lib/Transforms/InstCombine/InstCombineShifts.cpp

   1 //===- InstCombineShifts.cpp ----------------------------------------------===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This file implements the visitShl, visitLShr, and visitAShr functions.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 #include "InstCombine.h"
  15 #include "llvm/IntrinsicInst.h"
  16 #include "llvm/Analysis/InstructionSimplify.h"
  17 #include "llvm/Support/PatternMatch.h"
  18 using namespace llvm;
  19 using namespace PatternMatch;
  20
  21 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
  22   assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
  23   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
  24
  25   // See if we can fold away this shift.
  26   if (SimplifyDemandedInstructionBits(I))
  27     return &I;
  28
  29   // Try to fold constant and into select arguments.
  30   if (isa<Constant>(Op0))
  31     if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
  32       if (Instruction *R = FoldOpIntoSelect(I, SI))
  33         return R;
  34
  35   if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
  36     if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
  37       return Res;
  38
  39   // X shift (A srem B) -> X shift (A and B-1) iff B is a power of 2.
  40   // Because shifts by negative values (which could occur if A were negative)
  41   // are undefined.
  42   Value *A; const APInt *B;
  43   if (Op1->hasOneUse() && match(Op1, m_SRem(m_Value(A), m_Power2(B)))) {
  44     // FIXME: Should this get moved into SimplifyDemandedBits by saying we don't
  45     // demand the sign bit (and many others) here??
  46     Value *Rem = Builder->CreateAnd(A, ConstantInt::get(I.getType(), *B-1),
  47                                     Op1->getName());
  48     I.setOperand(1, Rem);
  49     return &I;
  50   }
  51
  52   return 0;
  53 }
  54
  55 /// CanEvaluateShifted - See if we can compute the specified value, but shifted
  56 /// logically to the left or right by some number of bits.  This should return
  57 /// true if the expression can be computed for the same cost as the current
  58 /// expression tree.  This is used to eliminate extraneous shifting from things
  59 /// like:
  60 ///      %C = shl i128 %A, 64
  61 ///      %D = shl i128 %B, 96
  62 ///      %E = or i128 %C, %D
  63 ///      %F = lshr i128 %E, 64
  64 /// where the client will ask if E can be computed shifted right by 64-bits.  If
  65 /// this succeeds, the GetShiftedValue function will be called to produce the
  66 /// value.
  67 static bool CanEvaluateShifted(Value *V, unsigned NumBits, bool isLeftShift,
  68                                InstCombiner &IC) {
  69   // We can always evaluate constants shifted.
  70   if (isa<Constant>(V))
  71     return true;
  72
  73   Instruction *I = dyn_cast<Instruction>(V);
  74   if (!I) return false;
  75
  76   // If this is the opposite shift, we can directly reuse the input of the shift
  77   // if the needed bits are already zero in the input.  This allows us to reuse
  78   // the value which means that we don't care if the shift has multiple uses.
  79   //  TODO:  Handle opposite shift by exact value.
  80   ConstantInt *CI = 0;
  81   if ((isLeftShift && match(I, m_LShr(m_Value(), m_ConstantInt(CI)))) ||
  82       (!isLeftShift && match(I, m_Shl(m_Value(), m_ConstantInt(CI))))) {
  83     if (CI->getZExtValue() == NumBits) {
  84       // TODO: Check that the input bits are already zero with MaskedValueIsZero
  85 #if 0
  86       // If this is a truncate of a logical shr, we can truncate it to a smaller
  87       // lshr iff we know that the bits we would otherwise be shifting in are
  88       // already zeros.
  89       uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
  90       uint32_t BitWidth = Ty->getScalarSizeInBits();
  91       if (MaskedValueIsZero(I->getOperand(0),
  92             APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
  93           CI->getLimitedValue(BitWidth) < BitWidth) {
  94         return CanEvaluateTruncated(I->getOperand(0), Ty);
  95       }
  96 #endif
  97
  98     }
  99   }
 100
 101   // We can't mutate something that has multiple uses: doing so would
 102   // require duplicating the instruction in general, which isn't profitable.
