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   // If we have X+0, exit early (simplifying logic below) and let it get folded
00744   // elsewhere.   icmp X+0, X  -> icmp X, X
00745   if (CI->isZero()) {
00746     bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
00747     return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
00748   }
00749 
00750   // (X+4) == X -> false.
00751   if (Pred == ICmpInst::ICMP_EQ)
00752     return ReplaceInstUsesWith(ICI, Builder->getFalse());
00753 
00754   // (X+4) != X -> true.
00755   if (Pred == ICmpInst::ICMP_NE)
00756     return ReplaceInstUsesWith(ICI, Builder->getTrue());
00757 
00758   // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
00759   // so the values can never be equal.  Similarly for all other "or equals"
00760   // operators.
00761 
00762   // (X+1) <u X        --> X >u (MAXUINT-1)        --> X == 255
00763   // (X+2) <u X        --> X >u (MAXUINT-2)        --> X > 253
00764   // (X+MAXUINT) <u X  --> X >u (MAXUINT-MAXUINT)  --> X != 0
00765   if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
00766     Value *R =
00767       ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
00768     return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
00769   }
00770 
00771   // (X+1) >u X        --> X <u (0-1)        --> X != 255
00772   // (X+2) >u X        --> X <u (0-2)        --> X <u 254
00773   // (X+MAXUINT) >u X  --> X <u (0-MAXUINT)  --> X <u 1  --> X == 0
00774   if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
00775     return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
00776 
00777   unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
00778   ConstantInt *SMax = ConstantInt::get(X->getContext(),
00779                                        APInt::getSignedMaxValue(BitWidth));
00780 
00781   // (X+ 1) <s X       --> X >s (MAXSINT-1)          --> X == 127
00782   // (X+ 2) <s X       --> X >s (MAXSINT-2)          --> X >s 125
00783   // (X+MAXSINT) <s X  --> X >s (MAXSINT-MAXSINT)    --> X >s 0
00784   // (X+MINSINT) <s X  --> X >s (MAXSINT-MINSINT)    --> X >s -1
00785   // (X+ -2) <s X      --> X >s (MAXSINT- -2)        --> X >s 126
00786   // (X+ -1) <s X      --> X >s (MAXSINT- -1)        --> X != 127
00787   if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
00788     return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
00789 
00790   // (X+ 1) >s X       --> X <s (MAXSINT-(1-1))       --> X != 127
00791   // (X+ 2) >s X       --> X <s (MAXSINT-(2-1))       --> X <s 126
00792   // (X+MAXSINT) >s X  --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
00793   // (X+MINSINT) >s X  --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
00794   // (X+ -2) >s X      --> X <s (MAXSINT-(-2-1))      --> X <s -126
00795   // (X+ -1) >s X      --> X <s (MAXSINT-(-1-1))      --> X == -128
00796 
00797   assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
00798   Constant *C = Builder->getInt(CI->getValue()-1);
00799   return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
00800 }
00801 
00802 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
00803 /// and CmpRHS are both known to be integer constants.
00804 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
00805                                           ConstantInt *DivRHS) {
00806   ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
00807   const APInt &CmpRHSV = CmpRHS->getValue();
00808 
00809   // FIXME: If the operand types don't match the type of the divide
00810   // then don't attempt this transform. The code below doesn't have the
00811   // logic to deal with a signed divide and an unsigned compare (and
00812   // vice versa). This is because (x /s C1) <s C2  produces different
00813   // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
00814   // (x /u C1) <u C2.  Simply casting the operands and result won't
00815   // work. :(  The if statement below tests that condition and bails
00816   // if it finds it.
00817   bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
00818   if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
00819     return nullptr;
00820   if (DivRHS->isZero())
00821     return nullptr; // The ProdOV computation fails on divide by zero.
00822   if (DivIsSigned && DivRHS->isAllOnesValue())
00823     return nullptr; // The overflow computation also screws up here
00824   if (DivRHS->isOne()) {
00825     // This eliminates some funny cases with INT_MIN.
00826     ICI.setOperand(0, DivI->getOperand(0));   // X/1 == X.
00827     return &ICI;
00828   }
00829 
00830   // Compute Prod = CI * DivRHS. We are essentially solving an equation
00831   // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
00832   // C2 (CI). By solving for X we can turn this into a range check
00833   // instead of computing a divide.
00834   Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
00835 
00836   // Determine if the product overflows by seeing if the product is
00837   // not equal to the divide. Make sure we do the same kind of divide
00838   // as in the LHS instruction that we're folding.
00839   bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
00840                  ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
00841 
00842   // Get the ICmp opcode
00843   ICmpInst::Predicate Pred = ICI.getPredicate();
00844 
00845   /// If the division is known to be exact, then there is no remainder from the
00846   /// divide, so the covered range size is unit, otherwise it is the divisor.
00847   ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
00848 
00849   // Figure out the interval that is being checked.  For example, a comparison
00850   // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
00851   // Compute this interval based on the constants involved and the signedness of
00852   // the compare/divide.  This computes a half-open interval, keeping track of
00853   // whether either value in the interval overflows.  After analysis each
00854   // overflow variable is set to 0 if it's corresponding bound variable is valid
00855   // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
00856   int LoOverflow = 0, HiOverflow = 0;
00857   Constant *LoBound = nullptr, *HiBound = nullptr;
00858 
00859   if (!DivIsSigned) {  // udiv
00860     // e.g. X/5 op 3  --> [15, 20)
00861     LoBound = Prod;
00862     HiOverflow = LoOverflow = ProdOV;
00863     if (!HiOverflow) {
00864       // If this is not an exact divide, then many values in the range collapse
00865       // to the same result value.
00866       HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
00867     }
00868 
00869   } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
00870     if (CmpRHSV == 0) {       // (X / pos) op 0
00871       // Can't overflow.  e.g.  X/2 op 0 --> [-1, 2)
00872       LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
00873       HiBound = RangeSize;
00874     } else if (CmpRHSV.isStrictlyPositive()) {   // (X / pos) op pos
00875       LoBound = Prod;     // e.g.   X/5 op 3 --> [15, 20)
00876       HiOverflow = LoOverflow = ProdOV;
00877       if (!HiOverflow)
00878         HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
00879     } else {                       // (X / pos) op neg
00880       // e.g. X/5 op -3  --> [-15-4, -15+1) --> [-19, -14)
00881       HiBound = AddOne(Prod);
00882       LoOverflow = HiOverflow = ProdOV ? -1 : 0;
00883       if (!LoOverflow) {
00884         ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
00885         LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
00886       }
00887     }
00888   } else if (DivRHS->isNegative()) { // Divisor is < 0.
00889     if (DivI->isExact())
00890       RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
00891     if (CmpRHSV == 0) {       // (X / neg) op 0
00892       // e.g. X/-5 op 0  --> [-4, 5)
00893       LoBound = AddOne(RangeSize);
00894       HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
00895       if (HiBound == DivRHS) {     // -INTMIN = INTMIN
00896         HiOverflow = 1;            // [INTMIN+1, overflow)
00897         HiBound = nullptr;         // e.g. X/INTMIN = 0 --> X > INTMIN
00898       }
00899     } else if (CmpRHSV.isStrictlyPositive()) {   // (X / neg) op pos
00900       // e.g. X/-5 op 3  --> [-19, -14)
00901       HiBound = AddOne(Prod);
00902       HiOverflow = LoOverflow = ProdOV ? -1 : 0;
00903       if (!LoOverflow)
00904         LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
00905     } else {                       // (X / neg) op neg
00906       LoBound = Prod;       // e.g. X/-5 op -3  --> [15, 20)
00907       LoOverflow = HiOverflow = ProdOV;
00908       if (!HiOverflow)
00909         HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
00910     }
00911 
00912     // Dividing by a negative swaps the condition.  LT <-> GT
00913     Pred = ICmpInst::getSwappedPredicate(Pred);
00914   }
00915 
00916   Value *X = DivI->getOperand(0);
00917   switch (Pred) {
00918   default: llvm_unreachable("Unhandled icmp opcode!");
00919   case ICmpInst::ICMP_EQ:
00920     if (LoOverflow && HiOverflow)
00921       return ReplaceInstUsesWith(ICI, Builder->getFalse());
00922     if (HiOverflow)
00923       return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
00924                           ICmpInst::ICMP_UGE, X, LoBound);
00925     if (LoOverflow)
00926       return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
00927                           ICmpInst::ICMP_ULT, X, HiBound);
00928     return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
00929                                                     DivIsSigned, true));
00930   case ICmpInst::ICMP_NE:
00931     if (LoOverflow && HiOverflow)
00932       return ReplaceInstUsesWith(ICI, Builder->getTrue());
00933     if (HiOverflow)
00934       return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
00935                           ICmpInst::ICMP_ULT, X, LoBound);
00936     if (LoOverflow)
00937       return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
00938                           ICmpInst::ICMP_UGE, X, HiBound);
00939     return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
00940                                                     DivIsSigned, false));
00941   case ICmpInst::ICMP_ULT:
00942   case ICmpInst::ICMP_SLT:
00943     if (LoOverflow == +1)   // Low bound is greater than input range.
00944       return ReplaceInstUsesWith(ICI, Builder->getTrue());
00945     if (LoOverflow == -1)   // Low bound is less than input range.
00946       return ReplaceInstUsesWith(ICI, Builder->getFalse());
00947     return new ICmpInst(Pred, X, LoBound);
00948   case ICmpInst::ICMP_UGT:
00949   case ICmpInst::ICMP_SGT:
00950     if (HiOverflow == +1)       // High bound greater than input range.
00951       return ReplaceInstUsesWith(ICI, Builder->getFalse());
00952     if (HiOverflow == -1)       // High bound less than input range.
00953       return ReplaceInstUsesWith(ICI, Builder->getTrue());
00954     if (Pred == ICmpInst::ICMP_UGT)
00955       return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
00956     return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
00957   }
00958 }
00959 
00960 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
00961 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
00962                                           ConstantInt *ShAmt) {
00963   const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
00964 
00965   // Check that the shift amount is in range.  If not, don't perform
00966   // undefined shifts.  When the shift is visited it will be
00967   // simplified.
00968   uint32_t TypeBits = CmpRHSV.getBitWidth();
00969   uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
00970   if (ShAmtVal >= TypeBits || ShAmtVal == 0)
00971     return nullptr;
00972 
00973   if (!ICI.isEquality()) {
00974     // If we have an unsigned comparison and an ashr, we can't simplify this.
00975     // Similarly for signed comparisons with lshr.
00976     if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
00977       return nullptr;
00978 
00979     // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
00980     // by a power of 2.  Since we already have logic to simplify these,
00981     // transform to div and then simplify the resultant comparison.
00982     if (Shr->getOpcode() == Instruction::AShr &&
00983         (!Shr->isExact() || ShAmtVal == TypeBits - 1))
00984       return nullptr;
00985 
00986     // Revisit the shift (to delete it).
00987     Worklist.Add(Shr);
00988 
00989     Constant *DivCst =
00990       ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
00991 
00992     Value *Tmp =
00993       Shr->getOpcode() == Instruction::AShr ?
00994       Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
00995       Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
00996 
00997     ICI.setOperand(0, Tmp);
00998 
00999     // If the builder folded the binop, just return it.
01000     BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
01001     if (!TheDiv)
01002       return &ICI;
01003 
01004     // Otherwise, fold this div/compare.
01005     assert(TheDiv->getOpcode() == Instruction::SDiv ||
01006            TheDiv->getOpcode() == Instruction::UDiv);
01007 
01008     Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
01009     assert(Res && "This div/cst should have folded!");
01010     return Res;
01011   }
01012 
01013 
01014   // If we are comparing against bits always shifted out, the
01015   // comparison cannot succeed.
01016   APInt Comp = CmpRHSV << ShAmtVal;
01017   ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
01018   if (Shr->getOpcode() == Instruction::LShr)
01019     Comp = Comp.lshr(ShAmtVal);
01020   else
01021     Comp = Comp.ashr(ShAmtVal);
01022 
01023   if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
01024     bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
01025     Constant *Cst = Builder->getInt1(IsICMP_NE);
01026     return ReplaceInstUsesWith(ICI, Cst);
01027   }
01028 
01029   // Otherwise, check to see if the bits shifted out are known to be zero.
01030   // If so, we can compare against the unshifted value:
01031   //  (X & 4) >> 1 == 2  --> (X & 4) == 4.
01032   if (Shr->hasOneUse() && Shr->isExact())
01033     return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
01034 
01035   if (Shr->hasOneUse()) {
01036     // Otherwise strength reduce the shift into an and.
01037     APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
01038     Constant *Mask = Builder->getInt(Val);
01039 
01040     Value *And = Builder->CreateAnd(Shr->getOperand(0),
01041                                     Mask, Shr->getName()+".mask");
01042     return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
01043   }
01044   return nullptr;
01045 }
01046 
01047 /// FoldICmpCstShrCst - Handle "(icmp eq/ne (ashr/lshr const2, A), const1)" ->
01048 /// (icmp eq/ne A, Log2(const2/const1)) ->
01049 /// (icmp eq/ne A, Log2(const2) - Log2(const1)).
01050 Instruction *InstCombiner::FoldICmpCstShrCst(ICmpInst &I, Value *Op, Value *A,
01051                                              ConstantInt *CI1,
01052                                              ConstantInt *CI2) {
01053   assert(I.isEquality() && "Cannot fold icmp gt/lt");
01054 
01055   auto getConstant = [&I, this](bool IsTrue) {
01056     if (I.getPredicate() == I.ICMP_NE)
01057       IsTrue = !IsTrue;
01058     return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
01059   };
01060 
01061   auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
01062     if (I.getPredicate() == I.ICMP_NE)
01063       Pred = CmpInst::getInversePredicate(Pred);
01064     return new ICmpInst(Pred, LHS, RHS);
01065   };
01066 
01067   APInt AP1 = CI1->getValue();
01068   APInt AP2 = CI2->getValue();
01069 
01070   if (!AP1) {
01071     if (!AP2) {
01072       // Both Constants are 0.
01073       return getConstant(true);
01074     }
01075 
01076     if (cast<BinaryOperator>(Op)->isExact())
01077       return getConstant(false);
01078 
01079     if (AP2.isNegative()) {
01080       // MSB is set, so a lshr with a large enough 'A' would be undefined.
01081       return getConstant(false);
01082     }
01083 
01084     // 'A' must be large enough to shift out the highest set bit.
