LLVM API Documentation

InstCombineAddSub.cpp
Go to the documentation of this file.
00001 //===- InstCombineAddSub.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 visit functions for add, fadd, sub, and fsub.
00011 //
00012 //===----------------------------------------------------------------------===//
00013 
00014 #include "InstCombine.h"
00015 #include "llvm/ADT/STLExtras.h"
00016 #include "llvm/Analysis/InstructionSimplify.h"
00017 #include "llvm/IR/DataLayout.h"
00018 #include "llvm/IR/GetElementPtrTypeIterator.h"
00019 #include "llvm/IR/PatternMatch.h"
00020 using namespace llvm;
00021 using namespace PatternMatch;
00022 
00023 #define DEBUG_TYPE "instcombine"
00024 
00025 namespace {
00026 
00027   /// Class representing coefficient of floating-point addend.
00028   /// This class needs to be highly efficient, which is especially true for
00029   /// the constructor. As of I write this comment, the cost of the default
00030   /// constructor is merely 4-byte-store-zero (Assuming compiler is able to
00031   /// perform write-merging).
00032   ///
00033   class FAddendCoef {
00034   public:
00035     // The constructor has to initialize a APFloat, which is uncessary for
00036     // most addends which have coefficient either 1 or -1. So, the constructor
00037     // is expensive. In order to avoid the cost of the constructor, we should
00038     // reuse some instances whenever possible. The pre-created instances
00039     // FAddCombine::Add[0-5] embodies this idea.
00040     //
00041     FAddendCoef() : IsFp(false), BufHasFpVal(false), IntVal(0) {}
00042     ~FAddendCoef();
00043 
00044     void set(short C) {
00045       assert(!insaneIntVal(C) && "Insane coefficient");
00046       IsFp = false; IntVal = C;
00047     }
00048 
00049     void set(const APFloat& C);
00050 
00051     void negate();
00052 
00053     bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
00054     Value *getValue(Type *) const;
00055 
00056     // If possible, don't define operator+/operator- etc because these
00057     // operators inevitably call FAddendCoef's constructor which is not cheap.
00058     void operator=(const FAddendCoef &A);
00059     void operator+=(const FAddendCoef &A);
00060     void operator-=(const FAddendCoef &A);
00061     void operator*=(const FAddendCoef &S);
00062 
00063     bool isOne() const { return isInt() && IntVal == 1; }
00064     bool isTwo() const { return isInt() && IntVal == 2; }
00065     bool isMinusOne() const { return isInt() && IntVal == -1; }
00066     bool isMinusTwo() const { return isInt() && IntVal == -2; }
00067 
00068   private:
00069     bool insaneIntVal(int V) { return V > 4 || V < -4; }
00070     APFloat *getFpValPtr(void)
00071       { return reinterpret_cast<APFloat*>(&FpValBuf.buffer[0]); }
00072     const APFloat *getFpValPtr(void) const
00073       { return reinterpret_cast<const APFloat*>(&FpValBuf.buffer[0]); }
00074 
00075     const APFloat &getFpVal(void) const {
00076       assert(IsFp && BufHasFpVal && "Incorret state");
00077       return *getFpValPtr();
00078     }
00079 
00080     APFloat &getFpVal(void) {
00081       assert(IsFp && BufHasFpVal && "Incorret state");
00082       return *getFpValPtr();
00083     }
00084 
00085     bool isInt() const { return !IsFp; }
00086 
00087     // If the coefficient is represented by an integer, promote it to a
00088     // floating point.
00089     void convertToFpType(const fltSemantics &Sem);
00090 
00091     // Construct an APFloat from a signed integer.
00092     // TODO: We should get rid of this function when APFloat can be constructed
00093     //       from an *SIGNED* integer.
00094     APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);
00095   private:
00096 
00097     bool IsFp;
00098 
00099     // True iff FpValBuf contains an instance of APFloat.
00100     bool BufHasFpVal;
00101 
00102     // The integer coefficient of an individual addend is either 1 or -1,
00103     // and we try to simplify at most 4 addends from neighboring at most
00104     // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
00105     // is overkill of this end.
00106     short IntVal;
00107 
00108     AlignedCharArrayUnion<APFloat> FpValBuf;
00109   };
00110 
00111   /// FAddend is used to represent floating-point addend. An addend is
00112   /// represented as <C, V>, where the V is a symbolic value, and C is a
00113   /// constant coefficient. A constant addend is represented as <C, 0>.
00114   ///
00115   class FAddend {
00116   public:
00117     FAddend() { Val = 0; }
00118 
00119     Value *getSymVal (void) const { return Val; }
00120     const FAddendCoef &getCoef(void) const { return Coeff; }
00121 
00122     bool isConstant() const { return Val == 0; }
00123     bool isZero() const { return Coeff.isZero(); }
00124 
00125     void set(short Coefficient, Value *V) { Coeff.set(Coefficient), Val = V; }
00126     void set(const APFloat& Coefficient, Value *V)
00127       { Coeff.set(Coefficient); Val = V; }
00128     void set(const ConstantFP* Coefficient, Value *V)
00129       { Coeff.set(Coefficient->getValueAPF()); Val = V; }
00130 
00131     void negate() { Coeff.negate(); }
00132 
00133     /// Drill down the U-D chain one step to find the definition of V, and
00134     /// try to break the definition into one or two addends.
00135     static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);
00136 
00137     /// Similar to FAddend::drillDownOneStep() except that the value being
00138     /// splitted is the addend itself.
00139     unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;
00140 
00141     void operator+=(const FAddend &T) {
00142       assert((Val == T.Val) && "Symbolic-values disagree");
00143       Coeff += T.Coeff;
00144     }
00145 
00146   private:
00147     void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }
00148 
00149     // This addend has the value of "Coeff * Val".
00150     Value *Val;
00151     FAddendCoef Coeff;
00152   };
00153 
00154   /// FAddCombine is the class for optimizing an unsafe fadd/fsub along
00155   /// with its neighboring at most two instructions.
00156   ///
00157   class FAddCombine {
00158   public:
00159     FAddCombine(InstCombiner::BuilderTy *B) : Builder(B), Instr(0) {}
00160     Value *simplify(Instruction *FAdd);
00161 
00162   private:
00163     typedef SmallVector<const FAddend*, 4> AddendVect;
00164 
00165     Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);
00166 
00167     Value *performFactorization(Instruction *I);
00168 
00169     /// Convert given addend to a Value
00170     Value *createAddendVal(const FAddend &A, bool& NeedNeg);
00171 
00172     /// Return the number of instructions needed to emit the N-ary addition.
00173     unsigned calcInstrNumber(const AddendVect& Vect);
00174     Value *createFSub(Value *Opnd0, Value *Opnd1);
00175     Value *createFAdd(Value *Opnd0, Value *Opnd1);
00176     Value *createFMul(Value *Opnd0, Value *Opnd1);
00177     Value *createFDiv(Value *Opnd0, Value *Opnd1);
00178     Value *createFNeg(Value *V);
00179     Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
00180     void createInstPostProc(Instruction *NewInst, bool NoNumber = false);
00181 
00182     InstCombiner::BuilderTy *Builder;
00183     Instruction *Instr;
00184 
00185   private:
00186      // Debugging stuff are clustered here.
00187     #ifndef NDEBUG
00188       unsigned CreateInstrNum;
00189       void initCreateInstNum() { CreateInstrNum = 0; }
00190       void incCreateInstNum() { CreateInstrNum++; }
00191     #else
00192       void initCreateInstNum() {}
00193       void incCreateInstNum() {}
00194     #endif
00195   };
00196 }
00197 
00198 //===----------------------------------------------------------------------===//
00199 //
00200 // Implementation of
00201 //    {FAddendCoef, FAddend, FAddition, FAddCombine}.