 103   if (!I->hasOneUse()) return false;
 104
 105   switch (I->getOpcode()) {
 106   default: return false;
 107   case Instruction::And:
 108   case Instruction::Or:
 109   case Instruction::Xor:
 110     // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
 111     return CanEvaluateShifted(I->getOperand(0), NumBits, isLeftShift, IC) &&
 112            CanEvaluateShifted(I->getOperand(1), NumBits, isLeftShift, IC);
 113
 114   case Instruction::Shl: {
 115     // We can often fold the shift into shifts-by-a-constant.
 116     CI = dyn_cast<ConstantInt>(I->getOperand(1));
 117     if (CI == 0) return false;
 118
 119     // We can always fold shl(c1)+shl(c2) -> shl(c1+c2).
 120     if (isLeftShift) return true;
 121
 122     // We can always turn shl(c)+shr(c) -> and(c2).
 123     if (CI->getValue() == NumBits) return true;
 124
 125     unsigned TypeWidth = I->getType()->getScalarSizeInBits();
 126
 127     // We can turn shl(c1)+shr(c2) -> shl(c3)+and(c4), but it isn't
 128     // profitable unless we know the and'd out bits are already zero.
 129     if (CI->getZExtValue() > NumBits) {
 130       unsigned LowBits = TypeWidth - CI->getZExtValue();
 131       if (MaskedValueIsZero(I->getOperand(0),
 132                        APInt::getLowBitsSet(TypeWidth, NumBits) << LowBits))
 133         return true;
 134     }
 135
 136     return false;
 137   }
 138   case Instruction::LShr: {
 139     // We can often fold the shift into shifts-by-a-constant.
 140     CI = dyn_cast<ConstantInt>(I->getOperand(1));
 141     if (CI == 0) return false;
 142
 143     // We can always fold lshr(c1)+lshr(c2) -> lshr(c1+c2).
 144     if (!isLeftShift) return true;
 145
 146     // We can always turn lshr(c)+shl(c) -> and(c2).
 147     if (CI->getValue() == NumBits) return true;
 148
 149     unsigned TypeWidth = I->getType()->getScalarSizeInBits();
 150
 151     // We can always turn lshr(c1)+shl(c2) -> lshr(c3)+and(c4), but it isn't
 152     // profitable unless we know the and'd out bits are already zero.
 153     if (CI->getZExtValue() > NumBits) {
 154       unsigned LowBits = CI->getZExtValue() - NumBits;
 155       if (MaskedValueIsZero(I->getOperand(0),
 156                           APInt::getLowBitsSet(TypeWidth, NumBits) << LowBits))
 157         return true;
 158     }
 159
 160     return false;
 161   }
 162   case Instruction::Select: {
 163     SelectInst *SI = cast<SelectInst>(I);
 164     return CanEvaluateShifted(SI->getTrueValue(), NumBits, isLeftShift, IC) &&
 165            CanEvaluateShifted(SI->getFalseValue(), NumBits, isLeftShift, IC);
 166   }
 167   case Instruction::PHI: {
 168     // We can change a phi if we can change all operands.  Note that we never
 169     // get into trouble with cyclic PHIs here because we only consider
 170     // instructions with a single use.
 171     PHINode *PN = cast<PHINode>(I);
 172     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
 173       if (!CanEvaluateShifted(PN->getIncomingValue(i), NumBits, isLeftShift,IC))
 174         return false;
 175     return true;
 176   }
 177   }
 178 }
 179
 180 /// GetShiftedValue - When CanEvaluateShifted returned true for an expression,
 181 /// this value inserts the new computation that produces the shifted value.
 182 static Value *GetShiftedValue(Value *V, unsigned NumBits, bool isLeftShift,
 183                               InstCombiner &IC) {
 184   // We can always evaluate constants shifted.
