LLVM API Documentation

ValueTracking.cpp

Go to the documentation of this file.
00001 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
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 contains routines that help analyze properties that chains of
00011 // computations have.
00012 //
00013 //===----------------------------------------------------------------------===//
00014 
00015 #include "llvm/Analysis/ValueTracking.h"
00016 #include "llvm/Constants.h"
00017 #include "llvm/Instructions.h"
00018 #include "llvm/GlobalVariable.h"
00019 #include "llvm/IntrinsicInst.h"
00020 #include "llvm/LLVMContext.h"
00021 #include "llvm/Operator.h"
00022 #include "llvm/Target/TargetData.h"
00023 #include "llvm/Support/GetElementPtrTypeIterator.h"
00024 #include "llvm/Support/MathExtras.h"
00025 #include <cstring>
00026 using namespace llvm;
00027 
00028 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
00029 /// known to be either zero or one and return them in the KnownZero/KnownOne
00030 /// bit sets.  This code only analyzes bits in Mask, in order to short-circuit
00031 /// processing.
00032 /// NOTE: we cannot consider 'undef' to be "IsZero" here.  The problem is that
00033 /// we cannot optimize based on the assumption that it is zero without changing
00034 /// it to be an explicit zero.  If we don't change it to zero, other code could
00035 /// optimized based on the contradictory assumption that it is non-zero.
00036 /// Because instcombine aggressively folds operations with undef args anyway,
00037 /// this won't lose us code quality.
00038 void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
00039                              APInt &KnownZero, APInt &KnownOne,
00040                              TargetData *TD, unsigned Depth) {
00041   const unsigned MaxDepth = 6;
00042   assert(V && "No Value?");
00043   assert(Depth <= MaxDepth && "Limit Search Depth");
00044   unsigned BitWidth = Mask.getBitWidth();
00045   assert((V->getType()->isIntOrIntVector() || isa<PointerType>(V->getType())) &&
00046          "Not integer or pointer type!");
00047   assert((!TD ||
00048           TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
00049          (!V->getType()->isIntOrIntVector() ||
00050           V->getType()->getScalarSizeInBits() == BitWidth) &&
00051          KnownZero.getBitWidth() == BitWidth && 
00052          KnownOne.getBitWidth() == BitWidth &&
00053          "V, Mask, KnownOne and KnownZero should have same BitWidth");
00054 
00055   if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
00056     // We know all of the bits for a constant!
00057     KnownOne = CI->getValue() & Mask;
00058     KnownZero = ~KnownOne & Mask;
00059     return;
00060   }
00061   // Null and aggregate-zero are all-zeros.
00062   if (isa<ConstantPointerNull>(V) ||
00063       isa<ConstantAggregateZero>(V)) {
00064     KnownOne.clear();
00065     KnownZero = Mask;
00066     return;
00067   }
00068   // Handle a constant vector by taking the intersection of the known bits of
00069   // each element.
00070   if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
00071     KnownZero.set(); KnownOne.set();
00072     for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
00073       APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
00074       ComputeMaskedBits(CV->getOperand(i), Mask, KnownZero2, KnownOne2,
00075                         TD, Depth);
00076       KnownZero &= KnownZero2;
00077       KnownOne &= KnownOne2;
00078     }
00079     return;
00080   }
00081   // The address of an aligned GlobalValue has trailing zeros.
00082   if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
00083     unsigned Align = GV->getAlignment();
00084     if (Align == 0 && TD && GV->getType()->getElementType()->isSized()) {
00085       const Type *ObjectType = GV->getType()->getElementType();
00086       // If the object is defined in the current Module, we'll be giving
00087       // it the preferred alignment. Otherwise, we have to assume that it
00088       // may only have the minimum ABI alignment.
00089       if (!GV->isDeclaration() && !GV->mayBeOverridden())
00090         Align = TD->getPrefTypeAlignment(ObjectType);
00091       else
00092         Align = TD->getABITypeAlignment(ObjectType);
00093     }
00094     if (Align > 0)
00095       KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
00096                                               CountTrailingZeros_32(Align));
00097     else
00098       KnownZero.clear();
00099     KnownOne.clear();
00100     return;
00101   }
00102 
00103   KnownZero.clear(); KnownOne.clear();   // Start out not knowing anything.
00104 
00105   if (Depth == MaxDepth || Mask == 0)
00106     return;  // Limit search depth.
00107 
00108   Operator *I = dyn_cast<Operator>(V);
00109   if (!I) return;
00110 
00111   APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
00112   switch (I->getOpcode()) {
00113   default: break;
00114   case Instruction::And: {
00115     // If either the LHS or the RHS are Zero, the result is zero.
00116     ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
00117     APInt Mask2(Mask & ~KnownZero);
00118     ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
00119                       Depth+1);
00120     assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 
00121     assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 
00122     
00123     // Output known-1 bits are only known if set in both the LHS & RHS.
00124     KnownOne &= KnownOne2;
00125     // Output known-0 are known to be clear if zero in either the LHS | RHS.
00126     KnownZero |= KnownZero2;
00127     return;
00128   }
00129   case Instruction::Or: {
00130     ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
00131     APInt Mask2(Mask & ~KnownOne);
00132     ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
00133                       Depth+1);
00134     assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 
00135     assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 
00136     
00137     // Output known-0 bits are only known if clear in both the LHS & RHS.