00202 //
00203 //===----------------------------------------------------------------------===//
00204 FAddendCoef::~FAddendCoef() {
00205   if (BufHasFpVal)
00206     getFpValPtr()->~APFloat();
00207 }
00208 
00209 void FAddendCoef::set(const APFloat& C) {
00210   APFloat *P = getFpValPtr();
00211 
00212   if (isInt()) {
00213     // As the buffer is meanless byte stream, we cannot call
00214     // APFloat::operator=().
00215     new(P) APFloat(C);
00216   } else
00217     *P = C;
00218 
00219   IsFp = BufHasFpVal = true;
00220 }
00221 
00222 void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
00223   if (!isInt())
00224     return;
00225 
00226   APFloat *P = getFpValPtr();
00227   if (IntVal > 0)
00228     new(P) APFloat(Sem, IntVal);
00229   else {
00230     new(P) APFloat(Sem, 0 - IntVal);
00231     P->changeSign();
00232   }
00233   IsFp = BufHasFpVal = true;
00234 }
00235 
00236 APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
00237   if (Val >= 0)
00238     return APFloat(Sem, Val);
00239 
00240   APFloat T(Sem, 0 - Val);
00241   T.changeSign();
00242 
00243   return T;
00244 }
00245 
00246 void FAddendCoef::operator=(const FAddendCoef &That) {
00247   if (That.isInt())
00248     set(That.IntVal);
00249   else
00250     set(That.getFpVal());
00251 }
00252 
00253 void FAddendCoef::operator+=(const FAddendCoef &That) {
00254   enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven;
00255   if (isInt() == That.isInt()) {
00256     if (isInt())
00257       IntVal += That.IntVal;
00258     else
00259       getFpVal().add(That.getFpVal(), RndMode);
00260     return;
00261   }
00262 
00263   if (isInt()) {
00264     const APFloat &T = That.getFpVal();
00265     convertToFpType(T.getSemantics());
00266     getFpVal().add(T, RndMode);
00267     return;
00268   }
00269 
00270   APFloat &T = getFpVal();
00271   T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
00272 }
00273 
00274 void FAddendCoef::operator-=(const FAddendCoef &That) {
00275   enum APFloat::roundingMode RndMode = APFloat::rmNearestTiesToEven;
00276   if (isInt() == That.isInt()) {
00277     if (isInt())
00278       IntVal -= That.IntVal;
00279     else
00280       getFpVal().subtract(That.getFpVal(), RndMode);
00281     return;
00282   }
00283 
00284   if (isInt()) {
00285     const APFloat &T = That.getFpVal();
00286     convertToFpType(T.getSemantics());
00287     getFpVal().subtract(T, RndMode);
00288     return;
00289   }
00290 
00291   APFloat &T = getFpVal();
00292   T.subtract(createAPFloatFromInt(T.getSemantics(), IntVal), RndMode);
00293 }
00294 
00295 void FAddendCoef::operator*=(const FAddendCoef &That) {
00296   if (That.isOne())
00297     return;
00298 
00299   if (That.isMinusOne()) {
00300     negate();
00301     return;
00302   }
00303 
00304   if (isInt() && That.isInt()) {
00305     int Res = IntVal * (int)That.IntVal;
00306     assert(!insaneIntVal(Res) && "Insane int value");
00307     IntVal = Res;
00308     return;
00309   }
00310 
00311   const fltSemantics &Semantic =
00312     isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();
00313 
00314   if (isInt())
00315     convertToFpType(Semantic);
00316   APFloat &F0 = getFpVal();
00317 
00318   if (That.isInt())
00319     F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
00320                 APFloat::rmNearestTiesToEven);
00321   else
00322     F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
00323 
00324   return;
00325 }
00326 
00327 void FAddendCoef::negate() {
00328   if (isInt())
00329     IntVal = 0 - IntVal;
00330   else
00331     getFpVal().changeSign();
00332 }
00333 
00334 Value *FAddendCoef::getValue(Type *Ty) const {
00335   return isInt() ?
00336     ConstantFP::get(Ty, float(IntVal)) :
00337     ConstantFP::get(Ty->getContext(), getFpVal());
00338 }
00339 
00340 // The definition of <Val>     Addends
00341 // =========================================
00342 //  A + B                     <1, A>, <1,B>
00343 //  A - B                     <1, A>, <1,B>
00344 //  0 - B                     <-1, B>
00345 //  C * A,                    <C, A>
00346 //  A + C                     <1, A> <C, NULL>
00347 //  0 +/- 0                   <0, NULL> (corner case)
00348 //
00349 // Legend: A and B are not constant, C is constant
00350 //
00351 unsigned FAddend::drillValueDownOneStep
00352   (Value *Val, FAddend &Addend0, FAddend &Addend1) {
00353   Instruction *I = 0;
00354   if (Val == 0 || !(I = dyn_cast<Instruction>(Val)))
00355     return 0;
00356 
00357   unsigned Opcode = I->getOpcode();
00358 
00359   if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
00360     ConstantFP *C0, *C1;
00361     Value *Opnd0 = I->getOperand(0);
00362     Value *Opnd1 = I->getOperand(1);
00363     if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
00364       Opnd0 = 0;
00365 
00366     if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
00367       Opnd1 = 0;
00368 
00369     if (Opnd0) {
00370       if (!C0)
00371         Addend0.set(1, Opnd0);
00372       else
00373         Addend0.set(C0, 0);
00374     }
00375 
00376     if (Opnd1) {
00377       FAddend &Addend = Opnd0 ? Addend1 : Addend0;
00378       if (!C1)
00379         Addend.set(1, Opnd1);
00380       else
00381         Addend.set(C1, 0);
00382       if (Opcode == Instruction::FSub)
00383         Addend.negate();
00384     }
00385 
00386     if (Opnd0 || Opnd1)
00387       return Opnd0 && Opnd1 ? 2 : 1;
00388 
00389     // Both operands are zero. Weird!
00390     Addend0.set(APFloat(C0->getValueAPF().getSemantics()), 0);
00391     return 1;
00392   }
00393 
00394   if (I->getOpcode() == Instruction::FMul) {
00395     Value *V0 = I->getOperand(0);
00396     Value *V1 = I->getOperand(1);
00397     if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
00398       Addend0.set(C, V1);
00399       return 1;
00400     }
00401 
00402     if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
00403       Addend0.set(C, V0);
00404       return 1;
00405     }
00406   }
00407 
00408   return 0;
00409 }
00410 
00411 // Try to break *this* addend into two addends. e.g. Suppose this addend is
00412 // <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
00413 // i.e. <2.3, X> and <2.3, Y>.
00414 //
00415 unsigned FAddend::drillAddendDownOneStep
00416   (FAddend &Addend0, FAddend &Addend1) const {
00417   if (isConstant())
00418     return 0;
00419 
00420   unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
00421   if (!BreakNum || Coeff.isOne())
00422     return BreakNum;
00423 
00424   Addend0.Scale(Coeff);
00425 
00426   if (BreakNum == 2)
00427     Addend1.Scale(Coeff);
00428 
00429   return BreakNum;
00430 }
00431 
00432 // Try to perform following optimization on the input instruction I. Return the
00433 // simplified expression if was successful; otherwise, return 0.
