972   Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
01973   Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
01974                                                 Res1, DestTy);
01975 
01976   // If the re-extended constant didn't change...
01977   if (Res2 == CI) {
01978     // Deal with equality cases early.
01979     if (ICI.isEquality())
01980       return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
01981 
01982     // A signed comparison of sign extended values simplifies into a
01983     // signed comparison.
01984     if (isSignedExt && isSignedCmp)
01985       return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
01986 
01987     // The other three cases all fold into an unsigned comparison.
01988     return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
01989   }
01990 
01991   // The re-extended constant changed so the constant cannot be represented
01992   // in the shorter type. Consequently, we cannot emit a simple comparison.
01993   // All the cases that fold to true or false will have already been handled
01994   // by SimplifyICmpInst, so only deal with the tricky case.
01995 
01996   if (isSignedCmp || !isSignedExt)
01997     return nullptr;
01998 
01999   // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
02000   // should have been folded away previously and not enter in here.
02001 
02002   // We're performing an unsigned comp with a sign extended value.
02003   // This is true if the input is >= 0. [aka >s -1]
02004   Constant *NegOne = Constant::getAllOnesValue(SrcTy);
02005   Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
02006 
02007   // Finally, return the value computed.
02008   if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
02009     return ReplaceInstUsesWith(ICI, Result);
02010 
02011   assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
02012   return BinaryOperator::CreateNot(Result);
02013 }
02014 
02015 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
02016 ///   I = icmp ugt (add (add A, B), CI2), CI1
02017 /// If this is of the form:
02018 ///   sum = a + b
02019 ///   if (sum+128 >u 255)
02020 /// Then replace it with llvm.sadd.with.overflow.i8.
02021 ///
02022 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
02023                                           ConstantInt *CI2, ConstantInt *CI1,
02024                                           InstCombiner &IC) {
02025   // The transformation we're trying to do here is to transform this into an
02026   // llvm.sadd.with.overflow.  To do this, we have to replace the original add
02027   // with a narrower add, and discard the add-with-constant that is part of the
02028   // range check (if we can't eliminate it, this isn't profitable).
02029 
02030   // In order to eliminate the add-with-constant, the compare can be its only
02031   // use.
02032   Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
02033   if (!AddWithCst->hasOneUse()) return nullptr;
02034 
02035   // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
02036   if (!CI2->getValue().isPowerOf2()) return nullptr;
02037   unsigned NewWidth = CI2->getValue().countTrailingZeros();
02038   if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return nullptr;
02039 
02040   // The width of the new add formed is 1 more than the bias.
02041   ++NewWidth;
02042 
02043   // Check to see that CI1 is an all-ones value with NewWidth bits.
02044   if (CI1->getBitWidth() == NewWidth ||
02045       CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
02046     return nullptr;
02047 
02048   // This is only really a signed overflow check if the inputs have been
02049   // sign-extended; check for that condition. For example, if CI2 is 2^31 and
02050   // the operands of the add are 64 bits wide, we need at least 33 sign bits.
02051   unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
02052   if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
02053       IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
02054     return nullptr;
02055 
02056   // In order to replace the original add with a narrower
02057   // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
02058   // and truncates that discard the high bits of the add.  Verify that this is
02059   // the case.
02060   Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
02061   for (User *U : OrigAdd->users()) {
02062     if (U == AddWithCst) continue;
02063 
02064     // Only accept truncates for now.  We would really like a nice recursive
02065     // predicate like SimplifyDemandedBits, but which goes downwards the use-def
02066     // chain to see which bits of a value are actually demanded.  If the
02067     // original add had another add which was then immediately truncated, we
02068     // could still do the transformation.
02069     TruncInst *TI = dyn_cast<TruncInst>(U);
02070     if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
02071       return nullptr;
02072   }
02073 
02074   // If the pattern matches, truncate the inputs to the narrower type and
02075   // use the sadd_with_overflow intrinsic to efficiently compute both the
02076   // result and the overflow bit.
02077   Module *M = I.getParent()->getParent()->getParent();
02078 
02079   Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
02080   Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
02081                                        NewType);
02082 
02083   InstCombiner::BuilderTy *Builder = IC.Builder;
02084 
02085   // Put the new code above the original add, in case there are any uses of the
02086   // add between the add and the compare.
02087   Builder->SetInsertPoint(OrigAdd);
02088 
02089   Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
02090   Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
02091   CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
02092   Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
02093   Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
02094 
02095   // The inner add was the result of the narrow add, zero extended to the
02096   // wider type.  Replace it with the result computed by the intrinsic.
02097   IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
02098 
02099   // The original icmp gets replaced with the overflow value.
02100   return ExtractValueInst::Create(Call, 1, "sadd.overflow");
02101 }
02102 
02103 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
02104                                      InstCombiner &IC) {
02105   // Don't bother doing this transformation for pointers, don't do it for
02106   // vectors.
02107   if (!isa<IntegerType>(OrigAddV->getType())) return nullptr;
02108 
02109   // If the add is a constant expr, then we don't bother transforming it.
02110   Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
02111   if (!OrigAdd) return nullptr;
02112 
02113   Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
02114 
02115   // Put the new code above the original add, in case there are any uses of the
02116   // add between the add and the compare.
02117   InstCombiner::BuilderTy *Builder = IC.Builder;
02118   Builder->SetInsertPoint(OrigAdd);
02119 
02120   Module *M = I.getParent()->getParent()->getParent();
02121   Type *Ty = LHS->getType();
02122   Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
02123   CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
02124   Value *Add = Builder->CreateExtractValue(Call, 0);
02125 
02126   IC.ReplaceInstUsesWith(*OrigAdd, Add);
02127 
02128   // The original icmp gets replaced with the overflow value.
02129   return ExtractValueInst::Create(Call, 1, "uadd.overflow");
02130 }
02131 
02132 /// \brief Recognize and process idiom involving test for multiplication
02133 /// overflow.
02134 ///
02135 /// The caller has matched a pattern of the form:
02136 ///   I = cmp u (mul(zext A, zext B), V
02137 /// The function checks if this is a test for overflow and if so replaces
02138 /// multiplication with call to 'mul.with.overflow' intrinsic.
02139 ///
02140 /// \param I Compare instruction.
02141 /// \param MulVal Result of 'mult' instruction.  It is one of the arguments of
02142 ///               the compare instruction.  Must be of integer type.
02143 /// \param OtherVal The other argument of compare instruction.
02144 /// \returns Instruction which must replace the compare instruction, NULL if no
02145 ///          replacement required.
02146 static Instruction *ProcessUMulZExtIdiom(ICmpInst &I, Value *MulVal,
02147                                          Value *OtherVal, InstCombiner &IC) {
02148   // Don't bother doing this transformation for pointers, don't do it for
02149   // vectors.
02150   if (!isa<IntegerType>(MulVal->getType()))
02151     return nullptr;
02152 
02153   assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
02154   assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
02155   Instruction *MulInstr = cast<Instruction>(MulVal);
02156   assert(MulInstr->getOpcode() == Instruction::Mul);
02157 
02158   Instruction *LHS = cast<Instruction>(MulInstr->getOperand(0)),
02159               *RHS = cast<Instruction>(MulInstr->getOperand(1));
02160   assert(LHS->getOpcode() == Instruction::ZExt);
02161   assert(RHS->getOpcode() == Instruction::ZExt);
02162   Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
02163 
02164   // Calculate type and width of the result produced by mul.with.overflow.
02165   Type *TyA = A->getType(), *TyB = B->getType();
02166   unsigned WidthA = TyA->getPrimitiveSizeInBits(),
02167            WidthB = TyB->getPrimitiveSizeInBits();
02168   unsigned MulWidth;
02169   Type *MulType;
02170   if (WidthB > WidthA) {
02171     MulWidth = WidthB;
02172     MulType = TyB;
02173   } else {
02174     MulWidth = WidthA;
02175     MulType = TyA;
02176   }
02177 
02178   // In order to replace the original mul with a narrower mul.with.overflow,
02179   // all uses must ignore upper bits of the product.  The number of used low
02180   // bits must be not greater than the width of mul.with.overflow.
02181   if (MulVal->hasNUsesOrMore(2))
02182     for (User *U : MulVal->users()) {
02183       if (U == &I)
02184         continue;
02185       if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
02186         // Check if truncation ignores bits above MulWidth.
02187         unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
02188         if (TruncWidth > MulWidth)
02189           return nullptr;
02190       } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
02191         // Check if AND ignores bits above MulWidth.
02192         if (BO->getOpcode() != Instruction::And)
02193           return nullptr;
02194         if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
02195           const APInt &CVal = CI->getValue();
02196           if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
02197             return nullptr;
02198         }
02199       } else {
02200         // Other uses prohibit this transformation.