 185   if (Constant *C = dyn_cast<Constant>(V)) {
 186     if (isLeftShift)
 187       V = IC.Builder->CreateShl(C, NumBits);
 188     else
 189       V = IC.Builder->CreateLShr(C, NumBits);
 190     // If we got a constantexpr back, try to simplify it with TD info.
 191     if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
 192       V = ConstantFoldConstantExpression(CE, IC.getTargetData());
 193     return V;
 194   }
 195
 196   Instruction *I = cast<Instruction>(V);
 197   IC.Worklist.Add(I);
 198
 199   switch (I->getOpcode()) {
 200   default: assert(0 && "Inconsistency with CanEvaluateShifted");
 201   case Instruction::And:
 202   case Instruction::Or:
 203   case Instruction::Xor:
 204     // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
 205     I->setOperand(0, GetShiftedValue(I->getOperand(0), NumBits,isLeftShift,IC));
 206     I->setOperand(1, GetShiftedValue(I->getOperand(1), NumBits,isLeftShift,IC));
 207     return I;
 208
 209   case Instruction::Shl: {
 210     unsigned TypeWidth = I->getType()->getScalarSizeInBits();
 211
 212     // We only accept shifts-by-a-constant in CanEvaluateShifted.
 213     ConstantInt *CI = cast<ConstantInt>(I->getOperand(1));
 214
 215     // We can always fold shl(c1)+shl(c2) -> shl(c1+c2).
 216     if (isLeftShift) {
 217       // If this is oversized composite shift, then unsigned shifts get 0.
 218       unsigned NewShAmt = NumBits+CI->getZExtValue();
 219       if (NewShAmt >= TypeWidth)
 220         return Constant::getNullValue(I->getType());
 221
 222       I->setOperand(1, ConstantInt::get(I->getType(), NewShAmt));
 223       return I;
 224     }
 225
 226     // We turn shl(c)+lshr(c) -> and(c2) if the input doesn't already have
 227     // zeros.
 228     if (CI->getValue() == NumBits) {
 229       APInt Mask(APInt::getLowBitsSet(TypeWidth, TypeWidth - NumBits));
 230       V = IC.Builder->CreateAnd(I->getOperand(0),
 231                                 ConstantInt::get(I->getContext(), Mask));
 232       if (Instruction *VI = dyn_cast<Instruction>(V)) {
 233         VI->moveBefore(I);
 234         VI->takeName(I);
 235       }
 236       return V;
 237     }
 238
 239     // We turn shl(c1)+shr(c2) -> shl(c3)+and(c4), but only when we know that
 240     // the and won't be needed.
 241     assert(CI->getZExtValue() > NumBits);
 242     I->setOperand(1, ConstantInt::get(I->getType(),
 243                                       CI->getZExtValue() - NumBits));
 244     return I;
 245   }
 246   case Instruction::LShr: {
 247     unsigned TypeWidth = I->getType()->getScalarSizeInBits();
 248     // We only accept shifts-by-a-constant in CanEvaluateShifted.
 249     ConstantInt *CI = cast<ConstantInt>(I->getOperand(1));
 250
 251     // We can always fold lshr(c1)+lshr(c2) -> lshr(c1+c2).
 252     if (!isLeftShift) {
 253       // If this is oversized composite shift, then unsigned shifts get 0.
 254       unsigned NewShAmt = NumBits+CI->getZExtValue();
 255       if (NewShAmt >= TypeWidth)
 256         return Constant::getNullValue(I->getType());
 257
 258       I->setOperand(1, ConstantInt::get(I->getType(), NewShAmt));
 259       return I;
 260     }
 261
 262     // We turn lshr(c)+shl(c) -> and(c2) if the input doesn't already have
 263     // zeros.
