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   if (!AP1) {
01056     if (!AP2) {
01057       // Both Constants are 0.
01058       return getConstant(true);
01059     }
01060 
01061     if (cast<BinaryOperator>(Op)->isExact())
01062       return getConstant(false);
01063 
01064     if (AP2.isNegative()) {
01065       // MSB is set, so a lshr with a large enough 'A' would be undefined.
01066       return getConstant(false);
01067     }
01068 
01069     // 'A' must be large enough to shift out the highest set bit.
01070     return getICmp(I.ICMP_UGT, A,
01071                    ConstantInt::get(A->getType(), AP2.logBase2()));
01072   }
01073 
01074   if (!AP2) {
01075     // Shifting 0 by any value gives 0.
01076     return getConstant(false);
01077   }
01078 
01079   bool IsAShr = isa<AShrOperator>(Op);
01080   if (AP1 == AP2) {
01081     if (AP1.isAllOnesValue() && IsAShr) {
01082       // Arithmatic shift of -1 is always -1.
01083       return getConstant(true);
01084     }
01085     return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
01086   }
01087 
01088   bool IsNegative = false;
01089   if (IsAShr) {
01090     if (AP1.isNegative() != AP2.isNegative()) {
01091       // Arithmetic shift will never change the sign.
01092       return getConstant(false);
01093     }
01094     // Both the constants are negative, take their positive to calculate log.
01095     if (AP1.isNegative()) {
01096       if (AP1.slt(AP2))
01097         // Right-shifting won't increase the magnitude.
01098         return getConstant(false);
01099       IsNegative = true;
01100     }
01101   }
01102 
01103   if (!IsNegative && AP1.ugt(AP2))
01104     // Right-shifting will not increase the value.
01105     return getConstant(false);
01106 
01107   // Get the distance between the highest bit that's set.
01108   int Shift;
01109   if (IsNegative)
01110     // Get the ones' complement of AP2 and AP1 when computing the distance.
01111     Shift = (~AP2).logBase2() - (~AP1).logBase2();
01112   else
01113     Shift = AP2.logBase2() - AP1.logBase2();
01114 
01115   if (IsAShr ? AP1 == AP2.ashr(Shift) : AP1 == AP2.lshr(Shift))
01116     return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
01117 
01118   // Shifting const2 will never be equal to const1.
01119   return getConstant(false);
01120 }
01121 
01122 /// FoldICmpCstShlCst - Handle "(icmp eq/ne (shl const2, A), const1)" ->
01123 /// (icmp eq/ne A, TrailingZeros(const1) - TrailingZeros(const2)).
01124 Instruction *InstCombiner::FoldICmpCstShlCst(ICmpInst &I, Value *Op, Value *A,
01125                                              ConstantInt *CI1,
01126                                              ConstantInt *CI2) {
01127   assert(I.isEquality() && "Cannot fold icmp gt/lt");
01128 
01129   auto getConstant = [&I, this](bool IsTrue) {
01130     if (I.getPredicate() == I.ICMP_NE)
01131       IsTrue = !IsTrue;
01132     return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
01133   };
01134 
01135   auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
01136     if (I.getPredicate() == I.ICMP_NE)
01137       Pred = CmpInst::getInversePredicate(Pred);
01138     return new ICmpInst(Pred, LHS, RHS);
01139   };
01140 
01141   APInt AP1 = CI1->getValue();
01142   APInt AP2 = CI2->getValue();
01143 
01144   assert(AP2 != 0 && "Handled in InstSimplify");
01145 
01146   unsigned AP2TrailingZeros = AP2.countTrailingZeros();
01147 
01148   if (!AP1 && AP2TrailingZeros != 0)
01149     return getICmp(I.ICMP_UGE, A,
01150                    ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
01151 
01152   if (AP1 == AP2)
01153     return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
01154 
01155   // Get the distance between the lowest bits that are set.
01156   int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
01157 
01158   if (Shift > 0 && AP2.shl(Shift) == AP1)
01159     return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
01160 
01161   // Shifting const2 will never be equal to const1.
01162   return getConstant(false);
01163 }
01164 
01165 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
01166 ///
01167 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
01168                                                           Instruction *LHSI,
01169                                                           ConstantInt *RHS) {
01170   const APInt &RHSV = RHS->getValue();
01171 
01172   switch (LHSI->getOpcode()) {
01173   case Instruction::Trunc:
01174     if (ICI.isEquality() && LHSI->hasOneUse()) {
01175       // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
01176       // of the high bits truncated out of x are known.
01177       unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
01178              SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
01179       APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
01180       computeKnownBits(LHSI->getOperand(0), KnownZero, KnownOne, 0, &ICI);
01181 
01182       // If all the high bits are known, we can do this xform.
01183       if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
01184         // Pull in the high bits from known-ones set.
01185         APInt NewRHS = RHS->getValue().zext(SrcBits);
01186         NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
01187         return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
01188                             Builder->getInt(NewRHS));
01189       }
01190     }
01191     break;
01192 
01193   case Instruction::Xor:         // (icmp pred (xor X, XorCst), CI)
01194     if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
01195       // If this is a comparison that tests the signbit (X < 0) or (x > -1),
01196       // fold the xor.
01197       if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
01198           (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
01199         Value *CompareVal = LHSI->getOperand(0);
01200 
01201         // If the sign bit of the XorCst is not set, there is no change to
01202         // the operation, just stop using the Xor.
01203         if (!XorCst->isNegative()) {
01204           ICI.setOperand(0, CompareVal);
01205           Worklist.Add(LHSI);
01206           return &ICI;
01207         }
01208 
01209         // Was the old condition true if the operand is positive?
01210         bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
01211 
01212         // If so, the new one isn't.
01213         isTrueIfPositive ^= true;
01214 
01215         if (isTrueIfPositive)
01216           return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
01217                               SubOne(RHS));
01218         else
01219           return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
01220                               AddOne(RHS));
01221       }
01222 
01223       if (LHSI->hasOneUse()) {
01224         // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
01225         if (!ICI.isEquality() && XorCst->getValue().isSignBit()) {
01226           const APInt &SignBit = XorCst->getValue();
01227           ICmpInst::Predicate Pred = ICI.isSigned()
01228                                          ? ICI.getUnsignedPredicate()
01229                                          : ICI.getSignedPredicate();
01230           return new ICmpInst(Pred, LHSI->getOperand(0),
01231                               Builder->getInt(RHSV ^ SignBit));
01232         }
01233 
01234         // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
01235         if (!ICI.isEquality() && XorCst->isMaxValue(true)) {
01236           const APInt &NotSignBit = XorCst->getValue();
01237           ICmpInst::Predicate Pred = ICI.isSigned()
01238                                          ? ICI.getUnsignedPredicate()
01239                                          : ICI.getSignedPredicate();
01240           Pred = ICI.getSwappedPredicate(Pred);
01241           return new ICmpInst(Pred, LHSI->getOperand(0),
01242                               Builder->getInt(RHSV ^ NotSignBit));
01243         }
01244       }
01245 
01246       // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
01247       //   iff -C is a power of 2
01248       if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
01249           XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
01250         return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst);
01251 
01252       // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
01253       //   iff -C is a power of 2
01254       if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
01255           XorCst->getValue() == -RHSV && RHSV.isPowerOf2())
01256         return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst);
01257     }
01258     break;
01259   case Instruction::And:         // (icmp pred (and X, AndCst), RHS)
01260     if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
01261         LHSI->getOperand(0)->hasOneUse()) {
01262       ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1));
01263 
01264       // If the LHS is an AND of a truncating cast, we can widen the
01265       // and/compare to be the input width without changing the value
01266       // produced, eliminating a cast.
01267       if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
01268         // We can do this transformation if either the AND constant does not
01269         // have its sign bit set or if it is an equality comparison.
01270         // Extending a relational comparison when we're checking the sign
01271         // bit would not work.
01272         if (ICI.isEquality() ||
01273             (!AndCst->isNegative() && RHSV.isNonNegative())) {
01274           Value *NewAnd =
01275             Builder->CreateAnd(Cast->getOperand(0),
01276                                ConstantExpr::getZExt(AndCst, Cast->getSrcTy()));
01277           NewAnd->takeName(LHSI);
01278           return new ICmpInst(ICI.getPredicate(), NewAnd,
01279                               ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
01280         }
01281       }
01282 
01283       // If the LHS is an AND of a zext, and we have an equality compare, we can
01284       // shrink the and/compare to the smaller type, eliminating the cast.
01285       if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
01286         IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
01287         // Make sure we don't compare the upper bits, SimplifyDemandedBits
01288         // should fold the icmp to true/false in that case.
01289         if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
01290           Value *NewAnd =
01291             Builder->CreateAnd(Cast->getOperand(0),
01292                                ConstantExpr::getTrunc(AndCst, Ty));
01293           NewAnd->takeName(LHSI);
01294           return new ICmpInst(ICI.getPredicate(), NewAnd,
01295                               ConstantExpr::getTrunc(RHS, Ty));
01296         }
01297       }
01298 
01299       // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
01300       // could exist), turn it into (X & (C2 << C1)) != (C3 << C1).  This
01301       // happens a LOT in code produced by the C front-end, for bitfield
01302       // access.
01303       BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
01304       if (Shift && !Shift->isShift())
01305         Shift = nullptr;
01306 
01307       ConstantInt *ShAmt;
01308       ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : nullptr;
01309 
01310       // This seemingly simple opportunity to fold away a shift turns out to
01311       // be rather complicated. See PR17827
01312       // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details.
01313       if (ShAmt) {
01314         bool CanFold = false;
01315         unsigned ShiftOpcode = Shift->getOpcode();
01316         if (ShiftOpcode == Instruction::AShr) {
01317           // There may be some constraints that make this possible,
01318           // but nothing simple has been discovered yet.
01319           CanFold = false;
01320         } else if (ShiftOpcode == Instruction::Shl) {
01321           // For a left shift, we can fold if the comparison is not signed.
01322           // We can also fold a signed comparison if the mask value and
01323           // comparison value are not negative. These constraints may not be
01324           // obvious, but we can prove that they are correct using an SMT
01325           // solver.
01326           if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative()))
01327             CanFold = true;
01328         } else if (ShiftOpcode == Instruction::LShr) {
01329           // For a logical right shift, we can fold if the comparison is not
01330           // signed. We can also fold a signed comparison if the shifted mask
01331           // value and the shifted comparison value are not negative.
01332           // These constraints may not be obvious, but we can prove that they
01333           // are correct using an SMT solver.
01334           if (!ICI.isSigned())
01335             CanFold = true;
01336           else {
01337             ConstantInt *ShiftedAndCst =
01338               cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt));
01339             ConstantInt *ShiftedRHSCst =
01340               cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt));
01341             
01342             if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative())
01343               CanFold = true;
01344           }
01345         }
01346 
01347         if (CanFold) {
01348           Constant *NewCst;
01349           if (ShiftOpcode == Instruction::Shl)
01350             NewCst = ConstantExpr::getLShr(RHS, ShAmt);
01351           else
01352             NewCst = ConstantExpr::getShl(RHS, ShAmt);
01353 
01354           // Check to see if we are shifting out any of the bits being
01355           // compared.
01356           if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) {
01357             // If we shifted bits out, the fold is not going to work out.
01358             // As a special case, check to see if this means that the
01359             // result is always true or false now.
01360             if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
01361               return ReplaceInstUsesWith(ICI, Builder->getFalse());
01362             if (ICI.getPredicate() == ICmpInst::ICMP_NE)
01363               return ReplaceInstUsesWith(ICI, Builder->getTrue());
01364           } else {
01365             ICI.setOperand(1, NewCst);
01366             Constant *NewAndCst;
01367             if (ShiftOpcode == Instruction::Shl)
01368               NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt);
01369             else
01370               NewAndCst = ConstantExpr::getShl(AndCst, ShAmt);
01371             LHSI->setOperand(1, NewAndCst);
01372             LHSI->setOperand(0, Shift->getOperand(0));
01373             Worklist.Add(Shift); // Shift is dead.
01374             return &ICI;
01375           }
01376         }
01377       }
01378 
01379       // Turn ((X >> Y) & C) == 0  into  (X & (C << Y)) == 0.  The later is
01380       // preferable because it allows the C<<Y expression to be hoisted out
01381       // of a loop if Y is invariant and X is not.
01382       if (Shift && Shift->hasOneUse() && RHSV == 0 &&
01383           ICI.isEquality() && !Shift->isArithmeticShift() &&
01384           !isa<Constant>(Shift->getOperand(0))) {
01385         // Compute C << Y.
01386         Value *NS;
01387         if (Shift->getOpcode() == Instruction::LShr) {
01388           NS = Builder->CreateShl(AndCst, Shift->getOperand(1));
01389         } else {
01390           // Insert a logical shift.
01391           NS = Builder->CreateLShr(AndCst, Shift->getOperand(1));
01392         }
01393 
01394         // Compute X & (C << Y).