01085     return getICmp(I.ICMP_UGT, A,
01086                    ConstantInt::get(A->getType(), AP2.logBase2()));
01087   }
01088 
01089   if (!AP2) {
01090     // Shifting 0 by any value gives 0.
01091     return getConstant(false);
01092   }
01093 
01094   bool IsAShr = isa<AShrOperator>(Op);
01095   if (AP1 == AP2) {
01096     if (AP1.isAllOnesValue() && IsAShr) {
01097       // Arithmatic shift of -1 is always -1.
01098       return getConstant(true);
01099     }
01100     return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
01101   }
01102 
01103   if (IsAShr) {
01104     if (AP1.isNegative() != AP2.isNegative()) {
01105       // Arithmetic shift will never change the sign.
01106       return getConstant(false);
01107     }
01108     // Both the constants are negative, take their positive to calculate
01109     // log.
01110     if (AP1.isNegative()) {
01111       AP1 = -AP1;
01112       AP2 = -AP2;
01113     }
01114   }
01115 
01116   if (AP1.ugt(AP2)) {
01117     // Right-shifting will not increase the value.
01118     return getConstant(false);
01119   }
01120 
01121   // Get the distance between the highest bit that's set.
01122   int Shift = AP2.logBase2() - AP1.logBase2();
01123 
01124   // Use lshr here, since we've canonicalized to +ve numbers.
01125   if (AP1 == AP2.lshr(Shift))
01126     return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
01127 
01128   // Shifting const2 will never be equal to const1.
01129   return getConstant(false);
01130 }
01131 
01132 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
01133 ///
01134 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
01135                                                           Instruction *LHSI,
01136                                                           ConstantInt *RHS) {
01137   const APInt &RHSV = RHS->getValue();
01138 
01139   switch (LHSI->getOpcode()) {
01140   case Instruction::Trunc:
01141     if (ICI.isEquality() && LHSI->hasOneUse()) {
01142       // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
01143       // of the high bits truncated out of x are known.
01144       unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
01145              SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
01146       APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
01147       computeKnownBits(LHSI->getOperand(0), KnownZero, KnownOne);
01148 
01149       // If all the high bits are known, we can do this xform.
01150       if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
01151         // Pull in the high bits from known-ones set.
01152         APInt NewRHS = RHS->getValue().zext(SrcBits);
01153         NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
01154         return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
01155                             Builder->getInt(NewRHS));
01156       }
01157     }
01158     break;
01159 
01160   case Instruction::Xor:         // (icmp pred (xor X, XorCst), CI)
01161     if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
01162       // If this is a comparison that tests the signbit (X < 0) or (x > -1),
01163       // fold the xor.
01164       if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
01165           (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
01166         Value *CompareVal = LHSI->getOperand(0);
01167 
01168         // If the sign bit of the XorCst is not set, there is no change to
01169         // the operation, just stop using the Xor.
01170         if (!XorCst->isNegative()) {
01171           ICI.setOperand(0, CompareVal);
01172           Worklist.Add(LHSI);
01173           return &ICI;
01174         }
01175 
01176         // Was the old condition true if the operand is positive?
01177         bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
01178 
01179         // If so, the new one isn't.
01180         isTrueIfPositive ^= true;
01181 
01182         if (isTrueIfPositive)
01183           return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
01184                               SubOne(RHS));
01185         else
01186           return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
01187                               AddOne(RHS));
01188       }
01189 
01190       if (LHSI->hasOneUse()) {
01191         // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
01192         if (!ICI.isEquality() && XorCst->getValue().isSignBit()) {
01193           const APInt &SignBit = XorCst->getValue();
01194           ICmpInst::Predicate Pred = ICI.isSigned()
01195                                          ? ICI.getUnsignedPredicate()
01196                                          : ICI.getSignedPredicate();
01197           return new ICmpInst(Pred, LHSI->getOperand(0),
01198                               Builder->getInt(RHSV ^ SignBit));
01199         }
01200 
01201         // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
01202         if (!ICI.isEquality() && XorCst->isMaxValue(true)) {
01203           const APInt &NotSignBit = XorCst->getValue();
01204           ICmpInst::Predicate Pred = ICI.isSigned()
01205                                          ? ICI.getUnsignedPredicate()
01206                                          : ICI.getSignedPredicate();
01207           Pred = ICI.getSwappedPredicate(Pred);
01208           return new ICmpInst(Pred, LHSI->getOperand(0),
01209                               Builder->getInt(RHSV ^ NotSignBit));
01210         }
01211       }
01212 
01213       // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
01214       //   iff -C is a power of 2
01215       if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
01216           XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
01217         return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst);
01218 
01219       // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
01220       //   iff -C is a power of 2
01221       if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
01222           XorCst->getValue() == -RHSV && RHSV.isPowerOf2())
01223         return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst);
01224     }
01225     break;
01226   case Instruction::And:         // (icmp pred (and X, AndCst), RHS)
01227     if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
01228         LHSI->getOperand(0)->hasOneUse()) {
01229       ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1));
01230 
01231       // If the LHS is an AND of a truncating cast, we can widen the
01232       // and/compare to be the input width without changing the value
01233       // produced, eliminating a cast.
01234       if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
01235         // We can do this transformation if either the AND constant does not
01236         // have its sign bit set or if it is an equality comparison.
01237         // Extending a relational comparison when we're checking the sign
01238         // bit would not work.
01239         if (ICI.isEquality() ||
01240             (!AndCst->isNegative() && RHSV.isNonNegative())) {
01241           Value *NewAnd =
01242             Builder->CreateAnd(Cast->getOperand(0),
01243                                ConstantExpr::getZExt(AndCst, Cast->getSrcTy()));
01244           NewAnd->takeName(LHSI);
01245           return new ICmpInst(ICI.getPredicate(), NewAnd,
01246                               ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
01247         }
01248       }
01249 
01250       // If the LHS is an AND of a zext, and we have an equality compare, we can
01251       // shrink the and/compare to the smaller type, eliminating the cast.
01252       if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
01253         IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
01254         // Make sure we don't compare the upper bits, SimplifyDemandedBits
01255         // should fold the icmp to true/false in that case.
01256         if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
01257           Value *NewAnd =
01258             Builder->CreateAnd(Cast->getOperand(0),
01259                                ConstantExpr::getTrunc(AndCst, Ty));
01260           NewAnd->takeName(LHSI);
01261           return new ICmpInst(ICI.getPredicate(), NewAnd,
01262                               ConstantExpr::getTrunc(RHS, Ty));
01263         }
01264       }
01265 
01266       // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
01267       // could exist), turn it into (X & (C2 << C1)) != (C3 << C1).  This
01268       // happens a LOT in code produced by the C front-end, for bitfield
01269       // access.
01270       BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
01271       if (Shift && !Shift->isShift())
01272         Shift = nullptr;
01273 
01274       ConstantInt *ShAmt;
01275       ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : nullptr;
01276 
01277       // This seemingly simple opportunity to fold away a shift turns out to
01278       // be rather complicated. See PR17827
01279       // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details.
01280       if (ShAmt) {
01281         bool CanFold = false;
01282         unsigned ShiftOpcode = Shift->getOpcode();
01283         if (ShiftOpcode == Instruction::AShr) {
01284           // There may be some constraints that make this possible,
01285           // but nothing simple has been discovered yet.
01286           CanFold = false;
01287         } else if (ShiftOpcode == Instruction::Shl) {
01288           // For a left shift, we can fold if the comparison is not signed.
01289           // We can also fold a signed comparison if the mask value and
01290           // comparison value are not negative. These constraints may not be
01291           // obvious, but we can prove that they are correct using an SMT
01292           // solver.
01293           if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative()))
01294             CanFold = true;
01295         } else if (ShiftOpcode == Instruction::LShr) {
01296           // For a logical right shift, we can fold if the comparison is not
01297           // signed. We can also fold a signed comparison if the shifted mask
01298           // value and the shifted comparison value are not negative.
01299           // These constraints may not be obvious, but we can prove that they
01300           // are correct using an SMT solver.
01301           if (!ICI.isSigned())
01302             CanFold = true;
01303           else {
01304             ConstantInt *ShiftedAndCst =
01305               cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt));
01306             ConstantInt *ShiftedRHSCst =
01307               cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt));
01308             
01309             if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative())
01310               CanFold = true;
01311           }
01312         }
01313 
01314         if (CanFold) {
01315           Constant *NewCst;
01316           if (ShiftOpcode == Instruction::Shl)
01317             NewCst = ConstantExpr::getLShr(RHS, ShAmt);
01318           else
01319             NewCst = ConstantExpr::getShl(RHS, ShAmt);
01320 
01321           // Check to see if we are shifting out any of the bits being
01322           // compared.
01323           if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) {
01324             // If we shifted bits out, the fold is not going to work out.
01325             // As a special case, check to see if this means that the
01326             // result is always true or false now.
01327             if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
01328               return ReplaceInstUsesWith(ICI, Builder->getFalse());
01329             if (ICI.getPredicate() == ICmpInst::ICMP_NE)
01330               return ReplaceInstUsesWith(ICI, Builder->getTrue());
01331           } else {
01332             ICI.setOperand(1, NewCst);
01333             Constant *NewAndCst;
01334             if (ShiftOpcode == Instruction::Shl)
01335               NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt);
01336             else
01337               NewAndCst = ConstantExpr::getShl(AndCst, ShAmt);
01338             LHSI->setOperand(1, NewAndCst);
01339             LHSI->setOperand(0, Shift->getOperand(0));
01340             Worklist.Add(Shift); // Shift is dead.
01341             return &ICI;
01342           }
01343         }
01344       }
01345 
01346       // Turn ((X >> Y) & C) == 0  into  (X & (C << Y)) == 0.  The later is
01347       // preferable because it allows the C<<Y expression to be hoisted out
01348       // of a loop if Y is invariant and X is not.
01349       if (Shift && Shift->hasOneUse() && RHSV == 0 &&
01350           ICI.isEquality() && !Shift->isArithmeticShift() &&
01351           !isa<Constant>(Shift->getOperand(0))) {
01352         // Compute C << Y.
01353         Value *NS;
01354         if (Shift->getOpcode() == Instruction::LShr) {
01355           NS = Builder->CreateShl(AndCst, Shift->getOperand(1));
01356         } else {
01357           // Insert a logical shift.
01358           NS = Builder->CreateLShr(AndCst, Shift->getOperand(1));
01359         }
01360 
01361         // Compute X & (C << Y).
01362         Value *NewAnd =
01363           Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
01364 
01365         ICI.setOperand(0, NewAnd);
01366         return &ICI;
01367       }
01368 
01369       // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any
01370       // bit set in (X & AndCst) will produce a result greater than RHSV.
01371       if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
01372         unsigned NTZ = AndCst->getValue().countTrailingZeros();
01373         if ((NTZ < AndCst->getBitWidth()) &&
01374             APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV))
01375           return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
01376                               Constant::getNullValue(RHS->getType()));
01377       }
01378     }
01379 
01380     // Try to optimize things like "A[i]&42 == 0" to index computations.
01381     if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
01382       if (GetElementPtrInst *GEP =
01383           dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
01384         if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
01385           if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
01386               !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
01387             ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
01388             if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
01389               return Res;
01390           }
01391     }
01392 
01393     // X & -C == -C -> X >  u ~C
01394     // X & -C != -C -> X <= u ~C
01395     //   iff C is a power of 2
01396     if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
01397       return new ICmpInst(
01398           ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
01399                                                   : ICmpInst::ICMP_ULE,
01400           LHSI->getOperand(0), SubOne(RHS));
01401     break;
01402 
01403   case Instruction::Or: {
01404     if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
01405       break;
01406     Value *P, *Q;
01407     if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
01408       // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
01409       // -> and (icmp eq P, null), (icmp eq Q, null).
01410       Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
01411                                         Constant::getNullValue(P->getType()));
01412       Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
01413                                         Constant::getNullValue(Q->getType()));
01414       Instruction *Op;
01415       if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
01416         Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
01417       else
01418         Op = BinaryOperator::CreateOr(ICIP, ICIQ);
01419       return Op;
01420     }
01421     break;
01422   }
01423 
01424   case Instruction::Mul: {       // (icmp pred (mul X, Val), CI)
01425     ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
01426     if (!Val) break;
01427 
01428     // If this is a signed comparison to 0 and the mul is sign preserving,
01429     // use the mul LHS operand instead.
01430     ICmpInst::Predicate pred = ICI.getPredicate();
01431     if (isSignTest(pred, RHS) && !Val->isZero() &&
01432         cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
01433       return new ICmpInst(Val->isNegative() ?