02201         return nullptr;
02202       }
02203     }
02204 
02205   // Recognize patterns
02206   switch (I.getPredicate()) {
02207   case ICmpInst::ICMP_EQ:
02208   case ICmpInst::ICMP_NE:
02209     // Recognize pattern:
02210     //   mulval = mul(zext A, zext B)
02211     //   cmp eq/neq mulval, zext trunc mulval
02212     if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
02213       if (Zext->hasOneUse()) {
02214         Value *ZextArg = Zext->getOperand(0);
02215         if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
02216           if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
02217             break; //Recognized
02218       }
02219 
02220     // Recognize pattern:
02221     //   mulval = mul(zext A, zext B)
02222     //   cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
02223     ConstantInt *CI;
02224     Value *ValToMask;
02225     if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
02226       if (ValToMask != MulVal)
02227         return nullptr;
02228       const APInt &CVal = CI->getValue() + 1;
02229       if (CVal.isPowerOf2()) {
02230         unsigned MaskWidth = CVal.logBase2();
02231         if (MaskWidth == MulWidth)
02232           break; // Recognized
02233       }
02234     }
02235     return nullptr;
02236 
02237   case ICmpInst::ICMP_UGT:
02238     // Recognize pattern:
02239     //   mulval = mul(zext A, zext B)
02240     //   cmp ugt mulval, max
02241     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
02242       APInt MaxVal = APInt::getMaxValue(MulWidth);
02243       MaxVal = MaxVal.zext(CI->getBitWidth());
02244       if (MaxVal.eq(CI->getValue()))
02245         break; // Recognized
02246     }
02247     return nullptr;
02248 
02249   case ICmpInst::ICMP_UGE:
02250     // Recognize pattern:
02251     //   mulval = mul(zext A, zext B)
02252     //   cmp uge mulval, max+1
02253     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
02254       APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
02255       if (MaxVal.eq(CI->getValue()))
02256         break; // Recognized
02257     }
02258     return nullptr;
02259 
02260   case ICmpInst::ICMP_ULE:
02261     // Recognize pattern:
02262     //   mulval = mul(zext A, zext B)
02263     //   cmp ule mulval, max
02264     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
02265       APInt MaxVal = APInt::getMaxValue(MulWidth);
02266       MaxVal = MaxVal.zext(CI->getBitWidth());
02267       if (MaxVal.eq(CI->getValue()))
02268         break; // Recognized
02269     }
02270     return nullptr;
02271 
02272   case ICmpInst::ICMP_ULT:
02273     // Recognize pattern:
02274     //   mulval = mul(zext A, zext B)
02275     //   cmp ule mulval, max + 1
02276     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
02277       APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
02278       if (MaxVal.eq(CI->getValue()))
02279         break; // Recognized
02280     }
02281     return nullptr;
02282 
02283   default:
02284     return nullptr;
02285   }
02286 
02287   InstCombiner::BuilderTy *Builder = IC.Builder;
02288   Builder->SetInsertPoint(MulInstr);
02289   Module *M = I.getParent()->getParent()->getParent();
02290 
02291   // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
02292   Value *MulA = A, *MulB = B;
02293   if (WidthA < MulWidth)
02294     MulA = Builder->CreateZExt(A, MulType);
02295   if (WidthB < MulWidth)
02296     MulB = Builder->CreateZExt(B, MulType);
02297   Value *F =
02298       Intrinsic::getDeclaration(M, Intrinsic::umul_with_overflow, MulType);
02299   CallInst *Call = Builder->CreateCall2(F, MulA, MulB, "umul");
02300   IC.Worklist.Add(MulInstr);
02301 
02302   // If there are uses of mul result other than the comparison, we know that
02303   // they are truncation or binary AND. Change them to use result of
02304   // mul.with.overflow and adjust properly mask/size.
02305   if (MulVal->hasNUsesOrMore(2)) {
02306     Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value");
02307     for (User *U : MulVal->users()) {
02308       if (U == &I || U == OtherVal)
02309         continue;
02310       if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
02311         if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
02312           IC.ReplaceInstUsesWith(*TI, Mul);
02313         else
02314           TI->setOperand(0, Mul);
02315       } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
02316         assert(BO->getOpcode() == Instruction::And);
02317         // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
02318         ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
02319         APInt ShortMask = CI->getValue().trunc(MulWidth);
02320         Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask);
02321         Instruction *Zext =
02322             cast<Instruction>(Builder->CreateZExt(ShortAnd, BO->getType()));
02323         IC.Worklist.Add(Zext);
02324         IC.ReplaceInstUsesWith(*BO, Zext);
02325       } else {
02326         llvm_unreachable("Unexpected Binary operation");
02327       }
02328       IC.Worklist.Add(cast<Instruction>(U));
02329     }
02330   }
02331   if (isa<Instruction>(OtherVal))
02332     IC.Worklist.Add(cast<Instruction>(OtherVal));
02333 
02334   // The original icmp gets replaced with the overflow value, maybe inverted
02335   // depending on predicate.
02336   bool Inverse = false;
02337   switch (I.getPredicate()) {
02338   case ICmpInst::ICMP_NE:
02339     break;
02340   case ICmpInst::ICMP_EQ:
02341     Inverse = true;
02342     break;
02343   case ICmpInst::ICMP_UGT:
02344   case ICmpInst::ICMP_UGE:
02345     if (I.getOperand(0) == MulVal)
02346       break;
02347     Inverse = true;
02348     break;
02349   case ICmpInst::ICMP_ULT:
02350   case ICmpInst::ICMP_ULE:
02351     if (I.getOperand(1) == MulVal)
02352       break;
02353     Inverse = true;
02354     break;
02355   default:
02356     llvm_unreachable("Unexpected predicate");
02357   }
02358   if (Inverse) {
02359     Value *Res = Builder->CreateExtractValue(Call, 1);
02360     return BinaryOperator::CreateNot(Res);
02361   }
02362 
02363   return ExtractValueInst::Create(Call, 1);
02364 }
02365 
02366 // DemandedBitsLHSMask - When performing a comparison against a constant,
02367 // it is possible that not all the bits in the LHS are demanded.  This helper
02368 // method computes the mask that IS demanded.
02369 static APInt DemandedBitsLHSMask(ICmpInst &I,
02370                                  unsigned BitWidth, bool isSignCheck) {
02371   if (isSignCheck)
02372     return APInt::getSignBit(BitWidth);
02373 
02374   ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
02375   if (!CI) return APInt::getAllOnesValue(BitWidth);
02376   const APInt &RHS = CI->getValue();
02377 
02378   switch (I.getPredicate()) {
02379   // For a UGT comparison, we don't care about any bits that
02380   // correspond to the trailing ones of the comparand.  The value of these
02381   // bits doesn't impact the outcome of the comparison, because any value
02382   // greater than the RHS must differ in a bit higher than these due to carry.
02383   case ICmpInst::ICMP_UGT: {
02384     unsigned trailingOnes = RHS.countTrailingOnes();
02385     APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
02386     return ~lowBitsSet;
02387   }
02388 
02389   // Similarly, for a ULT comparison, we don't care about the trailing zeros.
02390   // Any value less than the RHS must differ in a higher bit because of carries.
02391   case ICmpInst::ICMP_ULT: {
02392     unsigned trailingZeros = RHS.countTrailingZeros();
02393     APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
02394     return ~lowBitsSet;
02395   }
02396 
02397   default:
02398     return APInt::getAllOnesValue(BitWidth);
02399   }
02400 
02401 }
02402 
02403 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
02404 /// should be swapped.
02405 /// The decision is based on how many times these two operands are reused
02406 /// as subtract operands and their positions in those instructions.
02407 /// The rational is that several architectures use the same instruction for
02408 /// both subtract and cmp, thus it is better if the order of those operands
02409 /// match.
02410 /// \return true if Op0 and Op1 should be swapped.
02411 static bool swapMayExposeCSEOpportunities(const Value * Op0,
02412                                           const Value * Op1) {
02413   // Filter out pointer value as those cannot appears directly in subtract.
02414   // FIXME: we may want to go through inttoptrs or bitcasts.
02415   if (Op0->getType()->isPointerTy())
02416     return false;
02417   // Count every uses of both Op0 and Op1 in a subtract.
02418   // Each time Op0 is the first operand, count -1: swapping is bad, the
02419   // subtract has already the same layout as the compare.
02420   // Each time Op0 is the second operand, count +1: swapping is good, the
02421   // subtract has a different layout as the compare.
02422   // At the end, if the benefit is greater than 0, Op0 should come second to
02423   // expose more CSE opportunities.
02424   int GlobalSwapBenefits = 0;
02425   for (const User *U : Op0->users()) {
02426     const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
02427     if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
02428       continue;
02429     // If Op0 is the first argument, this is not beneficial to swap the
02430     // arguments.
