// 1. If a binary operator has a constant operand, it is moved to the RHS
// 2. Bitwise operators with constant operands are always grouped so that
// shifts are performed first, then or's, then and's, then xor's.
-// 3. SetCC instructions are converted from <,>,<=,>= to ==,!= if possible
-// 4. All SetCC instructions on boolean values are replaced with logical ops
+// 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
+// 4. All cmp instructions on boolean values are replaced with logical ops
// 5. add X, X is represented as (X*2) => (X << 1)
// 6. Multiplies with a power-of-two constant argument are transformed into
// shifts.
#include "llvm/Support/InstVisitor.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/PatternMatch.h"
+#include "llvm/Support/Compiler.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
#include <algorithm>
-#include <iostream>
using namespace llvm;
using namespace llvm::PatternMatch;
+STATISTIC(NumCombined , "Number of insts combined");
+STATISTIC(NumConstProp, "Number of constant folds");
+STATISTIC(NumDeadInst , "Number of dead inst eliminated");
+STATISTIC(NumDeadStore, "Number of dead stores eliminated");
+STATISTIC(NumSunkInst , "Number of instructions sunk");
+
namespace {
- Statistic<> NumCombined ("instcombine", "Number of insts combined");
- Statistic<> NumConstProp("instcombine", "Number of constant folds");
- Statistic<> NumDeadInst ("instcombine", "Number of dead inst eliminated");
- Statistic<> NumDeadStore("instcombine", "Number of dead stores eliminated");
- Statistic<> NumSunkInst ("instcombine", "Number of instructions sunk");
-
- class InstCombiner : public FunctionPass,
- public InstVisitor<InstCombiner, Instruction*> {
+ class VISIBILITY_HIDDEN InstCombiner
+ : public FunctionPass,
+ public InstVisitor<InstCombiner, Instruction*> {
// Worklist of all of the instructions that need to be simplified.
std::vector<Instruction*> WorkList;
TargetData *TD;
if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
WorkList.push_back(Op);
}
+
+ /// AddSoonDeadInstToWorklist - The specified instruction is about to become
+ /// dead. Add all of its operands to the worklist, turning them into
+ /// undef's to reduce the number of uses of those instructions.
+ ///
+ /// Return the specified operand before it is turned into an undef.
+ ///
+ Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
+ Value *R = I.getOperand(op);
+
+ for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
+ if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
+ WorkList.push_back(Op);
+ // Set the operand to undef to drop the use.
+ I.setOperand(i, UndefValue::get(Op->getType()));
+ }
+
+ return R;
+ }
// removeFromWorkList - remove all instances of I from the worklist.
void removeFromWorkList(Instruction *I);
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<TargetData>();
+ AU.addPreservedID(LCSSAID);
AU.setPreservesCFG();
}
Instruction *visitAdd(BinaryOperator &I);
Instruction *visitSub(BinaryOperator &I);
Instruction *visitMul(BinaryOperator &I);
- Instruction *visitDiv(BinaryOperator &I);
- Instruction *visitRem(BinaryOperator &I);
+ Instruction *visitURem(BinaryOperator &I);
+ Instruction *visitSRem(BinaryOperator &I);
+ Instruction *visitFRem(BinaryOperator &I);
+ Instruction *commonRemTransforms(BinaryOperator &I);
+ Instruction *commonIRemTransforms(BinaryOperator &I);
+ Instruction *commonDivTransforms(BinaryOperator &I);
+ Instruction *commonIDivTransforms(BinaryOperator &I);
+ Instruction *visitUDiv(BinaryOperator &I);
+ Instruction *visitSDiv(BinaryOperator &I);
+ Instruction *visitFDiv(BinaryOperator &I);
Instruction *visitAnd(BinaryOperator &I);
Instruction *visitOr (BinaryOperator &I);
Instruction *visitXor(BinaryOperator &I);
- Instruction *visitSetCondInst(SetCondInst &I);
- Instruction *visitSetCondInstWithCastAndCast(SetCondInst &SCI);
+ Instruction *visitFCmpInst(FCmpInst &I);
+ Instruction *visitICmpInst(ICmpInst &I);
+ Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
- Instruction *FoldGEPSetCC(User *GEPLHS, Value *RHS,
- Instruction::BinaryOps Cond, Instruction &I);
+ Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
+ ICmpInst::Predicate Cond, Instruction &I);
Instruction *visitShiftInst(ShiftInst &I);
- Instruction *FoldShiftByConstant(Value *Op0, ConstantUInt *Op1,
+ Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
ShiftInst &I);
- Instruction *visitCastInst(CastInst &CI);
+ Instruction *commonCastTransforms(CastInst &CI);
+ Instruction *commonIntCastTransforms(CastInst &CI);
+ Instruction *visitTrunc(CastInst &CI);
+ Instruction *visitZExt(CastInst &CI);
+ Instruction *visitSExt(CastInst &CI);
+ Instruction *visitFPTrunc(CastInst &CI);
+ Instruction *visitFPExt(CastInst &CI);
+ Instruction *visitFPToUI(CastInst &CI);
+ Instruction *visitFPToSI(CastInst &CI);
+ Instruction *visitUIToFP(CastInst &CI);
+ Instruction *visitSIToFP(CastInst &CI);
+ Instruction *visitPtrToInt(CastInst &CI);
+ Instruction *visitIntToPtr(CastInst &CI);
+ Instruction *visitBitCast(CastInst &CI);
Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
Instruction *FI);
Instruction *visitSelectInst(SelectInst &CI);
/// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
/// This also adds the cast to the worklist. Finally, this returns the
/// cast.
- Value *InsertCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
+ Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
+ Instruction &Pos) {
if (V->getType() == Ty) return V;
if (Constant *CV = dyn_cast<Constant>(V))
- return ConstantExpr::getCast(CV, Ty);
+ return ConstantExpr::getCast(opc, CV, Ty);
- Instruction *C = new CastInst(V, Ty, V->getName(), &Pos);
+ Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
WorkList.push_back(C);
return C;
}
/// InsertBefore instruction. This is specialized a bit to avoid inserting
/// casts that are known to not do anything...
///
- Value *InsertOperandCastBefore(Value *V, const Type *DestTy,
+ Value *InsertOperandCastBefore(Instruction::CastOps opcode,
+ Value *V, const Type *DestTy,
Instruction *InsertBefore);
- // SimplifyCommutative - This performs a few simplifications for commutative
- // operators.
+ /// SimplifyCommutative - This performs a few simplifications for
+ /// commutative operators.
bool SimplifyCommutative(BinaryOperator &I);
+ /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
+ /// most-complex to least-complex order.
+ bool SimplifyCompare(CmpInst &I);
+
bool SimplifyDemandedBits(Value *V, uint64_t Mask,
uint64_t &KnownZero, uint64_t &KnownOne,
unsigned Depth = 0);
+ Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
+ uint64_t &UndefElts, unsigned Depth = 0);
+
// FoldOpIntoPhi - Given a binary operator or cast instruction which has a
// PHI node as operand #0, see if we can fold the instruction into the PHI
// (which is only possible if all operands to the PHI are constants).
// operator and they all are only used by the PHI, PHI together their
// inputs, and do the operation once, to the result of the PHI.
Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
-
+ Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
+
+
Instruction *OptAndOp(Instruction *Op, ConstantIntegral *OpRHS,
ConstantIntegral *AndRHS, BinaryOperator &TheAnd);
Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantIntegral *Mask,
bool isSub, Instruction &I);
Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
- bool Inside, Instruction &IB);
+ bool isSigned, bool Inside, Instruction &IB);
Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
-
- Value *EvaluateInDifferentType(Value *V, const Type *Ty);
+ Instruction *MatchBSwap(BinaryOperator &I);
+
+ Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
};
- RegisterOpt<InstCombiner> X("instcombine", "Combine redundant instructions");
+ RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
}
// getComplexity: Assign a complexity or rank value to LLVM Values...
}
}
-/// isCast - If the specified operand is a CastInst or a constant expr cast,
-/// return the operand value, otherwise return null.
-static Value *isCast(Value *V) {
- if (CastInst *I = dyn_cast<CastInst>(V))
+/// getBitCastOperand - If the specified operand is a CastInst or a constant
+/// expression bitcast, return the operand value, otherwise return null.
+static Value *getBitCastOperand(Value *V) {
+ if (BitCastInst *I = dyn_cast<BitCastInst>(V))
return I->getOperand(0);
else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
- if (CE->getOpcode() == Instruction::Cast)
+ if (CE->getOpcode() == Instruction::BitCast)
return CE->getOperand(0);
return 0;
}
-enum CastType {
- Noop = 0,
- Truncate = 1,
- Signext = 2,
- Zeroext = 3
-};
-
-/// getCastType - In the future, we will split the cast instruction into these
-/// various types. Until then, we have to do the analysis here.
-static CastType getCastType(const Type *Src, const Type *Dest) {
- assert(Src->isIntegral() && Dest->isIntegral() &&
- "Only works on integral types!");
- unsigned SrcSize = Src->getPrimitiveSizeInBits();
- unsigned DestSize = Dest->getPrimitiveSizeInBits();
+/// This function is a wrapper around CastInst::isEliminableCastPair. It
+/// simply extracts arguments and returns what that function returns.
+/// @Determine if it is valid to eliminate a Convert pair
+static Instruction::CastOps
+isEliminableCastPair(
+ const CastInst *CI, ///< The first cast instruction
+ unsigned opcode, ///< The opcode of the second cast instruction
+ const Type *DstTy, ///< The target type for the second cast instruction
+ TargetData *TD ///< The target data for pointer size
+) {
- if (SrcSize == DestSize) return Noop;
- if (SrcSize > DestSize) return Truncate;
- if (Src->isSigned()) return Signext;
- return Zeroext;
-}
+ const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
+ const Type *MidTy = CI->getType(); // B from above
+ // Get the opcodes of the two Cast instructions
+ Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
+ Instruction::CastOps secondOp = Instruction::CastOps(opcode);
-// isEliminableCastOfCast - Return true if it is valid to eliminate the CI
-// instruction.
-//
-static bool isEliminableCastOfCast(const Type *SrcTy, const Type *MidTy,
- const Type *DstTy, TargetData *TD) {
-
- // It is legal to eliminate the instruction if casting A->B->A if the sizes
- // are identical and the bits don't get reinterpreted (for example
- // int->float->int would not be allowed).
- if (SrcTy == DstTy && SrcTy->isLosslesslyConvertibleTo(MidTy))
- return true;
-
- // If we are casting between pointer and integer types, treat pointers as
- // integers of the appropriate size for the code below.
- if (isa<PointerType>(SrcTy)) SrcTy = TD->getIntPtrType();
- if (isa<PointerType>(MidTy)) MidTy = TD->getIntPtrType();
- if (isa<PointerType>(DstTy)) DstTy = TD->getIntPtrType();
-
- // Allow free casting and conversion of sizes as long as the sign doesn't
- // change...
- if (SrcTy->isIntegral() && MidTy->isIntegral() && DstTy->isIntegral()) {
- CastType FirstCast = getCastType(SrcTy, MidTy);
- CastType SecondCast = getCastType(MidTy, DstTy);
-
- // Capture the effect of these two casts. If the result is a legal cast,
- // the CastType is stored here, otherwise a special code is used.
- static const unsigned CastResult[] = {
- // First cast is noop
- 0, 1, 2, 3,
- // First cast is a truncate
- 1, 1, 4, 4, // trunc->extend is not safe to eliminate
- // First cast is a sign ext
- 2, 5, 2, 4, // signext->zeroext never ok
- // First cast is a zero ext
- 3, 5, 3, 3,
- };
-
- unsigned Result = CastResult[FirstCast*4+SecondCast];
- switch (Result) {
- default: assert(0 && "Illegal table value!");
- case 0:
- case 1:
- case 2:
- case 3:
- // FIXME: in the future, when LLVM has explicit sign/zeroextends and
- // truncates, we could eliminate more casts.
- return (unsigned)getCastType(SrcTy, DstTy) == Result;
- case 4:
- return false; // Not possible to eliminate this here.
- case 5:
- // Sign or zero extend followed by truncate is always ok if the result
- // is a truncate or noop.
- CastType ResultCast = getCastType(SrcTy, DstTy);
- if (ResultCast == Noop || ResultCast == Truncate)
- return true;
- // Otherwise we are still growing the value, we are only safe if the
- // result will match the sign/zeroextendness of the result.
- return ResultCast == FirstCast;
- }
- }
-
- // If this is a cast from 'float -> double -> integer', cast from
- // 'float -> integer' directly, as the value isn't changed by the
- // float->double conversion.
- if (SrcTy->isFloatingPoint() && MidTy->isFloatingPoint() &&
- DstTy->isIntegral() &&
- SrcTy->getPrimitiveSize() < MidTy->getPrimitiveSize())
- return true;
-
- // Packed type conversions don't modify bits.
- if (isa<PackedType>(SrcTy) && isa<PackedType>(MidTy) &&isa<PackedType>(DstTy))
- return true;
-
- return false;
+ return Instruction::CastOps(
+ CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
+ DstTy, TD->getIntPtrType()));
}
/// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
/// in any code being generated. It does not require codegen if V is simple
/// enough or if the cast can be folded into other casts.
-static bool ValueRequiresCast(const Value *V, const Type *Ty, TargetData *TD) {
+static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
+ const Type *Ty, TargetData *TD) {
if (V->getType() == Ty || isa<Constant>(V)) return false;
// If this is a noop cast, it isn't real codegen.
- if (V->getType()->isLosslesslyConvertibleTo(Ty))
+ if (V->getType()->canLosslesslyBitCastTo(Ty))
return false;
// If this is another cast that can be eliminated, it isn't codegen either.
if (const CastInst *CI = dyn_cast<CastInst>(V))
- if (isEliminableCastOfCast(CI->getOperand(0)->getType(), CI->getType(), Ty,
- TD))
+ if (isEliminableCastPair(CI, opcode, Ty, TD))
return false;
return true;
}
/// InsertBefore instruction. This is specialized a bit to avoid inserting
/// casts that are known to not do anything...
///
-Value *InstCombiner::InsertOperandCastBefore(Value *V, const Type *DestTy,
+Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
+ Value *V, const Type *DestTy,
Instruction *InsertBefore) {
if (V->getType() == DestTy) return V;
if (Constant *C = dyn_cast<Constant>(V))
- return ConstantExpr::getCast(C, DestTy);
+ return ConstantExpr::getCast(opcode, C, DestTy);
- CastInst *CI = new CastInst(V, DestTy, V->getName());
- InsertNewInstBefore(CI, *InsertBefore);
- return CI;
+ return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
}
// SimplifyCommutative - This performs a few simplifications for commutative
return Changed;
}
+/// SimplifyCompare - For a CmpInst this function just orders the operands
+/// so that theyare listed from right (least complex) to left (most complex).
+/// This puts constants before unary operators before binary operators.
+bool InstCombiner::SimplifyCompare(CmpInst &I) {
+ if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
+ return false;
+ I.swapOperands();
+ // Compare instructions are not associative so there's nothing else we can do.
+ return true;
+}
+
// dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
// if the LHS is a constant zero (which is the 'negate' form).
//
/// GetConstantInType - Return a ConstantInt with the specified type and value.
///
static ConstantIntegral *GetConstantInType(const Type *Ty, uint64_t Val) {
- if (Ty->isUnsigned())
- return ConstantUInt::get(Ty, Val);
+ if (Ty->isUnsigned())
+ return ConstantInt::get(Ty, Val);
else if (Ty->getTypeID() == Type::BoolTyID)
return ConstantBool::get(Val);
int64_t SVal = Val;
SVal <<= 64-Ty->getPrimitiveSizeInBits();
SVal >>= 64-Ty->getPrimitiveSizeInBits();
- return ConstantSInt::get(Ty, SVal);
+ return ConstantInt::get(Ty, SVal);
}
KnownOne &= KnownOne2;
KnownZero &= KnownZero2;
return;
- case Instruction::Cast: {
+ case Instruction::FPTrunc:
+ case Instruction::FPExt:
+ case Instruction::FPToUI:
+ case Instruction::FPToSI:
+ case Instruction::SIToFP:
+ case Instruction::PtrToInt:
+ case Instruction::UIToFP:
+ case Instruction::IntToPtr:
+ return; // Can't work with floating point or pointers
+ case Instruction::Trunc:
+ // All these have integer operands
+ ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
+ return;
+ case Instruction::BitCast: {
const Type *SrcTy = I->getOperand(0)->getType();
- if (!SrcTy->isIntegral()) return;
-
- // If this is an integer truncate or noop, just look in the input.
- if (SrcTy->getPrimitiveSizeInBits() >=
- I->getType()->getPrimitiveSizeInBits()) {
+ if (SrcTy->isIntegral()) {
ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
return;
}
-
- // Sign or Zero extension. Compute the bits in the result that are not
- // present in the input.
+ break;
+ }
+ case Instruction::ZExt: {
+ // Compute the bits in the result that are not present in the input.
+ const Type *SrcTy = I->getOperand(0)->getType();
uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
- // Handle zero extension.
- if (!SrcTy->isSigned()) {
- Mask &= SrcTy->getIntegralTypeMask();
- ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
- assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
- // The top bits are known to be zero.
- KnownZero |= NewBits;
- } else {
- // Sign extension.
- Mask &= SrcTy->getIntegralTypeMask();
- ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
- assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ Mask &= SrcTy->getIntegralTypeMask();
+ ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ // The top bits are known to be zero.
+ KnownZero |= NewBits;
+ return;
+ }
+ case Instruction::SExt: {
+ // Compute the bits in the result that are not present in the input.
+ const Type *SrcTy = I->getOperand(0)->getType();
+ uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
+ uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
+
+ Mask &= SrcTy->getIntegralTypeMask();
+ ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
- // If the sign bit of the input is known set or clear, then we know the
- // top bits of the result.
- uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
- if (KnownZero & InSignBit) { // Input sign bit known zero
- KnownZero |= NewBits;
- KnownOne &= ~NewBits;
- } else if (KnownOne & InSignBit) { // Input sign bit known set
- KnownOne |= NewBits;
- KnownZero &= ~NewBits;
- } else { // Input sign bit unknown
- KnownZero &= ~NewBits;
- KnownOne &= ~NewBits;
- }
+ // If the sign bit of the input is known set or clear, then we know the
+ // top bits of the result.
+ uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
+ if (KnownZero & InSignBit) { // Input sign bit known zero
+ KnownZero |= NewBits;
+ KnownOne &= ~NewBits;
+ } else if (KnownOne & InSignBit) { // Input sign bit known set
+ KnownOne |= NewBits;
+ KnownZero &= ~NewBits;
+ } else { // Input sign bit unknown
+ KnownZero &= ~NewBits;
+ KnownOne &= ~NewBits;
}
return;
}
case Instruction::Shl:
// (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
- if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
- Mask >>= SA->getValue();
+ if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ uint64_t ShiftAmt = SA->getZExtValue();
+ Mask >>= ShiftAmt;
ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
- KnownZero <<= SA->getValue();
- KnownOne <<= SA->getValue();
- KnownZero |= (1ULL << SA->getValue())-1; // low bits known zero.
+ KnownZero <<= ShiftAmt;
+ KnownOne <<= ShiftAmt;
+ KnownZero |= (1ULL << ShiftAmt)-1; // low bits known zero.
return;
}
break;
- case Instruction::Shr:
+ case Instruction::LShr:
// (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
- if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
+ if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
// Compute the new bits that are at the top now.
- uint64_t HighBits = (1ULL << SA->getValue())-1;
- HighBits <<= I->getType()->getPrimitiveSizeInBits()-SA->getValue();
+ uint64_t ShiftAmt = SA->getZExtValue();
+ uint64_t HighBits = (1ULL << ShiftAmt)-1;
+ HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
- if (I->getType()->isUnsigned()) { // Unsigned shift right.
- Mask <<= SA->getValue();
- ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
- assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
- KnownZero >>= SA->getValue();
- KnownOne >>= SA->getValue();
- KnownZero |= HighBits; // high bits known zero.
- } else {
- Mask <<= SA->getValue();
- ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
- assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
- KnownZero >>= SA->getValue();
- KnownOne >>= SA->getValue();
+ // Unsigned shift right.
+ Mask <<= ShiftAmt;
+ ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
+ assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
+ KnownZero >>= ShiftAmt;
+ KnownOne >>= ShiftAmt;
+ KnownZero |= HighBits; // high bits known zero.
+ return;
+ }
+ break;
+ case Instruction::AShr:
+ // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
+ if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ // Compute the new bits that are at the top now.
+ uint64_t ShiftAmt = SA->getZExtValue();
+ uint64_t HighBits = (1ULL << ShiftAmt)-1;
+ HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
+
+ // Signed shift right.
+ Mask <<= ShiftAmt;
+ ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
+ assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
+ KnownZero >>= ShiftAmt;
+ KnownOne >>= ShiftAmt;
- // Handle the sign bits.
- uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
- SignBit >>= SA->getValue(); // Adjust to where it is now in the mask.
+ // Handle the sign bits.
+ uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
+ SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
- if (KnownZero & SignBit) { // New bits are known zero.
- KnownZero |= HighBits;
- } else if (KnownOne & SignBit) { // New bits are known one.
- KnownOne |= HighBits;
- }
+ if (KnownZero & SignBit) { // New bits are known zero.
+ KnownZero |= HighBits;
+ } else if (KnownOne & SignBit) { // New bits are known one.
+ KnownOne |= HighBits;
}
return;
}
DemandedMask &= V->getType()->getIntegralTypeMask();
- uint64_t KnownZero2, KnownOne2;
+ uint64_t KnownZero2 = 0, KnownOne2 = 0;
switch (I->getOpcode()) {
default: break;
case Instruction::And:
return true;
assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
- // If all of the demanded bits are known one on one side, return the other.
+ // If all of the demanded bits are known 1 on one side, return the other.
// These bits cannot contribute to the result of the 'and'.
if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
return UpdateValueUsesWith(I, I->getOperand(0));
// Output known-1 are known to be set if set in only one of the LHS, RHS.
uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
- // If all of the unknown bits are known to be zero on one side or the other
- // (but not both) turn this into an *inclusive* or.
+ // If all of the demanded bits are known to be zero on one side or the
+ // other, turn this into an *inclusive* or.
// e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
- if (uint64_t UnknownBits = DemandedMask & ~(KnownZeroOut|KnownOneOut)) {
- if ((UnknownBits & (KnownZero|KnownZero2)) == UnknownBits) {
- Instruction *Or =
- BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
- I->getName());
- InsertNewInstBefore(Or, *I);
- return UpdateValueUsesWith(I, Or);
- }
+ if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0) {
+ Instruction *Or =
+ BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
+ I->getName());
+ InsertNewInstBefore(Or, *I);
+ return UpdateValueUsesWith(I, Or);
}
// If all of the demanded bits on one side are known, and all of the set
KnownOne &= KnownOne2;
KnownZero &= KnownZero2;
break;
- case Instruction::Cast: {
+ case Instruction::Trunc:
+ if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
+ KnownZero, KnownOne, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ break;
+ case Instruction::BitCast:
+ if (!I->getOperand(0)->getType()->isIntegral())
+ return false;
+
+ if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
+ KnownZero, KnownOne, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ break;
+ case Instruction::ZExt: {
+ // Compute the bits in the result that are not present in the input.
const Type *SrcTy = I->getOperand(0)->getType();
- if (!SrcTy->isIntegral()) return false;
-
- // If this is an integer truncate or noop, just look in the input.
- if (SrcTy->getPrimitiveSizeInBits() >=
- I->getType()->getPrimitiveSizeInBits()) {
- if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
- KnownZero, KnownOne, Depth+1))
- return true;
- assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
- break;
- }
+ uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
+ uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
- // Sign or Zero extension. Compute the bits in the result that are not
- // present in the input.
+ DemandedMask &= SrcTy->getIntegralTypeMask();
+ if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
+ KnownZero, KnownOne, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ // The top bits are known to be zero.
+ KnownZero |= NewBits;
+ break;
+ }
+ case Instruction::SExt: {
+ // Compute the bits in the result that are not present in the input.
+ const Type *SrcTy = I->getOperand(0)->getType();
uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
- // Handle zero extension.
- if (!SrcTy->isSigned()) {
- DemandedMask &= SrcTy->getIntegralTypeMask();
- if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
- KnownZero, KnownOne, Depth+1))
- return true;
- assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
- // The top bits are known to be zero.
- KnownZero |= NewBits;
- } else {
- // Sign extension.
- uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
- int64_t InputDemandedBits = DemandedMask & SrcTy->getIntegralTypeMask();
-
- // If any of the sign extended bits are demanded, we know that the sign
- // bit is demanded.
- if (NewBits & DemandedMask)
- InputDemandedBits |= InSignBit;
+ // Get the sign bit for the source type
+ uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
+ int64_t InputDemandedBits = DemandedMask & SrcTy->getIntegralTypeMask();
+
+ // If any of the sign extended bits are demanded, we know that the sign
+ // bit is demanded.
+ if (NewBits & DemandedMask)
+ InputDemandedBits |= InSignBit;
- if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
- KnownZero, KnownOne, Depth+1))
+ if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
+ KnownZero, KnownOne, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+
+ // If the sign bit of the input is known set or clear, then we know the
+ // top bits of the result.
+
+ // If the input sign bit is known zero, or if the NewBits are not demanded
+ // convert this into a zero extension.
+ if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
+ // Convert to ZExt cast
+ CastInst *NewCast = CastInst::create(
+ Instruction::ZExt, I->getOperand(0), I->getType(), I->getName(), I);
+ return UpdateValueUsesWith(I, NewCast);
+ } else if (KnownOne & InSignBit) { // Input sign bit known set
+ KnownOne |= NewBits;
+ KnownZero &= ~NewBits;
+ } else { // Input sign bit unknown
+ KnownZero &= ~NewBits;
+ KnownOne &= ~NewBits;
+ }
+ break;
+ }
+ case Instruction::Add:
+ // If there is a constant on the RHS, there are a variety of xformations
+ // we can do.
+ if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ // If null, this should be simplified elsewhere. Some of the xforms here
+ // won't work if the RHS is zero.
+ if (RHS->isNullValue())
+ break;
+
+ // Figure out what the input bits are. If the top bits of the and result
+ // are not demanded, then the add doesn't demand them from its input
+ // either.
+
+ // Shift the demanded mask up so that it's at the top of the uint64_t.
+ unsigned BitWidth = I->getType()->getPrimitiveSizeInBits();
+ unsigned NLZ = CountLeadingZeros_64(DemandedMask << (64-BitWidth));
+
+ // If the top bit of the output is demanded, demand everything from the
+ // input. Otherwise, we demand all the input bits except NLZ top bits.
+ uint64_t InDemandedBits = ~0ULL >> 64-BitWidth+NLZ;
+
+ // Find information about known zero/one bits in the input.
+ if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
+ KnownZero2, KnownOne2, Depth+1))
return true;
- assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+
+ // If the RHS of the add has bits set that can't affect the input, reduce
+ // the constant.
+ if (ShrinkDemandedConstant(I, 1, InDemandedBits))
+ return UpdateValueUsesWith(I, I);
- // If the sign bit of the input is known set or clear, then we know the
- // top bits of the result.
-
- // If the input sign bit is known zero, or if the NewBits are not demanded
- // convert this into a zero extension.
- if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
- // Convert to unsigned first.
- Instruction *NewVal;
- NewVal = new CastInst(I->getOperand(0), SrcTy->getUnsignedVersion(),
- I->getOperand(0)->getName());
- InsertNewInstBefore(NewVal, *I);
- // Then cast that to the destination type.
- NewVal = new CastInst(NewVal, I->getType(), I->getName());
- InsertNewInstBefore(NewVal, *I);
- return UpdateValueUsesWith(I, NewVal);
- } else if (KnownOne & InSignBit) { // Input sign bit known set
- KnownOne |= NewBits;
- KnownZero &= ~NewBits;
- } else { // Input sign bit unknown
- KnownZero &= ~NewBits;
- KnownOne &= ~NewBits;
+ // Avoid excess work.
+ if (KnownZero2 == 0 && KnownOne2 == 0)
+ break;
+
+ // Turn it into OR if input bits are zero.
+ if ((KnownZero2 & RHS->getZExtValue()) == RHS->getZExtValue()) {
+ Instruction *Or =
+ BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
+ I->getName());
+ InsertNewInstBefore(Or, *I);
+ return UpdateValueUsesWith(I, Or);
+ }
+
+ // We can say something about the output known-zero and known-one bits,
+ // depending on potential carries from the input constant and the
+ // unknowns. For example if the LHS is known to have at most the 0x0F0F0
+ // bits set and the RHS constant is 0x01001, then we know we have a known
+ // one mask of 0x00001 and a known zero mask of 0xE0F0E.
+
+ // To compute this, we first compute the potential carry bits. These are
+ // the bits which may be modified. I'm not aware of a better way to do
+ // this scan.
+ uint64_t RHSVal = RHS->getZExtValue();
+
+ bool CarryIn = false;
+ uint64_t CarryBits = 0;
+ uint64_t CurBit = 1;
+ for (unsigned i = 0; i != BitWidth; ++i, CurBit <<= 1) {
+ // Record the current carry in.
+ if (CarryIn) CarryBits |= CurBit;
+
+ bool CarryOut;
+
+ // This bit has a carry out unless it is "zero + zero" or
+ // "zero + anything" with no carry in.