01434                           ICmpInst::getSwappedPredicate(pred) : pred,
01435                           LHSI->getOperand(0),
01436                           Constant::getNullValue(RHS->getType()));
01437 
01438     break;
01439   }
01440 
01441   case Instruction::Shl: {       // (icmp pred (shl X, ShAmt), CI)
01442     uint32_t TypeBits = RHSV.getBitWidth();
01443     ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
01444     if (!ShAmt) {
01445       Value *X;
01446       // (1 << X) pred P2 -> X pred Log2(P2)
01447       if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
01448         bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
01449         ICmpInst::Predicate Pred = ICI.getPredicate();
01450         if (ICI.isUnsigned()) {
01451           if (!RHSVIsPowerOf2) {
01452             // (1 << X) <  30 -> X <= 4
01453             // (1 << X) <= 30 -> X <= 4
01454             // (1 << X) >= 30 -> X >  4
01455             // (1 << X) >  30 -> X >  4
01456             if (Pred == ICmpInst::ICMP_ULT)
01457               Pred = ICmpInst::ICMP_ULE;
01458             else if (Pred == ICmpInst::ICMP_UGE)
01459               Pred = ICmpInst::ICMP_UGT;
01460           }
01461           unsigned RHSLog2 = RHSV.logBase2();
01462 
01463           // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
01464           // (1 << X) >  2147483648 -> X >  31 -> false
01465           // (1 << X) <= 2147483648 -> X <= 31 -> true
01466           // (1 << X) <  2147483648 -> X <  31 -> X != 31
01467           if (RHSLog2 == TypeBits-1) {
01468             if (Pred == ICmpInst::ICMP_UGE)
01469               Pred = ICmpInst::ICMP_EQ;
01470             else if (Pred == ICmpInst::ICMP_UGT)
01471               return ReplaceInstUsesWith(ICI, Builder->getFalse());
01472             else if (Pred == ICmpInst::ICMP_ULE)
01473               return ReplaceInstUsesWith(ICI, Builder->getTrue());
01474             else if (Pred == ICmpInst::ICMP_ULT)
01475               Pred = ICmpInst::ICMP_NE;
01476           }
01477 
01478           return new ICmpInst(Pred, X,
01479                               ConstantInt::get(RHS->getType(), RHSLog2));
01480         } else if (ICI.isSigned()) {
01481           if (RHSV.isAllOnesValue()) {
01482             // (1 << X) <= -1 -> X == 31
01483             if (Pred == ICmpInst::ICMP_SLE)
01484               return new ICmpInst(ICmpInst::ICMP_EQ, X,
01485                                   ConstantInt::get(RHS->getType(), TypeBits-1));
01486 
01487             // (1 << X) >  -1 -> X != 31
01488             if (Pred == ICmpInst::ICMP_SGT)
01489               return new ICmpInst(ICmpInst::ICMP_NE, X,
01490                                   ConstantInt::get(RHS->getType(), TypeBits-1));
01491           } else if (!RHSV) {
01492             // (1 << X) <  0 -> X == 31
01493             // (1 << X) <= 0 -> X == 31
01494             if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
01495               return new ICmpInst(ICmpInst::ICMP_EQ, X,
01496                                   ConstantInt::get(RHS->getType(), TypeBits-1));
01497 
01498             // (1 << X) >= 0 -> X != 31
01499             // (1 << X) >  0 -> X != 31
01500             if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
01501               return new ICmpInst(ICmpInst::ICMP_NE, X,
01502                                   ConstantInt::get(RHS->getType(), TypeBits-1));
01503           }
01504         } else if (ICI.isEquality()) {
01505           if (RHSVIsPowerOf2)
01506             return new ICmpInst(
01507                 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
01508 
01509           return ReplaceInstUsesWith(
01510               ICI, Pred == ICmpInst::ICMP_EQ ? Builder->getFalse()
01511                                              : Builder->getTrue());
01512         }
01513       }
01514       break;
01515     }
01516 
01517     // Check that the shift amount is in range.  If not, don't perform
01518     // undefined shifts.  When the shift is visited it will be
01519     // simplified.
01520     if (ShAmt->uge(TypeBits))
01521       break;
01522 
01523     if (ICI.isEquality()) {
01524       // If we are comparing against bits always shifted out, the
01525       // comparison cannot succeed.
01526       Constant *Comp =
01527         ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
01528                                                                  ShAmt);
01529       if (Comp != RHS) {// Comparing against a bit that we know is zero.
01530         bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
01531         Constant *Cst = Builder->getInt1(IsICMP_NE);
01532         return ReplaceInstUsesWith(ICI, Cst);
01533       }
01534 
01535       // If the shift is NUW, then it is just shifting out zeros, no need for an
01536       // AND.
01537       if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
01538         return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
01539                             ConstantExpr::getLShr(RHS, ShAmt));
01540 
01541       // If the shift is NSW and we compare to 0, then it is just shifting out
01542       // sign bits, no need for an AND either.
01543       if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
01544         return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
01545                             ConstantExpr::getLShr(RHS, ShAmt));
01546 
01547       if (LHSI->hasOneUse()) {
01548         // Otherwise strength reduce the shift into an and.
01549         uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
01550         Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
01551                                                           TypeBits - ShAmtVal));
01552 
01553         Value *And =
01554           Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
01555         return new ICmpInst(ICI.getPredicate(), And,
01556                             ConstantExpr::getLShr(RHS, ShAmt));
01557       }
01558     }
01559 
01560     // If this is a signed comparison to 0 and the shift is sign preserving,
01561     // use the shift LHS operand instead.
01562     ICmpInst::Predicate pred = ICI.getPredicate();
01563     if (isSignTest(pred, RHS) &&
01564         cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
01565       return new ICmpInst(pred,
01566                           LHSI->getOperand(0),
01567                           Constant::getNullValue(RHS->getType()));
01568 
01569     // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
01570     bool TrueIfSigned = false;
01571     if (LHSI->hasOneUse() &&
01572         isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
01573       // (X << 31) <s 0  --> (X&1) != 0
01574       Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
01575                                         APInt::getOneBitSet(TypeBits,
01576                                             TypeBits-ShAmt->getZExtValue()-1));
01577       Value *And =
01578         Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
01579       return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
01580                           And, Constant::getNullValue(And->getType()));
01581     }
01582 
01583     // Transform (icmp pred iM (shl iM %v, N), CI)
01584     // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
01585     // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
01586     // This enables to get rid of the shift in favor of a trunc which can be
01587     // free on the target. It has the additional benefit of comparing to a
01588     // smaller constant, which will be target friendly.
01589     unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
01590     if (LHSI->hasOneUse() &&
01591         Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
01592       Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
01593       Constant *NCI = ConstantExpr::getTrunc(
01594                         ConstantExpr::getAShr(RHS,
01595                           ConstantInt::get(RHS->getType(), Amt)),
01596                         NTy);
01597       return new ICmpInst(ICI.getPredicate(),
01598                           Builder->CreateTrunc(LHSI->getOperand(0), NTy),
01599                           NCI);
01600     }
01601 
01602     break;
01603   }
01604 
01605   case Instruction::LShr:         // (icmp pred (shr X, ShAmt), CI)
01606   case Instruction::AShr: {
01607     // Handle equality comparisons of shift-by-constant.
01608     BinaryOperator *BO = cast<BinaryOperator>(LHSI);
01609     if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
01610       if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
01611         return Res;
01612     }
01613 
01614     // Handle exact shr's.
01615     if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
01616       if (RHSV.isMinValue())
01617         return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
01618     }
01619     break;
01620   }
01621 
01622   case Instruction::SDiv:
01623   case Instruction::UDiv:
01624     // Fold: icmp pred ([us]div X, C1), C2 -> range test
01625     // Fold this div into the comparison, producing a range check.
01626     // Determine, based on the divide type, what the range is being
01627     // checked.  If there is an overflow on the low or high side, remember
01628     // it, otherwise compute the range [low, hi) bounding the new value.
01629     // See: InsertRangeTest above for the kinds of replacements possible.
01630     if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
01631       if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
01632                                           DivRHS))
01633         return R;
01634     break;
01635 
01636   case Instruction::Sub: {
01637     ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
01638     if (!LHSC) break;
01639     const APInt &LHSV = LHSC->getValue();
01640 
01641     // C1-X <u C2 -> (X|(C2-1)) == C1
01642     //   iff C1 & (C2-1) == C2-1
01643     //       C2 is a power of 2
01644     if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
01645         RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
01646       return new ICmpInst(ICmpInst::ICMP_EQ,
01647                           Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
01648                           LHSC);
01649 
01650     // C1-X >u C2 -> (X|C2) != C1
01651     //   iff C1 & C2 == C2
01652     //       C2+1 is a power of 2
01653     if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
01654         (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
01655       return new ICmpInst(ICmpInst::ICMP_NE,
01656                           Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
01657     break;
01658   }
01659 
01660   case Instruction::Add:
01661     // Fold: icmp pred (add X, C1), C2
01662     if (!ICI.isEquality()) {
01663       ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
01664       if (!LHSC) break;
01665       const APInt &LHSV = LHSC->getValue();
01666 
01667       ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
01668                             .subtract(LHSV);
01669 
01670       if (ICI.isSigned()) {
01671         if (CR.getLower().isSignBit()) {
01672           return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
01673                               Builder->getInt(CR.getUpper()));
01674         } else if (CR.getUpper().isSignBit()) {
01675           return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
01676                               Builder->getInt(CR.getLower()));
01677         }
01678       } else {
01679         if (CR.getLower().isMinValue()) {
01680           return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
01681                               Builder->getInt(CR.getUpper()));
01682         } else if (CR.getUpper().isMinValue()) {
01683           return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
01684                               Builder->getInt(CR.getLower()));
01685         }
01686       }
01687 
01688       // X-C1 <u C2 -> (X & -C2) == C1
01689       //   iff C1 & (C2-1) == 0
01690       //       C2 is a power of 2
01691       if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
01692           RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
01693         return new ICmpInst(ICmpInst::ICMP_EQ,
01694                             Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
01695                             ConstantExpr::getNeg(LHSC));
01696 
01697       // X-C1 >u C2 -> (X & ~C2) != C1
01698       //   iff C1 & C2 == 0
01699       //       C2+1 is a power of 2
01700       if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
01701           (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
01702         return new ICmpInst(ICmpInst::ICMP_NE,
01703                             Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
01704                             ConstantExpr::getNeg(LHSC));
01705     }
01706     break;
01707   }
01708 
01709   // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
01710   if (ICI.isEquality()) {
01711     bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
01712 
01713     // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
01714     // the second operand is a constant, simplify a bit.
01715     if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
01716       switch (BO->getOpcode()) {
01717       case Instruction::SRem:
01718         // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
01719         if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
01720           const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
01721           if (V.sgt(1) && V.isPowerOf2()) {
01722             Value *NewRem =
01723               Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
01724                                   BO->getName());
01725             return new ICmpInst(ICI.getPredicate(), NewRem,
01726                                 Constant::getNullValue(BO->getType()));
01727           }
01728         }
01729         break;
01730       case Instruction::Add:
01731         // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
01732         if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
01733           if (BO->hasOneUse())
01734             return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01735                                 ConstantExpr::getSub(RHS, BOp1C));
01736         } else if (RHSV == 0) {
01737           // Replace ((add A, B) != 0) with (A != -B) if A or B is
01738           // efficiently invertible, or if the add has just this one use.
01739           Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
01740 
01741           if (Value *NegVal = dyn_castNegVal(BOp1))
01742             return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
01743           if (Value *NegVal = dyn_castNegVal(BOp0))
01744             return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
01745           if (BO->hasOneUse()) {
01746             Value *Neg = Builder->CreateNeg(BOp1);
01747             Neg->takeName(BO);
01748             return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
01749           }
01750         }
01751         break;
01752       case Instruction::Xor:
01753         // For the xor case, we can xor two constants together, eliminating
01754         // the explicit xor.
01755         if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
01756           return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01757                               ConstantExpr::getXor(RHS, BOC));
01758         } else if (RHSV == 0) {
01759           // Replace ((xor A, B) != 0) with (A != B)
01760           return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01761                               BO->getOperand(1));
01762         }
01763         break;
01764       case Instruction::Sub:
01765         // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
01766         if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
01767           if (BO->hasOneUse())
01768             return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
01769                                 ConstantExpr::getSub(BOp0C, RHS));
01770         } else if (RHSV == 0) {
01771           // Replace ((sub A, B) != 0) with (A != B)
01772           return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01773                               BO->getOperand(1));
01774         }
01775         break;
01776       case Instruction::Or:
01777         // If bits are being or'd in that are not present in the constant we
01778         // are comparing against, then the comparison could never succeed!
01779         if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
01780           Constant *NotCI = ConstantExpr::getNot(RHS);
01781           if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
01782             return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
01783         }
01784         break;
01785 
01786       case Instruction::And:
01787         if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
01788           // If bits are being compared against that are and'd out, then the
01789           // comparison can never succeed!
01790           if ((RHSV & ~BOC->getValue()) != 0)
01791             return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
01792 
01793           // If we have ((X & C) == C), turn it into ((X & C) != 0).
01794           if (RHS == BOC && RHSV.isPowerOf2())
01795             return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
01796                                 ICmpInst::ICMP_NE, LHSI,
01797                                 Constant::getNullValue(RHS->getType()));
01798 
01799           // Don't perform the following transforms if the AND has multiple uses
01800           if (!BO->hasOneUse())
01801             break;
01802 
01803           // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
01804           if (BOC->getValue().isSignBit()) {
01805             Value *X = BO->getOperand(0);
01806             Constant *Zero = Constant::getNullValue(X->getType());
01807             ICmpInst::Predicate pred = isICMP_NE ?
01808               ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
01809             return new ICmpInst(pred, X, Zero);
01810           }
01811 
01812           // ((X & ~7) == 0) --> X < 8
01813           if (RHSV == 0 && isHighOnes(BOC)) {
01814             Value *X = BO->getOperand(0);
01815             Constant *NegX = ConstantExpr::getNeg(BOC);
01816             ICmpInst::Predicate pred = isICMP_NE ?
01817               ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
01818             return new ICmpInst(pred, X, NegX);
01819           }
01820         }
01821         break;
01822       case Instruction::Mul:
01823         if (RHSV == 0 && BO->hasNoSignedWrap()) {
01824           if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
01825             // The trivial case (mul X, 0) is handled by InstSimplify
01826             // General case : (mul X, C) != 0 iff X != 0
01827             //                (mul X, C) == 0 iff X == 0
01828             if (!BOC->isZero())
01829               return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
01830                                   Constant::getNullValue(RHS->getType()));
01831           }
01832         }
01833         break;
01834       default: break;
01835       }
01836     } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
01837       // Handle icmp {eq|ne} <intrinsic>, intcst.
01838       switch (II->getIntrinsicID()) {
01839       case Intrinsic::bswap:
01840         Worklist.Add(II);
01841         ICI.setOperand(0, II->getArgOperand(0));
01842         ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
01843         return &ICI;
01844       case Intrinsic::ctlz:
01845       case Intrinsic::cttz:
01846         // ctz(A) == bitwidth(a)  ->  A == 0 and likewise for !=
01847         if (RHSV == RHS->getType()->getBitWidth()) {
01848           Worklist.Add(II);
01849           ICI.setOperand(0, II->getArgOperand(0));
01850           ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
01851           return &ICI;
01852         }
01853         break;
01854       case Intrinsic::ctpop:
01855         // popcount(A) == 0  ->  A == 0 and likewise for !=
01856         if (RHS->isZero()) {
01857           Worklist.Add(II);
01858           ICI.setOperand(0, II->getArgOperand(0));
01859           ICI.setOperand(1, RHS);
01860           return &ICI;
01861         }
01862         break;
01863       default:
01864         break;
01865       }
01866     }
01867   }
01868   return nullptr;
01869 }
01870 
01871 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
01872 /// We only handle extending casts so far.
01873 ///
01874 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
01875   const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
01876   Value *LHSCIOp        = LHSCI->getOperand(0);
01877   Type *SrcTy     = LHSCIOp->getType();
01878   Type *DestTy    = LHSCI->getType();
01879   Value *RHSCIOp;
01880 
01881   // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
01882   // integer type is the same size as the pointer type.