00138     KnownZero &= KnownZero2;
00139     // Output known-1 are known to be set if set in either the LHS | RHS.
00140     KnownOne |= KnownOne2;
00141     return;
00142   }
00143   case Instruction::Xor: {
00144     ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
00145     ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, TD,
00146                       Depth+1);
00147     assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 
00148     assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 
00149     
00150     // Output known-0 bits are known if clear or set in both the LHS & RHS.
00151     APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
00152     // Output known-1 are known to be set if set in only one of the LHS, RHS.
00153     KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
00154     KnownZero = KnownZeroOut;
00155     return;
00156   }
00157   case Instruction::Mul: {
00158     APInt Mask2 = APInt::getAllOnesValue(BitWidth);
00159     ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
00160     ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
00161                       Depth+1);
00162     assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 
00163     assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 
00164     
00165     // If low bits are zero in either operand, output low known-0 bits.
00166     // Also compute a conserative estimate for high known-0 bits.
00167     // More trickiness is possible, but this is sufficient for the
00168     // interesting case of alignment computation.
00169     KnownOne.clear();
00170     unsigned TrailZ = KnownZero.countTrailingOnes() +
00171                       KnownZero2.countTrailingOnes();
00172     unsigned LeadZ =  std::max(KnownZero.countLeadingOnes() +
00173                                KnownZero2.countLeadingOnes(),
00174                                BitWidth) - BitWidth;
00175 
00176     TrailZ = std::min(TrailZ, BitWidth);
00177     LeadZ = std::min(LeadZ, BitWidth);
00178     KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
00179                 APInt::getHighBitsSet(BitWidth, LeadZ);
00180     KnownZero &= Mask;
00181     return;
00182   }
00183   case Instruction::UDiv: {
00184     // For the purposes of computing leading zeros we can conservatively
00185     // treat a udiv as a logical right shift by the power of 2 known to
00186     // be less than the denominator.
00187     APInt AllOnes = APInt::getAllOnesValue(BitWidth);
00188     ComputeMaskedBits(I->getOperand(0),
00189                       AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
00190     unsigned LeadZ = KnownZero2.countLeadingOnes();
00191 
00192     KnownOne2.clear();
00193     KnownZero2.clear();
00194     ComputeMaskedBits(I->getOperand(1),
00195                       AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
00196     unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
00197     if (RHSUnknownLeadingOnes != BitWidth)
00198       LeadZ = std::min(BitWidth,
00199                        LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
00200 
00201     KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
00202     return;
00203   }
00204   case Instruction::Select:
00205     ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, TD, Depth+1);
00206     ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, TD,
00207                       Depth+1);
00208     assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 
00209     assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 
00210 
00211     // Only known if known in both the LHS and RHS.
00212     KnownOne &= KnownOne2;
00213     KnownZero &= KnownZero2;
00214     return;
00215   case Instruction::FPTrunc:
00216   case Instruction::FPExt:
00217   case Instruction::FPToUI:
00218   case Instruction::FPToSI:
00219   case Instruction::SIToFP:
00220   case Instruction::UIToFP:
00221     return; // Can't work with floating point.
00222   case Instruction::PtrToInt:
00223   case Instruction::IntToPtr:
00224     // We can't handle these if we don't know the pointer size.
00225     if (!TD) return;
00226     // FALL THROUGH and handle them the same as zext/trunc.
00227   case Instruction::ZExt:
00228   case Instruction::Trunc: {
00229     // Note that we handle pointer operands here because of inttoptr/ptrtoint
00230     // which fall through here.
00231     const Type *SrcTy = I->getOperand(0)->getType();
00232     unsigned SrcBitWidth = TD ?
00233       TD->getTypeSizeInBits(SrcTy) :
00234       SrcTy->getScalarSizeInBits();
00235     APInt MaskIn(Mask);
00236     MaskIn.zextOrTrunc(SrcBitWidth);
00237     KnownZero.zextOrTrunc(SrcBitWidth);
00238     KnownOne.zextOrTrunc(SrcBitWidth);
00239     ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
00240                       Depth+1);
00241     KnownZero.zextOrTrunc(BitWidth);
00242     KnownOne.zextOrTrunc(BitWidth);
00243     // Any top bits are known to be zero.
00244     if (BitWidth > SrcBitWidth)
00245       KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
00246     return;
00247   }
00248   case Instruction::BitCast: {
00249     const Type *SrcTy = I->getOperand(0)->getType();
00250     if ((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
00251         // TODO: For now, not handling conversions like:
00252         // (bitcast i64 %x to <2 x i32>)
00253         !isa<VectorType>(I->getType())) {
00254       ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, TD,
00255                         Depth+1);
00256       return;
00257     }
00258     break;
00259   }
00260   case Instruction::SExt: {
00261     // Compute the bits in the result that are not present in the input.
00262     const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
00263     unsigned SrcBitWidth = SrcTy->getBitWidth();
00264       
00265     APInt MaskIn(Mask); 
00266     MaskIn.trunc(SrcBitWidth);
00267     KnownZero.trunc(SrcBitWidth);
00268     KnownOne.trunc(SrcBitWidth);
00269     ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
00270                       Depth+1);
00271     assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 
00272     KnownZero.zext(BitWidth);
00273     KnownOne.zext(BitWidth);
00274 
00275     // If the sign bit of the input is known set or clear, then we know the
00276     // top bits of the result.