+ if ((KnownZero2 & CurBit) && ((RHSVal & CurBit) == 0)) {
+ CarryOut = false; // 0 + 0 has no carry out, even with carry in.
+ } else if (!CarryIn &&
+ ((KnownZero2 & CurBit) || ((RHSVal & CurBit) == 0))) {
+ CarryOut = false; // 0 + anything has no carry out if no carry in.
+ } else {
+ // Otherwise, we have to assume we have a carry out.
+ CarryOut = true;
+ }
+
+ // This stage's carry out becomes the next stage's carry-in.
+ CarryIn = CarryOut;
}
+
+ // Now that we know which bits have carries, compute the known-1/0 sets.
+
+ // Bits are known one if they are known zero in one operand and one in the
+ // other, and there is no input carry.
+ KnownOne = ((KnownZero2 & RHSVal) | (KnownOne2 & ~RHSVal)) & ~CarryBits;
+
+ // Bits are known zero if they are known zero in both operands and there
+ // is no input carry.
+ KnownZero = KnownZero2 & ~RHSVal & ~CarryBits;
}
break;
- }
case Instruction::Shl:
- if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
- if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> SA->getValue(),
+ if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ uint64_t ShiftAmt = SA->getZExtValue();
+ if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> ShiftAmt,
KnownZero, KnownOne, Depth+1))
return true;
assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
- KnownZero <<= SA->getValue();
- KnownOne <<= SA->getValue();
- KnownZero |= (1ULL << SA->getValue())-1; // low bits known zero.
+ KnownZero <<= ShiftAmt;
+ KnownOne <<= ShiftAmt;
+ KnownZero |= (1ULL << ShiftAmt) - 1; // low bits known zero.
}
break;
- case Instruction::Shr:
- if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
- unsigned ShAmt = SA->getValue();
+ case Instruction::LShr:
+ // For a logical shift right
+ if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ unsigned ShiftAmt = SA->getZExtValue();
// Compute the new bits that are at the top now.
- uint64_t HighBits = (1ULL << ShAmt)-1;
- HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShAmt;
+ uint64_t HighBits = (1ULL << ShiftAmt)-1;
+ HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
uint64_t TypeMask = I->getType()->getIntegralTypeMask();
- if (I->getType()->isUnsigned()) { // Unsigned shift right.
- if (SimplifyDemandedBits(I->getOperand(0),
- (DemandedMask << ShAmt) & TypeMask,
- KnownZero, KnownOne, Depth+1))
- return true;
- assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
- KnownZero &= TypeMask;
- KnownOne &= TypeMask;
- KnownZero >>= ShAmt;
- KnownOne >>= ShAmt;
- KnownZero |= HighBits; // high bits known zero.
- } else { // Signed shift right.
- if (SimplifyDemandedBits(I->getOperand(0),
- (DemandedMask << ShAmt) & TypeMask,
- KnownZero, KnownOne, Depth+1))
- return true;
- assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
- KnownZero &= TypeMask;
- KnownOne &= TypeMask;
- KnownZero >>= SA->getValue();
- KnownOne >>= SA->getValue();
+ // Unsigned shift right.
+ if (SimplifyDemandedBits(I->getOperand(0),
+ (DemandedMask << ShiftAmt) & TypeMask,
+ KnownZero, KnownOne, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ KnownZero &= TypeMask;
+ KnownOne &= TypeMask;
+ KnownZero >>= ShiftAmt;
+ KnownOne >>= ShiftAmt;
+ KnownZero |= HighBits; // high bits known zero.
+ }
+ break;
+ case Instruction::AShr:
+ // If this is an arithmetic shift right and only the low-bit is set, we can
+ // always convert this into a logical shr, even if the shift amount is
+ // variable. The low bit of the shift cannot be an input sign bit unless
+ // the shift amount is >= the size of the datatype, which is undefined.
+ if (DemandedMask == 1) {
+ // Perform the logical shift right.
+ Value *NewVal = new ShiftInst(Instruction::LShr, I->getOperand(0),
+ I->getOperand(1), I->getName());
+ InsertNewInstBefore(cast<Instruction>(NewVal), *I);
+ return UpdateValueUsesWith(I, NewVal);
+ }
+
+ if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ unsigned ShiftAmt = SA->getZExtValue();
+
+ // Compute the new bits that are at the top now.
+ uint64_t HighBits = (1ULL << ShiftAmt)-1;
+ HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
+ uint64_t TypeMask = I->getType()->getIntegralTypeMask();
+ // Signed shift right.
+ if (SimplifyDemandedBits(I->getOperand(0),
+ (DemandedMask << ShiftAmt) & TypeMask,
+ KnownZero, KnownOne, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ KnownZero &= TypeMask;
+ KnownOne &= TypeMask;
+ KnownZero >>= ShiftAmt;
+ KnownOne >>= ShiftAmt;
- // Handle the sign bits.
- uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
- SignBit >>= SA->getValue(); // Adjust to where it is now in the mask.
+ // Handle the sign bits.
+ uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
+ SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
- // If the input sign bit is known to be zero, or if none of the top bits
- // are demanded, turn this into an unsigned shift right.
- if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
- // Convert the input to unsigned.
- Instruction *NewVal;
- NewVal = new CastInst(I->getOperand(0),
- I->getType()->getUnsignedVersion(),
- I->getOperand(0)->getName());
- InsertNewInstBefore(NewVal, *I);
- // Perform the unsigned shift right.
- NewVal = new ShiftInst(Instruction::Shr, NewVal, SA, I->getName());
- InsertNewInstBefore(NewVal, *I);
- // Then cast that to the destination type.
- NewVal = new CastInst(NewVal, I->getType(), I->getName());
- InsertNewInstBefore(NewVal, *I);
- return UpdateValueUsesWith(I, NewVal);
- } else if (KnownOne & SignBit) { // New bits are known one.
- KnownOne |= HighBits;
- }
+ // If the input sign bit is known to be zero, or if none of the top bits
+ // are demanded, turn this into an unsigned shift right.
+ if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
+ // Perform the logical shift right.
+ Value *NewVal = new ShiftInst(Instruction::LShr, I->getOperand(0),
+ SA, I->getName());
+ InsertNewInstBefore(cast<Instruction>(NewVal), *I);
+ return UpdateValueUsesWith(I, NewVal);
+ } else if (KnownOne & SignBit) { // New bits are known one.
+ KnownOne |= HighBits;
}
}
break;
return false;
}
-// isTrueWhenEqual - Return true if the specified setcondinst instruction is
-// true when both operands are equal...
-//
-static bool isTrueWhenEqual(Instruction &I) {
- return I.getOpcode() == Instruction::SetEQ ||
- I.getOpcode() == Instruction::SetGE ||
- I.getOpcode() == Instruction::SetLE;
+
+/// SimplifyDemandedVectorElts - The specified value producecs a vector with
+/// 64 or fewer elements. DemandedElts contains the set of elements that are
+/// actually used by the caller. This method analyzes which elements of the
+/// operand are undef and returns that information in UndefElts.
+///
+/// If the information about demanded elements can be used to simplify the
+/// operation, the operation is simplified, then the resultant value is
+/// returned. This returns null if no change was made.
+Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
+ uint64_t &UndefElts,
+ unsigned Depth) {
+ unsigned VWidth = cast<PackedType>(V->getType())->getNumElements();
+ assert(VWidth <= 64 && "Vector too wide to analyze!");
+ uint64_t EltMask = ~0ULL >> (64-VWidth);
+ assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
+ "Invalid DemandedElts!");
+
+ if (isa<UndefValue>(V)) {
+ // If the entire vector is undefined, just return this info.
+ UndefElts = EltMask;
+ return 0;
+ } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
+ UndefElts = EltMask;
+ return UndefValue::get(V->getType());
+ }
+
+ UndefElts = 0;
+ if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V)) {
+ const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
+ Constant *Undef = UndefValue::get(EltTy);
+
+ std::vector<Constant*> Elts;
+ for (unsigned i = 0; i != VWidth; ++i)
+ if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
+ Elts.push_back(Undef);
+ UndefElts |= (1ULL << i);
+ } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
+ Elts.push_back(Undef);
+ UndefElts |= (1ULL << i);
+ } else { // Otherwise, defined.
+ Elts.push_back(CP->getOperand(i));
+ }
+
+ // If we changed the constant, return it.
+ Constant *NewCP = ConstantPacked::get(Elts);
+ return NewCP != CP ? NewCP : 0;
+ } else if (isa<ConstantAggregateZero>(V)) {
+ // Simplify the CAZ to a ConstantPacked where the non-demanded elements are
+ // set to undef.
+ const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
+ Constant *Zero = Constant::getNullValue(EltTy);
+ Constant *Undef = UndefValue::get(EltTy);
+ std::vector<Constant*> Elts;
+ for (unsigned i = 0; i != VWidth; ++i)
+ Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
+ UndefElts = DemandedElts ^ EltMask;
+ return ConstantPacked::get(Elts);
+ }
+
+ if (!V->hasOneUse()) { // Other users may use these bits.
+ if (Depth != 0) { // Not at the root.
+ // TODO: Just compute the UndefElts information recursively.
+ return false;
+ }
+ return false;
+ } else if (Depth == 10) { // Limit search depth.
+ return false;
+ }
+
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (!I) return false; // Only analyze instructions.
+
+ bool MadeChange = false;
+ uint64_t UndefElts2;
+ Value *TmpV;
+ switch (I->getOpcode()) {
+ default: break;
+
+ case Instruction::InsertElement: {
+ // If this is a variable index, we don't know which element it overwrites.
+ // demand exactly the same input as we produce.
+ ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
+ if (Idx == 0) {
+ // Note that we can't propagate undef elt info, because we don't know
+ // which elt is getting updated.
+ TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
+ UndefElts2, Depth+1);
+ if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
+ break;
+ }
+
+ // If this is inserting an element that isn't demanded, remove this
+ // insertelement.
+ unsigned IdxNo = Idx->getZExtValue();
+ if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
+ return AddSoonDeadInstToWorklist(*I, 0);
+
+ // Otherwise, the element inserted overwrites whatever was there, so the
+ // input demanded set is simpler than the output set.
+ TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
+ DemandedElts & ~(1ULL << IdxNo),
+ UndefElts, Depth+1);
+ if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
+
+ // The inserted element is defined.
+ UndefElts |= 1ULL << IdxNo;
+ break;
+ }
+
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor:
+ case Instruction::Add:
+ case Instruction::Sub:
+ case Instruction::Mul:
+ // div/rem demand all inputs, because they don't want divide by zero.
+ TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
+ UndefElts, Depth+1);
+ if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
+ TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
+ UndefElts2, Depth+1);
+ if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
+
+ // Output elements are undefined if both are undefined. Consider things
+ // like undef&0. The result is known zero, not undef.
+ UndefElts &= UndefElts2;
+ break;
+
+ case Instruction::Call: {
+ IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
+ if (!II) break;
+ switch (II->getIntrinsicID()) {
+ default: break;
+
+ // Binary vector operations that work column-wise. A dest element is a
+ // function of the corresponding input elements from the two inputs.
+ case Intrinsic::x86_sse_sub_ss:
+ case Intrinsic::x86_sse_mul_ss:
+ case Intrinsic::x86_sse_min_ss:
+ case Intrinsic::x86_sse_max_ss:
+ case Intrinsic::x86_sse2_sub_sd:
+ case Intrinsic::x86_sse2_mul_sd:
+ case Intrinsic::x86_sse2_min_sd:
+ case Intrinsic::x86_sse2_max_sd:
+ TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
+ UndefElts, Depth+1);
+ if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
+ TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
+ UndefElts2, Depth+1);
+ if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
+
+ // If only the low elt is demanded and this is a scalarizable intrinsic,
+ // scalarize it now.
+ if (DemandedElts == 1) {
+ switch (II->getIntrinsicID()) {
+ default: break;
+ case Intrinsic::x86_sse_sub_ss:
+ case Intrinsic::x86_sse_mul_ss:
+ case Intrinsic::x86_sse2_sub_sd:
+ case Intrinsic::x86_sse2_mul_sd:
+ // TODO: Lower MIN/MAX/ABS/etc
+ Value *LHS = II->getOperand(1);
+ Value *RHS = II->getOperand(2);
+ // Extract the element as scalars.
+ LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
+ RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
+
+ switch (II->getIntrinsicID()) {
+ default: assert(0 && "Case stmts out of sync!");
+ case Intrinsic::x86_sse_sub_ss:
+ case Intrinsic::x86_sse2_sub_sd:
+ TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
+ II->getName()), *II);
+ break;
+ case Intrinsic::x86_sse_mul_ss:
+ case Intrinsic::x86_sse2_mul_sd:
+ TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
+ II->getName()), *II);
+ break;
+ }
+
+ Instruction *New =
+ new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
+ II->getName());
+ InsertNewInstBefore(New, *II);
+ AddSoonDeadInstToWorklist(*II, 0);
+ return New;
+ }
+ }
+
+ // Output elements are undefined if both are undefined. Consider things
+ // like undef&0. The result is known zero, not undef.
+ UndefElts &= UndefElts2;
+ break;
+ }
+ break;
+ }
+ }
+ return MadeChange ? I : 0;
+}
+
+/// @returns true if the specified compare instruction is
+/// true when both operands are equal...
+/// @brief Determine if the ICmpInst returns true if both operands are equal
+static bool isTrueWhenEqual(ICmpInst &ICI) {
+ ICmpInst::Predicate pred = ICI.getPredicate();
+ return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
+ pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
+ pred == ICmpInst::ICMP_SLE;
+}
+
+/// @returns true if the specified compare instruction is
+/// true when both operands are equal...
+/// @brief Determine if the FCmpInst returns true if both operands are equal
+static bool isTrueWhenEqual(FCmpInst &FCI) {
+ FCmpInst::Predicate pred = FCI.getPredicate();
+ return pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ ||
+ pred == FCmpInst::FCMP_OGE || pred == FCmpInst::FCMP_UGE ||
+ pred == FCmpInst::FCMP_OLE || pred == FCmpInst::FCMP_ULE;
}
/// AssociativeOpt - Perform an optimization on an associative operator. This
static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
InstCombiner *IC) {
- if (isa<CastInst>(I)) {
+ if (CastInst *CI = dyn_cast<CastInst>(&I)) {
if (Constant *SOC = dyn_cast<Constant>(SO))
- return ConstantExpr::getCast(SOC, I.getType());
+ return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
- return IC->InsertNewInstBefore(new CastInst(SO, I.getType(),
- SO->getName() + ".cast"), I);
+ return IC->InsertNewInstBefore(CastInst::create(
+ CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
}
// Figure out if the constant is the left or the right argument.
Instruction *New;
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
+ else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
+ New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
+ SO->getName()+".cmp");
else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
New = new ShiftInst(SI->getOpcode(), Op0, Op1, SO->getName()+".sh");
else {
Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
PHINode *PN = cast<PHINode>(I.getOperand(0));
unsigned NumPHIValues = PN->getNumIncomingValues();
- if (!PN->hasOneUse() || NumPHIValues == 0 ||
- !isa<Constant>(PN->getIncomingValue(0))) return 0;
-
- // Check to see if all of the operands of the PHI are constants. If not, we
- // cannot do the transformation.
- for (unsigned i = 1; i != NumPHIValues; ++i)
- if (!isa<Constant>(PN->getIncomingValue(i)))
- return 0;
+ if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
+
+ // Check to see if all of the operands of the PHI are constants. If there is
+ // one non-constant value, remember the BB it is. If there is more than one
+ // bail out.
+ BasicBlock *NonConstBB = 0;
+ for (unsigned i = 0; i != NumPHIValues; ++i)
+ if (!isa<Constant>(PN->getIncomingValue(i))) {
+ if (NonConstBB) return 0; // More than one non-const value.
+ NonConstBB = PN->getIncomingBlock(i);
+
+ // If the incoming non-constant value is in I's block, we have an infinite
+ // loop.
+ if (NonConstBB == I.getParent())
+ return 0;
+ }
+
+ // If there is exactly one non-constant value, we can insert a copy of the
+ // operation in that block. However, if this is a critical edge, we would be
+ // inserting the computation one some other paths (e.g. inside a loop). Only
+ // do this if the pred block is unconditionally branching into the phi block.
+ if (NonConstBB) {
+ BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
+ if (!BI || !BI->isUnconditional()) return 0;
+ }
// Okay, we can do the transformation: create the new PHI node.
PHINode *NewPN = new PHINode(I.getType(), I.getName());
if (I.getNumOperands() == 2) {
Constant *C = cast<Constant>(I.getOperand(1));
for (unsigned i = 0; i != NumPHIValues; ++i) {
- Constant *InV = cast<Constant>(PN->getIncomingValue(i));
- NewPN->addIncoming(ConstantExpr::get(I.getOpcode(), InV, C),
- PN->getIncomingBlock(i));
+ Value *InV;
+ if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
+ if (CmpInst *CI = dyn_cast<CmpInst>(&I))
+ InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
+ else
+ InV = ConstantExpr::get(I.getOpcode(), InC, C);
+ } else {
+ assert(PN->getIncomingBlock(i) == NonConstBB);
+ if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
+ InV = BinaryOperator::create(BO->getOpcode(),
+ PN->getIncomingValue(i), C, "phitmp",
+ NonConstBB->getTerminator());
+ else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
+ InV = CmpInst::create(CI->getOpcode(),
+ CI->getPredicate(),
+ PN->getIncomingValue(i), C, "phitmp",
+ NonConstBB->getTerminator());
+ else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
+ InV = new ShiftInst(SI->getOpcode(),
+ PN->getIncomingValue(i), C, "phitmp",
+ NonConstBB->getTerminator());
+ else
+ assert(0 && "Unknown binop!");
+
+ WorkList.push_back(cast<Instruction>(InV));
+ }
+ NewPN->addIncoming(InV, PN->getIncomingBlock(i));
}
- } else {
- assert(isa<CastInst>(I) && "Unary op should be a cast!");
- const Type *RetTy = I.getType();
+ } else {
+ CastInst *CI = cast<CastInst>(&I);
+ const Type *RetTy = CI->getType();
for (unsigned i = 0; i != NumPHIValues; ++i) {
- Constant *InV = cast<Constant>(PN->getIncomingValue(i));
- NewPN->addIncoming(ConstantExpr::getCast(InV, RetTy),
- PN->getIncomingBlock(i));
+ Value *InV;
+ if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
+ InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
+ } else {
+ assert(PN->getIncomingBlock(i) == NonConstBB);
+ InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
+ I.getType(), "phitmp",
+ NonConstBB->getTerminator());
+ WorkList.push_back(cast<Instruction>(InV));
+ }
+ NewPN->addIncoming(InV, PN->getIncomingBlock(i));
}
}
return ReplaceInstUsesWith(I, NewPN);
return ReplaceInstUsesWith(I, RHS);
// X + 0 --> X
- if (!I.getType()->isFloatingPoint()) { // NOTE: -0 + +0 = +0.
+ if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
if (RHSC->isNullValue())
return ReplaceInstUsesWith(I, LHS);
} else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
return ReplaceInstUsesWith(I, LHS);
}
- // X + (signbit) --> X ^ signbit
if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
+ // X + (signbit) --> X ^ signbit
uint64_t Val = CI->getZExtValue();
if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
return BinaryOperator::createXor(LHS, RHS);
+
+ // See if SimplifyDemandedBits can simplify this. This handles stuff like
+ // (X & 254)+1 -> (X&254)|1
+ uint64_t KnownZero, KnownOne;
+ if (!isa<PackedType>(I.getType()) &&
+ SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
+ KnownZero, KnownOne))
+ return &I;
}
if (isa<PHINode>(LHS))
case 8: MiddleType = Type::SByteTy; break;
}
if (MiddleType) {
- Instruction *NewTrunc = new CastInst(XorLHS, MiddleType, "sext");
+ Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
InsertNewInstBefore(NewTrunc, I);
- return new CastInst(NewTrunc, I.getType());
+ return new SExtInst(NewTrunc, I.getType());
}
}
}
if (Anded == CRHS) {
// See if all bits from the first bit set in the Add RHS up are included
// in the mask. First, get the rightmost bit.
- uint64_t AddRHSV = CRHS->getRawValue();
+ uint64_t AddRHSV = CRHS->getZExtValue();
// Form a mask of all bits from the lowest bit added through the top.
uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
AddRHSHighBits &= C2->getType()->getIntegralTypeMask();
// See if the and mask includes all of these bits.
- uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getRawValue();
+ uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
if (AddRHSHighBits == AddRHSHighBitsAnd) {
// Okay, the xform is safe. Insert the new add pronto.
return R;
}
+ // add (cast *A to intptrtype) B ->
+ // cast (GEP (cast *A to sbyte*) B) ->
+ // intptrtype
+ {
+ CastInst *CI = dyn_cast<CastInst>(LHS);
+ Value *Other = RHS;
+ if (!CI) {
+ CI = dyn_cast<CastInst>(RHS);
+ Other = LHS;
+ }
+ if (CI && CI->getType()->isSized() &&
+ (CI->getType()->getPrimitiveSize() ==
+ TD->getIntPtrType()->getPrimitiveSize())
+ && isa<PointerType>(CI->getOperand(0)->getType())) {
+ Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
+ PointerType::get(Type::SByteTy), I);
+ I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
+ return new PtrToIntInst(I2, CI->getType());
+ }
+ }
+
return Changed ? &I : 0;
}
// highest order bit set.
static bool isSignBit(ConstantInt *CI) {
unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
- return (CI->getRawValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
+ return (CI->getZExtValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
}
/// RemoveNoopCast - Strip off nonconverting casts from the value.
if (C->isNullValue()) {
Value *NoopCastedRHS = RemoveNoopCast(Op1);
if (ShiftInst *SI = dyn_cast<ShiftInst>(NoopCastedRHS))
- if (SI->getOpcode() == Instruction::Shr)
- if (ConstantUInt *CU = dyn_cast<ConstantUInt>(SI->getOperand(1))) {
- const Type *NewTy;
- if (SI->getType()->isSigned())
- NewTy = SI->getType()->getUnsignedVersion();
- else
- NewTy = SI->getType()->getSignedVersion();
+ if (SI->getOpcode() == Instruction::LShr) {
+ if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
// Check to see if we are shifting out everything but the sign bit.
- if (CU->getValue() == SI->getType()->getPrimitiveSizeInBits()-1) {
- // Ok, the transformation is safe. Insert a cast of the incoming
- // value, then the new shift, then the new cast.
- Instruction *FirstCast = new CastInst(SI->getOperand(0), NewTy,
- SI->getOperand(0)->getName());
- Value *InV = InsertNewInstBefore(FirstCast, I);
- Instruction *NewShift = new ShiftInst(Instruction::Shr, FirstCast,
- CU, SI->getName());
- if (NewShift->getType() == I.getType())
- return NewShift;
- else {
- InV = InsertNewInstBefore(NewShift, I);
- return new CastInst(NewShift, I.getType());
- }
+ if (CU->getZExtValue() ==
+ SI->getType()->getPrimitiveSizeInBits()-1) {
+ // Ok, the transformation is safe. Insert AShr.
+ return new ShiftInst(Instruction::AShr, SI->getOperand(0),
+ CU, SI->getName());
}
}
- }
-
- // Try to fold constant sub into select arguments.
- if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
+ }
+ else if (SI->getOpcode() == Instruction::AShr) {
+ if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
+ // Check to see if we are shifting out everything but the sign bit.
+ if (CU->getZExtValue() ==
+ SI->getType()->getPrimitiveSizeInBits()-1) {
+ // Ok, the transformation is safe. Insert LShr.
+ return new ShiftInst(Instruction::LShr, SI->getOperand(0),
+ CU, SI->getName());
+ }
+ }
+ }
+ }
+
+ // Try to fold constant sub into select arguments.
+ if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
if (Instruction *R = FoldOpIntoSelect(I, SI, this))
return R;
if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
if (Op1I->getOpcode() == Instruction::Add &&
- !Op0->getType()->isFloatingPoint()) {
+ !Op0->getType()->isFPOrFPVector()) {
if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
// is not used by anyone else...
//
if (Op1I->getOpcode() == Instruction::Sub &&
- !Op1I->getType()->isFloatingPoint()) {
+ !Op1I->getType()->isFPOrFPVector()) {
// Swap the two operands of the subexpr...
Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
Op1I->setOperand(0, IIOp1);
return BinaryOperator::createAnd(Op0, NewNot);
}
- // -(X sdiv C) -> (X sdiv -C)
- if (Op1I->getOpcode() == Instruction::Div)
- if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
+ // 0 - (X sdiv C) -> (X sdiv -C)
+ if (Op1I->getOpcode() == Instruction::SDiv)
+ if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
if (CSI->isNullValue())
if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
- return BinaryOperator::createDiv(Op1I->getOperand(0),
+ return BinaryOperator::createSDiv(Op1I->getOperand(0),
ConstantExpr::getNeg(DivRHS));
// X - X*C --> X * (1-C)
}
}
- if (!Op0->getType()->isFloatingPoint())
+ if (!Op0->getType()->isFPOrFPVector())
if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
if (Op0I->getOpcode() == Instruction::Add) {
if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
return 0;
}
-/// isSignBitCheck - Given an exploded setcc instruction, return true if it is
+/// isSignBitCheck - Given an exploded icmp instruction, return true if it
/// really just returns true if the most significant (sign) bit is set.
-static bool isSignBitCheck(unsigned Opcode, Value *LHS, ConstantInt *RHS) {
- if (RHS->getType()->isSigned()) {
- // True if source is LHS < 0 or LHS <= -1
- return Opcode == Instruction::SetLT && RHS->isNullValue() ||
- Opcode == Instruction::SetLE && RHS->isAllOnesValue();
- } else {
- ConstantUInt *RHSC = cast<ConstantUInt>(RHS);
- // True if source is LHS > 127 or LHS >= 128, where the constants depend on
- // the size of the integer type.
- if (Opcode == Instruction::SetGE)
- return RHSC->getValue() ==
- 1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1);
- if (Opcode == Instruction::SetGT)
- return RHSC->getValue() ==
+static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
+ switch (pred) {
+ case ICmpInst::ICMP_SLT:
+ // True if LHS s< RHS and RHS == 0
+ return RHS->isNullValue();
+ case ICmpInst::ICMP_SLE:
+ // True if LHS s<= RHS and RHS == -1
+ return RHS->isAllOnesValue();
+ case ICmpInst::ICMP_UGE:
+ // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
+ return RHS->getZExtValue() == (1ULL <<
+ (RHS->getType()->getPrimitiveSizeInBits()-1));
+ case ICmpInst::ICMP_UGT:
+ // True if LHS u> RHS and RHS == high-bit-mask - 1
+ return RHS->getZExtValue() ==
(1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
+ default:
+ return false;
}
- return false;
}
Instruction *InstCombiner::visitMul(BinaryOperator &I) {
if (CI->isAllOnesValue()) // X * -1 == 0 - X
return BinaryOperator::createNeg(Op0, I.getName());
- int64_t Val = (int64_t)cast<ConstantInt>(CI)->getRawValue();
+ int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
uint64_t C = Log2_64(Val);
return new ShiftInst(Instruction::Shl, Op0,
- ConstantUInt::get(Type::UByteTy, C));
+ ConstantInt::get(Type::UByteTy, C));
}
} else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
if (Op1F->isNullValue())
// See if we can simplify things based on how the boolean was originally
// formed.
CastInst *BoolCast = 0;
- if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(0)))
+ if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
if (CI->getOperand(0)->getType() == Type::BoolTy)
BoolCast = CI;
if (!BoolCast)
- if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(1)))
+ if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
if (CI->getOperand(0)->getType() == Type::BoolTy)
BoolCast = CI;
if (BoolCast) {
- if (SetCondInst *SCI = dyn_cast<SetCondInst>(BoolCast->getOperand(0))) {
+ if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
const Type *SCOpTy = SCIOp0->getType();
- // If the setcc is true iff the sign bit of X is set, then convert this
+ // If the icmp is true iff the sign bit of X is set, then convert this
// multiply into a shift/and combination.
if (isa<ConstantInt>(SCIOp1) &&
- isSignBitCheck(SCI->getOpcode(), SCIOp0, cast<ConstantInt>(SCIOp1))) {
+ isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
// Shift the X value right to turn it into "all signbits".
- Constant *Amt = ConstantUInt::get(Type::UByteTy,
+ Constant *Amt = ConstantInt::get(Type::UByteTy,
SCOpTy->getPrimitiveSizeInBits()-1);
- if (SCIOp0->getType()->isUnsigned()) {
- const Type *NewTy = SCIOp0->getType()->getSignedVersion();
- SCIOp0 = InsertNewInstBefore(new CastInst(SCIOp0, NewTy,
- SCIOp0->getName()), I);
- }
-
Value *V =
- InsertNewInstBefore(new ShiftInst(Instruction::Shr, SCIOp0, Amt,
+ InsertNewInstBefore(new ShiftInst(Instruction::AShr, SCIOp0, Amt,
BoolCast->getOperand(0)->getName()+
".mask"), I);
// If the multiply type is not the same as the source type, sign extend
// or truncate to the multiply type.
