lib/Transforms/InstCombine/InstCombineShifts.cpp

   1 //===- InstCombineShifts.cpp ----------------------------------------------===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9 // This file implements the visitShl, visitLShr, and visitAShr functions.
  10 //
  11 //===----------------------------------------------------------------------===//
  12
  13 #include "InstCombineInternal.h"
  14 #include "llvm/Analysis/ConstantFolding.h"
  15 #include "llvm/Analysis/InstructionSimplify.h"
  16 #include "llvm/IR/IntrinsicInst.h"
  17 #include "llvm/IR/PatternMatch.h"
  18 using namespace llvm;
  19 using namespace PatternMatch;
  20
  21 #define DEBUG_TYPE "instcombine"
  22
  23 // Given pattern:
  24 //   (x shiftopcode Q) shiftopcode K
  25 // we should rewrite it as
  26 //   x shiftopcode (Q+K)  iff (Q+K) u< bitwidth(x)
  27 // This is valid for any shift, but they must be identical.
  28 static Instruction *
  29 reassociateShiftAmtsOfTwoSameDirectionShifts(BinaryOperator *Sh0,
  30                                              const SimplifyQuery &SQ,
  31                                              InstCombiner::BuilderTy &Builder) {
  32   // Look for a shift of some instruction, ignore zext of shift amount if any.
  33   Instruction *Sh0Op0;
  34   Value *ShAmt0;
  35   if (!match(Sh0,
  36              m_Shift(m_Instruction(Sh0Op0), m_ZExtOrSelf(m_Value(ShAmt0)))))
  37     return nullptr;
  38
  39   // If there is a truncation between the two shifts, we must make note of it
  40   // and look through it. The truncation imposes additional constraints on the
  41   // transform.
  42   Instruction *Sh1;
  43   Value *Trunc = nullptr;
  44   match(Sh0Op0,
  45         m_CombineOr(m_CombineAnd(m_Trunc(m_Instruction(Sh1)), m_Value(Trunc)),
  46                     m_Instruction(Sh1)));
  47
  48   // Inner shift: (x shiftopcode ShAmt1)
  49   // Like with other shift, ignore zext of shift amount if any.
  50   Value *X, *ShAmt1;
  51   if (!match(Sh1, m_Shift(m_Value(X), m_ZExtOrSelf(m_Value(ShAmt1)))))
  52     return nullptr;
  53
  54   // We have two shift amounts from two different shifts. The types of those
  55   // shift amounts may not match. If that's the case let's bailout now..
  56   if (ShAmt0->getType() != ShAmt1->getType())
  57     return nullptr;
  58
  59   // The shift opcodes must be identical.
  60   Instruction::BinaryOps ShiftOpcode = Sh0->getOpcode();
  61   if (ShiftOpcode != Sh1->getOpcode())
  62     return nullptr;
  63
  64   // Did we match a pattern with truncation ?
  65   if (Trunc) {
  66     // For right-shifts we can't do any such simplifications. Leave as-is.
  67     if (ShiftOpcode != Instruction::BinaryOps::Shl)
  68       return nullptr; // FIXME: still could perform constant-folding.
  69     // If we saw truncation, we'll need to produce extra instruction,
  70     // and for that one of the operands of the shift must be one-use.
  71     if (!match(Sh0, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
  72       return nullptr;
  73   }
  74
  75   // Can we fold (ShAmt0+ShAmt1) ?
  76   auto *NewShAmt = dyn_cast_or_null<Constant>(
  77       SimplifyAddInst(ShAmt0, ShAmt1, /*isNSW=*/false, /*isNUW=*/false,
  78                       SQ.getWithInstruction(Sh0)));
  79   if (!NewShAmt)
  80     return nullptr; // Did not simplify.
  81   // Is the new shift amount smaller than the bit width of inner shift?
  82   if (!match(NewShAmt, m_SpecificInt_ICMP(
  83                            ICmpInst::Predicate::ICMP_ULT,
  84                            APInt(NewShAmt->getType()->getScalarSizeInBits(),
  85                                  X->getType()->getScalarSizeInBits()))))
  86     return nullptr; // FIXME: could perform constant-folding.
  87
  88   // All good, we can do this fold.
  89   NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, X->getType());
  90
  91   BinaryOperator *NewShift = BinaryOperator::Create(ShiftOpcode, X, NewShAmt);
  92
  93   // The flags can only be propagated if there wasn't a trunc.
  94   if (!Trunc) {
  95     // If the pattern did not involve trunc, and both of the original shifts
  96     // had the same flag set, preserve the flag.
  97     if (ShiftOpcode == Instruction::BinaryOps::Shl) {
  98       NewShift->setHasNoUnsignedWrap(Sh0->hasNoUnsignedWrap() &&
  99                                      Sh1->hasNoUnsignedWrap());
 100       NewShift->setHasNoSignedWrap(Sh0->hasNoSignedWrap() &&
 101                                    Sh1->hasNoSignedWrap());
 102     } else {
 103       NewShift->setIsExact(Sh0->isExact() && Sh1->isExact());
 104     }
 105   }
 106
 107   Instruction *Ret = NewShift;
 108   if (Trunc) {
 109     Builder.Insert(NewShift);
 110     Ret = CastInst::Create(Instruction::Trunc, NewShift, Sh0->getType());
 111   }
 112
 113   return Ret;
 114 }
 115
 116 // If we have some pattern that leaves only some low bits set, and then performs
 117 // left-shift of those bits, if none of the bits that are left after the final
 118 // shift are modified by the mask, we can omit the mask.