00277     if (KnownZero[SrcBitWidth-1])             // Input sign bit known zero
00278       KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
00279     else if (KnownOne[SrcBitWidth-1])           // Input sign bit known set
00280       KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
00281     return;
00282   }
00283   case Instruction::Shl:
00284     // (shl X, C1) & C2 == 0   iff   (X & C2 >>u C1) == 0
00285     if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
00286       uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
00287       APInt Mask2(Mask.lshr(ShiftAmt));
00288       ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
00289                         Depth+1);
00290       assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 
00291       KnownZero <<= ShiftAmt;
00292       KnownOne  <<= ShiftAmt;
00293       KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
00294       return;
00295     }
00296     break;
00297   case Instruction::LShr:
00298     // (ushr X, C1) & C2 == 0   iff  (-1 >> C1) & C2 == 0
00299     if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
00300       // Compute the new bits that are at the top now.
00301       uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
00302       
00303       // Unsigned shift right.
00304       APInt Mask2(Mask.shl(ShiftAmt));
00305       ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne, TD,
00306                         Depth+1);
00307       assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?"); 
00308       KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
00309       KnownOne  = APIntOps::lshr(KnownOne, ShiftAmt);
00310       // high bits known zero.
00311       KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
00312       return;
00313     }
00314     break;
00315   case Instruction::AShr:
00316     // (ashr X, C1) & C2 == 0   iff  (-1 >> C1) & C2 == 0
00317     if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
00318       // Compute the new bits that are at the top now.
00319       uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
00320       
00321       // Signed shift right.
00322       APInt Mask2(Mask.shl(ShiftAmt));
00323       ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
00324                         Depth+1);
00325       assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?"); 
00326       KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
00327       KnownOne  = APIntOps::lshr(KnownOne, ShiftAmt);
00328         
00329       APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
00330       if (KnownZero[BitWidth-ShiftAmt-1])    // New bits are known zero.
00331         KnownZero |= HighBits;
00332       else if (KnownOne[BitWidth-ShiftAmt-1])  // New bits are known one.
00333         KnownOne |= HighBits;
00334       return;
00335     }
00336     break;
00337   case Instruction::Sub: {
00338     if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
00339       // We know that the top bits of C-X are clear if X contains less bits
00340       // than C (i.e. no wrap-around can happen).  For example, 20-X is
00341       // positive if we can prove that X is >= 0 and < 16.
00342       if (!CLHS->getValue().isNegative()) {
00343         unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
00344         // NLZ can't be BitWidth with no sign bit
00345         APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
00346         ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
00347                           TD, Depth+1);
00348     
00349         // If all of the MaskV bits are known to be zero, then we know the
00350         // output top bits are zero, because we now know that the output is
00351         // from [0-C].
00352         if ((KnownZero2 & MaskV) == MaskV) {
00353           unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
00354           // Top bits known zero.
00355           KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
00356         }
00357       }        
00358     }
00359   }
00360   // fall through
00361   case Instruction::Add: {
00362     // If one of the operands has trailing zeros, than the bits that the
00363     // other operand has in those bit positions will be preserved in the
00364     // result. For an add, this works with either operand. For a subtract,
00365     // this only works if the known zeros are in the right operand.
00366     APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
00367     APInt Mask2 = APInt::getLowBitsSet(BitWidth,
00368                                        BitWidth - Mask.countLeadingZeros());
00369     ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
00370                       Depth+1);
00371     assert((LHSKnownZero & LHSKnownOne) == 0 &&
00372            "Bits known to be one AND zero?");
00373     unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
00374 
00375     ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, TD, 
00376                       Depth+1);
00377     assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 
00378     unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
00379 
00380     // Determine which operand has more trailing zeros, and use that
00381     // many bits from the other operand.
00382     if (LHSKnownZeroOut > RHSKnownZeroOut) {
00383       if (I->getOpcode() == Instruction::Add) {
00384         APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
00385         KnownZero |= KnownZero2 & Mask;
00386         KnownOne  |= KnownOne2 & Mask;
00387       } else {
00388         // If the known zeros are in the left operand for a subtract,
00389         // fall back to the minimum known zeros in both operands.
00390         KnownZero |= APInt::getLowBitsSet(BitWidth,
00391                                           std::min(LHSKnownZeroOut,
00392                                                    RHSKnownZeroOut));
00393       }
00394     } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
00395       APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
00396       KnownZero |= LHSKnownZero & Mask;
00397       KnownOne  |= LHSKnownOne & Mask;
00398     }
00399     return;
00400   }
00401   case Instruction::SRem:
00402     if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
00403       APInt RA = Rem->getValue();
00404       if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
00405         APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
00406         APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
00407         ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD, 
00408                           Depth+1);
00409 
00410         // If the sign bit of the first operand is zero, the sign bit of
00411         // the result is zero. If the first operand has no one bits below
00412         // the second operand's single 1 bit, its sign will be zero.
00413         if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
00414           KnownZero2 |= ~LowBits;
00415 
00416         KnownZero |= KnownZero2 & Mask;
00417 
00418         assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?"); 
00419       }
00420     }
00421     break;
00422   case Instruction::URem: {
00423     if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
00424       APInt RA = Rem->getValue();
00425       if (RA.isPowerOf2()) {
00426         APInt LowBits = (RA - 1);
00427         APInt Mask2 = LowBits & Mask;
00428         KnownZero |= ~LowBits & Mask;
00429         ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
00430                           Depth+1);
00431         assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
00432         break;
00433       }
00434     }
00435 
00436     // Since the result is less than or equal to either operand, any leading
00437     // zero bits in either operand must also exist in the result.