- if (I.getType() != V->getType())
- V = InsertNewInstBefore(new CastInst(V, I.getType(), V->getName()),I);
+ if (I.getType() != V->getType()) {
+ unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
+ unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
+ Instruction::CastOps opcode =
+ (SrcBits == DstBits ? Instruction::BitCast :
+ (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
+ V = InsertCastBefore(opcode, V, I.getType(), I);
+ }
Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
return BinaryOperator::createAnd(V, OtherOp);
return Changed ? &I : 0;
}
-Instruction *InstCombiner::visitDiv(BinaryOperator &I) {
+/// This function implements the transforms on div instructions that work
+/// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
+/// used by the visitors to those instructions.
+/// @brief Transforms common to all three div instructions
+Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
- if (isa<UndefValue>(Op0)) // undef / X -> 0
+ // undef / X -> 0
+ if (isa<UndefValue>(Op0))
return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
+
+ // X / undef -> undef
if (isa<UndefValue>(Op1))
- return ReplaceInstUsesWith(I, Op1); // X / undef -> undef
+ return ReplaceInstUsesWith(I, Op1);
+
+ // Handle cases involving: div X, (select Cond, Y, Z)
+ if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
+ // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
+ // same basic block, then we replace the select with Y, and the condition
+ // of the select with false (if the cond value is in the same BB). If the
+ // select has uses other than the div, this allows them to be simplified
+ // also. Note that div X, Y is just as good as div X, 0 (undef)
+ if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
+ if (ST->isNullValue()) {
+ Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
+ if (CondI && CondI->getParent() == I.getParent())
+ UpdateValueUsesWith(CondI, ConstantBool::getFalse());
+ else if (I.getParent() != SI->getParent() || SI->hasOneUse())
+ I.setOperand(1, SI->getOperand(2));
+ else
+ UpdateValueUsesWith(SI, SI->getOperand(2));
+ return &I;
+ }
+
+ // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
+ if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
+ if (ST->isNullValue()) {
+ Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
+ if (CondI && CondI->getParent() == I.getParent())
+ UpdateValueUsesWith(CondI, ConstantBool::getTrue());
+ else if (I.getParent() != SI->getParent() || SI->hasOneUse())
+ I.setOperand(1, SI->getOperand(1));
+ else
+ UpdateValueUsesWith(SI, SI->getOperand(1));
+ return &I;
+ }
+ }
+
+ return 0;
+}
+
+/// This function implements the transforms common to both integer division
+/// instructions (udiv and sdiv). It is called by the visitors to those integer
+/// division instructions.
+/// @brief Common integer divide transforms
+Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
+ Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
+
+ if (Instruction *Common = commonDivTransforms(I))
+ return Common;
if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
// div X, 1 == X
if (RHS->equalsInt(1))
return ReplaceInstUsesWith(I, Op0);
- // div X, -1 == -X
- if (RHS->isAllOnesValue())
- return BinaryOperator::createNeg(Op0);
-
+ // (X / C1) / C2 -> X / (C1*C2)
if (Instruction *LHS = dyn_cast<Instruction>(Op0))
- if (LHS->getOpcode() == Instruction::Div)
+ if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
- // (X / C1) / C2 -> X / (C1*C2)
- return BinaryOperator::createDiv(LHS->getOperand(0),
- ConstantExpr::getMul(RHS, LHSRHS));
+ return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
+ ConstantExpr::getMul(RHS, LHSRHS));
}
- // Check to see if this is an unsigned division with an exact power of 2,
- // if so, convert to a right shift.
- if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
- if (uint64_t Val = C->getValue()) // Don't break X / 0
- if (isPowerOf2_64(Val)) {
- uint64_t C = Log2_64(Val);
- return new ShiftInst(Instruction::Shr, Op0,
- ConstantUInt::get(Type::UByteTy, C));
- }
-
- // -X/C -> X/-C
- if (RHS->getType()->isSigned())
- if (Value *LHSNeg = dyn_castNegVal(Op0))
- return BinaryOperator::createDiv(LHSNeg, ConstantExpr::getNeg(RHS));
-
- if (!RHS->isNullValue()) {
+ if (!RHS->isNullValue()) { // avoid X udiv 0
if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
if (Instruction *R = FoldOpIntoSelect(I, SI, this))
return R;
}
}
- // If this is 'udiv X, (Cond ? C1, C2)' where C1&C2 are powers of two,
- // transform this into: '(Cond ? (udiv X, C1) : (udiv X, C2))'.
- if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
- if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
- if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
- if (STO->getValue() == 0) { // Couldn't be this argument.
- I.setOperand(1, SFO);
- return &I;
- } else if (SFO->getValue() == 0) {
- I.setOperand(1, STO);
- return &I;
- }
-
- uint64_t TVA = STO->getValue(), FVA = SFO->getValue();
- if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
- unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
- Constant *TC = ConstantUInt::get(Type::UByteTy, TSA);
- Instruction *TSI = new ShiftInst(Instruction::Shr, Op0,
- TC, SI->getName()+".t");
- TSI = InsertNewInstBefore(TSI, I);
-
- Constant *FC = ConstantUInt::get(Type::UByteTy, FSA);
- Instruction *FSI = new ShiftInst(Instruction::Shr, Op0,
- FC, SI->getName()+".f");
- FSI = InsertNewInstBefore(FSI, I);
- return new SelectInst(SI->getOperand(0), TSI, FSI);
- }
- }
-
// 0 / X == 0, we don't need to preserve faults!
if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
if (LHS->equalsInt(0))
return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
- if (I.getType()->isSigned()) {
- // If the sign bits of both operands are zero (i.e. we can prove they are
- // unsigned inputs), turn this into a udiv.
- uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
- if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
- const Type *NTy = Op0->getType()->getUnsignedVersion();
- Instruction *LHS = new CastInst(Op0, NTy, Op0->getName());
- InsertNewInstBefore(LHS, I);
- Value *RHS;
- if (Constant *R = dyn_cast<Constant>(Op1))
- RHS = ConstantExpr::getCast(R, NTy);
- else
- RHS = InsertNewInstBefore(new CastInst(Op1, NTy, Op1->getName()), I);
- Instruction *Div = BinaryOperator::createDiv(LHS, RHS, I.getName());
- InsertNewInstBefore(Div, I);
- return new CastInst(Div, I.getType());
- }
- } else {
- // Known to be an unsigned division.
- if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
- // Turn A / (C1 << N), where C1 is "1<<C2" into A >> (N+C2) [udiv only].
- if (RHSI->getOpcode() == Instruction::Shl &&
- isa<ConstantUInt>(RHSI->getOperand(0))) {
- unsigned C1 = cast<ConstantUInt>(RHSI->getOperand(0))->getRawValue();
- if (isPowerOf2_64(C1)) {
- unsigned C2 = Log2_64(C1);
- Value *Add = RHSI->getOperand(1);
- if (C2) {
- Constant *C2V = ConstantUInt::get(Add->getType(), C2);
- Add = InsertNewInstBefore(BinaryOperator::createAdd(Add, C2V,
- "tmp"), I);
- }
- return new ShiftInst(Instruction::Shr, Op0, Add);
+ return 0;
+}
+
+Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
+ Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
+
+ // Handle the integer div common cases
+ if (Instruction *Common = commonIDivTransforms(I))
+ return Common;
+
+ // X udiv C^2 -> X >> C
+ // Check to see if this is an unsigned division with an exact power of 2,
+ // if so, convert to a right shift.
+ if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
+ if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
+ if (isPowerOf2_64(Val)) {
+ uint64_t ShiftAmt = Log2_64(Val);
+ return new ShiftInst(Instruction::LShr, Op0,
+ ConstantInt::get(Type::UByteTy, ShiftAmt));
+ }
+ }
+
+ // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
+ if (ShiftInst *RHSI = dyn_cast<ShiftInst>(I.getOperand(1))) {
+ if (RHSI->getOpcode() == Instruction::Shl &&
+ isa<ConstantInt>(RHSI->getOperand(0))) {
+ uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
+ if (isPowerOf2_64(C1)) {
+ Value *N = RHSI->getOperand(1);
+ const Type *NTy = N->getType();
+ if (uint64_t C2 = Log2_64(C1)) {
+ Constant *C2V = ConstantInt::get(NTy, C2);
+ N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
}
+ return new ShiftInst(Instruction::LShr, Op0, N);
}
}
}
+ // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
+ // where C1&C2 are powers of two.
+ if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
+ if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
+ if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2)))
+ if (!STO->isNullValue() && !STO->isNullValue()) {
+ uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
+ if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
+ // Compute the shift amounts
+ unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
+ // Construct the "on true" case of the select
+ Constant *TC = ConstantInt::get(Type::UByteTy, TSA);
+ Instruction *TSI =
+ new ShiftInst(Instruction::LShr, Op0, TC, SI->getName()+".t");
+ TSI = InsertNewInstBefore(TSI, I);
+
+ // Construct the "on false" case of the select
+ Constant *FC = ConstantInt::get(Type::UByteTy, FSA);
+ Instruction *FSI =
+ new ShiftInst(Instruction::LShr, Op0, FC, SI->getName()+".f");
+ FSI = InsertNewInstBefore(FSI, I);
+
+ // construct the select instruction and return it.
+ return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
+ }
+ }
+ }
+ return 0;
+}
+
+Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
+ Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
+
+ // Handle the integer div common cases
+ if (Instruction *Common = commonIDivTransforms(I))
+ return Common;
+
+ if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
+ // sdiv X, -1 == -X
+ if (RHS->isAllOnesValue())
+ return BinaryOperator::createNeg(Op0);
+
+ // -X/C -> X/-C
+ if (Value *LHSNeg = dyn_castNegVal(Op0))
+ return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
+ }
+
+ // If the sign bits of both operands are zero (i.e. we can prove they are
+ // unsigned inputs), turn this into a udiv.
+ if (I.getType()->isInteger()) {
+ uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
+ if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
+ return BinaryOperator::createUDiv(Op0, Op1, I.getName());
+ }
+ }
+
return 0;
}
+Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
+ return commonDivTransforms(I);
+}
/// GetFactor - If we can prove that the specified value is at least a multiple
/// of some factor, return that factor.
unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
if (Zeros != V->getType()->getPrimitiveSizeInBits())
return ConstantExpr::getShl(Result,
- ConstantUInt::get(Type::UByteTy, Zeros));
+ ConstantInt::get(Type::UByteTy, Zeros));
}
- } else if (I->getOpcode() == Instruction::Cast) {
- Value *Op = I->getOperand(0);
+ } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
// Only handle int->int casts.
- if (!Op->getType()->isInteger()) return Result;
- return ConstantExpr::getCast(GetFactor(Op), V->getType());
+ if (!CI->isIntegerCast())
+ return Result;
+ Value *Op = CI->getOperand(0);
+ return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
}
return Result;
}
-Instruction *InstCombiner::visitRem(BinaryOperator &I) {
+/// This function implements the transforms on rem instructions that work
+/// regardless of the kind of rem instruction it is (urem, srem, or frem). It
+/// is used by the visitors to those instructions.
+/// @brief Transforms common to all three rem instructions
+Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
-
+
// 0 % X == 0, we don't need to preserve faults!
if (Constant *LHS = dyn_cast<Constant>(Op0))
if (LHS->isNullValue())
return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
if (isa<UndefValue>(Op1))
return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
-
- if (I.getType()->isSigned()) {
- if (Value *RHSNeg = dyn_castNegVal(Op1))
- if (!isa<ConstantSInt>(RHSNeg) ||
- cast<ConstantSInt>(RHSNeg)->getValue() > 0) {
- // X % -Y -> X % Y
- AddUsesToWorkList(I);
- I.setOperand(1, RHSNeg);
+
+ // Handle cases involving: rem X, (select Cond, Y, Z)
+ if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
+ // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
+ // the same basic block, then we replace the select with Y, and the
+ // condition of the select with false (if the cond value is in the same
+ // BB). If the select has uses other than the div, this allows them to be
+ // simplified also.
+ if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
+ if (ST->isNullValue()) {
+ Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
+ if (CondI && CondI->getParent() == I.getParent())
+ UpdateValueUsesWith(CondI, ConstantBool::getFalse());
+ else if (I.getParent() != SI->getParent() || SI->hasOneUse())
+ I.setOperand(1, SI->getOperand(2));
+ else
+ UpdateValueUsesWith(SI, SI->getOperand(2));
+ return &I;
+ }
+ // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
+ if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
+ if (ST->isNullValue()) {
+ Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
+ if (CondI && CondI->getParent() == I.getParent())
+ UpdateValueUsesWith(CondI, ConstantBool::getTrue());
+ else if (I.getParent() != SI->getParent() || SI->hasOneUse())
+ I.setOperand(1, SI->getOperand(1));
+ else
+ UpdateValueUsesWith(SI, SI->getOperand(1));
return &I;
}
-
- // If the top bits of both operands are zero (i.e. we can prove they are
- // unsigned inputs), turn this into a urem.
- uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
- if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
- const Type *NTy = Op0->getType()->getUnsignedVersion();
- Instruction *LHS = new CastInst(Op0, NTy, Op0->getName());
- InsertNewInstBefore(LHS, I);
- Value *RHS;
- if (Constant *R = dyn_cast<Constant>(Op1))
- RHS = ConstantExpr::getCast(R, NTy);
- else
- RHS = InsertNewInstBefore(new CastInst(Op1, NTy, Op1->getName()), I);
- Instruction *Rem = BinaryOperator::createRem(LHS, RHS, I.getName());
- InsertNewInstBefore(Rem, I);
- return new CastInst(Rem, I.getType());
- }
}
+ return 0;
+}
+
+/// This function implements the transforms common to both integer remainder
+/// instructions (urem and srem). It is called by the visitors to those integer
+/// remainder instructions.
+/// @brief Common integer remainder transforms
+Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
+ Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
+
+ if (Instruction *common = commonRemTransforms(I))
+ return common;
+
if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
// X % 0 == undef, we don't need to preserve faults!
if (RHS->equalsInt(0))
if (RHS->equalsInt(1)) // X % 1 == 0
return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
- // Check to see if this is an unsigned remainder with an exact power of 2,
- // if so, convert to a bitwise and.
- if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
- if (isPowerOf2_64(C->getValue()))
- return BinaryOperator::createAnd(Op0, SubOne(C));
-
if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
if (Instruction *R = FoldOpIntoSelect(I, SI, this))
if (Instruction *NV = FoldOpIntoPhi(I))
return NV;
}
-
- // X*C1%C2 --> 0 iff C1%C2 == 0
- if (ConstantExpr::getRem(GetFactor(Op0I), RHS)->isNullValue())
+ // (X * C1) % C2 --> 0 iff C1 % C2 == 0
+ if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
}
}
+ return 0;
+}
+
+Instruction *InstCombiner::visitURem(BinaryOperator &I) {
+ Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
+
+ if (Instruction *common = commonIRemTransforms(I))
+ return common;
+
+ if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
+ // X urem C^2 -> X and C
+ // Check to see if this is an unsigned remainder with an exact power of 2,
+ // if so, convert to a bitwise and.
+ if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
+ if (isPowerOf2_64(C->getZExtValue()))
+ return BinaryOperator::createAnd(Op0, SubOne(C));
+ }
+
if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
- // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1) [urem only].
- if (I.getType()->isUnsigned() &&
- RHSI->getOpcode() == Instruction::Shl &&
- isa<ConstantUInt>(RHSI->getOperand(0))) {
- unsigned C1 = cast<ConstantUInt>(RHSI->getOperand(0))->getRawValue();
+ // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
+ if (RHSI->getOpcode() == Instruction::Shl &&
+ isa<ConstantInt>(RHSI->getOperand(0))) {
+ unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
if (isPowerOf2_64(C1)) {
Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
return BinaryOperator::createAnd(Op0, Add);
}
}
-
- // If this is 'urem X, (Cond ? C1, C2)' where C1&C2 are powers of two,
- // transform this into: '(Cond ? (urem X, C1) : (urem X, C2))'.
- if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
- if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
- if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
- if (STO->getValue() == 0) { // Couldn't be this argument.
- I.setOperand(1, SFO);
- return &I;
- } else if (SFO->getValue() == 0) {
- I.setOperand(1, STO);
- return &I;
- }
-
- if (isPowerOf2_64(STO->getValue()) && isPowerOf2_64(SFO->getValue())){
- Value *TrueAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
- SubOne(STO), SI->getName()+".t"), I);
- Value *FalseAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
- SubOne(SFO), SI->getName()+".f"), I);
- return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
- }
+ }
+
+ // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
+ // where C1&C2 are powers of two.
+ if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
+ if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
+ if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
+ // STO == 0 and SFO == 0 handled above.
+ if (isPowerOf2_64(STO->getZExtValue()) &&
+ isPowerOf2_64(SFO->getZExtValue())) {
+ Value *TrueAnd = InsertNewInstBefore(
+ BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
+ Value *FalseAnd = InsertNewInstBefore(
+ BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
+ return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
}
+ }
}
return 0;
}
-// isMaxValueMinusOne - return true if this is Max-1
-static bool isMaxValueMinusOne(const ConstantInt *C) {
- if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C))
- return CU->getValue() == C->getType()->getIntegralTypeMask()-1;
+Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
+ Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
- const ConstantSInt *CS = cast<ConstantSInt>(C);
+ if (Instruction *common = commonIRemTransforms(I))
+ return common;
+
+ if (Value *RHSNeg = dyn_castNegVal(Op1))
+ if (!isa<ConstantInt>(RHSNeg) ||
+ cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
+ // X % -Y -> X % Y
+ AddUsesToWorkList(I);
+ I.setOperand(1, RHSNeg);
+ return &I;
+ }
+
+ // If the top bits of both operands are zero (i.e. we can prove they are
+ // unsigned inputs), turn this into a urem.
+ uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
+ if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
+ // X srem Y -> X urem Y, iff X and Y don't have sign bit set
+ return BinaryOperator::createURem(Op0, Op1, I.getName());
+ }
- // Calculate 0111111111..11111
- unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
- int64_t Val = INT64_MAX; // All ones
- Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
- return CS->getValue() == Val-1;
+ return 0;
}
-// isMinValuePlusOne - return true if this is Min+1
-static bool isMinValuePlusOne(const ConstantInt *C) {
- if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C))
- return CU->getValue() == 1;
+Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
+ return commonRemTransforms(I);
+}
- const ConstantSInt *CS = cast<ConstantSInt>(C);
+// isMaxValueMinusOne - return true if this is Max-1
+static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
+ if (isSigned) {
+ // Calculate 0111111111..11111
+ unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
+ int64_t Val = INT64_MAX; // All ones
+ Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
+ return C->getSExtValue() == Val-1;
+ }
+ return C->getZExtValue() == C->getType()->getIntegralTypeMask()-1;
+}
- // Calculate 1111111111000000000000
- unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
- int64_t Val = -1; // All ones
- Val <<= TypeBits-1; // Shift over to the right spot
- return CS->getValue() == Val+1;
+// isMinValuePlusOne - return true if this is Min+1
+static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
+ if (isSigned) {
+ // Calculate 1111111111000000000000
+ unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
+ int64_t Val = -1; // All ones
+ Val <<= TypeBits-1; // Shift over to the right spot
+ return C->getSExtValue() == Val+1;
+ }
+ return C->getZExtValue() == 1; // unsigned
}
// isOneBitSet - Return true if there is exactly one bit set in the specified
// constant.
static bool isOneBitSet(const ConstantInt *CI) {
- uint64_t V = CI->getRawValue();
+ uint64_t V = CI->getZExtValue();
return V && (V & (V-1)) == 0;
}
#if 0 // Currently unused
// isLowOnes - Return true if the constant is of the form 0+1+.
static bool isLowOnes(const ConstantInt *CI) {
- uint64_t V = CI->getRawValue();
+ uint64_t V = CI->getZExtValue();
// There won't be bits set in parts that the type doesn't contain.
- V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
+ V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
uint64_t U = V+1; // If it is low ones, this should be a power of two.
return U && V && (U & V) == 0;
// isHighOnes - Return true if the constant is of the form 1+0+.
// This is the same as lowones(~X).
static bool isHighOnes(const ConstantInt *CI) {
- uint64_t V = ~CI->getRawValue();
+ uint64_t V = ~CI->getZExtValue();
if (~V == 0) return false; // 0's does not match "1+"
// There won't be bits set in parts that the type doesn't contain.
- V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
+ V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
uint64_t U = V+1; // If it is low ones, this should be a power of two.
return U && V && (U & V) == 0;
}
-
-/// getSetCondCode - Encode a setcc opcode into a three bit mask. These bits
+/// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
/// are carefully arranged to allow folding of expressions such as:
///
/// (A < B) | (A > B) --> (A != B)
///
-/// Bit value '4' represents that the comparison is true if A > B, bit value '2'
-/// represents that the comparison is true if A == B, and bit value '1' is true
-/// if A < B.
+/// Note that this is only valid if the first and second predicates have the
+/// same sign. Is illegal to do: (A u< B) | (A s> B)
+///
+/// Three bits are used to represent the condition, as follows:
+/// 0 A > B
+/// 1 A == B
+/// 2 A < B
///
-static unsigned getSetCondCode(const SetCondInst *SCI) {
- switch (SCI->getOpcode()) {
+/// <=> Value Definition
+/// 000 0 Always false
+/// 001 1 A > B
+/// 010 2 A == B
+/// 011 3 A >= B
+/// 100 4 A < B
+/// 101 5 A != B
+/// 110 6 A <= B
+/// 111 7 Always true
+///
+static unsigned getICmpCode(const ICmpInst *ICI) {
+ switch (ICI->getPredicate()) {
// False -> 0
- case Instruction::SetGT: return 1;
- case Instruction::SetEQ: return 2;
- case Instruction::SetGE: return 3;
- case Instruction::SetLT: return 4;
- case Instruction::SetNE: return 5;
- case Instruction::SetLE: return 6;
+ case ICmpInst::ICMP_UGT: return 1; // 001
+ case ICmpInst::ICMP_SGT: return 1; // 001
+ case ICmpInst::ICMP_EQ: return 2; // 010
+ case ICmpInst::ICMP_UGE: return 3; // 011
+ case ICmpInst::ICMP_SGE: return 3; // 011
+ case ICmpInst::ICMP_ULT: return 4; // 100
+ case ICmpInst::ICMP_SLT: return 4; // 100
+ case ICmpInst::ICMP_NE: return 5; // 101
+ case ICmpInst::ICMP_ULE: return 6; // 110
+ case ICmpInst::ICMP_SLE: return 6; // 110
// True -> 7
default:
- assert(0 && "Invalid SetCC opcode!");
+ assert(0 && "Invalid ICmp predicate!");
return 0;
}
}
-/// getSetCCValue - This is the complement of getSetCondCode, which turns an
-/// opcode and two operands into either a constant true or false, or a brand new
-/// SetCC instruction.
-static Value *getSetCCValue(unsigned Opcode, Value *LHS, Value *RHS) {
- switch (Opcode) {
- case 0: return ConstantBool::False;
- case 1: return new SetCondInst(Instruction::SetGT, LHS, RHS);
- case 2: return new SetCondInst(Instruction::SetEQ, LHS, RHS);
- case 3: return new SetCondInst(Instruction::SetGE, LHS, RHS);
- case 4: return new SetCondInst(Instruction::SetLT, LHS, RHS);
- case 5: return new SetCondInst(Instruction::SetNE, LHS, RHS);
- case 6: return new SetCondInst(Instruction::SetLE, LHS, RHS);
- case 7: return ConstantBool::True;
- default: assert(0 && "Illegal SetCCCode!"); return 0;
+/// getICmpValue - This is the complement of getICmpCode, which turns an
+/// opcode and two operands into either a constant true or false, or a brand
+/// new /// ICmp instruction. The sign is passed in to determine which kind
+/// of predicate to use in new icmp instructions.
+static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
+ switch (code) {
+ default: assert(0 && "Illegal ICmp code!");
+ case 0: return ConstantBool::getFalse();
+ case 1:
+ if (sign)
+ return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
+ else
+ return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
+ case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
+ case 3:
+ if (sign)
+ return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
+ else
+ return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
+ case 4:
+ if (sign)
+ return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
+ else
+ return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
+ case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
+ case 6:
+ if (sign)
+ return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
+ else
+ return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
+ case 7: return ConstantBool::getTrue();
}
}
-// FoldSetCCLogical - Implements (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
-struct FoldSetCCLogical {
+static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
+ return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
+ (ICmpInst::isSignedPredicate(p1) &&
+ (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
+ (ICmpInst::isSignedPredicate(p2) &&
+ (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
+}
+
+namespace {
+// FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
+struct FoldICmpLogical {
InstCombiner &IC;
Value *LHS, *RHS;
- FoldSetCCLogical(InstCombiner &ic, SetCondInst *SCI)
- : IC(ic), LHS(SCI->getOperand(0)), RHS(SCI->getOperand(1)) {}
+ ICmpInst::Predicate pred;
+ FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
+ : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
+ pred(ICI->getPredicate()) {}
bool shouldApply(Value *V) const {
- if (SetCondInst *SCI = dyn_cast<SetCondInst>(V))
- return (SCI->getOperand(0) == LHS && SCI->getOperand(1) == RHS ||
- SCI->getOperand(0) == RHS && SCI->getOperand(1) == LHS);
+ if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
+ if (PredicatesFoldable(pred, ICI->getPredicate()))
+ return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
+ ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
return false;
}
- Instruction *apply(BinaryOperator &Log) const {
- SetCondInst *SCI = cast<SetCondInst>(Log.getOperand(0));
- if (SCI->getOperand(0) != LHS) {
- assert(SCI->getOperand(1) == LHS);
- SCI->swapOperands(); // Swap the LHS and RHS of the SetCC
+ Instruction *apply(Instruction &Log) const {
+ ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
+ if (ICI->getOperand(0) != LHS) {
+ assert(ICI->getOperand(1) == LHS);
+ ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
}
- unsigned LHSCode = getSetCondCode(SCI);
- unsigned RHSCode = getSetCondCode(cast<SetCondInst>(Log.getOperand(1)));
+ unsigned LHSCode = getICmpCode(ICI);
+ unsigned RHSCode = getICmpCode(cast<ICmpInst>(Log.getOperand(1)));
unsigned Code;
switch (Log.getOpcode()) {
case Instruction::And: Code = LHSCode & RHSCode; break;
default: assert(0 && "Illegal logical opcode!"); return 0;
}
- Value *RV = getSetCCValue(Code, LHS, RHS);
+ Value *RV = getICmpValue(ICmpInst::isSignedPredicate(pred), Code, LHS, RHS);
if (Instruction *I = dyn_cast<Instruction>(RV))
return I;
// Otherwise, it's a constant boolean value...
return IC.ReplaceInstUsesWith(Log, RV);
}
};
+} // end anonymous namespace
// OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
// the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
// Adding a one to a single bit bit-field should be turned into an XOR
// of the bit. First thing to check is to see if this AND is with a
// single bit constant.
- uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getRawValue();
+ uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
// Clear bits that are not part of the constant.
AndRHSV &= AndRHS->getType()->getIntegralTypeMask();
// Ok, at this point, we know that we are masking the result of the
// ADD down to exactly one bit. If the constant we are adding has
// no bits set below this bit, then we can eliminate the ADD.
- uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getRawValue();
+ uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
// Check to see if any bits below the one bit set in AndRHSV are set.
if ((AddRHS & (AndRHSV-1)) == 0) {
}
break;
}
- case Instruction::Shr:
+ case Instruction::LShr:
+ {
// We know that the AND will not produce any of the bits shifted in, so if
// the anded constant includes them, clear them now! This only applies to
// unsigned shifts, because a signed shr may bring in set bits!
//
- if (AndRHS->getType()->isUnsigned()) {
+ Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
+ Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
+ Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
+
+ if (CI == ShrMask) { // Masking out bits that the shift already masks.
+ return ReplaceInstUsesWith(TheAnd, Op);
+ } else if (CI != AndRHS) {
+ TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
+ return &TheAnd;
+ }
+ break;
+ }
+ case Instruction::AShr:
+ // Signed shr.
+ // See if this is shifting in some sign extension, then masking it out
+ // with an and.
+ if (Op->hasOneUse()) {
Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
- Constant *ShrMask = ConstantExpr::getShr(AllOne, OpRHS);
- Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
-
- if (CI == ShrMask) { // Masking out bits that the shift already masks.
- return ReplaceInstUsesWith(TheAnd, Op);
- } else if (CI != AndRHS) {
- TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
- return &TheAnd;
- }
- } else { // Signed shr.
- // See if this is shifting in some sign extension, then masking it out
- // with an and.
- if (Op->hasOneUse()) {
- Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
- Constant *ShrMask = ConstantExpr::getUShr(AllOne, OpRHS);
- Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
- if (CI == AndRHS) { // Masking out bits shifted in.
- // Make the argument unsigned.
- Value *ShVal = Op->getOperand(0);
- ShVal = InsertCastBefore(ShVal,
- ShVal->getType()->getUnsignedVersion(),
- TheAnd);
- ShVal = InsertNewInstBefore(new ShiftInst(Instruction::Shr, ShVal,
- OpRHS, Op->getName()),
- TheAnd);
- Value *AndRHS2 = ConstantExpr::getCast(AndRHS, ShVal->getType());
- ShVal = InsertNewInstBefore(BinaryOperator::createAnd(ShVal, AndRHS2,
- TheAnd.getName()),
- TheAnd);
- return new CastInst(ShVal, Op->getType());
- }
+ Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
+ Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
+ if (C == AndRHS) { // Masking out bits shifted in.
+ // (Val ashr C1) & C2 -> (Val lshr C1) & C2
+ // Make the argument unsigned.
