Esempio n. 1
0
bool AMDGPUCodeGenPrepare::promoteUniformOpToI32(BinaryOperator &I) const {
  assert(needsPromotionToI32(I.getType()) &&
         "I does not need promotion to i32");

  if (I.getOpcode() == Instruction::SDiv ||
      I.getOpcode() == Instruction::UDiv)
    return false;

  IRBuilder<> Builder(&I);
  Builder.SetCurrentDebugLocation(I.getDebugLoc());

  Type *I32Ty = getI32Ty(Builder, I.getType());
  Value *ExtOp0 = nullptr;
  Value *ExtOp1 = nullptr;
  Value *ExtRes = nullptr;
  Value *TruncRes = nullptr;

  if (isSigned(I)) {
    ExtOp0 = Builder.CreateSExt(I.getOperand(0), I32Ty);
    ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
  } else {
    ExtOp0 = Builder.CreateZExt(I.getOperand(0), I32Ty);
    ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
  }
  ExtRes = copyFlags(I, Builder.CreateBinOp(I.getOpcode(), ExtOp0, ExtOp1));
  TruncRes = Builder.CreateTrunc(ExtRes, I.getType());

  I.replaceAllUsesWith(TruncRes);
  I.eraseFromParent();

  return true;
}
Esempio n. 2
0
Value *X86RegisterSema::computeEFLAGSForDef(Value *Def, Value *OldEFLAGS,
                                            bool DontUpdateCF) {
  // FIXME: This describes the general semantics of EFLAGS update, but this
  // needs to handle the differences between instructions.
  // This would be done by keeping more information on the instruction with
  // LastEFLAGSChangingDef.
  // For now we only do DontUpdateCF, for INC/DEC instructions.

  setSF(X86::ZF, Builder->CreateIsNull(Def));

  setSF(X86::SF,
        Builder->CreateICmpSLT(Def, ConstantInt::getNullValue(Def->getType())));

  // FIXME: We need to generate AF as well.
  setSF(X86::AF, Builder->getFalse());

  // FIXME: CF/OF need a smarter trick.
  Intrinsic::ID OverflowIntrinsic = Intrinsic::not_intrinsic,
                CarryIntrinsic = Intrinsic::not_intrinsic;

  BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Def);
  if (BinOp && BinOp->getOpcode() == BinaryOperator::Add) {
    OverflowIntrinsic = Intrinsic::sadd_with_overflow;
    CarryIntrinsic = Intrinsic::uadd_with_overflow;
  } else if (BinOp && BinOp->getOpcode() == BinaryOperator::Sub) {
    OverflowIntrinsic = Intrinsic::ssub_with_overflow;
    CarryIntrinsic = Intrinsic::usub_with_overflow;
  }

  if (BinOp && OverflowIntrinsic && CarryIntrinsic) {
    Value *Args[] = { BinOp->getOperand(0), BinOp->getOperand(1) };
    setSF(X86::OF, Builder->CreateExtractValue(
                       Builder->CreateCall(
                           Intrinsic::getDeclaration(
                               TheModule, OverflowIntrinsic, BinOp->getType()),
                           Args),
                       1));
    if (!DontUpdateCF)
      setSF(X86::CF, Builder->CreateExtractValue(
                         Builder->CreateCall(
                             Intrinsic::getDeclaration(
                                 TheModule, CarryIntrinsic, BinOp->getType()),
                             Args),
                         1));
  } else {
    if (!DontUpdateCF)
      setSF(X86::CF, Builder->getFalse());
    setSF(X86::OF, Builder->getFalse());
  }

  Type *I8Ty = Builder->getInt8Ty();
  setSF(X86::PF, Builder->CreateIsNull(Builder->CreateTrunc(
                     Builder->CreateCall(Intrinsic::getDeclaration(
                                             TheModule, Intrinsic::ctpop, I8Ty),
                                         {Builder->CreateTrunc(Def, I8Ty)}),
                     Builder->getInt1Ty())));
  return createEFLAGSFromSFs(OldEFLAGS);
}
Esempio n. 3
0
//--------------------------------------------------------- 
void unrollRecordAssigns(StmtEditor& editor, Stmt* S)
{    
  const RecordType *RT;
  for (stmt_iterator<BinaryOperator> i = stmt_ibegin(S), 
       e = stmt_iend(S); i != e;)
  {
    BinaryOperator* BO = *i;
    if (BO->getOpcode() == BO_Assign && 
        (RT = BO->getType()->getAsStructureType()) != 0 && 
        editor.getStatementOfExpression(BO) == BO)  // ensures top level assign
    {
      CompoundStmt* CS = editor.ensureCompoundParent(BO);
      std::vector<Stmt*> compoundStmts(CS->child_begin(), CS->child_end());
      std::vector<Stmt*>::iterator insertPos = 
        std::find(compoundStmts.begin(), compoundStmts.end(), BO);
      assert(insertPos != compoundStmts.end());
      insertPos = compoundStmts.erase(insertPos);
      RecordAssignUnroller unroller(editor, RT->getDecl(), BO);
      compoundStmts.insert(insertPos, 
        unroller.compoundStmts.begin(), unroller.compoundStmts.end());
      editor.replaceStmts(CS, &compoundStmts[0], compoundStmts.size());
      i = stmt_ibegin(S);
    }
    else
    {
      ++i;
    }
  }
}
  void ModuloSchedulerDriverPass::foldAddInstructions(Instruction* add) {
      if (dyn_cast<BinaryOperator>(add) && add->getOpcode() == Instruction::Add) {
            BinaryOperator *bin = dyn_cast<BinaryOperator>(add);
            unsigned int bitWidth = cast<IntegerType>(bin->getType())->getBitWidth();