00438     APInt AllOnes = APInt::getAllOnesValue(BitWidth);
00439     ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
00440                       TD, Depth+1);
00441     ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
00442                       TD, Depth+1);
00443 
00444     unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
00445                                 KnownZero2.countLeadingOnes());
00446     KnownOne.clear();
00447     KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
00448     break;
00449   }
00450 
00451   case Instruction::Alloca:
00452   case Instruction::Malloc: {
00453     AllocationInst *AI = cast<AllocationInst>(V);
00454     unsigned Align = AI->getAlignment();
00455     if (Align == 0 && TD) {
00456       if (isa<AllocaInst>(AI))
00457         Align = TD->getABITypeAlignment(AI->getType()->getElementType());
00458       else if (isa<MallocInst>(AI)) {
00459         // Malloc returns maximally aligned memory.
00460         Align = TD->getABITypeAlignment(AI->getType()->getElementType());
00461         Align =
00462           std::max(Align,
00463                    (unsigned)TD->getABITypeAlignment(
00464                      Type::getDoubleTy(V->getContext())));
00465         Align =
00466           std::max(Align,
00467                    (unsigned)TD->getABITypeAlignment(
00468                       Type::getInt64Ty(V->getContext())));
00469       }
00470     }
00471     
00472     if (Align > 0)
00473       KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
00474                                               CountTrailingZeros_32(Align));
00475     break;
00476   }
00477   case Instruction::GetElementPtr: {
00478     // Analyze all of the subscripts of this getelementptr instruction
00479     // to determine if we can prove known low zero bits.
00480     APInt LocalMask = APInt::getAllOnesValue(BitWidth);
00481     APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
00482     ComputeMaskedBits(I->getOperand(0), LocalMask,
00483                       LocalKnownZero, LocalKnownOne, TD, Depth+1);
00484     unsigned TrailZ = LocalKnownZero.countTrailingOnes();
00485 
00486     gep_type_iterator GTI = gep_type_begin(I);
00487     for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
00488       Value *Index = I->getOperand(i);
00489       if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
00490         // Handle struct member offset arithmetic.
00491         if (!TD) return;
00492         const StructLayout *SL = TD->getStructLayout(STy);
00493         unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
00494         uint64_t Offset = SL->getElementOffset(Idx);
00495         TrailZ = std::min(TrailZ,
00496                           CountTrailingZeros_64(Offset));
00497       } else {
00498         // Handle array index arithmetic.
00499         const Type *IndexedTy = GTI.getIndexedType();
00500         if (!IndexedTy->isSized()) return;
00501         unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
00502         uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
00503         LocalMask = APInt::getAllOnesValue(GEPOpiBits);
00504         LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
00505         ComputeMaskedBits(Index, LocalMask,
00506                           LocalKnownZero, LocalKnownOne, TD, Depth+1);
00507         TrailZ = std::min(TrailZ,
00508                           unsigned(CountTrailingZeros_64(TypeSize) +
00509                                    LocalKnownZero.countTrailingOnes()));
00510       }
00511     }
00512     
00513     KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
00514     break;
00515   }
00516   case Instruction::PHI: {
00517     PHINode *P = cast<PHINode>(I);
00518     // Handle the case of a simple two-predecessor recurrence PHI.
00519     // There's a lot more that could theoretically be done here, but
00520     // this is sufficient to catch some interesting cases.
00521     if (P->getNumIncomingValues() == 2) {
00522       for (unsigned i = 0; i != 2; ++i) {
00523         Value *L = P->getIncomingValue(i);
00524         Value *R = P->getIncomingValue(!i);
00525         Operator *LU = dyn_cast<Operator>(L);
00526         if (!LU)
00527           continue;
00528         unsigned Opcode = LU->getOpcode();
00529         // Check for operations that have the property that if
00530         // both their operands have low zero bits, the result
00531         // will have low zero bits.
00532         if (Opcode == Instruction::Add ||
00533             Opcode == Instruction::Sub ||
00534             Opcode == Instruction::And ||
00535             Opcode == Instruction::Or ||
00536             Opcode == Instruction::Mul) {
00537           Value *LL = LU->getOperand(0);
00538           Value *LR = LU->getOperand(1);
00539           // Find a recurrence.
00540           if (LL == I)
00541             L = LR;
00542           else if (LR == I)
00543             L = LL;
00544           else
00545             break;
00546           // Ok, we have a PHI of the form L op= R. Check for low
00547           // zero bits.
00548           APInt Mask2 = APInt::getAllOnesValue(BitWidth);
00549           ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
00550           Mask2 = APInt::getLowBitsSet(BitWidth,
00551                                        KnownZero2.countTrailingOnes());
00552 
00553           // We need to take the minimum number of known bits
00554           APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
00555           ComputeMaskedBits(L, Mask2, KnownZero3, KnownOne3, TD, Depth+1);
00556 
00557           KnownZero = Mask &
00558                       APInt::getLowBitsSet(BitWidth,
00559                                            std::min(KnownZero2.countTrailingOnes(),
00560                                                     KnownZero3.countTrailingOnes()));
00561           break;
00562         }
00563       }
00564     }
00565 
00566     // Otherwise take the unions of the known bit sets of the operands,
00567     // taking conservative care to avoid excessive recursion.