+ Value *ShVal = Op->getOperand(0);
+ ShVal = InsertNewInstBefore(new ShiftInst(Instruction::LShr, ShVal,
+ OpRHS, Op->getName()), TheAnd);
+ return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
}
}
break;
/// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
/// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
-/// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. IB is the location to
+/// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
+/// whether to treat the V, Lo and HI as signed or not. IB is the location to
/// insert new instructions.
Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
- bool Inside, Instruction &IB) {
- assert(cast<ConstantBool>(ConstantExpr::getSetLE(Lo, Hi))->getValue() &&
+ bool isSigned, bool Inside,
+ Instruction &IB) {
+ assert(cast<ConstantBool>(ConstantExpr::getICmp((isSigned ?
+ ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getValue() &&
"Lo is not <= Hi in range emission code!");
+
if (Inside) {
if (Lo == Hi) // Trivially false.
- return new SetCondInst(Instruction::SetNE, V, V);
- if (cast<ConstantIntegral>(Lo)->isMinValue())
- return new SetCondInst(Instruction::SetLT, V, Hi);
+ return new ICmpInst(ICmpInst::ICMP_NE, V, V);
+
+ // V >= Min && V < Hi --> V < Hi
+ if (cast<ConstantIntegral>(Lo)->isMinValue(isSigned)) {
+ ICmpInst::Predicate pred = (isSigned ?
+ ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
+ return new ICmpInst(pred, V, Hi);
+ }
- Constant *AddCST = ConstantExpr::getNeg(Lo);
- Instruction *Add = BinaryOperator::createAdd(V, AddCST,V->getName()+".off");
+ // Emit V-Lo <u Hi-Lo
+ Constant *NegLo = ConstantExpr::getNeg(Lo);
+ Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
InsertNewInstBefore(Add, IB);
- // Convert to unsigned for the comparison.
- const Type *UnsType = Add->getType()->getUnsignedVersion();
- Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
- AddCST = ConstantExpr::getAdd(AddCST, Hi);
- AddCST = ConstantExpr::getCast(AddCST, UnsType);
- return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
+ Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
+ return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
}
if (Lo == Hi) // Trivially true.
- return new SetCondInst(Instruction::SetEQ, V, V);
+ return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
+ // V < Min || V >= Hi ->'V > Hi-1'
Hi = SubOne(cast<ConstantInt>(Hi));
- if (cast<ConstantIntegral>(Lo)->isMinValue()) // V < 0 || V >= Hi ->'V > Hi-1'
- return new SetCondInst(Instruction::SetGT, V, Hi);
+ if (cast<ConstantIntegral>(Lo)->isMinValue(isSigned)) {
+ ICmpInst::Predicate pred = (isSigned ?
+ ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
+ return new ICmpInst(pred, V, Hi);
+ }
- // Emit X-Lo > Hi-Lo-1
- Constant *AddCST = ConstantExpr::getNeg(Lo);
- Instruction *Add = BinaryOperator::createAdd(V, AddCST, V->getName()+".off");
+ // Emit V-Lo > Hi-1-Lo
+ Constant *NegLo = ConstantExpr::getNeg(Lo);
+ Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
InsertNewInstBefore(Add, IB);
- // Convert to unsigned for the comparison.
- const Type *UnsType = Add->getType()->getUnsignedVersion();
- Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
- AddCST = ConstantExpr::getAdd(AddCST, Hi);
- AddCST = ConstantExpr::getCast(AddCST, UnsType);
- return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
+ Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
+ return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
}
// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
// MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
// not, since all 1s are not contiguous.
static bool isRunOfOnes(ConstantIntegral *Val, unsigned &MB, unsigned &ME) {
- uint64_t V = Val->getRawValue();
+ uint64_t V = Val->getZExtValue();
if (!isShiftedMask_64(V)) return false;
// look for the first zero bit after the run of ones
case Instruction::And:
if (ConstantExpr::getAnd(N, Mask) == Mask) {
// If the AndRHS is a power of two minus one (0+1+), this is simple.
- if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0)
+ if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
break;
// Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
case Instruction::Or:
case Instruction::Xor:
// If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
- if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0 &&
+ if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
ConstantExpr::getAnd(N, Mask)->isNullValue())
break;
return 0;
if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
return Res;
} else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
- const Type *SrcTy = CI->getOperand(0)->getType();
-
// If this is an integer truncation or change from signed-to-unsigned, and
// if the source is an and/or with immediate, transform it. This
// frequently occurs for bitfield accesses.
if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
- if (SrcTy->getPrimitiveSizeInBits() >=
- I.getType()->getPrimitiveSizeInBits() &&
+ if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
CastOp->getNumOperands() == 2)
if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
if (CastOp->getOpcode() == Instruction::And) {
// Change: and (cast (and X, C1) to T), C2
- // into : and (cast X to T), trunc(C1)&C2
- // This will folds the two ands together, which may allow other
- // simplifications.
- Instruction *NewCast =
- new CastInst(CastOp->getOperand(0), I.getType(),
- CastOp->getName()+".shrunk");
+ // into : and (cast X to T), trunc_or_bitcast(C1)&C2
+ // This will fold the two constants together, which may allow
+ // other simplifications.
+ Instruction *NewCast = CastInst::createTruncOrBitCast(
+ CastOp->getOperand(0), I.getType(),
+ CastOp->getName()+".shrunk");
NewCast = InsertNewInstBefore(NewCast, I);
-
- Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
- C3 = ConstantExpr::getAnd(C3, AndRHS); // trunc(C1)&C2
+ // trunc_or_bitcast(C1)&C2
+ Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
+ C3 = ConstantExpr::getAnd(C3, AndRHS);
return BinaryOperator::createAnd(NewCast, C3);
} else if (CastOp->getOpcode() == Instruction::Or) {
// Change: and (cast (or X, C1) to T), C2
// into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
- Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
+ Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
return ReplaceInstUsesWith(I, AndRHS);
}
{
Value *A = 0, *B = 0;
- ConstantInt *C1 = 0, *C2 = 0;
if (match(Op0, m_Or(m_Value(A), m_Value(B))))
if (A == Op1 || B == Op1) // (A | ?) & A --> A
return ReplaceInstUsesWith(I, Op1);
}
}
-
- if (SetCondInst *RHS = dyn_cast<SetCondInst>(Op1)) {
- // (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
- if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
+ if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
+ // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
+ if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
return R;
Value *LHSVal, *RHSVal;
ConstantInt *LHSCst, *RHSCst;
- Instruction::BinaryOps LHSCC, RHSCC;
- if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
- if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
- if (LHSVal == RHSVal && // Found (X setcc C1) & (X setcc C2)
- // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
- LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
- RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
+ ICmpInst::Predicate LHSCC, RHSCC;
+ if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
+ if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
+ if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
+ // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
+ LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
+ RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
+ LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
+ RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
// Ensure that the larger constant is on the RHS.
- Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
- SetCondInst *LHS = cast<SetCondInst>(Op0);
+ ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
+ ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
+ Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
+ ICmpInst *LHS = cast<ICmpInst>(Op0);
if (cast<ConstantBool>(Cmp)->getValue()) {
std::swap(LHS, RHS);
std::swap(LHSCst, RHSCst);
std::swap(LHSCC, RHSCC);
}
- // At this point, we know we have have two setcc instructions
+ // At this point, we know we have have two icmp instructions
// comparing a value against two constants and and'ing the result
// together. Because of the above check, we know that we only have
- // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
- // FoldSetCCLogical check above), that the two constants are not
- // equal.
+ // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
+ // (from the FoldICmpLogical check above), that the two constants
+ // are not equal and that the larger constant is on the RHS
assert(LHSCst != RHSCst && "Compares not folded above?");
switch (LHSCC) {
default: assert(0 && "Unknown integer condition code!");
- case Instruction::SetEQ:
+ case ICmpInst::ICMP_EQ:
switch (RHSCC) {
default: assert(0 && "Unknown integer condition code!");
- case Instruction::SetEQ: // (X == 13 & X == 15) -> false
- case Instruction::SetGT: // (X == 13 & X > 15) -> false
- return ReplaceInstUsesWith(I, ConstantBool::False);
- case Instruction::SetNE: // (X == 13 & X != 15) -> X == 13
- case Instruction::SetLT: // (X == 13 & X < 15) -> X == 13
+ case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
+ case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
+ case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
+ case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
+ case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
return ReplaceInstUsesWith(I, LHS);
}
- case Instruction::SetNE:
+ case ICmpInst::ICMP_NE:
switch (RHSCC) {
default: assert(0 && "Unknown integer condition code!");
- case Instruction::SetLT:
- if (LHSCst == SubOne(RHSCst)) // (X != 13 & X < 14) -> X < 13
- return new SetCondInst(Instruction::SetLT, LHSVal, LHSCst);
- break; // (X != 13 & X < 15) -> no change
- case Instruction::SetEQ: // (X != 13 & X == 15) -> X == 15
- case Instruction::SetGT: // (X != 13 & X > 15) -> X > 15
+ case ICmpInst::ICMP_ULT:
+ if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
+ return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
+ break; // (X != 13 & X u< 15) -> no change
+ case ICmpInst::ICMP_SLT:
+ if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
+ return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
+ break; // (X != 13 & X s< 15) -> no change
+ case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
+ case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
+ case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
return ReplaceInstUsesWith(I, RHS);
- case Instruction::SetNE:
- if (LHSCst == SubOne(RHSCst)) {// (X != 13 & X != 14) -> X-13 >u 1
+ case ICmpInst::ICMP_NE:
+ if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
Constant *AddCST = ConstantExpr::getNeg(LHSCst);
Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
LHSVal->getName()+".off");
InsertNewInstBefore(Add, I);
const Type *UnsType = Add->getType()->getUnsignedVersion();
- Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
+ Value *OffsetVal = InsertCastBefore(Instruction::BitCast, Add,
+ UnsType, I);
AddCST = ConstantExpr::getSub(RHSCst, LHSCst);
- AddCST = ConstantExpr::getCast(AddCST, UnsType);
- return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
+ AddCST = ConstantExpr::getBitCast(AddCST, UnsType);
+ return new ICmpInst(ICmpInst::ICMP_UGT, OffsetVal, AddCST);
}
break; // (X != 13 & X != 15) -> no change
}
break;
- case Instruction::SetLT:
+ case ICmpInst::ICMP_ULT:
+ switch (RHSCC) {
+ default: assert(0 && "Unknown integer condition code!");
+ case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
+ case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
+ break;
+ case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
+ case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
+ return ReplaceInstUsesWith(I, LHS);
+ case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
+ break;
+ }
+ break;
+ case ICmpInst::ICMP_SLT:
switch (RHSCC) {
default: assert(0 && "Unknown integer condition code!");
- case Instruction::SetEQ: // (X < 13 & X == 15) -> false
- case Instruction::SetGT: // (X < 13 & X > 15) -> false
- return ReplaceInstUsesWith(I, ConstantBool::False);
- case Instruction::SetNE: // (X < 13 & X != 15) -> X < 13
- case Instruction::SetLT: // (X < 13 & X < 15) -> X < 13
+ case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
+ case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
+ break;
+ case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
+ case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
+ return ReplaceInstUsesWith(I, LHS);
+ case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
+ break;
+ }
+ break;
+ case ICmpInst::ICMP_UGT:
+ switch (RHSCC) {
+ default: assert(0 && "Unknown integer condition code!");
+ case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
return ReplaceInstUsesWith(I, LHS);
+ case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
+ return ReplaceInstUsesWith(I, RHS);
+ case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
+ break;
+ case ICmpInst::ICMP_NE:
+ if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
+ return new ICmpInst(LHSCC, LHSVal, RHSCst);
+ break; // (X u> 13 & X != 15) -> no change
+ case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
+ return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
+ true, I);
+ case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
+ break;
}
- case Instruction::SetGT:
+ break;
+ case ICmpInst::ICMP_SGT:
switch (RHSCC) {
default: assert(0 && "Unknown integer condition code!");
- case Instruction::SetEQ: // (X > 13 & X == 15) -> X > 13
+ case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
return ReplaceInstUsesWith(I, LHS);
- case Instruction::SetGT: // (X > 13 & X > 15) -> X > 15
+ case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
return ReplaceInstUsesWith(I, RHS);
- case Instruction::SetNE:
- if (RHSCst == AddOne(LHSCst)) // (X > 13 & X != 14) -> X > 14
- return new SetCondInst(Instruction::SetGT, LHSVal, RHSCst);
- break; // (X > 13 & X != 15) -> no change
- case Instruction::SetLT: // (X > 13 & X < 15) -> (X-14) <u 1
- return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true, I);
+ case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
+ break;
+ case ICmpInst::ICMP_NE:
+ if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
+ return new ICmpInst(LHSCC, LHSVal, RHSCst);
+ break; // (X s> 13 & X != 15) -> no change
+ case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
+ return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
+ true, I);
+ case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
+ break;
}
+ break;
}
}
}
// fold (and (cast A), (cast B)) -> (cast (and A, B))
- if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
- const Type *SrcTy = Op0C->getOperand(0)->getType();
+ if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
- if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
- // Only do this if the casts both really cause code to be generated.
- ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
- ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
- Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
- Op1C->getOperand(0),
- I.getName());
- InsertNewInstBefore(NewOp, I);
- return new CastInst(NewOp, I.getType());
+ if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
+ const Type *SrcTy = Op0C->getOperand(0)->getType();
+ if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
+ // Only do this if the casts both really cause code to be generated.
+ ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
+ I.getType(), TD) &&
+ ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
+ I.getType(), TD)) {
+ Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
+ Op1C->getOperand(0),
+ I.getName());
+ InsertNewInstBefore(NewOp, I);
+ return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
+ }
+ }
+
+ // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
+ if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
+ if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
+ if (SI0->getOpcode() == SI1->getOpcode() &&
+ SI0->getOperand(1) == SI1->getOperand(1) &&
+ (SI0->hasOneUse() || SI1->hasOneUse())) {
+ Instruction *NewOp =
+ InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
+ SI1->getOperand(0),
+ SI0->getName()), I);
+ return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
}
}
return Changed ? &I : 0;
}
+/// CollectBSwapParts - Look to see if the specified value defines a single byte
+/// in the result. If it does, and if the specified byte hasn't been filled in
+/// yet, fill it in and return false.
+static bool CollectBSwapParts(Value *V, std::vector<Value*> &ByteValues) {
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (I == 0) return true;
+
+ // If this is an or instruction, it is an inner node of the bswap.
+ if (I->getOpcode() == Instruction::Or)
+ return CollectBSwapParts(I->getOperand(0), ByteValues) ||
+ CollectBSwapParts(I->getOperand(1), ByteValues);
+
+ // If this is a shift by a constant int, and it is "24", then its operand
+ // defines a byte. We only handle unsigned types here.
+ if (isa<ShiftInst>(I) && isa<ConstantInt>(I->getOperand(1))) {
+ // Not shifting the entire input by N-1 bytes?
+ if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
+ 8*(ByteValues.size()-1))
+ return true;
+
+ unsigned DestNo;
+ if (I->getOpcode() == Instruction::Shl) {
+ // X << 24 defines the top byte with the lowest of the input bytes.
+ DestNo = ByteValues.size()-1;
+ } else {
+ // X >>u 24 defines the low byte with the highest of the input bytes.
+ DestNo = 0;
+ }
+
+ // If the destination byte value is already defined, the values are or'd
+ // together, which isn't a bswap (unless it's an or of the same bits).
+ if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
+ return true;
+ ByteValues[DestNo] = I->getOperand(0);
+ return false;
+ }
+
+ // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
+ // don't have this.
+ Value *Shift = 0, *ShiftLHS = 0;
+ ConstantInt *AndAmt = 0, *ShiftAmt = 0;
+ if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
+ !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
+ return true;
+ Instruction *SI = cast<Instruction>(Shift);
+
+ // Make sure that the shift amount is by a multiple of 8 and isn't too big.
+ if (ShiftAmt->getZExtValue() & 7 ||
+ ShiftAmt->getZExtValue() > 8*ByteValues.size())
+ return true;
+
+ // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
+ unsigned DestByte;
+ for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
+ if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
+ break;
+ // Unknown mask for bswap.
+ if (DestByte == ByteValues.size()) return true;
+
+ unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
+ unsigned SrcByte;
+ if (SI->getOpcode() == Instruction::Shl)
+ SrcByte = DestByte - ShiftBytes;
+ else
+ SrcByte = DestByte + ShiftBytes;
+
+ // If the SrcByte isn't a bswapped value from the DestByte, reject it.
+ if (SrcByte != ByteValues.size()-DestByte-1)
+ return true;
+
+ // If the destination byte value is already defined, the values are or'd
+ // together, which isn't a bswap (unless it's an or of the same bits).
+ if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
+ return true;
+ ByteValues[DestByte] = SI->getOperand(0);
+ return false;
+}
+
+/// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
+/// If so, insert the new bswap intrinsic and return it.
+Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
+ // We can only handle bswap of unsigned integers, and cannot bswap one byte.
+ if (!I.getType()->isUnsigned() || I.getType() == Type::UByteTy)
+ return 0;
+
+ /// ByteValues - For each byte of the result, we keep track of which value
+ /// defines each byte.
+ std::vector<Value*> ByteValues;
+ ByteValues.resize(I.getType()->getPrimitiveSize());
+
+ // Try to find all the pieces corresponding to the bswap.
+ if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
+ CollectBSwapParts(I.getOperand(1), ByteValues))
+ return 0;
+
+ // Check to see if all of the bytes come from the same value.
+ Value *V = ByteValues[0];
+ if (V == 0) return 0; // Didn't find a byte? Must be zero.
+
+ // Check to make sure that all of the bytes come from the same value.
+ for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
+ if (ByteValues[i] != V)
+ return 0;
+
+ // If they do then *success* we can turn this into a bswap. Figure out what
+ // bswap to make it into.
+ Module *M = I.getParent()->getParent()->getParent();
+ const char *FnName = 0;
+ if (I.getType() == Type::UShortTy)
+ FnName = "llvm.bswap.i16";
+ else if (I.getType() == Type::UIntTy)
+ FnName = "llvm.bswap.i32";
+ else if (I.getType() == Type::ULongTy)
+ FnName = "llvm.bswap.i64";
+ else
+ assert(0 && "Unknown integer type!");
+ Function *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
+
+ return new CallInst(F, V);
+}
+
+
Instruction *InstCombiner::visitOr(BinaryOperator &I) {
bool Changed = SimplifyCommutative(I);
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
if (A == Op0 || B == Op0) // A | (A & ?) --> A
return ReplaceInstUsesWith(I, Op0);
+ // (A | B) | C and A | (B | C) -> bswap if possible.
+ // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
+ if (match(Op0, m_Or(m_Value(), m_Value())) ||
+ match(Op1, m_Or(m_Value(), m_Value())) ||
+ (match(Op0, m_Shift(m_Value(), m_Value())) &&
+ match(Op1, m_Shift(m_Value(), m_Value())))) {
+ if (Instruction *BSwap = MatchBSwap(I))
+ return BSwap;
+ }
+
// (X^C)|Y -> (X|Y)^C iff Y&C == 0
if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
MaskedValueIsZero(Op1, C1->getZExtValue())) {
// replace with V+N.
if (C1 == ConstantExpr::getNot(C2)) {
Value *V1 = 0, *V2 = 0;
- if ((C2->getRawValue() & (C2->getRawValue()+1)) == 0 && // C2 == 0+1+
+ if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
match(A, m_Add(m_Value(V1), m_Value(V2)))) {
// Add commutes, try both ways.
if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
return ReplaceInstUsesWith(I, A);
}
// Or commutes, try both ways.
- if ((C1->getRawValue() & (C1->getRawValue()+1)) == 0 &&
+ if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
match(B, m_Add(m_Value(V1), m_Value(V2)))) {
// Add commutes, try both ways.
if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
}
}
}
+
+ // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
+ if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
+ if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
+ if (SI0->getOpcode() == SI1->getOpcode() &&
+ SI0->getOperand(1) == SI1->getOperand(1) &&
+ (SI0->hasOneUse() || SI1->hasOneUse())) {
+ Instruction *NewOp =
+ InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
+ SI1->getOperand(0),
+ SI0->getName()), I);
+ return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
+ }
+ }
if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
if (A == Op1) // ~A | A == -1
}
}
- // (setcc1 A, B) | (setcc2 A, B) --> (setcc3 A, B)
- if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1))) {
- if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
+ // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
+ if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
+ if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
return R;
Value *LHSVal, *RHSVal;
ConstantInt *LHSCst, *RHSCst;
- Instruction::BinaryOps LHSCC, RHSCC;
- if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
- if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
- if (LHSVal == RHSVal && // Found (X setcc C1) | (X setcc C2)
- // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
- LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
- RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
+ ICmpInst::Predicate LHSCC, RHSCC;
+ if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
+ if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
+ if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
+ // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
+ LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
+ RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
+ LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
+ RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
// Ensure that the larger constant is on the RHS.
- Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
- SetCondInst *LHS = cast<SetCondInst>(Op0);
+ ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
+ ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
+ Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
+ ICmpInst *LHS = cast<ICmpInst>(Op0);
if (cast<ConstantBool>(Cmp)->getValue()) {
std::swap(LHS, RHS);
std::swap(LHSCst, RHSCst);
std::swap(LHSCC, RHSCC);
}
- // At this point, we know we have have two setcc instructions
+ // At this point, we know we have have two icmp instructions
// comparing a value against two constants and or'ing the result
// together. Because of the above check, we know that we only have
- // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
- // FoldSetCCLogical check above), that the two constants are not
+ // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
+ // FoldICmpLogical check above), that the two constants are not
// equal.
assert(LHSCst != RHSCst && "Compares not folded above?");
switch (LHSCC) {
default: assert(0 && "Unknown integer condition code!");
- case Instruction::SetEQ:
+ case ICmpInst::ICMP_EQ:
switch (RHSCC) {
default: assert(0 && "Unknown integer condition code!");
- case Instruction::SetEQ:
+ case ICmpInst::ICMP_EQ:
if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
Constant *AddCST = ConstantExpr::getNeg(LHSCst);
Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
LHSVal->getName()+".off");
InsertNewInstBefore(Add, I);
- const Type *UnsType = Add->getType()->getUnsignedVersion();
- Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
- AddCST = ConstantExpr::getCast(AddCST, UnsType);
- return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
+ return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
}
- break; // (X == 13 | X == 15) -> no change
-
- case Instruction::SetGT: // (X == 13 | X > 14) -> no change
+ break; // (X == 13 | X == 15) -> no change
+ case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
+ case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
break;
- case Instruction::SetNE: // (X == 13 | X != 15) -> X != 15
- case Instruction::SetLT: // (X == 13 | X < 15) -> X < 15
+ case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
+ case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
+ case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
return ReplaceInstUsesWith(I, RHS);
}
break;
- case Instruction::SetNE:
+ case ICmpInst::ICMP_NE:
switch (RHSCC) {
default: assert(0 && "Unknown integer condition code!");
- case Instruction::SetEQ: // (X != 13 | X == 15) -> X != 13
- case Instruction::SetGT: // (X != 13 | X > 15) -> X != 13
+ case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
+ case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
+ case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
return ReplaceInstUsesWith(I, LHS);
- case Instruction::SetNE: // (X != 13 | X != 15) -> true
- case Instruction::SetLT: // (X != 13 | X < 15) -> true
- return ReplaceInstUsesWith(I, ConstantBool::True);
+ case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
+ case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
+ case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
+ }
+ break;
+ case ICmpInst::ICMP_ULT:
+ switch (RHSCC) {
+ default: assert(0 && "Unknown integer condition code!");
+ case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
+ break;
+ case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
+ return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
+ false, I);
+ case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
+ break;
+ case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
+ case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
+ return ReplaceInstUsesWith(I, RHS);
+ case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
+ break;
}
break;
- case Instruction::SetLT:
+ case ICmpInst::ICMP_SLT:
switch (RHSCC) {
default: assert(0 && "Unknown integer condition code!");
- case Instruction::SetEQ: // (X < 13 | X == 14) -> no change
+ case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
+ break;
+ case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
+ return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
+ false, I);
+ case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
break;
- case Instruction::SetGT: // (X < 13 | X > 15) -> (X-13) > 2
- return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false, I);
- case Instruction::SetNE: // (X < 13 | X != 15) -> X != 15
- case Instruction::SetLT: // (X < 13 | X < 15) -> X < 15
+ case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
+ case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
return ReplaceInstUsesWith(I, RHS);
+ case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
+ break;
}
break;
- case Instruction::SetGT:
+ case ICmpInst::ICMP_UGT:
switch (RHSCC) {
default: assert(0 && "Unknown integer condition code!");
- case Instruction::SetEQ: // (X > 13 | X == 15) -> X > 13
- case Instruction::SetGT: // (X > 13 | X > 15) -> X > 13
+ case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
+ case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
return ReplaceInstUsesWith(I, LHS);
- case Instruction::SetNE: // (X > 13 | X != 15) -> true
- case Instruction::SetLT: // (X > 13 | X < 15) -> true
- return ReplaceInstUsesWith(I, ConstantBool::True);
+ case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
+ break;
+ case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
+ case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
+ case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
+ break;
}
- }
- }
- }
-
- // fold (or (cast A), (cast B)) -> (cast (or A, B))
- if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
- const Type *SrcTy = Op0C->getOperand(0)->getType();
+ break;
+ case ICmpInst::ICMP_SGT:
+ switch (RHSCC) {
+ default: assert(0 && "Unknown integer condition code!");
+ case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
+ case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
+ return ReplaceInstUsesWith(I, LHS);
+ case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
+ break;
+ case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
+ case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
+ case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
+ break;
+ }
+ break;
+ }
+ }
+ }
+
+ // fold (or (cast A), (cast B)) -> (cast (or A, B))
+ if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
- if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
- // Only do this if the casts both really cause code to be generated.
- ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
- ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
- Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
- Op1C->getOperand(0),
- I.getName());
- InsertNewInstBefore(NewOp, I);
- return new CastInst(NewOp, I.getType());
+ if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
+ const Type *SrcTy = Op0C->getOperand(0)->getType();
+ if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
+ // Only do this if the casts both really cause code to be generated.
+ ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
+ I.getType(), TD) &&
+ ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
+ I.getType(), TD)) {
+ Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
+ Op1C->getOperand(0),
+ I.getName());
+ InsertNewInstBefore(NewOp, I);
+ return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
+ }
}
- }
return Changed ? &I : 0;
return &I;
if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
- if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
- // xor (setcc A, B), true = not (setcc A, B) = setncc A, B
- if (SetCondInst *SCI = dyn_cast<SetCondInst>(Op0I))
- if (RHS == ConstantBool::True && SCI->hasOneUse())
- return new SetCondInst(SCI->getInverseCondition(),
- SCI->getOperand(0), SCI->getOperand(1));
+ // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
+ if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
+ if (RHS == ConstantBool::getTrue() && ICI->hasOneUse())
+ return new ICmpInst(ICI->getInversePredicate(),
+ ICI->getOperand(0), ICI->getOperand(1));
+ if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
// ~(c-X) == X-c-1 == X+(-c-1)
if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
}
}
- // (setcc1 A, B) ^ (setcc2 A, B) --> (setcc3 A, B)
- if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1)))
- if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
+ // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
+ if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
+ if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
return R;
// fold (xor (cast A), (cast B)) -> (cast (xor A, B))
- if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
- const Type *SrcTy = Op0C->getOperand(0)->getType();
+ if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
- if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
- // Only do this if the casts both really cause code to be generated.
- ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
- ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
- Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
- Op1C->getOperand(0),
- I.getName());
- InsertNewInstBefore(NewOp, I);
- return new CastInst(NewOp, I.getType());
+ if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
+ const Type *SrcTy = Op0C->getOperand(0)->getType();
+ if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
+ // Only do this if the casts both really cause code to be generated.
+ ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
+ I.getType(), TD) &&
+ ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
+ I.getType(), TD)) {
+ Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
+ Op1C->getOperand(0),
+ I.getName());
+ InsertNewInstBefore(NewOp, I);
+ return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
+ }
+ }
+
+ // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
+ if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
+ if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
+ if (SI0->getOpcode() == SI1->getOpcode() &&
+ SI0->getOperand(1) == SI1->getOperand(1) &&
+ (SI0->hasOneUse() || SI1->hasOneUse())) {
+ Instruction *NewOp =
+ InsertNewInstBefore(BinaryOperator::createXor(SI0->getOperand(0),
+ SI1->getOperand(0),
+ SI0->getName()), I);
+ return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
}
}
return Changed ? &I : 0;
}
-/// MulWithOverflow - Compute Result = In1*In2, returning true if the result
-/// overflowed for this type.