01395         Value *NewAnd =
01396           Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
01397 
01398         ICI.setOperand(0, NewAnd);
01399         return &ICI;
01400       }
01401 
01402       // (icmp pred (and (or (lshr X, Y), X), 1), 0) -->
01403       //    (icmp pred (and X, (or (shl 1, Y), 1), 0))
01404       //
01405       // iff pred isn't signed
01406       {
01407         Value *X, *Y, *LShr;
01408         if (!ICI.isSigned() && RHSV == 0) {
01409           if (match(LHSI->getOperand(1), m_One())) {
01410             Constant *One = cast<Constant>(LHSI->getOperand(1));
01411             Value *Or = LHSI->getOperand(0);
01412             if (match(Or, m_Or(m_Value(LShr), m_Value(X))) &&
01413                 match(LShr, m_LShr(m_Specific(X), m_Value(Y)))) {
01414               unsigned UsesRemoved = 0;
01415               if (LHSI->hasOneUse())
01416                 ++UsesRemoved;
01417               if (Or->hasOneUse())
01418                 ++UsesRemoved;
01419               if (LShr->hasOneUse())
01420                 ++UsesRemoved;
01421               Value *NewOr = nullptr;
01422               // Compute X & ((1 << Y) | 1)
01423               if (auto *C = dyn_cast<Constant>(Y)) {
01424                 if (UsesRemoved >= 1)
01425                   NewOr =
01426                       ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
01427               } else {
01428                 if (UsesRemoved >= 3)
01429                   NewOr = Builder->CreateOr(Builder->CreateShl(One, Y,
01430                                                                LShr->getName(),
01431                                                                /*HasNUW=*/true),
01432                                             One, Or->getName());
01433               }
01434               if (NewOr) {
01435                 Value *NewAnd = Builder->CreateAnd(X, NewOr, LHSI->getName());
01436                 ICI.setOperand(0, NewAnd);
01437                 return &ICI;
01438               }
01439             }
01440           }
01441         }
01442       }
01443 
01444       // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any
01445       // bit set in (X & AndCst) will produce a result greater than RHSV.
01446       if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
01447         unsigned NTZ = AndCst->getValue().countTrailingZeros();
01448         if ((NTZ < AndCst->getBitWidth()) &&
01449             APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV))
01450           return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
01451                               Constant::getNullValue(RHS->getType()));
01452       }
01453     }
01454 
01455     // Try to optimize things like "A[i]&42 == 0" to index computations.
01456     if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
01457       if (GetElementPtrInst *GEP =
01458           dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
01459         if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
01460           if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
01461               !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
01462             ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
01463             if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
01464               return Res;
01465           }
01466     }
01467 
01468     // X & -C == -C -> X >  u ~C
01469     // X & -C != -C -> X <= u ~C
01470     //   iff C is a power of 2
01471     if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
01472       return new ICmpInst(
01473           ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
01474                                                   : ICmpInst::ICMP_ULE,
01475           LHSI->getOperand(0), SubOne(RHS));
01476     break;
01477 
01478   case Instruction::Or: {
01479     if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
01480       break;
01481     Value *P, *Q;
01482     if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
01483       // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
01484       // -> and (icmp eq P, null), (icmp eq Q, null).
01485       Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
01486                                         Constant::getNullValue(P->getType()));
01487       Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
01488                                         Constant::getNullValue(Q->getType()));
01489       Instruction *Op;
01490       if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
01491         Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
01492       else
01493         Op = BinaryOperator::CreateOr(ICIP, ICIQ);
01494       return Op;
01495     }
01496     break;
01497   }
01498 
01499   case Instruction::Mul: {       // (icmp pred (mul X, Val), CI)
01500     ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
01501     if (!Val) break;
01502 
01503     // If this is a signed comparison to 0 and the mul is sign preserving,
01504     // use the mul LHS operand instead.
01505     ICmpInst::Predicate pred = ICI.getPredicate();
01506     if (isSignTest(pred, RHS) && !Val->isZero() &&
01507         cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
01508       return new ICmpInst(Val->isNegative() ?
01509                           ICmpInst::getSwappedPredicate(pred) : pred,
01510                           LHSI->getOperand(0),
01511                           Constant::getNullValue(RHS->getType()));
01512 
01513     break;
01514   }
01515 
01516   case Instruction::Shl: {       // (icmp pred (shl X, ShAmt), CI)
01517     uint32_t TypeBits = RHSV.getBitWidth();
01518     ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
01519     if (!ShAmt) {
01520       Value *X;
01521       // (1 << X) pred P2 -> X pred Log2(P2)
01522       if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
01523         bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
01524         ICmpInst::Predicate Pred = ICI.getPredicate();
01525         if (ICI.isUnsigned()) {
01526           if (!RHSVIsPowerOf2) {
01527             // (1 << X) <  30 -> X <= 4
01528             // (1 << X) <= 30 -> X <= 4
01529             // (1 << X) >= 30 -> X >  4
01530             // (1 << X) >  30 -> X >  4
01531             if (Pred == ICmpInst::ICMP_ULT)
01532               Pred = ICmpInst::ICMP_ULE;
01533             else if (Pred == ICmpInst::ICMP_UGE)
01534               Pred = ICmpInst::ICMP_UGT;
01535           }
01536           unsigned RHSLog2 = RHSV.logBase2();
01537 
01538           // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
01539           // (1 << X) <  2147483648 -> X <  31 -> X != 31
01540           if (RHSLog2 == TypeBits-1) {
01541             if (Pred == ICmpInst::ICMP_UGE)
01542               Pred = ICmpInst::ICMP_EQ;
01543             else if (Pred == ICmpInst::ICMP_ULT)
01544               Pred = ICmpInst::ICMP_NE;
01545           }
01546 
01547           return new ICmpInst(Pred, X,
01548                               ConstantInt::get(RHS->getType(), RHSLog2));
01549         } else if (ICI.isSigned()) {
01550           if (RHSV.isAllOnesValue()) {
01551             // (1 << X) <= -1 -> X == 31
01552             if (Pred == ICmpInst::ICMP_SLE)
01553               return new ICmpInst(ICmpInst::ICMP_EQ, X,
01554                                   ConstantInt::get(RHS->getType(), TypeBits-1));
01555 
01556             // (1 << X) >  -1 -> X != 31
01557             if (Pred == ICmpInst::ICMP_SGT)
01558               return new ICmpInst(ICmpInst::ICMP_NE, X,
01559                                   ConstantInt::get(RHS->getType(), TypeBits-1));
01560           } else if (!RHSV) {
01561             // (1 << X) <  0 -> X == 31
01562             // (1 << X) <= 0 -> X == 31
01563             if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
01564               return new ICmpInst(ICmpInst::ICMP_EQ, X,
01565                                   ConstantInt::get(RHS->getType(), TypeBits-1));
01566 
01567             // (1 << X) >= 0 -> X != 31
01568             // (1 << X) >  0 -> X != 31
01569             if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
01570               return new ICmpInst(ICmpInst::ICMP_NE, X,
01571                                   ConstantInt::get(RHS->getType(), TypeBits-1));
01572           }
01573         } else if (ICI.isEquality()) {
01574           if (RHSVIsPowerOf2)
01575             return new ICmpInst(
01576                 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
01577         }
01578       }
01579       break;
01580     }
01581 
01582     // Check that the shift amount is in range.  If not, don't perform
01583     // undefined shifts.  When the shift is visited it will be
01584     // simplified.
01585     if (ShAmt->uge(TypeBits))
01586       break;
01587 
01588     if (ICI.isEquality()) {
01589       // If we are comparing against bits always shifted out, the
01590       // comparison cannot succeed.
01591       Constant *Comp =
01592         ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
01593                                                                  ShAmt);
01594       if (Comp != RHS) {// Comparing against a bit that we know is zero.
01595         bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
01596         Constant *Cst = Builder->getInt1(IsICMP_NE);
01597         return ReplaceInstUsesWith(ICI, Cst);
01598       }
01599 
01600       // If the shift is NUW, then it is just shifting out zeros, no need for an
01601       // AND.
01602       if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
01603         return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
01604                             ConstantExpr::getLShr(RHS, ShAmt));
01605 
01606       // If the shift is NSW and we compare to 0, then it is just shifting out
01607       // sign bits, no need for an AND either.
01608       if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
01609         return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
01610                             ConstantExpr::getLShr(RHS, ShAmt));
01611 
01612       if (LHSI->hasOneUse()) {
01613         // Otherwise strength reduce the shift into an and.
01614         uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
01615         Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
01616                                                           TypeBits - ShAmtVal));
01617 
01618         Value *And =
01619           Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
01620         return new ICmpInst(ICI.getPredicate(), And,
01621                             ConstantExpr::getLShr(RHS, ShAmt));
01622       }
01623     }
01624 
01625     // If this is a signed comparison to 0 and the shift is sign preserving,
01626     // use the shift LHS operand instead.
01627     ICmpInst::Predicate pred = ICI.getPredicate();
01628     if (isSignTest(pred, RHS) &&
01629         cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
01630       return new ICmpInst(pred,
01631                           LHSI->getOperand(0),
01632                           Constant::getNullValue(RHS->getType()));
01633 
01634     // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
01635     bool TrueIfSigned = false;
01636     if (LHSI->hasOneUse() &&
01637         isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
01638       // (X << 31) <s 0  --> (X&1) != 0
01639       Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
01640                                         APInt::getOneBitSet(TypeBits,
01641                                             TypeBits-ShAmt->getZExtValue()-1));
01642       Value *And =
01643         Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
01644       return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
01645                           And, Constant::getNullValue(And->getType()));
01646     }
01647 
01648     // Transform (icmp pred iM (shl iM %v, N), CI)
01649     // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
01650     // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
01651     // This enables to get rid of the shift in favor of a trunc which can be
01652     // free on the target. It has the additional benefit of comparing to a
01653     // smaller constant, which will be target friendly.
01654     unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
01655     if (LHSI->hasOneUse() &&
01656         Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
01657       Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
01658       Constant *NCI = ConstantExpr::getTrunc(
01659                         ConstantExpr::getAShr(RHS,
01660                           ConstantInt::get(RHS->getType(), Amt)),
01661                         NTy);
01662       return new ICmpInst(ICI.getPredicate(),
01663                           Builder->CreateTrunc(LHSI->getOperand(0), NTy),
01664                           NCI);
01665     }
01666 
01667     break;
01668   }
01669 
01670   case Instruction::LShr:         // (icmp pred (shr X, ShAmt), CI)
01671   case Instruction::AShr: {
01672     // Handle equality comparisons of shift-by-constant.
01673     BinaryOperator *BO = cast<BinaryOperator>(LHSI);
01674     if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
01675       if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
01676         return Res;
01677     }
01678 
01679     // Handle exact shr's.
01680     if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
01681       if (RHSV.isMinValue())
01682         return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
01683     }
01684     break;
01685   }
01686 
01687   case Instruction::SDiv:
01688   case Instruction::UDiv:
01689     // Fold: icmp pred ([us]div X, C1), C2 -> range test
01690     // Fold this div into the comparison, producing a range check.
01691     // Determine, based on the divide type, what the range is being
01692     // checked.  If there is an overflow on the low or high side, remember
01693     // it, otherwise compute the range [low, hi) bounding the new value.
01694     // See: InsertRangeTest above for the kinds of replacements possible.
01695     if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
01696       if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
01697                                           DivRHS))
01698         return R;
01699     break;
01700 
01701   case Instruction::Sub: {
01702     ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
01703     if (!LHSC) break;
01704     const APInt &LHSV = LHSC->getValue();
01705 
01706     // C1-X <u C2 -> (X|(C2-1)) == C1
01707     //   iff C1 & (C2-1) == C2-1
01708     //       C2 is a power of 2
01709     if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
01710         RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
01711       return new ICmpInst(ICmpInst::ICMP_EQ,
01712                           Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
01713                           LHSC);
01714 
01715     // C1-X >u C2 -> (X|C2) != C1
01716     //   iff C1 & C2 == C2
01717     //       C2+1 is a power of 2
01718     if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
01719         (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
01720       return new ICmpInst(ICmpInst::ICMP_NE,
01721                           Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
01722     break;
01723   }
01724 
01725   case Instruction::Add:
01726     // Fold: icmp pred (add X, C1), C2
01727     if (!ICI.isEquality()) {
01728       ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
01729       if (!LHSC) break;
01730       const APInt &LHSV = LHSC->getValue();
01731 
01732       ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
01733                             .subtract(LHSV);
01734 
01735       if (ICI.isSigned()) {
01736         if (CR.getLower().isSignBit()) {
01737           return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
01738                               Builder->getInt(CR.getUpper()));
01739         } else if (CR.getUpper().isSignBit()) {
01740           return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
01741                               Builder->getInt(CR.getLower()));
01742         }
01743       } else {
01744         if (CR.getLower().isMinValue()) {
01745           return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
01746                               Builder->getInt(CR.getUpper()));
01747         } else if (CR.getUpper().isMinValue()) {
01748           return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
01749                               Builder->getInt(CR.getLower()));
01750         }
01751       }
01752 
01753       // X-C1 <u C2 -> (X & -C2) == C1
01754       //   iff C1 & (C2-1) == 0
01755       //       C2 is a power of 2
01756       if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
01757           RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
01758         return new ICmpInst(ICmpInst::ICMP_EQ,
01759                             Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
01760                             ConstantExpr::getNeg(LHSC));
01761 
01762       // X-C1 >u C2 -> (X & ~C2) != C1
01763       //   iff C1 & C2 == 0
01764       //       C2+1 is a power of 2
01765       if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
01766           (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
01767         return new ICmpInst(ICmpInst::ICMP_NE,
01768                             Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
01769                             ConstantExpr::getNeg(LHSC));
01770     }
01771     break;
01772   }
01773 
01774   // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
01775   if (ICI.isEquality()) {
01776     bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
01777 
01778     // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
01779     // the second operand is a constant, simplify a bit.