01883   if (DL && LHSCI->getOpcode() == Instruction::PtrToInt &&
01884       DL->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
01885     Value *RHSOp = nullptr;
01886     if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
01887       RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
01888     } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
01889       RHSOp = RHSC->getOperand(0);
01890       // If the pointer types don't match, insert a bitcast.
01891       if (LHSCIOp->getType() != RHSOp->getType())
01892         RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
01893     }
01894 
01895     if (RHSOp)
01896       return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
01897   }
01898 
01899   // The code below only handles extension cast instructions, so far.
01900   // Enforce this.
01901   if (LHSCI->getOpcode() != Instruction::ZExt &&
01902       LHSCI->getOpcode() != Instruction::SExt)
01903     return nullptr;
01904 
01905   bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
01906   bool isSignedCmp = ICI.isSigned();
01907 
01908   if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
01909     // Not an extension from the same type?
01910     RHSCIOp = CI->getOperand(0);
01911     if (RHSCIOp->getType() != LHSCIOp->getType())
01912       return nullptr;
01913 
01914     // If the signedness of the two casts doesn't agree (i.e. one is a sext
01915     // and the other is a zext), then we can't handle this.
01916     if (CI->getOpcode() != LHSCI->getOpcode())
01917       return nullptr;
01918 
01919     // Deal with equality cases early.
01920     if (ICI.isEquality())
01921       return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
01922 
01923     // A signed comparison of sign extended values simplifies into a
01924     // signed comparison.
01925     if (isSignedCmp && isSignedExt)
01926       return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
01927 
01928     // The other three cases all fold into an unsigned comparison.
01929     return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
01930   }
01931 
01932   // If we aren't dealing with a constant on the RHS, exit early
01933   ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
01934   if (!CI)
01935     return nullptr;
01936 
01937   // Compute the constant that would happen if we truncated to SrcTy then
01938   // reextended to DestTy.
01939   Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
01940   Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
01941                                                 Res1, DestTy);
01942 
01943   // If the re-extended constant didn't change...
01944   if (Res2 == CI) {
01945     // Deal with equality cases early.
01946     if (ICI.isEquality())
01947       return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
01948 
01949     // A signed comparison of sign extended values simplifies into a
01950     // signed comparison.
01951     if (isSignedExt && isSignedCmp)
01952       return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
01953 
01954     // The other three cases all fold into an unsigned comparison.
01955     return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
01956   }
01957 
01958   // The re-extended constant changed so the constant cannot be represented
01959   // in the shorter type. Consequently, we cannot emit a simple comparison.
01960   // All the cases that fold to true or false will have already been handled
01961   // by SimplifyICmpInst, so only deal with the tricky case.
01962 
01963   if (isSignedCmp || !isSignedExt)
01964     return nullptr;
01965 
01966   // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
01967   // should have been folded away previously and not enter in here.
01968 
01969   // We're performing an unsigned comp with a sign extended value.
01970   // This is true if the input is >= 0. [aka >s -1]
01971   Constant *NegOne = Constant::getAllOnesValue(SrcTy);
01972   Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
01973 
01974   // Finally, return the value computed.
01975   if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
01976     return ReplaceInstUsesWith(ICI, Result);
01977 
01978   assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
01979   return BinaryOperator::CreateNot(Result);
01980 }
01981 
01982 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
01983 ///   I = icmp ugt (add (add A, B), CI2), CI1
01984 /// If this is of the form:
01985 ///   sum = a + b
01986 ///   if (sum+128 >u 255)
01987 /// Then replace it with llvm.sadd.with.overflow.i8.
01988 ///
01989 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
01990                                           ConstantInt *CI2, ConstantInt *CI1,
01991                                           InstCombiner &IC) {
01992   // The transformation we're trying to do here is to transform this into an
01993   // llvm.sadd.with.overflow.  To do this, we have to replace the original add
01994   // with a narrower add, and discard the add-with-constant that is part of the
01995   // range check (if we can't eliminate it, this isn't profitable).
01996 
01997   // In order to eliminate the add-with-constant, the compare can be its only
01998   // use.
01999   Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
02000   if (!AddWithCst->hasOneUse()) return nullptr;
02001 
02002   // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
02003   if (!CI2->getValue().isPowerOf2()) return nullptr;
02004   unsigned NewWidth = CI2->getValue().countTrailingZeros();
02005   if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return nullptr;
02006 
02007   // The width of the new add formed is 1 more than the bias.
02008   ++NewWidth;
02009 
02010   // Check to see that CI1 is an all-ones value with NewWidth bits.
02011   if (CI1->getBitWidth() == NewWidth ||
02012       CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
02013     return nullptr;
02014 
02015   // This is only really a signed overflow check if the inputs have been
02016   // sign-extended; check for that condition. For example, if CI2 is 2^31 and
02017   // the operands of the add are 64 bits wide, we need at least 33 sign bits.
02018   unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
02019   if (IC.ComputeNumSignBits(A) < NeededSignBits ||
02020       IC.ComputeNumSignBits(B) < NeededSignBits)
02021     return nullptr;
02022 
02023   // In order to replace the original add with a narrower
02024   // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
02025   // and truncates that discard the high bits of the add.  Verify that this is
02026   // the case.
02027   Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
02028   for (User *U : OrigAdd->users()) {
02029     if (U == AddWithCst) continue;
02030 
02031     // Only accept truncates for now.  We would really like a nice recursive
02032     // predicate like SimplifyDemandedBits, but which goes downwards the use-def
02033     // chain to see which bits of a value are actually demanded.  If the
02034     // original add had another add which was then immediately truncated, we
02035     // could still do the transformation.
02036     TruncInst *TI = dyn_cast<TruncInst>(U);
02037     if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
02038       return nullptr;
02039   }
02040 
02041   // If the pattern matches, truncate the inputs to the narrower type and
02042   // use the sadd_with_overflow intrinsic to efficiently compute both the
02043   // result and the overflow bit.
02044   Module *M = I.getParent()->getParent()->getParent();
02045 
02046   Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
02047   Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
02048                                        NewType);
02049 
02050   InstCombiner::BuilderTy *Builder = IC.Builder;
02051 
02052   // Put the new code above the original add, in case there are any uses of the
02053   // add between the add and the compare.
02054   Builder->SetInsertPoint(OrigAdd);
02055 
02056   Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
02057   Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
02058   CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
02059   Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
02060   Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
02061 
02062   // The inner add was the result of the narrow add, zero extended to the
02063   // wider type.  Replace it with the result computed by the intrinsic.
02064   IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
02065 
02066   // The original icmp gets replaced with the overflow value.
02067   return ExtractValueInst::Create(Call, 1, "sadd.overflow");
02068 }
02069 
02070 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
02071                                      InstCombiner &IC) {
02072   // Don't bother doing this transformation for pointers, don't do it for
02073   // vectors.
02074   if (!isa<IntegerType>(OrigAddV->getType())) return nullptr;
02075 
02076   // If the add is a constant expr, then we don't bother transforming it.
02077   Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
02078   if (!OrigAdd) return nullptr;
02079 
02080   Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
02081 
02082   // Put the new code above the original add, in case there are any uses of the
02083   // add between the add and the compare.
02084   InstCombiner::BuilderTy *Builder = IC.Builder;
02085   Builder->SetInsertPoint(OrigAdd);
02086 
02087   Module *M = I.getParent()->getParent()->getParent();
02088   Type *Ty = LHS->getType();
02089   Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
02090   CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
02091   Value *Add = Builder->CreateExtractValue(Call, 0);
02092 
02093   IC.ReplaceInstUsesWith(*OrigAdd, Add);
02094 
02095   // The original icmp gets replaced with the overflow value.
02096   return ExtractValueInst::Create(Call, 1, "uadd.overflow");
02097 }
02098 
02099 /// \brief Recognize and process idiom involving test for multiplication
02100 /// overflow.
02101 ///
02102 /// The caller has matched a pattern of the form:
02103 ///   I = cmp u (mul(zext A, zext B), V
02104 /// The function checks if this is a test for overflow and if so replaces
02105 /// multiplication with call to 'mul.with.overflow' intrinsic.
02106 ///
02107 /// \param I Compare instruction.
02108 /// \param MulVal Result of 'mult' instruction.  It is one of the arguments of
02109 ///               the compare instruction.  Must be of integer type.
02110 /// \param OtherVal The other argument of compare instruction.
02111 /// \returns Instruction which must replace the compare instruction, NULL if no
02112 ///          replacement required.
02113 static Instruction *ProcessUMulZExtIdiom(ICmpInst &I, Value *MulVal,
02114                                          Value *OtherVal, InstCombiner &IC) {
02115   // Don't bother doing this transformation for pointers, don't do it for
02116   // vectors.
02117   if (!isa<IntegerType>(MulVal->getType()))
02118     return nullptr;
02119 
02120   assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
02121   assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
02122   Instruction *MulInstr = cast<Instruction>(MulVal);
02123   assert(MulInstr->getOpcode() == Instruction::Mul);
02124 
02125   Instruction *LHS = cast<Instruction>(MulInstr->getOperand(0)),
02126               *RHS = cast<Instruction>(MulInstr->getOperand(1));
02127   assert(LHS->getOpcode() == Instruction::ZExt);
02128   assert(RHS->getOpcode() == Instruction::ZExt);
02129   Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
02130 
02131   // Calculate type and width of the result produced by mul.with.overflow.
02132   Type *TyA = A->getType(), *TyB = B->getType();
02133   unsigned WidthA = TyA->getPrimitiveSizeInBits(),
02134            WidthB = TyB->getPrimitiveSizeInBits();
02135   unsigned MulWidth;
02136   Type *MulType;
02137   if (WidthB > WidthA) {
02138     MulWidth = WidthB;
02139     MulType = TyB;
02140   } else {
02141     MulWidth = WidthA;
02142     MulType = TyA;
02143   }
02144 
02145   // In order to replace the original mul with a narrower mul.with.overflow,
02146   // all uses must ignore upper bits of the product.  The number of used low
02147   // bits must be not greater than the width of mul.with.overflow.
02148   if (MulVal->hasNUsesOrMore(2))
02149     for (User *U : MulVal->users()) {
02150       if (U == &I)
02151         continue;
02152       if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
02153         // Check if truncation ignores bits above MulWidth.
02154         unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
02155         if (TruncWidth > MulWidth)
02156           return nullptr;
02157       } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
02158         // Check if AND ignores bits above MulWidth.
02159         if (BO->getOpcode() != Instruction::And)
02160           return nullptr;
02161         if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
02162           const APInt &CVal = CI->getValue();
02163           if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
02164             return nullptr;
02165         }
02166       } else {
02167         // Other uses prohibit this transformation.
02168         return nullptr;
02169       }
02170     }
02171 
02172   // Recognize patterns
02173   switch (I.getPredicate()) {
02174   case ICmpInst::ICMP_EQ:
02175   case ICmpInst::ICMP_NE:
02176     // Recognize pattern:
02177     //   mulval = mul(zext A, zext B)
02178     //   cmp eq/neq mulval, zext trunc mulval
02179     if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
02180       if (Zext->hasOneUse()) {
02181         Value *ZextArg = Zext->getOperand(0);
02182         if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
02183           if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
02184             break; //Recognized
02185       }
02186 
02187     // Recognize pattern:
02188     //   mulval = mul(zext A, zext B)
02189     //   cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
02190     ConstantInt *CI;
02191     Value *ValToMask;
02192     if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
02193       if (ValToMask != MulVal)
02194         return nullptr;
02195       const APInt &CVal = CI->getValue() + 1;
02196       if (CVal.isPowerOf2()) {
02197         unsigned MaskWidth = CVal.logBase2();
02198         if (MaskWidth == MulWidth)
02199           break; // Recognized
02200       }
02201     }
02202     return nullptr;
02203 
02204   case ICmpInst::ICMP_UGT:
02205     // Recognize pattern:
02206     //   mulval = mul(zext A, zext B)
02207     //   cmp ugt mulval, max
02208     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
02209       APInt MaxVal = APInt::getMaxValue(MulWidth);
02210       MaxVal = MaxVal.zext(CI->getBitWidth());
02211       if (MaxVal.eq(CI->getValue()))
02212         break; // Recognized
02213     }
02214     return nullptr;
02215 
02216   case ICmpInst::ICMP_UGE:
02217     // Recognize pattern:
02218     //   mulval = mul(zext A, zext B)
02219     //   cmp uge mulval, max+1
02220     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
02221       APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
02222       if (MaxVal.eq(CI->getValue()))
02223         break; // Recognized
02224     }
02225     return nullptr;
02226 
02227   case ICmpInst::ICMP_ULE:
02228     // Recognize pattern:
02229     //   mulval = mul(zext A, zext B)
02230     //   cmp ule mulval, max
02231     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
02232       APInt MaxVal = APInt::getMaxValue(MulWidth);
02233       MaxVal = MaxVal.zext(CI->getBitWidth());
02234       if (MaxVal.eq(CI->getValue()))
02235         break; // Recognized
02236     }
02237     return nullptr;
02238 
02239   case ICmpInst::ICMP_ULT:
02240     // Recognize pattern:
02241     //   mulval = mul(zext A, zext B)
02242     //   cmp ule mulval, max + 1
02243     if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
02244       APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
02245       if (MaxVal.eq(CI->getValue()))
02246         break; // Recognized
02247     }
02248     return nullptr;
02249 
02250   default:
02251     return nullptr;
02252   }
02253 
02254   InstCombiner::BuilderTy *Builder = IC.Builder;
02255   Builder->SetInsertPoint(MulInstr);
02256   Module *M = I.getParent()->getParent()->getParent();
02257 
02258   // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
02259   Value *MulA = A, *MulB = B;
02260   if (WidthA < MulWidth)
02261     MulA = Builder->CreateZExt(A, MulType);
02262   if (WidthB < MulWidth)
02263     MulB = Builder->CreateZExt(B, MulType);
02264   Value *F =
02265       Intrinsic::getDeclaration(M, Intrinsic::umul_with_overflow, MulType);
02266   CallInst *Call = Builder->CreateCall2(F, MulA, MulB, "umul");
02267   IC.Worklist.Add(MulInstr);
02268 
02269   // If there are uses of mul result other than the comparison, we know that
02270   // they are truncation or binary AND. Change them to use result of
02271   // mul.with.overflow and adjust properly mask/size.