            Value *p0 = add->getOperand(0);//param1
            Value *p1 = add->getOperand(1);//param0
            if (dyn_cast<ConstantInt>(p0) && dyn_cast<ConstantInt>(p1)) {
                    ConstantInt *c0 = dyn_cast<ConstantInt>(p0);
                    ConstantInt *c1 = dyn_cast<ConstantInt>(p1);

                    //TODO:May overflow
                    unsigned int val = (c0->getValue().getZExtValue() + c1->getValue().getZExtValue());
                    ConstantInt *ncon = ConstantInt::get(APInt(bitWidth, val));
                    add->replaceAllUsesWith(ncon);
                    add->eraseFromParent();
                    return;
            }
            
            if (ConstantInt *c0 = dyn_cast<ConstantInt>(p0)) {
                    if (0 == c0->getValue().getZExtValue()) {
                         add->replaceAllUsesWith(p1);
                         add->eraseFromParent();
                         return;
                    }
            } 
            if (ConstantInt *c1 = dyn_cast<ConstantInt>(p1)) {
                    if (0 == c1->getValue().getZExtValue()) {
                        add->replaceAllUsesWith(p0);
                        add->eraseFromParent();
                        return;
                    }
            }
      }
  }
bool AMDGPUCodeGenPrepare::promoteUniformOpToI32(BinaryOperator &I) const {
  assert(needsPromotionToI32(I.getType()) &&
         "I does not need promotion to i32");

  if (I.getOpcode() == Instruction::SDiv ||
      I.getOpcode() == Instruction::UDiv ||
      I.getOpcode() == Instruction::SRem ||
      I.getOpcode() == Instruction::URem)
    return false;

  IRBuilder<> Builder(&I);
  Builder.SetCurrentDebugLocation(I.getDebugLoc());

  Type *I32Ty = getI32Ty(Builder, I.getType());
  Value *ExtOp0 = nullptr;
  Value *ExtOp1 = nullptr;
  Value *ExtRes = nullptr;
  Value *TruncRes = nullptr;

  if (isSigned(I)) {
    ExtOp0 = Builder.CreateSExt(I.getOperand(0), I32Ty);
    ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
  } else {
    ExtOp0 = Builder.CreateZExt(I.getOperand(0), I32Ty);
    ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
  }

  ExtRes = Builder.CreateBinOp(I.getOpcode(), ExtOp0, ExtOp1);
  if (Instruction *Inst = dyn_cast<Instruction>(ExtRes)) {
    if (promotedOpIsNSW(cast<Instruction>(I)))
      Inst->setHasNoSignedWrap();

    if (promotedOpIsNUW(cast<Instruction>(I)))
      Inst->setHasNoUnsignedWrap();

    if (const auto *ExactOp = dyn_cast<PossiblyExactOperator>(&I))
      Inst->setIsExact(ExactOp->isExact());
  }

  TruncRes = Builder.CreateTrunc(ExtRes, I.getType());

  I.replaceAllUsesWith(TruncRes);
  I.eraseFromParent();

  return true;
}
Esempio n. 6
0
void IndVarSimplify::EliminateIVRemainders() {
  // Look for SRem and URem users.
  for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
    IVStrideUse &UI = *I;
    BinaryOperator *Rem = dyn_cast<BinaryOperator>(UI.getUser());
    if (!Rem) continue;

    bool isSigned = Rem->getOpcode() == Instruction::SRem;
    if (!isSigned && Rem->getOpcode() != Instruction::URem)
      continue;

    // We're only interested in the case where we know something about
    // the numerator.
    if (UI.getOperandValToReplace() != Rem->getOperand(0))
      continue;

    // Get the SCEVs for the ICmp operands.
    const SCEV *S = SE->getSCEV(Rem->getOperand(0));
    const SCEV *X = SE->getSCEV(Rem->getOperand(1));

    // Simplify unnecessary loops away.
    const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
    S = SE->getSCEVAtScope(S, ICmpLoop);
    X = SE->getSCEVAtScope(X, ICmpLoop);