00434 //
00435 //   Instruction "I" is                Simplified into
00436 // -------------------------------------------------------
00437 //   (x * y) +/- (x * z)               x * (y +/- z)
00438 //   (y / x) +/- (z / x)               (y +/- z) / x
00439 //
00440 Value *FAddCombine::performFactorization(Instruction *I) {
00441   assert((I->getOpcode() == Instruction::FAdd ||
00442           I->getOpcode() == Instruction::FSub) && "Expect add/sub");
00443 
00444   Instruction *I0 = dyn_cast<Instruction>(I->getOperand(0));
00445   Instruction *I1 = dyn_cast<Instruction>(I->getOperand(1));
00446 
00447   if (!I0 || !I1 || I0->getOpcode() != I1->getOpcode())
00448     return 0;
00449 
00450   bool isMpy = false;
00451   if (I0->getOpcode() == Instruction::FMul)
00452     isMpy = true;
00453   else if (I0->getOpcode() != Instruction::FDiv)
00454     return 0;
00455 
00456   Value *Opnd0_0 = I0->getOperand(0);
00457   Value *Opnd0_1 = I0->getOperand(1);
00458   Value *Opnd1_0 = I1->getOperand(0);
00459   Value *Opnd1_1 = I1->getOperand(1);
00460 
00461   //  Input Instr I       Factor   AddSub0  AddSub1
00462   //  ----------------------------------------------
00463   // (x*y) +/- (x*z)        x        y         z
00464   // (y/x) +/- (z/x)        x        y         z
00465   //
00466   Value *Factor = 0;
00467   Value *AddSub0 = 0, *AddSub1 = 0;
00468 
00469   if (isMpy) {
00470     if (Opnd0_0 == Opnd1_0 || Opnd0_0 == Opnd1_1)
00471       Factor = Opnd0_0;
00472     else if (Opnd0_1 == Opnd1_0 || Opnd0_1 == Opnd1_1)
00473       Factor = Opnd0_1;
00474 
00475     if (Factor) {
00476       AddSub0 = (Factor == Opnd0_0) ? Opnd0_1 : Opnd0_0;
00477       AddSub1 = (Factor == Opnd1_0) ? Opnd1_1 : Opnd1_0;
00478     }
00479   } else if (Opnd0_1 == Opnd1_1) {
00480     Factor = Opnd0_1;
00481     AddSub0 = Opnd0_0;
00482     AddSub1 = Opnd1_0;
00483   }
00484 
00485   if (!Factor)
00486     return 0;
00487 
00488   FastMathFlags Flags;
00489   Flags.setUnsafeAlgebra();
00490   if (I0) Flags &= I->getFastMathFlags();
00491   if (I1) Flags &= I->getFastMathFlags();
00492 
00493   // Create expression "NewAddSub = AddSub0 +/- AddsSub1"
00494   Value *NewAddSub = (I->getOpcode() == Instruction::FAdd) ?
00495                       createFAdd(AddSub0, AddSub1) :
00496                       createFSub(AddSub0, AddSub1);
00497   if (ConstantFP *CFP = dyn_cast<ConstantFP>(NewAddSub)) {
00498     const APFloat &F = CFP->getValueAPF();
00499     if (!F.isNormal())
00500       return 0;
00501   } else if (Instruction *II = dyn_cast<Instruction>(NewAddSub))
00502     II->setFastMathFlags(Flags);
00503 
00504   if (isMpy) {
00505     Value *RI = createFMul(Factor, NewAddSub);
00506     if (Instruction *II = dyn_cast<Instruction>(RI))
00507       II->setFastMathFlags(Flags);
00508     return RI;
00509   }
00510 
00511   Value *RI = createFDiv(NewAddSub, Factor);
00512   if (Instruction *II = dyn_cast<Instruction>(RI))
00513     II->setFastMathFlags(Flags);
00514   return RI;
00515 }
00516 
00517 Value *FAddCombine::simplify(Instruction *I) {
00518   assert(I->hasUnsafeAlgebra() && "Should be in unsafe mode");
00519 
00520   // Currently we are not able to handle vector type.
00521   if (I->getType()->isVectorTy())
00522     return 0;
00523 
00524   assert((I->getOpcode() == Instruction::FAdd ||
00525           I->getOpcode() == Instruction::FSub) && "Expect add/sub");
00526 
00527   // Save the instruction before calling other member-functions.
00528   Instr = I;
00529 
00530   FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;
00531 
00532   unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);
00533 
00534   // Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
00535   unsigned Opnd0_ExpNum = 0;
00536   unsigned Opnd1_ExpNum = 0;
00537 
00538   if (!Opnd0.isConstant())
00539     Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);
00540 
00541   // Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
00542   if (OpndNum == 2 && !Opnd1.isConstant())
00543     Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);
00544 
00545   // Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
00546   if (Opnd0_ExpNum && Opnd1_ExpNum) {
00547     AddendVect AllOpnds;
00548     AllOpnds.push_back(&Opnd0_0);
00549     AllOpnds.push_back(&Opnd1_0);
00550     if (Opnd0_ExpNum == 2)
00551       AllOpnds.push_back(&Opnd0_1);
00552     if (Opnd1_ExpNum == 2)
00553       AllOpnds.push_back(&Opnd1_1);
00554 
00555     // Compute instruction quota. We should save at least one instruction.
00556     unsigned InstQuota = 0;
00557 
00558     Value *V0 = I->getOperand(0);
00559     Value *V1 = I->getOperand(1);
00560     InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
00561                  (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;
00562 
00563     if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
00564       return R;
00565   }
00566 
00567   if (OpndNum != 2) {
00568     // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
00569     // splitted into two addends, say "V = X - Y", the instruction would have
00570     // been optimized into "I = Y - X" in the previous steps.
00571     //
00572     const FAddendCoef &CE = Opnd0.getCoef();
00573     return CE.isOne() ? Opnd0.getSymVal() : 0;
00574   }
00575 
00576   // step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
00577   if (Opnd1_ExpNum) {
00578     AddendVect AllOpnds;
00579     AllOpnds.push_back(&Opnd0);
00580     AllOpnds.push_back(&Opnd1_0);
00581     if (Opnd1_ExpNum == 2)
00582       AllOpnds.push_back(&Opnd1_1);
00583 
00584     if (Value *R = simplifyFAdd(AllOpnds, 1))
00585       return R;
00586   }
00587 
00588   // step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
00589   if (Opnd0_ExpNum) {
00590     AddendVect AllOpnds;
00591     AllOpnds.push_back(&Opnd1);
00592     AllOpnds.push_back(&Opnd0_0);
00593     if (Opnd0_ExpNum == 2)
00594       AllOpnds.push_back(&Opnd0_1);
00595 
00596     if (Value *R = simplifyFAdd(AllOpnds, 1))
00597       return R;
00598   }
00599 
00600   // step 6: Try factorization as the last resort,
00601   return performFactorization(I);
00602 }
00603 
00604 Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
00605 
00606   unsigned AddendNum = Addends.size();
00607   assert(AddendNum <= 4 && "Too many addends");
00608 
00609   // For saving intermediate results;
00610   unsigned NextTmpIdx = 0;
00611   FAddend TmpResult[3];
00612 
00613   // Points to the constant addend of the resulting simplified expression.
00614   // If the resulting expr has constant-addend, this constant-addend is
00615   // desirable to reside at the top of the resulting expression tree. Placing
00616   // constant close to supper-expr(s) will potentially reveal some optimization
00617   // opportunities in super-expr(s).
00618   //
00619   const FAddend *ConstAdd = 0;
00620 
00621   // Simplified addends are placed <SimpVect>.
00622   AddendVect SimpVect;
00623 
00624   // The outer loop works on one symbolic-value at a time. Suppose the input
00625   // addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
00626   // The symbolic-values will be processed in this order: x, y, z.
00627   //
00628   for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {
00629 
00630     const FAddend *ThisAddend = Addends[SymIdx];
00631     if (!ThisAddend) {
00632       // This addend was processed before.