01780     if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
01781       switch (BO->getOpcode()) {
01782       case Instruction::SRem:
01783         // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
01784         if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
01785           const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
01786           if (V.sgt(1) && V.isPowerOf2()) {
01787             Value *NewRem =
01788               Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
01789                                   BO->getName());
01790             return new ICmpInst(ICI.getPredicate(), NewRem,
01791                                 Constant::getNullValue(BO->getType()));
01792           }
01793         }
01794         break;
01795       case Instruction::Add:
01796         // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
01797         if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
01798           if (BO->hasOneUse())
01799             return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01800                                 ConstantExpr::getSub(RHS, BOp1C));
01801         } else if (RHSV == 0) {
01802           // Replace ((add A, B) != 0) with (A != -B) if A or B is
01803           // efficiently invertible, or if the add has just this one use.
01804           Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
01805 
01806           if (Value *NegVal = dyn_castNegVal(BOp1))
01807             return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
01808           if (Value *NegVal = dyn_castNegVal(BOp0))
01809             return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
01810           if (BO->hasOneUse()) {
01811             Value *Neg = Builder->CreateNeg(BOp1);
01812             Neg->takeName(BO);
01813             return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
01814           }
01815         }
01816         break;
01817       case Instruction::Xor:
01818         // For the xor case, we can xor two constants together, eliminating
01819         // the explicit xor.
01820         if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
01821           return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01822                               ConstantExpr::getXor(RHS, BOC));
01823         } else if (RHSV == 0) {
01824           // Replace ((xor A, B) != 0) with (A != B)
01825           return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01826                               BO->getOperand(1));
01827         }
01828         break;
01829       case Instruction::Sub:
01830         // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
01831         if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
01832           if (BO->hasOneUse())
01833             return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
01834                                 ConstantExpr::getSub(BOp0C, RHS));
01835         } else if (RHSV == 0) {
01836           // Replace ((sub A, B) != 0) with (A != B)
01837           return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01838                               BO->getOperand(1));
01839         }
01840         break;
01841       case Instruction::Or:
01842         // If bits are being or'd in that are not present in the constant we
01843         // are comparing against, then the comparison could never succeed!
01844         if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
01845           Constant *NotCI = ConstantExpr::getNot(RHS);
01846           if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
01847             return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
01848         }
01849         break;
01850 
01851       case Instruction::And:
01852         if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
01853           // If bits are being compared against that are and'd out, then the
01854           // comparison can never succeed!
01855           if ((RHSV & ~BOC->getValue()) != 0)
01856             return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
01857 
01858           // If we have ((X & C) == C), turn it into ((X & C) != 0).
01859           if (RHS == BOC && RHSV.isPowerOf2())
01860             return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
01861                                 ICmpInst::ICMP_NE, LHSI,
01862                                 Constant::getNullValue(RHS->getType()));
01863 
01864           // Don't perform the following transforms if the AND has multiple uses
01865           if (!BO->hasOneUse())
01866             break;
01867 
01868           // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
01869           if (BOC->getValue().isSignBit()) {
01870             Value *X = BO->getOperand(0);
01871             Constant *Zero = Constant::getNullValue(X->getType());
01872             ICmpInst::Predicate pred = isICMP_NE ?
01873               ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
01874             return new ICmpInst(pred, X, Zero);
01875           }
01876 
01877           // ((X & ~7) == 0) --> X < 8
01878           if (RHSV == 0 && isHighOnes(BOC)) {
01879             Value *X = BO->getOperand(0);
01880             Constant *NegX = ConstantExpr::getNeg(BOC);
01881             ICmpInst::Predicate pred = isICMP_NE ?
01882               ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
01883             return new ICmpInst(pred, X, NegX);
01884           }
01885         }
01886         break;
01887       case Instruction::Mul:
01888         if (RHSV == 0 && BO->hasNoSignedWrap()) {
01889           if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
01890             // The trivial case (mul X, 0) is handled by InstSimplify
01891             // General case : (mul X, C) != 0 iff X != 0
01892             //                (mul X, C) == 0 iff X == 0
01893             if (!BOC->isZero())
01894               return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01895                                   Constant::getNullValue(RHS->getType()));
01896           }
01897         }
01898         break;
01899       default: break;
01900       }
01901     } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
01902       // Handle icmp {eq|ne} <intrinsic>, intcst.
01903       switch (II->getIntrinsicID()) {
01904       case Intrinsic::bswap:
01905         Worklist.Add(II);
01906         ICI.setOperand(0, II->getArgOperand(0));
01907         ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
01908         return &ICI;
01909       case Intrinsic::ctlz:
01910       case Intrinsic::cttz:
01911         // ctz(A) == bitwidth(a)  ->  A == 0 and likewise for !=
01912         if (RHSV == RHS->getType()->getBitWidth()) {
01913           Worklist.Add(II);
01914           ICI.setOperand(0, II->getArgOperand(0));
01915           ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
01916           return &ICI;
01917         }
01918         break;
01919       case Intrinsic::ctpop:
01920         // popcount(A) == 0  ->  A == 0 and likewise for !=
01921         if (RHS->isZero()) {
01922           Worklist.Add(II);
01923           ICI.setOperand(0, II->getArgOperand(0));
01924           ICI.setOperand(1, RHS);
01925           return &ICI;
01926         }
01927         break;
01928       default:
01929         break;
01930       }
01931     }
01932   }
01933   return nullptr;
01934 }
01935 
01936 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
01937 /// We only handle extending casts so far.
01938 ///
01939 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
01940   const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
01941   Value *LHSCIOp        = LHSCI->getOperand(0);
01942   Type *SrcTy     = LHSCIOp->getType();
01943   Type *DestTy    = LHSCI->getType();
01944   Value *RHSCIOp;
01945 
01946   // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
01947   // integer type is the same size as the pointer type.
01948   if (DL && LHSCI->getOpcode() == Instruction::PtrToInt &&
01949       DL->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
01950     Value *RHSOp = nullptr;
01951     if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
01952       RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
01953     } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
01954       RHSOp = RHSC->getOperand(0);
01955       // If the pointer types don't match, insert a bitcast.
01956       if (LHSCIOp->getType() != RHSOp->getType())
01957         RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
01958     }
01959 
01960     if (RHSOp)
01961       return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
01962   }
01963 
01964   // The code below only handles extension cast instructions, so far.
01965   // Enforce this.
01966   if (LHSCI->getOpcode() != Instruction::ZExt &&
01967       LHSCI->getOpcode() != Instruction::SExt)
01968     return nullptr;
01969 
01970   bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
01971   bool isSignedCmp = ICI.isSigned();
01972 
01973   if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
01974     // Not an extension from the same type?
01975     RHSCIOp = CI->getOperand(0);
01976     if (RHSCIOp->getType() != LHSCIOp->getType())
01977       return nullptr;
01978 
01979     // If the signedness of the two casts doesn't agree (i.e. one is a sext
01980     // and the other is a zext), then we can't handle this.
01981     if (CI->getOpcode() != LHSCI->getOpcode())
01982       return nullptr;
01983 
01984     // Deal with equality cases early.
01985     if (ICI.isEquality())
01986       return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
01987 
01988     // A signed comparison of sign extended values simplifies into a
01989     // signed comparison.
01990     if (isSignedCmp && isSignedExt)
01991       return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
01992 
01993     // The other three cases all fold into an unsigned comparison.
01994     return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
01995   }
01996 
01997   // If we aren't dealing with a constant on the RHS, exit early
01998   ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
01999   if (!CI)
02000     return nullptr;
02001 
02002   // Compute the constant that would happen if we truncated to SrcTy then
02003   // reextended to DestTy.
02004   Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
02005   Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
02006                                                 Res1, DestTy);
02007 
02008   // If the re-extended constant didn't change...
02009   if (Res2 == CI) {
02010     // Deal with equality cases early.
02011     if (ICI.isEquality())
02012       return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
02013 
02014     // A signed comparison of sign extended values simplifies into a
02015     // signed comparison.
02016     if (isSignedExt && isSignedCmp)
02017       return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
02018 
02019     // The other three cases all fold into an unsigned comparison.
02020     return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
02021   }
02022 
02023   // The re-extended constant changed so the constant cannot be represented
02024   // in the shorter type. Consequently, we cannot emit a simple comparison.
02025   // All the cases that fold to true or false will have already been handled
02026   // by SimplifyICmpInst, so only deal with the tricky case.
02027 
02028   if (isSignedCmp || !isSignedExt)
02029     return nullptr;
02030 
02031   // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
02032   // should have been folded away previously and not enter in here.
02033 
02034   // We're performing an unsigned comp with a sign extended value.
02035   // This is true if the input is >= 0. [aka >s -1]
02036   Constant *NegOne = Constant::getAllOnesValue(SrcTy);
02037   Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
02038 
02039   // Finally, return the value computed.
02040   if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
02041     return ReplaceInstUsesWith(ICI, Result);
02042 
02043   assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
02044   return BinaryOperator::CreateNot(Result);
02045 }
02046 
02047 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
02048 ///   I = icmp ugt (add (add A, B), CI2), CI1
02049 /// If this is of the form:
02050 ///   sum = a + b
02051 ///   if (sum+128 >u 255)
02052 /// Then replace it with llvm.sadd.with.overflow.i8.
02053 ///
02054 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
02055                                           ConstantInt *CI2, ConstantInt *CI1,
02056                                           InstCombiner &IC) {
02057   // The transformation we're trying to do here is to transform this into an
02058   // llvm.sadd.with.overflow.  To do this, we have to replace the original add
02059   // with a narrower add, and discard the add-with-constant that is part of the
02060   // range check (if we can't eliminate it, this isn't profitable).
02061 
02062   // In order to eliminate the add-with-constant, the compare can be its only
02063   // use.
02064   Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
02065   if (!AddWithCst->hasOneUse()) return nullptr;
02066 
02067   // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
02068   if (!CI2->getValue().isPowerOf2()) return nullptr;
02069   unsigned NewWidth = CI2->getValue().countTrailingZeros();
02070   if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return nullptr;
02071 
02072   // The width of the new add formed is 1 more than the bias.
02073   ++NewWidth;
02074 
02075   // Check to see that CI1 is an all-ones value with NewWidth bits.
02076   if (CI1->getBitWidth() == NewWidth ||
02077       CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
02078     return nullptr;
02079 
02080   // This is only really a signed overflow check if the inputs have been
02081   // sign-extended; check for that condition. For example, if CI2 is 2^31 and
02082   // the operands of the add are 64 bits wide, we need at least 33 sign bits.
02083   unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
02084   if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
02085       IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
02086     return nullptr;
02087 
02088   // In order to replace the original add with a narrower
02089   // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
02090   // and truncates that discard the high bits of the add.  Verify that this is
02091   // the case.
02092   Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
02093   for (User *U : OrigAdd->users()) {
02094     if (U == AddWithCst) continue;
02095 
02096     // Only accept truncates for now.  We would really like a nice recursive
02097     // predicate like SimplifyDemandedBits, but which goes downwards the use-def
02098     // chain to see which bits of a value are actually demanded.  If the
02099     // original add had another add which was then immediately truncated, we
02100     // could still do the transformation.
02101     TruncInst *TI = dyn_cast<TruncInst>(U);
02102     if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
02103       return nullptr;
02104   }
02105 
02106   // If the pattern matches, truncate the inputs to the narrower type and
02107   // use the sadd_with_overflow intrinsic to efficiently compute both the
02108   // result and the overflow bit.
02109   Module *M = I.getParent()->getParent()->getParent();
02110 
02111   Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
02112   Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
02113                                        NewType);
02114 
02115   InstCombiner::BuilderTy *Builder = IC.Builder;
02116 
02117   // Put the new code above the original add, in case there are any uses of the
02118   // add between the add and the compare.
02119   Builder->SetInsertPoint(OrigAdd);
02120 
02121   Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
02122   Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
02123   CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
02124   Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
02125   Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
02126 
02127   // The inner add was the result of the narrow add, zero extended to the
02128   // wider type.  Replace it with the result computed by the intrinsic.
02129   IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
02130 
02131   // The original icmp gets replaced with the overflow value.
02132   return ExtractValueInst::Create(Call, 1, "sadd.overflow");
02133 }
02134 
02135 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
02136                                      InstCombiner &IC) {
02137   // Don't bother doing this transformation for pointers, don't do it for
02138   // vectors.
02139   if (!isa<IntegerType>(OrigAddV->getType())) return nullptr;
02140 
02141   // If the add is a constant expr, then we don't bother transforming it.