02431     int LocalSwapBenefits = -1;
02432     unsigned Op1Idx = 1;
02433     if (BinOp->getOperand(Op1Idx) == Op0) {
02434       Op1Idx = 0;
02435       LocalSwapBenefits = 1;
02436     }
02437     if (BinOp->getOperand(Op1Idx) != Op1)
02438       continue;
02439     GlobalSwapBenefits += LocalSwapBenefits;
02440   }
02441   return GlobalSwapBenefits > 0;
02442 }
02443 
02444 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
02445   bool Changed = false;
02446   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
02447   unsigned Op0Cplxity = getComplexity(Op0);
02448   unsigned Op1Cplxity = getComplexity(Op1);
02449 
02450   /// Orders the operands of the compare so that they are listed from most
02451   /// complex to least complex.  This puts constants before unary operators,
02452   /// before binary operators.
02453   if (Op0Cplxity < Op1Cplxity ||
02454         (Op0Cplxity == Op1Cplxity &&
02455          swapMayExposeCSEOpportunities(Op0, Op1))) {
02456     I.swapOperands();
02457     std::swap(Op0, Op1);
02458     Changed = true;
02459   }
02460 
02461   if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL, TLI, DT, AT))
02462     return ReplaceInstUsesWith(I, V);
02463 
02464   // comparing -val or val with non-zero is the same as just comparing val
02465   // ie, abs(val) != 0 -> val != 0
02466   if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
02467   {
02468     Value *Cond, *SelectTrue, *SelectFalse;
02469     if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
02470                             m_Value(SelectFalse)))) {
02471       if (Value *V = dyn_castNegVal(SelectTrue)) {
02472         if (V == SelectFalse)
02473           return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
02474       }
02475       else if (Value *V = dyn_castNegVal(SelectFalse)) {
02476         if (V == SelectTrue)
02477           return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
02478       }
02479     }
02480   }
02481 
02482   Type *Ty = Op0->getType();
02483 
02484   // icmp's with boolean values can always be turned into bitwise operations
02485   if (Ty->isIntegerTy(1)) {
02486     switch (I.getPredicate()) {
02487     default: llvm_unreachable("Invalid icmp instruction!");
02488     case ICmpInst::ICMP_EQ: {               // icmp eq i1 A, B -> ~(A^B)
02489       Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
02490       return BinaryOperator::CreateNot(Xor);
02491     }
02492     case ICmpInst::ICMP_NE:                  // icmp eq i1 A, B -> A^B
02493       return BinaryOperator::CreateXor(Op0, Op1);
02494 
02495     case ICmpInst::ICMP_UGT:
02496       std::swap(Op0, Op1);                   // Change icmp ugt -> icmp ult
02497       // FALL THROUGH
02498     case ICmpInst::ICMP_ULT:{               // icmp ult i1 A, B -> ~A & B
02499       Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
02500       return BinaryOperator::CreateAnd(Not, Op1);
02501     }
02502     case ICmpInst::ICMP_SGT:
02503       std::swap(Op0, Op1);                   // Change icmp sgt -> icmp slt
02504       // FALL THROUGH
02505     case ICmpInst::ICMP_SLT: {               // icmp slt i1 A, B -> A & ~B
02506       Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
02507       return BinaryOperator::CreateAnd(Not, Op0);
02508     }
02509     case ICmpInst::ICMP_UGE:
02510       std::swap(Op0, Op1);                   // Change icmp uge -> icmp ule
02511       // FALL THROUGH
02512     case ICmpInst::ICMP_ULE: {               //  icmp ule i1 A, B -> ~A | B
02513       Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
02514       return BinaryOperator::CreateOr(Not, Op1);
02515     }
02516     case ICmpInst::ICMP_SGE:
02517       std::swap(Op0, Op1);                   // Change icmp sge -> icmp sle
02518       // FALL THROUGH
02519     case ICmpInst::ICMP_SLE: {               //  icmp sle i1 A, B -> A | ~B
02520       Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
02521       return BinaryOperator::CreateOr(Not, Op0);
02522     }
02523     }
02524   }
02525 
02526   unsigned BitWidth = 0;
02527   if (Ty->isIntOrIntVectorTy())
02528     BitWidth = Ty->getScalarSizeInBits();
02529   else if (DL)  // Pointers require DL info to get their size.
02530     BitWidth = DL->getTypeSizeInBits(Ty->getScalarType());
02531 
02532   bool isSignBit = false;
02533 
02534   // See if we are doing a comparison with a constant.
02535   if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02536     Value *A = nullptr, *B = nullptr;
02537 
02538     // Match the following pattern, which is a common idiom when writing
02539     // overflow-safe integer arithmetic function.  The source performs an
02540     // addition in wider type, and explicitly checks for overflow using
02541     // comparisons against INT_MIN and INT_MAX.  Simplify this by using the
02542     // sadd_with_overflow intrinsic.
02543     //
02544     // TODO: This could probably be generalized to handle other overflow-safe
02545     // operations if we worked out the formulas to compute the appropriate
02546     // magic constants.
02547     //
02548     // sum = a + b
02549     // if (sum+128 >u 255)  ...  -> llvm.sadd.with.overflow.i8
02550     {
02551     ConstantInt *CI2;    // I = icmp ugt (add (add A, B), CI2), CI
02552     if (I.getPredicate() == ICmpInst::ICMP_UGT &&
02553         match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
02554       if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
02555         return Res;
02556     }
02557 
02558     // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
02559     if (I.isEquality() && CI->isZero() &&
02560         match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
02561       // (icmp cond A B) if cond is equality
02562       return new ICmpInst(I.getPredicate(), A, B);
02563     }
02564 
02565     // If we have an icmp le or icmp ge instruction, turn it into the
02566     // appropriate icmp lt or icmp gt instruction.  This allows us to rely on
02567     // them being folded in the code below.  The SimplifyICmpInst code has
02568     // already handled the edge cases for us, so we just assert on them.
02569     switch (I.getPredicate()) {
02570     default: break;
02571     case ICmpInst::ICMP_ULE:
02572       assert(!CI->isMaxValue(false));                 // A <=u MAX -> TRUE
02573       return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
02574                           Builder->getInt(CI->getValue()+1));
02575     case ICmpInst::ICMP_SLE:
02576       assert(!CI->isMaxValue(true));                  // A <=s MAX -> TRUE
02577       return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
02578                           Builder->getInt(CI->getValue()+1));
02579     case ICmpInst::ICMP_UGE:
02580       assert(!CI->isMinValue(false));                 // A >=u MIN -> TRUE
02581       return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
02582                           Builder->getInt(CI->getValue()-1));
02583     case ICmpInst::ICMP_SGE:
02584       assert(!CI->isMinValue(true));                  // A >=s MIN -> TRUE
02585       return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
02586                           Builder->getInt(CI->getValue()-1));
02587     }
02588 
02589     if (I.isEquality()) {
02590       ConstantInt *CI2;
02591       if (match(Op0, m_AShr(m_ConstantInt(CI2), m_Value(A))) ||
02592           match(Op0, m_LShr(m_ConstantInt(CI2), m_Value(A)))) {
02593         // (icmp eq/ne (ashr/lshr const2, A), const1)
02594         return FoldICmpCstShrCst(I, Op0, A, CI, CI2);
02595       }
02596       if (match(Op0, m_Shl(m_ConstantInt(CI2), m_Value(A)))) {
02597         // (icmp eq/ne (shl const2, A), const1)
02598         return FoldICmpCstShlCst(I, Op0, A, CI, CI2);
02599       }
02600     }
02601 
02602     // If this comparison is a normal comparison, it demands all
02603     // bits, if it is a sign bit comparison, it only demands the sign bit.
02604     bool UnusedBit;
02605     isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
02606   }
02607 
02608   // See if we can fold the comparison based on range information we can get
02609   // by checking whether bits are known to be zero or one in the input.
02610   if (BitWidth != 0) {
02611     APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
02612     APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
02613 
02614     if (SimplifyDemandedBits(I.getOperandUse(0),
02615                              DemandedBitsLHSMask(I, BitWidth, isSignBit),
02616                              Op0KnownZero, Op0KnownOne, 0))
02617       return &I;
02618     if (SimplifyDemandedBits(I.getOperandUse(1),
02619                              APInt::getAllOnesValue(BitWidth),
02620                              Op1KnownZero, Op1KnownOne, 0))
02621       return &I;
02622 
02623     // Given the known and unknown bits, compute a range that the LHS could be
02624     // in.  Compute the Min, Max and RHS values based on the known bits. For the
02625     // EQ and NE we use unsigned values.
02626     APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
02627     APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
02628     if (I.isSigned()) {
02629       ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
02630                                              Op0Min, Op0Max);
02631       ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
02632                                              Op1Min, Op1Max);
02633     } else {
02634       ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
02635                                                Op0Min, Op0Max);
02636       ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
02637                                                Op1Min, Op1Max);
02638     }
02639 
02640     // If Min and Max are known to be the same, then SimplifyDemandedBits
02641     // figured out that the LHS is a constant.  Just constant fold this now so
02642     // that code below can assume that Min != Max.