00568     if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
00569       KnownZero = APInt::getAllOnesValue(BitWidth);
00570       KnownOne = APInt::getAllOnesValue(BitWidth);
00571       for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
00572         // Skip direct self references.
00573         if (P->getIncomingValue(i) == P) continue;
00574 
00575         KnownZero2 = APInt(BitWidth, 0);
00576         KnownOne2 = APInt(BitWidth, 0);
00577         // Recurse, but cap the recursion to one level, because we don't
00578         // want to waste time spinning around in loops.
00579         ComputeMaskedBits(P->getIncomingValue(i), KnownZero | KnownOne,
00580                           KnownZero2, KnownOne2, TD, MaxDepth-1);
00581         KnownZero &= KnownZero2;
00582         KnownOne &= KnownOne2;
00583         // If all bits have been ruled out, there's no need to check
00584         // more operands.
00585         if (!KnownZero && !KnownOne)
00586           break;
00587       }
00588     }
00589     break;
00590   }
00591   case Instruction::Call:
00592     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
00593       switch (II->getIntrinsicID()) {
00594       default: break;
00595       case Intrinsic::ctpop:
00596       case Intrinsic::ctlz:
00597       case Intrinsic::cttz: {
00598         unsigned LowBits = Log2_32(BitWidth)+1;
00599         KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
00600         break;
00601       }
00602       }
00603     }
00604     break;
00605   }
00606 }
00607 
00608 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero.  We use
00609 /// this predicate to simplify operations downstream.  Mask is known to be zero
00610 /// for bits that V cannot have.
00611 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
00612                              TargetData *TD, unsigned Depth) {
00613   APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
00614   ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
00615   assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 
00616   return (KnownZero & Mask) == Mask;
00617 }
00618 
00619 
00620 
00621 /// ComputeNumSignBits - Return the number of times the sign bit of the
00622 /// register is replicated into the other bits.  We know that at least 1 bit
00623 /// is always equal to the sign bit (itself), but other cases can give us
00624 /// information.  For example, immediately after an "ashr X, 2", we know that
00625 /// the top 3 bits are all equal to each other, so we return 3.
00626 ///
00627 /// 'Op' must have a scalar integer type.
00628 ///
00629 unsigned llvm::ComputeNumSignBits(Value *V, TargetData *TD, unsigned Depth) {
00630   assert((TD || V->getType()->isIntOrIntVector()) &&
00631          "ComputeNumSignBits requires a TargetData object to operate "
00632          "on non-integer values!");
00633   const Type *Ty = V->getType();
00634   unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
00635                          Ty->getScalarSizeInBits();
00636   unsigned Tmp, Tmp2;
00637   unsigned FirstAnswer = 1;
00638 
00639   // Note that ConstantInt is handled by the general ComputeMaskedBits case
00640   // below.
00641 
00642   if (Depth == 6)
00643     return 1;  // Limit search depth.
00644   
00645   Operator *U = dyn_cast<Operator>(V);
00646   switch (Operator::getOpcode(V)) {
00647   default: break;
00648   case Instruction::SExt:
00649     Tmp = TyBits-cast<IntegerType>(U->getOperand(0)->getType())->getBitWidth();
00650     return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
00651     
00652   case Instruction::AShr:
00653     Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
00654     // ashr X, C   -> adds C sign bits.
00655     if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
00656       Tmp += C->getZExtValue();
00657       if (Tmp > TyBits) Tmp = TyBits;
00658     }
00659     return Tmp;
00660   case Instruction::Shl:
00661     if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
00662       // shl destroys sign bits.
00663       Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
00664       if (C->getZExtValue() >= TyBits ||      // Bad shift.
00665           C->getZExtValue() >= Tmp) break;    // Shifted all sign bits out.
00666       return Tmp - C->getZExtValue();
00667     }
00668     break;
00669   case Instruction::And:
00670   case Instruction::Or:
00671   case Instruction::Xor:    // NOT is handled here.
00672     // Logical binary ops preserve the number of sign bits at the worst.
00673     Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
00674     if (Tmp != 1) {
00675       Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
00676       FirstAnswer = std::min(Tmp, Tmp2);
00677       // We computed what we know about the sign bits as our first
00678       // answer. Now proceed to the generic code that uses
00679       // ComputeMaskedBits, and pick whichever answer is better.
00680     }
00681     break;
00682 
00683   case Instruction::Select:
00684     Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
00685     if (Tmp == 1) return 1;  // Early out.
00686     Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
00687     return std::min(Tmp, Tmp2);
00688     
00689   case Instruction::Add:
00690     // Add can have at most one carry bit.  Thus we know that the output
00691     // is, at worst, one more bit than the inputs.
00692     Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
00693     if (Tmp == 1) return 1;  // Early out.
00694       
00695     // Special case decrementing a value (ADD X, -1):
00696     if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
00697       if (CRHS->isAllOnesValue()) {
00698         APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
00699         APInt Mask = APInt::getAllOnesValue(TyBits);
00700         ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, TD,
00701                           Depth+1);
00702         
00703         // If the input is known to be 0 or 1, the output is 0/-1, which is all
00704         // sign bits set.
00705         if ((KnownZero | APInt(TyBits, 1)) == Mask)
00706           return TyBits;
00707         
00708         // If we are subtracting one from a positive number, there is no carry
00709         // out of the result.