02272   if (MulVal->hasNUsesOrMore(2)) {
02273     Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value");
02274     for (User *U : MulVal->users()) {
02275       if (U == &I || U == OtherVal)
02276         continue;
02277       if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
02278         if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
02279           IC.ReplaceInstUsesWith(*TI, Mul);
02280         else
02281           TI->setOperand(0, Mul);
02282       } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
02283         assert(BO->getOpcode() == Instruction::And);
02284         // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
02285         ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
02286         APInt ShortMask = CI->getValue().trunc(MulWidth);
02287         Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask);
02288         Instruction *Zext =
02289             cast<Instruction>(Builder->CreateZExt(ShortAnd, BO->getType()));
02290         IC.Worklist.Add(Zext);
02291         IC.ReplaceInstUsesWith(*BO, Zext);
02292       } else {
02293         llvm_unreachable("Unexpected Binary operation");
02294       }
02295       IC.Worklist.Add(cast<Instruction>(U));
02296     }
02297   }
02298   if (isa<Instruction>(OtherVal))
02299     IC.Worklist.Add(cast<Instruction>(OtherVal));
02300 
02301   // The original icmp gets replaced with the overflow value, maybe inverted
02302   // depending on predicate.
02303   bool Inverse = false;
02304   switch (I.getPredicate()) {
02305   case ICmpInst::ICMP_NE:
02306     break;
02307   case ICmpInst::ICMP_EQ:
02308     Inverse = true;
02309     break;
02310   case ICmpInst::ICMP_UGT:
02311   case ICmpInst::ICMP_UGE:
02312     if (I.getOperand(0) == MulVal)
02313       break;
02314     Inverse = true;
02315     break;
02316   case ICmpInst::ICMP_ULT:
02317   case ICmpInst::ICMP_ULE:
02318     if (I.getOperand(1) == MulVal)
02319       break;
02320     Inverse = true;
02321     break;
02322   default:
02323     llvm_unreachable("Unexpected predicate");
02324   }
02325   if (Inverse) {
02326     Value *Res = Builder->CreateExtractValue(Call, 1);
02327     return BinaryOperator::CreateNot(Res);
02328   }
02329 
02330   return ExtractValueInst::Create(Call, 1);
02331 }
02332 
02333 // DemandedBitsLHSMask - When performing a comparison against a constant,
02334 // it is possible that not all the bits in the LHS are demanded.  This helper
02335 // method computes the mask that IS demanded.
02336 static APInt DemandedBitsLHSMask(ICmpInst &I,
02337                                  unsigned BitWidth, bool isSignCheck) {
02338   if (isSignCheck)
02339     return APInt::getSignBit(BitWidth);
02340 
02341   ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
02342   if (!CI) return APInt::getAllOnesValue(BitWidth);
02343   const APInt &RHS = CI->getValue();
02344 
02345   switch (I.getPredicate()) {
02346   // For a UGT comparison, we don't care about any bits that
02347   // correspond to the trailing ones of the comparand.  The value of these
02348   // bits doesn't impact the outcome of the comparison, because any value
02349   // greater than the RHS must differ in a bit higher than these due to carry.
02350   case ICmpInst::ICMP_UGT: {
02351     unsigned trailingOnes = RHS.countTrailingOnes();
02352     APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
02353     return ~lowBitsSet;
02354   }
02355 
02356   // Similarly, for a ULT comparison, we don't care about the trailing zeros.
02357   // Any value less than the RHS must differ in a higher bit because of carries.
02358   case ICmpInst::ICMP_ULT: {
02359     unsigned trailingZeros = RHS.countTrailingZeros();
02360     APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
02361     return ~lowBitsSet;
02362   }
02363 
02364   default:
02365     return APInt::getAllOnesValue(BitWidth);
02366   }
02367 
02368 }
02369 
02370 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
02371 /// should be swapped.
02372 /// The decision is based on how many times these two operands are reused
02373 /// as subtract operands and their positions in those instructions.
02374 /// The rational is that several architectures use the same instruction for
02375 /// both subtract and cmp, thus it is better if the order of those operands
02376 /// match.
02377 /// \return true if Op0 and Op1 should be swapped.
02378 static bool swapMayExposeCSEOpportunities(const Value * Op0,
02379                                           const Value * Op1) {
02380   // Filter out pointer value as those cannot appears directly in subtract.
02381   // FIXME: we may want to go through inttoptrs or bitcasts.
02382   if (Op0->getType()->isPointerTy())
02383     return false;
02384   // Count every uses of both Op0 and Op1 in a subtract.
02385   // Each time Op0 is the first operand, count -1: swapping is bad, the
02386   // subtract has already the same layout as the compare.
02387   // Each time Op0 is the second operand, count +1: swapping is good, the
02388   // subtract has a different layout as the compare.
02389   // At the end, if the benefit is greater than 0, Op0 should come second to
02390   // expose more CSE opportunities.
02391   int GlobalSwapBenefits = 0;
02392   for (const User *U : Op0->users()) {
02393     const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
02394     if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
02395       continue;
02396     // If Op0 is the first argument, this is not beneficial to swap the
02397     // arguments.
02398     int LocalSwapBenefits = -1;
02399     unsigned Op1Idx = 1;
02400     if (BinOp->getOperand(Op1Idx) == Op0) {
02401       Op1Idx = 0;
02402       LocalSwapBenefits = 1;
02403     }
02404     if (BinOp->getOperand(Op1Idx) != Op1)
02405       continue;
02406     GlobalSwapBenefits += LocalSwapBenefits;
02407   }
02408   return GlobalSwapBenefits > 0;
02409 }
02410 
02411 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
02412   bool Changed = false;
02413   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
02414   unsigned Op0Cplxity = getComplexity(Op0);
02415   unsigned Op1Cplxity = getComplexity(Op1);
02416 
02417   /// Orders the operands of the compare so that they are listed from most
02418   /// complex to least complex.  This puts constants before unary operators,
02419   /// before binary operators.
02420   if (Op0Cplxity < Op1Cplxity ||
02421         (Op0Cplxity == Op1Cplxity &&
02422          swapMayExposeCSEOpportunities(Op0, Op1))) {
02423     I.swapOperands();
02424     std::swap(Op0, Op1);
02425     Changed = true;
02426   }
02427 
02428   if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL))
02429     return ReplaceInstUsesWith(I, V);
02430 
02431   // comparing -val or val with non-zero is the same as just comparing val
02432   // ie, abs(val) != 0 -> val != 0
02433   if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
02434   {
02435     Value *Cond, *SelectTrue, *SelectFalse;
02436     if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
02437                             m_Value(SelectFalse)))) {
02438       if (Value *V = dyn_castNegVal(SelectTrue)) {
02439         if (V == SelectFalse)
02440           return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
02441       }
02442       else if (Value *V = dyn_castNegVal(SelectFalse)) {
02443         if (V == SelectTrue)
02444           return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
02445       }
02446     }
02447   }
02448 
02449   Type *Ty = Op0->getType();
02450 
02451   // icmp's with boolean values can always be turned into bitwise operations
02452   if (Ty->isIntegerTy(1)) {
02453     switch (I.getPredicate()) {
02454     default: llvm_unreachable("Invalid icmp instruction!");
02455     case ICmpInst::ICMP_EQ: {               // icmp eq i1 A, B -> ~(A^B)
02456       Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
02457       return BinaryOperator::CreateNot(Xor);
02458     }
02459     case ICmpInst::ICMP_NE:                  // icmp eq i1 A, B -> A^B
02460       return BinaryOperator::CreateXor(Op0, Op1);
02461 
02462     case ICmpInst::ICMP_UGT:
02463       std::swap(Op0, Op1);                   // Change icmp ugt -> icmp ult
02464       // FALL THROUGH
02465     case ICmpInst::ICMP_ULT:{               // icmp ult i1 A, B -> ~A & B
02466       Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
02467       return BinaryOperator::CreateAnd(Not, Op1);
02468     }
02469     case ICmpInst::ICMP_SGT:
02470       std::swap(Op0, Op1);                   // Change icmp sgt -> icmp slt
02471       // FALL THROUGH
02472     case ICmpInst::ICMP_SLT: {               // icmp slt i1 A, B -> A & ~B
02473       Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
02474       return BinaryOperator::CreateAnd(Not, Op0);
02475     }
02476     case ICmpInst::ICMP_UGE:
02477       std::swap(Op0, Op1);                   // Change icmp uge -> icmp ule
02478       // FALL THROUGH
02479     case ICmpInst::ICMP_ULE: {               //  icmp ule i1 A, B -> ~A | B
02480       Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
02481       return BinaryOperator::CreateOr(Not, Op1);
02482     }
02483     case ICmpInst::ICMP_SGE:
02484       std::swap(Op0, Op1);                   // Change icmp sge -> icmp sle
02485       // FALL THROUGH
02486     case ICmpInst::ICMP_SLE: {               //  icmp sle i1 A, B -> A | ~B
02487       Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
02488       return BinaryOperator::CreateOr(Not, Op0);
02489     }
02490     }
02491   }
02492 
02493   unsigned BitWidth = 0;
02494   if (Ty->isIntOrIntVectorTy())
02495     BitWidth = Ty->getScalarSizeInBits();
02496   else if (DL)  // Pointers require DL info to get their size.
02497     BitWidth = DL->getTypeSizeInBits(Ty->getScalarType());
02498 
02499   bool isSignBit = false;
02500 
02501   // See if we are doing a comparison with a constant.
02502   if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02503     Value *A = nullptr, *B = nullptr;
02504 
02505     // Match the following pattern, which is a common idiom when writing
02506     // overflow-safe integer arithmetic function.  The source performs an
02507     // addition in wider type, and explicitly checks for overflow using
02508     // comparisons against INT_MIN and INT_MAX.  Simplify this by using the
02509     // sadd_with_overflow intrinsic.
02510     //
02511     // TODO: This could probably be generalized to handle other overflow-safe
02512     // operations if we worked out the formulas to compute the appropriate
02513     // magic constants.
02514     //
02515     // sum = a + b
02516     // if (sum+128 >u 255)  ...  -> llvm.sadd.with.overflow.i8
02517     {
02518     ConstantInt *CI2;    // I = icmp ugt (add (add A, B), CI2), CI
02519     if (I.getPredicate() == ICmpInst::ICMP_UGT &&
02520         match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
02521       if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
02522         return Res;
02523     }
02524 
02525     // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
02526     if (I.isEquality() && CI->isZero() &&
02527         match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
02528       // (icmp cond A B) if cond is equality
02529       return new ICmpInst(I.getPredicate(), A, B);
02530     }
02531 
02532     // If we have an icmp le or icmp ge instruction, turn it into the
02533     // appropriate icmp lt or icmp gt instruction.  This allows us to rely on
02534     // them being folded in the code below.  The SimplifyICmpInst code has
02535     // already handled the edge cases for us, so we just assert on them.
02536     switch (I.getPredicate()) {
02537     default: break;
02538     case ICmpInst::ICMP_ULE:
02539       assert(!CI->isMaxValue(false));                 // A <=u MAX -> TRUE
02540       return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
02541                           Builder->getInt(CI->getValue()+1));
02542     case ICmpInst::ICMP_SLE:
02543       assert(!CI->isMaxValue(true));                  // A <=s MAX -> TRUE
02544       return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
02545                           Builder->getInt(CI->getValue()+1));
02546     case ICmpInst::ICMP_UGE:
02547       assert(!CI->isMinValue(false));                 // A >=u MIN -> TRUE
02548       return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
02549                           Builder->getInt(CI->getValue()-1));
02550     case ICmpInst::ICMP_SGE:
02551       assert(!CI->isMinValue(true));                  // A >=s MIN -> TRUE
02552       return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
02553                           Builder->getInt(CI->getValue()-1));
02554     }
02555 
02556     // (icmp eq/ne (ashr/lshr const2, A), const1)
02557     if (I.isEquality()) {
02558       ConstantInt *CI2;
02559       if (match(Op0, m_AShr(m_ConstantInt(CI2), m_Value(A))) ||
02560           match(Op0, m_LShr(m_ConstantInt(CI2), m_Value(A)))) {
02561         return FoldICmpCstShrCst(I, Op0, A, CI, CI2);
02562       }
02563     }
02564 
02565     // If this comparison is a normal comparison, it demands all
02566     // bits, if it is a sign bit comparison, it only demands the sign bit.
02567     bool UnusedBit;
02568     isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
02569   }
02570 
02571   // See if we can fold the comparison based on range information we can get
02572   // by checking whether bits are known to be zero or one in the input.
02573   if (BitWidth != 0) {
02574     APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
02575     APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
02576 
02577     if (SimplifyDemandedBits(I.getOperandUse(0),
02578                              DemandedBitsLHSMask(I, BitWidth, isSignBit),
02579                              Op0KnownZero, Op0KnownOne, 0))
02580       return &I;
02581     if (SimplifyDemandedBits(I.getOperandUse(1),
02582                              APInt::getAllOnesValue(BitWidth),
02583                              Op1KnownZero, Op1KnownOne, 0))
02584       return &I;
02585 
02586     // Given the known and unknown bits, compute a range that the LHS could be
02587     // in.  Compute the Min, Max and RHS values based on the known bits. For the
02588     // EQ and NE we use unsigned values.
02589     APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
02590     APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
02591     if (I.isSigned()) {
02592       ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
02593                                              Op0Min, Op0Max);
02594       ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
02595                                              Op1Min, Op1Max);
02596     } else {
02597       ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
02598                                                Op0Min, Op0Max);
02599       ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
02600                                                Op1Min, Op1Max);
02601     }
02602 
02603     // If Min and Max are known to be the same, then SimplifyDemandedBits
02604     // figured out that the LHS is a constant.  Just constant fold this now so
02605     // that code below can assume that Min != Max.
02606     if (!isa<Constant>(Op0) && Op0Min == Op0Max)
02607       return new ICmpInst(I.getPredicate(),
02608                           ConstantInt::get(Op0->getType(), Op0Min), Op1);
02609     if (!isa<Constant>(Op1) && Op1Min == Op1Max)
02610       return new ICmpInst(I.getPredicate(), Op0,
02611                           ConstantInt::get(Op1->getType(), Op1Min));
02612 
02613     // Based on the range information we know about the LHS, see if we can
02614     // simplify this comparison.  For example, (x&4) < 8 is always true.
02615     switch (I.getPredicate()) {
02616     default: llvm_unreachable("Unknown icmp opcode!");
02617     case ICmpInst::ICMP_EQ: {
02618       if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
02619         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02620 
02621       // If all bits are known zero except for one, then we know at most one
02622       // bit is set.   If the comparison is against zero, then this is a check
02623       // to see if *that* bit is set.