    // i % n  -->  i  if i is in [0,n).
    if ((!isSigned || SE->isKnownNonNegative(S)) &&
        SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
                             S, X))
      Rem->replaceAllUsesWith(Rem->getOperand(0));
    else {
      // (i+1) % n  -->  (i+1)==n?0:(i+1)  if i is in [0,n).
      const SCEV *LessOne =
        SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
      if ((!isSigned || SE->isKnownNonNegative(LessOne)) &&
          SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
                               LessOne, X)) {
        ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
                                      Rem->getOperand(0), Rem->getOperand(1),
                                      "tmp");
        SelectInst *Sel =
          SelectInst::Create(ICmp,
                             ConstantInt::get(Rem->getType(), 0),
                             Rem->getOperand(0), "tmp", Rem);
        Rem->replaceAllUsesWith(Sel);
      } else
        continue;
    }

    // Inform IVUsers about the new users.
    if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0)))
      IU->AddUsersIfInteresting(I);

    DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
    DeadInsts.push_back(Rem);
  }
}
Esempio n. 7
0
bool AMDGPUCodeGenPrepare::visitBinaryOperator(BinaryOperator &I) {
  bool Changed = false;

  if (ST->has16BitInsts() && needsPromotionToI32(I.getType()) &&
      DA->isUniform(&I))
    Changed |= promoteUniformOpToI32(I);

  return Changed;
}
bool AMDGPUCodeGenPrepare::visitBinaryOperator(BinaryOperator &I) {
  if (ST->has16BitInsts() && needsPromotionToI32(I.getType()) &&
      DA->isUniform(&I) && promoteUniformOpToI32(I))
    return true;

  bool Changed = false;
  Instruction::BinaryOps Opc = I.getOpcode();
  Type *Ty = I.getType();
  Value *NewDiv = nullptr;
  if ((Opc == Instruction::URem || Opc == Instruction::UDiv ||
       Opc == Instruction::SRem || Opc == Instruction::SDiv) &&
      Ty->getScalarSizeInBits() <= 32) {
    Value *Num = I.getOperand(0);
    Value *Den = I.getOperand(1);
    IRBuilder<> Builder(&I);
    Builder.SetCurrentDebugLocation(I.getDebugLoc());

    if (VectorType *VT = dyn_cast<VectorType>(Ty)) {
      NewDiv = UndefValue::get(VT);

      for (unsigned N = 0, E = VT->getNumElements(); N != E; ++N) {
        Value *NumEltN = Builder.CreateExtractElement(Num, N);
        Value *DenEltN = Builder.CreateExtractElement(Den, N);
        Value *NewElt = expandDivRem32(Builder, I, NumEltN, DenEltN);
        if (!NewElt)
          NewElt = Builder.CreateBinOp(Opc, NumEltN, DenEltN);
        NewDiv = Builder.CreateInsertElement(NewDiv, NewElt, N);
      }
    } else {
      NewDiv = expandDivRem32(Builder, I, Num, Den);
    }

    if (NewDiv) {
      I.replaceAllUsesWith(NewDiv);
      I.eraseFromParent();
      Changed = true;
    }
  }

  return Changed;
}
Esempio n. 9
0
bool IRTranslator::translateBinaryOp(unsigned Opcode,
                                     const BinaryOperator &Inst) {
  // FIXME: handle signed/unsigned wrapping flags.

  // Get or create a virtual register for each value.
  // Unless the value is a Constant => loadimm cst?
  // or inline constant each time?
  // Creation of a virtual register needs to have a size.
  unsigned Op0 = getOrCreateVReg(*Inst.getOperand(0));
  unsigned Op1 = getOrCreateVReg(*Inst.getOperand(1));
  unsigned Res = getOrCreateVReg(Inst);
  MIRBuilder.buildInstr(Opcode, LLT{*Inst.getType()})
      .addDef(Res)
      .addUse(Op0)
      .addUse(Op1);
  return true;
}
Esempio n. 10
0
void Lint::visitShl(BinaryOperator &I) {
  if (ConstantInt *CI =
          dyn_cast<ConstantInt>(findValue(I.getOperand(1), /*OffsetOk=*/false)))
    Assert(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
           "Undefined result: Shift count out of range", &I);
}
Esempio n. 11
0
/// HandleFloatingPointIV - If the loop has floating induction variable
/// then insert corresponding integer induction variable if possible.
/// For example,
/// for(double i = 0; i < 10000; ++i)
///   bar(i)
/// is converted into
/// for(int i = 0; i < 10000; ++i)
///   bar((double)i);
///
void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
  unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
  unsigned BackEdge     = IncomingEdge^1;

  // Check incoming value.
  ConstantFP *InitValueVal =
    dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));

  int64_t InitValue;
  if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
    return;

  // Check IV increment. Reject this PN if increment operation is not
  // an add or increment value can not be represented by an integer.
  BinaryOperator *Incr =
    dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
  if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
  
  // If this is not an add of the PHI with a constantfp, or if the constant fp
  // is not an integer, bail out.
  ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
  int64_t IncValue;
  if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
      !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
    return;