 264     if (CI->getValue() == NumBits) {
 265       APInt Mask(APInt::getHighBitsSet(TypeWidth, TypeWidth - NumBits));
 266       V = IC.Builder->CreateAnd(I->getOperand(0),
 267                                 ConstantInt::get(I->getContext(), Mask));
 268       if (Instruction *VI = dyn_cast<Instruction>(V)) {
 269         VI->moveBefore(I);
 270         VI->takeName(I);
 271       }
 272       return V;
 273     }
 274
 275     // We turn lshr(c1)+shl(c2) -> lshr(c3)+and(c4), but only when we know that
 276     // the and won't be needed.
 277     assert(CI->getZExtValue() > NumBits);
 278     I->setOperand(1, ConstantInt::get(I->getType(),
 279                                       CI->getZExtValue() - NumBits));
 280     return I;
 281   }
 282
 283   case Instruction::Select:
 284     I->setOperand(1, GetShiftedValue(I->getOperand(1), NumBits,isLeftShift,IC));
 285     I->setOperand(2, GetShiftedValue(I->getOperand(2), NumBits,isLeftShift,IC));
 286     return I;
 287   case Instruction::PHI: {
 288     // We can change a phi if we can change all operands.  Note that we never
 289     // get into trouble with cyclic PHIs here because we only consider
 290     // instructions with a single use.
 291     PHINode *PN = cast<PHINode>(I);
 292     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
 293       PN->setIncomingValue(i, GetShiftedValue(PN->getIncomingValue(i),
 294                                               NumBits, isLeftShift, IC));
 295     return PN;
 296   }
 297   }
 298 }
 299
 300
 301
 302 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
 303                                                BinaryOperator &I) {
 304   bool isLeftShift = I.getOpcode() == Instruction::Shl;
 305
 306
 307   // See if we can propagate this shift into the input, this covers the trivial
 308   // cast of lshr(shl(x,c1),c2) as well as other more complex cases.
 309   if (I.getOpcode() != Instruction::AShr &&
 310       CanEvaluateShifted(Op0, Op1->getZExtValue(), isLeftShift, *this)) {
 311     DEBUG(dbgs() << "ICE: GetShiftedValue propagating shift through expression"
 312               " to eliminate shift:\n  IN: " << *Op0 << "\n  SH: " << I <<"\n");
 313
 314     return ReplaceInstUsesWith(I,
 315                  GetShiftedValue(Op0, Op1->getZExtValue(), isLeftShift, *this));
 316   }
 317
 318
 319   // See if we can simplify any instructions used by the instruction whose sole
 320   // purpose is to compute bits we don't care about.
 321   uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
 322
 323   // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
 324   // a signed shift.
 325   //
 326   if (Op1->uge(TypeBits)) {
 327     if (I.getOpcode() != Instruction::AShr)
 328       return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
 329     // ashr i32 X, 32 --> ashr i32 X, 31
 330     I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
 331     return &I;
 332   }
 333
 334   // ((X*C1) << C2) == (X * (C1 << C2))
 335   if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
 336     if (BO->getOpcode() == Instruction::Mul && isLeftShift)
 337       if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
 338         return BinaryOperator::CreateMul(BO->getOperand(0),
 339                                         ConstantExpr::getShl(BOOp, Op1));
 340
 341   // Try to fold constant and into select arguments.
 342   if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
 343     if (Instruction *R = FoldOpIntoSelect(I, SI))
 344       return R;
 345   if (isa<PHINode>(Op0))
 346     if (Instruction *NV = FoldOpIntoPhi(I))
 347       return NV;
 348
 349   // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
 350   if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
 351     Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
 352     // If 'shift2' is an ashr, we would have to get the sign bit into a funny
 353     // place.  Don't try to do this transformation in this case.  Also, we
 354     // require that the input operand is a shift-by-constant so that we have
 355     // confidence that the shifts will get folded together.  We could do this
 356     // xform in more cases, but it is unlikely to be profitable.
 357     if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
 358         isa<ConstantInt>(TrOp->getOperand(1))) {
 359       // Okay, we'll do this xform.  Make the shift of shift.