02142   Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
02143   if (!OrigAdd) return nullptr;
02144 
02145   Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
02146 
02147   // Put the new code above the original add, in case there are any uses of the
02148   // add between the add and the compare.
02149   InstCombiner::BuilderTy *Builder = IC.Builder;
02150   Builder->SetInsertPoint(OrigAdd);
02151 
02152   Module *M = I.getParent()->getParent()->getParent();
02153   Type *Ty = LHS->getType();
02154   Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
02155   CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
02156   Value *Add = Builder->CreateExtractValue(Call, 0);
02157 
02158   IC.ReplaceInstUsesWith(*OrigAdd, Add);
02159 
02160   // The original icmp gets replaced with the overflow value.
02161   return ExtractValueInst::Create(Call, 1, "uadd.overflow");
02162 }
02163 
02164 /// \brief Recognize and process idiom involving test for multiplication
02165 /// overflow.
02166 ///
02167 /// The caller has matched a pattern of the form:
02168 ///   I = cmp u (mul(zext A, zext B), V
02169 /// The function checks if this is a test for overflow and if so replaces
02170 /// multiplication with call to 'mul.with.overflow' intrinsic.
02171 ///
02172 /// \param I Compare instruction.
02173 /// \param MulVal Result of 'mult' instruction.  It is one of the arguments of
02174 ///               the compare instruction.  Must be of integer type.
02175 /// \param OtherVal The other argument of compare instruction.
02176 /// \returns Instruction which must replace the compare instruction, NULL if no
02177 ///          replacement required.
02178 static Instruction *ProcessUMulZExtIdiom(ICmpInst &I, Value *MulVal,
02179                                          Value *OtherVal, InstCombiner &IC) {
02180   // Don't bother doing this transformation for pointers, don't do it for
02181   // vectors.
02182   if (!isa<IntegerType>(MulVal->getType()))
02183     return nullptr;
02184 
02185   assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
02186   assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
02187   Instruction *MulInstr = cast<Instruction>(MulVal);
02188   assert(MulInstr->getOpcode() == Instruction::Mul);
02189 
02190   Instruction *LHS = cast<Instruction>(MulInstr->getOperand(0)),
02191               *RHS = cast<Instruction>(MulInstr->getOperand(1));
02192   assert(LHS->getOpcode() == Instruction::ZExt);
02193   assert(RHS->getOpcode() == Instruction::ZExt);
02194   Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
02195 
02196   // Calculate type and width of the result produced by mul.with.overflow.
02197   Type *TyA = A->getType(), *TyB = B->getType();
02198   unsigned WidthA = TyA->getPrimitiveSizeInBits(),
02199            WidthB = TyB->getPrimitiveSizeInBits();
02200   unsigned MulWidth;
02201   Type *MulType;
02202   if (WidthB > WidthA) {
02203     MulWidth = WidthB;
02204     MulType = TyB;
02205   } else {
02206     MulWidth = WidthA;
02207     MulType = TyA;
02208   }
02209 
02210   // In order to replace the original mul with a narrower mul.with.overflow,
02211   // all uses must ignore upper bits of the product.  The number of used low
02212   // bits must be not greater than the width of mul.with.overflow.
02213   if (MulVal->hasNUsesOrMore(2))
02214     for (User *U : MulVal->users()) {
02215       if (U == &I)
02216         continue;
02217       if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
02218         // Check if truncation ignores bits above MulWidth.
02219         unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
02220         if (TruncWidth > MulWidth)
02221           return nullptr;
02222       } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
02223         // Check if AND ignores bits above MulWidth.
02224         if (BO->getOpcode() != Instruction::And)
02225           return nullptr;
02226         if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
02227           const APInt &CVal = CI->getValue();
02228           if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
02229             return nullptr;
02230         }
02231       } else {
02232         // Other uses prohibit this transformation.
02233         return nullptr;
02234       }
02235     }
02236 
02237   // Recognize patterns
02238   switch (I.getPredicate()) {
02239   case ICmpInst::ICMP_EQ:
02240   case ICmpInst::ICMP_NE:
02241     // Recognize pattern:
02242     //   mulval = mul(zext A, zext B)
02243     //   cmp eq/neq mulval, zext trunc mulval
02244     if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
02245       if (Zext->hasOneUse()) {
02246         Value *ZextArg = Zext->getOperand(0);
02247         if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
02248           if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
02249             break; //Recognized
02250       }
02251 
02252     // Recognize pattern:
02253     //   mulval = mul(zext A, zext B)
02254     //   cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
02255     ConstantInt *CI;
02256     Value *ValToMask;
02257     if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
02258       if (ValToMask != MulVal)
02259         return nullptr;
02260       const APInt &CVal = CI->getValue() + 1;
02261       if (CVal.isPowerOf2()) {
02262         unsigned MaskWidth = CVal.logBase2();
02263         if (MaskWidth == MulWidth)
02264           break; // Recognized
02265       }
02266     }
02267     return nullptr;
02268 
02269   case ICmpInst::ICMP_UGT:
02270     // Recognize pattern:
02271     //   mulval = mul(zext A, zext B)
02272     //   cmp ugt mulval, max
02273     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
02274       APInt MaxVal = APInt::getMaxValue(MulWidth);
02275       MaxVal = MaxVal.zext(CI->getBitWidth());
02276       if (MaxVal.eq(CI->getValue()))
02277         break; // Recognized
02278     }
02279     return nullptr;
02280 
02281   case ICmpInst::ICMP_UGE:
02282     // Recognize pattern:
02283     //   mulval = mul(zext A, zext B)
02284     //   cmp uge mulval, max+1
02285     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
02286       APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
02287       if (MaxVal.eq(CI->getValue()))
02288         break; // Recognized
02289     }
02290     return nullptr;
02291 
02292   case ICmpInst::ICMP_ULE:
02293     // Recognize pattern:
02294     //   mulval = mul(zext A, zext B)
02295     //   cmp ule mulval, max
02296     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
02297       APInt MaxVal = APInt::getMaxValue(MulWidth);
02298       MaxVal = MaxVal.zext(CI->getBitWidth());
02299       if (MaxVal.eq(CI->getValue()))
02300         break; // Recognized
02301     }
02302     return nullptr;
02303 
02304   case ICmpInst::ICMP_ULT:
02305     // Recognize pattern:
02306     //   mulval = mul(zext A, zext B)
02307     //   cmp ule mulval, max + 1
02308     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
02309       APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
02310       if (MaxVal.eq(CI->getValue()))
02311         break; // Recognized
02312     }
02313     return nullptr;
02314 
02315   default:
02316     return nullptr;
02317   }
02318 
02319   InstCombiner::BuilderTy *Builder = IC.Builder;
02320   Builder->SetInsertPoint(MulInstr);
02321   Module *M = I.getParent()->getParent()->getParent();
02322 
02323   // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
02324   Value *MulA = A, *MulB = B;
02325   if (WidthA < MulWidth)
02326     MulA = Builder->CreateZExt(A, MulType);
02327   if (WidthB < MulWidth)
02328     MulB = Builder->CreateZExt(B, MulType);
02329   Value *F =
02330       Intrinsic::getDeclaration(M, Intrinsic::umul_with_overflow, MulType);
02331   CallInst *Call = Builder->CreateCall2(F, MulA, MulB, "umul");
02332   IC.Worklist.Add(MulInstr);
02333 
02334   // If there are uses of mul result other than the comparison, we know that
02335   // they are truncation or binary AND. Change them to use result of
02336   // mul.with.overflow and adjust properly mask/size.
02337   if (MulVal->hasNUsesOrMore(2)) {
02338     Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value");
02339     for (User *U : MulVal->users()) {
02340       if (U == &I || U == OtherVal)
02341         continue;
02342       if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
02343         if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
02344           IC.ReplaceInstUsesWith(*TI, Mul);
02345         else
02346           TI->setOperand(0, Mul);
02347       } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
02348         assert(BO->getOpcode() == Instruction::And);
02349         // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
02350         ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
02351         APInt ShortMask = CI->getValue().trunc(MulWidth);
02352         Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask);
02353         Instruction *Zext =
02354             cast<Instruction>(Builder->CreateZExt(ShortAnd, BO->getType()));
02355         IC.Worklist.Add(Zext);
02356         IC.ReplaceInstUsesWith(*BO, Zext);
02357       } else {
02358         llvm_unreachable("Unexpected Binary operation");
02359       }
02360       IC.Worklist.Add(cast<Instruction>(U));
02361     }
02362   }
02363   if (isa<Instruction>(OtherVal))
02364     IC.Worklist.Add(cast<Instruction>(OtherVal));
02365 
02366   // The original icmp gets replaced with the overflow value, maybe inverted
02367   // depending on predicate.
02368   bool Inverse = false;
02369   switch (I.getPredicate()) {
02370   case ICmpInst::ICMP_NE:
02371     break;
02372   case ICmpInst::ICMP_EQ:
02373     Inverse = true;
02374     break;
02375   case ICmpInst::ICMP_UGT:
02376   case ICmpInst::ICMP_UGE:
02377     if (I.getOperand(0) == MulVal)
02378       break;
02379     Inverse = true;
02380     break;
02381   case ICmpInst::ICMP_ULT:
02382   case ICmpInst::ICMP_ULE:
02383     if (I.getOperand(1) == MulVal)
02384       break;
02385     Inverse = true;
02386     break;
02387   default:
02388     llvm_unreachable("Unexpected predicate");
02389   }
02390   if (Inverse) {
02391     Value *Res = Builder->CreateExtractValue(Call, 1);
02392     return BinaryOperator::CreateNot(Res);
02393   }
02394 
02395   return ExtractValueInst::Create(Call, 1);
02396 }
02397 
02398 // DemandedBitsLHSMask - When performing a comparison against a constant,
02399 // it is possible that not all the bits in the LHS are demanded.  This helper
02400 // method computes the mask that IS demanded.
02401 static APInt DemandedBitsLHSMask(ICmpInst &I,
02402                                  unsigned BitWidth, bool isSignCheck) {
02403   if (isSignCheck)
02404     return APInt::getSignBit(BitWidth);
02405 
02406   ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
02407   if (!CI) return APInt::getAllOnesValue(BitWidth);
02408   const APInt &RHS = CI->getValue();
02409 
02410   switch (I.getPredicate()) {
02411   // For a UGT comparison, we don't care about any bits that
02412   // correspond to the trailing ones of the comparand.  The value of these
02413   // bits doesn't impact the outcome of the comparison, because any value
02414   // greater than the RHS must differ in a bit higher than these due to carry.
02415   case ICmpInst::ICMP_UGT: {
02416     unsigned trailingOnes = RHS.countTrailingOnes();
02417     APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
02418     return ~lowBitsSet;
02419   }
02420 
02421   // Similarly, for a ULT comparison, we don't care about the trailing zeros.
02422   // Any value less than the RHS must differ in a higher bit because of carries.
02423   case ICmpInst::ICMP_ULT: {
02424     unsigned trailingZeros = RHS.countTrailingZeros();
02425     APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
02426     return ~lowBitsSet;
02427   }
02428 
02429   default:
02430     return APInt::getAllOnesValue(BitWidth);
02431   }
02432 
02433 }
02434 
02435 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
02436 /// should be swapped.
02437 /// The decision is based on how many times these two operands are reused
02438 /// as subtract operands and their positions in those instructions.
02439 /// The rational is that several architectures use the same instruction for
02440 /// both subtract and cmp, thus it is better if the order of those operands
02441 /// match.
02442 /// \return true if Op0 and Op1 should be swapped.
02443 static bool swapMayExposeCSEOpportunities(const Value * Op0,
02444                                           const Value * Op1) {
02445   // Filter out pointer value as those cannot appears directly in subtract.
02446   // FIXME: we may want to go through inttoptrs or bitcasts.
02447   if (Op0->getType()->isPointerTy())
02448     return false;
02449   // Count every uses of both Op0 and Op1 in a subtract.
02450   // Each time Op0 is the first operand, count -1: swapping is bad, the
02451   // subtract has already the same layout as the compare.
02452   // Each time Op0 is the second operand, count +1: swapping is good, the
02453   // subtract has a different layout as the compare.
02454   // At the end, if the benefit is greater than 0, Op0 should come second to
02455   // expose more CSE opportunities.
02456   int GlobalSwapBenefits = 0;
02457   for (const User *U : Op0->users()) {
02458     const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
02459     if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
02460       continue;
02461     // If Op0 is the first argument, this is not beneficial to swap the
02462     // arguments.
02463     int LocalSwapBenefits = -1;
02464     unsigned Op1Idx = 1;
02465     if (BinOp->getOperand(Op1Idx) == Op0) {
02466       Op1Idx = 0;
02467       LocalSwapBenefits = 1;
02468     }
02469     if (BinOp->getOperand(Op1Idx) != Op1)
02470       continue;
02471     GlobalSwapBenefits += LocalSwapBenefits;
02472   }
02473   return GlobalSwapBenefits > 0;
02474 }
02475 
02476 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
02477   bool Changed = false;
02478   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
02479   unsigned Op0Cplxity = getComplexity(Op0);
02480   unsigned Op1Cplxity = getComplexity(Op1);
02481 
02482   /// Orders the operands of the compare so that they are listed from most
02483   /// complex to least complex.  This puts constants before unary operators,
02484   /// before binary operators.