00633       continue;
00634     }
00635 
00636     Value *Val = ThisAddend->getSymVal();
00637     unsigned StartIdx = SimpVect.size();
00638     SimpVect.push_back(ThisAddend);
00639 
00640     // The inner loop collects addends sharing same symbolic-value, and these
00641     // addends will be later on folded into a single addend. Following above
00642     // example, if the symbolic value "y" is being processed, the inner loop
00643     // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
00644     // be later on folded into "<b1+b2, y>".
00645     //
00646     for (unsigned SameSymIdx = SymIdx + 1;
00647          SameSymIdx < AddendNum; SameSymIdx++) {
00648       const FAddend *T = Addends[SameSymIdx];
00649       if (T && T->getSymVal() == Val) {
00650         // Set null such that next iteration of the outer loop will not process
00651         // this addend again.
00652         Addends[SameSymIdx] = 0;
00653         SimpVect.push_back(T);
00654       }
00655     }
00656 
00657     // If multiple addends share same symbolic value, fold them together.
00658     if (StartIdx + 1 != SimpVect.size()) {
00659       FAddend &R = TmpResult[NextTmpIdx ++];
00660       R = *SimpVect[StartIdx];
00661       for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
00662         R += *SimpVect[Idx];
00663 
00664       // Pop all addends being folded and push the resulting folded addend.
00665       SimpVect.resize(StartIdx);
00666       if (Val != 0) {
00667         if (!R.isZero()) {
00668           SimpVect.push_back(&R);
00669         }
00670       } else {
00671         // Don't push constant addend at this time. It will be the last element
00672         // of <SimpVect>.
00673         ConstAdd = &R;
00674       }
00675     }
00676   }
00677 
00678   assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&
00679          "out-of-bound access");
00680 
00681   if (ConstAdd)
00682     SimpVect.push_back(ConstAdd);
00683 
00684   Value *Result;
00685   if (!SimpVect.empty())
00686     Result = createNaryFAdd(SimpVect, InstrQuota);
00687   else {
00688     // The addition is folded to 0.0.
00689     Result = ConstantFP::get(Instr->getType(), 0.0);
00690   }
00691 
00692   return Result;
00693 }
00694 
00695 Value *FAddCombine::createNaryFAdd
00696   (const AddendVect &Opnds, unsigned InstrQuota) {
00697   assert(!Opnds.empty() && "Expect at least one addend");
00698 
00699   // Step 1: Check if the # of instructions needed exceeds the quota.
00700   //
00701   unsigned InstrNeeded = calcInstrNumber(Opnds);
00702   if (InstrNeeded > InstrQuota)
00703     return 0;
00704 
00705   initCreateInstNum();
00706 
00707   // step 2: Emit the N-ary addition.
00708   // Note that at most three instructions are involved in Fadd-InstCombine: the
00709   // addition in question, and at most two neighboring instructions.
00710   // The resulting optimized addition should have at least one less instruction
00711   // than the original addition expression tree. This implies that the resulting
00712   // N-ary addition has at most two instructions, and we don't need to worry
00713   // about tree-height when constructing the N-ary addition.
00714 
00715   Value *LastVal = 0;
00716   bool LastValNeedNeg = false;
00717 
00718   // Iterate the addends, creating fadd/fsub using adjacent two addends.
00719   for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end();
00720        I != E; I++) {
00721     bool NeedNeg;
00722     Value *V = createAddendVal(**I, NeedNeg);
00723     if (!LastVal) {
00724       LastVal = V;
00725       LastValNeedNeg = NeedNeg;
00726       continue;
00727     }
00728 
00729     if (LastValNeedNeg == NeedNeg) {
00730       LastVal = createFAdd(LastVal, V);
00731       continue;
00732     }
00733 
00734     if (LastValNeedNeg)
00735       LastVal = createFSub(V, LastVal);
00736     else
00737       LastVal = createFSub(LastVal, V);
00738 
00739     LastValNeedNeg = false;
00740   }
00741 
00742   if (LastValNeedNeg) {
00743     LastVal = createFNeg(LastVal);
00744   }
00745 
00746   #ifndef NDEBUG
00747     assert(CreateInstrNum == InstrNeeded &&
00748            "Inconsistent in instruction numbers");
00749   #endif
00750 
00751   return LastVal;
00752 }
00753 
00754 Value *FAddCombine::createFSub
00755   (Value *Opnd0, Value *Opnd1) {
00756   Value *V = Builder->CreateFSub(Opnd0, Opnd1);
00757   if (Instruction *I = dyn_cast<Instruction>(V))
00758     createInstPostProc(I);
00759   return V;
00760 }
00761 
00762 Value *FAddCombine::createFNeg(Value *V) {
00763   Value *Zero = cast<Value>(ConstantFP::get(V->getType(), 0.0));
00764   Value *NewV = createFSub(Zero, V);
00765   if (Instruction *I = dyn_cast<Instruction>(NewV))
00766     createInstPostProc(I, true); // fneg's don't receive instruction numbers.
00767   return NewV;
00768 }
00769 
00770 Value *FAddCombine::createFAdd
00771   (Value *Opnd0, Value *Opnd1) {
00772   Value *V = Builder->CreateFAdd(Opnd0, Opnd1);
00773   if (Instruction *I = dyn_cast<Instruction>(V))
00774     createInstPostProc(I);
00775   return V;
00776 }
00777 
00778 Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
00779   Value *V = Builder->CreateFMul(Opnd0, Opnd1);
00780   if (Instruction *I = dyn_cast<Instruction>(V))
00781     createInstPostProc(I);
00782   return V;
00783 }
00784 
00785 Value *FAddCombine::createFDiv(Value *Opnd0, Value *Opnd1) {
00786   Value *V = Builder->CreateFDiv(Opnd0, Opnd1);
00787   if (Instruction *I = dyn_cast<Instruction>(V))
00788     createInstPostProc(I);
00789   return V;
00790 }
00791 
00792 void FAddCombine::createInstPostProc(Instruction *NewInstr,
00793                                      bool NoNumber) {
00794   NewInstr->setDebugLoc(Instr->getDebugLoc());
00795 
00796   // Keep track of the number of instruction created.
00797   if (!NoNumber)
00798     incCreateInstNum();
00799 
00800   // Propagate fast-math flags
00801   NewInstr->setFastMathFlags(Instr->getFastMathFlags());
00802 }
00803 
00804 // Return the number of instruction needed to emit the N-ary addition.
00805 // NOTE: Keep this function in sync with createAddendVal().
00806 unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
00807   unsigned OpndNum = Opnds.size();
00808   unsigned InstrNeeded = OpndNum - 1;
00809 
00810   // The number of addends in the form of "(-1)*x".
00811   unsigned NegOpndNum = 0;
00812 
00813   // Adjust the number of instructions needed to emit the N-ary add.
00814   for (AddendVect::const_iterator I = Opnds.begin(), E = Opnds.end();
00815        I != E; I++) {
00816     const FAddend *Opnd = *I;
00817     if (Opnd->isConstant())
00818       continue;
00819 
00820     const FAddendCoef &CE = Opnd->getCoef();
00821     if (CE.isMinusOne() || CE.isMinusTwo())
00822       NegOpndNum++;
00823 
00824     // Let the addend be "c * x". If "c == +/-1", the value of the addend
00825     // is immediately available; otherwise, it needs exactly one instruction
00826     // to evaluate the value.
