LLVM API Documentation

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