02643     if (!isa<Constant>(Op0) && Op0Min == Op0Max)
02644       return new ICmpInst(I.getPredicate(),
02645                           ConstantInt::get(Op0->getType(), Op0Min), Op1);
02646     if (!isa<Constant>(Op1) && Op1Min == Op1Max)
02647       return new ICmpInst(I.getPredicate(), Op0,
02648                           ConstantInt::get(Op1->getType(), Op1Min));
02649 
02650     // Based on the range information we know about the LHS, see if we can
02651     // simplify this comparison.  For example, (x&4) < 8 is always true.
02652     switch (I.getPredicate()) {
02653     default: llvm_unreachable("Unknown icmp opcode!");
02654     case ICmpInst::ICMP_EQ: {
02655       if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
02656         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02657 
02658       // If all bits are known zero except for one, then we know at most one
02659       // bit is set.   If the comparison is against zero, then this is a check
02660       // to see if *that* bit is set.
02661       APInt Op0KnownZeroInverted = ~Op0KnownZero;
02662       if (~Op1KnownZero == 0) {
02663         // If the LHS is an AND with the same constant, look through it.
02664         Value *LHS = nullptr;
02665         ConstantInt *LHSC = nullptr;
02666         if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
02667             LHSC->getValue() != Op0KnownZeroInverted)
02668           LHS = Op0;
02669 
02670         // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
02671         // then turn "((1 << x)&8) == 0" into "x != 3".
02672         // or turn "((1 << x)&7) == 0" into "x > 2".
02673         Value *X = nullptr;
02674         if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
02675           APInt ValToCheck = Op0KnownZeroInverted;
02676           if (ValToCheck.isPowerOf2()) {
02677             unsigned CmpVal = ValToCheck.countTrailingZeros();
02678             return new ICmpInst(ICmpInst::ICMP_NE, X,
02679                                 ConstantInt::get(X->getType(), CmpVal));
02680           } else if ((++ValToCheck).isPowerOf2()) {
02681             unsigned CmpVal = ValToCheck.countTrailingZeros() - 1;
02682             return new ICmpInst(ICmpInst::ICMP_UGT, X,
02683                                 ConstantInt::get(X->getType(), CmpVal));
02684           }
02685         }
02686 
02687         // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
02688         // then turn "((8 >>u x)&1) == 0" into "x != 3".
02689         const APInt *CI;
02690         if (Op0KnownZeroInverted == 1 &&
02691             match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
02692           return new ICmpInst(ICmpInst::ICMP_NE, X,
02693                               ConstantInt::get(X->getType(),
02694                                                CI->countTrailingZeros()));
02695       }
02696 
02697       break;
02698     }
02699     case ICmpInst::ICMP_NE: {
02700       if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
02701         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02702 
02703       // If all bits are known zero except for one, then we know at most one
02704       // bit is set.   If the comparison is against zero, then this is a check
02705       // to see if *that* bit is set.
02706       APInt Op0KnownZeroInverted = ~Op0KnownZero;
02707       if (~Op1KnownZero == 0) {
02708         // If the LHS is an AND with the same constant, look through it.
02709         Value *LHS = nullptr;
02710         ConstantInt *LHSC = nullptr;
02711         if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
02712             LHSC->getValue() != Op0KnownZeroInverted)
02713           LHS = Op0;
02714 
02715         // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
02716         // then turn "((1 << x)&8) != 0" into "x == 3".
02717         // or turn "((1 << x)&7) != 0" into "x < 3".
02718         Value *X = nullptr;
02719         if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
02720           APInt ValToCheck = Op0KnownZeroInverted;
02721           if (ValToCheck.isPowerOf2()) {
02722             unsigned CmpVal = ValToCheck.countTrailingZeros();
02723             return new ICmpInst(ICmpInst::ICMP_EQ, X,
02724                                 ConstantInt::get(X->getType(), CmpVal));
02725           } else if ((++ValToCheck).isPowerOf2()) {
02726             unsigned CmpVal = ValToCheck.countTrailingZeros();
02727             return new ICmpInst(ICmpInst::ICMP_ULT, X,
02728                                 ConstantInt::get(X->getType(), CmpVal));
02729           }
02730         }
02731 
02732         // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
02733         // then turn "((8 >>u x)&1) != 0" into "x == 3".
02734         const APInt *CI;
02735         if (Op0KnownZeroInverted == 1 &&
02736             match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
02737           return new ICmpInst(ICmpInst::ICMP_EQ, X,
02738                               ConstantInt::get(X->getType(),
02739                                                CI->countTrailingZeros()));
02740       }
02741 
02742       break;
02743     }
02744     case ICmpInst::ICMP_ULT:
02745       if (Op0Max.ult(Op1Min))          // A <u B -> true if max(A) < min(B)
02746         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02747       if (Op0Min.uge(Op1Max))          // A <u B -> false if min(A) >= max(B)
02748         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02749       if (Op1Min == Op0Max)            // A <u B -> A != B if max(A) == min(B)
02750         return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
02751       if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02752         if (Op1Max == Op0Min+1)        // A <u C -> A == C-1 if min(A)+1 == C
02753           return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
02754                               Builder->getInt(CI->getValue()-1));
02755 
02756         // (x <u 2147483648) -> (x >s -1)  -> true if sign bit clear
02757         if (CI->isMinValue(true))
02758           return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
02759                            Constant::getAllOnesValue(Op0->getType()));
02760       }
02761       break;
02762     case ICmpInst::ICMP_UGT:
02763       if (Op0Min.ugt(Op1Max))          // A >u B -> true if min(A) > max(B)
02764         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02765       if (Op0Max.ule(Op1Min))          // A >u B -> false if max(A) <= max(B)
02766         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02767 
02768       if (Op1Max == Op0Min)            // A >u B -> A != B if min(A) == max(B)
02769         return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
02770       if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02771         if (Op1Min == Op0Max-1)        // A >u C -> A == C+1 if max(a)-1 == C
02772           return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
02773                               Builder->getInt(CI->getValue()+1));
02774 
02775         // (x >u 2147483647) -> (x <s 0)  -> true if sign bit set
02776         if (CI->isMaxValue(true))
02777           return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
02778                               Constant::getNullValue(Op0->getType()));
02779       }
02780       break;
02781     case ICmpInst::ICMP_SLT:
02782       if (Op0Max.slt(Op1Min))          // A <s B -> true if max(A) < min(C)
02783         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02784       if (Op0Min.sge(Op1Max))          // A <s B -> false if min(A) >= max(C)
02785         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02786       if (Op1Min == Op0Max)            // A <s B -> A != B if max(A) == min(B)
02787         return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
02788       if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02789         if (Op1Max == Op0Min+1)        // A <s C -> A == C-1 if min(A)+1 == C
02790           return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
02791                               Builder->getInt(CI->getValue()-1));
02792       }
02793       break;
02794     case ICmpInst::ICMP_SGT:
02795       if (Op0Min.sgt(Op1Max))          // A >s B -> true if min(A) > max(B)
02796         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02797       if (Op0Max.sle(Op1Min))          // A >s B -> false if max(A) <= min(B)
02798         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02799 
02800       if (Op1Max == Op0Min)            // A >s B -> A != B if min(A) == max(B)
02801         return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
02802       if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02803         if (Op1Min == Op0Max-1)        // A >s C -> A == C+1 if max(A)-1 == C
02804           return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
02805                               Builder->getInt(CI->getValue()+1));
02806       }
02807       break;
02808     case ICmpInst::ICMP_SGE:
02809       assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
02810       if (Op0Min.sge(Op1Max))          // A >=s B -> true if min(A) >= max(B)
02811         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02812       if (Op0Max.slt(Op1Min))          // A >=s B -> false if max(A) < min(B)
02813         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02814       break;
02815     case ICmpInst::ICMP_SLE:
02816       assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
02817       if (Op0Max.sle(Op1Min))          // A <=s B -> true if max(A) <= min(B)
02818         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02819       if (Op0Min.sgt(Op1Max))          // A <=s B -> false if min(A) > max(B)
02820         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02821       break;
02822     case ICmpInst::ICMP_UGE:
02823       assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
02824       if (Op0Min.uge(Op1Max))          // A >=u B -> true if min(A) >= max(B)
02825         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02826       if (Op0Max.ult(Op1Min))          // A >=u B -> false if max(A) < min(B)
02827         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02828       break;
02829     case ICmpInst::ICMP_ULE:
02830       assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
02831       if (Op0Max.ule(Op1Min))          // A <=u B -> true if max(A) <= min(B)
02832         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02833       if (Op0Min.ugt(Op1Max))          // A <=u B -> false if min(A) > max(B)
02834         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02835       break;
02836     }
02837 
02838     // Turn a signed comparison into an unsigned one if both operands
02839     // are known to have the same sign.
02840     if (I.isSigned() &&
02841         ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
02842          (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
02843       return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
02844   }
02845 
02846   // Test if the ICmpInst instruction is used exclusively by a select as
02847   // part of a minimum or maximum operation. If so, refrain from doing
02848   // any other folding. This helps out other analyses which understand
02849   // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
02850   // and CodeGen. And in this case, at least one of the comparison
02851   // operands has at least one user besides the compare (the select),
02852   // which would often largely negate the benefit of folding anyway.