  // Check Incr uses. One user is PN and the other user is an exit condition
  // used by the conditional terminator.
  Value::use_iterator IncrUse = Incr->use_begin();
  Instruction *U1 = cast<Instruction>(IncrUse++);
  if (IncrUse == Incr->use_end()) return;
  Instruction *U2 = cast<Instruction>(IncrUse++);
  if (IncrUse != Incr->use_end()) return;

  // Find exit condition, which is an fcmp.  If it doesn't exist, or if it isn't
  // only used by a branch, we can't transform it.
  FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
  if (!Compare)
    Compare = dyn_cast<FCmpInst>(U2);
  if (Compare == 0 || !Compare->hasOneUse() ||
      !isa<BranchInst>(Compare->use_back()))
    return;
  
  BranchInst *TheBr = cast<BranchInst>(Compare->use_back());

  // We need to verify that the branch actually controls the iteration count
  // of the loop.  If not, the new IV can overflow and no one will notice.
  // The branch block must be in the loop and one of the successors must be out
  // of the loop.
  assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
  if (!L->contains(TheBr->getParent()) ||
      (L->contains(TheBr->getSuccessor(0)) &&
       L->contains(TheBr->getSuccessor(1))))
    return;
  
  
  // If it isn't a comparison with an integer-as-fp (the exit value), we can't
  // transform it.
  ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
  int64_t ExitValue;
  if (ExitValueVal == 0 ||
      !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
    return;
  
  // Find new predicate for integer comparison.
  CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
  switch (Compare->getPredicate()) {
  default: return;  // Unknown comparison.
  case CmpInst::FCMP_OEQ:
  case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
  case CmpInst::FCMP_ONE:
  case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
  case CmpInst::FCMP_OGT:
  case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
  case CmpInst::FCMP_OGE:
  case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
  case CmpInst::FCMP_OLT:
  case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
  case CmpInst::FCMP_OLE:
  case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
  }
  
  // We convert the floating point induction variable to a signed i32 value if
  // we can.  This is only safe if the comparison will not overflow in a way
  // that won't be trapped by the integer equivalent operations.  Check for this
  // now.
  // TODO: We could use i64 if it is native and the range requires it.
  
  // The start/stride/exit values must all fit in signed i32.
  if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
    return;

  // If not actually striding (add x, 0.0), avoid touching the code.
  if (IncValue == 0)
    return;

  // Positive and negative strides have different safety conditions.
  if (IncValue > 0) {
    // If we have a positive stride, we require the init to be less than the
    // exit value and an equality or less than comparison.
    if (InitValue >= ExitValue ||
        NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
      return;
    
    uint32_t Range = uint32_t(ExitValue-InitValue);
    if (NewPred == CmpInst::ICMP_SLE) {
      // Normalize SLE -> SLT, check for infinite loop.
      if (++Range == 0) return;  // Range overflows.
    }
    
    unsigned Leftover = Range % uint32_t(IncValue);
    
    // If this is an equality comparison, we require that the strided value
    // exactly land on the exit value, otherwise the IV condition will wrap
    // around and do things the fp IV wouldn't.
    if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
        Leftover != 0)
      return;
    
    // If the stride would wrap around the i32 before exiting, we can't
    // transform the IV.
    if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
      return;
    
  } else {
    // If we have a negative stride, we require the init to be greater than the
    // exit value and an equality or greater than comparison.
    if (InitValue >= ExitValue ||
        NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
      return;
    
    uint32_t Range = uint32_t(InitValue-ExitValue);
    if (NewPred == CmpInst::ICMP_SGE) {
      // Normalize SGE -> SGT, check for infinite loop.
      if (++Range == 0) return;  // Range overflows.
    }
    
    unsigned Leftover = Range % uint32_t(-IncValue);
    
    // If this is an equality comparison, we require that the strided value
    // exactly land on the exit value, otherwise the IV condition will wrap
    // around and do things the fp IV wouldn't.
    if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
        Leftover != 0)
      return;
    
    // If the stride would wrap around the i32 before exiting, we can't
    // transform the IV.
    if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
      return;
  }
  
  const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());

  // Insert new integer induction variable.
  PHINode *NewPHI = PHINode::Create(Int32Ty, PN->getName()+".int", PN);
  NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
                      PN->getIncomingBlock(IncomingEdge));

  Value *NewAdd =
    BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
                              Incr->getName()+".int", Incr);
  NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));

  ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
                                      ConstantInt::get(Int32Ty, ExitValue),
                                      Compare->getName());

  // In the following deletions, PN may become dead and may be deleted.
  // Use a WeakVH to observe whether this happens.
  WeakVH WeakPH = PN;

  // Delete the old floating point exit comparison.  The branch starts using the
  // new comparison.
  NewCompare->takeName(Compare);
  Compare->replaceAllUsesWith(NewCompare);
  RecursivelyDeleteTriviallyDeadInstructions(Compare);