00827     if (!CE.isMinusOne() && !CE.isOne())
00828       InstrNeeded++;
00829   }
00830   if (NegOpndNum == OpndNum)
00831     InstrNeeded++;
00832   return InstrNeeded;
00833 }
00834 
00835 // Input Addend        Value           NeedNeg(output)
00836 // ================================================================
00837 // Constant C          C               false
00838 // <+/-1, V>           V               coefficient is -1
00839 // <2/-2, V>          "fadd V, V"      coefficient is -2
00840 // <C, V>             "fmul V, C"      false
00841 //
00842 // NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
00843 Value *FAddCombine::createAddendVal
00844   (const FAddend &Opnd, bool &NeedNeg) {
00845   const FAddendCoef &Coeff = Opnd.getCoef();
00846 
00847   if (Opnd.isConstant()) {
00848     NeedNeg = false;
00849     return Coeff.getValue(Instr->getType());
00850   }
00851 
00852   Value *OpndVal = Opnd.getSymVal();
00853 
00854   if (Coeff.isMinusOne() || Coeff.isOne()) {
00855     NeedNeg = Coeff.isMinusOne();
00856     return OpndVal;
00857   }
00858 
00859   if (Coeff.isTwo() || Coeff.isMinusTwo()) {
00860     NeedNeg = Coeff.isMinusTwo();
00861     return createFAdd(OpndVal, OpndVal);
00862   }
00863 
00864   NeedNeg = false;
00865   return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
00866 }
00867 
00868 // dyn_castFoldableMul - If this value is a multiply that can be folded into
00869 // other computations (because it has a constant operand), return the
00870 // non-constant operand of the multiply, and set CST to point to the multiplier.
00871 // Otherwise, return null.
00872 //
00873 static inline Value *dyn_castFoldableMul(Value *V, Constant *&CST) {
00874   if (!V->hasOneUse() || !V->getType()->isIntOrIntVectorTy())
00875     return 0;
00876 
00877   Instruction *I = dyn_cast<Instruction>(V);
00878   if (I == 0) return 0;
00879 
00880   if (I->getOpcode() == Instruction::Mul)
00881     if ((CST = dyn_cast<Constant>(I->getOperand(1))))
00882       return I->getOperand(0);
00883   if (I->getOpcode() == Instruction::Shl)
00884     if ((CST = dyn_cast<Constant>(I->getOperand(1)))) {
00885       // The multiplier is really 1 << CST.
00886       CST = ConstantExpr::getShl(ConstantInt::get(V->getType(), 1), CST);
00887       return I->getOperand(0);
00888     }
00889   return 0;
00890 }
00891 
00892 
00893 /// WillNotOverflowSignedAdd - Return true if we can prove that:
00894 ///    (sext (add LHS, RHS))  === (add (sext LHS), (sext RHS))
00895 /// This basically requires proving that the add in the original type would not
00896 /// overflow to change the sign bit or have a carry out.
00897 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
00898   // There are different heuristics we can use for this.  Here are some simple
00899   // ones.
00900 
00901   // Add has the property that adding any two 2's complement numbers can only
00902   // have one carry bit which can change a sign.  As such, if LHS and RHS each
00903   // have at least two sign bits, we know that the addition of the two values
00904   // will sign extend fine.
00905   if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
00906     return true;
00907 
00908 
00909   // If one of the operands only has one non-zero bit, and if the other operand
00910   // has a known-zero bit in a more significant place than it (not including the
00911   // sign bit) the ripple may go up to and fill the zero, but won't change the
00912   // sign.  For example, (X & ~4) + 1.
00913 
00914   // TODO: Implement.
00915 
00916   return false;
00917 }
00918 
00919 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
00920   bool Changed = SimplifyAssociativeOrCommutative(I);
00921   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
00922 
00923   if (Value *V = SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(),
00924                                  I.hasNoUnsignedWrap(), DL))
00925     return ReplaceInstUsesWith(I, V);
00926 
00927   // (A*B)+(A*C) -> A*(B+C) etc
00928   if (Value *V = SimplifyUsingDistributiveLaws(I))
00929     return ReplaceInstUsesWith(I, V);
00930 
00931   if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
00932     // X + (signbit) --> X ^ signbit
00933     const APInt &Val = CI->getValue();
00934     if (Val.isSignBit())
00935       return BinaryOperator::CreateXor(LHS, RHS);
00936 
00937     // See if SimplifyDemandedBits can simplify this.  This handles stuff like
00938     // (X & 254)+1 -> (X&254)|1
00939     if (SimplifyDemandedInstructionBits(I))
00940       return &I;
00941 
00942     // zext(bool) + C -> bool ? C + 1 : C
00943     if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
00944       if (ZI->getSrcTy()->isIntegerTy(1))
00945         return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
00946 
00947     Value *XorLHS = 0; ConstantInt *XorRHS = 0;
00948     if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
00949       uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
00950       const APInt &RHSVal = CI->getValue();
00951       unsigned ExtendAmt = 0;
00952       // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
00953       // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
00954       if (XorRHS->getValue() == -RHSVal) {
00955         if (RHSVal.isPowerOf2())
00956           ExtendAmt = TySizeBits - RHSVal.logBase2() - 1;
00957         else if (XorRHS->getValue().isPowerOf2())
00958           ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1;
00959       }
00960 
00961       if (ExtendAmt) {
00962         APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt);
00963         if (!MaskedValueIsZero(XorLHS, Mask))
00964           ExtendAmt = 0;
00965       }
00966 
00967       if (ExtendAmt) {
00968         Constant *ShAmt = ConstantInt::get(I.getType(), ExtendAmt);
00969         Value *NewShl = Builder->CreateShl(XorLHS, ShAmt, "sext");
00970         return BinaryOperator::CreateAShr(NewShl, ShAmt);
00971       }
00972 
00973       // If this is a xor that was canonicalized from a sub, turn it back into
00974       // a sub and fuse this add with it.
00975       if (LHS->hasOneUse() && (XorRHS->getValue()+1).isPowerOf2()) {
00976         IntegerType *IT = cast<IntegerType>(I.getType());
00977         APInt LHSKnownOne(IT->getBitWidth(), 0);
00978         APInt LHSKnownZero(IT->getBitWidth(), 0);
00979         ComputeMaskedBits(XorLHS, LHSKnownZero, LHSKnownOne);
00980         if ((XorRHS->getValue() | LHSKnownZero).isAllOnesValue())
00981           return BinaryOperator::CreateSub(ConstantExpr::getAdd(XorRHS, CI),
00982                                            XorLHS);
00983       }
00984       // (X + signbit) + C could have gotten canonicalized to (X ^ signbit) + C,
00985       // transform them into (X + (signbit ^ C))
00986       if (XorRHS->getValue().isSignBit())
00987           return BinaryOperator::CreateAdd(XorLHS,
00988                                            ConstantExpr::getXor(XorRHS, CI));
00989     }
00990   }
00991 
00992   if (isa<Constant>(RHS) && isa<PHINode>(LHS))
00993     if (Instruction *NV = FoldOpIntoPhi(I))
00994       return NV;
00995 
00996   if (I.getType()->getScalarType()->isIntegerTy(1))
00997     return BinaryOperator::CreateXor(LHS, RHS);
00998 
00999   // X + X --> X << 1
01000   if (LHS == RHS) {
01001     BinaryOperator *New =
01002       BinaryOperator::CreateShl(LHS, ConstantInt::get(I.getType(), 1));
01003     New->setHasNoSignedWrap(I.hasNoSignedWrap());
01004     New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
01005     return New;
01006   }
01007 
01008   // -A + B  -->  B - A
01009   // -A + -B  -->  -(A + B)
01010   if (Value *LHSV = dyn_castNegVal(LHS)) {
01011     if (!isa<Constant>(RHS))
01012       if (Value *RHSV = dyn_castNegVal(RHS)) {
01013         Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum");
01014         return BinaryOperator::CreateNeg(NewAdd);
01015       }
01016 
01017     return BinaryOperator::CreateSub(RHS, LHSV);
01018   }
01019 
01020   // A + -B  -->  A - B
01021   if (!isa<Constant>(RHS))
01022     if (Value *V = dyn_castNegVal(RHS))
01023       return BinaryOperator::CreateSub(LHS, V);
01024 
01025 
01026   {
01027     Constant *C2;
01028     if (Value *X = dyn_castFoldableMul(LHS, C2)) {
01029       if (X == RHS) // X*C + X --> X * (C+1)
01030         return BinaryOperator::CreateMul(RHS, AddOne(C2));
01031 
01032       // X*C1 + X*C2 --> X * (C1+C2)
01033       Constant *C1;
01034       if (X == dyn_castFoldableMul(RHS, C1))
01035         return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2));
01036     }
01037 
01038     // X + X*C --> X * (C+1)
01039     if (dyn_castFoldableMul(RHS, C2) == LHS)
01040       return BinaryOperator::CreateMul(LHS, AddOne(C2));
01041   }
01042 
01043   // A+B --> A|B iff A and B have no bits set in common.