02853   if (I.hasOneUse())
02854     if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
02855       if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
02856           (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
02857         return nullptr;
02858 
02859   // See if we are doing a comparison between a constant and an instruction that
02860   // can be folded into the comparison.
02861   if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02862     // Since the RHS is a ConstantInt (CI), if the left hand side is an
02863     // instruction, see if that instruction also has constants so that the
02864     // instruction can be folded into the icmp
02865     if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
02866       if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
02867         return Res;
02868   }
02869 
02870   // Handle icmp with constant (but not simple integer constant) RHS
02871   if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
02872     if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
02873       switch (LHSI->getOpcode()) {
02874       case Instruction::GetElementPtr:
02875           // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
02876         if (RHSC->isNullValue() &&
02877             cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
02878           return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
02879                   Constant::getNullValue(LHSI->getOperand(0)->getType()));
02880         break;
02881       case Instruction::PHI:
02882         // Only fold icmp into the PHI if the phi and icmp are in the same
02883         // block.  If in the same block, we're encouraging jump threading.  If
02884         // not, we are just pessimizing the code by making an i1 phi.
02885         if (LHSI->getParent() == I.getParent())
02886           if (Instruction *NV = FoldOpIntoPhi(I))
02887             return NV;
02888         break;
02889       case Instruction::Select: {
02890         // If either operand of the select is a constant, we can fold the
02891         // comparison into the select arms, which will cause one to be
02892         // constant folded and the select turned into a bitwise or.
02893         Value *Op1 = nullptr, *Op2 = nullptr;
02894         if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
02895           Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
02896         if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
02897           Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
02898 
02899         // We only want to perform this transformation if it will not lead to
02900         // additional code. This is true if either both sides of the select
02901         // fold to a constant (in which case the icmp is replaced with a select
02902         // which will usually simplify) or this is the only user of the
02903         // select (in which case we are trading a select+icmp for a simpler
02904         // select+icmp).
02905         if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
02906           if (!Op1)
02907             Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
02908                                       RHSC, I.getName());
02909           if (!Op2)
02910             Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
02911                                       RHSC, I.getName());
02912           return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
02913         }
02914         break;
02915       }
02916       case Instruction::IntToPtr:
02917         // icmp pred inttoptr(X), null -> icmp pred X, 0
02918         if (RHSC->isNullValue() && DL &&
02919             DL->getIntPtrType(RHSC->getType()) ==
02920                LHSI->getOperand(0)->getType())
02921           return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
02922                         Constant::getNullValue(LHSI->getOperand(0)->getType()));
02923         break;
02924 
02925       case Instruction::Load:
02926         // Try to optimize things like "A[i] > 4" to index computations.
02927         if (GetElementPtrInst *GEP =
02928               dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
02929           if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
02930             if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
02931                 !cast<LoadInst>(LHSI)->isVolatile())
02932               if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
02933                 return Res;
02934         }
02935         break;
02936       }
02937   }
02938 
02939   // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
02940   if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
02941     if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
02942       return NI;
02943   if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
02944     if (Instruction *NI = FoldGEPICmp(GEP, Op0,
02945                            ICmpInst::getSwappedPredicate(I.getPredicate()), I))
02946       return NI;
02947 
02948   // Test to see if the operands of the icmp are casted versions of other
02949   // values.  If the ptr->ptr cast can be stripped off both arguments, we do so
02950   // now.
02951   if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
02952     if (Op0->getType()->isPointerTy() &&
02953         (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
02954       // We keep moving the cast from the left operand over to the right
02955       // operand, where it can often be eliminated completely.
02956       Op0 = CI->getOperand(0);
02957 
02958       // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
02959       // so eliminate it as well.
02960       if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
02961         Op1 = CI2->getOperand(0);
02962 
02963       // If Op1 is a constant, we can fold the cast into the constant.
02964       if (Op0->getType() != Op1->getType()) {
02965         if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
02966           Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
02967         } else {
02968           // Otherwise, cast the RHS right before the icmp
02969           Op1 = Builder->CreateBitCast(Op1, Op0->getType());
02970         }
02971       }
02972       return new ICmpInst(I.getPredicate(), Op0, Op1);
02973     }
02974   }
02975 
02976   if (isa<CastInst>(Op0)) {
02977     // Handle the special case of: icmp (cast bool to X), <cst>
02978     // This comes up when you have code like
02979     //   int X = A < B;
02980     //   if (X) ...
02981     // For generality, we handle any zero-extension of any operand comparison
02982     // with a constant or another cast from the same type.
02983     if (isa<Constant>(Op1) || isa<CastInst>(Op1))
02984       if (Instruction *R = visitICmpInstWithCastAndCast(I))
02985         return R;
02986   }
02987 
02988   // Special logic for binary operators.
02989   BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
02990   BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
02991   if (BO0 || BO1) {
02992     CmpInst::Predicate Pred = I.getPredicate();
02993     bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
02994     if (BO0 && isa<OverflowingBinaryOperator>(BO0))
02995       NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
02996         (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
02997         (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
02998     if (BO1 && isa<OverflowingBinaryOperator>(BO1))
02999       NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
03000         (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
03001         (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
03002 
03003     // Analyze the case when either Op0 or Op1 is an add instruction.
03004     // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
03005     Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
03006     if (BO0 && BO0->getOpcode() == Instruction::Add)
03007       A = BO0->getOperand(0), B = BO0->getOperand(1);
03008     if (BO1 && BO1->getOpcode() == Instruction::Add)
03009       C = BO1->getOperand(0), D = BO1->getOperand(1);
03010 
03011     // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
03012     if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
03013       return new ICmpInst(Pred, A == Op1 ? B : A,
03014                           Constant::getNullValue(Op1->getType()));
03015 
03016     // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
03017     if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
03018       return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
03019                           C == Op0 ? D : C);
03020 
03021     // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
03022     if (A && C && (A == C || A == D || B == C || B == D) &&
03023         NoOp0WrapProblem && NoOp1WrapProblem &&
03024         // Try not to increase register pressure.
03025         BO0->hasOneUse() && BO1->hasOneUse()) {
03026       // Determine Y and Z in the form icmp (X+Y), (X+Z).
03027       Value *Y, *Z;
03028       if (A == C) {
03029         // C + B == C + D  ->  B == D
03030         Y = B;
03031         Z = D;
03032       } else if (A == D) {
03033         // D + B == C + D  ->  B == C
03034         Y = B;
03035         Z = C;
03036       } else if (B == C) {
03037         // A + C == C + D  ->  A == D
03038         Y = A;
03039         Z = D;
03040       } else {
03041         assert(B == D);
03042         // A + D == C + D  ->  A == C
03043         Y = A;
03044         Z = C;
03045       }
03046       return new ICmpInst(Pred, Y, Z);
03047     }
03048 
03049     // icmp slt (X + -1), Y -> icmp sle X, Y
03050     if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
03051         match(B, m_AllOnes()))
03052       return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
03053 
03054     // icmp sge (X + -1), Y -> icmp sgt X, Y
03055     if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
03056         match(B, m_AllOnes()))
03057       return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
03058 
03059     // icmp sle (X + 1), Y -> icmp slt X, Y
03060     if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
03061         match(B, m_One()))
03062       return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
03063 
03064     // icmp sgt (X + 1), Y -> icmp sge X, Y
03065     if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
03066         match(B, m_One()))
03067       return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
03068 
03069     // if C1 has greater magnitude than C2:
03070     //  icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
03071     //  s.t. C3 = C1 - C2
03072     //
03073     // if C2 has greater magnitude than C1:
03074     //  icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
03075     //  s.t. C3 = C2 - C1
03076     if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
03077         (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
03078       if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
03079         if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
03080           const APInt &AP1 = C1->getValue();
03081           const APInt &AP2 = C2->getValue();
03082           if (AP1.isNegative() == AP2.isNegative()) {
03083             APInt AP1Abs = C1->getValue().abs();
03084             APInt AP2Abs = C2->getValue().abs();
03085             if (AP1Abs.uge(AP2Abs)) {
03086               ConstantInt *C3 = Builder->getInt(AP1 - AP2);
03087               Value *NewAdd = Builder->CreateNSWAdd(A, C3);
03088               return new ICmpInst(Pred, NewAdd, C);
03089             } else {
03090               ConstantInt *C3 = Builder->getInt(AP2 - AP1);
03091               Value *NewAdd = Builder->CreateNSWAdd(C, C3);
03092               return new ICmpInst(Pred, A, NewAdd);
03093             }
03094           }
03095         }
03096 
03097 
03098     // Analyze the case when either Op0 or Op1 is a sub instruction.