 119 //
 120 // There are many variants to this pattern:
 121 //   a)  (x & ((1 << MaskShAmt) - 1)) << ShiftShAmt
 122 //   b)  (x & (~(-1 << MaskShAmt))) << ShiftShAmt
 123 //   c)  (x & (-1 >> MaskShAmt)) << ShiftShAmt
 124 //   d)  (x & ((-1 << MaskShAmt) >> MaskShAmt)) << ShiftShAmt
 125 //   e)  ((x << MaskShAmt) l>> MaskShAmt) << ShiftShAmt
 126 //   f)  ((x << MaskShAmt) a>> MaskShAmt) << ShiftShAmt
 127 // All these patterns can be simplified to just:
 128 //   x << ShiftShAmt
 129 // iff:
 130 //   a,b)     (MaskShAmt+ShiftShAmt) u>= bitwidth(x)
 131 //   c,d,e,f) (ShiftShAmt-MaskShAmt) s>= 0 (i.e. ShiftShAmt u>= MaskShAmt)
 132 static Instruction *
 133 dropRedundantMaskingOfLeftShiftInput(BinaryOperator *OuterShift,
 134                                      const SimplifyQuery &SQ) {
 135   assert(OuterShift->getOpcode() == Instruction::BinaryOps::Shl &&
 136          "The input must be 'shl'!");
 137
 138   Value *Masked = OuterShift->getOperand(0);
 139   Value *ShiftShAmt = OuterShift->getOperand(1);
 140
 141   Value *MaskShAmt;
 142
 143   // ((1 << MaskShAmt) - 1)
 144   auto MaskA = m_Add(m_Shl(m_One(), m_Value(MaskShAmt)), m_AllOnes());
 145   // (~(-1 << maskNbits))
 146   auto MaskB = m_Xor(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_AllOnes());
 147   // (-1 >> MaskShAmt)
 148   auto MaskC = m_Shr(m_AllOnes(), m_Value(MaskShAmt));
 149   // ((-1 << MaskShAmt) >> MaskShAmt)
 150   auto MaskD =
 151       m_Shr(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_Deferred(MaskShAmt));
 152
 153   Value *X;
 154   if (match(Masked, m_c_And(m_CombineOr(MaskA, MaskB), m_Value(X)))) {
 155     // Can we simplify (MaskShAmt+ShiftShAmt) ?
 156     Value *SumOfShAmts =
 157         SimplifyAddInst(MaskShAmt, ShiftShAmt, /*IsNSW=*/false, /*IsNUW=*/false,
 158                         SQ.getWithInstruction(OuterShift));
 159     if (!SumOfShAmts)
 160       return nullptr; // Did not simplify.
 161     // Is the total shift amount *not* smaller than the bit width?
 162     // FIXME: could also rely on ConstantRange.
 163     unsigned BitWidth = X->getType()->getScalarSizeInBits();
 164     if (!match(SumOfShAmts, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_UGE,
 165                                                APInt(BitWidth, BitWidth))))
 166       return nullptr;
 167     // All good, we can do this fold.
 168   } else if (match(Masked, m_c_And(m_CombineOr(MaskC, MaskD), m_Value(X))) ||
 169              match(Masked, m_Shr(m_Shl(m_Value(X), m_Value(MaskShAmt)),
 170                                  m_Deferred(MaskShAmt)))) {
 171     // Can we simplify (ShiftShAmt-MaskShAmt) ?
 172     Value *ShAmtsDiff =
 173         SimplifySubInst(ShiftShAmt, MaskShAmt, /*IsNSW=*/false, /*IsNUW=*/false,
 174                         SQ.getWithInstruction(OuterShift));
 175     if (!ShAmtsDiff)
 176       return nullptr; // Did not simplify.
 177     // Is the difference non-negative? (is ShiftShAmt u>= MaskShAmt ?)
 178     // FIXME: could also rely on ConstantRange.
 179     if (!match(ShAmtsDiff, m_NonNegative()))
 180       return nullptr;
 181     // All good, we can do this fold.
 182   } else
 183     return nullptr; // Don't know anything about this pattern.
 184
 185   // No 'NUW'/'NSW'!
 186   // We no longer know that we won't shift-out non-0 bits.
 187   return BinaryOperator::Create(OuterShift->getOpcode(), X, ShiftShAmt);
 188 }
 189
 190 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
 191   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
 192   assert(Op0->getType() == Op1->getType());
 193
 194   // See if we can fold away this shift.
 195   if (SimplifyDemandedInstructionBits(I))
 196     return &I;
 197
 198   // Try to fold constant and into select arguments.
 199   if (isa<Constant>(Op0))
 200     if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
 201       if (Instruction *R = FoldOpIntoSelect(I, SI))
 202         return R;
 203
 204   if (Constant *CUI = dyn_cast<Constant>(Op1))
 205     if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
 206       return Res;
 207
 208   if (Instruction *NewShift =
 209           reassociateShiftAmtsOfTwoSameDirectionShifts(&I, SQ, Builder))
 210     return NewShift;
 211
 212   // (C1 shift (A add C2)) -> (C1 shift C2) shift A)
 213   // iff A and C2 are both positive.
 214   Value *A;
 215   Constant *C;
 216   if (match(Op0, m_Constant()) && match(Op1, m_Add(m_Value(A), m_Constant(C))))
 217     if (isKnownNonNegative(A, DL, 0, &AC, &I, &DT) &&
 218         isKnownNonNegative(C, DL, 0, &AC, &I, &DT))
 219       return BinaryOperator::Create(
 220           I.getOpcode(), Builder.CreateBinOp(I.getOpcode(), Op0, C), A);
 221
 222   // X shift (A srem B) -> X shift (A and B-1) iff B is a power of 2.
 223   // Because shifts by negative values (which could occur if A were negative)
 224   // are undefined.
 225   const APInt *B;
 226   if (Op1->hasOneUse() && match(Op1, m_SRem(m_Value(A), m_Power2(B)))) {
 227     // FIXME: Should this get moved into SimplifyDemandedBits by saying we don't
 228     // demand the sign bit (and many others) here??
 229     Value *Rem = Builder.CreateAnd(A, ConstantInt::get(I.getType(), *B - 1),
 230                                    Op1->getName());
 231     I.setOperand(1, Rem);
 232     return &I;
 233   }
 234
 235   return nullptr;
 236 }
 237
 238 /// Return true if we can simplify two logical (either left or right) shifts
 239 /// that have constant shift amounts: OuterShift (InnerShift X, C1), C2.
 240 static bool canEvaluateShiftedShift(unsigned OuterShAmt, bool IsOuterShl,
 241                                     Instruction *InnerShift, InstCombiner &IC,
 242                                     Instruction *CxtI) {
 243   assert(InnerShift->isLogicalShift() && "Unexpected instruction type");
 244
 245   // We need constant scalar or constant splat shifts.