  // Delete the old floating point increment.
  Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
  RecursivelyDeleteTriviallyDeadInstructions(Incr);

  // If the FP induction variable still has uses, this is because something else
  // in the loop uses its value.  In order to canonicalize the induction
  // variable, we chose to eliminate the IV and rewrite it in terms of an
  // int->fp cast.
  //
  // We give preference to sitofp over uitofp because it is faster on most
  // platforms.
  if (WeakPH) {
    Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
                                 PN->getParent()->getFirstNonPHI());
    PN->replaceAllUsesWith(Conv);
    RecursivelyDeleteTriviallyDeadInstructions(PN);
  }

  // Add a new IVUsers entry for the newly-created integer PHI.
  IU->AddUsersIfInteresting(NewPHI);
}
Esempio n. 12
0
// Insert an intrinsic for fast fdiv for safe math situations where we can
// reduce precision. Leave fdiv for situations where the generic node is
// expected to be optimized.
bool AMDGPUCodeGenPrepare::visitFDiv(BinaryOperator &FDiv) {
  Type *Ty = FDiv.getType();

  // TODO: Handle half
  if (!Ty->getScalarType()->isFloatTy())
    return false;

  MDNode *FPMath = FDiv.getMetadata(LLVMContext::MD_fpmath);
  if (!FPMath)
    return false;

  const FPMathOperator *FPOp = cast<const FPMathOperator>(&FDiv);
  float ULP = FPOp->getFPAccuracy();
  if (ULP < 2.5f)
    return false;

  FastMathFlags FMF = FPOp->getFastMathFlags();
  bool UnsafeDiv = HasUnsafeFPMath || FMF.unsafeAlgebra() ||
                                      FMF.allowReciprocal();
  if (ST->hasFP32Denormals() && !UnsafeDiv)
    return false;

  IRBuilder<> Builder(FDiv.getParent(), std::next(FDiv.getIterator()), FPMath);
  Builder.setFastMathFlags(FMF);
  Builder.SetCurrentDebugLocation(FDiv.getDebugLoc());

  const AMDGPUIntrinsicInfo *II = TM->getIntrinsicInfo();
  Function *Decl
    = II->getDeclaration(Mod, AMDGPUIntrinsic::amdgcn_fdiv_fast, {});

  Value *Num = FDiv.getOperand(0);
  Value *Den = FDiv.getOperand(1);

  Value *NewFDiv = nullptr;

  if (VectorType *VT = dyn_cast<VectorType>(Ty)) {
    NewFDiv = UndefValue::get(VT);

    // FIXME: Doesn't do the right thing for cases where the vector is partially
    // constant. This works when the scalarizer pass is run first.
    for (unsigned I = 0, E = VT->getNumElements(); I != E; ++I) {
      Value *NumEltI = Builder.CreateExtractElement(Num, I);
      Value *DenEltI = Builder.CreateExtractElement(Den, I);
      Value *NewElt;

      if (shouldKeepFDivF32(NumEltI, UnsafeDiv)) {
        NewElt = Builder.CreateFDiv(NumEltI, DenEltI);
      } else {
        NewElt = Builder.CreateCall(Decl, { NumEltI, DenEltI });
      }

      NewFDiv = Builder.CreateInsertElement(NewFDiv, NewElt, I);
    }
  } else {
    if (!shouldKeepFDivF32(Num, UnsafeDiv))
      NewFDiv = Builder.CreateCall(Decl, { Num, Den });
  }

  if (NewFDiv) {
    FDiv.replaceAllUsesWith(NewFDiv);
    NewFDiv->takeName(&FDiv);
    FDiv.eraseFromParent();
  }

  return true;
}
Esempio n. 13
0
void Lint::visitShl(BinaryOperator &I) {
  if (ConstantInt *CI =
        dyn_cast<ConstantInt>(I.getOperand(1)->stripPointerCasts()))
    Assert1(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
            "Undefined result: Shift count out of range", &I);
}
Esempio n. 14
0
// Peephole optimize the following instructions:
// %t1 = cast ? to x *
// %t2 = add x * %SP, %t1              ;; Constant must be 2nd operand
//
// Into: %t3 = getelementptr {<...>} * %SP, <element indices>
//       %t2 = cast <eltype> * %t3 to {<...>}*
//
static bool HandleCastToPointer(BasicBlock::iterator BI,
                                const PointerType *DestPTy,
                                const TargetData &TD) {
  CastInst &CI = cast<CastInst>(*BI);
  if (CI.use_empty()) return false;

  // Scan all of the uses, looking for any uses that are not add or sub
  // instructions.  If we have non-adds, do not make this transformation.
  //
  bool HasSubUse = false;  // Keep track of any subtracts...
  for (Value::use_iterator I = CI.use_begin(), E = CI.use_end();
       I != E; ++I)
    if (BinaryOperator *BO = dyn_cast<BinaryOperator>(*I)) {
      if ((BO->getOpcode() != Instruction::Add &&
           BO->getOpcode() != Instruction::Sub) ||
          // Avoid add sbyte* %X, %X cases...
          BO->getOperand(0) == BO->getOperand(1))
        return false;
      else
        HasSubUse |= BO->getOpcode() == Instruction::Sub;
    } else {
      return false;
    }

  std::vector<Value*> Indices;
  Value *Src = CI.getOperand(0);
  const Type *Result = ConvertibleToGEP(DestPTy, Src, Indices, TD, &BI);
  if (Result == 0) return false;  // Not convertible...