03099     // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
03100     A = nullptr; B = nullptr; C = nullptr; D = nullptr;
03101     if (BO0 && BO0->getOpcode() == Instruction::Sub)
03102       A = BO0->getOperand(0), B = BO0->getOperand(1);
03103     if (BO1 && BO1->getOpcode() == Instruction::Sub)
03104       C = BO1->getOperand(0), D = BO1->getOperand(1);
03105 
03106     // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
03107     if (A == Op1 && NoOp0WrapProblem)
03108       return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
03109 
03110     // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
03111     if (C == Op0 && NoOp1WrapProblem)
03112       return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
03113 
03114     // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
03115     if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
03116         // Try not to increase register pressure.
03117         BO0->hasOneUse() && BO1->hasOneUse())
03118       return new ICmpInst(Pred, A, C);
03119 
03120     // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
03121     if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
03122         // Try not to increase register pressure.
03123         BO0->hasOneUse() && BO1->hasOneUse())
03124       return new ICmpInst(Pred, D, B);
03125 
03126     // icmp (0-X) < cst --> x > -cst
03127     if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
03128       Value *X;
03129       if (match(BO0, m_Neg(m_Value(X))))
03130         if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
03131           if (!RHSC->isMinValue(/*isSigned=*/true))
03132             return new ICmpInst(I.getSwappedPredicate(), X,
03133                                 ConstantExpr::getNeg(RHSC));
03134     }
03135 
03136     BinaryOperator *SRem = nullptr;
03137     // icmp (srem X, Y), Y
03138     if (BO0 && BO0->getOpcode() == Instruction::SRem &&
03139         Op1 == BO0->getOperand(1))
03140       SRem = BO0;
03141     // icmp Y, (srem X, Y)
03142     else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
03143              Op0 == BO1->getOperand(1))
03144       SRem = BO1;
03145     if (SRem) {
03146       // We don't check hasOneUse to avoid increasing register pressure because
03147       // the value we use is the same value this instruction was already using.
03148       switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
03149         default: break;
03150         case ICmpInst::ICMP_EQ:
03151           return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
03152         case ICmpInst::ICMP_NE:
03153           return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
03154         case ICmpInst::ICMP_SGT:
03155         case ICmpInst::ICMP_SGE:
03156           return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
03157                               Constant::getAllOnesValue(SRem->getType()));
03158         case ICmpInst::ICMP_SLT:
03159         case ICmpInst::ICMP_SLE:
03160           return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
03161                               Constant::getNullValue(SRem->getType()));
03162       }
03163     }
03164 
03165     if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
03166         BO0->hasOneUse() && BO1->hasOneUse() &&
03167         BO0->getOperand(1) == BO1->getOperand(1)) {
03168       switch (BO0->getOpcode()) {
03169       default: break;
03170       case Instruction::Add:
03171       case Instruction::Sub:
03172       case Instruction::Xor:
03173         if (I.isEquality())    // a+x icmp eq/ne b+x --> a icmp b
03174           return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
03175                               BO1->getOperand(0));
03176         // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
03177         if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
03178           if (CI->getValue().isSignBit()) {
03179             ICmpInst::Predicate Pred = I.isSigned()
03180                                            ? I.getUnsignedPredicate()
03181                                            : I.getSignedPredicate();
03182             return new ICmpInst(Pred, BO0->getOperand(0),
03183                                 BO1->getOperand(0));
03184           }
03185 
03186           if (CI->isMaxValue(true)) {
03187             ICmpInst::Predicate Pred = I.isSigned()
03188                                            ? I.getUnsignedPredicate()
03189                                            : I.getSignedPredicate();
03190             Pred = I.getSwappedPredicate(Pred);
03191             return new ICmpInst(Pred, BO0->getOperand(0),
03192                                 BO1->getOperand(0));
03193           }
03194         }
03195         break;
03196       case Instruction::Mul:
03197         if (!I.isEquality())
03198           break;
03199 
03200         if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
03201           // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
03202           // Mask = -1 >> count-trailing-zeros(Cst).
03203           if (!CI->isZero() && !CI->isOne()) {
03204             const APInt &AP = CI->getValue();
03205             ConstantInt *Mask = ConstantInt::get(I.getContext(),
03206                                     APInt::getLowBitsSet(AP.getBitWidth(),
03207                                                          AP.getBitWidth() -
03208                                                     AP.countTrailingZeros()));
03209             Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
03210             Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
03211             return new ICmpInst(I.getPredicate(), And1, And2);
03212           }
03213         }
03214         break;
03215       case Instruction::UDiv:
03216       case Instruction::LShr:
03217         if (I.isSigned())
03218           break;
03219         // fall-through
03220       case Instruction::SDiv:
03221       case Instruction::AShr:
03222         if (!BO0->isExact() || !BO1->isExact())
03223           break;
03224         return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
03225                             BO1->getOperand(0));
03226       case Instruction::Shl: {
03227         bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
03228         bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
03229         if (!NUW && !NSW)
03230           break;
03231         if (!NSW && I.isSigned())
03232           break;
03233         return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
03234                             BO1->getOperand(0));
03235       }
03236       }
03237     }
03238   }
03239 
03240   { Value *A, *B;
03241     // Transform (A & ~B) == 0 --> (A & B) != 0
03242     // and       (A & ~B) != 0 --> (A & B) == 0
03243     // if A is a power of 2.
03244     if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
03245         match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A, false,
03246                                                        0, AT, &I, DT) &&
03247                                 I.isEquality())
03248       return new ICmpInst(I.getInversePredicate(),
03249                           Builder->CreateAnd(A, B),
03250                           Op1);
03251 
03252     // ~x < ~y --> y < x
03253     // ~x < cst --> ~cst < x
03254     if (match(Op0, m_Not(m_Value(A)))) {
03255       if (match(Op1, m_Not(m_Value(B))))
03256         return new ICmpInst(I.getPredicate(), B, A);
03257       if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
03258         return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
03259     }
03260 
03261     // (a+b) <u a  --> llvm.uadd.with.overflow.
03262     // (a+b) <u b  --> llvm.uadd.with.overflow.
03263     if (I.getPredicate() == ICmpInst::ICMP_ULT &&
03264         match(Op0, m_Add(m_Value(A), m_Value(B))) &&
03265         (Op1 == A || Op1 == B))
03266       if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
03267         return R;
03268 
03269     // a >u (a+b)  --> llvm.uadd.with.overflow.
03270     // b >u (a+b)  --> llvm.uadd.with.overflow.
03271     if (I.getPredicate() == ICmpInst::ICMP_UGT &&
03272         match(Op1, m_Add(m_Value(A), m_Value(B))) &&
03273         (Op0 == A || Op0 == B))
03274       if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
03275         return R;
03276 
03277     // (zext a) * (zext b)  --> llvm.umul.with.overflow.
03278     if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
03279       if (Instruction *R = ProcessUMulZExtIdiom(I, Op0, Op1, *this))
03280         return R;
03281     }
03282     if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
03283       if (Instruction *R = ProcessUMulZExtIdiom(I, Op1, Op0, *this))
03284         return R;
03285     }
03286   }
03287 
03288   if (I.isEquality()) {
03289     Value *A, *B, *C, *D;
03290 
03291     if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
03292       if (A == Op1 || B == Op1) {    // (A^B) == A  ->  B == 0
03293         Value *OtherVal = A == Op1 ? B : A;
03294         return new ICmpInst(I.getPredicate(), OtherVal,
03295                             Constant::getNullValue(A->getType()));
03296       }
03297 
03298       if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
03299         // A^c1 == C^c2 --> A == C^(c1^c2)
03300         ConstantInt *C1, *C2;
03301         if (match(B, m_ConstantInt(C1)) &&
03302             match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
03303           Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
03304           Value *Xor = Builder->CreateXor(C, NC);
03305           return new ICmpInst(I.getPredicate(), A, Xor);
03306         }
03307 
03308         // A^B == A^D -> B == D
03309         if (A == C) return new ICmpInst(I.getPredicate(), B, D);
03310         if (A == D) return new ICmpInst(I.getPredicate(), B, C);
03311         if (B == C) return new ICmpInst(I.getPredicate(), A, D);
03312         if (B == D) return new ICmpInst(I.getPredicate(), A, C);
03313       }
03314     }
03315 
03316     if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
03317         (A == Op0 || B == Op0)) {
03318       // A == (A^B)  ->  B == 0
03319       Value *OtherVal = A == Op0 ? B : A;
03320       return new ICmpInst(I.getPredicate(), OtherVal,
03321                           Constant::getNullValue(A->getType()));
03322     }
03323 
03324     // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
03325     if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
03326         match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
03327       Value *X = nullptr, *Y = nullptr, *Z = nullptr;
03328 
03329       if (A == C) {
03330         X = B; Y = D; Z = A;
03331       } else if (A == D) {
03332         X = B; Y = C; Z = A;
03333       } else if (B == C) {
03334         X = A; Y = D; Z = B;
03335       } else if (B == D) {
03336         X = A; Y = C; Z = B;
03337       }
03338 
03339       if (X) {   // Build (X^Y) & Z
03340         Op1 = Builder->CreateXor(X, Y);
03341         Op1 = Builder->CreateAnd(Op1, Z);
03342         I.setOperand(0, Op1);
03343         I.setOperand(1, Constant::getNullValue(Op1->getType()));
03344         return &I;
03345       }
03346     }
03347 
03348     // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
03349     // and       (B & (1<<X)-1) == (zext A) --> A == (trunc B)
03350     ConstantInt *Cst1;
03351     if ((Op0->hasOneUse() &&
03352          match(Op0, m_ZExt(m_Value(A))) &&
03353          match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
03354         (Op1->hasOneUse() &&
03355          match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
03356          match(Op1, m_ZExt(m_Value(A))))) {
03357       APInt Pow2 = Cst1->getValue() + 1;
03358       if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
03359           Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
03360         return new ICmpInst(I.getPredicate(), A,
03361                             Builder->CreateTrunc(B, A->getType()));
03362     }
03363 
03364     // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
03365     // For lshr and ashr pairs.