 246   const APInt *InnerShiftConst;
 247   if (!match(InnerShift->getOperand(1), m_APInt(InnerShiftConst)))
 248     return false;
 249
 250   // Two logical shifts in the same direction:
 251   // shl (shl X, C1), C2 -->  shl X, C1 + C2
 252   // lshr (lshr X, C1), C2 --> lshr X, C1 + C2
 253   bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl;
 254   if (IsInnerShl == IsOuterShl)
 255     return true;
 256
 257   // Equal shift amounts in opposite directions become bitwise 'and':
 258   // lshr (shl X, C), C --> and X, C'
 259   // shl (lshr X, C), C --> and X, C'
 260   if (*InnerShiftConst == OuterShAmt)
 261     return true;
 262
 263   // If the 2nd shift is bigger than the 1st, we can fold:
 264   // lshr (shl X, C1), C2 -->  and (shl X, C1 - C2), C3
 265   // shl (lshr X, C1), C2 --> and (lshr X, C1 - C2), C3
 266   // but it isn't profitable unless we know the and'd out bits are already zero.
 267   // Also, check that the inner shift is valid (less than the type width) or
 268   // we'll crash trying to produce the bit mask for the 'and'.
 269   unsigned TypeWidth = InnerShift->getType()->getScalarSizeInBits();
 270   if (InnerShiftConst->ugt(OuterShAmt) && InnerShiftConst->ult(TypeWidth)) {
 271     unsigned InnerShAmt = InnerShiftConst->getZExtValue();
 272     unsigned MaskShift =
 273         IsInnerShl ? TypeWidth - InnerShAmt : InnerShAmt - OuterShAmt;
 274     APInt Mask = APInt::getLowBitsSet(TypeWidth, OuterShAmt) << MaskShift;
 275     if (IC.MaskedValueIsZero(InnerShift->getOperand(0), Mask, 0, CxtI))
 276       return true;
 277   }
 278
 279   return false;
 280 }
 281
 282 /// See if we can compute the specified value, but shifted logically to the left
 283 /// or right by some number of bits. This should return true if the expression
 284 /// can be computed for the same cost as the current expression tree. This is
 285 /// used to eliminate extraneous shifting from things like:
 286 ///      %C = shl i128 %A, 64
 287 ///      %D = shl i128 %B, 96
 288 ///      %E = or i128 %C, %D
 289 ///      %F = lshr i128 %E, 64
 290 /// where the client will ask if E can be computed shifted right by 64-bits. If
 291 /// this succeeds, getShiftedValue() will be called to produce the value.
 292 static bool canEvaluateShifted(Value *V, unsigned NumBits, bool IsLeftShift,
 293                                InstCombiner &IC, Instruction *CxtI) {
 294   // We can always evaluate constants shifted.
 295   if (isa<Constant>(V))
 296     return true;
 297
 298   Instruction *I = dyn_cast<Instruction>(V);
 299   if (!I) return false;
 300
 301   // If this is the opposite shift, we can directly reuse the input of the shift
 302   // if the needed bits are already zero in the input.  This allows us to reuse
 303   // the value which means that we don't care if the shift has multiple uses.
 304   //  TODO:  Handle opposite shift by exact value.
 305   ConstantInt *CI = nullptr;
 306   if ((IsLeftShift && match(I, m_LShr(m_Value(), m_ConstantInt(CI)))) ||
 307       (!IsLeftShift && match(I, m_Shl(m_Value(), m_ConstantInt(CI))))) {
 308     if (CI->getValue() == NumBits) {
 309       // TODO: Check that the input bits are already zero with MaskedValueIsZero
 310 #if 0
 311       // If this is a truncate of a logical shr, we can truncate it to a smaller
 312       // lshr iff we know that the bits we would otherwise be shifting in are
 313       // already zeros.
 314       uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
 315       uint32_t BitWidth = Ty->getScalarSizeInBits();
 316       if (MaskedValueIsZero(I->getOperand(0),
 317             APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
 318           CI->getLimitedValue(BitWidth) < BitWidth) {
 319         return CanEvaluateTruncated(I->getOperand(0), Ty);
 320       }
 321 #endif
 322
 323     }
 324   }
 325
 326   // We can't mutate something that has multiple uses: doing so would
 327   // require duplicating the instruction in general, which isn't profitable.
 328   if (!I->hasOneUse()) return false;
 329
 330   switch (I->getOpcode()) {
 331   default: return false;
 332   case Instruction::And:
 333   case Instruction::Or:
 334   case Instruction::Xor:
 335     // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
 336     return canEvaluateShifted(I->getOperand(0), NumBits, IsLeftShift, IC, I) &&
 337            canEvaluateShifted(I->getOperand(1), NumBits, IsLeftShift, IC, I);
 338
 339   case Instruction::Shl:
 340   case Instruction::LShr:
 341     return canEvaluateShiftedShift(NumBits, IsLeftShift, I, IC, CxtI);
 342
 343   case Instruction::Select: {
 344     SelectInst *SI = cast<SelectInst>(I);
 345     Value *TrueVal = SI->getTrueValue();
 346     Value *FalseVal = SI->getFalseValue();
 347     return canEvaluateShifted(TrueVal, NumBits, IsLeftShift, IC, SI) &&
 348            canEvaluateShifted(FalseVal, NumBits, IsLeftShift, IC, SI);
 349   }
 350   case Instruction::PHI: {
 351     // We can change a phi if we can change all operands.  Note that we never
 352     // get into trouble with cyclic PHIs here because we only consider
 353     // instructions with a single use.
 354     PHINode *PN = cast<PHINode>(I);
 355     for (Value *IncValue : PN->incoming_values())
 356       if (!canEvaluateShifted(IncValue, NumBits, IsLeftShift, IC, PN))
 357         return false;
 358     return true;
 359   }
 360   }
 361 }
 362
 363 /// Fold OuterShift (InnerShift X, C1), C2.
 364 /// See canEvaluateShiftedShift() for the constraints on these instructions.
 365 static Value *foldShiftedShift(BinaryOperator *InnerShift, unsigned OuterShAmt,
 366                                bool IsOuterShl,
 367                                InstCombiner::BuilderTy &Builder) {
 368   bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl;
 369   Type *ShType = InnerShift->getType();
 370   unsigned TypeWidth = ShType->getScalarSizeInBits();
 371
 372   // We only accept shifts-by-a-constant in canEvaluateShifted().
 373   const APInt *C1;
 374   match(InnerShift->getOperand(1), m_APInt(C1));
 375   unsigned InnerShAmt = C1->getZExtValue();
 376
 377   // Change the shift amount and clear the appropriate IR flags.