 360       Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
 361       // (shift2 (shift1 & 0x00FF), c2)
 362       Value *NSh = Builder->CreateBinOp(I.getOpcode(), TrOp, ShAmt,I.getName());
 363
 364       // For logical shifts, the truncation has the effect of making the high
 365       // part of the register be zeros.  Emulate this by inserting an AND to
 366       // clear the top bits as needed.  This 'and' will usually be zapped by
 367       // other xforms later if dead.
 368       unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
 369       unsigned DstSize = TI->getType()->getScalarSizeInBits();
 370       APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
 371
 372       // The mask we constructed says what the trunc would do if occurring
 373       // between the shifts.  We want to know the effect *after* the second
 374       // shift.  We know that it is a logical shift by a constant, so adjust the
 375       // mask as appropriate.
 376       if (I.getOpcode() == Instruction::Shl)
 377         MaskV <<= Op1->getZExtValue();
 378       else {
 379         assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
 380         MaskV = MaskV.lshr(Op1->getZExtValue());
 381       }
 382
 383       // shift1 & 0x00FF
 384       Value *And = Builder->CreateAnd(NSh,
 385                                       ConstantInt::get(I.getContext(), MaskV),
 386                                       TI->getName());
 387
 388       // Return the value truncated to the interesting size.
 389       return new TruncInst(And, I.getType());
 390     }
 391   }
 392
 393   if (Op0->hasOneUse()) {
 394     if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
 395       // Turn ((X >> C) + Y) << C  ->  (X + (Y << C)) & (~0 << C)
 396       Value *V1, *V2;
 397       ConstantInt *CC;
 398       switch (Op0BO->getOpcode()) {
 399       default: break;
 400       case Instruction::Add:
 401       case Instruction::And:
 402       case Instruction::Or:
 403       case Instruction::Xor: {
 404         // These operators commute.
 405         // Turn (Y + (X >> C)) << C  ->  (X + (Y << C)) & (~0 << C)
 406         if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
 407             match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
 408                   m_Specific(Op1)))) {
 409           Value *YS =         // (Y << C)
 410             Builder->CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
 411           // (X + (Y << C))
 412           Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), YS, V1,
 413                                           Op0BO->getOperand(1)->getName());
 414           uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
 415           return BinaryOperator::CreateAnd(X, ConstantInt::get(I.getContext(),
 416                      APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
 417         }
 418
 419         // Turn (Y + ((X >> C) & CC)) << C  ->  ((X & (CC << C)) + (Y << C))
 420         Value *Op0BOOp1 = Op0BO->getOperand(1);
 421         if (isLeftShift && Op0BOOp1->hasOneUse() &&
 422             match(Op0BOOp1,
 423                   m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
 424                         m_ConstantInt(CC))) &&
 425             cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
 426           Value *YS =   // (Y << C)
 427             Builder->CreateShl(Op0BO->getOperand(0), Op1,
 428                                          Op0BO->getName());
 429           // X & (CC << C)
 430           Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
 431                                          V1->getName()+".mask");
 432           return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
 433         }
 434       }
 435
 436       // FALL THROUGH.
 437       case Instruction::Sub: {
 438         // Turn ((X >> C) + Y) << C  ->  (X + (Y << C)) & (~0 << C)
 439         if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
 440             match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
 441                   m_Specific(Op1)))) {
 442           Value *YS =  // (Y << C)
 443             Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
 444           // (X + (Y << C))
 445           Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), V1, YS,
 446                                           Op0BO->getOperand(0)->getName());
 447           uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
 448           return BinaryOperator::CreateAnd(X, ConstantInt::get(I.getContext(),
 449                      APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
 450         }
 451
 452         // Turn (((X >> C)&CC) + Y) << C  ->  (X + (Y << C)) & (CC << C)
 453         if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
 454             match(Op0BO->getOperand(0),
 455                   m_And(m_Shr(m_Value(V1), m_Value(V2)),
 456                         m_ConstantInt(CC))) && V2 == Op1 &&
 457             cast<BinaryOperator>(Op0BO->getOperand(0))
 458                 ->getOperand(0)->hasOneUse()) {
 459           Value *YS = // (Y << C)
 460             Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
 461           // X & (CC << C)
 462           Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
 463                                          V1->getName()+".mask");
 464
 465           return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
 466         }
 467
 468         break;
 469       }
 470       }
 471
 472
 473       // If the operand is an bitwise operator with a constant RHS, and the
 474       // shift is the only use, we can pull it out of the shift.