03366     if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
03367          match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
03368         (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
03369          match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
03370       unsigned TypeBits = Cst1->getBitWidth();
03371       unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
03372       if (ShAmt < TypeBits && ShAmt != 0) {
03373         ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE
03374                                        ? ICmpInst::ICMP_UGE
03375                                        : ICmpInst::ICMP_ULT;
03376         Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
03377         APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
03378         return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal));
03379       }
03380     }
03381 
03382     // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
03383     // "icmp (and X, mask), cst"
03384     uint64_t ShAmt = 0;
03385     if (Op0->hasOneUse() &&
03386         match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
03387                                            m_ConstantInt(ShAmt))))) &&
03388         match(Op1, m_ConstantInt(Cst1)) &&
03389         // Only do this when A has multiple uses.  This is most important to do
03390         // when it exposes other optimizations.
03391         !A->hasOneUse()) {
03392       unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
03393 
03394       if (ShAmt < ASize) {
03395         APInt MaskV =
03396           APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
03397         MaskV <<= ShAmt;
03398 
03399         APInt CmpV = Cst1->getValue().zext(ASize);
03400         CmpV <<= ShAmt;
03401 
03402         Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
03403         return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
03404       }
03405     }
03406   }
03407 
03408   {
03409     Value *X; ConstantInt *Cst;
03410     // icmp X+Cst, X
03411     if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
03412       return FoldICmpAddOpCst(I, X, Cst, I.getPredicate());
03413 
03414     // icmp X, X+Cst
03415     if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
03416       return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
03417   }
03418   return Changed ? &I : nullptr;
03419 }
03420 
03421 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
03422 ///
03423 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
03424                                                 Instruction *LHSI,
03425                                                 Constant *RHSC) {
03426   if (!isa<ConstantFP>(RHSC)) return nullptr;
03427   const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
03428 
03429   // Get the width of the mantissa.  We don't want to hack on conversions that
03430   // might lose information from the integer, e.g. "i64 -> float"
03431   int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
03432   if (MantissaWidth == -1) return nullptr;  // Unknown.
03433 
03434   // Check to see that the input is converted from an integer type that is small
03435   // enough that preserves all bits.  TODO: check here for "known" sign bits.
03436   // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
03437   unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
03438 
03439   // If this is a uitofp instruction, we need an extra bit to hold the sign.
03440   bool LHSUnsigned = isa<UIToFPInst>(LHSI);
03441   if (LHSUnsigned)
03442     ++InputSize;
03443 
03444   // If the conversion would lose info, don't hack on this.
03445   if ((int)InputSize > MantissaWidth)
03446     return nullptr;
03447 
03448   // Otherwise, we can potentially simplify the comparison.  We know that it
03449   // will always come through as an integer value and we know the constant is
03450   // not a NAN (it would have been previously simplified).
03451   assert(!RHS.isNaN() && "NaN comparison not already folded!");
03452 
03453   ICmpInst::Predicate Pred;
03454   switch (I.getPredicate()) {
03455   default: llvm_unreachable("Unexpected predicate!");
03456   case FCmpInst::FCMP_UEQ:
03457   case FCmpInst::FCMP_OEQ:
03458     Pred = ICmpInst::ICMP_EQ;
03459     break;
03460   case FCmpInst::FCMP_UGT:
03461   case FCmpInst::FCMP_OGT:
03462     Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
03463     break;
03464   case FCmpInst::FCMP_UGE:
03465   case FCmpInst::FCMP_OGE:
03466     Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
03467     break;
03468   case FCmpInst::FCMP_ULT:
03469   case FCmpInst::FCMP_OLT:
03470     Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
03471     break;
03472   case FCmpInst::FCMP_ULE:
03473   case FCmpInst::FCMP_OLE:
03474     Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
03475     break;
03476   case FCmpInst::FCMP_UNE:
03477   case FCmpInst::FCMP_ONE:
03478     Pred = ICmpInst::ICMP_NE;
03479     break;
03480   case FCmpInst::FCMP_ORD:
03481     return ReplaceInstUsesWith(I, Builder->getTrue());
03482   case FCmpInst::FCMP_UNO:
03483     return ReplaceInstUsesWith(I, Builder->getFalse());
03484   }
03485 
03486   IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
03487 
03488   // Now we know that the APFloat is a normal number, zero or inf.
03489 
03490   // See if the FP constant is too large for the integer.  For example,
03491   // comparing an i8 to 300.0.
03492   unsigned IntWidth = IntTy->getScalarSizeInBits();
03493 
03494   if (!LHSUnsigned) {
03495     // If the RHS value is > SignedMax, fold the comparison.  This handles +INF
03496     // and large values.
03497     APFloat SMax(RHS.getSemantics());
03498     SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
03499                           APFloat::rmNearestTiesToEven);
03500     if (SMax.compare(RHS) == APFloat::cmpLessThan) {  // smax < 13123.0
03501       if (Pred == ICmpInst::ICMP_NE  || Pred == ICmpInst::ICMP_SLT ||
03502           Pred == ICmpInst::ICMP_SLE)
03503         return ReplaceInstUsesWith(I, Builder->getTrue());
03504       return ReplaceInstUsesWith(I, Builder->getFalse());
03505     }
03506   } else {
03507     // If the RHS value is > UnsignedMax, fold the comparison. This handles
03508     // +INF and large values.
03509     APFloat UMax(RHS.getSemantics());
03510     UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
03511                           APFloat::rmNearestTiesToEven);
03512     if (UMax.compare(RHS) == APFloat::cmpLessThan) {  // umax < 13123.0
03513       if (Pred == ICmpInst::ICMP_NE  || Pred == ICmpInst::ICMP_ULT ||
03514           Pred == ICmpInst::ICMP_ULE)
03515         return ReplaceInstUsesWith(I, Builder->getTrue());
03516       return ReplaceInstUsesWith(I, Builder->getFalse());
03517     }
03518   }
03519 
03520   if (!LHSUnsigned) {
03521     // See if the RHS value is < SignedMin.
03522     APFloat SMin(RHS.getSemantics());
03523     SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
03524                           APFloat::rmNearestTiesToEven);
03525     if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
03526       if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
03527           Pred == ICmpInst::ICMP_SGE)
03528         return ReplaceInstUsesWith(I, Builder->getTrue());
03529       return ReplaceInstUsesWith(I, Builder->getFalse());
03530     }
03531   } else {
03532     // See if the RHS value is < UnsignedMin.
03533     APFloat SMin(RHS.getSemantics());
03534     SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
03535                           APFloat::rmNearestTiesToEven);
03536     if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
03537       if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
03538           Pred == ICmpInst::ICMP_UGE)
03539         return ReplaceInstUsesWith(I, Builder->getTrue());
03540       return ReplaceInstUsesWith(I, Builder->getFalse());
03541     }
03542   }
03543 
03544   // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
03545   // [0, UMAX], but it may still be fractional.  See if it is fractional by
03546   // casting the FP value to the integer value and back, checking for equality.
03547   // Don't do this for zero, because -0.0 is not fractional.
03548   Constant *RHSInt = LHSUnsigned
03549     ? ConstantExpr::getFPToUI(RHSC, IntTy)
03550     : ConstantExpr::getFPToSI(RHSC, IntTy);
03551   if (!RHS.isZero()) {
03552     bool Equal = LHSUnsigned
03553       ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
03554       : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
03555     if (!Equal) {
03556       // If we had a comparison against a fractional value, we have to adjust
03557       // the compare predicate and sometimes the value.  RHSC is rounded towards
03558       // zero at this point.