 378   auto NewInnerShift = [&](unsigned ShAmt) {
 379     InnerShift->setOperand(1, ConstantInt::get(ShType, ShAmt));
 380     if (IsInnerShl) {
 381       InnerShift->setHasNoUnsignedWrap(false);
 382       InnerShift->setHasNoSignedWrap(false);
 383     } else {
 384       InnerShift->setIsExact(false);
 385     }
 386     return InnerShift;
 387   };
 388
 389   // Two logical shifts in the same direction:
 390   // shl (shl X, C1), C2 -->  shl X, C1 + C2
 391   // lshr (lshr X, C1), C2 --> lshr X, C1 + C2
 392   if (IsInnerShl == IsOuterShl) {
 393     // If this is an oversized composite shift, then unsigned shifts get 0.
 394     if (InnerShAmt + OuterShAmt >= TypeWidth)
 395       return Constant::getNullValue(ShType);
 396
 397     return NewInnerShift(InnerShAmt + OuterShAmt);
 398   }
 399
 400   // Equal shift amounts in opposite directions become bitwise 'and':
 401   // lshr (shl X, C), C --> and X, C'
 402   // shl (lshr X, C), C --> and X, C'
 403   if (InnerShAmt == OuterShAmt) {
 404     APInt Mask = IsInnerShl
 405                      ? APInt::getLowBitsSet(TypeWidth, TypeWidth - OuterShAmt)
 406                      : APInt::getHighBitsSet(TypeWidth, TypeWidth - OuterShAmt);
 407     Value *And = Builder.CreateAnd(InnerShift->getOperand(0),
 408                                    ConstantInt::get(ShType, Mask));
 409     if (auto *AndI = dyn_cast<Instruction>(And)) {
 410       AndI->moveBefore(InnerShift);
 411       AndI->takeName(InnerShift);
 412     }
 413     return And;
 414   }
 415
 416   assert(InnerShAmt > OuterShAmt &&
 417          "Unexpected opposite direction logical shift pair");
 418
 419   // In general, we would need an 'and' for this transform, but
 420   // canEvaluateShiftedShift() guarantees that the masked-off bits are not used.
 421   // lshr (shl X, C1), C2 -->  shl X, C1 - C2
 422   // shl (lshr X, C1), C2 --> lshr X, C1 - C2
 423   return NewInnerShift(InnerShAmt - OuterShAmt);
 424 }
 425
 426 /// When canEvaluateShifted() returns true for an expression, this function
 427 /// inserts the new computation that produces the shifted value.
 428 static Value *getShiftedValue(Value *V, unsigned NumBits, bool isLeftShift,
 429                               InstCombiner &IC, const DataLayout &DL) {
 430   // We can always evaluate constants shifted.
 431   if (Constant *C = dyn_cast<Constant>(V)) {
 432     if (isLeftShift)
 433       V = IC.Builder.CreateShl(C, NumBits);
 434     else
 435       V = IC.Builder.CreateLShr(C, NumBits);
 436     // If we got a constantexpr back, try to simplify it with TD info.
 437     if (auto *C = dyn_cast<Constant>(V))
 438       if (auto *FoldedC =
 439               ConstantFoldConstant(C, DL, &IC.getTargetLibraryInfo()))
 440         V = FoldedC;
 441     return V;
 442   }
 443
 444   Instruction *I = cast<Instruction>(V);
 445   IC.Worklist.Add(I);
 446
 447   switch (I->getOpcode()) {
 448   default: llvm_unreachable("Inconsistency with CanEvaluateShifted");
 449   case Instruction::And:
 450   case Instruction::Or:
 451   case Instruction::Xor:
 452     // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
 453     I->setOperand(
 454         0, getShiftedValue(I->getOperand(0), NumBits, isLeftShift, IC, DL));
 455     I->setOperand(
 456         1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL));
 457     return I;
 458
 459   case Instruction::Shl:
 460   case Instruction::LShr:
 461     return foldShiftedShift(cast<BinaryOperator>(I), NumBits, isLeftShift,
 462                             IC.Builder);
 463
 464   case Instruction::Select:
 465     I->setOperand(
 466         1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL));
 467     I->setOperand(
 468         2, getShiftedValue(I->getOperand(2), NumBits, isLeftShift, IC, DL));
 469     return I;
 470   case Instruction::PHI: {
 471     // We can change a phi if we can change all operands.  Note that we never
 472     // get into trouble with cyclic PHIs here because we only consider
 473     // instructions with a single use.
 474     PHINode *PN = cast<PHINode>(I);
 475     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
 476       PN->setIncomingValue(i, getShiftedValue(PN->getIncomingValue(i), NumBits,
 477                                               isLeftShift, IC, DL));
 478     return PN;
 479   }
 480   }
 481 }
 482
 483 // If this is a bitwise operator or add with a constant RHS we might be able
 484 // to pull it through a shift.
 485 static bool canShiftBinOpWithConstantRHS(BinaryOperator &Shift,
 486                                          BinaryOperator *BO) {
 487   switch (BO->getOpcode()) {
 488   default:
 489     return false; // Do not perform transform!
 490   case Instruction::Add:
 491     return Shift.getOpcode() == Instruction::Shl;
 492   case Instruction::Or:
 493   case Instruction::Xor:
 494   case Instruction::And:
 495     return true;
 496   }
 497 }
 498
 499 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, Constant *Op1,
 500                                                BinaryOperator &I) {
 501   bool isLeftShift = I.getOpcode() == Instruction::Shl;
 502
 503   const APInt *Op1C;
 504   if (!match(Op1, m_APInt(Op1C)))
 505     return nullptr;
 506
 507   // See if we can propagate this shift into the input, this covers the trivial
 508   // cast of lshr(shl(x,c1),c2) as well as other more complex cases.
 509   if (I.getOpcode() != Instruction::AShr &&
 510       canEvaluateShifted(Op0, Op1C->getZExtValue(), isLeftShift, *this, &I)) {
 511     LLVM_DEBUG(
 512         dbgs() << "ICE: GetShiftedValue propagating shift through expression"
 513                   " to eliminate shift:\n  IN: "
 514                << *Op0 << "\n  SH: " << I << "\n");
 515
 516     return replaceInstUsesWith(
 517         I, getShiftedValue(Op0, Op1C->getZExtValue(), isLeftShift, *this, DL));
 518   }
 519
 520   // See if we can simplify any instructions used by the instruction whose sole
 521   // purpose is to compute bits we don't care about.