-static bool MulWithOverflow(ConstantInt *&Result, ConstantInt *In1,
- ConstantInt *In2) {
- Result = cast<ConstantInt>(ConstantExpr::getMul(In1, In2));
- return !In2->isNullValue() && ConstantExpr::getDiv(Result, In2) != In1;
-}
-
static bool isPositive(ConstantInt *C) {
- return cast<ConstantSInt>(C)->getValue() >= 0;
+ return C->getSExtValue() >= 0;
}
/// AddWithOverflow - Compute Result = In1+In2, returning true if the result
Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
if (In1->getType()->isUnsigned())
- return cast<ConstantUInt>(Result)->getValue() <
- cast<ConstantUInt>(In1)->getValue();
+ return cast<ConstantInt>(Result)->getZExtValue() <
+ cast<ConstantInt>(In1)->getZExtValue();
if (isPositive(In1) != isPositive(In2))
return false;
if (isPositive(In1))
- return cast<ConstantSInt>(Result)->getValue() <
- cast<ConstantSInt>(In1)->getValue();
- return cast<ConstantSInt>(Result)->getValue() >
- cast<ConstantSInt>(In1)->getValue();
+ return cast<ConstantInt>(Result)->getSExtValue() <
+ cast<ConstantInt>(In1)->getSExtValue();
+ return cast<ConstantInt>(Result)->getSExtValue() >
+ cast<ConstantInt>(In1)->getSExtValue();
}
/// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
TargetData &TD = IC.getTargetData();
gep_type_iterator GTI = gep_type_begin(GEP);
- const Type *UIntPtrTy = TD.getIntPtrType();
- const Type *SIntPtrTy = UIntPtrTy->getSignedVersion();
- Value *Result = Constant::getNullValue(SIntPtrTy);
+ const Type *IntPtrTy = TD.getIntPtrType();
+ Value *Result = Constant::getNullValue(IntPtrTy);
// Build a mask for high order bits.
uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
Value *Op = GEP->getOperand(i);
uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
- Constant *Scale = ConstantExpr::getCast(ConstantUInt::get(UIntPtrTy, Size),
- SIntPtrTy);
+ Constant *Scale = ConstantInt::get(IntPtrTy, Size);
if (Constant *OpC = dyn_cast<Constant>(Op)) {
if (!OpC->isNullValue()) {
- OpC = ConstantExpr::getCast(OpC, SIntPtrTy);
+ OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
Scale = ConstantExpr::getMul(OpC, Scale);
if (Constant *RC = dyn_cast<Constant>(Result))
Result = ConstantExpr::getAdd(RC, Scale);
}
} else {
// Convert to correct type.
- Op = IC.InsertNewInstBefore(new CastInst(Op, SIntPtrTy,
+ Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
Op->getName()+".c"), I);
if (Size != 1)
// We'll let instcombine(mul) convert this to a shl if possible.
return Result;
}
-/// FoldGEPSetCC - Fold comparisons between a GEP instruction and something
+/// FoldGEPICmp - Fold comparisons between a GEP instruction and something
/// else. At this point we know that the GEP is on the LHS of the comparison.
-Instruction *InstCombiner::FoldGEPSetCC(User *GEPLHS, Value *RHS,
- Instruction::BinaryOps Cond,
- Instruction &I) {
+Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
+ ICmpInst::Predicate Cond,
+ Instruction &I) {
assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
if (CastInst *CI = dyn_cast<CastInst>(RHS))
Value *PtrBase = GEPLHS->getOperand(0);
if (PtrBase == RHS) {
// As an optimization, we don't actually have to compute the actual value of
- // OFFSET if this is a seteq or setne comparison, just return whether each
- // index is zero or not.
- if (Cond == Instruction::SetEQ || Cond == Instruction::SetNE) {
+ // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
+ // each index is zero or not.
+ if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
Instruction *InVal = 0;
gep_type_iterator GTI = gep_type_begin(GEPLHS);
for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
EmitIt = false; // This is indexing into a zero sized array?
} else if (isa<ConstantInt>(C))
return ReplaceInstUsesWith(I, // No comparison is needed here.
- ConstantBool::get(Cond == Instruction::SetNE));
+ ConstantBool::get(Cond == ICmpInst::ICMP_NE));
}
if (EmitIt) {
Instruction *Comp =
- new SetCondInst(Cond, GEPLHS->getOperand(i),
+ new ICmpInst(Cond, GEPLHS->getOperand(i),
Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
if (InVal == 0)
InVal = Comp;
else {
InVal = InsertNewInstBefore(InVal, I);
InsertNewInstBefore(Comp, I);
- if (Cond == Instruction::SetNE) // True if any are unequal
+ if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
InVal = BinaryOperator::createOr(InVal, Comp);
else // True if all are equal
InVal = BinaryOperator::createAnd(InVal, Comp);
if (InVal)
return InVal;
else
- ReplaceInstUsesWith(I, // No comparison is needed here, all indexes = 0
- ConstantBool::get(Cond == Instruction::SetEQ));
+ // No comparison is needed here, all indexes = 0
+ ReplaceInstUsesWith(I, ConstantBool::get(Cond == ICmpInst::ICMP_EQ));
}
- // Only lower this if the setcc is the only user of the GEP or if we expect
+ // Only lower this if the icmp is the only user of the GEP or if we expect
// the result to fold to a constant!
if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
// ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
- return new SetCondInst(Cond, Offset,
- Constant::getNullValue(Offset->getType()));
+ return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
+ Constant::getNullValue(Offset->getType()));
}
} else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
// If the base pointers are different, but the indices are the same, just
// If all indices are the same, just compare the base pointers.
if (IndicesTheSame)
- return new SetCondInst(Cond, GEPLHS->getOperand(0),
- GEPRHS->getOperand(0));
+ return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
+ GEPLHS->getOperand(0), GEPRHS->getOperand(0));
// Otherwise, the base pointers are different and the indices are
// different, bail out.
break;
}
if (AllZeros)
- return FoldGEPSetCC(GEPRHS, GEPLHS->getOperand(0),
- SetCondInst::getSwappedCondition(Cond), I);
+ return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
+ ICmpInst::getSwappedPredicate(Cond), I);
// If the other GEP has all zero indices, recurse.
AllZeros = true;
break;
}
if (AllZeros)
- return FoldGEPSetCC(GEPLHS, GEPRHS->getOperand(0), Cond, I);
+ return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
// If the GEPs only differ by one index, compare it.
if (NumDifferences == 0) // SAME GEP?
return ReplaceInstUsesWith(I, // No comparison is needed here.
- ConstantBool::get(Cond == Instruction::SetEQ));
+ ConstantBool::get(Cond == ICmpInst::ICMP_EQ));
else if (NumDifferences == 1) {
Value *LHSV = GEPLHS->getOperand(DiffOperand);
Value *RHSV = GEPRHS->getOperand(DiffOperand);
-
- // Convert the operands to signed values to make sure to perform a
- // signed comparison.
- const Type *NewTy = LHSV->getType()->getSignedVersion();
- if (LHSV->getType() != NewTy)
- LHSV = InsertNewInstBefore(new CastInst(LHSV, NewTy,
- LHSV->getName()), I);
- if (RHSV->getType() != NewTy)
- RHSV = InsertNewInstBefore(new CastInst(RHSV, NewTy,
- RHSV->getName()), I);
- return new SetCondInst(Cond, LHSV, RHSV);
+ if (LHSV->getType() != RHSV->getType())
+ // Doesn't matter which one we bitconvert here.
+ LHSV = InsertCastBefore(Instruction::BitCast, LHSV, RHSV->getType(),
+ I);
+ // Make sure we do a signed comparison here.
+ return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
}
}
- // Only lower this if the setcc is the only user of the GEP or if we expect
+ // Only lower this if the icmp is the only user of the GEP or if we expect
// the result to fold to a constant!
if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
(isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
// ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
Value *L = EmitGEPOffset(GEPLHS, I, *this);
Value *R = EmitGEPOffset(GEPRHS, I, *this);
- return new SetCondInst(Cond, L, R);
+ return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
}
}
return 0;
}
+Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
+ bool Changed = SimplifyCompare(I);
+ Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
-Instruction *InstCombiner::visitSetCondInst(SetCondInst &I) {
- bool Changed = SimplifyCommutative(I);
+ // fcmp pred X, X
+ if (Op0 == Op1)
+ return ReplaceInstUsesWith(I, ConstantBool::get(isTrueWhenEqual(I)));
+
+ if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
+ return ReplaceInstUsesWith(I, UndefValue::get(Type::BoolTy));
+
+ // Handle fcmp with constant RHS
+ if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
+ if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
+ switch (LHSI->getOpcode()) {
+ case Instruction::PHI:
+ if (Instruction *NV = FoldOpIntoPhi(I))
+ return NV;
+ break;
+ case Instruction::Select:
+ // If either operand of the select is a constant, we can fold the
+ // comparison into the select arms, which will cause one to be
+ // constant folded and the select turned into a bitwise or.
+ Value *Op1 = 0, *Op2 = 0;
+ if (LHSI->hasOneUse()) {
+ if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
+ // Fold the known value into the constant operand.
+ Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
+ // Insert a new FCmp of the other select operand.
+ Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
+ LHSI->getOperand(2), RHSC,
+ I.getName()), I);
+ } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
+ // Fold the known value into the constant operand.
+ Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
+ // Insert a new FCmp of the other select operand.
+ Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
+ LHSI->getOperand(1), RHSC,
+ I.getName()), I);
+ }
+ }
+
+ if (Op1)
+ return new SelectInst(LHSI->getOperand(0), Op1, Op2);
+ break;
+ }
+ }
+
+ return Changed ? &I : 0;
+}
+
+Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
+ bool Changed = SimplifyCompare(I);
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
const Type *Ty = Op0->getType();
- // setcc X, X
+ // icmp X, X
if (Op0 == Op1)
return ReplaceInstUsesWith(I, ConstantBool::get(isTrueWhenEqual(I)));
- if (isa<UndefValue>(Op1)) // X setcc undef -> undef
+ if (isa<UndefValue>(Op1)) // X icmp undef -> undef
return ReplaceInstUsesWith(I, UndefValue::get(Type::BoolTy));
- // setcc <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
+ // icmp of GlobalValues can never equal each other as long as they aren't
+ // external weak linkage type.
+ if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
+ if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
+ if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
+ return ReplaceInstUsesWith(I, ConstantBool::get(!isTrueWhenEqual(I)));
+
+ // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
// addresses never equal each other! We already know that Op0 != Op1.
if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
isa<ConstantPointerNull>(Op0)) &&
isa<ConstantPointerNull>(Op1)))
return ReplaceInstUsesWith(I, ConstantBool::get(!isTrueWhenEqual(I)));
- // setcc's with boolean values can always be turned into bitwise operations
+ // icmp's with boolean values can always be turned into bitwise operations
if (Ty == Type::BoolTy) {
- switch (I.getOpcode()) {
- default: assert(0 && "Invalid setcc instruction!");
- case Instruction::SetEQ: { // seteq bool %A, %B -> ~(A^B)
+ switch (I.getPredicate()) {
+ default: assert(0 && "Invalid icmp instruction!");
+ case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
InsertNewInstBefore(Xor, I);
return BinaryOperator::createNot(Xor);
}
- case Instruction::SetNE:
+ case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
return BinaryOperator::createXor(Op0, Op1);
- case Instruction::SetGT:
- std::swap(Op0, Op1); // Change setgt -> setlt
+ case ICmpInst::ICMP_UGT:
+ case ICmpInst::ICMP_SGT:
+ std::swap(Op0, Op1); // Change icmp gt -> icmp lt
// FALL THROUGH
- case Instruction::SetLT: { // setlt bool A, B -> ~X & Y
+ case ICmpInst::ICMP_ULT:
+ case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
InsertNewInstBefore(Not, I);
return BinaryOperator::createAnd(Not, Op1);
}
- case Instruction::SetGE:
- std::swap(Op0, Op1); // Change setge -> setle
+ case ICmpInst::ICMP_UGE:
+ case ICmpInst::ICMP_SGE:
+ std::swap(Op0, Op1); // Change icmp ge -> icmp le
// FALL THROUGH
- case Instruction::SetLE: { // setle bool %A, %B -> ~A | B
+ case ICmpInst::ICMP_ULE:
+ case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
InsertNewInstBefore(Not, I);
return BinaryOperator::createOr(Not, Op1);
// See if we are doing a comparison between a constant and an instruction that
// can be folded into the comparison.
if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
- // Check to see if we are comparing against the minimum or maximum value...
- if (CI->isMinValue()) {
- if (I.getOpcode() == Instruction::SetLT) // A < MIN -> FALSE
- return ReplaceInstUsesWith(I, ConstantBool::False);
- if (I.getOpcode() == Instruction::SetGE) // A >= MIN -> TRUE
- return ReplaceInstUsesWith(I, ConstantBool::True);
- if (I.getOpcode() == Instruction::SetLE) // A <= MIN -> A == MIN
- return BinaryOperator::createSetEQ(Op0, Op1);
- if (I.getOpcode() == Instruction::SetGT) // A > MIN -> A != MIN
- return BinaryOperator::createSetNE(Op0, Op1);
-
- } else if (CI->isMaxValue()) {
- if (I.getOpcode() == Instruction::SetGT) // A > MAX -> FALSE
- return ReplaceInstUsesWith(I, ConstantBool::False);
- if (I.getOpcode() == Instruction::SetLE) // A <= MAX -> TRUE
- return ReplaceInstUsesWith(I, ConstantBool::True);
- if (I.getOpcode() == Instruction::SetGE) // A >= MAX -> A == MAX
- return BinaryOperator::createSetEQ(Op0, Op1);
- if (I.getOpcode() == Instruction::SetLT) // A < MAX -> A != MAX
- return BinaryOperator::createSetNE(Op0, Op1);
-
- // Comparing against a value really close to min or max?
- } else if (isMinValuePlusOne(CI)) {
- if (I.getOpcode() == Instruction::SetLT) // A < MIN+1 -> A == MIN
- return BinaryOperator::createSetEQ(Op0, SubOne(CI));
- if (I.getOpcode() == Instruction::SetGE) // A >= MIN-1 -> A != MIN
- return BinaryOperator::createSetNE(Op0, SubOne(CI));
-
- } else if (isMaxValueMinusOne(CI)) {
- if (I.getOpcode() == Instruction::SetGT) // A > MAX-1 -> A == MAX
- return BinaryOperator::createSetEQ(Op0, AddOne(CI));
- if (I.getOpcode() == Instruction::SetLE) // A <= MAX-1 -> A != MAX
- return BinaryOperator::createSetNE(Op0, AddOne(CI));
- }
-
- // If we still have a setle or setge instruction, turn it into the
- // appropriate setlt or setgt instruction. Since the border cases have
+ switch (I.getPredicate()) {
+ default: break;
+ case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
+ if (CI->isMinValue(false))
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
+ return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
+ if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
+ return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
+ break;
+
+ case ICmpInst::ICMP_SLT:
+ if (CI->isMinValue(true)) // A <s MIN -> FALSE
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
+ return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
+ if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
+ return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
+ break;
+
+ case ICmpInst::ICMP_UGT:
+ if (CI->isMaxValue(false)) // A >u MAX -> FALSE
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ if (CI->isMinValue(false)) // A >u MIN -> A != MIN
+ return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
+ if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
+ return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
+ break;
+
+ case ICmpInst::ICMP_SGT:
+ if (CI->isMaxValue(true)) // A >s MAX -> FALSE
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ if (CI->isMinValue(true)) // A >s MIN -> A != MIN
+ return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
+ if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
+ return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
+ break;
+
+ case ICmpInst::ICMP_ULE:
+ if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
+ if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
+ return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
+ if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
+ return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
+ break;
+
+ case ICmpInst::ICMP_SLE:
+ if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
+ if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
+ return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
+ if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
+ return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
+ break;
+
+ case ICmpInst::ICMP_UGE:
+ if (CI->isMinValue(false)) // A >=u MIN -> TRUE
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
+ if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
+ return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
+ if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
+ return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
+ break;
+
+ case ICmpInst::ICMP_SGE:
+ if (CI->isMinValue(true)) // A >=s MIN -> TRUE
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
+ if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
+ return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
+ if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
+ return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
+ break;
+ }
+
+ // If we still have a icmp le or icmp ge instruction, turn it into the
+ // appropriate icmp lt or icmp gt instruction. Since the border cases have
// already been handled above, this requires little checking.
//
- if (I.getOpcode() == Instruction::SetLE)
- return BinaryOperator::createSetLT(Op0, AddOne(CI));
- if (I.getOpcode() == Instruction::SetGE)
- return BinaryOperator::createSetGT(Op0, SubOne(CI));
-
+ if (I.getPredicate() == ICmpInst::ICMP_ULE)
+ return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
+ if (I.getPredicate() == ICmpInst::ICMP_SLE)
+ return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
+ if (I.getPredicate() == ICmpInst::ICMP_UGE)
+ return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
+ if (I.getPredicate() == ICmpInst::ICMP_SGE)
+ return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
// See if we can fold the comparison based on bits known to be zero or one
// in the input.
// Given the known and unknown bits, compute a range that the LHS could be
// in.
if (KnownOne | KnownZero) {
- if (Ty->isUnsigned()) { // Unsigned comparison.
- uint64_t Min, Max;
- uint64_t RHSVal = CI->getZExtValue();
- ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
- Min, Max);
- switch (I.getOpcode()) { // LE/GE have been folded already.
- default: assert(0 && "Unknown setcc opcode!");
- case Instruction::SetEQ:
- if (Max < RHSVal || Min > RHSVal)
- return ReplaceInstUsesWith(I, ConstantBool::False);
- break;
- case Instruction::SetNE:
- if (Max < RHSVal || Min > RHSVal)
- return ReplaceInstUsesWith(I, ConstantBool::True);
- break;
- case Instruction::SetLT:
- if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
- if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
- break;
- case Instruction::SetGT:
- if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
- if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
- break;
- }
- } else { // Signed comparison.
- int64_t Min, Max;
- int64_t RHSVal = CI->getSExtValue();
- ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
- Min, Max);
- switch (I.getOpcode()) { // LE/GE have been folded already.
- default: assert(0 && "Unknown setcc opcode!");
- case Instruction::SetEQ:
- if (Max < RHSVal || Min > RHSVal)
- return ReplaceInstUsesWith(I, ConstantBool::False);
- break;
- case Instruction::SetNE:
- if (Max < RHSVal || Min > RHSVal)
- return ReplaceInstUsesWith(I, ConstantBool::True);
- break;
- case Instruction::SetLT:
- if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
- if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
- break;
- case Instruction::SetGT:
- if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
- if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
- break;
- }
+ // Compute the Min, Max and RHS values based on the known bits. For the
+ // EQ and NE we use unsigned values.
+ uint64_t UMin, UMax, URHSVal;
+ int64_t SMin, SMax, SRHSVal;
+ if (ICmpInst::isSignedPredicate(I.getPredicate())) {
+ SRHSVal = CI->getSExtValue();
+ ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, SMin,
+ SMax);
+ } else {
+ URHSVal = CI->getZExtValue();
+ ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, UMin,
+ UMax);
+ }
+ switch (I.getPredicate()) { // LE/GE have been folded already.
+ default: assert(0 && "Unknown icmp opcode!");
+ case ICmpInst::ICMP_EQ:
+ if (UMax < URHSVal || UMin > URHSVal)
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ break;
+ case ICmpInst::ICMP_NE:
+ if (UMax < URHSVal || UMin > URHSVal)
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
+ break;
+ case ICmpInst::ICMP_ULT:
+ if (UMax < URHSVal)
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
+ if (UMin > URHSVal)
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ break;
+ case ICmpInst::ICMP_UGT:
+ if (UMin > URHSVal)
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
+ if (UMax < URHSVal)
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ break;
+ case ICmpInst::ICMP_SLT:
+ if (SMax < SRHSVal)
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
+ if (SMin > SRHSVal)
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ break;
+ case ICmpInst::ICMP_SGT:
+ if (SMin > SRHSVal)
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
+ if (SMax < SRHSVal)
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ break;
}
}
-
+ // Since the RHS is a ConstantInt (CI), if the left hand side is an
+ // instruction, see if that instruction also has constants so that the
+ // instruction can be folded into the icmp
if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
switch (LHSI->getOpcode()) {
case Instruction::And:
if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
LHSI->getOperand(0)->hasOneUse()) {
+ ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
+
+ // If the LHS is an AND of a truncating cast, we can widen the
+ // and/compare to be the input width without changing the value
+ // produced, eliminating a cast.
+ if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
+ // We can do this transformation if either the AND constant does not
+ // have its sign bit set or if it is an equality comparison.
+ // Extending a relational comparison when we're checking the sign
+ // bit would not work.
+ if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
+ (I.isEquality() ||
+ (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
+ (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
+ ConstantInt *NewCST;
+ ConstantInt *NewCI;
+ if (Cast->getOperand(0)->getType()->isSigned()) {
+ NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
+ AndCST->getZExtValue());
+ NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
+ CI->getZExtValue());
+ } else {
+ NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
+ AndCST->getZExtValue());
+ NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
+ CI->getZExtValue());
+ }
+ Instruction *NewAnd =
+ BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
+ LHSI->getName());
+ InsertNewInstBefore(NewAnd, I);
+ return new ICmpInst(I.getPredicate(), NewAnd, NewCI);
+ }
+ }
+
// If this is: (X >> C1) & C2 != C3 (where any shift and any compare
// could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
// happens a LOT in code produced by the C front-end, for bitfield
// access.
ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
- ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
// Check to see if there is a noop-cast between the shift and the and.
if (!Shift) {
if (CastInst *CI = dyn_cast<CastInst>(LHSI->getOperand(0)))
- if (CI->getOperand(0)->getType()->isIntegral() &&
- CI->getOperand(0)->getType()->getPrimitiveSizeInBits() ==
- CI->getType()->getPrimitiveSizeInBits())
+ if (CI->getOpcode() == Instruction::BitCast)
Shift = dyn_cast<ShiftInst>(CI->getOperand(0));
}
-
- ConstantUInt *ShAmt;
- ShAmt = Shift ? dyn_cast<ConstantUInt>(Shift->getOperand(1)) : 0;
+
+ ConstantInt *ShAmt;
+ ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
const Type *AndTy = AndCST->getType(); // Type of the and.
// into the mask. This can only happen with signed shift
// rights, as they sign-extend.
if (ShAmt) {
- bool CanFold = Shift->getOpcode() != Instruction::Shr ||
- Ty->isUnsigned();
+ bool CanFold = Shift->isLogicalShift();
if (!CanFold) {
// To test for the bad case of the signed shr, see if any
// of the bits shifted in could be tested after the mask.
- int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getValue();
+ int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
- Constant *OShAmt = ConstantUInt::get(Type::UByteTy, ShAmtVal);
+ Constant *OShAmt = ConstantInt::get(Type::UByteTy, ShAmtVal);
Constant *ShVal =
ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
OShAmt);
if (CanFold) {
Constant *NewCst;
if (Shift->getOpcode() == Instruction::Shl)
- NewCst = ConstantExpr::getUShr(CI, ShAmt);
+ NewCst = ConstantExpr::getLShr(CI, ShAmt);
else
NewCst = ConstantExpr::getShl(CI, ShAmt);
// If we shifted bits out, the fold is not going to work out.
// As a special case, check to see if this means that the
// result is always true or false now.
- if (I.getOpcode() == Instruction::SetEQ)
- return ReplaceInstUsesWith(I, ConstantBool::False);
- if (I.getOpcode() == Instruction::SetNE)
- return ReplaceInstUsesWith(I, ConstantBool::True);
+ if (I.getPredicate() == ICmpInst::ICMP_EQ)
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ if (I.getPredicate() == ICmpInst::ICMP_NE)
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
} else {
I.setOperand(1, NewCst);
Constant *NewAndCST;
if (Shift->getOpcode() == Instruction::Shl)
- NewAndCST = ConstantExpr::getUShr(AndCST, ShAmt);
+ NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
else
NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
LHSI->setOperand(1, NewAndCST);
if (AndTy == Ty)
LHSI->setOperand(0, Shift->getOperand(0));
else {
- Value *NewCast = InsertCastBefore(Shift->getOperand(0), AndTy,
+ Value *NewCast = InsertCastBefore(Instruction::BitCast,
+ Shift->getOperand(0), AndTy,
*Shift);
LHSI->setOperand(0, NewCast);
}
}
}
}
+
+ // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
+ // preferable because it allows the C<<Y expression to be hoisted out
+ // of a loop if Y is invariant and X is not.
+ if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
+ I.isEquality() && !Shift->isArithmeticShift() &&
+ isa<Instruction>(Shift->getOperand(0))) {
+ // Compute C << Y.
+ Value *NS;
+ if (Shift->getOpcode() == Instruction::LShr) {
+ NS = new ShiftInst(Instruction::Shl, AndCST, Shift->getOperand(1),
+ "tmp");
+ } else {
+ // Insert a logical shift.
+ NS = new ShiftInst(Instruction::LShr, AndCST,
+ Shift->getOperand(1), "tmp");
+ }
+ InsertNewInstBefore(cast<Instruction>(NS), I);
+
+ // If C's sign doesn't agree with the and, insert a cast now.
+ if (NS->getType() != LHSI->getType())
+ NS = InsertCastBefore(Instruction::BitCast, NS, LHSI->getType(),
+ I);
+
+ Value *ShiftOp = Shift->getOperand(0);
+ if (ShiftOp->getType() != LHSI->getType())
+ ShiftOp = InsertCastBefore(Instruction::BitCast, ShiftOp,
+ LHSI->getType(), I);
+
+ // Compute X & (C << Y).
+ Instruction *NewAnd =
+ BinaryOperator::createAnd(ShiftOp, NS, LHSI->getName());
+ InsertNewInstBefore(NewAnd, I);
+
+ I.setOperand(0, NewAnd);
+ return &I;
+ }
}
break;
- case Instruction::Shl: // (setcc (shl X, ShAmt), CI)
- if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
- switch (I.getOpcode()) {
- default: break;
- case Instruction::SetEQ:
- case Instruction::SetNE: {
+ case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
+ if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
+ if (I.isEquality()) {
unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
// Check that the shift amount is in range. If not, don't perform
// undefined shifts. When the shift is visited it will be
// simplified.
- if (ShAmt->getValue() >= TypeBits)
+ if (ShAmt->getZExtValue() >= TypeBits)
break;
// If we are comparing against bits always shifted out, the
// comparison cannot succeed.
Constant *Comp =
- ConstantExpr::getShl(ConstantExpr::getShr(CI, ShAmt), ShAmt);
+ ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
if (Comp != CI) {// Comparing against a bit that we know is zero.
- bool IsSetNE = I.getOpcode() == Instruction::SetNE;
- Constant *Cst = ConstantBool::get(IsSetNE);
+ bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
+ Constant *Cst = ConstantBool::get(IsICMP_NE);
return ReplaceInstUsesWith(I, Cst);
}
if (LHSI->hasOneUse()) {
// Otherwise strength reduce the shift into an and.
- unsigned ShAmtVal = (unsigned)ShAmt->getValue();
+ unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
Constant *Mask;
if (CI->getType()->isUnsigned()) {
- Mask = ConstantUInt::get(CI->getType(), Val);
+ Mask = ConstantInt::get(CI->getType(), Val);
} else if (ShAmtVal != 0) {
- Mask = ConstantSInt::get(CI->getType(), Val);
+ Mask = ConstantInt::get(CI->getType(), Val);
} else {
Mask = ConstantInt::getAllOnesValue(CI->getType());
}
BinaryOperator::createAnd(LHSI->getOperand(0),
Mask, LHSI->getName()+".mask");
Value *And = InsertNewInstBefore(AndI, I);
- return new SetCondInst(I.getOpcode(), And,
- ConstantExpr::getUShr(CI, ShAmt));
+ return new ICmpInst(I.getPredicate(), And,
+ ConstantExpr::getLShr(CI, ShAmt));
}
}
- }
}
break;
- case Instruction::Shr: // (setcc (shr X, ShAmt), CI)
- if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
- switch (I.getOpcode()) {
- default: break;
- case Instruction::SetEQ:
- case Instruction::SetNE: {
-
+ case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
+ case Instruction::AShr:
+ if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
+ if (I.isEquality()) {
// Check that the shift amount is in range. If not, don't perform
// undefined shifts. When the shift is visited it will be
// simplified.
unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
- if (ShAmt->getValue() >= TypeBits)
+ if (ShAmt->getZExtValue() >= TypeBits)
break;
// If we are comparing against bits always shifted out, the
// comparison cannot succeed.
- Constant *Comp =
- ConstantExpr::getShr(ConstantExpr::getShl(CI, ShAmt), ShAmt);
+ Constant *Comp;
+ if (CI->getType()->isUnsigned())
+ Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
+ ShAmt);
+ else
+ Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
+ ShAmt);
if (Comp != CI) {// Comparing against a bit that we know is zero.
- bool IsSetNE = I.getOpcode() == Instruction::SetNE;
- Constant *Cst = ConstantBool::get(IsSetNE);
+ bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
+ Constant *Cst = ConstantBool::get(IsICMP_NE);
return ReplaceInstUsesWith(I, Cst);
}
if (LHSI->hasOneUse() || CI->isNullValue()) {
- unsigned ShAmtVal = (unsigned)ShAmt->getValue();
+ unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
// Otherwise strength reduce the shift into an and.
uint64_t Val = ~0ULL; // All ones.
Constant *Mask;
if (CI->getType()->isUnsigned()) {
Val &= ~0ULL >> (64-TypeBits);
- Mask = ConstantUInt::get(CI->getType(), Val);
+ Mask = ConstantInt::get(CI->getType(), Val);
} else {
- Mask = ConstantSInt::get(CI->getType(), Val);
+ Mask = ConstantInt::get(CI->getType(), Val);
}
Instruction *AndI =
BinaryOperator::createAnd(LHSI->getOperand(0),
Mask, LHSI->getName()+".mask");
Value *And = InsertNewInstBefore(AndI, I);
- return new SetCondInst(I.getOpcode(), And,
+ return new ICmpInst(I.getPredicate(), And,
ConstantExpr::getShl(CI, ShAmt));
}
- break;
- }
}
}
break;
- case Instruction::Div:
- // Fold: (div X, C1) op C2 -> range check
+ case Instruction::SDiv:
+ case Instruction::UDiv:
+ // Fold: icmp pred ([us]div X, C1), C2 -> range test
+ // Fold this div into the comparison, producing a range check.