02485   if (Op0Cplxity < Op1Cplxity ||
02486         (Op0Cplxity == Op1Cplxity &&
02487          swapMayExposeCSEOpportunities(Op0, Op1))) {
02488     I.swapOperands();
02489     std::swap(Op0, Op1);
02490     Changed = true;
02491   }
02492 
02493   if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL, TLI, DT, AT))
02494     return ReplaceInstUsesWith(I, V);
02495 
02496   // comparing -val or val with non-zero is the same as just comparing val
02497   // ie, abs(val) != 0 -> val != 0
02498   if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
02499   {
02500     Value *Cond, *SelectTrue, *SelectFalse;
02501     if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
02502                             m_Value(SelectFalse)))) {
02503       if (Value *V = dyn_castNegVal(SelectTrue)) {
02504         if (V == SelectFalse)
02505           return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
02506       }
02507       else if (Value *V = dyn_castNegVal(SelectFalse)) {
02508         if (V == SelectTrue)
02509           return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
02510       }
02511     }
02512   }
02513 
02514   Type *Ty = Op0->getType();
02515 
02516   // icmp's with boolean values can always be turned into bitwise operations
02517   if (Ty->isIntegerTy(1)) {
02518     switch (I.getPredicate()) {
02519     default: llvm_unreachable("Invalid icmp instruction!");
02520     case ICmpInst::ICMP_EQ: {               // icmp eq i1 A, B -> ~(A^B)
02521       Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
02522       return BinaryOperator::CreateNot(Xor);
02523     }
02524     case ICmpInst::ICMP_NE:                  // icmp eq i1 A, B -> A^B
02525       return BinaryOperator::CreateXor(Op0, Op1);
02526 
02527     case ICmpInst::ICMP_UGT:
02528       std::swap(Op0, Op1);                   // Change icmp ugt -> icmp ult
02529       // FALL THROUGH
02530     case ICmpInst::ICMP_ULT:{               // icmp ult i1 A, B -> ~A & B
02531       Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
02532       return BinaryOperator::CreateAnd(Not, Op1);
02533     }
02534     case ICmpInst::ICMP_SGT:
02535       std::swap(Op0, Op1);                   // Change icmp sgt -> icmp slt
02536       // FALL THROUGH
02537     case ICmpInst::ICMP_SLT: {               // icmp slt i1 A, B -> A & ~B
02538       Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
02539       return BinaryOperator::CreateAnd(Not, Op0);
02540     }
02541     case ICmpInst::ICMP_UGE:
02542       std::swap(Op0, Op1);                   // Change icmp uge -> icmp ule
02543       // FALL THROUGH
02544     case ICmpInst::ICMP_ULE: {               //  icmp ule i1 A, B -> ~A | B
02545       Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
02546       return BinaryOperator::CreateOr(Not, Op1);
02547     }
02548     case ICmpInst::ICMP_SGE:
02549       std::swap(Op0, Op1);                   // Change icmp sge -> icmp sle
02550       // FALL THROUGH
02551     case ICmpInst::ICMP_SLE: {               //  icmp sle i1 A, B -> A | ~B
02552       Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
02553       return BinaryOperator::CreateOr(Not, Op0);
02554     }
02555     }
02556   }
02557 
02558   unsigned BitWidth = 0;
02559   if (Ty->isIntOrIntVectorTy())
02560     BitWidth = Ty->getScalarSizeInBits();
02561   else if (DL)  // Pointers require DL info to get their size.
02562     BitWidth = DL->getTypeSizeInBits(Ty->getScalarType());
02563 
02564   bool isSignBit = false;
02565 
02566   // See if we are doing a comparison with a constant.
02567   if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02568     Value *A = nullptr, *B = nullptr;
02569 
02570     // Match the following pattern, which is a common idiom when writing
02571     // overflow-safe integer arithmetic function.  The source performs an
02572     // addition in wider type, and explicitly checks for overflow using
02573     // comparisons against INT_MIN and INT_MAX.  Simplify this by using the
02574     // sadd_with_overflow intrinsic.
02575     //
02576     // TODO: This could probably be generalized to handle other overflow-safe
02577     // operations if we worked out the formulas to compute the appropriate
02578     // magic constants.
02579     //
02580     // sum = a + b
02581     // if (sum+128 >u 255)  ...  -> llvm.sadd.with.overflow.i8
02582     {
02583     ConstantInt *CI2;    // I = icmp ugt (add (add A, B), CI2), CI
02584     if (I.getPredicate() == ICmpInst::ICMP_UGT &&
02585         match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
02586       if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
02587         return Res;
02588     }
02589 
02590     // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
02591     if (I.isEquality() && CI->isZero() &&
02592         match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
02593       // (icmp cond A B) if cond is equality
02594       return new ICmpInst(I.getPredicate(), A, B);
02595     }
02596 
02597     // If we have an icmp le or icmp ge instruction, turn it into the
02598     // appropriate icmp lt or icmp gt instruction.  This allows us to rely on
02599     // them being folded in the code below.  The SimplifyICmpInst code has
02600     // already handled the edge cases for us, so we just assert on them.
02601     switch (I.getPredicate()) {
02602     default: break;
02603     case ICmpInst::ICMP_ULE:
02604       assert(!CI->isMaxValue(false));                 // A <=u MAX -> TRUE
02605       return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
02606                           Builder->getInt(CI->getValue()+1));
02607     case ICmpInst::ICMP_SLE:
02608       assert(!CI->isMaxValue(true));                  // A <=s MAX -> TRUE
02609       return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
02610                           Builder->getInt(CI->getValue()+1));
02611     case ICmpInst::ICMP_UGE:
02612       assert(!CI->isMinValue(false));                 // A >=u MIN -> TRUE
02613       return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
02614                           Builder->getInt(CI->getValue()-1));
02615     case ICmpInst::ICMP_SGE:
02616       assert(!CI->isMinValue(true));                  // A >=s MIN -> TRUE
02617       return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
02618                           Builder->getInt(CI->getValue()-1));
02619     }
02620 
02621     if (I.isEquality()) {
02622       ConstantInt *CI2;
02623       if (match(Op0, m_AShr(m_ConstantInt(CI2), m_Value(A))) ||
02624           match(Op0, m_LShr(m_ConstantInt(CI2), m_Value(A)))) {
02625         // (icmp eq/ne (ashr/lshr const2, A), const1)
02626         return FoldICmpCstShrCst(I, Op0, A, CI, CI2);
02627       }
02628       if (match(Op0, m_Shl(m_ConstantInt(CI2), m_Value(A)))) {
02629         // (icmp eq/ne (shl const2, A), const1)
02630         return FoldICmpCstShlCst(I, Op0, A, CI, CI2);
02631       }
02632     }
02633 
02634     // If this comparison is a normal comparison, it demands all
02635     // bits, if it is a sign bit comparison, it only demands the sign bit.
02636     bool UnusedBit;
02637     isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
02638   }
02639 
02640   // See if we can fold the comparison based on range information we can get
02641   // by checking whether bits are known to be zero or one in the input.
02642   if (BitWidth != 0) {
02643     APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
02644     APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
02645 
02646     if (SimplifyDemandedBits(I.getOperandUse(0),
02647                              DemandedBitsLHSMask(I, BitWidth, isSignBit),
02648                              Op0KnownZero, Op0KnownOne, 0))
02649       return &I;
02650     if (SimplifyDemandedBits(I.getOperandUse(1),
02651                              APInt::getAllOnesValue(BitWidth),
02652                              Op1KnownZero, Op1KnownOne, 0))
02653       return &I;
02654 
02655     // Given the known and unknown bits, compute a range that the LHS could be
02656     // in.  Compute the Min, Max and RHS values based on the known bits. For the
02657     // EQ and NE we use unsigned values.
02658     APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
02659     APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
02660     if (I.isSigned()) {
02661       ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
02662                                              Op0Min, Op0Max);
02663       ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
02664                                              Op1Min, Op1Max);
02665     } else {
02666       ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
02667                                                Op0Min, Op0Max);
02668       ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
02669                                                Op1Min, Op1Max);
02670     }
02671 
02672     // If Min and Max are known to be the same, then SimplifyDemandedBits
02673     // figured out that the LHS is a constant.  Just constant fold this now so
02674     // that code below can assume that Min != Max.
02675     if (!isa<Constant>(Op0) && Op0Min == Op0Max)
02676       return new ICmpInst(I.getPredicate(),
02677                           ConstantInt::get(Op0->getType(), Op0Min), Op1);
02678     if (!isa<Constant>(Op1) && Op1Min == Op1Max)
02679       return new ICmpInst(I.getPredicate(), Op0,
02680                           ConstantInt::get(Op1->getType(), Op1Min));
02681 
02682     // Based on the range information we know about the LHS, see if we can
02683     // simplify this comparison.  For example, (x&4) < 8 is always true.
02684     switch (I.getPredicate()) {
02685     default: llvm_unreachable("Unknown icmp opcode!");
02686     case ICmpInst::ICMP_EQ: {
02687       if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
02688         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02689 
02690       // If all bits are known zero except for one, then we know at most one
02691       // bit is set.   If the comparison is against zero, then this is a check
02692       // to see if *that* bit is set.
02693       APInt Op0KnownZeroInverted = ~Op0KnownZero;
02694       if (~Op1KnownZero == 0) {
02695         // If the LHS is an AND with the same constant, look through it.
02696         Value *LHS = nullptr;
02697         ConstantInt *LHSC = nullptr;
02698         if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
02699             LHSC->getValue() != Op0KnownZeroInverted)
02700           LHS = Op0;
02701 
02702         // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
02703         // then turn "((1 << x)&8) == 0" into "x != 3".
02704         // or turn "((1 << x)&7) == 0" into "x > 2".
02705         Value *X = nullptr;
02706         if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
02707           APInt ValToCheck = Op0KnownZeroInverted;
02708           if (ValToCheck.isPowerOf2()) {
02709             unsigned CmpVal = ValToCheck.countTrailingZeros();
02710             return new ICmpInst(ICmpInst::ICMP_NE, X,
02711                                 ConstantInt::get(X->getType(), CmpVal));
02712           } else if ((++ValToCheck).isPowerOf2()) {
02713             unsigned CmpVal = ValToCheck.countTrailingZeros() - 1;
02714             return new ICmpInst(ICmpInst::ICMP_UGT, X,
02715                                 ConstantInt::get(X->getType(), CmpVal));
02716           }
02717         }
02718 
02719         // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
02720         // then turn "((8 >>u x)&1) == 0" into "x != 3".
02721         const APInt *CI;
02722         if (Op0KnownZeroInverted == 1 &&
02723             match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
02724           return new ICmpInst(ICmpInst::ICMP_NE, X,
02725                               ConstantInt::get(X->getType(),
02726                                                CI->countTrailingZeros()));
02727       }
02728 
02729       break;
02730     }
02731     case ICmpInst::ICMP_NE: {
02732       if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
02733         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02734 
02735       // If all bits are known zero except for one, then we know at most one
02736       // bit is set.   If the comparison is against zero, then this is a check
02737       // to see if *that* bit is set.
02738       APInt Op0KnownZeroInverted = ~Op0KnownZero;
02739       if (~Op1KnownZero == 0) {
02740         // If the LHS is an AND with the same constant, look through it.
02741         Value *LHS = nullptr;
02742         ConstantInt *LHSC = nullptr;
02743         if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
02744             LHSC->getValue() != Op0KnownZeroInverted)
02745           LHS = Op0;
02746 
02747         // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
02748         // then turn "((1 << x)&8) != 0" into "x == 3".
02749         // or turn "((1 << x)&7) != 0" into "x < 3".
02750         Value *X = nullptr;
02751         if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
02752           APInt ValToCheck = Op0KnownZeroInverted;
02753           if (ValToCheck.isPowerOf2()) {
02754             unsigned CmpVal = ValToCheck.countTrailingZeros();
02755             return new ICmpInst(ICmpInst::ICMP_EQ, X,
02756                                 ConstantInt::get(X->getType(), CmpVal));
02757           } else if ((++ValToCheck).isPowerOf2()) {
02758             unsigned CmpVal = ValToCheck.countTrailingZeros();
02759             return new ICmpInst(ICmpInst::ICMP_ULT, X,
02760                                 ConstantInt::get(X->getType(), CmpVal));
02761           }
02762         }
02763 
02764         // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
02765         // then turn "((8 >>u x)&1) != 0" into "x == 3".