 522   unsigned TypeBits = Op0->getType()->getScalarSizeInBits();
 523
 524   assert(!Op1C->uge(TypeBits) &&
 525          "Shift over the type width should have been removed already");
 526
 527   if (Instruction *FoldedShift = foldBinOpIntoSelectOrPhi(I))
 528     return FoldedShift;
 529
 530   // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
 531   if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
 532     Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
 533     // If 'shift2' is an ashr, we would have to get the sign bit into a funny
 534     // place.  Don't try to do this transformation in this case.  Also, we
 535     // require that the input operand is a shift-by-constant so that we have
 536     // confidence that the shifts will get folded together.  We could do this
 537     // xform in more cases, but it is unlikely to be profitable.
 538     if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
 539         isa<ConstantInt>(TrOp->getOperand(1))) {
 540       // Okay, we'll do this xform.  Make the shift of shift.
 541       Constant *ShAmt =
 542           ConstantExpr::getZExt(cast<Constant>(Op1), TrOp->getType());
 543       // (shift2 (shift1 & 0x00FF), c2)
 544       Value *NSh = Builder.CreateBinOp(I.getOpcode(), TrOp, ShAmt, I.getName());
 545
 546       // For logical shifts, the truncation has the effect of making the high
 547       // part of the register be zeros.  Emulate this by inserting an AND to
 548       // clear the top bits as needed.  This 'and' will usually be zapped by
 549       // other xforms later if dead.
 550       unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
 551       unsigned DstSize = TI->getType()->getScalarSizeInBits();
 552       APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
 553
 554       // The mask we constructed says what the trunc would do if occurring
 555       // between the shifts.  We want to know the effect *after* the second
 556       // shift.  We know that it is a logical shift by a constant, so adjust the
 557       // mask as appropriate.
 558       if (I.getOpcode() == Instruction::Shl)
 559         MaskV <<= Op1C->getZExtValue();
 560       else {
 561         assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
 562         MaskV.lshrInPlace(Op1C->getZExtValue());
 563       }
 564
 565       // shift1 & 0x00FF
 566       Value *And = Builder.CreateAnd(NSh,
 567                                      ConstantInt::get(I.getContext(), MaskV),
 568                                      TI->getName());
 569
 570       // Return the value truncated to the interesting size.
 571       return new TruncInst(And, I.getType());
 572     }
 573   }
 574
 575   if (Op0->hasOneUse()) {
 576     if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
 577       // Turn ((X >> C) + Y) << C  ->  (X + (Y << C)) & (~0 << C)
 578       Value *V1, *V2;
 579       ConstantInt *CC;
 580       switch (Op0BO->getOpcode()) {
 581       default: break;
 582       case Instruction::Add:
 583       case Instruction::And:
 584       case Instruction::Or:
 585       case Instruction::Xor: {
 586         // These operators commute.
 587         // Turn (Y + (X >> C)) << C  ->  (X + (Y << C)) & (~0 << C)
 588         if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
 589             match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
 590                   m_Specific(Op1)))) {
 591           Value *YS =         // (Y << C)
 592             Builder.CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
 593           // (X + (Y << C))
 594           Value *X = Builder.CreateBinOp(Op0BO->getOpcode(), YS, V1,
 595                                          Op0BO->getOperand(1)->getName());
 596           unsigned Op1Val = Op1C->getLimitedValue(TypeBits);
 597
 598           APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val);
 599           Constant *Mask = ConstantInt::get(I.getContext(), Bits);
 600           if (VectorType *VT = dyn_cast<VectorType>(X->getType()))
 601             Mask = ConstantVector::getSplat(VT->getNumElements(), Mask);
 602           return BinaryOperator::CreateAnd(X, Mask);
 603         }
 604
 605         // Turn (Y + ((X >> C) & CC)) << C  ->  ((X & (CC << C)) + (Y << C))
 606         Value *Op0BOOp1 = Op0BO->getOperand(1);
 607         if (isLeftShift && Op0BOOp1->hasOneUse() &&
 608             match(Op0BOOp1,
 609                   m_And(m_OneUse(m_Shr(m_Value(V1), m_Specific(Op1))),
 610                         m_ConstantInt(CC)))) {
 611           Value *YS =   // (Y << C)
 612             Builder.CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
 613           // X & (CC << C)
 614           Value *XM = Builder.CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
 615                                         V1->getName()+".mask");
 616           return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
 617         }
 618         LLVM_FALLTHROUGH;
 619       }
 620
 621       case Instruction::Sub: {
 622         // Turn ((X >> C) + Y) << C  ->  (X + (Y << C)) & (~0 << C)
 623         if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
 624             match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
 625                   m_Specific(Op1)))) {
 626           Value *YS =  // (Y << C)
 627             Builder.CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
 628           // (X + (Y << C))
 629           Value *X = Builder.CreateBinOp(Op0BO->getOpcode(), V1, YS,
 630                                          Op0BO->getOperand(0)->getName());
 631           unsigned Op1Val = Op1C->getLimitedValue(TypeBits);
 632
 633           APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val);
 634           Constant *Mask = ConstantInt::get(I.getContext(), Bits);
 635           if (VectorType *VT = dyn_cast<VectorType>(X->getType()))
 636             Mask = ConstantVector::getSplat(VT->getNumElements(), Mask);
 637           return BinaryOperator::CreateAnd(X, Mask);
 638         }
 639
 640         // Turn (((X >> C)&CC) + Y) << C  ->  (X + (Y << C)) & (CC << C)
 641         if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
 642             match(Op0BO->getOperand(0),
 643                   m_And(m_OneUse(m_Shr(m_Value(V1), m_Value(V2))),
 644                         m_ConstantInt(CC))) && V2 == Op1) {
 645           Value *YS = // (Y << C)
 646             Builder.CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
 647           // X & (CC << C)
 648           Value *XM = Builder.CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
 649                                         V1->getName()+".mask");
 650
 651           return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
 652         }
 653
 654         break;
 655       }
 656       }
 657
 658
 659       // If the operand is a bitwise operator with a constant RHS, and the
 660       // shift is the only use, we can pull it out of the shift.
 661       const APInt *Op0C;
 662       if (match(Op0BO->getOperand(1), m_APInt(Op0C))) {
 663         if (canShiftBinOpWithConstantRHS(I, Op0BO)) {
 664           Constant *NewRHS = ConstantExpr::get(I.getOpcode(),
 665                                      cast<Constant>(Op0BO->getOperand(1)), Op1);
 666
 667           Value *NewShift =
 668             Builder.CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
 669           NewShift->takeName(Op0BO);
 670
 671           return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
 672                                         NewRHS);
 673         }
 674       }
 675
 676       // If the operand is a subtract with a constant LHS, and the shift
 677       // is the only use, we can pull it out of the shift.