01044   if (IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
01045     APInt LHSKnownOne(IT->getBitWidth(), 0);
01046     APInt LHSKnownZero(IT->getBitWidth(), 0);
01047     ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne);
01048     if (LHSKnownZero != 0) {
01049       APInt RHSKnownOne(IT->getBitWidth(), 0);
01050       APInt RHSKnownZero(IT->getBitWidth(), 0);
01051       ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne);
01052 
01053       // No bits in common -> bitwise or.
01054       if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
01055         return BinaryOperator::CreateOr(LHS, RHS);
01056     }
01057   }
01058 
01059   // W*X + Y*Z --> W * (X+Z)  iff W == Y
01060   {
01061     Value *W, *X, *Y, *Z;
01062     if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
01063         match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
01064       if (W != Y) {
01065         if (W == Z) {
01066           std::swap(Y, Z);
01067         } else if (Y == X) {
01068           std::swap(W, X);
01069         } else if (X == Z) {
01070           std::swap(Y, Z);
01071           std::swap(W, X);
01072         }
01073       }
01074 
01075       if (W == Y) {
01076         Value *NewAdd = Builder->CreateAdd(X, Z, LHS->getName());
01077         return BinaryOperator::CreateMul(W, NewAdd);
01078       }
01079     }
01080   }
01081 
01082   if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
01083     Value *X;
01084     if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
01085       return BinaryOperator::CreateSub(SubOne(CRHS), X);
01086   }
01087 
01088   if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
01089     // (X & FF00) + xx00  -> (X+xx00) & FF00
01090     Value *X;
01091     ConstantInt *C2;
01092     if (LHS->hasOneUse() &&
01093         match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) &&
01094         CRHS->getValue() == (CRHS->getValue() & C2->getValue())) {
01095       // See if all bits from the first bit set in the Add RHS up are included
01096       // in the mask.  First, get the rightmost bit.
01097       const APInt &AddRHSV = CRHS->getValue();
01098 
01099       // Form a mask of all bits from the lowest bit added through the top.
01100       APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
01101 
01102       // See if the and mask includes all of these bits.
01103       APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
01104 
01105       if (AddRHSHighBits == AddRHSHighBitsAnd) {
01106         // Okay, the xform is safe.  Insert the new add pronto.
01107         Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName());
01108         return BinaryOperator::CreateAnd(NewAdd, C2);
01109       }
01110     }
01111 
01112     // Try to fold constant add into select arguments.
01113     if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
01114       if (Instruction *R = FoldOpIntoSelect(I, SI))
01115         return R;
01116   }
01117 
01118   // add (select X 0 (sub n A)) A  -->  select X A n
01119   {
01120     SelectInst *SI = dyn_cast<SelectInst>(LHS);
01121     Value *A = RHS;
01122     if (!SI) {
01123       SI = dyn_cast<SelectInst>(RHS);
01124       A = LHS;
01125     }
01126     if (SI && SI->hasOneUse()) {
01127       Value *TV = SI->getTrueValue();
01128       Value *FV = SI->getFalseValue();
01129       Value *N;
01130 
01131       // Can we fold the add into the argument of the select?
01132       // We check both true and false select arguments for a matching subtract.
01133       if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
01134         // Fold the add into the true select value.
01135         return SelectInst::Create(SI->getCondition(), N, A);
01136 
01137       if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
01138         // Fold the add into the false select value.
01139         return SelectInst::Create(SI->getCondition(), A, N);
01140     }
01141   }
01142 
01143   // Check for (add (sext x), y), see if we can merge this into an
01144   // integer add followed by a sext.
01145   if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
01146     // (add (sext x), cst) --> (sext (add x, cst'))
01147     if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
01148       Constant *CI =
01149         ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
01150       if (LHSConv->hasOneUse() &&
01151           ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
01152           WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
01153         // Insert the new, smaller add.
01154         Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
01155                                               CI, "addconv");
01156         return new SExtInst(NewAdd, I.getType());
01157       }
01158     }
01159 
01160     // (add (sext x), (sext y)) --> (sext (add int x, y))
01161     if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
01162       // Only do this if x/y have the same type, if at last one of them has a
01163       // single use (so we don't increase the number of sexts), and if the
01164       // integer add will not overflow.
01165       if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
01166           (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
01167           WillNotOverflowSignedAdd(LHSConv->getOperand(0),
01168                                    RHSConv->getOperand(0))) {
01169         // Insert the new integer add.
01170         Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
01171                                              RHSConv->getOperand(0), "addconv");
01172         return new SExtInst(NewAdd, I.getType());
01173       }
01174     }
01175   }
01176 
01177   // Check for (x & y) + (x ^ y)
01178   {
01179     Value *A = 0, *B = 0;
01180     if (match(RHS, m_Xor(m_Value(A), m_Value(B))) &&
01181         (match(LHS, m_And(m_Specific(A), m_Specific(B))) ||
01182          match(LHS, m_And(m_Specific(B), m_Specific(A)))))
01183       return BinaryOperator::CreateOr(A, B);
01184 
01185     if (match(LHS, m_Xor(m_Value(A), m_Value(B))) &&
01186         (match(RHS, m_And(m_Specific(A), m_Specific(B))) ||
01187          match(RHS, m_And(m_Specific(B), m_Specific(A)))))
01188       return BinaryOperator::CreateOr(A, B);
01189   }
01190 
01191   return Changed ? &I : 0;
01192 }
01193 
01194 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
01195   bool Changed = SimplifyAssociativeOrCommutative(I);
01196   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
01197 
01198   if (Value *V = SimplifyFAddInst(LHS, RHS, I.getFastMathFlags(), DL))
01199     return ReplaceInstUsesWith(I, V);
01200 
01201   if (isa<Constant>(RHS)) {
01202     if (isa<PHINode>(LHS))
01203       if (Instruction *NV = FoldOpIntoPhi(I))
01204         return NV;
01205 
01206     if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
01207       if (Instruction *NV = FoldOpIntoSelect(I, SI))
01208         return NV;
01209   }
01210 
01211   // -A + B  -->  B - A
01212   // -A + -B  -->  -(A + B)
01213   if (Value *LHSV = dyn_castFNegVal(LHS)) {
01214     Instruction *RI = BinaryOperator::CreateFSub(RHS, LHSV);
01215     RI->copyFastMathFlags(&I);
01216     return RI;
01217   }
01218 
01219   // A + -B  -->  A - B
01220   if (!isa<Constant>(RHS))
01221     if (Value *V = dyn_castFNegVal(RHS)) {
01222       Instruction *RI = BinaryOperator::CreateFSub(LHS, V);
01223       RI->copyFastMathFlags(&I);
01224       return RI;
01225     }
01226 
01227   // Check for (fadd double (sitofp x), y), see if we can merge this into an
01228   // integer add followed by a promotion.
