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