  // Cannot handle subtracts if there is more than one index required...
  if (HasSubUse && Indices.size() != 1) return false;

  PRINT_PEEPHOLE2("cast-add-to-gep:in", *Src, CI);

  // If we have a getelementptr capability... transform all of the 
  // add instruction uses into getelementptr's.
  while (!CI.use_empty()) {
    BinaryOperator *I = cast<BinaryOperator>(*CI.use_begin());
    assert((I->getOpcode() == Instruction::Add ||
            I->getOpcode() == Instruction::Sub) && 
           "Use is not a valid add instruction!");
    
    // Get the value added to the cast result pointer...
    Value *OtherPtr = I->getOperand((I->getOperand(0) == &CI) ? 1 : 0);

    Instruction *GEP = new GetElementPtrInst(OtherPtr, Indices, I->getName());
    PRINT_PEEPHOLE1("cast-add-to-gep:i", *I);

    // If the instruction is actually a subtract, we are guaranteed to only have
    // one index (from code above), so we just need to negate the pointer index
    // long value.
    if (I->getOpcode() == Instruction::Sub) {
      Instruction *Neg = BinaryOperator::createNeg(GEP->getOperand(1), 
                                       GEP->getOperand(1)->getName()+".neg", I);
      GEP->setOperand(1, Neg);
    }

    if (GEP->getType() == I->getType()) {
      // Replace the old add instruction with the shiny new GEP inst
      ReplaceInstWithInst(I, GEP);
    } else {
      // If the type produced by the gep instruction differs from the original
      // add instruction type, insert a cast now.
      //

      // Insert the GEP instruction before the old add instruction...
      I->getParent()->getInstList().insert(I, GEP);

      PRINT_PEEPHOLE1("cast-add-to-gep:o", *GEP);
      GEP = new CastInst(GEP, I->getType());

      // Replace the old add instruction with the shiny new GEP inst
      ReplaceInstWithInst(I, GEP);
    }

    PRINT_PEEPHOLE1("cast-add-to-gep:o", *GEP);
  }
  return true;
}
Esempio n. 15
0
/// GetShiftedValue - When CanEvaluateShifted returned true for an expression,
/// this value inserts the new computation that produces the shifted value.
static Value *GetShiftedValue(Value *V, unsigned NumBits, bool isLeftShift,
                              InstCombiner &IC) {
  // We can always evaluate constants shifted.
  if (Constant *C = dyn_cast<Constant>(V)) {
    if (isLeftShift)
      V = IC.Builder->CreateShl(C, NumBits);
    else
      V = IC.Builder->CreateLShr(C, NumBits);
    // If we got a constantexpr back, try to simplify it with TD info.
    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
      V = ConstantFoldConstantExpression(CE, IC.getDataLayout(),
                                         IC.getTargetLibraryInfo());
    return V;
  }

  Instruction *I = cast<Instruction>(V);
  IC.Worklist.Add(I);

  switch (I->getOpcode()) {
  default: llvm_unreachable("Inconsistency with CanEvaluateShifted");
  case Instruction::And:
  case Instruction::Or:
  case Instruction::Xor:
    // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
    I->setOperand(0, GetShiftedValue(I->getOperand(0), NumBits,isLeftShift,IC));
    I->setOperand(1, GetShiftedValue(I->getOperand(1), NumBits,isLeftShift,IC));
    return I;

  case Instruction::Shl: {
    BinaryOperator *BO = cast<BinaryOperator>(I);
    unsigned TypeWidth = BO->getType()->getScalarSizeInBits();

    // We only accept shifts-by-a-constant in CanEvaluateShifted.
    ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));

    // We can always fold shl(c1)+shl(c2) -> shl(c1+c2).
    if (isLeftShift) {
      // If this is oversized composite shift, then unsigned shifts get 0.
      unsigned NewShAmt = NumBits+CI->getZExtValue();
      if (NewShAmt >= TypeWidth)
        return Constant::getNullValue(I->getType());

      BO->setOperand(1, ConstantInt::get(BO->getType(), NewShAmt));
      BO->setHasNoUnsignedWrap(false);
      BO->setHasNoSignedWrap(false);
      return I;
    }