02766         const APInt *CI;
02767         if (Op0KnownZeroInverted == 1 &&
02768             match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
02769           return new ICmpInst(ICmpInst::ICMP_EQ, X,
02770                               ConstantInt::get(X->getType(),
02771                                                CI->countTrailingZeros()));
02772       }
02773 
02774       break;
02775     }
02776     case ICmpInst::ICMP_ULT:
02777       if (Op0Max.ult(Op1Min))          // A <u B -> true if max(A) < min(B)
02778         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02779       if (Op0Min.uge(Op1Max))          // A <u B -> false if min(A) >= max(B)
02780         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02781       if (Op1Min == Op0Max)            // A <u B -> A != B if max(A) == min(B)
02782         return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
02783       if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02784         if (Op1Max == Op0Min+1)        // A <u C -> A == C-1 if min(A)+1 == C
02785           return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
02786                               Builder->getInt(CI->getValue()-1));
02787 
02788         // (x <u 2147483648) -> (x >s -1)  -> true if sign bit clear
02789         if (CI->isMinValue(true))
02790           return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
02791                            Constant::getAllOnesValue(Op0->getType()));
02792       }
02793       break;
02794     case ICmpInst::ICMP_UGT:
02795       if (Op0Min.ugt(Op1Max))          // A >u B -> true if min(A) > max(B)
02796         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02797       if (Op0Max.ule(Op1Min))          // A >u B -> false if max(A) <= max(B)
02798         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02799 
02800       if (Op1Max == Op0Min)            // A >u 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 >u 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         // (x >u 2147483647) -> (x <s 0)  -> true if sign bit set
02808         if (CI->isMaxValue(true))
02809           return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
02810                               Constant::getNullValue(Op0->getType()));
02811       }
02812       break;
02813     case ICmpInst::ICMP_SLT:
02814       if (Op0Max.slt(Op1Min))          // A <s B -> true if max(A) < min(C)
02815         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02816       if (Op0Min.sge(Op1Max))          // A <s B -> false if min(A) >= max(C)
02817         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02818       if (Op1Min == Op0Max)            // A <s B -> A != B if max(A) == min(B)
02819         return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
02820       if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02821         if (Op1Max == Op0Min+1)        // A <s C -> A == C-1 if min(A)+1 == C
02822           return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
02823                               Builder->getInt(CI->getValue()-1));
02824       }
02825       break;
02826     case ICmpInst::ICMP_SGT:
02827       if (Op0Min.sgt(Op1Max))          // A >s B -> true if min(A) > max(B)
02828         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02829       if (Op0Max.sle(Op1Min))          // A >s B -> false if max(A) <= min(B)
02830         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02831 
02832       if (Op1Max == Op0Min)            // A >s B -> A != B if min(A) == max(B)
02833         return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
02834       if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02835         if (Op1Min == Op0Max-1)        // A >s C -> A == C+1 if max(A)-1 == C
02836           return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
02837                               Builder->getInt(CI->getValue()+1));
02838       }
02839       break;
02840     case ICmpInst::ICMP_SGE:
02841       assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
02842       if (Op0Min.sge(Op1Max))          // A >=s B -> true if min(A) >= max(B)
02843         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02844       if (Op0Max.slt(Op1Min))          // A >=s B -> false if max(A) < min(B)
02845         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02846       break;
02847     case ICmpInst::ICMP_SLE:
02848       assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
02849       if (Op0Max.sle(Op1Min))          // A <=s B -> true if max(A) <= min(B)
02850         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02851       if (Op0Min.sgt(Op1Max))          // A <=s B -> false if min(A) > max(B)
02852         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02853       break;
02854     case ICmpInst::ICMP_UGE:
02855       assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
02856       if (Op0Min.uge(Op1Max))          // A >=u B -> true if min(A) >= max(B)
02857         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02858       if (Op0Max.ult(Op1Min))          // A >=u B -> false if max(A) < min(B)
02859         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02860       break;
02861     case ICmpInst::ICMP_ULE:
02862       assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
02863       if (Op0Max.ule(Op1Min))          // A <=u B -> true if max(A) <= min(B)
02864         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02865       if (Op0Min.ugt(Op1Max))          // A <=u B -> false if min(A) > max(B)
02866         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02867       break;
02868     }
02869 
02870     // Turn a signed comparison into an unsigned one if both operands
02871     // are known to have the same sign.
02872     if (I.isSigned() &&
02873         ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
02874          (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
02875       return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
02876   }
02877 
02878   // Test if the ICmpInst instruction is used exclusively by a select as
02879   // part of a minimum or maximum operation. If so, refrain from doing
02880   // any other folding. This helps out other analyses which understand
02881   // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
02882   // and CodeGen. And in this case, at least one of the comparison
02883   // operands has at least one user besides the compare (the select),
02884   // which would often largely negate the benefit of folding anyway.
02885   if (I.hasOneUse())
02886     if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
02887       if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
02888           (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
02889         return nullptr;
02890 
02891   // See if we are doing a comparison between a constant and an instruction that
02892   // can be folded into the comparison.
02893   if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02894     // Since the RHS is a ConstantInt (CI), if the left hand side is an
02895     // instruction, see if that instruction also has constants so that the
02896     // instruction can be folded into the icmp
02897     if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
02898       if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
02899         return Res;
02900   }
02901 
02902   // Handle icmp with constant (but not simple integer constant) RHS
02903   if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
02904     if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
02905       switch (LHSI->getOpcode()) {
02906       case Instruction::GetElementPtr:
02907           // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
02908         if (RHSC->isNullValue() &&
02909             cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
02910           return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
02911                   Constant::getNullValue(LHSI->getOperand(0)->getType()));
02912         break;
02913       case Instruction::PHI:
02914         // Only fold icmp into the PHI if the phi and icmp are in the same
02915         // block.  If in the same block, we're encouraging jump threading.  If
02916         // not, we are just pessimizing the code by making an i1 phi.
02917         if (LHSI->getParent() == I.getParent())
02918           if (Instruction *NV = FoldOpIntoPhi(I))
02919             return NV;
02920         break;
02921       case Instruction::Select: {
02922         // If either operand of the select is a constant, we can fold the
02923         // comparison into the select arms, which will cause one to be
02924         // constant folded and the select turned into a bitwise or.
02925         Value *Op1 = nullptr, *Op2 = nullptr;
02926         if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
02927           Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
02928         if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
02929           Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
02930 
02931         // We only want to perform this transformation if it will not lead to
02932         // additional code. This is true if either both sides of the select
02933         // fold to a constant (in which case the icmp is replaced with a select
02934         // which will usually simplify) or this is the only user of the
02935         // select (in which case we are trading a select+icmp for a simpler
02936         // select+icmp).
02937         if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
02938           if (!Op1)
02939             Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
02940                                       RHSC, I.getName());
02941           if (!Op2)
02942             Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
02943                                       RHSC, I.getName());
02944           return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
02945         }
02946         break;
02947       }
02948       case Instruction::IntToPtr:
02949         // icmp pred inttoptr(X), null -> icmp pred X, 0
02950         if (RHSC->isNullValue() && DL &&
02951             DL->getIntPtrType(RHSC->getType()) ==
02952                LHSI->getOperand(0)->getType())
02953           return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
02954                         Constant::getNullValue(LHSI->getOperand(0)->getType()));
02955         break;
02956 
02957       case Instruction::Load:
02958         // Try to optimize things like "A[i] > 4" to index computations.
02959         if (GetElementPtrInst *GEP =
02960               dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
02961           if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
02962             if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
02963                 !cast<LoadInst>(LHSI)->isVolatile())
02964               if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
02965                 return Res;
02966         }
02967         break;
02968       }
02969   }
02970 
02971   // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
02972   if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
02973     if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
02974       return NI;
02975   if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
02976     if (Instruction *NI = FoldGEPICmp(GEP, Op0,
02977                            ICmpInst::getSwappedPredicate(I.getPredicate()), I))
02978       return NI;
02979 
02980   // Test to see if the operands of the icmp are casted versions of other
02981   // values.  If the ptr->ptr cast can be stripped off both arguments, we do so
02982   // now.
02983   if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
02984     if (Op0->getType()->isPointerTy() &&
02985         (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
02986       // We keep moving the cast from the left operand over to the right
02987       // operand, where it can often be eliminated completely.
02988       Op0 = CI->getOperand(0);
02989 
02990       // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
02991       // so eliminate it as well.
02992       if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
02993         Op1 = CI2->getOperand(0);
02994 
02995       // If Op1 is a constant, we can fold the cast into the constant.
02996       if (Op0->getType() != Op1->getType()) {
02997         if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
02998           Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
02999         } else {
03000           // Otherwise, cast the RHS right before the icmp
03001           Op1 = Builder->CreateBitCast(Op1, Op0->getType());
03002         }
03003       }
03004       return new ICmpInst(I.getPredicate(), Op0, Op1);
03005     }
03006   }
03007 
03008   if (isa<CastInst>(Op0)) {
03009     // Handle the special case of: icmp (cast bool to X), <cst>
03010     // This comes up when you have code like
03011     //   int X = A < B;
03012     //   if (X) ...
03013     // For generality, we handle any zero-extension of any operand comparison
03014     // with a constant or another cast from the same type.
03015     if (isa<Constant>(Op1) || isa<CastInst>(Op1))
03016       if (Instruction *R = visitICmpInstWithCastAndCast(I))
03017         return R;
03018   }
03019 
03020   // Special logic for binary operators.
03021   BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
03022   BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
03023   if (BO0 || BO1) {
03024     CmpInst::Predicate Pred = I.getPredicate();
03025     bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
03026     if (BO0 && isa<OverflowingBinaryOperator>(BO0))
03027       NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
03028         (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
03029         (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
03030     if (BO1 && isa<OverflowingBinaryOperator>(BO1))
03031       NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
03032         (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
03033         (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
03034 
03035     // Analyze the case when either Op0 or Op1 is an add instruction.
03036     // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
03037     Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
03038     if (BO0 && BO0->getOpcode() == Instruction::Add)
03039       A = BO0->getOperand(0), B = BO0->getOperand(1);
03040     if (BO1 && BO1->getOpcode() == Instruction::Add)
03041       C = BO1->getOperand(0), D = BO1->getOperand(1);
03042 
03043     // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
03044     if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
03045       return new ICmpInst(Pred, A == Op1 ? B : A,
03046                           Constant::getNullValue(Op1->getType()));
03047 
03048     // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
03049     if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
03050       return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
03051                           C == Op0 ? D : C);
03052 
03053     // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
03054     if (A && C && (A == C || A == D || B == C || B == D) &&
03055         NoOp0WrapProblem && NoOp1WrapProblem &&
03056         // Try not to increase register pressure.
03057         BO0->hasOneUse() && BO1->hasOneUse()) {
03058       // Determine Y and Z in the form icmp (X+Y), (X+Z).
03059       Value *Y, *Z;
03060       if (A == C) {
03061         // C + B == C + D  ->  B == D
03062         Y = B;
03063         Z = D;
03064       } else if (A == D) {
03065         // D + B == C + D  ->  B == C
03066         Y = B;
03067         Z = C;
03068       } else if (B == C) {
03069         // A + C == C + D  ->  A == D
03070         Y = A;
03071         Z = D;
03072       } else {
03073         assert(B == D);
03074         // A + D == C + D  ->  A == C
03075         Y = A;
03076         Z = C;
03077       }
03078       return new ICmpInst(Pred, Y, Z);
03079     }
03080 
03081     // icmp slt (X + -1), Y -> icmp sle X, Y
03082     if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
03083         match(B, m_AllOnes()))
03084       return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
03085 
03086     // icmp sge (X + -1), Y -> icmp sgt X, Y
03087     if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
03088         match(B, m_AllOnes()))
03089       return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
03090 
03091     // icmp sle (X + 1), Y -> icmp slt X, Y
03092     if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
03093         match(B, m_One()))
03094       return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
03095 
03096     // icmp sgt (X + 1), Y -> icmp sge X, Y
03097     if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
03098         match(B, m_One()))
03099       return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
03100 
03101     // if C1 has greater magnitude than C2:
03102     //  icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
03103     //  s.t. C3 = C1 - C2
03104     //
03105     // if C2 has greater magnitude than C1:
03106     //  icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
03107     //  s.t. C3 = C2 - C1
03108     if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
03109         (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
03110       if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
03111         if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
03112           const APInt &AP1 = C1->getValue();
03113           const APInt &AP2 = C2->getValue();
03114           if (AP1.isNegative() == AP2.isNegative()) {
03115             APInt AP1Abs = C1->getValue().abs();
03116             APInt AP2Abs = C2->getValue().abs();
03117             if (AP1Abs.uge(AP2Abs)) {
03118               ConstantInt *C3 = Builder->getInt(AP1 - AP2);
03119               Value *NewAdd = Builder->CreateNSWAdd(A, C3);
03120               return new ICmpInst(Pred, NewAdd, C);
03121             } else {
03122               ConstantInt *C3 = Builder->getInt(AP2 - AP1);
03123               Value *NewAdd = Builder->CreateNSWAdd(C, C3);
03124               return new ICmpInst(Pred, A, NewAdd);
03125             }
03126           }
03127         }
03128 
03129 
03130     // Analyze the case when either Op0 or Op1 is a sub instruction.
03131     // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
03132     A = nullptr; B = nullptr; C = nullptr; D = nullptr;
03133     if (BO0 && BO0->getOpcode() == Instruction::Sub)
03134       A = BO0->getOperand(0), B = BO0->getOperand(1);
03135     if (BO1 && BO1->getOpcode() == Instruction::Sub)
03136       C = BO1->getOperand(0), D = BO1->getOperand(1);
03137 
03138     // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
03139     if (A == Op1 && NoOp0WrapProblem)
03140       return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
03141 
03142     // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
03143     if (C == Op0 && NoOp1WrapProblem)
03144       return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
03145 
03146     // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
03147     if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
03148         // Try not to increase register pressure.