00710         if (KnownZero.isNegative())
00711           return Tmp;
00712       }
00713       
00714     Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
00715     if (Tmp2 == 1) return 1;
00716       return std::min(Tmp, Tmp2)-1;
00717     break;
00718     
00719   case Instruction::Sub:
00720     Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
00721     if (Tmp2 == 1) return 1;
00722       
00723     // Handle NEG.
00724     if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
00725       if (CLHS->isNullValue()) {
00726         APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
00727         APInt Mask = APInt::getAllOnesValue(TyBits);
00728         ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne, 
00729                           TD, Depth+1);
00730         // If the input is known to be 0 or 1, the output is 0/-1, which is all
00731         // sign bits set.
00732         if ((KnownZero | APInt(TyBits, 1)) == Mask)
00733           return TyBits;
00734         
00735         // If the input is known to be positive (the sign bit is known clear),
00736         // the output of the NEG has the same number of sign bits as the input.
00737         if (KnownZero.isNegative())
00738           return Tmp2;
00739         
00740         // Otherwise, we treat this like a SUB.
00741       }
00742     
00743     // Sub can have at most one carry bit.  Thus we know that the output
00744     // is, at worst, one more bit than the inputs.
00745     Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
00746     if (Tmp == 1) return 1;  // Early out.
00747       return std::min(Tmp, Tmp2)-1;
00748     break;
00749   case Instruction::Trunc:
00750     // FIXME: it's tricky to do anything useful for this, but it is an important
00751     // case for targets like X86.
00752     break;
00753   }
00754   
00755   // Finally, if we can prove that the top bits of the result are 0's or 1's,
00756   // use this information.
00757   APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
00758   APInt Mask = APInt::getAllOnesValue(TyBits);
00759   ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
00760   
00761   if (KnownZero.isNegative()) {        // sign bit is 0
00762     Mask = KnownZero;
00763   } else if (KnownOne.isNegative()) {  // sign bit is 1;
00764     Mask = KnownOne;
00765   } else {
00766     // Nothing known.
00767     return FirstAnswer;
00768   }
00769   
00770   // Okay, we know that the sign bit in Mask is set.  Use CLZ to determine
00771   // the number of identical bits in the top of the input value.
00772   Mask = ~Mask;
00773   Mask <<= Mask.getBitWidth()-TyBits;
00774   // Return # leading zeros.  We use 'min' here in case Val was zero before
00775   // shifting.  We don't want to return '64' as for an i32 "0".
00776   return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
00777 }
00778 
00779 /// CannotBeNegativeZero - Return true if we can prove that the specified FP 
00780 /// value is never equal to -0.0.
00781 ///
00782 /// NOTE: this function will need to be revisited when we support non-default
00783 /// rounding modes!
00784 ///
00785 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
00786   if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
00787     return !CFP->getValueAPF().isNegZero();
00788   
00789   if (Depth == 6)
00790     return 1;  // Limit search depth.
00791 
00792   const Operator *I = dyn_cast<Operator>(V);
00793   if (I == 0) return false;
00794   
00795   // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
00796   if (I->getOpcode() == Instruction::FAdd &&
00797       isa<ConstantFP>(I->getOperand(1)) && 
00798       cast<ConstantFP>(I->getOperand(1))->isNullValue())
00799     return true;
00800     
00801   // sitofp and uitofp turn into +0.0 for zero.
00802   if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
00803     return true;
00804   
00805   if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
00806     // sqrt(-0.0) = -0.0, no other negative results are possible.
00807     if (II->getIntrinsicID() == Intrinsic::sqrt)
00808       return CannotBeNegativeZero(II->getOperand(1), Depth+1);
00809   
00810   if (const CallInst *CI = dyn_cast<CallInst>(I))
00811     if (const Function *F = CI->getCalledFunction()) {
00812       if (F->isDeclaration()) {
00813         // abs(x) != -0.0
00814         if (F->getName() == "abs") return true;
00815         // abs[lf](x) != -0.0
00816         if (F->getName() == "absf") return true;
00817         if (F->getName() == "absl") return true;
00818       }
00819     }
00820   
00821   return false;
00822 }
00823 
00824 // This is the recursive version of BuildSubAggregate. It takes a few different
00825 // arguments. Idxs is the index within the nested struct From that we are
00826 // looking at now (which is of type IndexedType). IdxSkip is the number of
00827 // indices from Idxs that should be left out when inserting into the resulting
00828 // struct. To is the result struct built so far, new insertvalue instructions
00829 // build on that.
00830 static Value *BuildSubAggregate(Value *From, Value* To, const Type *IndexedType,
00831                                 SmallVector<unsigned, 10> &Idxs,
00832                                 unsigned IdxSkip,
00833                                 LLVMContext &Context,
00834                                 Instruction *InsertBefore) {
00835   const llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
00836   if (STy) {
00837     // Save the original To argument so we can modify it
00838     Value *OrigTo = To;
00839     // General case, the type indexed by Idxs is a struct
00840     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
00841       // Process each struct element recursively
00842       Idxs.push_back(i);
00843       Value *PrevTo = To;
00844       To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
00845                              Context, InsertBefore);
00846       Idxs.pop_back();
00847       if (!To) {
00848         // Couldn't find any inserted value for this index? Cleanup
00849         while (PrevTo != OrigTo) {
00850           InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
00851           PrevTo = Del->getAggregateOperand();
00852           Del->eraseFromParent();
00853         }
00854         // Stop processing elements
00855         break;
00856       }
00857     }
00858     // If we succesfully found a value for each of our subaggregates 
00859     if (To)
00860       return To;
00861   }
00862   // Base case, the type indexed by SourceIdxs is not a struct, or not all of
00863   // the struct's elements had a value that was inserted directly. In the latter
00864   // case, perhaps we can't determine each of the subelements individually, but
00865   // we might be able to find the complete struct somewhere.