    // We turn shl(c)+lshr(c) -> and(c2) if the input doesn't already have
    // zeros.
    if (CI->getValue() == NumBits) {
      APInt Mask(APInt::getLowBitsSet(TypeWidth, TypeWidth - NumBits));
      V = IC.Builder->CreateAnd(BO->getOperand(0),
                                ConstantInt::get(BO->getContext(), Mask));
      if (Instruction *VI = dyn_cast<Instruction>(V)) {
        VI->moveBefore(BO);
        VI->takeName(BO);
      }
      return V;
    }

    // We turn shl(c1)+shr(c2) -> shl(c3)+and(c4), but only when we know that
    // the and won't be needed.
    assert(CI->getZExtValue() > NumBits);
    BO->setOperand(1, ConstantInt::get(BO->getType(),
                                       CI->getZExtValue() - NumBits));
    BO->setHasNoUnsignedWrap(false);
    BO->setHasNoSignedWrap(false);
    return BO;
  }
  case Instruction::LShr: {
    BinaryOperator *BO = cast<BinaryOperator>(I);
    unsigned TypeWidth = BO->getType()->getScalarSizeInBits();
    // We only accept shifts-by-a-constant in CanEvaluateShifted.
    ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));

    // We can always fold lshr(c1)+lshr(c2) -> lshr(c1+c2).
    if (!isLeftShift) {
      // If this is oversized composite shift, then unsigned shifts get 0.
      unsigned NewShAmt = NumBits+CI->getZExtValue();
      if (NewShAmt >= TypeWidth)
        return Constant::getNullValue(BO->getType());

      BO->setOperand(1, ConstantInt::get(BO->getType(), NewShAmt));
      BO->setIsExact(false);
      return I;
    }

    // We turn lshr(c)+shl(c) -> and(c2) if the input doesn't already have
    // zeros.
    if (CI->getValue() == NumBits) {
      APInt Mask(APInt::getHighBitsSet(TypeWidth, TypeWidth - NumBits));
      V = IC.Builder->CreateAnd(I->getOperand(0),
                                ConstantInt::get(BO->getContext(), Mask));
      if (Instruction *VI = dyn_cast<Instruction>(V)) {
        VI->moveBefore(I);
        VI->takeName(I);
      }
      return V;
    }

    // We turn lshr(c1)+shl(c2) -> lshr(c3)+and(c4), but only when we know that
    // the and won't be needed.
    assert(CI->getZExtValue() > NumBits);
    BO->setOperand(1, ConstantInt::get(BO->getType(),
                                       CI->getZExtValue() - NumBits));
    BO->setIsExact(false);
    return BO;
  }

  case Instruction::Select:
    I->setOperand(1, GetShiftedValue(I->getOperand(1), NumBits,isLeftShift,IC));
    I->setOperand(2, GetShiftedValue(I->getOperand(2), NumBits,isLeftShift,IC));
    return I;
  case Instruction::PHI: {
    // We can change a phi if we can change all operands.  Note that we never
    // get into trouble with cyclic PHIs here because we only consider
    // instructions with a single use.
    PHINode *PN = cast<PHINode>(I);
    for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
      PN->setIncomingValue(i, GetShiftedValue(PN->getIncomingValue(i),
                                              NumBits, isLeftShift, IC));
    return PN;
  }
  }
}
/// HandleFloatingPointIV - If the loop has floating induction variable
/// then insert corresponding integer induction variable if possible.
/// For example,
/// for(double i = 0; i < 10000; ++i)
///   bar(i)
/// is converted into
/// for(int i = 0; i < 10000; ++i)
///   bar((double)i);
///
void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PH) {

  unsigned IncomingEdge = L->contains(PH->getIncomingBlock(0));
  unsigned BackEdge     = IncomingEdge^1;

  // Check incoming value.
  ConstantFP *InitValue = dyn_cast<ConstantFP>(PH->getIncomingValue(IncomingEdge));
  if (!InitValue) return;
  uint64_t newInitValue =
              Type::getInt32Ty(PH->getContext())->getPrimitiveSizeInBits();
  if (!convertToInt(InitValue->getValueAPF(), &newInitValue))
    return;

  // Check IV increment. Reject this PH if increment operation is not
  // an add or increment value can not be represented by an integer.
  BinaryOperator *Incr =
    dyn_cast<BinaryOperator>(PH->getIncomingValue(BackEdge));
  if (!Incr) return;
  if (Incr->getOpcode() != Instruction::FAdd) return;
  ConstantFP *IncrValue = NULL;
  unsigned IncrVIndex = 1;
  if (Incr->getOperand(1) == PH)
    IncrVIndex = 0;
  IncrValue = dyn_cast<ConstantFP>(Incr->getOperand(IncrVIndex));
  if (!IncrValue) return;
  uint64_t newIncrValue =
              Type::getInt32Ty(PH->getContext())->getPrimitiveSizeInBits();
  if (!convertToInt(IncrValue->getValueAPF(), &newIncrValue))
    return;