 475       if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
 476         bool isValid = true;     // Valid only for And, Or, Xor
 477         bool highBitSet = false; // Transform if high bit of constant set?
 478
 479         switch (Op0BO->getOpcode()) {
 480         default: isValid = false; break;   // Do not perform transform!
 481         case Instruction::Add:
 482           isValid = isLeftShift;
 483           break;
 484         case Instruction::Or:
 485         case Instruction::Xor:
 486           highBitSet = false;
 487           break;
 488         case Instruction::And:
 489           highBitSet = true;
 490           break;
 491         }
 492
 493         // If this is a signed shift right, and the high bit is modified
 494         // by the logical operation, do not perform the transformation.
 495         // The highBitSet boolean indicates the value of the high bit of
 496         // the constant which would cause it to be modified for this
 497         // operation.
 498         //
 499         if (isValid && I.getOpcode() == Instruction::AShr)
 500           isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
 501
 502         if (isValid) {
 503           Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
 504
 505           Value *NewShift =
 506             Builder->CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
 507           NewShift->takeName(Op0BO);
 508
 509           return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
 510                                         NewRHS);
 511         }
 512       }
 513     }
 514   }
 515
 516   // Find out if this is a shift of a shift by a constant.
 517   BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
 518   if (ShiftOp && !ShiftOp->isShift())
 519     ShiftOp = 0;
 520
 521   if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
 522     ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
 523     uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
 524     uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
 525     assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
 526     if (ShiftAmt1 == 0) return 0;  // Will be simplified in the future.
 527     Value *X = ShiftOp->getOperand(0);
 528
 529     uint32_t AmtSum = ShiftAmt1+ShiftAmt2;   // Fold into one big shift.
 530
 531     const IntegerType *Ty = cast<IntegerType>(I.getType());
 532
 533     // Check for (X << c1) << c2  and  (X >> c1) >> c2
 534     if (I.getOpcode() == ShiftOp->getOpcode()) {
 535       // If this is oversized composite shift, then unsigned shifts get 0, ashr
 536       // saturates.
 537       if (AmtSum >= TypeBits) {
 538         if (I.getOpcode() != Instruction::AShr)
 539           return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
 540         AmtSum = TypeBits-1;  // Saturate to 31 for i32 ashr.
 541       }
 542
 543       return BinaryOperator::Create(I.getOpcode(), X,
 544                                     ConstantInt::get(Ty, AmtSum));
 545     }
 546
 547     if (ShiftAmt1 == ShiftAmt2) {
 548       // If we have ((X >>? C) << C), turn this into X & (-1 << C).