02624       APInt Op0KnownZeroInverted = ~Op0KnownZero;
02625       if (~Op1KnownZero == 0) {
02626         // If the LHS is an AND with the same constant, look through it.
02627         Value *LHS = nullptr;
02628         ConstantInt *LHSC = nullptr;
02629         if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
02630             LHSC->getValue() != Op0KnownZeroInverted)
02631           LHS = Op0;
02632 
02633         // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
02634         // then turn "((1 << x)&8) == 0" into "x != 3".
02635         // or turn "((1 << x)&7) == 0" into "x > 2".
02636         Value *X = nullptr;
02637         if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
02638           APInt ValToCheck = Op0KnownZeroInverted;
02639           if (ValToCheck.isPowerOf2()) {
02640             unsigned CmpVal = ValToCheck.countTrailingZeros();
02641             return new ICmpInst(ICmpInst::ICMP_NE, X,
02642                                 ConstantInt::get(X->getType(), CmpVal));
02643           } else if ((++ValToCheck).isPowerOf2()) {
02644             unsigned CmpVal = ValToCheck.countTrailingZeros() - 1;
02645             return new ICmpInst(ICmpInst::ICMP_UGT, X,
02646                                 ConstantInt::get(X->getType(), CmpVal));
02647           }
02648         }
02649 
02650         // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
02651         // then turn "((8 >>u x)&1) == 0" into "x != 3".
02652         const APInt *CI;
02653         if (Op0KnownZeroInverted == 1 &&
02654             match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
02655           return new ICmpInst(ICmpInst::ICMP_NE, X,
02656                               ConstantInt::get(X->getType(),
02657                                                CI->countTrailingZeros()));
02658       }
02659 
02660       break;
02661     }
02662     case ICmpInst::ICMP_NE: {
02663       if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
02664         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02665 
02666       // If all bits are known zero except for one, then we know at most one
02667       // bit is set.   If the comparison is against zero, then this is a check
02668       // to see if *that* bit is set.
02669       APInt Op0KnownZeroInverted = ~Op0KnownZero;
02670       if (~Op1KnownZero == 0) {
02671         // If the LHS is an AND with the same constant, look through it.
02672         Value *LHS = nullptr;
02673         ConstantInt *LHSC = nullptr;
02674         if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
02675             LHSC->getValue() != Op0KnownZeroInverted)
02676           LHS = Op0;
02677 
02678         // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
02679         // then turn "((1 << x)&8) != 0" into "x == 3".
02680         // or turn "((1 << x)&7) != 0" into "x < 3".
02681         Value *X = nullptr;
02682         if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
02683           APInt ValToCheck = Op0KnownZeroInverted;
02684           if (ValToCheck.isPowerOf2()) {
02685             unsigned CmpVal = ValToCheck.countTrailingZeros();
02686             return new ICmpInst(ICmpInst::ICMP_EQ, X,
02687                                 ConstantInt::get(X->getType(), CmpVal));
02688           } else if ((++ValToCheck).isPowerOf2()) {
02689             unsigned CmpVal = ValToCheck.countTrailingZeros();
02690             return new ICmpInst(ICmpInst::ICMP_ULT, X,
02691                                 ConstantInt::get(X->getType(), CmpVal));
02692           }
02693         }
02694 
02695         // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
02696         // then turn "((8 >>u x)&1) != 0" into "x == 3".
02697         const APInt *CI;
02698         if (Op0KnownZeroInverted == 1 &&
02699             match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
02700           return new ICmpInst(ICmpInst::ICMP_EQ, X,
02701                               ConstantInt::get(X->getType(),
02702                                                CI->countTrailingZeros()));
02703       }
02704 
02705       break;
02706     }
02707     case ICmpInst::ICMP_ULT:
02708       if (Op0Max.ult(Op1Min))          // A <u B -> true if max(A) < min(B)
02709         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02710       if (Op0Min.uge(Op1Max))          // A <u B -> false if min(A) >= max(B)
02711         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02712       if (Op1Min == Op0Max)            // A <u B -> A != B if max(A) == min(B)
02713         return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
02714       if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02715         if (Op1Max == Op0Min+1)        // A <u C -> A == C-1 if min(A)+1 == C
02716           return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
02717                               Builder->getInt(CI->getValue()-1));
02718 
02719         // (x <u 2147483648) -> (x >s -1)  -> true if sign bit clear
02720         if (CI->isMinValue(true))
02721           return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
02722                            Constant::getAllOnesValue(Op0->getType()));
02723       }
02724       break;
02725     case ICmpInst::ICMP_UGT:
02726       if (Op0Min.ugt(Op1Max))          // A >u B -> true if min(A) > max(B)
02727         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02728       if (Op0Max.ule(Op1Min))          // A >u B -> false if max(A) <= max(B)
02729         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02730 
02731       if (Op1Max == Op0Min)            // A >u B -> A != B if min(A) == max(B)
02732         return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
02733       if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02734         if (Op1Min == Op0Max-1)        // A >u C -> A == C+1 if max(a)-1 == C
02735           return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
02736                               Builder->getInt(CI->getValue()+1));
02737 
02738         // (x >u 2147483647) -> (x <s 0)  -> true if sign bit set
02739         if (CI->isMaxValue(true))
02740           return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
02741                               Constant::getNullValue(Op0->getType()));
02742       }
02743       break;
02744     case ICmpInst::ICMP_SLT:
02745       if (Op0Max.slt(Op1Min))          // A <s B -> true if max(A) < min(C)
02746         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02747       if (Op0Min.sge(Op1Max))          // A <s B -> false if min(A) >= max(C)
02748         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02749       if (Op1Min == Op0Max)            // A <s B -> A != B if max(A) == min(B)
02750         return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
02751       if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02752         if (Op1Max == Op0Min+1)        // A <s C -> A == C-1 if min(A)+1 == C
02753           return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
02754                               Builder->getInt(CI->getValue()-1));
02755       }
02756       break;
02757     case ICmpInst::ICMP_SGT:
02758       if (Op0Min.sgt(Op1Max))          // A >s B -> true if min(A) > max(B)
02759         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02760       if (Op0Max.sle(Op1Min))          // A >s B -> false if max(A) <= min(B)
02761         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02762 
02763       if (Op1Max == Op0Min)            // A >s B -> A != B if min(A) == max(B)
02764         return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
02765       if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02766         if (Op1Min == Op0Max-1)        // A >s C -> A == C+1 if max(A)-1 == C
02767           return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
02768                               Builder->getInt(CI->getValue()+1));
02769       }
02770       break;
02771     case ICmpInst::ICMP_SGE:
02772       assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
02773       if (Op0Min.sge(Op1Max))          // A >=s B -> true if min(A) >= max(B)
02774         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02775       if (Op0Max.slt(Op1Min))          // A >=s B -> false if max(A) < min(B)
02776         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02777       break;
02778     case ICmpInst::ICMP_SLE:
02779       assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
02780       if (Op0Max.sle(Op1Min))          // A <=s B -> true if max(A) <= min(B)
02781         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02782       if (Op0Min.sgt(Op1Max))          // A <=s B -> false if min(A) > max(B)
02783         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02784       break;
02785     case ICmpInst::ICMP_UGE:
02786       assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
02787       if (Op0Min.uge(Op1Max))          // A >=u B -> true if min(A) >= max(B)
02788         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02789       if (Op0Max.ult(Op1Min))          // A >=u B -> false if max(A) < min(B)
02790         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02791       break;
02792     case ICmpInst::ICMP_ULE:
02793       assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
02794       if (Op0Max.ule(Op1Min))          // A <=u B -> true if max(A) <= min(B)
02795         return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
02796       if (Op0Min.ugt(Op1Max))          // A <=u B -> false if min(A) > max(B)
02797         return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
02798       break;
02799     }
02800 
02801     // Turn a signed comparison into an unsigned one if both operands
02802     // are known to have the same sign.
02803     if (I.isSigned() &&
02804         ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
02805          (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
02806       return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
02807   }
02808 
02809   // Test if the ICmpInst instruction is used exclusively by a select as
02810   // part of a minimum or maximum operation. If so, refrain from doing
02811   // any other folding. This helps out other analyses which understand
02812   // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
02813   // and CodeGen. And in this case, at least one of the comparison
02814   // operands has at least one user besides the compare (the select),
02815   // which would often largely negate the benefit of folding anyway.
02816   if (I.hasOneUse())
02817     if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
02818       if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
02819           (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
02820         return nullptr;
02821 
02822   // See if we are doing a comparison between a constant and an instruction that
02823   // can be folded into the comparison.
02824   if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
02825     // Since the RHS is a ConstantInt (CI), if the left hand side is an
02826     // instruction, see if that instruction also has constants so that the
02827     // instruction can be folded into the icmp
02828     if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
02829       if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
02830         return Res;
02831   }
02832 
02833   // Handle icmp with constant (but not simple integer constant) RHS
02834   if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
02835     if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
02836       switch (LHSI->getOpcode()) {
02837       case Instruction::GetElementPtr:
02838           // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
02839         if (RHSC->isNullValue() &&
02840             cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
02841           return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
02842                   Constant::getNullValue(LHSI->getOperand(0)->getType()));
02843         break;
02844       case Instruction::PHI:
02845         // Only fold icmp into the PHI if the phi and icmp are in the same
02846         // block.  If in the same block, we're encouraging jump threading.  If
02847         // not, we are just pessimizing the code by making an i1 phi.
02848         if (LHSI->getParent() == I.getParent())
02849           if (Instruction *NV = FoldOpIntoPhi(I))
02850             return NV;
02851         break;
02852       case Instruction::Select: {
02853         // If either operand of the select is a constant, we can fold the
02854         // comparison into the select arms, which will cause one to be
02855         // constant folded and the select turned into a bitwise or.
02856         Value *Op1 = nullptr, *Op2 = nullptr;
02857         if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
02858           Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
02859         if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
02860           Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
02861 
02862         // We only want to perform this transformation if it will not lead to
02863         // additional code. This is true if either both sides of the select
02864         // fold to a constant (in which case the icmp is replaced with a select
02865         // which will usually simplify) or this is the only user of the
02866         // select (in which case we are trading a select+icmp for a simpler
02867         // select+icmp).
02868         if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
02869           if (!Op1)
02870             Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
02871                                       RHSC, I.getName());
02872           if (!Op2)
02873             Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
02874                                       RHSC, I.getName());
02875           return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
02876         }
02877         break;
02878       }
02879       case Instruction::IntToPtr:
02880         // icmp pred inttoptr(X), null -> icmp pred X, 0
02881         if (RHSC->isNullValue() && DL &&
02882             DL->getIntPtrType(RHSC->getType()) ==
02883                LHSI->getOperand(0)->getType())
02884           return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
02885                         Constant::getNullValue(LHSI->getOperand(0)->getType()));
02886         break;
02887 
02888       case Instruction::Load:
02889         // Try to optimize things like "A[i] > 4" to index computations.
02890         if (GetElementPtrInst *GEP =
02891               dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
02892           if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
02893             if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
02894                 !cast<LoadInst>(LHSI)->isVolatile())
02895               if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
02896                 return Res;
02897         }
02898         break;
02899       }
02900   }
02901 
02902   // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
02903   if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
02904     if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
02905       return NI;
02906   if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
02907     if (Instruction *NI = FoldGEPICmp(GEP, Op0,
02908                            ICmpInst::getSwappedPredicate(I.getPredicate()), I))
02909       return NI;
02910 
02911   // Test to see if the operands of the icmp are casted versions of other
02912   // values.  If the ptr->ptr cast can be stripped off both arguments, we do so
02913   // now.
02914   if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
02915     if (Op0->getType()->isPointerTy() &&
02916         (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
02917       // We keep moving the cast from the left operand over to the right
02918       // operand, where it can often be eliminated completely.
02919       Op0 = CI->getOperand(0);
02920 
02921       // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
02922       // so eliminate it as well.
02923       if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
02924         Op1 = CI2->getOperand(0);
02925 
02926       // If Op1 is a constant, we can fold the cast into the constant.
02927       if (Op0->getType() != Op1->getType()) {
02928         if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
02929           Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
02930         } else {
02931           // Otherwise, cast the RHS right before the icmp
02932           Op1 = Builder->CreateBitCast(Op1, Op0->getType());
02933         }
02934       }
02935       return new ICmpInst(I.getPredicate(), Op0, Op1);
02936     }
02937   }
02938 
02939   if (isa<CastInst>(Op0)) {
02940     // Handle the special case of: icmp (cast bool to X), <cst>
02941     // This comes up when you have code like
02942     //   int X = A < B;
02943     //   if (X) ...
02944     // For generality, we handle any zero-extension of any operand comparison
02945     // with a constant or another cast from the same type.
02946     if (isa<Constant>(Op1) || isa<CastInst>(Op1))
02947       if (Instruction *R = visitICmpInstWithCastAndCast(I))
02948         return R;
02949   }
02950 
02951   // Special logic for binary operators.
02952   BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
02953   BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
02954   if (BO0 || BO1) {
02955     CmpInst::Predicate Pred = I.getPredicate();
02956     bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
02957     if (BO0 && isa<OverflowingBinaryOperator>(BO0))
02958       NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
02959         (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
02960         (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
02961     if (BO1 && isa<OverflowingBinaryOperator>(BO1))
02962       NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
02963         (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
02964         (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
02965 
02966     // Analyze the case when either Op0 or Op1 is an add instruction.
02967     // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
02968     Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
02969     if (BO0 && BO0->getOpcode() == Instruction::Add)
02970       A = BO0->getOperand(0), B = BO0->getOperand(1);
02971     if (BO1 && BO1->getOpcode() == Instruction::Add)
02972       C = BO1->getOperand(0), D = BO1->getOperand(1);
02973 
02974     // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
02975     if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
02976       return new ICmpInst(Pred, A == Op1 ? B : A,
02977                           Constant::getNullValue(Op1->getType()));
02978 
02979     // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
02980     if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
02981       return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
02982                           C == Op0 ? D : C);
02983 
02984     // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
02985     if (A && C && (A == C || A == D || B == C || B == D) &&
02986         NoOp0WrapProblem && NoOp1WrapProblem &&
02987         // Try not to increase register pressure.
02988         BO0->hasOneUse() && BO1->hasOneUse()) {
02989       // Determine Y and Z in the form icmp (X+Y), (X+Z).