 678       // This folds (shl (sub C1, X), C2) -> (sub (C1 << C2), (shl X, C2))
 679       if (isLeftShift && Op0BO->getOpcode() == Instruction::Sub &&
 680           match(Op0BO->getOperand(0), m_APInt(Op0C))) {
 681         Constant *NewRHS = ConstantExpr::get(I.getOpcode(),
 682                                    cast<Constant>(Op0BO->getOperand(0)), Op1);
 683
 684         Value *NewShift = Builder.CreateShl(Op0BO->getOperand(1), Op1);
 685         NewShift->takeName(Op0BO);
 686
 687         return BinaryOperator::CreateSub(NewRHS, NewShift);
 688       }
 689     }
 690
 691     // If we have a select that conditionally executes some binary operator,
 692     // see if we can pull it the select and operator through the shift.
 693     //
 694     // For example, turning:
 695     //   shl (select C, (add X, C1), X), C2
 696     // Into:
 697     //   Y = shl X, C2
 698     //   select C, (add Y, C1 << C2), Y
 699     Value *Cond;
 700     BinaryOperator *TBO;
 701     Value *FalseVal;
 702     if (match(Op0, m_Select(m_Value(Cond), m_OneUse(m_BinOp(TBO)),
 703                             m_Value(FalseVal)))) {
 704       const APInt *C;
 705       if (!isa<Constant>(FalseVal) && TBO->getOperand(0) == FalseVal &&
 706           match(TBO->getOperand(1), m_APInt(C)) &&
 707           canShiftBinOpWithConstantRHS(I, TBO)) {
 708         Constant *NewRHS = ConstantExpr::get(I.getOpcode(),
 709                                        cast<Constant>(TBO->getOperand(1)), Op1);
 710
 711         Value *NewShift =
 712           Builder.CreateBinOp(I.getOpcode(), FalseVal, Op1);
 713         Value *NewOp = Builder.CreateBinOp(TBO->getOpcode(), NewShift,
 714                                            NewRHS);
 715         return SelectInst::Create(Cond, NewOp, NewShift);
 716       }
 717     }
 718
 719     BinaryOperator *FBO;
 720     Value *TrueVal;
 721     if (match(Op0, m_Select(m_Value(Cond), m_Value(TrueVal),
 722                             m_OneUse(m_BinOp(FBO))))) {
 723       const APInt *C;
 724       if (!isa<Constant>(TrueVal) && FBO->getOperand(0) == TrueVal &&
 725           match(FBO->getOperand(1), m_APInt(C)) &&
 726           canShiftBinOpWithConstantRHS(I, FBO)) {
 727         Constant *NewRHS = ConstantExpr::get(I.getOpcode(),
 728                                        cast<Constant>(FBO->getOperand(1)), Op1);
 729
 730         Value *NewShift =
 731           Builder.CreateBinOp(I.getOpcode(), TrueVal, Op1);
 732         Value *NewOp = Builder.CreateBinOp(FBO->getOpcode(), NewShift,
 733                                            NewRHS);
 734         return SelectInst::Create(Cond, NewShift, NewOp);
 735       }
 736     }
 737   }
 738
 739   return nullptr;
 740 }
 741
 742 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
 743   if (Value *V = SimplifyShlInst(I.getOperand(0), I.getOperand(1),
 744                                  I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
 745                                  SQ.getWithInstruction(&I)))
 746     return replaceInstUsesWith(I, V);
 747
 748   if (Instruction *X = foldVectorBinop(I))
 749     return X;
 750
 751   if (Instruction *V = commonShiftTransforms(I))
 752     return V;
 753
 754   if (Instruction *V = dropRedundantMaskingOfLeftShiftInput(&I, SQ))
 755     return V;
 756
 757   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
 758   Type *Ty = I.getType();
 759   unsigned BitWidth = Ty->getScalarSizeInBits();
 760
 761   const APInt *ShAmtAPInt;
 762   if (match(Op1, m_APInt(ShAmtAPInt))) {
 763     unsigned ShAmt = ShAmtAPInt->getZExtValue();
 764
 765     // shl (zext X), ShAmt --> zext (shl X, ShAmt)
 766     // This is only valid if X would have zeros shifted out.
 767     Value *X;
 768     if (match(Op0, m_OneUse(m_ZExt(m_Value(X))))) {
 769       unsigned SrcWidth = X->getType()->getScalarSizeInBits();
 770       if (ShAmt < SrcWidth &&
 771           MaskedValueIsZero(X, APInt::getHighBitsSet(SrcWidth, ShAmt), 0, &I))
 772         return new ZExtInst(Builder.CreateShl(X, ShAmt), Ty);
 773     }
 774
 775     // (X >> C) << C --> X & (-1 << C)
 776     if (match(Op0, m_Shr(m_Value(X), m_Specific(Op1)))) {
 777       APInt Mask(APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt));
 778       return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask));
 779     }
 780
 781     // FIXME: we do not yet transform non-exact shr's. The backend (DAGCombine)
 782     // needs a few fixes for the rotate pattern recognition first.
 783     const APInt *ShOp1;
 784     if (match(Op0, m_Exact(m_Shr(m_Value(X), m_APInt(ShOp1))))) {
 785       unsigned ShrAmt = ShOp1->getZExtValue();
 786       if (ShrAmt < ShAmt) {
 787         // If C1 < C2: (X >>?,exact C1) << C2 --> X << (C2 - C1)
 788         Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShrAmt);
 789         auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
 790         NewShl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
 791         NewShl->setHasNoSignedWrap(I.hasNoSignedWrap());
 792         return NewShl;
 793       }
 794       if (ShrAmt > ShAmt) {
 795         // If C1 > C2: (X >>?exact C1) << C2 --> X >>?exact (C1 - C2)
 796         Constant *ShiftDiff = ConstantInt::get(Ty, ShrAmt - ShAmt);
 797         auto *NewShr = BinaryOperator::Create(
 798             cast<BinaryOperator>(Op0)->getOpcode(), X, ShiftDiff);
 799         NewShr->setIsExact(true);
 800         return NewShr;
 801       }
 802     }
 803
 804     if (match(Op0, m_Shl(m_Value(X), m_APInt(ShOp1)))) {
 805       unsigned AmtSum = ShAmt + ShOp1->getZExtValue();
 806       // Oversized shifts are simplified to zero in InstSimplify.