+ // Determine, based on the divide type, what the range is being
+ // checked. If there is an overflow on the low or high side, remember
+ // it, otherwise compute the range [low, hi) bounding the new value.
+ // See: InsertRangeTest above for the kinds of replacements possible.
if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
- // Fold this div into the comparison, producing a range check.
- // Determine, based on the divide type, what the range is being
- // checked. If there is an overflow on the low or high side, remember
- // it, otherwise compute the range [low, hi) bounding the new value.
- bool LoOverflow = false, HiOverflow = 0;
- ConstantInt *LoBound = 0, *HiBound = 0;
-
- ConstantInt *Prod;
- bool ProdOV = MulWithOverflow(Prod, CI, DivRHS);
+ // FIXME: If the operand types don't match the type of the divide
+ // then don't attempt this transform. The code below doesn't have the
+ // logic to deal with a signed divide and an unsigned compare (and
+ // vice versa). This is because (x /s C1) <s C2 produces different
+ // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
+ // (x /u C1) <u C2. Simply casting the operands and result won't
+ // work. :( The if statement below tests that condition and bails
+ // if it finds it.
+ bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
+ if (!I.isEquality() && DivIsSigned != I.isSignedPredicate())
+ break;
- Instruction::BinaryOps Opcode = I.getOpcode();
+ // Initialize the variables that will indicate the nature of the
+ // range check.
+ bool LoOverflow = false, HiOverflow = false;
+ ConstantInt *LoBound = 0, *HiBound = 0;
- if (DivRHS->isNullValue()) { // Don't hack on divide by zeros.
- } else if (LHSI->getType()->isUnsigned()) { // udiv
+ // Compute Prod = CI * DivRHS. We are essentially solving an equation
+ // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
+ // C2 (CI). By solving for X we can turn this into a range check
+ // instead of computing a divide.
+ ConstantInt *Prod =
+ cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
+
+ // Determine if the product overflows by seeing if the product is
+ // not equal to the divide. Make sure we do the same kind of divide
+ // as in the LHS instruction that we're folding.
+ bool ProdOV = !DivRHS->isNullValue() &&
+ (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
+ ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
+
+ // Get the ICmp opcode
+ ICmpInst::Predicate predicate = I.getPredicate();
+
+ if (DivRHS->isNullValue()) {
+ // Don't hack on divide by zeros!
+ } else if (!DivIsSigned) { // udiv
LoBound = Prod;
LoOverflow = ProdOV;
HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
- } else if (isPositive(DivRHS)) { // Divisor is > 0.
+ } else if (isPositive(DivRHS)) { // Divisor is > 0.
if (CI->isNullValue()) { // (X / pos) op 0
// Can't overflow.
LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
HiBound = Prod;
HiOverflow = ProdOV;
}
- } else { // Divisor is < 0.
+ } else { // Divisor is < 0.
if (CI->isNullValue()) { // (X / neg) op 0
LoBound = AddOne(DivRHS);
HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
if (HiBound == DivRHS)
- LoBound = 0; // - INTMIN = INTMIN
+ LoBound = 0; // - INTMIN = INTMIN
} else if (isPositive(CI)) { // (X / neg) op pos
HiOverflow = LoOverflow = ProdOV;
if (!LoOverflow)
}
// Dividing by a negate swaps the condition.
- Opcode = SetCondInst::getSwappedCondition(Opcode);
+ predicate = ICmpInst::getSwappedPredicate(predicate);
}
if (LoBound) {
Value *X = LHSI->getOperand(0);
- switch (Opcode) {
- default: assert(0 && "Unhandled setcc opcode!");
- case Instruction::SetEQ:
+ switch (predicate) {
+ default: assert(0 && "Unhandled icmp opcode!");
+ case ICmpInst::ICMP_EQ:
if (LoOverflow && HiOverflow)
- return ReplaceInstUsesWith(I, ConstantBool::False);
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
else if (HiOverflow)
- return new SetCondInst(Instruction::SetGE, X, LoBound);
+ return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
+ ICmpInst::ICMP_UGE, X, LoBound);
else if (LoOverflow)
- return new SetCondInst(Instruction::SetLT, X, HiBound);
+ return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
+ ICmpInst::ICMP_ULT, X, HiBound);
else
- return InsertRangeTest(X, LoBound, HiBound, true, I);
- case Instruction::SetNE:
+ return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
+ true, I);
+ case ICmpInst::ICMP_NE:
if (LoOverflow && HiOverflow)
- return ReplaceInstUsesWith(I, ConstantBool::True);
+ return ReplaceInstUsesWith(I, ConstantBool::getTrue());
else if (HiOverflow)
- return new SetCondInst(Instruction::SetLT, X, LoBound);
+ return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
+ ICmpInst::ICMP_ULT, X, LoBound);
else if (LoOverflow)
- return new SetCondInst(Instruction::SetGE, X, HiBound);
+ return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
+ ICmpInst::ICMP_UGE, X, HiBound);
else
- return InsertRangeTest(X, LoBound, HiBound, false, I);
- case Instruction::SetLT:
+ return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
+ false, I);
+ case ICmpInst::ICMP_ULT:
+ case ICmpInst::ICMP_SLT:
if (LoOverflow)
- return ReplaceInstUsesWith(I, ConstantBool::False);
- return new SetCondInst(Instruction::SetLT, X, LoBound);
- case Instruction::SetGT:
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ return new ICmpInst(predicate, X, LoBound);
+ case ICmpInst::ICMP_UGT:
+ case ICmpInst::ICMP_SGT:
if (HiOverflow)
- return ReplaceInstUsesWith(I, ConstantBool::False);
- return new SetCondInst(Instruction::SetGE, X, HiBound);
+ return ReplaceInstUsesWith(I, ConstantBool::getFalse());
+ if (predicate == ICmpInst::ICMP_UGT)
+ return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
+ else
+ return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
}
}
}
break;
}
- // Simplify seteq and setne instructions...
- if (I.getOpcode() == Instruction::SetEQ ||
- I.getOpcode() == Instruction::SetNE) {
- bool isSetNE = I.getOpcode() == Instruction::SetNE;
+ // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
+ if (I.isEquality()) {
+ bool isICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
- // If the first operand is (and|or|xor) with a constant, and the second
- // operand is a constant, simplify a bit.
+ // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
+ // the second operand is a constant, simplify a bit.
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
switch (BO->getOpcode()) {
- case Instruction::Rem:
+ case Instruction::SRem:
// If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
- if (CI->isNullValue() && isa<ConstantSInt>(BO->getOperand(1)) &&
- BO->hasOneUse() &&
- cast<ConstantSInt>(BO->getOperand(1))->getValue() > 1) {
- int64_t V = cast<ConstantSInt>(BO->getOperand(1))->getValue();
- if (isPowerOf2_64(V)) {
- unsigned L2 = Log2_64(V);
- const Type *UTy = BO->getType()->getUnsignedVersion();
- Value *NewX = InsertNewInstBefore(new CastInst(BO->getOperand(0),
- UTy, "tmp"), I);
- Constant *RHSCst = ConstantUInt::get(UTy, 1ULL << L2);
- Value *NewRem =InsertNewInstBefore(BinaryOperator::createRem(NewX,
- RHSCst, BO->getName()), I);
- return BinaryOperator::create(I.getOpcode(), NewRem,
- Constant::getNullValue(UTy));
+ if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
+ BO->hasOneUse()) {
+ int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
+ if (V > 1 && isPowerOf2_64(V)) {
+ Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
+ BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
+ return new ICmpInst(I.getPredicate(), NewRem,
+ Constant::getNullValue(BO->getType()));
}
}
break;
-
case Instruction::Add:
// Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
if (BO->hasOneUse())
- return new SetCondInst(I.getOpcode(), BO->getOperand(0),
- ConstantExpr::getSub(CI, BOp1C));
+ return new ICmpInst(I.getPredicate(), BO->getOperand(0),
+ ConstantExpr::getSub(CI, BOp1C));
} else if (CI->isNullValue()) {
// Replace ((add A, B) != 0) with (A != -B) if A or B is
// efficiently invertible, or if the add has just this one use.
Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
if (Value *NegVal = dyn_castNegVal(BOp1))
- return new SetCondInst(I.getOpcode(), BOp0, NegVal);
+ return new ICmpInst(I.getPredicate(), BOp0, NegVal);
else if (Value *NegVal = dyn_castNegVal(BOp0))
- return new SetCondInst(I.getOpcode(), NegVal, BOp1);
+ return new ICmpInst(I.getPredicate(), NegVal, BOp1);
else if (BO->hasOneUse()) {
Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
BO->setName("");
InsertNewInstBefore(Neg, I);
- return new SetCondInst(I.getOpcode(), BOp0, Neg);
+ return new ICmpInst(I.getPredicate(), BOp0, Neg);
}
}
break;
// For the xor case, we can xor two constants together, eliminating
// the explicit xor.
if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
- return BinaryOperator::create(I.getOpcode(), BO->getOperand(0),
- ConstantExpr::getXor(CI, BOC));
+ return new ICmpInst(I.getPredicate(), BO->getOperand(0),
+ ConstantExpr::getXor(CI, BOC));
// FALLTHROUGH
case Instruction::Sub:
// Replace (([sub|xor] A, B) != 0) with (A != B)
if (CI->isNullValue())
- return new SetCondInst(I.getOpcode(), BO->getOperand(0),
- BO->getOperand(1));
+ return new ICmpInst(I.getPredicate(), BO->getOperand(0),
+ BO->getOperand(1));
break;
case Instruction::Or:
if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
Constant *NotCI = ConstantExpr::getNot(CI);
if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
- return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
+ return ReplaceInstUsesWith(I, ConstantBool::get(isICMP_NE));
}
break;
// comparison can never succeed!
if (!ConstantExpr::getAnd(CI,
ConstantExpr::getNot(BOC))->isNullValue())
- return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
+ return ReplaceInstUsesWith(I, ConstantBool::get(isICMP_NE));
// If we have ((X & C) == C), turn it into ((X & C) != 0).
if (CI == BOC && isOneBitSet(CI))
- return new SetCondInst(isSetNE ? Instruction::SetEQ :
- Instruction::SetNE, Op0,
- Constant::getNullValue(CI->getType()));
+ return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
+ ICmpInst::ICMP_NE, Op0,
+ Constant::getNullValue(CI->getType()));
- // Replace (and X, (1 << size(X)-1) != 0) with x < 0, converting X
- // to be a signed value as appropriate.
+ // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
if (isSignBit(BOC)) {
Value *X = BO->getOperand(0);
// If 'X' is not signed, insert a cast now...
if (!BOC->getType()->isSigned()) {
const Type *DestTy = BOC->getType()->getSignedVersion();
- X = InsertCastBefore(X, DestTy, I);
+ X = InsertCastBefore(Instruction::BitCast, X, DestTy, I);
}
- return new SetCondInst(isSetNE ? Instruction::SetLT :
- Instruction::SetGE, X,
- Constant::getNullValue(X->getType()));
+ Constant *Zero = Constant::getNullValue(X->getType());
+ ICmpInst::Predicate pred = isICMP_NE ?
+ ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
+ return new ICmpInst(pred, X, Zero);
}
// ((X & ~7) == 0) --> X < 8
if (CI->isNullValue() && isHighOnes(BOC)) {
Value *X = BO->getOperand(0);
Constant *NegX = ConstantExpr::getNeg(BOC);
-
- // If 'X' is signed, insert a cast now.
- if (NegX->getType()->isSigned()) {
- const Type *DestTy = NegX->getType()->getUnsignedVersion();
- X = InsertCastBefore(X, DestTy, I);
- NegX = ConstantExpr::getCast(NegX, DestTy);
- }
-
- return new SetCondInst(isSetNE ? Instruction::SetGE :
- Instruction::SetLT, X, NegX);
+ ICmpInst::Predicate pred = isICMP_NE ?
+ ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
+ return new ICmpInst(pred, X, NegX);
}
}
default: break;
}
+ } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
+ // Handle set{eq|ne} <intrinsic>, intcst.
+ switch (II->getIntrinsicID()) {
+ default: break;
+ case Intrinsic::bswap_i16:
+ // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
+ WorkList.push_back(II); // Dead?
+ I.setOperand(0, II->getOperand(1));
+ I.setOperand(1, ConstantInt::get(Type::UShortTy,
+ ByteSwap_16(CI->getZExtValue())));
+ return &I;
+ case Intrinsic::bswap_i32:
+ // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
+ WorkList.push_back(II); // Dead?
+ I.setOperand(0, II->getOperand(1));
+ I.setOperand(1, ConstantInt::get(Type::UIntTy,
+ ByteSwap_32(CI->getZExtValue())));
+ return &I;
+ case Intrinsic::bswap_i64:
+ // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
+ WorkList.push_back(II); // Dead?
+ I.setOperand(0, II->getOperand(1));
+ I.setOperand(1, ConstantInt::get(Type::ULongTy,
+ ByteSwap_64(CI->getZExtValue())));
+ return &I;
+ }
}
- } else { // Not a SetEQ/SetNE
- // If the LHS is a cast from an integral value of the same size,
+ } else { // Not a ICMP_EQ/ICMP_NE
+ // If the LHS is a cast from an integral value of the same size, then
+ // since we know the RHS is a constant, try to simlify.
if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
Value *CastOp = Cast->getOperand(0);
const Type *SrcTy = CastOp->getType();
unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
- if (SrcTy != Cast->getType() && SrcTy->isInteger() &&
+ if (SrcTy->isInteger() &&
SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
- assert((SrcTy->isSigned() ^ Cast->getType()->isSigned()) &&
- "Source and destination signednesses should differ!");
- if (Cast->getType()->isSigned()) {
- // If this is a signed comparison, check for comparisons in the
- // vicinity of zero.
- if (I.getOpcode() == Instruction::SetLT && CI->isNullValue())
- // X < 0 => x > 127
- return BinaryOperator::createSetGT(CastOp,
- ConstantUInt::get(SrcTy, (1ULL << (SrcTySize-1))-1));
- else if (I.getOpcode() == Instruction::SetGT &&
- cast<ConstantSInt>(CI)->getValue() == -1)
- // X > -1 => x < 128
- return BinaryOperator::createSetLT(CastOp,
- ConstantUInt::get(SrcTy, 1ULL << (SrcTySize-1)));
- } else {
- ConstantUInt *CUI = cast<ConstantUInt>(CI);
- if (I.getOpcode() == Instruction::SetLT &&
- CUI->getValue() == 1ULL << (SrcTySize-1))
- // X < 128 => X > -1
- return BinaryOperator::createSetGT(CastOp,
- ConstantSInt::get(SrcTy, -1));
- else if (I.getOpcode() == Instruction::SetGT &&
- CUI->getValue() == (1ULL << (SrcTySize-1))-1)
- // X > 127 => X < 0
- return BinaryOperator::createSetLT(CastOp,
- Constant::getNullValue(SrcTy));
+ // If this is an unsigned comparison, try to make the comparison use
+ // smaller constant values.
+ switch (I.getPredicate()) {
+ default: break;
+ case ICmpInst::ICMP_ULT: { // X u< 128 => X s> -1
+ ConstantInt *CUI = cast<ConstantInt>(CI);
+ if (CUI->getZExtValue() == 1ULL << (SrcTySize-1))
+ return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
+ ConstantInt::get(SrcTy, -1));
+ break;
+ }
+ case ICmpInst::ICMP_UGT: { // X u> 127 => X s< 0
+ ConstantInt *CUI = cast<ConstantInt>(CI);
+ if (CUI->getZExtValue() == (1ULL << (SrcTySize-1))-1)
+ return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
+ Constant::getNullValue(SrcTy));
+ break;
+ }
}
+
}
}
}
}
- // Handle setcc with constant RHS's that can be integer, FP or pointer.
+ // Handle icmp with constant RHS
if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
switch (LHSI->getOpcode()) {
case Instruction::GetElementPtr:
if (RHSC->isNullValue()) {
- // Transform setcc GEP P, int 0, int 0, int 0, null -> setcc P, null
+ // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
bool isAllZeros = true;
for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
if (!isa<Constant>(LHSI->getOperand(i)) ||
break;
}
if (isAllZeros)
- return new SetCondInst(I.getOpcode(), LHSI->getOperand(0),
+ return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
Constant::getNullValue(LHSI->getOperand(0)->getType()));
}
break;
if (LHSI->hasOneUse()) {
if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
// Fold the known value into the constant operand.
- Op1 = ConstantExpr::get(I.getOpcode(), C, RHSC);
- // Insert a new SetCC of the other select operand.
- Op2 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
- LHSI->getOperand(2), RHSC,
- I.getName()), I);
+ Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
+ // Insert a new ICmp of the other select operand.
+ Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
+ LHSI->getOperand(2), RHSC,
+ I.getName()), I);
} else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
// Fold the known value into the constant operand.
- Op2 = ConstantExpr::get(I.getOpcode(), C, RHSC);
- // Insert a new SetCC of the other select operand.
- Op1 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
- LHSI->getOperand(1), RHSC,
- I.getName()), I);
+ Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
+ // Insert a new ICmp of the other select operand.
+ Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
+ LHSI->getOperand(1), RHSC,
+ I.getName()), I);
}
}
}
}
- // If we can optimize a 'setcc GEP, P' or 'setcc P, GEP', do so now.
+ // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
if (User *GEP = dyn_castGetElementPtr(Op0))
- if (Instruction *NI = FoldGEPSetCC(GEP, Op1, I.getOpcode(), I))
+ if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
return NI;
if (User *GEP = dyn_castGetElementPtr(Op1))
- if (Instruction *NI = FoldGEPSetCC(GEP, Op0,
- SetCondInst::getSwappedCondition(I.getOpcode()), I))
+ if (Instruction *NI = FoldGEPICmp(GEP, Op0,
+ ICmpInst::getSwappedPredicate(I.getPredicate()), I))
return NI;
- // Test to see if the operands of the setcc are casted versions of other
+ // Test to see if the operands of the icmp are casted versions of other
// values. If the cast can be stripped off both arguments, we do so now.
if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
Value *CastOp0 = CI->getOperand(0);
- if (CastOp0->getType()->isLosslesslyConvertibleTo(CI->getType()) &&
- (isa<Constant>(Op1) || isa<CastInst>(Op1)) &&
- (I.getOpcode() == Instruction::SetEQ ||
- I.getOpcode() == Instruction::SetNE)) {
+ if (CI->isLosslessCast() && I.isEquality() &&
+ (isa<Constant>(Op1) || isa<CastInst>(Op1))) {
// We keep moving the cast from the left operand over to the right
// operand, where it can often be eliminated completely.
Op0 = CastOp0;
// If operand #1 is a cast instruction, see if we can eliminate it as
// well.
- if (CastInst *CI2 = dyn_cast<CastInst>(Op1))
- if (CI2->getOperand(0)->getType()->isLosslesslyConvertibleTo(
- Op0->getType()))
- Op1 = CI2->getOperand(0);
+ if (CastInst *CI2 = dyn_cast<CastInst>(Op1)) {
+ Value *CI2Op0 = CI2->getOperand(0);
+ if (CI2Op0->getType()->canLosslesslyBitCastTo(Op0->getType()))
+ Op1 = CI2Op0;
+ }
// If Op1 is a constant, we can fold the cast into the constant.
if (Op1->getType() != Op0->getType())
if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
- Op1 = ConstantExpr::getCast(Op1C, Op0->getType());
+ Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
} else {
- // Otherwise, cast the RHS right before the setcc
- Op1 = new CastInst(Op1, Op0->getType(), Op1->getName());
- InsertNewInstBefore(cast<Instruction>(Op1), I);
+ // Otherwise, cast the RHS right before the icmp
+ Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
}
- return BinaryOperator::create(I.getOpcode(), Op0, Op1);
+ return new ICmpInst(I.getPredicate(), Op0, Op1);
}
- // Handle the special case of: setcc (cast bool to X), <cst>
+ // Handle the special case of: icmp (cast bool to X), <cst>
// This comes up when you have code like
// int X = A < B;
// if (X) ...
// For generality, we handle any zero-extension of any operand comparison
// with a constant or another cast from the same type.
if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
- if (Instruction *R = visitSetCondInstWithCastAndCast(I))
+ if (Instruction *R = visitICmpInstWithCastAndCast(I))
return R;
}
- if (I.getOpcode() == Instruction::SetNE ||
- I.getOpcode() == Instruction::SetEQ) {
+ if (I.isEquality()) {
Value *A, *B;
if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
(A == Op1 || B == Op1)) {
// (A^B) == A -> B == 0
Value *OtherVal = A == Op1 ? B : A;
- return BinaryOperator::create(I.getOpcode(), OtherVal,
- Constant::getNullValue(A->getType()));
+ return new ICmpInst(I.getPredicate(), OtherVal,
+ Constant::getNullValue(A->getType()));
} else if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
(A == Op0 || B == Op0)) {
// A == (A^B) -> B == 0
Value *OtherVal = A == Op0 ? B : A;
- return BinaryOperator::create(I.getOpcode(), OtherVal,
- Constant::getNullValue(A->getType()));
+ return new ICmpInst(I.getPredicate(), OtherVal,
+ Constant::getNullValue(A->getType()));
} else if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
// (A-B) == A -> B == 0
- return BinaryOperator::create(I.getOpcode(), B,
- Constant::getNullValue(B->getType()));
+ return new ICmpInst(I.getPredicate(), B,
+ Constant::getNullValue(B->getType()));
} else if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
// A == (A-B) -> B == 0
- return BinaryOperator::create(I.getOpcode(), B,
- Constant::getNullValue(B->getType()));
+ return new ICmpInst(I.getPredicate(), B,
+ Constant::getNullValue(B->getType()));
+ }
+
+ Value *C, *D;
+ // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
+ if (Op0->hasOneUse() && Op1->hasOneUse() &&
+ match(Op0, m_And(m_Value(A), m_Value(B))) &&
+ match(Op1, m_And(m_Value(C), m_Value(D)))) {
+ Value *X = 0, *Y = 0, *Z = 0;
+
+ if (A == C) {
+ X = B; Y = D; Z = A;
+ } else if (A == D) {
+ X = B; Y = C; Z = A;
+ } else if (B == C) {
+ X = A; Y = D; Z = B;
+ } else if (B == D) {
+ X = A; Y = C; Z = B;
+ }
+
+ if (X) { // Build (X^Y) & Z
+ Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
+ Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
+ I.setOperand(0, Op1);
+ I.setOperand(1, Constant::getNullValue(Op1->getType()));
+ return &I;
+ }
}
}
return Changed ? &I : 0;
}
-// visitSetCondInstWithCastAndCast - Handle setcond (cast x to y), (cast/cst).
+// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
// We only handle extending casts so far.
//
-Instruction *InstCombiner::visitSetCondInstWithCastAndCast(SetCondInst &SCI) {
- Value *LHSCIOp = cast<CastInst>(SCI.getOperand(0))->getOperand(0);
- const Type *SrcTy = LHSCIOp->getType();
- const Type *DestTy = SCI.getOperand(0)->getType();
+Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
+ const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
+ Value *LHSCIOp = LHSCI->getOperand(0);
+ const Type *SrcTy = LHSCIOp->getType();
+ const Type *DestTy = LHSCI->getType();
Value *RHSCIOp;
- if (!DestTy->isIntegral() || !SrcTy->isIntegral())
+ // We only handle extension cast instructions, so far. Enforce this.
+ if (LHSCI->getOpcode() != Instruction::ZExt &&
+ LHSCI->getOpcode() != Instruction::SExt)
return 0;
- unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
- unsigned DestBits = DestTy->getPrimitiveSizeInBits();
- if (SrcBits >= DestBits) return 0; // Only handle extending cast.
-
- // Is this a sign or zero extension?
- bool isSignSrc = SrcTy->isSigned();
- bool isSignDest = DestTy->isSigned();
+ bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
+ bool isSignedCmp = ICI.isSignedPredicate();
- if (CastInst *CI = dyn_cast<CastInst>(SCI.getOperand(1))) {
+ if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
// Not an extension from the same type?
RHSCIOp = CI->getOperand(0);
- if (RHSCIOp->getType() != LHSCIOp->getType()) return 0;
- } else if (ConstantInt *CI = dyn_cast<ConstantInt>(SCI.getOperand(1))) {
- // Compute the constant that would happen if we truncated to SrcTy then
- // reextended to DestTy.
- Constant *Res = ConstantExpr::getCast(CI, SrcTy);
-
- if (ConstantExpr::getCast(Res, DestTy) == CI) {
- RHSCIOp = Res;
- } else {
- // If the value cannot be represented in the shorter type, we cannot emit
- // a simple comparison.
- if (SCI.getOpcode() == Instruction::SetEQ)
- return ReplaceInstUsesWith(SCI, ConstantBool::False);
- if (SCI.getOpcode() == Instruction::SetNE)
- return ReplaceInstUsesWith(SCI, ConstantBool::True);
-
- // Evaluate the comparison for LT.
- Value *Result;
- if (DestTy->isSigned()) {
- // We're performing a signed comparison.
- if (isSignSrc) {
- // Signed extend and signed comparison.
- if (cast<ConstantSInt>(CI)->getValue() < 0) // X < (small) --> false
- Result = ConstantBool::False;
- else
- Result = ConstantBool::True; // X < (large) --> true
- } else {
- // Unsigned extend and signed comparison.
- if (cast<ConstantSInt>(CI)->getValue() < 0)
- Result = ConstantBool::False;
- else
- Result = ConstantBool::True;
- }
- } else {
- // We're performing an unsigned comparison.
- if (!isSignSrc) {
- // Unsigned extend & compare -> always true.
- Result = ConstantBool::True;
- } else {
- // We're performing an unsigned comp with a sign extended value.
- // This is true if the input is >= 0. [aka >s -1]
- Constant *NegOne = ConstantIntegral::getAllOnesValue(SrcTy);
- Result = InsertNewInstBefore(BinaryOperator::createSetGT(LHSCIOp,
- NegOne, SCI.getName()), SCI);
- }
- }
+ if (RHSCIOp->getType() != LHSCIOp->getType())
+ return 0;
+ else
+ // Okay, just insert a compare of the reduced operands now!
+ return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
+ }
- // Finally, return the value computed.
- if (SCI.getOpcode() == Instruction::SetLT) {
- return ReplaceInstUsesWith(SCI, Result);
- } else {
- assert(SCI.getOpcode()==Instruction::SetGT &&"SetCC should be folded!");
- if (Constant *CI = dyn_cast<Constant>(Result))
- return ReplaceInstUsesWith(SCI, ConstantExpr::getNot(CI));
- else
- return BinaryOperator::createNot(Result);
- }
- }
- } else {
+ // If we aren't dealing with a constant on the RHS, exit early
+ ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
+ if (!CI)
return 0;
+
+ // Compute the constant that would happen if we truncated to SrcTy then
+ // reextended to DestTy.
+ Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
+ Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
+
+ // If the re-extended constant didn't change...
+ if (Res2 == CI) {
+ // Make sure that sign of the Cmp and the sign of the Cast are the same.
+ // For example, we might have:
+ // %A = sext short %X to uint
+ // %B = icmp ugt uint %A, 1330
+ // It is incorrect to transform this into
+ // %B = icmp ugt short %X, 1330
+ // because %A may have negative value.
+ //
+ // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
+ // OR operation is EQ/NE.
+ if (isSignedExt == isSignedCmp || SrcTy == Type::BoolTy || ICI.isEquality())
+ return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
+ else
+ return 0;
+ }
+
+ // The re-extended constant changed so the constant cannot be represented
+ // in the shorter type. Consequently, we cannot emit a simple comparison.
+
+ // First, handle some easy cases. We know the result cannot be equal at this
+ // point so handle the ICI.isEquality() cases
+ if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
+ return ReplaceInstUsesWith(ICI, ConstantBool::getFalse());
+ if (ICI.getPredicate() == ICmpInst::ICMP_NE)
+ return ReplaceInstUsesWith(ICI, ConstantBool::getTrue());
+
+ // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
+ // should have been folded away previously and not enter in here.
+ Value *Result;
+ if (isSignedCmp) {
+ // We're performing a signed comparison.
+ if (cast<ConstantInt>(CI)->getSExtValue() < 0)
+ Result = ConstantBool::getFalse(); // X < (small) --> false
+ else
+ Result = ConstantBool::getTrue(); // X < (large) --> true
+ } else {
+ // We're performing an unsigned comparison.
+ if (isSignedExt) {
+ // We're performing an unsigned comp with a sign extended value.
+ // This is true if the input is >= 0. [aka >s -1]
+ Constant *NegOne = ConstantIntegral::getAllOnesValue(SrcTy);
+ Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
+ NegOne, ICI.getName()), ICI);
+ } else {
+ // Unsigned extend & unsigned compare -> always true.
+ Result = ConstantBool::getTrue();
+ }
}
- // Okay, just insert a compare of the reduced operands now!
- return BinaryOperator::create(SCI.getOpcode(), LHSCIOp, RHSCIOp);
+ // Finally, return the value computed.