02990       Value *Y, *Z;
02991       if (A == C) {
02992         // C + B == C + D  ->  B == D
02993         Y = B;
02994         Z = D;
02995       } else if (A == D) {
02996         // D + B == C + D  ->  B == C
02997         Y = B;
02998         Z = C;
02999       } else if (B == C) {
03000         // A + C == C + D  ->  A == D
03001         Y = A;
03002         Z = D;
03003       } else {
03004         assert(B == D);
03005         // A + D == C + D  ->  A == C
03006         Y = A;
03007         Z = C;
03008       }
03009       return new ICmpInst(Pred, Y, Z);
03010     }
03011 
03012     // icmp slt (X + -1), Y -> icmp sle X, Y
03013     if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
03014         match(B, m_AllOnes()))
03015       return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
03016 
03017     // icmp sge (X + -1), Y -> icmp sgt X, Y
03018     if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
03019         match(B, m_AllOnes()))
03020       return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
03021 
03022     // icmp sle (X + 1), Y -> icmp slt X, Y
03023     if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
03024         match(B, m_One()))
03025       return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
03026 
03027     // icmp sgt (X + 1), Y -> icmp sge X, Y
03028     if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
03029         match(B, m_One()))
03030       return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
03031 
03032     // if C1 has greater magnitude than C2:
03033     //  icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
03034     //  s.t. C3 = C1 - C2
03035     //
03036     // if C2 has greater magnitude than C1:
03037     //  icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
03038     //  s.t. C3 = C2 - C1
03039     if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
03040         (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
03041       if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
03042         if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
03043           const APInt &AP1 = C1->getValue();
03044           const APInt &AP2 = C2->getValue();
03045           if (AP1.isNegative() == AP2.isNegative()) {
03046             APInt AP1Abs = C1->getValue().abs();
03047             APInt AP2Abs = C2->getValue().abs();
03048             if (AP1Abs.uge(AP2Abs)) {
03049               ConstantInt *C3 = Builder->getInt(AP1 - AP2);
03050               Value *NewAdd = Builder->CreateNSWAdd(A, C3);
03051               return new ICmpInst(Pred, NewAdd, C);
03052             } else {
03053               ConstantInt *C3 = Builder->getInt(AP2 - AP1);
03054               Value *NewAdd = Builder->CreateNSWAdd(C, C3);
03055               return new ICmpInst(Pred, A, NewAdd);
03056             }
03057           }
03058         }
03059 
03060 
03061     // Analyze the case when either Op0 or Op1 is a sub instruction.
03062     // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
03063     A = nullptr; B = nullptr; C = nullptr; D = nullptr;
03064     if (BO0 && BO0->getOpcode() == Instruction::Sub)
03065       A = BO0->getOperand(0), B = BO0->getOperand(1);
03066     if (BO1 && BO1->getOpcode() == Instruction::Sub)
03067       C = BO1->getOperand(0), D = BO1->getOperand(1);
03068 
03069     // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
03070     if (A == Op1 && NoOp0WrapProblem)
03071       return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
03072 
03073     // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
03074     if (C == Op0 && NoOp1WrapProblem)
03075       return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
03076 
03077     // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
03078     if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
03079         // Try not to increase register pressure.
03080         BO0->hasOneUse() && BO1->hasOneUse())
03081       return new ICmpInst(Pred, A, C);
03082 
03083     // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
03084     if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
03085         // Try not to increase register pressure.
03086         BO0->hasOneUse() && BO1->hasOneUse())
03087       return new ICmpInst(Pred, D, B);
03088 
03089     // icmp (0-X) < cst --> x > -cst
03090     if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
03091       Value *X;
03092       if (match(BO0, m_Neg(m_Value(X))))
03093         if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
03094           if (!RHSC->isMinValue(/*isSigned=*/true))
03095             return new ICmpInst(I.getSwappedPredicate(), X,
03096                                 ConstantExpr::getNeg(RHSC));
03097     }
03098 
03099     BinaryOperator *SRem = nullptr;
03100     // icmp (srem X, Y), Y
03101     if (BO0 && BO0->getOpcode() == Instruction::SRem &&
03102         Op1 == BO0->getOperand(1))
03103       SRem = BO0;
03104     // icmp Y, (srem X, Y)
03105     else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
03106              Op0 == BO1->getOperand(1))
03107       SRem = BO1;
03108     if (SRem) {
03109       // We don't check hasOneUse to avoid increasing register pressure because
03110       // the value we use is the same value this instruction was already using.
03111       switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
03112         default: break;
03113         case ICmpInst::ICMP_EQ:
03114           return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
03115         case ICmpInst::ICMP_NE:
03116           return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
03117         case ICmpInst::ICMP_SGT:
03118         case ICmpInst::ICMP_SGE:
03119           return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
03120                               Constant::getAllOnesValue(SRem->getType()));
03121         case ICmpInst::ICMP_SLT:
03122         case ICmpInst::ICMP_SLE:
03123           return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
03124                               Constant::getNullValue(SRem->getType()));
03125       }
03126     }
03127 
03128     if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
03129         BO0->hasOneUse() && BO1->hasOneUse() &&
03130         BO0->getOperand(1) == BO1->getOperand(1)) {
03131       switch (BO0->getOpcode()) {
03132       default: break;
03133       case Instruction::Add:
03134       case Instruction::Sub:
03135       case Instruction::Xor:
03136         if (I.isEquality())    // a+x icmp eq/ne b+x --> a icmp b
03137           return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
03138                               BO1->getOperand(0));
03139         // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
03140         if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
03141           if (CI->getValue().isSignBit()) {
03142             ICmpInst::Predicate Pred = I.isSigned()
03143                                            ? I.getUnsignedPredicate()
03144                                            : I.getSignedPredicate();
03145             return new ICmpInst(Pred, BO0->getOperand(0),
03146                                 BO1->getOperand(0));
03147           }
03148 
03149           if (CI->isMaxValue(true)) {
03150             ICmpInst::Predicate Pred = I.isSigned()
03151                                            ? I.getUnsignedPredicate()
03152                                            : I.getSignedPredicate();
03153             Pred = I.getSwappedPredicate(Pred);
03154             return new ICmpInst(Pred, BO0->getOperand(0),
03155                                 BO1->getOperand(0));
03156           }
03157         }
03158         break;
03159       case Instruction::Mul:
03160         if (!I.isEquality())
03161           break;
03162 
03163         if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
03164           // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
03165           // Mask = -1 >> count-trailing-zeros(Cst).
03166           if (!CI->isZero() && !CI->isOne()) {
03167             const APInt &AP = CI->getValue();
03168             ConstantInt *Mask = ConstantInt::get(I.getContext(),
03169                                     APInt::getLowBitsSet(AP.getBitWidth(),
03170                                                          AP.getBitWidth() -
03171                                                     AP.countTrailingZeros()));
03172             Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
03173             Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
03174             return new ICmpInst(I.getPredicate(), And1, And2);
03175           }
03176         }
03177         break;
03178       case Instruction::UDiv:
03179       case Instruction::LShr:
03180         if (I.isSigned())
03181           break;
03182         // fall-through
03183       case Instruction::SDiv:
03184       case Instruction::AShr:
03185         if (!BO0->isExact() || !BO1->isExact())
03186           break;
03187         return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
03188                             BO1->getOperand(0));
03189       case Instruction::Shl: {
03190         bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
03191         bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
03192         if (!NUW && !NSW)
03193           break;
03194         if (!NSW && I.isSigned())
03195           break;
03196         return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
03197                             BO1->getOperand(0));
03198       }
03199       }
03200     }
03201   }
03202 
03203   { Value *A, *B;
03204     // Transform (A & ~B) == 0 --> (A & B) != 0
03205     // and       (A & ~B) != 0 --> (A & B) == 0
03206     // if A is a power of 2.
03207     if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
03208         match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A) && I.isEquality())
03209       return new ICmpInst(I.getInversePredicate(),
03210                           Builder->CreateAnd(A, B),
03211                           Op1);
03212 
03213     // ~x < ~y --> y < x
03214     // ~x < cst --> ~cst < x
03215     if (match(Op0, m_Not(m_Value(A)))) {
03216       if (match(Op1, m_Not(m_Value(B))))
03217         return new ICmpInst(I.getPredicate(), B, A);
03218       if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
03219         return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
03220     }
03221 
03222     // (a+b) <u a  --> llvm.uadd.with.overflow.
03223     // (a+b) <u b  --> llvm.uadd.with.overflow.
03224     if (I.getPredicate() == ICmpInst::ICMP_ULT &&
03225         match(Op0, m_Add(m_Value(A), m_Value(B))) &&
03226         (Op1 == A || Op1 == B))
03227       if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
03228         return R;
03229 
03230     // a >u (a+b)  --> llvm.uadd.with.overflow.
03231     // b >u (a+b)  --> llvm.uadd.with.overflow.
03232     if (I.getPredicate() == ICmpInst::ICMP_UGT &&
03233         match(Op1, m_Add(m_Value(A), m_Value(B))) &&
03234         (Op0 == A || Op0 == B))
03235       if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
03236         return R;
03237 
03238     // (zext a) * (zext b)  --> llvm.umul.with.overflow.
03239     if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
03240       if (Instruction *R = ProcessUMulZExtIdiom(I, Op0, Op1, *this))
03241         return R;
03242     }
03243     if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
03244       if (Instruction *R = ProcessUMulZExtIdiom(I, Op1, Op0, *this))
03245         return R;
03246     }
03247   }
03248 
03249   if (I.isEquality()) {
03250     Value *A, *B, *C, *D;
03251 
03252     if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
03253       if (A == Op1 || B == Op1) {    // (A^B) == A  ->  B == 0
03254         Value *OtherVal = A == Op1 ? B : A;
03255         return new ICmpInst(I.getPredicate(), OtherVal,
03256                             Constant::getNullValue(A->getType()));
03257       }
03258 
03259       if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
03260         // A^c1 == C^c2 --> A == C^(c1^c2)
03261         ConstantInt *C1, *C2;
03262         if (match(B, m_ConstantInt(C1)) &&
03263             match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
03264           Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
03265           Value *Xor = Builder->CreateXor(C, NC);
03266           return new ICmpInst(I.getPredicate(), A, Xor);
03267         }
03268 
03269         // A^B == A^D -> B == D
03270         if (A == C) return new ICmpInst(I.getPredicate(), B, D);
03271         if (A == D) return new ICmpInst(I.getPredicate(), B, C);
03272         if (B == C) return new ICmpInst(I.getPredicate(), A, D);
03273         if (B == D) return new ICmpInst(I.getPredicate(), A, C);
03274       }
03275     }
03276 
03277     if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
03278         (A == Op0 || B == Op0)) {
03279       // A == (A^B)  ->  B == 0
03280       Value *OtherVal = A == Op0 ? B : A;
03281       return new ICmpInst(I.getPredicate(), OtherVal,
03282                           Constant::getNullValue(A->getType()));
03283     }
03284 
03285     // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
03286     if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
03287         match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
03288       Value *X = nullptr, *Y = nullptr, *Z = nullptr;
03289 
03290       if (A == C) {
03291         X = B; Y = D; Z = A;
03292       } else if (A == D) {
03293         X = B; Y = C; Z = A;
03294       } else if (B == C) {
03295         X = A; Y = D; Z = B;
03296       } else if (B == D) {
03297         X = A; Y = C; Z = B;
03298       }
03299 
03300       if (X) {   // Build (X^Y) & Z
03301         Op1 = Builder->CreateXor(X, Y);
03302         Op1 = Builder->CreateAnd(Op1, Z);
03303         I.setOperand(0, Op1);
03304         I.setOperand(1, Constant::getNullValue(Op1->getType()));
03305         return &I;
03306       }
03307     }
03308 
03309     // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
03310     // and       (B & (1<<X)-1) == (zext A) --> A == (trunc B)
03311     ConstantInt *Cst1;
03312     if ((Op0->hasOneUse() &&
03313          match(Op0, m_ZExt(m_Value(A))) &&
03314          match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
03315         (Op1->hasOneUse() &&
03316          match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
03317          match(Op1, m_ZExt(m_Value(A))))) {
03318       APInt Pow2 = Cst1->getValue() + 1;
03319       if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
03320           Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
03321         return new ICmpInst(I.getPredicate(), A,
03322                             Builder->CreateTrunc(B, A->getType()));
03323     }
03324 
03325     // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
03326     // For lshr and ashr pairs.
03327     if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
03328          match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
03329         (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
03330          match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
03331       unsigned TypeBits = Cst1->getBitWidth();
03332       unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
03333       if (ShAmt < TypeBits && ShAmt != 0) {
03334         ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE
03335                                        ? ICmpInst::ICMP_UGE
03336                                        : ICmpInst::ICMP_ULT;
03337         Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
03338         APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
03339         return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal));
03340       }
03341     }
03342 
03343     // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
03344     // "icmp (and X, mask), cst"
03345     uint64_t ShAmt = 0;
03346     if (Op0->hasOneUse() &&
03347         match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
03348                                            m_ConstantInt(ShAmt))))) &&
03349         match(Op1, m_ConstantInt(Cst1)) &&
03350         // Only do this when A has multiple uses.  This is most important to do
03351         // when it exposes other optimizations.
03352         !A->hasOneUse()) {
03353       unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
03354 
03355       if (ShAmt < ASize) {
03356         APInt MaskV =
03357           APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
03358         MaskV <<= ShAmt;
03359 
03360         APInt CmpV = Cst1->getValue().zext(ASize);
03361         CmpV <<= ShAmt;
03362 
03363         Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
03364         return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
03365       }
03366     }
03367   }
03368 
03369   {
03370     Value *X; ConstantInt *Cst;
03371     // icmp X+Cst, X
03372     if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
03373       return FoldICmpAddOpCst(I, X, Cst, I.getPredicate());
03374 
03375     // icmp X, X+Cst
03376     if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
03377       return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
03378   }
03379   return Changed ? &I : nullptr;
03380 }
03381 
03382 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
03383 ///
03384 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
03385                                                 Instruction *LHSI,
03386                                                 Constant *RHSC) {
03387   if (!isa<ConstantFP>(RHSC)) return nullptr;
03388   const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
03389 
03390   // Get the width of the mantissa.  We don't want to hack on conversions that
03391   // might lose information from the integer, e.g. "i64 -> float"
03392   int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
03393   if (MantissaWidth == -1) return nullptr;  // Unknown.
03394 
03395   // Check to see that the input is converted from an integer type that is small
03396   // enough that preserves all bits.  TODO: check here for "known" sign bits.
03397   // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
03398   unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
03399 
03400   // If this is a uitofp instruction, we need an extra bit to hold the sign.