03149         BO0->hasOneUse() && BO1->hasOneUse())
03150       return new ICmpInst(Pred, A, C);
03151 
03152     // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
03153     if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
03154         // Try not to increase register pressure.
03155         BO0->hasOneUse() && BO1->hasOneUse())
03156       return new ICmpInst(Pred, D, B);
03157 
03158     // icmp (0-X) < cst --> x > -cst
03159     if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
03160       Value *X;
03161       if (match(BO0, m_Neg(m_Value(X))))
03162         if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
03163           if (!RHSC->isMinValue(/*isSigned=*/true))
03164             return new ICmpInst(I.getSwappedPredicate(), X,
03165                                 ConstantExpr::getNeg(RHSC));
03166     }
03167 
03168     BinaryOperator *SRem = nullptr;
03169     // icmp (srem X, Y), Y
03170     if (BO0 && BO0->getOpcode() == Instruction::SRem &&
03171         Op1 == BO0->getOperand(1))
03172       SRem = BO0;
03173     // icmp Y, (srem X, Y)
03174     else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
03175              Op0 == BO1->getOperand(1))
03176       SRem = BO1;
03177     if (SRem) {
03178       // We don't check hasOneUse to avoid increasing register pressure because
03179       // the value we use is the same value this instruction was already using.
03180       switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
03181         default: break;
03182         case ICmpInst::ICMP_EQ:
03183           return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
03184         case ICmpInst::ICMP_NE:
03185           return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
03186         case ICmpInst::ICMP_SGT:
03187         case ICmpInst::ICMP_SGE:
03188           return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
03189                               Constant::getAllOnesValue(SRem->getType()));
03190         case ICmpInst::ICMP_SLT:
03191         case ICmpInst::ICMP_SLE:
03192           return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
03193                               Constant::getNullValue(SRem->getType()));
03194       }
03195     }
03196 
03197     if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
03198         BO0->hasOneUse() && BO1->hasOneUse() &&
03199         BO0->getOperand(1) == BO1->getOperand(1)) {
03200       switch (BO0->getOpcode()) {
03201       default: break;
03202       case Instruction::Add:
03203       case Instruction::Sub:
03204       case Instruction::Xor:
03205         if (I.isEquality())    // a+x icmp eq/ne b+x --> a icmp b
03206           return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
03207                               BO1->getOperand(0));
03208         // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
03209         if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
03210           if (CI->getValue().isSignBit()) {
03211             ICmpInst::Predicate Pred = I.isSigned()
03212                                            ? I.getUnsignedPredicate()
03213                                            : I.getSignedPredicate();
03214             return new ICmpInst(Pred, BO0->getOperand(0),
03215                                 BO1->getOperand(0));
03216           }
03217 
03218           if (CI->isMaxValue(true)) {
03219             ICmpInst::Predicate Pred = I.isSigned()
03220                                            ? I.getUnsignedPredicate()
03221                                            : I.getSignedPredicate();
03222             Pred = I.getSwappedPredicate(Pred);
03223             return new ICmpInst(Pred, BO0->getOperand(0),
03224                                 BO1->getOperand(0));
03225           }
03226         }
03227         break;
03228       case Instruction::Mul:
03229         if (!I.isEquality())
03230           break;
03231 
03232         if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
03233           // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
03234           // Mask = -1 >> count-trailing-zeros(Cst).
03235           if (!CI->isZero() && !CI->isOne()) {
03236             const APInt &AP = CI->getValue();
03237             ConstantInt *Mask = ConstantInt::get(I.getContext(),
03238                                     APInt::getLowBitsSet(AP.getBitWidth(),
03239                                                          AP.getBitWidth() -
03240                                                     AP.countTrailingZeros()));
03241             Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
03242             Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
03243             return new ICmpInst(I.getPredicate(), And1, And2);
03244           }
03245         }
03246         break;
03247       case Instruction::UDiv:
03248       case Instruction::LShr:
03249         if (I.isSigned())
03250           break;
03251         // fall-through
03252       case Instruction::SDiv:
03253       case Instruction::AShr:
03254         if (!BO0->isExact() || !BO1->isExact())
03255           break;
03256         return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
03257                             BO1->getOperand(0));
03258       case Instruction::Shl: {
03259         bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
03260         bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
03261         if (!NUW && !NSW)
03262           break;
03263         if (!NSW && I.isSigned())
03264           break;
03265         return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
03266                             BO1->getOperand(0));
03267       }
03268       }
03269     }
03270   }
03271 
03272   { Value *A, *B;
03273     // Transform (A & ~B) == 0 --> (A & B) != 0
03274     // and       (A & ~B) != 0 --> (A & B) == 0
03275     // if A is a power of 2.
03276     if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
03277         match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A, false,
03278                                                        0, AT, &I, DT) &&
03279                                 I.isEquality())
03280       return new ICmpInst(I.getInversePredicate(),
03281                           Builder->CreateAnd(A, B),
03282                           Op1);
03283 
03284     // ~x < ~y --> y < x
03285     // ~x < cst --> ~cst < x
03286     if (match(Op0, m_Not(m_Value(A)))) {
03287       if (match(Op1, m_Not(m_Value(B))))
03288         return new ICmpInst(I.getPredicate(), B, A);
03289       if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
03290         return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
03291     }
03292 
03293     // (a+b) <u a  --> llvm.uadd.with.overflow.
03294     // (a+b) <u b  --> llvm.uadd.with.overflow.
03295     if (I.getPredicate() == ICmpInst::ICMP_ULT &&
03296         match(Op0, m_Add(m_Value(A), m_Value(B))) &&
03297         (Op1 == A || Op1 == B))
03298       if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
03299         return R;
03300 
03301     // a >u (a+b)  --> llvm.uadd.with.overflow.
03302     // b >u (a+b)  --> llvm.uadd.with.overflow.
03303     if (I.getPredicate() == ICmpInst::ICMP_UGT &&
03304         match(Op1, m_Add(m_Value(A), m_Value(B))) &&
03305         (Op0 == A || Op0 == B))
03306       if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
03307         return R;
03308 
03309     // (zext a) * (zext b)  --> llvm.umul.with.overflow.
03310     if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
03311       if (Instruction *R = ProcessUMulZExtIdiom(I, Op0, Op1, *this))
03312         return R;
03313     }
03314     if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
03315       if (Instruction *R = ProcessUMulZExtIdiom(I, Op1, Op0, *this))
03316         return R;
03317     }
03318   }
03319 
03320   if (I.isEquality()) {
03321     Value *A, *B, *C, *D;
03322 
03323     if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
03324       if (A == Op1 || B == Op1) {    // (A^B) == A  ->  B == 0
03325         Value *OtherVal = A == Op1 ? B : A;
03326         return new ICmpInst(I.getPredicate(), OtherVal,
03327                             Constant::getNullValue(A->getType()));
03328       }
03329 
03330       if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
03331         // A^c1 == C^c2 --> A == C^(c1^c2)
03332         ConstantInt *C1, *C2;
03333         if (match(B, m_ConstantInt(C1)) &&
03334             match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
03335           Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
03336           Value *Xor = Builder->CreateXor(C, NC);
03337           return new ICmpInst(I.getPredicate(), A, Xor);
03338         }
03339 
03340         // A^B == A^D -> B == D
03341         if (A == C) return new ICmpInst(I.getPredicate(), B, D);
03342         if (A == D) return new ICmpInst(I.getPredicate(), B, C);
03343         if (B == C) return new ICmpInst(I.getPredicate(), A, D);
03344         if (B == D) return new ICmpInst(I.getPredicate(), A, C);
03345       }
03346     }
03347 
03348     if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
03349         (A == Op0 || B == Op0)) {
03350       // A == (A^B)  ->  B == 0
03351       Value *OtherVal = A == Op0 ? B : A;
03352       return new ICmpInst(I.getPredicate(), OtherVal,
03353                           Constant::getNullValue(A->getType()));
03354     }
03355 
03356     // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
03357     if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
03358         match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
03359       Value *X = nullptr, *Y = nullptr, *Z = nullptr;
03360 
03361       if (A == C) {
03362         X = B; Y = D; Z = A;
03363       } else if (A == D) {
03364         X = B; Y = C; Z = A;
03365       } else if (B == C) {
03366         X = A; Y = D; Z = B;
03367       } else if (B == D) {
03368         X = A; Y = C; Z = B;
03369       }
03370 
03371       if (X) {   // Build (X^Y) & Z
03372         Op1 = Builder->CreateXor(X, Y);
03373         Op1 = Builder->CreateAnd(Op1, Z);
03374         I.setOperand(0, Op1);
03375         I.setOperand(1, Constant::getNullValue(Op1->getType()));
03376         return &I;
03377       }
03378     }
03379 
03380     // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
03381     // and       (B & (1<<X)-1) == (zext A) --> A == (trunc B)
03382     ConstantInt *Cst1;
03383     if ((Op0->hasOneUse() &&
03384          match(Op0, m_ZExt(m_Value(A))) &&
03385          match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
03386         (Op1->hasOneUse() &&
03387          match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
03388          match(Op1, m_ZExt(m_Value(A))))) {
03389       APInt Pow2 = Cst1->getValue() + 1;
03390       if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
03391           Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
03392         return new ICmpInst(I.getPredicate(), A,
03393                             Builder->CreateTrunc(B, A->getType()));
03394     }
03395 
03396     // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
03397     // For lshr and ashr pairs.
03398     if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
03399          match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
03400         (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
03401          match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
03402       unsigned TypeBits = Cst1->getBitWidth();
03403       unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
03404       if (ShAmt < TypeBits && ShAmt != 0) {
03405         ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE
03406                                        ? ICmpInst::ICMP_UGE
03407                                        : ICmpInst::ICMP_ULT;
03408         Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
03409         APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
03410         return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal));
03411       }
03412     }
03413 
03414     // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
03415     // "icmp (and X, mask), cst"
03416     uint64_t ShAmt = 0;
03417     if (Op0->hasOneUse() &&
03418         match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
03419                                            m_ConstantInt(ShAmt))))) &&
03420         match(Op1, m_ConstantInt(Cst1)) &&
03421         // Only do this when A has multiple uses.  This is most important to do
03422         // when it exposes other optimizations.
03423         !A->hasOneUse()) {
03424       unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
03425 
03426       if (ShAmt < ASize) {
03427         APInt MaskV =
03428           APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
03429         MaskV <<= ShAmt;
03430 
03431         APInt CmpV = Cst1->getValue().zext(ASize);
03432         CmpV <<= ShAmt;
03433 
03434         Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
03435         return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
03436       }
03437     }
03438   }
03439 
03440   {
03441     Value *X; ConstantInt *Cst;
03442     // icmp X+Cst, X
03443     if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
03444       return FoldICmpAddOpCst(I, X, Cst, I.getPredicate());
03445 
03446     // icmp X, X+Cst
03447     if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
03448       return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
03449   }
03450   return Changed ? &I : nullptr;
03451 }
03452 
03453 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
03454 ///
03455 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
03456                                                 Instruction *LHSI,
03457                                                 Constant *RHSC) {
03458   if (!isa<ConstantFP>(RHSC)) return nullptr;
03459   const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
03460 
03461   // Get the width of the mantissa.  We don't want to hack on conversions that
03462   // might lose information from the integer, e.g. "i64 -> float"
03463   int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
03464   if (MantissaWidth == -1) return nullptr;  // Unknown.
03465 
03466   // Check to see that the input is converted from an integer type that is small
03467   // enough that preserves all bits.  TODO: check here for "known" sign bits.
03468   // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
03469   unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
03470 
03471   // If this is a uitofp instruction, we need an extra bit to hold the sign.
03472   bool LHSUnsigned = isa<UIToFPInst>(LHSI);
03473   if (LHSUnsigned)
03474     ++InputSize;
03475 
03476   // If the conversion would lose info, don't hack on this.
03477   if ((int)InputSize > MantissaWidth)
03478     return nullptr;
03479 
03480   // Otherwise, we can potentially simplify the comparison.  We know that it
03481   // will always come through as an integer value and we know the constant is
03482   // not a NAN (it would have been previously simplified).
03483   assert(!RHS.isNaN() && "NaN comparison not already folded!");
03484 
03485   ICmpInst::Predicate Pred;
03486   switch (I.getPredicate()) {
03487   default: llvm_unreachable("Unexpected predicate!");
03488   case FCmpInst::FCMP_UEQ:
03489   case FCmpInst::FCMP_OEQ:
03490     Pred = ICmpInst::ICMP_EQ;
03491     break;
03492   case FCmpInst::FCMP_UGT:
03493   case FCmpInst::FCMP_OGT:
03494     Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
03495     break;
03496   case FCmpInst::FCMP_UGE:
03497   case FCmpInst::FCMP_OGE:
03498     Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
03499     break;
03500   case FCmpInst::FCMP_ULT:
03501   case FCmpInst::FCMP_OLT:
03502     Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
03503     break;
03504   case FCmpInst::FCMP_ULE:
03505   case FCmpInst::FCMP_OLE:
03506     Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
03507     break;
03508   case FCmpInst::FCMP_UNE:
03509   case FCmpInst::FCMP_ONE:
03510     Pred = ICmpInst::ICMP_NE;
03511     break;
03512   case FCmpInst::FCMP_ORD:
03513     return ReplaceInstUsesWith(I, Builder->getTrue());
03514   case FCmpInst::FCMP_UNO:
03515     return ReplaceInstUsesWith(I, Builder->getFalse());
03516   }
03517 
03518   IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
03519 
03520   // Now we know that the APFloat is a normal number, zero or inf.