03559       switch (Pred) {
03560       default: llvm_unreachable("Unexpected integer comparison!");
03561       case ICmpInst::ICMP_NE:  // (float)int != 4.4   --> true
03562         return ReplaceInstUsesWith(I, Builder->getTrue());
03563       case ICmpInst::ICMP_EQ:  // (float)int == 4.4   --> false
03564         return ReplaceInstUsesWith(I, Builder->getFalse());
03565       case ICmpInst::ICMP_ULE:
03566         // (float)int <= 4.4   --> int <= 4
03567         // (float)int <= -4.4  --> false
03568         if (RHS.isNegative())
03569           return ReplaceInstUsesWith(I, Builder->getFalse());
03570         break;
03571       case ICmpInst::ICMP_SLE:
03572         // (float)int <= 4.4   --> int <= 4
03573         // (float)int <= -4.4  --> int < -4
03574         if (RHS.isNegative())
03575           Pred = ICmpInst::ICMP_SLT;
03576         break;
03577       case ICmpInst::ICMP_ULT:
03578         // (float)int < -4.4   --> false
03579         // (float)int < 4.4    --> int <= 4
03580         if (RHS.isNegative())
03581           return ReplaceInstUsesWith(I, Builder->getFalse());
03582         Pred = ICmpInst::ICMP_ULE;
03583         break;
03584       case ICmpInst::ICMP_SLT:
03585         // (float)int < -4.4   --> int < -4
03586         // (float)int < 4.4    --> int <= 4
03587         if (!RHS.isNegative())
03588           Pred = ICmpInst::ICMP_SLE;
03589         break;
03590       case ICmpInst::ICMP_UGT:
03591         // (float)int > 4.4    --> int > 4
03592         // (float)int > -4.4   --> true
03593         if (RHS.isNegative())
03594           return ReplaceInstUsesWith(I, Builder->getTrue());
03595         break;
03596       case ICmpInst::ICMP_SGT:
03597         // (float)int > 4.4    --> int > 4
03598         // (float)int > -4.4   --> int >= -4
03599         if (RHS.isNegative())
03600           Pred = ICmpInst::ICMP_SGE;
03601         break;
03602       case ICmpInst::ICMP_UGE:
03603         // (float)int >= -4.4   --> true
03604         // (float)int >= 4.4    --> int > 4
03605         if (RHS.isNegative())
03606           return ReplaceInstUsesWith(I, Builder->getTrue());
03607         Pred = ICmpInst::ICMP_UGT;
03608         break;
03609       case ICmpInst::ICMP_SGE:
03610         // (float)int >= -4.4   --> int >= -4
03611         // (float)int >= 4.4    --> int > 4
03612         if (!RHS.isNegative())
03613           Pred = ICmpInst::ICMP_SGT;
03614         break;
03615       }
03616     }
03617   }
03618 
03619   // Lower this FP comparison into an appropriate integer version of the
03620   // comparison.
03621   return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
03622 }
03623 
03624 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
03625   bool Changed = false;
03626 
03627   /// Orders the operands of the compare so that they are listed from most
03628   /// complex to least complex.  This puts constants before unary operators,
03629   /// before binary operators.
03630   if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
03631     I.swapOperands();
03632     Changed = true;
03633   }
03634 
03635   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
03636 
03637   if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, DL, TLI, DT, AT))
03638     return ReplaceInstUsesWith(I, V);
03639 
03640   // Simplify 'fcmp pred X, X'
03641   if (Op0 == Op1) {
03642     switch (I.getPredicate()) {
03643     default: llvm_unreachable("Unknown predicate!");
03644     case FCmpInst::FCMP_UNO:    // True if unordered: isnan(X) | isnan(Y)
03645     case FCmpInst::FCMP_ULT:    // True if unordered or less than
03646     case FCmpInst::FCMP_UGT:    // True if unordered or greater than
03647     case FCmpInst::FCMP_UNE:    // True if unordered or not equal
03648       // Canonicalize these to be 'fcmp uno %X, 0.0'.
03649       I.setPredicate(FCmpInst::FCMP_UNO);
03650       I.setOperand(1, Constant::getNullValue(Op0->getType()));
03651       return &I;
03652 
03653     case FCmpInst::FCMP_ORD:    // True if ordered (no nans)
03654     case FCmpInst::FCMP_OEQ:    // True if ordered and equal
03655     case FCmpInst::FCMP_OGE:    // True if ordered and greater than or equal
03656     case FCmpInst::FCMP_OLE:    // True if ordered and less than or equal
03657       // Canonicalize these to be 'fcmp ord %X, 0.0'.
03658       I.setPredicate(FCmpInst::FCMP_ORD);
03659       I.setOperand(1, Constant::getNullValue(Op0->getType()));
03660       return &I;
03661     }
03662   }
03663 
03664   // Handle fcmp with constant RHS
03665   if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
03666     if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
03667       switch (LHSI->getOpcode()) {
03668       case Instruction::FPExt: {
03669         // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
03670         FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
03671         ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
03672         if (!RHSF)
03673           break;
03674 
03675         const fltSemantics *Sem;
03676         // FIXME: This shouldn't be here.
03677         if (LHSExt->getSrcTy()->isHalfTy())
03678           Sem = &APFloat::IEEEhalf;
03679         else if (LHSExt->getSrcTy()->isFloatTy())
03680           Sem = &APFloat::IEEEsingle;
03681         else if (LHSExt->getSrcTy()->isDoubleTy())
03682           Sem = &APFloat::IEEEdouble;
03683         else if (LHSExt->getSrcTy()->isFP128Ty())
03684           Sem = &APFloat::IEEEquad;
03685         else if (LHSExt->getSrcTy()->isX86_FP80Ty())
03686           Sem = &APFloat::x87DoubleExtended;
03687         else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
03688           Sem = &APFloat::PPCDoubleDouble;
03689         else
03690           break;
03691 
03692         bool Lossy;
03693         APFloat F = RHSF->getValueAPF();
03694         F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
03695 
03696         // Avoid lossy conversions and denormals. Zero is a special case
03697         // that's OK to convert.
03698         APFloat Fabs = F;
03699         Fabs.clearSign();
03700         if (!Lossy &&
03701             ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
03702                  APFloat::cmpLessThan) || Fabs.isZero()))
03703 
03704           return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
03705                               ConstantFP::get(RHSC->getContext(), F));
03706         break;
03707       }
03708       case Instruction::PHI:
03709         // Only fold fcmp into the PHI if the phi and fcmp are in the same
03710         // block.  If in the same block, we're encouraging jump threading.  If
03711         // not, we are just pessimizing the code by making an i1 phi.
03712         if (LHSI->getParent() == I.getParent())
03713           if (Instruction *NV = FoldOpIntoPhi(I))
03714             return NV;
03715         break;
03716       case Instruction::SIToFP:
03717       case Instruction::UIToFP:
03718         if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
03719           return NV;
03720         break;
03721       case Instruction::FSub: {
03722         // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
03723         Value *Op;
03724         if (match(LHSI, m_FNeg(m_Value(Op))))
03725           return new FCmpInst(I.getSwappedPredicate(), Op,
03726                               ConstantExpr::getFNeg(RHSC));
03727         break;
03728       }
03729       case Instruction::Load:
03730         if (GetElementPtrInst *GEP =
03731             dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
03732           if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
03733             if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
03734                 !cast<LoadInst>(LHSI)->isVolatile())
03735               if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
03736                 return Res;
03737         }
03738         break;
03739       case Instruction::Call: {
03740         CallInst *CI = cast<CallInst>(LHSI);
03741         LibFunc::Func Func;
03742         // Various optimization for fabs compared with zero.
03743         if (RHSC->isNullValue() && CI->getCalledFunction() &&
03744             TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
03745             TLI->has(Func)) {
03746           if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
03747               Func == LibFunc::fabsl) {
03748             switch (I.getPredicate()) {
03749             default: break;
03750             // fabs(x) < 0 --> false
03751             case FCmpInst::FCMP_OLT:
03752               return ReplaceInstUsesWith(I, Builder->getFalse());
03753             // fabs(x) > 0 --> x != 0
03754             case FCmpInst::FCMP_OGT:
03755               return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
03756                                   RHSC);
03757             // fabs(x) <= 0 --> x == 0
03758             case FCmpInst::FCMP_OLE:
03759               return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
03760                                   RHSC);
03761             // fabs(x) >= 0 --> !isnan(x)
03762             case FCmpInst::FCMP_OGE:
03763               return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
03764                                   RHSC);
03765             // fabs(x) == 0 --> x == 0
03766             // fabs(x) != 0 --> x != 0
03767             case FCmpInst::FCMP_OEQ:
03768             case FCmpInst::FCMP_UEQ:
03769             case FCmpInst::FCMP_ONE:
03770             case FCmpInst::FCMP_UNE:
03771               return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
03772                                   RHSC);
03773             }
03774           }
03775         }
03776       }
03777       }
03778   }
03779 
03780   // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
03781   Value *X, *Y;
03782   if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
03783     return new FCmpInst(I.getSwappedPredicate(), X, Y);
03784 
03785   // fcmp (fpext x), (fpext y) -> fcmp x, y
03786   if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
03787     if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
03788       if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
03789         return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
03790                             RHSExt->getOperand(0));
03791 
03792   return Changed ? &I : nullptr;
03793 }