03401   bool LHSUnsigned = isa<UIToFPInst>(LHSI);
03402   if (LHSUnsigned)
03403     ++InputSize;
03404 
03405   // If the conversion would lose info, don't hack on this.
03406   if ((int)InputSize > MantissaWidth)
03407     return nullptr;
03408 
03409   // Otherwise, we can potentially simplify the comparison.  We know that it
03410   // will always come through as an integer value and we know the constant is
03411   // not a NAN (it would have been previously simplified).
03412   assert(!RHS.isNaN() && "NaN comparison not already folded!");
03413 
03414   ICmpInst::Predicate Pred;
03415   switch (I.getPredicate()) {
03416   default: llvm_unreachable("Unexpected predicate!");
03417   case FCmpInst::FCMP_UEQ:
03418   case FCmpInst::FCMP_OEQ:
03419     Pred = ICmpInst::ICMP_EQ;
03420     break;
03421   case FCmpInst::FCMP_UGT:
03422   case FCmpInst::FCMP_OGT:
03423     Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
03424     break;
03425   case FCmpInst::FCMP_UGE:
03426   case FCmpInst::FCMP_OGE:
03427     Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
03428     break;
03429   case FCmpInst::FCMP_ULT:
03430   case FCmpInst::FCMP_OLT:
03431     Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
03432     break;
03433   case FCmpInst::FCMP_ULE:
03434   case FCmpInst::FCMP_OLE:
03435     Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
03436     break;
03437   case FCmpInst::FCMP_UNE:
03438   case FCmpInst::FCMP_ONE:
03439     Pred = ICmpInst::ICMP_NE;
03440     break;
03441   case FCmpInst::FCMP_ORD:
03442     return ReplaceInstUsesWith(I, Builder->getTrue());
03443   case FCmpInst::FCMP_UNO:
03444     return ReplaceInstUsesWith(I, Builder->getFalse());
03445   }
03446 
03447   IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
03448 
03449   // Now we know that the APFloat is a normal number, zero or inf.
03450 
03451   // See if the FP constant is too large for the integer.  For example,
03452   // comparing an i8 to 300.0.
03453   unsigned IntWidth = IntTy->getScalarSizeInBits();
03454 
03455   if (!LHSUnsigned) {
03456     // If the RHS value is > SignedMax, fold the comparison.  This handles +INF
03457     // and large values.
03458     APFloat SMax(RHS.getSemantics());
03459     SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
03460                           APFloat::rmNearestTiesToEven);
03461     if (SMax.compare(RHS) == APFloat::cmpLessThan) {  // smax < 13123.0
03462       if (Pred == ICmpInst::ICMP_NE  || Pred == ICmpInst::ICMP_SLT ||
03463           Pred == ICmpInst::ICMP_SLE)
03464         return ReplaceInstUsesWith(I, Builder->getTrue());
03465       return ReplaceInstUsesWith(I, Builder->getFalse());
03466     }
03467   } else {
03468     // If the RHS value is > UnsignedMax, fold the comparison. This handles
03469     // +INF and large values.
03470     APFloat UMax(RHS.getSemantics());
03471     UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
03472                           APFloat::rmNearestTiesToEven);
03473     if (UMax.compare(RHS) == APFloat::cmpLessThan) {  // umax < 13123.0
03474       if (Pred == ICmpInst::ICMP_NE  || Pred == ICmpInst::ICMP_ULT ||
03475           Pred == ICmpInst::ICMP_ULE)
03476         return ReplaceInstUsesWith(I, Builder->getTrue());
03477       return ReplaceInstUsesWith(I, Builder->getFalse());
03478     }
03479   }
03480 
03481   if (!LHSUnsigned) {
03482     // See if the RHS value is < SignedMin.
03483     APFloat SMin(RHS.getSemantics());
03484     SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
03485                           APFloat::rmNearestTiesToEven);
03486     if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
03487       if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
03488           Pred == ICmpInst::ICMP_SGE)
03489         return ReplaceInstUsesWith(I, Builder->getTrue());
03490       return ReplaceInstUsesWith(I, Builder->getFalse());
03491     }
03492   } else {
03493     // See if the RHS value is < UnsignedMin.
03494     APFloat SMin(RHS.getSemantics());
03495     SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
03496                           APFloat::rmNearestTiesToEven);
03497     if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
03498       if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
03499           Pred == ICmpInst::ICMP_UGE)
03500         return ReplaceInstUsesWith(I, Builder->getTrue());
03501       return ReplaceInstUsesWith(I, Builder->getFalse());
03502     }
03503   }
03504 
03505   // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
03506   // [0, UMAX], but it may still be fractional.  See if it is fractional by
03507   // casting the FP value to the integer value and back, checking for equality.
03508   // Don't do this for zero, because -0.0 is not fractional.
03509   Constant *RHSInt = LHSUnsigned
03510     ? ConstantExpr::getFPToUI(RHSC, IntTy)
03511     : ConstantExpr::getFPToSI(RHSC, IntTy);
03512   if (!RHS.isZero()) {
03513     bool Equal = LHSUnsigned
03514       ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
03515       : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
03516     if (!Equal) {
03517       // If we had a comparison against a fractional value, we have to adjust
03518       // the compare predicate and sometimes the value.  RHSC is rounded towards
03519       // zero at this point.
03520       switch (Pred) {
03521       default: llvm_unreachable("Unexpected integer comparison!");
03522       case ICmpInst::ICMP_NE:  // (float)int != 4.4   --> true
03523         return ReplaceInstUsesWith(I, Builder->getTrue());
03524       case ICmpInst::ICMP_EQ:  // (float)int == 4.4   --> false
03525         return ReplaceInstUsesWith(I, Builder->getFalse());
03526       case ICmpInst::ICMP_ULE:
03527         // (float)int <= 4.4   --> int <= 4
03528         // (float)int <= -4.4  --> false
03529         if (RHS.isNegative())
03530           return ReplaceInstUsesWith(I, Builder->getFalse());
03531         break;
03532       case ICmpInst::ICMP_SLE:
03533         // (float)int <= 4.4   --> int <= 4
03534         // (float)int <= -4.4  --> int < -4
03535         if (RHS.isNegative())
03536           Pred = ICmpInst::ICMP_SLT;
03537         break;
03538       case ICmpInst::ICMP_ULT:
03539         // (float)int < -4.4   --> false
03540         // (float)int < 4.4    --> int <= 4
03541         if (RHS.isNegative())
03542           return ReplaceInstUsesWith(I, Builder->getFalse());
03543         Pred = ICmpInst::ICMP_ULE;
03544         break;
03545       case ICmpInst::ICMP_SLT:
03546         // (float)int < -4.4   --> int < -4
03547         // (float)int < 4.4    --> int <= 4
03548         if (!RHS.isNegative())
03549           Pred = ICmpInst::ICMP_SLE;
03550         break;
03551       case ICmpInst::ICMP_UGT:
03552         // (float)int > 4.4    --> int > 4
03553         // (float)int > -4.4   --> true
03554         if (RHS.isNegative())
03555           return ReplaceInstUsesWith(I, Builder->getTrue());
03556         break;
03557       case ICmpInst::ICMP_SGT:
03558         // (float)int > 4.4    --> int > 4
03559         // (float)int > -4.4   --> int >= -4
03560         if (RHS.isNegative())
03561           Pred = ICmpInst::ICMP_SGE;
03562         break;
03563       case ICmpInst::ICMP_UGE:
03564         // (float)int >= -4.4   --> true
03565         // (float)int >= 4.4    --> int > 4
03566         if (RHS.isNegative())
03567           return ReplaceInstUsesWith(I, Builder->getTrue());
03568         Pred = ICmpInst::ICMP_UGT;
03569         break;
03570       case ICmpInst::ICMP_SGE:
03571         // (float)int >= -4.4   --> int >= -4
03572         // (float)int >= 4.4    --> int > 4
03573         if (!RHS.isNegative())
03574           Pred = ICmpInst::ICMP_SGT;
03575         break;
03576       }
03577     }
03578   }
03579 
03580   // Lower this FP comparison into an appropriate integer version of the
03581   // comparison.
03582   return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
03583 }
03584 
03585 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
03586   bool Changed = false;
03587 
03588   /// Orders the operands of the compare so that they are listed from most
03589   /// complex to least complex.  This puts constants before unary operators,
03590   /// before binary operators.
03591   if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
03592     I.swapOperands();
03593     Changed = true;
03594   }
03595 
03596   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
03597 
03598   if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, DL))
03599     return ReplaceInstUsesWith(I, V);
03600 
03601   // Simplify 'fcmp pred X, X'
03602   if (Op0 == Op1) {
03603     switch (I.getPredicate()) {
03604     default: llvm_unreachable("Unknown predicate!");
03605     case FCmpInst::FCMP_UNO:    // True if unordered: isnan(X) | isnan(Y)
03606     case FCmpInst::FCMP_ULT:    // True if unordered or less than
03607     case FCmpInst::FCMP_UGT:    // True if unordered or greater than
03608     case FCmpInst::FCMP_UNE:    // True if unordered or not equal
03609       // Canonicalize these to be 'fcmp uno %X, 0.0'.
03610       I.setPredicate(FCmpInst::FCMP_UNO);
03611       I.setOperand(1, Constant::getNullValue(Op0->getType()));
03612       return &I;
03613 
03614     case FCmpInst::FCMP_ORD:    // True if ordered (no nans)
03615     case FCmpInst::FCMP_OEQ:    // True if ordered and equal
03616     case FCmpInst::FCMP_OGE:    // True if ordered and greater than or equal
03617     case FCmpInst::FCMP_OLE:    // True if ordered and less than or equal
03618       // Canonicalize these to be 'fcmp ord %X, 0.0'.
03619       I.setPredicate(FCmpInst::FCMP_ORD);
03620       I.setOperand(1, Constant::getNullValue(Op0->getType()));
03621       return &I;
03622     }
03623   }
03624 
03625   // Handle fcmp with constant RHS
03626   if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
03627     if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
03628       switch (LHSI->getOpcode()) {
03629       case Instruction::FPExt: {
03630         // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
03631         FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
03632         ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
03633         if (!RHSF)
03634           break;
03635 
03636         const fltSemantics *Sem;
03637         // FIXME: This shouldn't be here.
03638         if (LHSExt->getSrcTy()->isHalfTy())
03639           Sem = &APFloat::IEEEhalf;
03640         else if (LHSExt->getSrcTy()->isFloatTy())
03641           Sem = &APFloat::IEEEsingle;
03642         else if (LHSExt->getSrcTy()->isDoubleTy())
03643           Sem = &APFloat::IEEEdouble;
03644         else if (LHSExt->getSrcTy()->isFP128Ty())
03645           Sem = &APFloat::IEEEquad;
03646         else if (LHSExt->getSrcTy()->isX86_FP80Ty())
03647           Sem = &APFloat::x87DoubleExtended;
03648         else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
03649           Sem = &APFloat::PPCDoubleDouble;
03650         else
03651           break;
03652 
03653         bool Lossy;
03654         APFloat F = RHSF->getValueAPF();
03655         F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
03656 
03657         // Avoid lossy conversions and denormals. Zero is a special case
03658         // that's OK to convert.
03659         APFloat Fabs = F;
03660         Fabs.clearSign();
03661         if (!Lossy &&
03662             ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
03663                  APFloat::cmpLessThan) || Fabs.isZero()))
03664 
03665           return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
03666                               ConstantFP::get(RHSC->getContext(), F));
03667         break;
03668       }
03669       case Instruction::PHI:
03670         // Only fold fcmp into the PHI if the phi and fcmp are in the same
03671         // block.  If in the same block, we're encouraging jump threading.  If
03672         // not, we are just pessimizing the code by making an i1 phi.
03673         if (LHSI->getParent() == I.getParent())
03674           if (Instruction *NV = FoldOpIntoPhi(I))
03675             return NV;
03676         break;
03677       case Instruction::SIToFP:
03678       case Instruction::UIToFP:
03679         if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
03680           return NV;
03681         break;
03682       case Instruction::FSub: {
03683         // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
03684         Value *Op;
03685         if (match(LHSI, m_FNeg(m_Value(Op))))
03686           return new FCmpInst(I.getSwappedPredicate(), Op,
03687                               ConstantExpr::getFNeg(RHSC));
03688         break;
03689       }
03690       case Instruction::Load:
03691         if (GetElementPtrInst *GEP =
03692             dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
03693           if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
03694             if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
03695                 !cast<LoadInst>(LHSI)->isVolatile())
03696               if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
03697                 return Res;
03698         }
03699         break;
03700       case Instruction::Call: {
03701         CallInst *CI = cast<CallInst>(LHSI);
03702         LibFunc::Func Func;
03703         // Various optimization for fabs compared with zero.
03704         if (RHSC->isNullValue() && CI->getCalledFunction() &&
03705             TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
03706             TLI->has(Func)) {
03707           if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
03708               Func == LibFunc::fabsl) {
03709             switch (I.getPredicate()) {
03710             default: break;
03711             // fabs(x) < 0 --> false
03712             case FCmpInst::FCMP_OLT:
03713               return ReplaceInstUsesWith(I, Builder->getFalse());
03714             // fabs(x) > 0 --> x != 0
03715             case FCmpInst::FCMP_OGT:
03716               return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
03717                                   RHSC);
03718             // fabs(x) <= 0 --> x == 0
03719             case FCmpInst::FCMP_OLE:
03720               return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
03721                                   RHSC);
03722             // fabs(x) >= 0 --> !isnan(x)
03723             case FCmpInst::FCMP_OGE:
03724               return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
03725                                   RHSC);
03726             // fabs(x) == 0 --> x == 0
03727             // fabs(x) != 0 --> x != 0
03728             case FCmpInst::FCMP_OEQ:
03729             case FCmpInst::FCMP_UEQ:
03730             case FCmpInst::FCMP_ONE:
03731             case FCmpInst::FCMP_UNE:
03732               return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
03733                                   RHSC);
03734             }
03735           }
03736         }
03737       }
03738       }
03739   }
03740 
03741   // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
03742   Value *X, *Y;
03743   if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
03744     return new FCmpInst(I.getSwappedPredicate(), X, Y);
03745 
03746   // fcmp (fpext x), (fpext y) -> fcmp x, y
03747   if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
03748     if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
03749       if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
03750         return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
03751                             RHSExt->getOperand(0));
03752 
03753   return Changed ? &I : nullptr;
03754 }