 549       if (I.getOpcode() == Instruction::Shl &&
 550           ShiftOp->getOpcode() != Instruction::Shl) {
 551         APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
 552         return BinaryOperator::CreateAnd(X,
 553                                          ConstantInt::get(I.getContext(),Mask));
 554       }
 555       // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
 556       if (I.getOpcode() == Instruction::LShr &&
 557           ShiftOp->getOpcode() == Instruction::Shl) {
 558         APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
 559         return BinaryOperator::CreateAnd(X,
 560                                         ConstantInt::get(I.getContext(), Mask));
 561       }
 562     } else if (ShiftAmt1 < ShiftAmt2) {
 563       uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
 564
 565       // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
 566       if (I.getOpcode() == Instruction::Shl &&
 567           ShiftOp->getOpcode() != Instruction::Shl) {
 568         assert(ShiftOp->getOpcode() == Instruction::LShr ||
 569                ShiftOp->getOpcode() == Instruction::AShr);
 570         Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
 571
 572         APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
 573         return BinaryOperator::CreateAnd(Shift,
 574                                          ConstantInt::get(I.getContext(),Mask));
 575       }
 576
 577       // (X << C1) >>u C2  --> X >>u (C2-C1) & (-1 >> C2)
 578       if (I.getOpcode() == Instruction::LShr &&
 579           ShiftOp->getOpcode() == Instruction::Shl) {
 580         assert(ShiftOp->getOpcode() == Instruction::Shl);
 581         Value *Shift = Builder->CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
 582
 583         APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
 584         return BinaryOperator::CreateAnd(Shift,
 585                                          ConstantInt::get(I.getContext(),Mask));
 586       }
 587
 588       // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
 589     } else {
 590       assert(ShiftAmt2 < ShiftAmt1);
 591       uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
 592
 593       // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
 594       if (I.getOpcode() == Instruction::Shl &&
 595           ShiftOp->getOpcode() != Instruction::Shl) {
 596         Value *Shift = Builder->CreateBinOp(ShiftOp->getOpcode(), X,
 597                                             ConstantInt::get(Ty, ShiftDiff));
 598
 599         APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
 600         return BinaryOperator::CreateAnd(Shift,
 601                                          ConstantInt::get(I.getContext(),Mask));
 602       }
 603
 604       // (X << C1) >>u C2  --> X << (C1-C2) & (-1 >> C2)
 605       if (I.getOpcode() == Instruction::LShr &&
 606           ShiftOp->getOpcode() == Instruction::Shl) {
 607         Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
 608
 609         APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
 610         return BinaryOperator::CreateAnd(Shift,
 611                                          ConstantInt::get(I.getContext(),Mask));
 612       }
 613
 614       // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
 615     }
 616   }
 617   return 0;
 618 }
 619
 620 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
 621   if (Value *V = SimplifyShlInst(I.getOperand(0), I.getOperand(1),
 622                                  I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
 623                                  TD))
 624     return ReplaceInstUsesWith(I, V);
 625
 626   if (Instruction *V = commonShiftTransforms(I))
 627     return V;
 628
 629   if (ConstantInt *Op1C = dyn_cast<ConstantInt>(I.getOperand(1))) {
 630     unsigned ShAmt = Op1C->getZExtValue();
 631
 632     // If the shifted-out value is known-zero, then this is a NUW shift.
 633     if (!I.hasNoUnsignedWrap() &&
 634         MaskedValueIsZero(I.getOperand(0),
 635                           APInt::getHighBitsSet(Op1C->getBitWidth(), ShAmt))) {
 636           I.setHasNoUnsignedWrap();
 637           return &I;
 638         }
 639
 640     // If the shifted out value is all signbits, this is a NSW shift.