+ if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
+ ICI.getPredicate() == ICmpInst::ICMP_SLT) {
+ return ReplaceInstUsesWith(ICI, Result);
+ } else {
+ assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
+ ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
+ "ICmp should be folded!");
+ if (Constant *CI = dyn_cast<Constant>(Result))
+ return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
+ else
+ return BinaryOperator::createNot(Result);
+ }
}
Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
assert(I.getOperand(1)->getType() == Type::UByteTy);
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
- bool isLeftShift = I.getOpcode() == Instruction::Shl;
// shl X, 0 == X and shr X, 0 == X
// shl 0, X == 0 and shr 0, X == 0
Op0 == Constant::getNullValue(Op0->getType()))
return ReplaceInstUsesWith(I, Op0);
- if (isa<UndefValue>(Op0)) { // undef >>s X -> undef
- if (!isLeftShift && I.getType()->isSigned())
+ if (isa<UndefValue>(Op0)) {
+ if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
return ReplaceInstUsesWith(I, Op0);
- else // undef << X -> 0 AND undef >>u X -> 0
+ else // undef << X -> 0, undef >>u X -> 0
return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
}
if (isa<UndefValue>(Op1)) {
- if (isLeftShift || I.getType()->isUnsigned())// X << undef, X >>u undef -> 0
+ if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
+ return ReplaceInstUsesWith(I, Op0);
+ else // X << undef, X >>u undef -> 0
return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
- else
- return ReplaceInstUsesWith(I, Op0); // X >>s undef -> X
}
- // shr int -1, X = -1 (for any arithmetic shift rights of ~0)
- if (!isLeftShift)
- if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
+ // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
+ if (I.getOpcode() == Instruction::AShr)
+ if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
if (CSI->isAllOnesValue())
return ReplaceInstUsesWith(I, CSI);
return R;
// See if we can turn a signed shr into an unsigned shr.
- if (!isLeftShift && I.getType()->isSigned()) {
+ if (I.isArithmeticShift()) {
if (MaskedValueIsZero(Op0,
1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
- Value *V = InsertCastBefore(Op0, I.getType()->getUnsignedVersion(), I);
- V = InsertNewInstBefore(new ShiftInst(Instruction::Shr, V, Op1,
- I.getName()), I);
- return new CastInst(V, I.getType());
+ return new ShiftInst(Instruction::LShr, Op0, Op1, I.getName());
}
}
- if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op1))
- if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
- return Res;
+ if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
+ if (CUI->getType()->isUnsigned())
+ if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
+ return Res;
return 0;
}
-Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantUInt *Op1,
+Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
ShiftInst &I) {
- bool isLeftShift = I.getOpcode() == Instruction::Shl;
- bool isSignedShift = Op0->getType()->isSigned();
+ bool isLeftShift = I.getOpcode() == Instruction::Shl;
+ bool isSignedShift = I.getOpcode() == Instruction::AShr;
bool isUnsignedShift = !isSignedShift;
// See if we can simplify any instructions used by the instruction whose sole
// of a signed value.
//
unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
- if (Op1->getValue() >= TypeBits) {
+ if (Op1->getZExtValue() >= TypeBits) {
if (isUnsignedShift || isLeftShift)
return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
else {
- I.setOperand(1, ConstantUInt::get(Type::UByteTy, TypeBits-1));
+ I.setOperand(1, ConstantInt::get(Type::UByteTy, TypeBits-1));
return &I;
}
}
Op0BO->getName());
InsertNewInstBefore(YS, I); // (Y << C)
Instruction *X =
- BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
+ BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
Op0BO->getOperand(0)->getName());
InsertNewInstBefore(X, I); // (X + (Y << C))
Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
return BinaryOperator::createAnd(X, C2);
}
+ // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
match(Op0BO->getOperand(0),
m_And(m_Shr(m_Value(V1), m_Value(V2)),
V1->getName()+".mask");
InsertNewInstBefore(XM, I); // X & (CC << C)
- return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
+ return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
}
break;
// operation.
//
if (isValid && !isLeftShift && isSignedShift) {
- uint64_t Val = Op0C->getRawValue();
+ uint64_t Val = Op0C->getZExtValue();
isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
}
ShiftInst *ShiftOp = 0;
if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
ShiftOp = Op0SI;
- else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
- // If this is a noop-integer case of a shift instruction, use the shift.
- if (CI->getOperand(0)->getType()->isInteger() &&
- CI->getOperand(0)->getType()->getPrimitiveSizeInBits() ==
- CI->getType()->getPrimitiveSizeInBits() &&
- isa<ShiftInst>(CI->getOperand(0))) {
+ else if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
+ // If this is a noop-integer cast of a shift instruction, use the shift.
+ if (isa<ShiftInst>(CI->getOperand(0))) {
ShiftOp = cast<ShiftInst>(CI->getOperand(0));
}
}
- if (ShiftOp && isa<ConstantUInt>(ShiftOp->getOperand(1))) {
+ if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
// Find the operands and properties of the input shift. Note that the
// signedness of the input shift may differ from the current shift if there
// is a noop cast between the two.
- bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
- bool isShiftOfSignedShift = ShiftOp->getType()->isSigned();
+ bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
+ bool isShiftOfSignedShift = ShiftOp->getOpcode() == Instruction::AShr;
bool isShiftOfUnsignedShift = !isShiftOfSignedShift;
- ConstantUInt *ShiftAmt1C = cast<ConstantUInt>(ShiftOp->getOperand(1));
+ ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
- unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getValue();
- unsigned ShiftAmt2 = (unsigned)Op1->getValue();
+ unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
+ unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
// Check for (A << c1) << c2 and (A >> c1) >> c2.
if (isLeftShift == isShiftOfLeftShift) {
Amt = Op0->getType()->getPrimitiveSizeInBits();
Value *Op = ShiftOp->getOperand(0);
- if (isShiftOfSignedShift != isSignedShift)
- Op = InsertNewInstBefore(new CastInst(Op, I.getType(), "tmp"), I);
- return new ShiftInst(I.getOpcode(), Op,
- ConstantUInt::get(Type::UByteTy, Amt));
+ ShiftInst *ShiftResult = new ShiftInst(I.getOpcode(), Op,
+ ConstantInt::get(Type::UByteTy, Amt));
+ if (I.getType() == ShiftResult->getType())
+ return ShiftResult;
+ InsertNewInstBefore(ShiftResult, I);
+ return CastInst::create(Instruction::BitCast, ShiftResult, I.getType());
}
// Check for (A << c1) >> c2 or (A >> c1) << c2. If we are dealing with
if (isLeftShift)
C = ConstantExpr::getShl(C, ShiftAmt1C);
else
- C = ConstantExpr::getUShr(C, ShiftAmt1C);
+ C = ConstantExpr::getLShr(C, ShiftAmt1C);
Value *Op = ShiftOp->getOperand(0);
- if (isShiftOfSignedShift != isSignedShift)
- Op = InsertNewInstBefore(new CastInst(Op, I.getType(),Op->getName()),I);
+ if (Op->getType() != C->getType())
+ Op = InsertCastBefore(Instruction::BitCast, Op, I.getType(), I);
Instruction *Mask =
BinaryOperator::createAnd(Op, C, Op->getName()+".mask");
return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
} else if (ShiftAmt1 < ShiftAmt2) {
return new ShiftInst(I.getOpcode(), Mask,
- ConstantUInt::get(Type::UByteTy, ShiftAmt2-ShiftAmt1));
+ ConstantInt::get(Type::UByteTy, ShiftAmt2-ShiftAmt1));
} else if (isShiftOfUnsignedShift || isShiftOfLeftShift) {
if (isShiftOfUnsignedShift && !isShiftOfLeftShift && isSignedShift) {
- // Make sure to emit an unsigned shift right, not a signed one.
- Mask = InsertNewInstBefore(new CastInst(Mask,
- Mask->getType()->getUnsignedVersion(),
- Op->getName()), I);
- Mask = new ShiftInst(Instruction::Shr, Mask,
- ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
- InsertNewInstBefore(Mask, I);
- return new CastInst(Mask, I.getType());
+ return new ShiftInst(Instruction::LShr, Mask,
+ ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
} else {
return new ShiftInst(ShiftOp->getOpcode(), Mask,
- ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
+ ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
}
} else {
// (X >>s C1) << C2 where C1 > C2 === (X >>s (C1-C2)) & mask
- Op = InsertNewInstBefore(new CastInst(Mask,
- I.getType()->getSignedVersion(),
- Mask->getName()), I);
Instruction *Shift =
- new ShiftInst(ShiftOp->getOpcode(), Op,
- ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
+ new ShiftInst(ShiftOp->getOpcode(), Mask,
+ ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
InsertNewInstBefore(Shift, I);
C = ConstantIntegral::getAllOnesValue(Shift->getType());
C = ConstantExpr::getShl(C, Op1);
- Mask = BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
- InsertNewInstBefore(Mask, I);
- return new CastInst(Mask, I.getType());
+ return BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
}
} else {
// We can handle signed (X << C1) >>s C2 if it's a sign extend. In
}
if (SExtType) {
- Instruction *NewTrunc = new CastInst(ShiftOp->getOperand(0),
- SExtType, "sext");
+ Instruction *NewTrunc =
+ new TruncInst(ShiftOp->getOperand(0), SExtType, "sext");
InsertNewInstBefore(NewTrunc, I);
- return new CastInst(NewTrunc, I.getType());
+ return new SExtInst(NewTrunc, I.getType());
}
}
}
static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
unsigned &Offset) {
assert(Val->getType() == Type::UIntTy && "Unexpected allocation size type!");
- if (ConstantUInt *CI = dyn_cast<ConstantUInt>(Val)) {
- Offset = CI->getValue();
- Scale = 1;
- return ConstantUInt::get(Type::UIntTy, 0);
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
+ if (CI->getType()->isUnsigned()) {
+ Offset = CI->getZExtValue();
+ Scale = 1;
+ return ConstantInt::get(Type::UIntTy, 0);
+ }
} else if (Instruction *I = dyn_cast<Instruction>(Val)) {
if (I->getNumOperands() == 2) {
- if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(I->getOperand(1))) {
- if (I->getOpcode() == Instruction::Shl) {
- // This is a value scaled by '1 << the shift amt'.
- Scale = 1U << CUI->getValue();
- Offset = 0;
- return I->getOperand(0);
- } else if (I->getOpcode() == Instruction::Mul) {
- // This value is scaled by 'CUI'.
- Scale = CUI->getValue();
- Offset = 0;
- return I->getOperand(0);
- } else if (I->getOpcode() == Instruction::Add) {
- // We have X+C. Check to see if we really have (X*C2)+C1, where C1 is
- // divisible by C2.
- unsigned SubScale;
- Value *SubVal = DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
- Offset);
- Offset += CUI->getValue();
- if (SubScale > 1 && (Offset % SubScale == 0)) {
- Scale = SubScale;
- return SubVal;
+ if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ if (CUI->getType()->isUnsigned()) {
+ if (I->getOpcode() == Instruction::Shl) {
+ // This is a value scaled by '1 << the shift amt'.
+ Scale = 1U << CUI->getZExtValue();
+ Offset = 0;
+ return I->getOperand(0);
+ } else if (I->getOpcode() == Instruction::Mul) {
+ // This value is scaled by 'CUI'.
+ Scale = CUI->getZExtValue();
+ Offset = 0;
+ return I->getOperand(0);
+ } else if (I->getOpcode() == Instruction::Add) {
+ // We have X+C. Check to see if we really have (X*C2)+C1,
+ // where C1 is divisible by C2.
+ unsigned SubScale;
+ Value *SubVal =
+ DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
+ Offset += CUI->getZExtValue();
+ if (SubScale > 1 && (Offset % SubScale == 0)) {
+ Scale = SubScale;
+ return SubVal;
+ }
}
}
}
// Add operands to the worklist.
AddUsesToWorkList(*User);
++NumDeadInst;
- DEBUG(std::cerr << "IC: DCE: " << *User);
+ DOUT << "IC: DCE: " << *User;
User->eraseFromParent();
removeFromWorkList(User);
const Type *CastElTy = PTy->getElementType();
if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
- unsigned AllocElTyAlign = TD->getTypeSize(AllocElTy);
- unsigned CastElTyAlign = TD->getTypeSize(CastElTy);
+ unsigned AllocElTyAlign = TD->getTypeAlignment(AllocElTy);
+ unsigned CastElTyAlign = TD->getTypeAlignment(CastElTy);
if (CastElTyAlign < AllocElTyAlign) return 0;
// If the allocation has multiple uses, only promote it if we are strictly
if (Scale == 1) {
Amt = NumElements;
} else {
- Amt = ConstantUInt::get(Type::UIntTy, Scale);
- if (ConstantUInt *CI = dyn_cast<ConstantUInt>(NumElements))
- Amt = ConstantExpr::getMul(CI, cast<ConstantUInt>(Amt));
+ // If the allocation size is constant, form a constant mul expression
+ Amt = ConstantInt::get(Type::UIntTy, Scale);
+ if (isa<ConstantInt>(NumElements) && NumElements->getType()->isUnsigned())
+ Amt = ConstantExpr::getMul(
+ cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
+ // otherwise multiply the amount and the number of elements
else if (Scale != 1) {
Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
Amt = InsertNewInstBefore(Tmp, AI);
}
if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
- Value *Off = ConstantUInt::get(Type::UIntTy, Offset);
+ Value *Off = ConstantInt::get(Type::UIntTy, Offset);
Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
Amt = InsertNewInstBefore(Tmp, AI);
}
// die soon.
if (!AI.hasOneUse()) {
AddUsesToWorkList(AI);
- CastInst *NewCast = new CastInst(New, AI.getType(), "tmpcast");
+ // New is the allocation instruction, pointer typed. AI is the original
+ // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
+ CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
InsertNewInstBefore(NewCast, AI);
AI.replaceAllUsesWith(NewCast);
}
// These operators can all arbitrarily be extended or truncated.
return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
- case Instruction::Cast:
+ case Instruction::AShr:
+ case Instruction::LShr:
+ case Instruction::Shl:
+ // If this is just a bitcast changing the sign of the operation, we can
+ // convert if the operand can be converted.
+ if (V->getType()->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
+ return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
+ break;
+ case Instruction::Trunc:
+ case Instruction::ZExt:
+ case Instruction::SExt:
+ case Instruction::BitCast:
// If this is a cast from the destination type, we can trivially eliminate
// it, and this will remove a cast overall.
if (I->getOperand(0)->getType() == Ty) {
+ // If the first operand is itself a cast, and is eliminable, do not count
+ // this as an eliminable cast. We would prefer to eliminate those two
+ // casts first.
+ if (isa<CastInst>(I->getOperand(0)))
+ return true;
+
++NumCastsRemoved;
return true;
}
+ break;
+ default:
// TODO: Can handle more cases here.
break;
}
/// EvaluateInDifferentType - Given an expression that
/// CanEvaluateInDifferentType returns true for, actually insert the code to
/// evaluate the expression.
-Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty) {
+Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
+ bool isSigned ) {
if (Constant *C = dyn_cast<Constant>(V))
- return ConstantExpr::getCast(C, Ty);
+ return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
// Otherwise, it must be an instruction.
Instruction *I = cast<Instruction>(V);
case Instruction::And:
case Instruction::Or:
case Instruction::Xor: {
- Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty);
- Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty);
+ Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
+ Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
LHS, RHS, I->getName());
break;
}
- case Instruction::Cast:
- // If this is a cast from the destination type, return the input.
+ case Instruction::AShr:
+ case Instruction::LShr:
+ case Instruction::Shl: {
+ Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
+ Res = new ShiftInst((Instruction::OtherOps)I->getOpcode(), LHS,
+ I->getOperand(1), I->getName());
+ break;
+ }
+ case Instruction::Trunc:
+ case Instruction::ZExt:
+ case Instruction::SExt:
+ case Instruction::BitCast:
+ // If the source type of the cast is the type we're trying for then we can
+ // just return the source. There's no need to insert it because its not new.
if (I->getOperand(0)->getType() == Ty)
return I->getOperand(0);
+ // Some other kind of cast, which shouldn't happen, so just ..
+ // FALL THROUGH
+ default:
// TODO: Can handle more cases here.
assert(0 && "Unreachable!");
break;
}
- return InsertNewInstBefore(Res, *I);
+ return InsertNewInstBefore(Res, *I);
+}
+
+/// @brief Implement the transforms common to all CastInst visitors.
+Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
+ Value *Src = CI.getOperand(0);
+
+ // Casting undef to anything results in undef so might as just replace it and
+ // get rid of the cast.
+ if (isa<UndefValue>(Src)) // cast undef -> undef
+ return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
+
+ // Many cases of "cast of a cast" are eliminable. If its eliminable we just
+ // eliminate it now.
+ if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
+ if (Instruction::CastOps opc =
+ isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
+ // The first cast (CSrc) is eliminable so we need to fix up or replace
+ // the second cast (CI). CSrc will then have a good chance of being dead.
+ return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
+ }
+ }
+
+ // If casting the result of a getelementptr instruction with no offset, turn
+ // this into a cast of the original pointer!
+ //
+ if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
+ bool AllZeroOperands = true;
+ for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
+ if (!isa<Constant>(GEP->getOperand(i)) ||
+ !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
+ AllZeroOperands = false;
+ break;
+ }
+ if (AllZeroOperands) {
+ // Changing the cast operand is usually not a good idea but it is safe
+ // here because the pointer operand is being replaced with another
+ // pointer operand so the opcode doesn't need to change.
+ CI.setOperand(0, GEP->getOperand(0));
+ return &CI;
+ }
+ }
+
+ // If we are casting a malloc or alloca to a pointer to a type of the same
+ // size, rewrite the allocation instruction to allocate the "right" type.
+ if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
+ if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
+ return V;
+
+ // If we are casting a select then fold the cast into the select
+ if (SelectInst *SI = dyn_cast<SelectInst>(Src))
+ if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
+ return NV;
+
+ // If we are casting a PHI then fold the cast into the PHI
+ if (isa<PHINode>(Src))
+ if (Instruction *NV = FoldOpIntoPhi(CI))
+ return NV;
+
+ return 0;
+}
+
+/// Only the TRUNC, ZEXT, SEXT, and BITCONVERT can have both operands as
+/// integers. This function implements the common transforms for all those
+/// cases.
+/// @brief Implement the transforms common to CastInst with integer operands
+Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
+ if (Instruction *Result = commonCastTransforms(CI))
+ return Result;
+
+ Value *Src = CI.getOperand(0);
+ const Type *SrcTy = Src->getType();
+ const Type *DestTy = CI.getType();
+ unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
+ unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
+
+ // See if we can simplify any instructions used by the LHS whose sole
+ // purpose is to compute bits we don't care about.
+ uint64_t KnownZero = 0, KnownOne = 0;
+ if (SimplifyDemandedBits(&CI, DestTy->getIntegralTypeMask(),
+ KnownZero, KnownOne))
+ return &CI;
+
+ // If the source isn't an instruction or has more than one use then we
+ // can't do anything more.
+ Instruction *SrcI = dyn_cast<Instruction>(Src);
+ if (!SrcI || !Src->hasOneUse())
+ return 0;
+
+ // Attempt to propagate the cast into the instruction.
+ int NumCastsRemoved = 0;
+ if (CanEvaluateInDifferentType(SrcI, DestTy, NumCastsRemoved)) {
+ // If this cast is a truncate, evaluting in a different type always
+ // eliminates the cast, so it is always a win. If this is a noop-cast
+ // this just removes a noop cast which isn't pointful, but simplifies
+ // the code. If this is a zero-extension, we need to do an AND to
+ // maintain the clear top-part of the computation, so we require that
+ // the input have eliminated at least one cast. If this is a sign
+ // extension, we insert two new casts (to do the extension) so we
+ // require that two casts have been eliminated.
+ bool DoXForm = CI.isNoopCast(TD->getIntPtrType());
+ if (!DoXForm) {
+ switch (CI.getOpcode()) {
+ case Instruction::Trunc:
+ DoXForm = true;
+ break;
+ case Instruction::ZExt:
+ DoXForm = NumCastsRemoved >= 1;
+ break;
+ case Instruction::SExt:
+ DoXForm = NumCastsRemoved >= 2;
+ break;
+ case Instruction::BitCast:
+ DoXForm = false;
+ break;
+ default:
+ // All the others use floating point so we shouldn't actually
+ // get here because of the check above.
+ assert(!"Unknown cast type .. unreachable");
+ break;
+ }
+ }
+
+ if (DoXForm) {
+ Value *Res = EvaluateInDifferentType(SrcI, DestTy,
+ CI.getOpcode() == Instruction::SExt);
+ assert(Res->getType() == DestTy);
+ switch (CI.getOpcode()) {
+ default: assert(0 && "Unknown cast type!");
+ case Instruction::Trunc:
+ case Instruction::BitCast:
+ // Just replace this cast with the result.
+ return ReplaceInstUsesWith(CI, Res);
+ case Instruction::ZExt: {
+ // We need to emit an AND to clear the high bits.
+ assert(SrcBitSize < DestBitSize && "Not a zext?");
+ Constant *C =
+ ConstantInt::get(Type::ULongTy, (1ULL << SrcBitSize)-1);
+ if (DestBitSize < 64)
+ C = ConstantExpr::getTrunc(C, DestTy);
+ else {
+ assert(DestBitSize == 64);
+ C = ConstantExpr::getBitCast(C, DestTy);
+ }
+ return BinaryOperator::createAnd(Res, C);
+ }
+ case Instruction::SExt:
+ // We need to emit a cast to truncate, then a cast to sext.
+ return CastInst::create(Instruction::SExt,
+ InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
+ CI), DestTy);
+ }
+ }
+ }
+
+ Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
+ Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
+
+ switch (SrcI->getOpcode()) {
+ case Instruction::Add:
+ case Instruction::Mul:
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor:
+ // If we are discarding information, or just changing the sign,
+ // rewrite.
+ if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
+ // Don't insert two casts if they cannot be eliminated. We allow
+ // two casts to be inserted if the sizes are the same. This could
+ // only be converting signedness, which is a noop.
+ if (DestBitSize == SrcBitSize ||
+ !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
+ !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
+ Instruction::CastOps opcode = CI.getOpcode();
+ Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
+ Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
+ return BinaryOperator::create(
+ cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
+ }
+ }
+
+ // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
+ if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
+ SrcI->getOpcode() == Instruction::Xor &&
+ Op1 == ConstantBool::getTrue() &&
+ (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
+ Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
+ return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
+ }
+ break;
+ case Instruction::SDiv:
+ case Instruction::UDiv:
+ case Instruction::SRem:
+ case Instruction::URem:
+ // If we are just changing the sign, rewrite.
+ if (DestBitSize == SrcBitSize) {
+ // Don't insert two casts if they cannot be eliminated. We allow
+ // two casts to be inserted if the sizes are the same. This could
+ // only be converting signedness, which is a noop.
+ if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
+ !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
+ Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
+ Op0, DestTy, SrcI);
+ Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
+ Op1, DestTy, SrcI);
+ return BinaryOperator::create(
+ cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
+ }
+ }
+ break;
+
+ case Instruction::Shl:
+ // Allow changing the sign of the source operand. Do not allow
+ // changing the size of the shift, UNLESS the shift amount is a
+ // constant. We must not change variable sized shifts to a smaller
+ // size, because it is undefined to shift more bits out than exist
+ // in the value.
+ if (DestBitSize == SrcBitSize ||
+ (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
+ Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
+ Instruction::BitCast : Instruction::Trunc);
+ Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
+ return new ShiftInst(Instruction::Shl, Op0c, Op1);
+ }
+ break;
+ case Instruction::AShr:
+ // If this is a signed shr, and if all bits shifted in are about to be
+ // truncated off, turn it into an unsigned shr to allow greater
+ // simplifications.
+ if (DestBitSize < SrcBitSize &&
+ isa<ConstantInt>(Op1)) {
+ unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
+ if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
+ // Insert the new logical shift right.
+ return new ShiftInst(Instruction::LShr, Op0, Op1);
+ }
+ }
+ break;
+
+ case Instruction::ICmp:
+ // If we are just checking for a icmp eq of a single bit and casting it
+ // to an integer, then shift the bit to the appropriate place and then
+ // cast to integer to avoid the comparison.
+ if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
+ uint64_t Op1CV = Op1C->getZExtValue();
+ // cast (X == 0) to int --> X^1 iff X has only the low bit set.
+ // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
+ // cast (X == 1) to int --> X iff X has only the low bit set.
+ // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
+ // cast (X != 0) to int --> X iff X has only the low bit set.
+ // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
+ // cast (X != 1) to int --> X^1 iff X has only the low bit set.
+ // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
+ if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
+ // If Op1C some other power of two, convert:
+ uint64_t KnownZero, KnownOne;
+ uint64_t TypeMask = Op1->getType()->getIntegralTypeMask();
+ ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
+
+ // This only works for EQ and NE
+ ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
+ if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
+ break;
+
+ if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly 1 possible 1?
+ bool isNE = pred == ICmpInst::ICMP_NE;
+ if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
+ // (X&4) == 2 --> false
+ // (X&4) != 2 --> true
+ Constant *Res = ConstantBool::get(isNE);
+ Res = ConstantExpr::getZExt(Res, CI.getType());
+ return ReplaceInstUsesWith(CI, Res);
+ }
+
+ unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
+ Value *In = Op0;
+ if (ShiftAmt) {
+ // Perform a logical shr by shiftamt.
+ // Insert the shift to put the result in the low bit.
+ In = InsertNewInstBefore(
+ new ShiftInst(Instruction::LShr, In,
+ ConstantInt::get(Type::UByteTy, ShiftAmt),
+ In->getName()+".lobit"), CI);
+ }
+
+ if ((Op1CV != 0) == isNE) { // Toggle the low bit.
+ Constant *One = ConstantInt::get(In->getType(), 1);
+ In = BinaryOperator::createXor(In, One, "tmp");
+ InsertNewInstBefore(cast<Instruction>(In), CI);
+ }
+
+ if (CI.getType() == In->getType())
+ return ReplaceInstUsesWith(CI, In);
+ else
+ return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
+ }
+ }
+ }
+ break;
+ }
+ return 0;
+}
+
+Instruction *InstCombiner::visitTrunc(CastInst &CI) {
+ if (Instruction *Result = commonIntCastTransforms(CI))
+ return Result;
+
+ Value *Src = CI.getOperand(0);
+ const Type *Ty = CI.getType();
+ unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
+
+ if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
+ switch (SrcI->getOpcode()) {
+ default: break;
+ case Instruction::LShr:
+ // We can shrink lshr to something smaller if we know the bits shifted in
+ // are already zeros.
+ if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
+ unsigned ShAmt = ShAmtV->getZExtValue();
+
+ // Get a mask for the bits shifting in.
+ uint64_t Mask = (~0ULL >> (64-ShAmt)) << DestBitWidth;
+ Value* SrcIOp0 = SrcI->getOperand(0);
+ if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
+ if (ShAmt >= DestBitWidth) // All zeros.
+ return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
+
+ // Okay, we can shrink this. Truncate the input, then return a new
+ // shift.
+ Value *V = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
+ return new ShiftInst(Instruction::LShr, V, SrcI->getOperand(1));
+ }
+ } else { // This is a variable shr.
+
+ // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
+ // more LLVM instructions, but allows '1 << Y' to be hoisted if
+ // loop-invariant and CSE'd.
+ if (CI.getType() == Type::BoolTy && SrcI->hasOneUse()) {
+ Value *One = ConstantInt::get(SrcI->getType(), 1);
+
+ Value *V = InsertNewInstBefore(new ShiftInst(Instruction::Shl, One,
+ SrcI->getOperand(1),
+ "tmp"), CI);
+ V = InsertNewInstBefore(BinaryOperator::createAnd(V,
+ SrcI->getOperand(0),
+ "tmp"), CI);
+ Value *Zero = Constant::getNullValue(V->getType());
+ return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
+ }
+ }
+ break;
+ }
+ }
+
+ return 0;
}
+Instruction *InstCombiner::visitZExt(CastInst &CI) {
+ // If one of the common conversion will work ..
+ if (Instruction *Result = commonIntCastTransforms(CI))
+ return Result;
-// CastInst simplification
-//
-Instruction *InstCombiner::visitCastInst(CastInst &CI) {
Value *Src = CI.getOperand(0);
- // If the user is casting a value to the same type, eliminate this cast
- // instruction...
- if (CI.getType() == Src->getType())
- return ReplaceInstUsesWith(CI, Src);
-
- if (isa<UndefValue>(Src)) // cast undef -> undef
- return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
-
- // If casting the result of another cast instruction, try to eliminate this
- // one!
- //
+ // If this is a cast of a cast
if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
- Value *A = CSrc->getOperand(0);
- if (isEliminableCastOfCast(A->getType(), CSrc->getType(),
- CI.getType(), TD)) {
- // This instruction now refers directly to the cast's src operand. This
- // has a good chance of making CSrc dead.
- CI.setOperand(0, CSrc->getOperand(0));
- return &CI;
- }
-
- // If this is an A->B->A cast, and we are dealing with integral types, try
- // to convert this into a logical 'and' instruction.