00866   
00867   // Find the value that is at that particular spot
00868   Value *V = FindInsertedValue(From, Idxs.begin(), Idxs.end(), Context);
00869 
00870   if (!V)
00871     return NULL;
00872 
00873   // Insert the value in the new (sub) aggregrate
00874   return llvm::InsertValueInst::Create(To, V, Idxs.begin() + IdxSkip,
00875                                        Idxs.end(), "tmp", InsertBefore);
00876 }
00877 
00878 // This helper takes a nested struct and extracts a part of it (which is again a
00879 // struct) into a new value. For example, given the struct:
00880 // { a, { b, { c, d }, e } }
00881 // and the indices "1, 1" this returns
00882 // { c, d }.
00883 //
00884 // It does this by inserting an insertvalue for each element in the resulting
00885 // struct, as opposed to just inserting a single struct. This will only work if
00886 // each of the elements of the substruct are known (ie, inserted into From by an
00887 // insertvalue instruction somewhere).
00888 //
00889 // All inserted insertvalue instructions are inserted before InsertBefore
00890 static Value *BuildSubAggregate(Value *From, const unsigned *idx_begin,
00891                                 const unsigned *idx_end, LLVMContext &Context,
00892                                 Instruction *InsertBefore) {
00893   assert(InsertBefore && "Must have someplace to insert!");
00894   const Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
00895                                                              idx_begin,
00896                                                              idx_end);
00897   Value *To = UndefValue::get(IndexedType);
00898   SmallVector<unsigned, 10> Idxs(idx_begin, idx_end);
00899   unsigned IdxSkip = Idxs.size();
00900 
00901   return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip,
00902                            Context, InsertBefore);
00903 }
00904 
00905 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
00906 /// the scalar value indexed is already around as a register, for example if it
00907 /// were inserted directly into the aggregrate.
00908 ///
00909 /// If InsertBefore is not null, this function will duplicate (modified)
00910 /// insertvalues when a part of a nested struct is extracted.
00911 Value *llvm::FindInsertedValue(Value *V, const unsigned *idx_begin,
00912                          const unsigned *idx_end, LLVMContext &Context,
00913                          Instruction *InsertBefore) {
00914   // Nothing to index? Just return V then (this is useful at the end of our
00915   // recursion)
00916   if (idx_begin == idx_end)
00917     return V;
00918   // We have indices, so V should have an indexable type
00919   assert((isa<StructType>(V->getType()) || isa<ArrayType>(V->getType()))
00920          && "Not looking at a struct or array?");
00921   assert(ExtractValueInst::getIndexedType(V->getType(), idx_begin, idx_end)
00922          && "Invalid indices for type?");
00923   const CompositeType *PTy = cast<CompositeType>(V->getType());
00924 
00925   if (isa<UndefValue>(V))
00926     return UndefValue::get(ExtractValueInst::getIndexedType(PTy,
00927                                                               idx_begin,
00928                                                               idx_end));
00929   else if (isa<ConstantAggregateZero>(V))
00930     return Constant::getNullValue(ExtractValueInst::getIndexedType(PTy, 
00931                                                                   idx_begin,
00932                                                                   idx_end));
00933   else if (Constant *C = dyn_cast<Constant>(V)) {
00934     if (isa<ConstantArray>(C) || isa<ConstantStruct>(C))
00935       // Recursively process this constant
00936       return FindInsertedValue(C->getOperand(*idx_begin), idx_begin + 1,
00937                                idx_end, Context, InsertBefore);
00938   } else if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
00939     // Loop the indices for the insertvalue instruction in parallel with the
00940     // requested indices
00941     const unsigned *req_idx = idx_begin;
00942     for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
00943          i != e; ++i, ++req_idx) {
00944       if (req_idx == idx_end) {
00945         if (InsertBefore)
00946           // The requested index identifies a part of a nested aggregate. Handle
00947           // this specially. For example,
00948           // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
00949           // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
00950           // %C = extractvalue {i32, { i32, i32 } } %B, 1
00951           // This can be changed into
00952           // %A = insertvalue {i32, i32 } undef, i32 10, 0
00953           // %C = insertvalue {i32, i32 } %A, i32 11, 1
00954           // which allows the unused 0,0 element from the nested struct to be
00955           // removed.
00956           return BuildSubAggregate(V, idx_begin, req_idx,
00957                                    Context, InsertBefore);
00958         else
00959           // We can't handle this without inserting insertvalues
00960           return 0;
00961       }
00962       
00963       // This insert value inserts something else than what we are looking for.
00964       // See if the (aggregrate) value inserted into has the value we are
00965       // looking for, then.
00966       if (*req_idx != *i)
00967         return FindInsertedValue(I->getAggregateOperand(), idx_begin, idx_end,
00968                                  Context, InsertBefore);
00969     }
00970     // If we end up here, the indices of the insertvalue match with those
00971     // requested (though possibly only partially). Now we recursively look at
00972     // the inserted value, passing any remaining indices.