03521 
03522   // See if the FP constant is too large for the integer.  For example,
03523   // comparing an i8 to 300.0.
03524   unsigned IntWidth = IntTy->getScalarSizeInBits();
03525 
03526   if (!LHSUnsigned) {
03527     // If the RHS value is > SignedMax, fold the comparison.  This handles +INF
03528     // and large values.
03529     APFloat SMax(RHS.getSemantics());
03530     SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
03531                           APFloat::rmNearestTiesToEven);
03532     if (SMax.compare(RHS) == APFloat::cmpLessThan) {  // smax < 13123.0
03533       if (Pred == ICmpInst::ICMP_NE  || Pred == ICmpInst::ICMP_SLT ||
03534           Pred == ICmpInst::ICMP_SLE)
03535         return ReplaceInstUsesWith(I, Builder->getTrue());
03536       return ReplaceInstUsesWith(I, Builder->getFalse());
03537     }
03538   } else {
03539     // If the RHS value is > UnsignedMax, fold the comparison. This handles
03540     // +INF and large values.
03541     APFloat UMax(RHS.getSemantics());
03542     UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
03543                           APFloat::rmNearestTiesToEven);
03544     if (UMax.compare(RHS) == APFloat::cmpLessThan) {  // umax < 13123.0
03545       if (Pred == ICmpInst::ICMP_NE  || Pred == ICmpInst::ICMP_ULT ||
03546           Pred == ICmpInst::ICMP_ULE)
03547         return ReplaceInstUsesWith(I, Builder->getTrue());
03548       return ReplaceInstUsesWith(I, Builder->getFalse());
03549     }
03550   }
03551 
03552   if (!LHSUnsigned) {
03553     // See if the RHS value is < SignedMin.
03554     APFloat SMin(RHS.getSemantics());
03555     SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
03556                           APFloat::rmNearestTiesToEven);
03557     if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
03558       if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
03559           Pred == ICmpInst::ICMP_SGE)
03560         return ReplaceInstUsesWith(I, Builder->getTrue());
03561       return ReplaceInstUsesWith(I, Builder->getFalse());
03562     }
03563   } else {
03564     // See if the RHS value is < UnsignedMin.
03565     APFloat SMin(RHS.getSemantics());
03566     SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
03567                           APFloat::rmNearestTiesToEven);
03568     if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
03569       if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
03570           Pred == ICmpInst::ICMP_UGE)
03571         return ReplaceInstUsesWith(I, Builder->getTrue());
03572       return ReplaceInstUsesWith(I, Builder->getFalse());
03573     }
03574   }
03575 
03576   // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
03577   // [0, UMAX], but it may still be fractional.  See if it is fractional by
03578   // casting the FP value to the integer value and back, checking for equality.
03579   // Don't do this for zero, because -0.0 is not fractional.
03580   Constant *RHSInt = LHSUnsigned
03581     ? ConstantExpr::getFPToUI(RHSC, IntTy)
03582     : ConstantExpr::getFPToSI(RHSC, IntTy);
03583   if (!RHS.isZero()) {
03584     bool Equal = LHSUnsigned
03585       ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
03586       : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
03587     if (!Equal) {
03588       // If we had a comparison against a fractional value, we have to adjust
03589       // the compare predicate and sometimes the value.  RHSC is rounded towards
03590       // zero at this point.
03591       switch (Pred) {
03592       default: llvm_unreachable("Unexpected integer comparison!");
03593       case ICmpInst::ICMP_NE:  // (float)int != 4.4   --> true
03594         return ReplaceInstUsesWith(I, Builder->getTrue());
03595       case ICmpInst::ICMP_EQ:  // (float)int == 4.4   --> false
03596         return ReplaceInstUsesWith(I, Builder->getFalse());
03597       case ICmpInst::ICMP_ULE:
03598         // (float)int <= 4.4   --> int <= 4
03599         // (float)int <= -4.4  --> false
03600         if (RHS.isNegative())
03601           return ReplaceInstUsesWith(I, Builder->getFalse());
03602         break;
03603       case ICmpInst::ICMP_SLE:
03604         // (float)int <= 4.4   --> int <= 4
03605         // (float)int <= -4.4  --> int < -4
03606         if (RHS.isNegative())
03607           Pred = ICmpInst::ICMP_SLT;
03608         break;
03609       case ICmpInst::ICMP_ULT:
03610         // (float)int < -4.4   --> false
03611         // (float)int < 4.4    --> int <= 4
03612         if (RHS.isNegative())
03613           return ReplaceInstUsesWith(I, Builder->getFalse());
03614         Pred = ICmpInst::ICMP_ULE;
03615         break;
03616       case ICmpInst::ICMP_SLT:
03617         // (float)int < -4.4   --> int < -4
03618         // (float)int < 4.4    --> int <= 4
03619         if (!RHS.isNegative())
03620           Pred = ICmpInst::ICMP_SLE;
03621         break;
03622       case ICmpInst::ICMP_UGT:
03623         // (float)int > 4.4    --> int > 4
03624         // (float)int > -4.4   --> true
03625         if (RHS.isNegative())
03626           return ReplaceInstUsesWith(I, Builder->getTrue());
03627         break;
03628       case ICmpInst::ICMP_SGT:
03629         // (float)int > 4.4    --> int > 4
03630         // (float)int > -4.4   --> int >= -4
03631         if (RHS.isNegative())
03632           Pred = ICmpInst::ICMP_SGE;
03633         break;
03634       case ICmpInst::ICMP_UGE:
03635         // (float)int >= -4.4   --> true
03636         // (float)int >= 4.4    --> int > 4
03637         if (RHS.isNegative())
03638           return ReplaceInstUsesWith(I, Builder->getTrue());
03639         Pred = ICmpInst::ICMP_UGT;
03640         break;
03641       case ICmpInst::ICMP_SGE:
03642         // (float)int >= -4.4   --> int >= -4
03643         // (float)int >= 4.4    --> int > 4
03644         if (!RHS.isNegative())
03645           Pred = ICmpInst::ICMP_SGT;
03646         break;
03647       }
03648     }
03649   }
03650 
03651   // Lower this FP comparison into an appropriate integer version of the
03652   // comparison.
03653   return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
03654 }
03655 
03656 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
03657   bool Changed = false;
03658 
03659   /// Orders the operands of the compare so that they are listed from most
03660   /// complex to least complex.  This puts constants before unary operators,
03661   /// before binary operators.
03662   if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
03663     I.swapOperands();
03664     Changed = true;
03665   }
03666 
03667   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
03668 
03669   if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, DL, TLI, DT, AT))
03670     return ReplaceInstUsesWith(I, V);
03671 
03672   // Simplify 'fcmp pred X, X'
03673   if (Op0 == Op1) {
03674     switch (I.getPredicate()) {
03675     default: llvm_unreachable("Unknown predicate!");
03676     case FCmpInst::FCMP_UNO:    // True if unordered: isnan(X) | isnan(Y)
03677     case FCmpInst::FCMP_ULT:    // True if unordered or less than
03678     case FCmpInst::FCMP_UGT:    // True if unordered or greater than
03679     case FCmpInst::FCMP_UNE:    // True if unordered or not equal
03680       // Canonicalize these to be 'fcmp uno %X, 0.0'.
03681       I.setPredicate(FCmpInst::FCMP_UNO);
03682       I.setOperand(1, Constant::getNullValue(Op0->getType()));
03683       return &I;
03684 
03685     case FCmpInst::FCMP_ORD:    // True if ordered (no nans)
03686     case FCmpInst::FCMP_OEQ:    // True if ordered and equal
03687     case FCmpInst::FCMP_OGE:    // True if ordered and greater than or equal
03688     case FCmpInst::FCMP_OLE:    // True if ordered and less than or equal
03689       // Canonicalize these to be 'fcmp ord %X, 0.0'.
03690       I.setPredicate(FCmpInst::FCMP_ORD);
03691       I.setOperand(1, Constant::getNullValue(Op0->getType()));
03692       return &I;
03693     }
03694   }
03695 
03696   // Handle fcmp with constant RHS
03697   if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
03698     if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
03699       switch (LHSI->getOpcode()) {
03700       case Instruction::FPExt: {
03701         // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
03702         FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
03703         ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
03704         if (!RHSF)
03705           break;
03706 
03707         const fltSemantics *Sem;
03708         // FIXME: This shouldn't be here.
03709         if (LHSExt->getSrcTy()->isHalfTy())
03710           Sem = &APFloat::IEEEhalf;
03711         else if (LHSExt->getSrcTy()->isFloatTy())
03712           Sem = &APFloat::IEEEsingle;
03713         else if (LHSExt->getSrcTy()->isDoubleTy())
03714           Sem = &APFloat::IEEEdouble;
03715         else if (LHSExt->getSrcTy()->isFP128Ty())
03716           Sem = &APFloat::IEEEquad;
03717         else if (LHSExt->getSrcTy()->isX86_FP80Ty())
03718           Sem = &APFloat::x87DoubleExtended;
03719         else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
03720           Sem = &APFloat::PPCDoubleDouble;
03721         else
03722           break;
03723 
03724         bool Lossy;
03725         APFloat F = RHSF->getValueAPF();
03726         F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
03727 
03728         // Avoid lossy conversions and denormals. Zero is a special case
03729         // that's OK to convert.
03730         APFloat Fabs = F;
03731         Fabs.clearSign();
03732         if (!Lossy &&
03733             ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
03734                  APFloat::cmpLessThan) || Fabs.isZero()))
03735 
03736           return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
03737                               ConstantFP::get(RHSC->getContext(), F));
03738         break;
03739       }
03740       case Instruction::PHI:
03741         // Only fold fcmp into the PHI if the phi and fcmp are in the same
03742         // block.  If in the same block, we're encouraging jump threading.  If
03743         // not, we are just pessimizing the code by making an i1 phi.
03744         if (LHSI->getParent() == I.getParent())
03745           if (Instruction *NV = FoldOpIntoPhi(I))
03746             return NV;
03747         break;
03748       case Instruction::SIToFP:
03749       case Instruction::UIToFP:
03750         if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
03751           return NV;
03752         break;
03753       case Instruction::FSub: {
03754         // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
03755         Value *Op;
03756         if (match(LHSI, m_FNeg(m_Value(Op))))
03757           return new FCmpInst(I.getSwappedPredicate(), Op,
03758                               ConstantExpr::getFNeg(RHSC));
03759         break;
03760       }
03761       case Instruction::Load:
03762         if (GetElementPtrInst *GEP =
03763             dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
03764           if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
03765             if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
03766                 !cast<LoadInst>(LHSI)->isVolatile())
03767               if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
03768                 return Res;
03769         }
03770         break;
03771       case Instruction::Call: {
03772         CallInst *CI = cast<CallInst>(LHSI);
03773         LibFunc::Func Func;
03774         // Various optimization for fabs compared with zero.
03775         if (RHSC->isNullValue() && CI->getCalledFunction() &&
03776             TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
03777             TLI->has(Func)) {
03778           if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
03779               Func == LibFunc::fabsl) {
03780             switch (I.getPredicate()) {
03781             default: break;
03782             // fabs(x) < 0 --> false
03783             case FCmpInst::FCMP_OLT:
03784               return ReplaceInstUsesWith(I, Builder->getFalse());
03785             // fabs(x) > 0 --> x != 0
03786             case FCmpInst::FCMP_OGT:
03787               return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
03788                                   RHSC);
03789             // fabs(x) <= 0 --> x == 0
03790             case FCmpInst::FCMP_OLE:
03791               return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
03792                                   RHSC);
03793             // fabs(x) >= 0 --> !isnan(x)
03794             case FCmpInst::FCMP_OGE:
03795               return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
03796                                   RHSC);
03797             // fabs(x) == 0 --> x == 0
03798             // fabs(x) != 0 --> x != 0
03799             case FCmpInst::FCMP_OEQ:
03800             case FCmpInst::FCMP_UEQ:
03801             case FCmpInst::FCMP_ONE:
03802             case FCmpInst::FCMP_UNE:
03803               return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
03804                                   RHSC);
03805             }
03806           }
03807         }
03808       }
03809       }
03810   }
03811 
03812   // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
03813   Value *X, *Y;
03814   if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
03815     return new FCmpInst(I.getSwappedPredicate(), X, Y);
03816 
03817   // fcmp (fpext x), (fpext y) -> fcmp x, y
03818   if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
03819     if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
03820       if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
03821         return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
03822                             RHSExt->getOperand(0));
03823 
03824   return Changed ? &I : nullptr;
03825 }