  // Check Incr uses. One user is PH and the other users is exit condition used
  // by the conditional terminator.
  Value::use_iterator IncrUse = Incr->use_begin();
  Instruction *U1 = cast<Instruction>(IncrUse++);
  if (IncrUse == Incr->use_end()) return;
  Instruction *U2 = cast<Instruction>(IncrUse++);
  if (IncrUse != Incr->use_end()) return;

  // Find exit condition.
  FCmpInst *EC = dyn_cast<FCmpInst>(U1);
  if (!EC)
    EC = dyn_cast<FCmpInst>(U2);
  if (!EC) return;

  if (BranchInst *BI = dyn_cast<BranchInst>(EC->getParent()->getTerminator())) {
    if (!BI->isConditional()) return;
    if (BI->getCondition() != EC) return;
  }

  // Find exit value. If exit value can not be represented as an integer then
  // do not handle this floating point PH.
  ConstantFP *EV = NULL;
  unsigned EVIndex = 1;
  if (EC->getOperand(1) == Incr)
    EVIndex = 0;
  EV = dyn_cast<ConstantFP>(EC->getOperand(EVIndex));
  if (!EV) return;
  uint64_t intEV = Type::getInt32Ty(PH->getContext())->getPrimitiveSizeInBits();
  if (!convertToInt(EV->getValueAPF(), &intEV))
    return;

  // Find new predicate for integer comparison.
  CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
  switch (EC->getPredicate()) {
  case CmpInst::FCMP_OEQ:
  case CmpInst::FCMP_UEQ:
    NewPred = CmpInst::ICMP_EQ;
    break;
  case CmpInst::FCMP_OGT:
  case CmpInst::FCMP_UGT:
    NewPred = CmpInst::ICMP_UGT;
    break;
  case CmpInst::FCMP_OGE:
  case CmpInst::FCMP_UGE:
    NewPred = CmpInst::ICMP_UGE;
    break;
  case CmpInst::FCMP_OLT:
  case CmpInst::FCMP_ULT:
    NewPred = CmpInst::ICMP_ULT;
    break;
  case CmpInst::FCMP_OLE:
  case CmpInst::FCMP_ULE:
    NewPred = CmpInst::ICMP_ULE;
    break;
  default:
    break;
  }
  if (NewPred == CmpInst::BAD_ICMP_PREDICATE) return;

  // Insert new integer induction variable.
  PHINode *NewPHI = PHINode::Create(Type::getInt32Ty(PH->getContext()),
                                    PH->getName()+".int", PH);
  NewPHI->addIncoming(ConstantInt::get(Type::getInt32Ty(PH->getContext()),
                                       newInitValue),
                      PH->getIncomingBlock(IncomingEdge));

  Value *NewAdd = BinaryOperator::CreateAdd(NewPHI,
                           ConstantInt::get(Type::getInt32Ty(PH->getContext()),
                                                             newIncrValue),
                                            Incr->getName()+".int", Incr);
  NewPHI->addIncoming(NewAdd, PH->getIncomingBlock(BackEdge));

  // The back edge is edge 1 of newPHI, whatever it may have been in the
  // original PHI.
  ConstantInt *NewEV = ConstantInt::get(Type::getInt32Ty(PH->getContext()),
                                        intEV);
  Value *LHS = (EVIndex == 1 ? NewPHI->getIncomingValue(1) : NewEV);
  Value *RHS = (EVIndex == 1 ? NewEV : NewPHI->getIncomingValue(1));
  ICmpInst *NewEC = new ICmpInst(EC->getParent()->getTerminator(),
                                 NewPred, LHS, RHS, EC->getName());

  // In the following deletions, PH may become dead and may be deleted.
  // Use a WeakVH to observe whether this happens.
  WeakVH WeakPH = PH;

  // Delete old, floating point, exit comparison instruction.
  NewEC->takeName(EC);
  EC->replaceAllUsesWith(NewEC);
  RecursivelyDeleteTriviallyDeadInstructions(EC);

  // Delete old, floating point, increment instruction.
  Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
  RecursivelyDeleteTriviallyDeadInstructions(Incr);

  // Replace floating induction variable, if it isn't already deleted.
  // Give SIToFPInst preference over UIToFPInst because it is faster on
  // platforms that are widely used.
  if (WeakPH && !PH->use_empty()) {
    if (useSIToFPInst(*InitValue, *EV, newInitValue, intEV)) {
      SIToFPInst *Conv = new SIToFPInst(NewPHI, PH->getType(), "indvar.conv",
                                        PH->getParent()->getFirstNonPHI());
      PH->replaceAllUsesWith(Conv);
    } else {
      UIToFPInst *Conv = new UIToFPInst(NewPHI, PH->getType(), "indvar.conv",
                                        PH->getParent()->getFirstNonPHI());
      PH->replaceAllUsesWith(Conv);
    }
    RecursivelyDeleteTriviallyDeadInstructions(PH);
  }

  // Add a new IVUsers entry for the newly-created integer PHI.
  IU->AddUsersIfInteresting(NewPHI);
}