 641     if (!I.hasNoSignedWrap() &&
 642         ComputeNumSignBits(I.getOperand(0)) > ShAmt) {
 643       I.setHasNoSignedWrap();
 644       return &I;
 645     }
 646   }
 647
 648   // (C1 << A) << C2 -> (C1 << C2) << A
 649   Constant *C1, *C2;
 650   Value *A;
 651   if (match(I.getOperand(0), m_OneUse(m_Shl(m_Constant(C1), m_Value(A)))) &&
 652       match(I.getOperand(1), m_Constant(C2)))
 653     return BinaryOperator::CreateShl(ConstantExpr::getShl(C1, C2), A);
 654
 655   return 0;
 656 }
 657
 658 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
 659   if (Value *V = SimplifyLShrInst(I.getOperand(0), I.getOperand(1),
 660                                   I.isExact(), TD))
 661     return ReplaceInstUsesWith(I, V);
 662
 663   if (Instruction *R = commonShiftTransforms(I))
 664     return R;
 665
 666   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
 667
 668   if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
 669     unsigned ShAmt = Op1C->getZExtValue();
 670
 671     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
 672       unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
 673       // ctlz.i32(x)>>5  --> zext(x == 0)
 674       // cttz.i32(x)>>5  --> zext(x == 0)
 675       // ctpop.i32(x)>>5 --> zext(x == -1)
 676       if ((II->getIntrinsicID() == Intrinsic::ctlz ||
 677            II->getIntrinsicID() == Intrinsic::cttz ||
 678            II->getIntrinsicID() == Intrinsic::ctpop) &&
 679           isPowerOf2_32(BitWidth) && Log2_32(BitWidth) == ShAmt) {
 680         bool isCtPop = II->getIntrinsicID() == Intrinsic::ctpop;
 681         Constant *RHS = ConstantInt::getSigned(Op0->getType(), isCtPop ? -1:0);
 682         Value *Cmp = Builder->CreateICmpEQ(II->getArgOperand(0), RHS);
 683         return new ZExtInst(Cmp, II->getType());
 684       }
 685     }
 686
 687     // If the shifted-out value is known-zero, then this is an exact shift.
 688     if (!I.isExact() &&
 689         MaskedValueIsZero(Op0,APInt::getLowBitsSet(Op1C->getBitWidth(),ShAmt))){
 690       I.setIsExact();
 691       return &I;
 692     }
 693   }
 694
 695   return 0;
 696 }
 697
 698 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
 699   if (Value *V = SimplifyAShrInst(I.getOperand(0), I.getOperand(1),
 700                                   I.isExact(), TD))
 701     return ReplaceInstUsesWith(I, V);
 702
 703   if (Instruction *R = commonShiftTransforms(I))
 704     return R;
 705
 706   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
 707
 708   if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
 709     unsigned ShAmt = Op1C->getZExtValue();
 710
 711     // If the input is a SHL by the same constant (ashr (shl X, C), C), then we
 712     // have a sign-extend idiom.
 713     Value *X;
 714     if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1)))) {
 715       // If the left shift is just shifting out partial signbits, delete the
 716       // extension.
 717       if (cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap())
 718         return ReplaceInstUsesWith(I, X);
 719
 720       // If the input is an extension from the shifted amount value, e.g.
 721       //   %x = zext i8 %A to i32
 722       //   %y = shl i32 %x, 24
 723       //   %z = ashr %y, 24
 724       // then turn this into "z = sext i8 A to i32".
 725       if (ZExtInst *ZI = dyn_cast<ZExtInst>(X)) {
 726         uint32_t SrcBits = ZI->getOperand(0)->getType()->getScalarSizeInBits();
 727         uint32_t DestBits = ZI->getType()->getScalarSizeInBits();
 728         if (Op1C->getZExtValue() == DestBits-SrcBits)
 729           return new SExtInst(ZI->getOperand(0), ZI->getType());
 730       }
 731     }
 732
 733     // If the shifted-out value is known-zero, then this is an exact shift.
 734     if (!I.isExact() &&
 735         MaskedValueIsZero(Op0,APInt::getLowBitsSet(Op1C->getBitWidth(),ShAmt))){
 736       I.setIsExact();
 737       return &I;
 738     }
 739   }
 740
 741   // See if we can turn a signed shr into an unsigned shr.
 742   if (MaskedValueIsZero(Op0,
 743                         APInt::getSignBit(I.getType()->getScalarSizeInBits())))
 744     return BinaryOperator::CreateLShr(Op0, Op1);
 745
 746   // Arithmetic shifting an all-sign-bit value is a no-op.
 747   unsigned NumSignBits = ComputeNumSignBits(Op0);
 748   if (NumSignBits == Op0->getType()->getScalarSizeInBits())
 749     return ReplaceInstUsesWith(I, Op0);
 750
 751   return 0;
 752 }
 753