 807       if (AmtSum < BitWidth)
 808         // (X << C1) << C2 --> X << (C1 + C2)
 809         return BinaryOperator::CreateShl(X, ConstantInt::get(Ty, AmtSum));
 810     }
 811
 812     // If the shifted-out value is known-zero, then this is a NUW shift.
 813     if (!I.hasNoUnsignedWrap() &&
 814         MaskedValueIsZero(Op0, APInt::getHighBitsSet(BitWidth, ShAmt), 0, &I)) {
 815       I.setHasNoUnsignedWrap();
 816       return &I;
 817     }
 818
 819     // If the shifted-out value is all signbits, then this is a NSW shift.
 820     if (!I.hasNoSignedWrap() && ComputeNumSignBits(Op0, 0, &I) > ShAmt) {
 821       I.setHasNoSignedWrap();
 822       return &I;
 823     }
 824   }
 825
 826   // Transform  (x >> y) << y  to  x & (-1 << y)
 827   // Valid for any type of right-shift.
 828   Value *X;
 829   if (match(Op0, m_OneUse(m_Shr(m_Value(X), m_Specific(Op1))))) {
 830     Constant *AllOnes = ConstantInt::getAllOnesValue(Ty);
 831     Value *Mask = Builder.CreateShl(AllOnes, Op1);
 832     return BinaryOperator::CreateAnd(Mask, X);
 833   }
 834
 835   Constant *C1;
 836   if (match(Op1, m_Constant(C1))) {
 837     Constant *C2;
 838     Value *X;
 839     // (C2 << X) << C1 --> (C2 << C1) << X
 840     if (match(Op0, m_OneUse(m_Shl(m_Constant(C2), m_Value(X)))))
 841       return BinaryOperator::CreateShl(ConstantExpr::getShl(C2, C1), X);
 842
 843     // (X * C2) << C1 --> X * (C2 << C1)
 844     if (match(Op0, m_Mul(m_Value(X), m_Constant(C2))))
 845       return BinaryOperator::CreateMul(X, ConstantExpr::getShl(C2, C1));
 846   }
 847
 848   // (1 << (C - x)) -> ((1 << C) >> x) if C is bitwidth - 1
 849   if (match(Op0, m_One()) &&
 850       match(Op1, m_Sub(m_SpecificInt(BitWidth - 1), m_Value(X))))
 851     return BinaryOperator::CreateLShr(
 852         ConstantInt::get(Ty, APInt::getSignMask(BitWidth)), X);
 853
 854   return nullptr;
 855 }
 856
 857 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
 858   if (Value *V = SimplifyLShrInst(I.getOperand(0), I.getOperand(1), I.isExact(),
 859                                   SQ.getWithInstruction(&I)))
 860     return replaceInstUsesWith(I, V);
 861
 862   if (Instruction *X = foldVectorBinop(I))
 863     return X;
 864
 865   if (Instruction *R = commonShiftTransforms(I))
 866     return R;
 867
 868   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
 869   Type *Ty = I.getType();
 870   const APInt *ShAmtAPInt;
 871   if (match(Op1, m_APInt(ShAmtAPInt))) {
 872     unsigned ShAmt = ShAmtAPInt->getZExtValue();
 873     unsigned BitWidth = Ty->getScalarSizeInBits();
 874     auto *II = dyn_cast<IntrinsicInst>(Op0);
 875     if (II && isPowerOf2_32(BitWidth) && Log2_32(BitWidth) == ShAmt &&
 876         (II->getIntrinsicID() == Intrinsic::ctlz ||
 877          II->getIntrinsicID() == Intrinsic::cttz ||
 878          II->getIntrinsicID() == Intrinsic::ctpop)) {
 879       // ctlz.i32(x)>>5  --> zext(x == 0)
 880       // cttz.i32(x)>>5  --> zext(x == 0)
 881       // ctpop.i32(x)>>5 --> zext(x == -1)
 882       bool IsPop = II->getIntrinsicID() == Intrinsic::ctpop;
 883       Constant *RHS = ConstantInt::getSigned(Ty, IsPop ? -1 : 0);
 884       Value *Cmp = Builder.CreateICmpEQ(II->getArgOperand(0), RHS);
 885       return new ZExtInst(Cmp, Ty);
 886     }
 887
 888     Value *X;
 889     const APInt *ShOp1;
 890     if (match(Op0, m_Shl(m_Value(X), m_APInt(ShOp1))) && ShOp1->ult(BitWidth)) {
 891       if (ShOp1->ult(ShAmt)) {
 892         unsigned ShlAmt = ShOp1->getZExtValue();
 893         Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt);
 894         if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) {
 895           // (X <<nuw C1) >>u C2 --> X >>u (C2 - C1)
 896           auto *NewLShr = BinaryOperator::CreateLShr(X, ShiftDiff);
 897           NewLShr->setIsExact(I.isExact());
 898           return NewLShr;
 899         }
 900         // (X << C1) >>u C2  --> (X >>u (C2 - C1)) & (-1 >> C2)
 901         Value *NewLShr = Builder.CreateLShr(X, ShiftDiff, "", I.isExact());
 902         APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt));
 903         return BinaryOperator::CreateAnd(NewLShr, ConstantInt::get(Ty, Mask));
 904       }
 905       if (ShOp1->ugt(ShAmt)) {
 906         unsigned ShlAmt = ShOp1->getZExtValue();
 907         Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt);
 908         if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) {
 909           // (X <<nuw C1) >>u C2 --> X <<nuw (C1 - C2)
 910           auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
 911           NewShl->setHasNoUnsignedWrap(true);
 912           return NewShl;
 913         }
 914         // (X << C1) >>u C2  --> X << (C1 - C2) & (-1 >> C2)
 915         Value *NewShl = Builder.CreateShl(X, ShiftDiff);
 916         APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt));
 917         return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask));
 918       }
 919       assert(*ShOp1 == ShAmt);
 920       // (X << C) >>u C --> X & (-1 >>u C)
 921       APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt));
 922       return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask));
 923     }
 924
 925     if (match(Op0, m_OneUse(m_ZExt(m_Value(X)))) &&
 926         (!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType()))) {
 927       assert(ShAmt < X->getType()->getScalarSizeInBits() &&
 928              "Big shift not simplified to zero?");
 929       // lshr (zext iM X to iN), C --> zext (lshr X, C) to iN
 930       Value *NewLShr = Builder.CreateLShr(X, ShAmt);
 931       return new ZExtInst(NewLShr, Ty);
 932     }
 933
 934     if (match(Op0, m_SExt(m_Value(X))) &&
 935         (!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType()))) {
 936       // Are we moving the sign bit to the low bit and widening with high zeros?