01229   if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
01230     // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
01231     // ... if the constant fits in the integer value.  This is useful for things
01232     // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
01233     // requires a constant pool load, and generally allows the add to be better
01234     // instcombined.
01235     if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
01236       Constant *CI =
01237       ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
01238       if (LHSConv->hasOneUse() &&
01239           ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
01240           WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
01241         // Insert the new integer add.
01242         Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
01243                                               CI, "addconv");
01244         return new SIToFPInst(NewAdd, I.getType());
01245       }
01246     }
01247 
01248     // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
01249     if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
01250       // Only do this if x/y have the same type, if at last one of them has a
01251       // single use (so we don't increase the number of int->fp conversions),
01252       // and if the integer add will not overflow.
01253       if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
01254           (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
01255           WillNotOverflowSignedAdd(LHSConv->getOperand(0),
01256                                    RHSConv->getOperand(0))) {
01257         // Insert the new integer add.
01258         Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
01259                                               RHSConv->getOperand(0),"addconv");
01260         return new SIToFPInst(NewAdd, I.getType());
01261       }
01262     }
01263   }
01264 
01265   // select C, 0, B + select C, A, 0 -> select C, A, B
01266   {
01267     Value *A1, *B1, *C1, *A2, *B2, *C2;
01268     if (match(LHS, m_Select(m_Value(C1), m_Value(A1), m_Value(B1))) &&
01269         match(RHS, m_Select(m_Value(C2), m_Value(A2), m_Value(B2)))) {
01270       if (C1 == C2) {
01271         Constant *Z1=0, *Z2=0;
01272         Value *A, *B, *C=C1;
01273         if (match(A1, m_AnyZero()) && match(B2, m_AnyZero())) {
01274             Z1 = dyn_cast<Constant>(A1); A = A2;
01275             Z2 = dyn_cast<Constant>(B2); B = B1;
01276         } else if (match(B1, m_AnyZero()) && match(A2, m_AnyZero())) {
01277             Z1 = dyn_cast<Constant>(B1); B = B2;
01278             Z2 = dyn_cast<Constant>(A2); A = A1; 
01279         }
01280         
01281         if (Z1 && Z2 && 
01282             (I.hasNoSignedZeros() || 
01283              (Z1->isNegativeZeroValue() && Z2->isNegativeZeroValue()))) {
01284           return SelectInst::Create(C, A, B);
01285         }
01286       }
01287     }
01288   }
01289 
01290   if (I.hasUnsafeAlgebra()) {
01291     if (Value *V = FAddCombine(Builder).simplify(&I))
01292       return ReplaceInstUsesWith(I, V);
01293   }
01294 
01295   return Changed ? &I : 0;
01296 }
01297 
01298 
01299 /// Optimize pointer differences into the same array into a size.  Consider:
01300 ///  &A[10] - &A[0]: we should compile this to "10".  LHS/RHS are the pointer
01301 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
01302 ///
01303 Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS,
01304                                                Type *Ty) {
01305   assert(DL && "Must have target data info for this");
01306 
01307   // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
01308   // this.
01309   bool Swapped = false;
01310   GEPOperator *GEP1 = 0, *GEP2 = 0;
01311 
01312   // For now we require one side to be the base pointer "A" or a constant
01313   // GEP derived from it.
01314   if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
01315     // (gep X, ...) - X
01316     if (LHSGEP->getOperand(0) == RHS) {
01317       GEP1 = LHSGEP;
01318       Swapped = false;
01319     } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
01320       // (gep X, ...) - (gep X, ...)
01321       if (LHSGEP->getOperand(0)->stripPointerCasts() ==
01322             RHSGEP->getOperand(0)->stripPointerCasts()) {
01323         GEP2 = RHSGEP;
01324         GEP1 = LHSGEP;
01325         Swapped = false;
01326       }
01327     }
01328   }
01329 
01330   if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
01331     // X - (gep X, ...)
01332     if (RHSGEP->getOperand(0) == LHS) {
01333       GEP1 = RHSGEP;
01334       Swapped = true;
01335     } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
01336       // (gep X, ...) - (gep X, ...)
01337       if (RHSGEP->getOperand(0)->stripPointerCasts() ==
01338             LHSGEP->getOperand(0)->stripPointerCasts()) {
01339         GEP2 = LHSGEP;
01340         GEP1 = RHSGEP;
01341         Swapped = true;
01342       }
01343     }
01344   }
01345 
01346   // Avoid duplicating the arithmetic if GEP2 has non-constant indices and
01347   // multiple users.
01348   if (GEP1 == 0 ||
01349       (GEP2 != 0 && !GEP2->hasAllConstantIndices() && !GEP2->hasOneUse()))
01350     return 0;
01351 
01352   // Emit the offset of the GEP and an intptr_t.
01353   Value *Result = EmitGEPOffset(GEP1);
01354 
01355   // If we had a constant expression GEP on the other side offsetting the
01356   // pointer, subtract it from the offset we have.
01357   if (GEP2) {
01358     Value *Offset = EmitGEPOffset(GEP2);
01359     Result = Builder->CreateSub(Result, Offset);
01360   }
01361 
01362   // If we have p - gep(p, ...)  then we have to negate the result.
01363   if (Swapped)
01364     Result = Builder->CreateNeg(Result, "diff.neg");
01365 
01366   return Builder->CreateIntCast(Result, Ty, true);
01367 }
01368 
01369 
01370 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
01371   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
01372 
01373   if (Value *V = SimplifySubInst(Op0, Op1, I.hasNoSignedWrap(),
01374                                  I.hasNoUnsignedWrap(), DL))
01375     return ReplaceInstUsesWith(I, V);
01376 
01377   // (A*B)-(A*C) -> A*(B-C) etc
01378   if (Value *V = SimplifyUsingDistributiveLaws(I))
01379     return ReplaceInstUsesWith(I, V);
01380 
01381   // If this is a 'B = x-(-A)', change to B = x+A.  This preserves NSW/NUW.
01382   if (Value *V = dyn_castNegVal(Op1)) {
01383     BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
01384     Res->setHasNoSignedWrap(I.hasNoSignedWrap());
01385     Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
01386     return Res;
01387   }
01388 
01389   if (I.getType()->isIntegerTy(1))
01390     return BinaryOperator::CreateXor(Op0, Op1);
01391 
01392   // Replace (-1 - A) with (~A).
01393   if (match(Op0, m_AllOnes()))
01394     return BinaryOperator::CreateNot(Op1);
01395 
01396   if (Constant *C = dyn_cast<Constant>(Op0)) {
01397     // C - ~X == X + (1+C)
01398     Value *X = 0;
01399     if (match(Op1, m_Not(m_Value(X))))
01400       return BinaryOperator::CreateAdd(X, AddOne(C));
01401 
01402     // Try to fold constant sub into select arguments.
01403     if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
01404       if (Instruction *R = FoldOpIntoSelect(I, SI))
01405         return R;
01406 
01407     // C-(X+C2) --> (C-C2)-X
01408     Constant *C2;
01409     if (match(Op1, m_Add(m_Value(X), m_Constant(C2))))
01410       return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
01411 
01412     if (SimplifyDemandedInstructionBits(I))
01413       return &I;
01414 
01415     // Fold (sub 0, (zext bool to B)) --> (sext bool to B)
01416     if (C->isNullValue() && match(Op1, m_ZExt(m_Value(X))))
01417       if (X->getType()->getScalarType()->isIntegerTy(1))
01418         return CastInst::CreateSExtOrBitCast(X, Op1->getType());
01419 
01420     // Fold (sub 0, (sext bool to B)) --> (zext bool to B)
01421     if (C->isNullValue() && match(Op1, m_SExt(m_Value(X))))
01422       if (X->getType()->getScalarType()->isIntegerTy(1))
01423         return CastInst::CreateZExtOrBitCast(X, Op1->getType());
01424   }
01425 
01426   if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
01427     // -(X >>u 31) -> (X >>s 31)
01428     // -(X >>s 31) -> (X >>u 31)
01429     if (C->isZero()) {
01430       Value *X; ConstantInt *CI;
01431       if (match(Op1, m_LShr(m_Value(X), m_ConstantInt(CI))) &&
01432           // Verify we are shifting out everything but the sign bit.