00973     return FindInsertedValue(I->getInsertedValueOperand(), req_idx, idx_end,
00974                              Context, InsertBefore);
00975   } else if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
00976     // If we're extracting a value from an aggregrate that was extracted from
00977     // something else, we can extract from that something else directly instead.
00978     // However, we will need to chain I's indices with the requested indices.
00979    
00980     // Calculate the number of indices required 
00981     unsigned size = I->getNumIndices() + (idx_end - idx_begin);
00982     // Allocate some space to put the new indices in
00983     SmallVector<unsigned, 5> Idxs;
00984     Idxs.reserve(size);
00985     // Add indices from the extract value instruction
00986     for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
00987          i != e; ++i)
00988       Idxs.push_back(*i);
00989     
00990     // Add requested indices
00991     for (const unsigned *i = idx_begin, *e = idx_end; i != e; ++i)
00992       Idxs.push_back(*i);
00993 
00994     assert(Idxs.size() == size 
00995            && "Number of indices added not correct?");
00996     
00997     return FindInsertedValue(I->getAggregateOperand(), Idxs.begin(), Idxs.end(),
00998                              Context, InsertBefore);
00999   }
01000   // Otherwise, we don't know (such as, extracting from a function return value
01001   // or load instruction)
01002   return 0;
01003 }
01004 
01005 /// GetConstantStringInfo - This function computes the length of a
01006 /// null-terminated C string pointed to by V.  If successful, it returns true
01007 /// and returns the string in Str.  If unsuccessful, it returns false.
01008 bool llvm::GetConstantStringInfo(Value *V, std::string &Str, uint64_t Offset,
01009                                  bool StopAtNul) {
01010   // If V is NULL then return false;
01011   if (V == NULL) return false;
01012 
01013   // Look through bitcast instructions.
01014   if (BitCastInst *BCI = dyn_cast<BitCastInst>(V))
01015     return GetConstantStringInfo(BCI->getOperand(0), Str, Offset, StopAtNul);
01016   
01017   // If the value is not a GEP instruction nor a constant expression with a
01018   // GEP instruction, then return false because ConstantArray can't occur
01019   // any other way
01020   User *GEP = 0;
01021   if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
01022     GEP = GEPI;
01023   } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
01024     if (CE->getOpcode() == Instruction::BitCast)
01025       return GetConstantStringInfo(CE->getOperand(0), Str, Offset, StopAtNul);
01026     if (CE->getOpcode() != Instruction::GetElementPtr)
01027       return false;
01028     GEP = CE;
01029   }
01030   
01031   if (GEP) {
01032     // Make sure the GEP has exactly three arguments.
01033     if (GEP->getNumOperands() != 3)
01034       return false;
01035     
01036     // Make sure the index-ee is a pointer to array of i8.
01037     const PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
01038     const ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
01039     if (AT == 0 || AT->getElementType() != Type::getInt8Ty(V->getContext()))
01040       return false;
01041     
01042     // Check to make sure that the first operand of the GEP is an integer and
01043     // has value 0 so that we are sure we're indexing into the initializer.
01044     ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
01045     if (FirstIdx == 0 || !FirstIdx->isZero())
01046       return false;
01047     
01048     // If the second index isn't a ConstantInt, then this is a variable index
01049     // into the array.  If this occurs, we can't say anything meaningful about
01050     // the string.
01051     uint64_t StartIdx = 0;
01052     if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
01053       StartIdx = CI->getZExtValue();
01054     else
01055       return false;
01056     return GetConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset,
01057                                  StopAtNul);
01058   }
01059   
01060   // The GEP instruction, constant or instruction, must reference a global
01061   // variable that is a constant and is initialized. The referenced constant
01062   // initializer is the array that we'll use for optimization.
01063   GlobalVariable* GV = dyn_cast<GlobalVariable>(V);
01064   if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
01065     return false;
01066   Constant *GlobalInit = GV->getInitializer();
01067   
01068   // Handle the ConstantAggregateZero case
01069   if (isa<ConstantAggregateZero>(GlobalInit)) {
01070     // This is a degenerate case. The initializer is constant zero so the
01071     // length of the string must be zero.
01072     Str.clear();
01073     return true;
01074   }
01075   
01076   // Must be a Constant Array
01077   ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
01078   if (Array == 0 ||
01079       Array->getType()->getElementType() != Type::getInt8Ty(V->getContext()))
01080     return false;
01081   
01082   // Get the number of elements in the array
01083   uint64_t NumElts = Array->getType()->getNumElements();
01084   
01085   if (Offset > NumElts)
01086     return false;
01087   
01088   // Traverse the constant array from 'Offset' which is the place the GEP refers
01089   // to in the array.
01090   Str.reserve(NumElts-Offset);
01091   for (unsigned i = Offset; i != NumElts; ++i) {
01092     Constant *Elt = Array->getOperand(i);
01093     ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
01094     if (!CI) // This array isn't suitable, non-int initializer.
01095       return false;
01096     if (StopAtNul && CI->isZero())
01097       return true; // we found end of string, success!
01098     Str += (char)CI->getZExtValue();
01099   }
01100   
01101   // The array isn't null terminated, but maybe this is a memcpy, not a strcpy.
01102   return true;
01103 }