 937       unsigned SrcTyBitWidth = X->getType()->getScalarSizeInBits();
 938       if (ShAmt == BitWidth - 1) {
 939         // lshr (sext i1 X to iN), N-1 --> zext X to iN
 940         if (SrcTyBitWidth == 1)
 941           return new ZExtInst(X, Ty);
 942
 943         // lshr (sext iM X to iN), N-1 --> zext (lshr X, M-1) to iN
 944         if (Op0->hasOneUse()) {
 945           Value *NewLShr = Builder.CreateLShr(X, SrcTyBitWidth - 1);
 946           return new ZExtInst(NewLShr, Ty);
 947         }
 948       }
 949
 950       // lshr (sext iM X to iN), N-M --> zext (ashr X, min(N-M, M-1)) to iN
 951       if (ShAmt == BitWidth - SrcTyBitWidth && Op0->hasOneUse()) {
 952         // The new shift amount can't be more than the narrow source type.
 953         unsigned NewShAmt = std::min(ShAmt, SrcTyBitWidth - 1);
 954         Value *AShr = Builder.CreateAShr(X, NewShAmt);
 955         return new ZExtInst(AShr, Ty);
 956       }
 957     }
 958
 959     if (match(Op0, m_LShr(m_Value(X), m_APInt(ShOp1)))) {
 960       unsigned AmtSum = ShAmt + ShOp1->getZExtValue();
 961       // Oversized shifts are simplified to zero in InstSimplify.
 962       if (AmtSum < BitWidth)
 963         // (X >>u C1) >>u C2 --> X >>u (C1 + C2)
 964         return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
 965     }
 966
 967     // If the shifted-out value is known-zero, then this is an exact shift.
 968     if (!I.isExact() &&
 969         MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmt), 0, &I)) {
 970       I.setIsExact();
 971       return &I;
 972     }
 973   }
 974
 975   // Transform  (x << y) >> y  to  x & (-1 >> y)
 976   Value *X;
 977   if (match(Op0, m_OneUse(m_Shl(m_Value(X), m_Specific(Op1))))) {
 978     Constant *AllOnes = ConstantInt::getAllOnesValue(Ty);
 979     Value *Mask = Builder.CreateLShr(AllOnes, Op1);
 980     return BinaryOperator::CreateAnd(Mask, X);
 981   }
 982
 983   return nullptr;
 984 }
 985
 986 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
 987   if (Value *V = SimplifyAShrInst(I.getOperand(0), I.getOperand(1), I.isExact(),
 988                                   SQ.getWithInstruction(&I)))
 989     return replaceInstUsesWith(I, V);
 990
 991   if (Instruction *X = foldVectorBinop(I))
 992     return X;
 993
 994   if (Instruction *R = commonShiftTransforms(I))
 995     return R;
 996
 997   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
 998   Type *Ty = I.getType();
 999   unsigned BitWidth = Ty->getScalarSizeInBits();
1000   const APInt *ShAmtAPInt;
1001   if (match(Op1, m_APInt(ShAmtAPInt)) && ShAmtAPInt->ult(BitWidth)) {
1002     unsigned ShAmt = ShAmtAPInt->getZExtValue();
1003
1004     // If the shift amount equals the difference in width of the destination
1005     // and source scalar types:
1006     // ashr (shl (zext X), C), C --> sext X
1007     Value *X;
1008     if (match(Op0, m_Shl(m_ZExt(m_Value(X)), m_Specific(Op1))) &&
1009         ShAmt == BitWidth - X->getType()->getScalarSizeInBits())
1010       return new SExtInst(X, Ty);
1011
1012     // We can't handle (X << C1) >>s C2. It shifts arbitrary bits in. However,
1013     // we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits.
1014     const APInt *ShOp1;
1015     if (match(Op0, m_NSWShl(m_Value(X), m_APInt(ShOp1))) &&
1016         ShOp1->ult(BitWidth)) {
1017       unsigned ShlAmt = ShOp1->getZExtValue();
1018       if (ShlAmt < ShAmt) {
1019         // (X <<nsw C1) >>s C2 --> X >>s (C2 - C1)
1020         Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt);
1021         auto *NewAShr = BinaryOperator::CreateAShr(X, ShiftDiff);
1022         NewAShr->setIsExact(I.isExact());
1023         return NewAShr;
1024       }
1025       if (ShlAmt > ShAmt) {
1026         // (X <<nsw C1) >>s C2 --> X <<nsw (C1 - C2)
1027         Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt);
1028         auto *NewShl = BinaryOperator::Create(Instruction::Shl, X, ShiftDiff);
1029         NewShl->setHasNoSignedWrap(true);
1030         return NewShl;
1031       }
1032     }
1033
1034     if (match(Op0, m_AShr(m_Value(X), m_APInt(ShOp1))) &&
1035         ShOp1->ult(BitWidth)) {
1036       unsigned AmtSum = ShAmt + ShOp1->getZExtValue();
1037       // Oversized arithmetic shifts replicate the sign bit.
1038       AmtSum = std::min(AmtSum, BitWidth - 1);
1039       // (X >>s C1) >>s C2 --> X >>s (C1 + C2)
1040       return BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
1041     }
1042
1043     if (match(Op0, m_OneUse(m_SExt(m_Value(X)))) &&
1044         (Ty->isVectorTy() || shouldChangeType(Ty, X->getType()))) {
1045       // ashr (sext X), C --> sext (ashr X, C')
1046       Type *SrcTy = X->getType();
1047       ShAmt = std::min(ShAmt, SrcTy->getScalarSizeInBits() - 1);
1048       Value *NewSh = Builder.CreateAShr(X, ConstantInt::get(SrcTy, ShAmt));
1049       return new SExtInst(NewSh, Ty);
1050     }
1051
1052     // If the shifted-out value is known-zero, then this is an exact shift.
1053     if (!I.isExact() &&
1054         MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmt), 0, &I)) {
1055       I.setIsExact();
1056       return &I;
1057     }
1058   }
1059
1060   // See if we can turn a signed shr into an unsigned shr.
1061   if (MaskedValueIsZero(Op0, APInt::getSignMask(BitWidth), 0, &I))
1062     return BinaryOperator::CreateLShr(Op0, Op1);
1063
1064   return nullptr;
1065 }