01433           CI->getValue() == I.getType()->getPrimitiveSizeInBits()-1)
01434         return BinaryOperator::CreateAShr(X, CI);
01435 
01436       if (match(Op1, m_AShr(m_Value(X), m_ConstantInt(CI))) &&
01437           // Verify we are shifting out everything but the sign bit.
01438           CI->getValue() == I.getType()->getPrimitiveSizeInBits()-1)
01439         return BinaryOperator::CreateLShr(X, CI);
01440     }
01441   }
01442 
01443 
01444   { Value *Y;
01445     // X-(X+Y) == -Y    X-(Y+X) == -Y
01446     if (match(Op1, m_Add(m_Specific(Op0), m_Value(Y))) ||
01447         match(Op1, m_Add(m_Value(Y), m_Specific(Op0))))
01448       return BinaryOperator::CreateNeg(Y);
01449 
01450     // (X-Y)-X == -Y
01451     if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
01452       return BinaryOperator::CreateNeg(Y);
01453   }
01454 
01455   if (Op1->hasOneUse()) {
01456     Value *X = 0, *Y = 0, *Z = 0;
01457     Constant *C = 0;
01458     Constant *CI = 0;
01459 
01460     // (X - (Y - Z))  -->  (X + (Z - Y)).
01461     if (match(Op1, m_Sub(m_Value(Y), m_Value(Z))))
01462       return BinaryOperator::CreateAdd(Op0,
01463                                       Builder->CreateSub(Z, Y, Op1->getName()));
01464 
01465     // (X - (X & Y))   -->   (X & ~Y)
01466     //
01467     if (match(Op1, m_And(m_Value(Y), m_Specific(Op0))) ||
01468         match(Op1, m_And(m_Specific(Op0), m_Value(Y))))
01469       return BinaryOperator::CreateAnd(Op0,
01470                                   Builder->CreateNot(Y, Y->getName() + ".not"));
01471 
01472     // 0 - (X sdiv C)  -> (X sdiv -C)
01473     if (match(Op1, m_SDiv(m_Value(X), m_Constant(C))) &&
01474         match(Op0, m_Zero()))
01475       return BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(C));
01476 
01477     // 0 - (X << Y)  -> (-X << Y)   when X is freely negatable.
01478     if (match(Op1, m_Shl(m_Value(X), m_Value(Y))) && match(Op0, m_Zero()))
01479       if (Value *XNeg = dyn_castNegVal(X))
01480         return BinaryOperator::CreateShl(XNeg, Y);
01481 
01482     // X - X*C --> X * (1-C)
01483     if (match(Op1, m_Mul(m_Specific(Op0), m_Constant(CI)))) {
01484       Constant *CP1 = ConstantExpr::getSub(ConstantInt::get(I.getType(),1), CI);
01485       return BinaryOperator::CreateMul(Op0, CP1);
01486     }
01487 
01488     // X - X<<C --> X * (1-(1<<C))
01489     if (match(Op1, m_Shl(m_Specific(Op0), m_Constant(CI)))) {
01490       Constant *One = ConstantInt::get(I.getType(), 1);
01491       C = ConstantExpr::getSub(One, ConstantExpr::getShl(One, CI));
01492       return BinaryOperator::CreateMul(Op0, C);
01493     }
01494 
01495     // X - A*-B -> X + A*B
01496     // X - -A*B -> X + A*B
01497     Value *A, *B;
01498     if (match(Op1, m_Mul(m_Value(A), m_Neg(m_Value(B)))) ||
01499         match(Op1, m_Mul(m_Neg(m_Value(A)), m_Value(B))))
01500       return BinaryOperator::CreateAdd(Op0, Builder->CreateMul(A, B));
01501 
01502     // X - A*CI -> X + A*-CI
01503     // X - CI*A -> X + A*-CI
01504     if (match(Op1, m_Mul(m_Value(A), m_Constant(CI))) ||
01505         match(Op1, m_Mul(m_Constant(CI), m_Value(A)))) {
01506       Value *NewMul = Builder->CreateMul(A, ConstantExpr::getNeg(CI));
01507       return BinaryOperator::CreateAdd(Op0, NewMul);
01508     }
01509   }
01510 
01511   Constant *C1;
01512   if (Value *X = dyn_castFoldableMul(Op0, C1)) {
01513     if (X == Op1)  // X*C - X --> X * (C-1)
01514       return BinaryOperator::CreateMul(Op1, SubOne(C1));
01515 
01516     Constant *C2;   // X*C1 - X*C2 -> X * (C1-C2)
01517     if (X == dyn_castFoldableMul(Op1, C2))
01518       return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2));
01519   }
01520 
01521   // Optimize pointer differences into the same array into a size.  Consider:
01522   //  &A[10] - &A[0]: we should compile this to "10".
01523   if (DL) {
01524     Value *LHSOp, *RHSOp;
01525     if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
01526         match(Op1, m_PtrToInt(m_Value(RHSOp))))
01527       if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
01528         return ReplaceInstUsesWith(I, Res);
01529 
01530     // trunc(p)-trunc(q) -> trunc(p-q)
01531     if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
01532         match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
01533       if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
01534         return ReplaceInstUsesWith(I, Res);
01535   }
01536 
01537   return 0;
01538 }
01539 
01540 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
01541   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
01542 
01543   if (Value *V = SimplifyFSubInst(Op0, Op1, I.getFastMathFlags(), DL))
01544     return ReplaceInstUsesWith(I, V);
01545 
01546   if (isa<Constant>(Op0))
01547     if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
01548       if (Instruction *NV = FoldOpIntoSelect(I, SI))
01549         return NV;
01550 
01551   // If this is a 'B = x-(-A)', change to B = x+A, potentially looking
01552   // through FP extensions/truncations along the way.
01553   if (Value *V = dyn_castFNegVal(Op1)) {
01554     Instruction *NewI = BinaryOperator::CreateFAdd(Op0, V);
01555     NewI->copyFastMathFlags(&I);
01556     return NewI;
01557   }
01558   if (FPTruncInst *FPTI = dyn_cast<FPTruncInst>(Op1)) {
01559     if (Value *V = dyn_castFNegVal(FPTI->getOperand(0))) {
01560       Value *NewTrunc = Builder->CreateFPTrunc(V, I.getType());
01561       Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewTrunc);
01562       NewI->copyFastMathFlags(&I);
01563       return NewI;
01564     }
01565   } else if (FPExtInst *FPEI = dyn_cast<FPExtInst>(Op1)) {
01566     if (Value *V = dyn_castFNegVal(FPEI->getOperand(0))) {
01567       Value *NewExt = Builder->CreateFPExt(V, I.getType());
01568       Instruction *NewI = BinaryOperator::CreateFAdd(Op0, NewExt);
01569       NewI->copyFastMathFlags(&I);
01570       return NewI;
01571     }
01572   }
01573 
01574   if (I.hasUnsafeAlgebra()) {
01575     if (Value *V = FAddCombine(Builder).simplify(&I))
01576       return ReplaceInstUsesWith(I, V);
01577   }
01578 
01579   return 0;
01580 }