- //
- if (A->getType()->isInteger() &&
- CI.getType()->isInteger() && CSrc->getType()->isInteger() &&
- CSrc->getType()->isUnsigned() && // B->A cast must zero extend
- CSrc->getType()->getPrimitiveSizeInBits() <
- CI.getType()->getPrimitiveSizeInBits()&&
- A->getType()->getPrimitiveSizeInBits() ==
- CI.getType()->getPrimitiveSizeInBits()) {
- assert(CSrc->getType() != Type::ULongTy &&
- "Cannot have type bigger than ulong!");
- uint64_t AndValue = CSrc->getType()->getIntegralTypeMask();
- Constant *AndOp = ConstantUInt::get(A->getType()->getUnsignedVersion(),
- AndValue);
- AndOp = ConstantExpr::getCast(AndOp, A->getType());
- Instruction *And = BinaryOperator::createAnd(CSrc->getOperand(0), AndOp);
- if (And->getType() != CI.getType()) {
- And->setName(CSrc->getName()+".mask");
- InsertNewInstBefore(And, CI);
- And = new CastInst(And, CI.getType());
- }
- return And;
+ // If this is a TRUNC followed by a ZEXT then we are dealing with integral
+ // types and if the sizes are just right we can convert this into a logical
+ // 'and' which will be much cheaper than the pair of casts.
+ if (isa<TruncInst>(CSrc)) {
+ // Get the sizes of the types involved
+ Value *A = CSrc->getOperand(0);
+ unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
+ unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
+ unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
+ // If we're actually extending zero bits and the trunc is a no-op
+ if (MidSize < DstSize && SrcSize == DstSize) {
+ // Replace both of the casts with an And of the type mask.
+ uint64_t AndValue = CSrc->getType()->getIntegralTypeMask();
+ Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
+ Instruction *And =
+ BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
+ // Unfortunately, if the type changed, we need to cast it back.
+ if (And->getType() != CI.getType()) {
+ And->setName(CSrc->getName()+".mask");
+ InsertNewInstBefore(And, CI);
+ And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
+ }
+ return And;
+ }
}
}
-
- // If this is a cast to bool, turn it into the appropriate setne instruction.
- if (CI.getType() == Type::BoolTy)
- return BinaryOperator::createSetNE(CI.getOperand(0),
- Constant::getNullValue(CI.getOperand(0)->getType()));
- // See if we can simplify any instructions used by the LHS whose sole
- // purpose is to compute bits we don't care about.
- if (CI.getType()->isInteger() && CI.getOperand(0)->getType()->isIntegral()) {
- uint64_t KnownZero, KnownOne;
- if (SimplifyDemandedBits(&CI, CI.getType()->getIntegralTypeMask(),
- KnownZero, KnownOne))
- return &CI;
- }
-
- // If casting the result of a getelementptr instruction with no offset, turn
- // this into a cast of the original pointer!
- //
- if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
- bool AllZeroOperands = true;
- for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
- if (!isa<Constant>(GEP->getOperand(i)) ||
- !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
- AllZeroOperands = false;
- break;
- }
- if (AllZeroOperands) {
- CI.setOperand(0, GEP->getOperand(0));
- return &CI;
- }
+ return 0;
+}
+
+Instruction *InstCombiner::visitSExt(CastInst &CI) {
+ return commonIntCastTransforms(CI);
+}
+
+Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
+ return commonCastTransforms(CI);
+}
+
+Instruction *InstCombiner::visitFPExt(CastInst &CI) {
+ return commonCastTransforms(CI);
+}
+
+Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
+ return commonCastTransforms(CI);
+}
+
+Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
+ return commonCastTransforms(CI);
+}
+
+Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
+ return commonCastTransforms(CI);
+}
+
+Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
+ return commonCastTransforms(CI);
+}
+
+Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
+ return commonCastTransforms(CI);
+}
+
+Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
+ return commonCastTransforms(CI);
+}
+
+Instruction *InstCombiner::visitBitCast(CastInst &CI) {
+
+ // If the operands are integer typed then apply the integer transforms,
+ // otherwise just apply the common ones.
+ Value *Src = CI.getOperand(0);
+ const Type *SrcTy = Src->getType();
+ const Type *DestTy = CI.getType();
+
+ if (SrcTy->isInteger() && DestTy->isInteger()) {
+ if (Instruction *Result = commonIntCastTransforms(CI))
+ return Result;
+ } else {
+ if (Instruction *Result = commonCastTransforms(CI))
+ return Result;
}
- // If we are casting a malloc or alloca to a pointer to a type of the same
- // size, rewrite the allocation instruction to allocate the "right" type.
- //
- if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
- if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
- return V;
- if (SelectInst *SI = dyn_cast<SelectInst>(Src))
- if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
- return NV;
- if (isa<PHINode>(Src))
- if (Instruction *NV = FoldOpIntoPhi(CI))
- return NV;
-
+ // Get rid of casts from one type to the same type. These are useless and can
+ // be replaced by the operand.
+ if (DestTy == Src->getType())
+ return ReplaceInstUsesWith(CI, Src);
+
// If the source and destination are pointers, and this cast is equivalent to
// a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
// This can enhance SROA and other transforms that want type-safe pointers.
- if (const PointerType *DstPTy = dyn_cast<PointerType>(CI.getType()))
- if (const PointerType *SrcPTy = dyn_cast<PointerType>(Src->getType())) {
- const Type *DstTy = DstPTy->getElementType();
- const Type *SrcTy = SrcPTy->getElementType();
+ if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
+ if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
+ const Type *DstElTy = DstPTy->getElementType();
+ const Type *SrcElTy = SrcPTy->getElementType();
Constant *ZeroUInt = Constant::getNullValue(Type::UIntTy);
unsigned NumZeros = 0;
- while (SrcTy != DstTy &&
- isa<CompositeType>(SrcTy) && !isa<PointerType>(SrcTy)) {
- SrcTy = cast<CompositeType>(SrcTy)->getTypeAtIndex(ZeroUInt);
+ while (SrcElTy != DstElTy &&
+ isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
+ SrcElTy->getNumContainedTypes() /* not "{}" */) {
+ SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
++NumZeros;
}
// If we found a path from the src to dest, create the getelementptr now.
- if (SrcTy == DstTy) {
+ if (SrcElTy == DstElTy) {
std::vector<Value*> Idxs(NumZeros+1, ZeroUInt);
return new GetElementPtrInst(Src, Idxs);
}
}
-
- // If the source value is an instruction with only this use, we can attempt to
- // propagate the cast into the instruction. Also, only handle integral types
- // for now.
- if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
- if (SrcI->hasOneUse() && Src->getType()->isIntegral() &&
- CI.getType()->isInteger()) { // Don't mess with casts to bool here
-
- int NumCastsRemoved = 0;
- if (CanEvaluateInDifferentType(SrcI, CI.getType(), NumCastsRemoved)) {
- // If this cast is a truncate, evaluting in a different type always
- // eliminates the cast, so it is always a win. If this is a noop-cast
- // this just removes a noop cast which isn't pointful, but simplifies
- // the code. If this is a zero-extension, we need to do an AND to
- // maintain the clear top-part of the computation, so we require that
- // the input have eliminated at least one cast. If this is a sign
- // extension, we insert two new casts (to do the extension) so we
- // require that two casts have been eliminated.
- bool DoXForm;
- switch (getCastType(Src->getType(), CI.getType())) {
- default: assert(0 && "Unknown cast type!");
- case Noop:
- case Truncate:
- DoXForm = true;
- break;
- case Zeroext:
- DoXForm = NumCastsRemoved >= 1;
- break;
- case Signext:
- DoXForm = NumCastsRemoved >= 2;
- break;
- }
-
- if (DoXForm) {
- Value *Res = EvaluateInDifferentType(SrcI, CI.getType());
- assert(Res->getType() == CI.getType());
- switch (getCastType(Src->getType(), CI.getType())) {
- default: assert(0 && "Unknown cast type!");
- case Noop:
- case Truncate:
- // Just replace this cast with the result.
- return ReplaceInstUsesWith(CI, Res);
- case Zeroext: {
- // We need to emit an AND to clear the high bits.
- unsigned SrcBitSize = Src->getType()->getPrimitiveSizeInBits();
- unsigned DestBitSize = CI.getType()->getPrimitiveSizeInBits();
- assert(SrcBitSize < DestBitSize && "Not a zext?");
- Constant *C = ConstantUInt::get(Type::ULongTy, (1 << SrcBitSize)-1);
- C = ConstantExpr::getCast(C, CI.getType());
- return BinaryOperator::createAnd(Res, C);
- }
- case Signext:
- // We need to emit a cast to truncate, then a cast to sext.
- return new CastInst(InsertCastBefore(Res, Src->getType(), CI),
- CI.getType());
- }
- }
- }
-
- const Type *DestTy = CI.getType();
- unsigned SrcBitSize = Src->getType()->getPrimitiveSizeInBits();
- unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
-
- Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
- Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
-
- switch (SrcI->getOpcode()) {
- case Instruction::Add:
- case Instruction::Mul:
- case Instruction::And:
- case Instruction::Or:
- case Instruction::Xor:
- // If we are discarding information, or just changing the sign, rewrite.
- if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
- // Don't insert two casts if they cannot be eliminated. We allow two
- // casts to be inserted if the sizes are the same. This could only be
- // converting signedness, which is a noop.
- if (DestBitSize == SrcBitSize || !ValueRequiresCast(Op1, DestTy,TD) ||
- !ValueRequiresCast(Op0, DestTy, TD)) {
- Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
- Value *Op1c = InsertOperandCastBefore(Op1, DestTy, SrcI);
- return BinaryOperator::create(cast<BinaryOperator>(SrcI)
- ->getOpcode(), Op0c, Op1c);
- }
- }
-
- // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
- if (SrcBitSize == 1 && SrcI->getOpcode() == Instruction::Xor &&
- Op1 == ConstantBool::True &&
- (!Op0->hasOneUse() || !isa<SetCondInst>(Op0))) {
- Value *New = InsertOperandCastBefore(Op0, DestTy, &CI);
- return BinaryOperator::createXor(New,
- ConstantInt::get(CI.getType(), 1));
- }
- break;
- case Instruction::Shl:
- // Allow changing the sign of the source operand. Do not allow changing
- // the size of the shift, UNLESS the shift amount is a constant. We
- // mush not change variable sized shifts to a smaller size, because it
- // is undefined to shift more bits out than exist in the value.
- if (DestBitSize == SrcBitSize ||
- (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
- Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
- return new ShiftInst(Instruction::Shl, Op0c, Op1);
- }
- break;
- case Instruction::Shr:
- // If this is a signed shr, and if all bits shifted in are about to be
- // truncated off, turn it into an unsigned shr to allow greater
- // simplifications.
- if (DestBitSize < SrcBitSize && Src->getType()->isSigned() &&
- isa<ConstantInt>(Op1)) {
- unsigned ShiftAmt = cast<ConstantUInt>(Op1)->getValue();
- if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
- // Convert to unsigned.
- Value *N1 = InsertOperandCastBefore(Op0,
- Op0->getType()->getUnsignedVersion(), &CI);
- // Insert the new shift, which is now unsigned.
- N1 = InsertNewInstBefore(new ShiftInst(Instruction::Shr, N1,
- Op1, Src->getName()), CI);
- return new CastInst(N1, CI.getType());
- }
- }
- break;
+ }
- case Instruction::SetEQ:
- case Instruction::SetNE:
- // We if we are just checking for a seteq of a single bit and casting it
- // to an integer. If so, shift the bit to the appropriate place then
- // cast to integer to avoid the comparison.
- if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
- uint64_t Op1CV = Op1C->getZExtValue();
- // cast (X == 0) to int --> X^1 iff X has only the low bit set.
- // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
- // cast (X == 1) to int --> X iff X has only the low bit set.
- // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
- // cast (X != 0) to int --> X iff X has only the low bit set.
- // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
- // cast (X != 1) to int --> X^1 iff X has only the low bit set.
- // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
- if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
- // If Op1C some other power of two, convert:
- uint64_t KnownZero, KnownOne;
- uint64_t TypeMask = Op1->getType()->getIntegralTypeMask();
- ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
-
- if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly one possible 1?
- bool isSetNE = SrcI->getOpcode() == Instruction::SetNE;
- if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
- // (X&4) == 2 --> false
- // (X&4) != 2 --> true
- Constant *Res = ConstantBool::get(isSetNE);
- Res = ConstantExpr::getCast(Res, CI.getType());
- return ReplaceInstUsesWith(CI, Res);
- }
-
- unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
- Value *In = Op0;
- if (ShiftAmt) {
- // Perform an unsigned shr by shiftamt. Convert input to
- // unsigned if it is signed.
- if (In->getType()->isSigned())
- In = InsertNewInstBefore(new CastInst(In,
- In->getType()->getUnsignedVersion(), In->getName()),CI);
- // Insert the shift to put the result in the low bit.
- In = InsertNewInstBefore(new ShiftInst(Instruction::Shr, In,
- ConstantInt::get(Type::UByteTy, ShiftAmt),
- In->getName()+".lobit"), CI);
- }
-
- if ((Op1CV != 0) == isSetNE) { // Toggle the low bit.
- Constant *One = ConstantInt::get(In->getType(), 1);
- In = BinaryOperator::createXor(In, One, "tmp");
- InsertNewInstBefore(cast<Instruction>(In), CI);
- }
-
- if (CI.getType() == In->getType())
- return ReplaceInstUsesWith(CI, In);
- else
- return new CastInst(In, CI.getType());
- }
- }
- }
- break;
- }
- }
-
- if (SrcI->hasOneUse()) {
- if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(SrcI)) {
- // Okay, we have (cast (shuffle ..)). We know this cast is a bitconvert
- // because the inputs are known to be a vector. Check to see if this is
- // a cast to a vector with the same # elts.
- if (isa<PackedType>(CI.getType()) &&
- cast<PackedType>(CI.getType())->getNumElements() ==
- SVI->getType()->getNumElements()) {
- CastInst *Tmp;
- // If either of the operands is a cast from CI.getType(), then
- // evaluating the shuffle in the casted destination's type will allow
- // us to eliminate at least one cast.
- if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
- Tmp->getOperand(0)->getType() == CI.getType()) ||
- ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
- Tmp->getOperand(1)->getType() == CI.getType())) {
- Value *LHS = InsertOperandCastBefore(SVI->getOperand(0),
- CI.getType(), &CI);
- Value *RHS = InsertOperandCastBefore(SVI->getOperand(1),
- CI.getType(), &CI);
- // Return a new shuffle vector. Use the same element ID's, as we
- // know the vector types match #elts.
- return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
- }
+ if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
+ if (SVI->hasOneUse()) {
+ // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
+ // a bitconvert to a vector with the same # elts.
+ if (isa<PackedType>(DestTy) &&
+ cast<PackedType>(DestTy)->getNumElements() ==
+ SVI->getType()->getNumElements()) {
+ CastInst *Tmp;
+ // If either of the operands is a cast from CI.getType(), then
+ // evaluating the shuffle in the casted destination's type will allow
+ // us to eliminate at least one cast.
+ if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
+ Tmp->getOperand(0)->getType() == DestTy) ||
+ ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
+ Tmp->getOperand(0)->getType() == DestTy)) {
+ Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
+ SVI->getOperand(0), DestTy, &CI);
+ Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
+ SVI->getOperand(1), DestTy, &CI);
+ // Return a new shuffle vector. Use the same element ID's, as we
+ // know the vector types match #elts.
+ return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
}
}
}
}
-
return 0;
}
return 3; // Can fold through either operand.
case Instruction::Sub: // Can only fold on the amount subtracted.
case Instruction::Shl: // Can only fold on the shift amount.
- case Instruction::Shr:
+ case Instruction::LShr:
+ case Instruction::AShr:
return 1;
default:
return 0; // Cannot fold
case Instruction::Xor:
return Constant::getNullValue(I->getType());
case Instruction::Shl:
- case Instruction::Shr:
+ case Instruction::LShr:
+ case Instruction::AShr:
return Constant::getNullValue(Type::UByteTy);
case Instruction::And:
return ConstantInt::getAllOnesValue(I->getType());
if (TI->getNumOperands() == 1) {
// If this is a non-volatile load or a cast from the same type,
// merge.
- if (TI->getOpcode() == Instruction::Cast) {
+ if (TI->isCast()) {
if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
return 0;
} else {
SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
FI->getOperand(0), SI.getName()+".v");
InsertNewInstBefore(NewSI, SI);
- return new CastInst(NewSI, TI->getType());
+ return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
+ TI->getType());
}
- // Only handle binary operators here.
- if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI))
+ // Only handle binary, compare and shift operators here.
+ if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI) && !isa<CmpInst>(TI))
+ return 0;
+
+ // If the CmpInst predicates don't match, then the instructions aren't the
+ // same and we can't continue.
+ if (isa<CmpInst>(TI) && isa<CmpInst>(FI) &&
+ (cast<CmpInst>(TI)->getPredicate() != cast<CmpInst>(FI)->getPredicate()))
return 0;
// Figure out if the operations have any operands in common.
// select true, X, Y -> X
// select false, X, Y -> Y
if (ConstantBool *C = dyn_cast<ConstantBool>(CondVal))
- if (C == ConstantBool::True)
- return ReplaceInstUsesWith(SI, TrueVal);
- else {
- assert(C == ConstantBool::False);
- return ReplaceInstUsesWith(SI, FalseVal);
- }
+ return ReplaceInstUsesWith(SI, C->getValue() ? TrueVal : FalseVal);
// select C, X, X -> X
if (TrueVal == FalseVal)
if (SI.getType() == Type::BoolTy)
if (ConstantBool *C = dyn_cast<ConstantBool>(TrueVal)) {
- if (C == ConstantBool::True) {
+ if (C->getValue()) {
// Change: A = select B, true, C --> A = or B, C
return BinaryOperator::createOr(CondVal, FalseVal);
} else {
return BinaryOperator::createAnd(NotCond, FalseVal);
}
} else if (ConstantBool *C = dyn_cast<ConstantBool>(FalseVal)) {
- if (C == ConstantBool::False) {
+ if (C->getValue() == false) {
// Change: A = select B, C, false --> A = and B, C
return BinaryOperator::createAnd(CondVal, TrueVal);
} else {
if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
// select C, 1, 0 -> cast C to int
- if (FalseValC->isNullValue() && TrueValC->getRawValue() == 1) {
- return new CastInst(CondVal, SI.getType());
- } else if (TrueValC->isNullValue() && FalseValC->getRawValue() == 1) {
+ if (FalseValC->isNullValue() && TrueValC->getZExtValue() == 1) {
+ return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
+ } else if (TrueValC->isNullValue() && FalseValC->getZExtValue() == 1) {
// select C, 0, 1 -> cast !C to int
Value *NotCond =
InsertNewInstBefore(BinaryOperator::createNot(CondVal,
"not."+CondVal->getName()), SI);
- return new CastInst(NotCond, SI.getType());
- }
-
- // If one of the constants is zero (we know they can't both be) and we
- // have a setcc instruction with zero, and we have an 'and' with the
- // non-constant value, eliminate this whole mess. This corresponds to
- // cases like this: ((X & 27) ? 27 : 0)
- if (TrueValC->isNullValue() || FalseValC->isNullValue())
- if (Instruction *IC = dyn_cast<Instruction>(SI.getCondition()))
- if ((IC->getOpcode() == Instruction::SetEQ ||
- IC->getOpcode() == Instruction::SetNE) &&
- isa<ConstantInt>(IC->getOperand(1)) &&
+ return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
+ }
+
+ if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
+
+ // (x <s 0) ? -1 : 0 -> ashr x, 31
+ // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
+ if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
+ if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
+ bool CanXForm = false;
+ if (IC->isSignedPredicate())
+ CanXForm = CmpCst->isNullValue() &&
+ IC->getPredicate() == ICmpInst::ICMP_SLT;
+ else {
+ unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
+ CanXForm = (CmpCst->getZExtValue() == ~0ULL >> (64-Bits+1)) &&
+ IC->getPredicate() == ICmpInst::ICMP_UGT;
+ }
+
+ if (CanXForm) {
+ // The comparison constant and the result are not neccessarily the
+ // same width. Make an all-ones value by inserting a AShr.
+ Value *X = IC->getOperand(0);
+ unsigned Bits = X->getType()->getPrimitiveSizeInBits();
+ Constant *ShAmt = ConstantInt::get(Type::UByteTy, Bits-1);
+ Instruction *SRA = new ShiftInst(Instruction::AShr, X,
+ ShAmt, "ones");
+ InsertNewInstBefore(SRA, SI);
+
+ // Finally, convert to the type of the select RHS. We figure out
+ // if this requires a SExt, Trunc or BitCast based on the sizes.
+ Instruction::CastOps opc = Instruction::BitCast;
+ unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
+ unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
+ if (SRASize < SISize)
+ opc = Instruction::SExt;
+ else if (SRASize > SISize)
+ opc = Instruction::Trunc;
+ return CastInst::create(opc, SRA, SI.getType());
+ }
+ }
+
+
+ // If one of the constants is zero (we know they can't both be) and we
+ // have a fcmp instruction with zero, and we have an 'and' with the
+ // non-constant value, eliminate this whole mess. This corresponds to
+ // cases like this: ((X & 27) ? 27 : 0)
+ if (TrueValC->isNullValue() || FalseValC->isNullValue())
+ if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
cast<Constant>(IC->getOperand(1))->isNullValue())
if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
if (ICA->getOpcode() == Instruction::And &&
ICA->getOperand(1) == FalseValC) &&
isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
// Okay, now we know that everything is set up, we just don't
- // know whether we have a setne or seteq and whether the true or
- // false val is the zero.
+ // know whether we have a icmp_ne or icmp_eq and whether the
+ // true or false val is the zero.
bool ShouldNotVal = !TrueValC->isNullValue();
- ShouldNotVal ^= IC->getOpcode() == Instruction::SetNE;
+ ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
Value *V = ICA;
if (ShouldNotVal)
V = InsertNewInstBefore(BinaryOperator::create(
Instruction::Xor, V, ICA->getOperand(1)), SI);
return ReplaceInstUsesWith(SI, V);
}
+ }
+ }
+
+ // See if we are selecting two values based on a comparison of the two values.
+ if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
+ if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
+ // Transform (X == Y) ? X : Y -> Y
+ if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
+ return ReplaceInstUsesWith(SI, FalseVal);
+ // Transform (X != Y) ? X : Y -> X
+ if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
+ return ReplaceInstUsesWith(SI, TrueVal);
+ // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
+
+ } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
+ // Transform (X == Y) ? Y : X -> X
+ if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
+ return ReplaceInstUsesWith(SI, FalseVal);
+ // Transform (X != Y) ? Y : X -> Y
+ if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
+ return ReplaceInstUsesWith(SI, TrueVal);
+ // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
}
+ }
// See if we are selecting two values based on a comparison of the two values.
- if (SetCondInst *SCI = dyn_cast<SetCondInst>(CondVal)) {
- if (SCI->getOperand(0) == TrueVal && SCI->getOperand(1) == FalseVal) {
+ if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
+ if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
// Transform (X == Y) ? X : Y -> Y
- if (SCI->getOpcode() == Instruction::SetEQ)
+ if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
return ReplaceInstUsesWith(SI, FalseVal);
// Transform (X != Y) ? X : Y -> X
- if (SCI->getOpcode() == Instruction::SetNE)
+ if (ICI->getPredicate() == ICmpInst::ICMP_NE)
return ReplaceInstUsesWith(SI, TrueVal);
// NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
- } else if (SCI->getOperand(0) == FalseVal && SCI->getOperand(1) == TrueVal){
+ } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
// Transform (X == Y) ? Y : X -> X
- if (SCI->getOpcode() == Instruction::SetEQ)
+ if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
return ReplaceInstUsesWith(SI, FalseVal);
// Transform (X != Y) ? Y : X -> Y
- if (SCI->getOpcode() == Instruction::SetNE)
+ if (ICI->getPredicate() == ICmpInst::ICMP_NE)
return ReplaceInstUsesWith(SI, TrueVal);
// NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
}
if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
if (TI->hasOneUse() && FI->hasOneUse()) {
- bool isInverse = false;
Instruction *AddOp = 0, *SubOp = 0;
// Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
}
}
return Align;
- } else if (isa<CastInst>(V) ||
+ } else if (isa<BitCastInst>(V) ||
(isa<ConstantExpr>(V) &&
- cast<ConstantExpr>(V)->getOpcode() == Instruction::Cast)) {
+ cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
User *CI = cast<User>(V);
if (isa<PointerType>(CI->getOperand(0)->getType()))
return GetKnownAlignment(CI->getOperand(0), TD);
if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
- if (CI->getRawValue() == 1) {
+ if (CI->getZExtValue() == 1) {
// Replace the instruction with just byte operations. We would
// transform other cases to loads/stores, but we don't know if
// alignment is sufficient.
if (GVSrc->isConstant()) {
Module *M = CI.getParent()->getParent()->getParent();
const char *Name;
- if (CI.getCalledFunction()->getFunctionType()->getParamType(3) ==
+ if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
Type::UIntTy)
Name = "llvm.memcpy.i32";
else
unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
unsigned Align = std::min(Alignment1, Alignment2);
- if (MI->getAlignment()->getRawValue() < Align) {
- MI->setAlignment(ConstantUInt::get(Type::UIntTy, Align));
+ if (MI->getAlignment()->getZExtValue() < Align) {
+ MI->setAlignment(ConstantInt::get(Type::UIntTy, Align));
Changed = true;
}
} else if (isa<MemSetInst>(MI)) {
unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
- if (MI->getAlignment()->getRawValue() < Alignment) {
- MI->setAlignment(ConstantUInt::get(Type::UIntTy, Alignment));
+ if (MI->getAlignment()->getZExtValue() < Alignment) {
+ MI->setAlignment(ConstantInt::get(Type::UIntTy, Alignment));
Changed = true;
}
}
// Turn PPC lvx -> load if the pointer is known aligned.
// Turn X86 loadups -> load if the pointer is known aligned.
if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
- Value *Ptr = InsertCastBefore(II->getOperand(1),
+ Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
PointerType::get(II->getType()), CI);
return new LoadInst(Ptr);
}
// Turn stvx -> store if the pointer is known aligned.
if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
- Value *Ptr = InsertCastBefore(II->getOperand(2), OpPtrTy, CI);
+ Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
+ OpPtrTy, CI);
return new StoreInst(II->getOperand(1), Ptr);
}
break;
// Turn X86 storeu -> store if the pointer is known aligned.
if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
- Value *Ptr = InsertCastBefore(II->getOperand(1), OpPtrTy, CI);
+ Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
+ OpPtrTy, CI);
return new StoreInst(II->getOperand(2), Ptr);
}
break;
+
+ case Intrinsic::x86_sse_cvttss2si: {
+ // These intrinsics only demands the 0th element of its input vector. If
+ // we can simplify the input based on that, do so now.
+ uint64_t UndefElts;
+ if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
+ UndefElts)) {
+ II->setOperand(1, V);
+ return II;
+ }
+ break;
+ }
+
case Intrinsic::ppc_altivec_vperm:
// Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
if (ConstantPacked *Mask = dyn_cast<ConstantPacked>(II->getOperand(3))) {
if (AllEltsOk) {
// Cast the input vectors to byte vectors.
- Value *Op0 = InsertCastBefore(II->getOperand(1), Mask->getType(), CI);
- Value *Op1 = InsertCastBefore(II->getOperand(2), Mask->getType(), CI);
+ Value *Op0 = InsertCastBefore(Instruction::BitCast,
+ II->getOperand(1), Mask->getType(), CI);
+ Value *Op1 = InsertCastBefore(Instruction::BitCast,
+ II->getOperand(2), Mask->getType(), CI);
Value *Result = UndefValue::get(Op0->getType());
// Only extract each element once.
for (unsigned i = 0; i != 16; ++i) {
if (isa<UndefValue>(Mask->getOperand(i)))
continue;
- unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getRawValue();
+ unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
Idx &= 31; // Match the hardware behavior.
if (ExtractedElts[Idx] == 0) {
Instruction *Elt =
- new ExtractElementInst(Idx < 16 ? Op0 : Op1,
- ConstantUInt::get(Type::UIntTy, Idx&15),
- "tmp");
+ new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
InsertNewInstBefore(Elt, CI);
ExtractedElts[Idx] = Elt;
}
// Insert this value into the result vector.
- Result = new InsertElementInst(Result, ExtractedElts[Idx],
- ConstantUInt::get(Type::UIntTy, i),
- "tmp");
+ Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
InsertNewInstBefore(cast<Instruction>(Result), CI);
}
- return new CastInst(Result, CI.getType());
+ return CastInst::create(Instruction::BitCast, Result, CI.getType());
}
}
break;
Instruction *OldCall = CS.getInstruction();
// If the call and callee calling conventions don't match, this call must
// be unreachable, as the call is undefined.
- new StoreInst(ConstantBool::True,
+ new StoreInst(ConstantBool::getTrue(),
UndefValue::get(PointerType::get(Type::BoolTy)), OldCall);
if (!OldCall->use_empty())
OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
// This instruction is not reachable, just remove it. We insert a store to
// undef so that we know that this code is not reachable, despite the fact
// that we can't modify the CFG here.
- new StoreInst(ConstantBool::True,
+ new StoreInst(ConstantBool::getTrue(),
UndefValue::get(PointerType::get(Type::BoolTy)),
CS.getInstruction());
if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
// Don't break the CFG, insert a dummy cond branch.
new BranchInst(II->getNormalDest(), II->getUnwindDest(),
- ConstantBool::True, II);
+ ConstantBool::getTrue(), II);
}
return EraseInstFromFunction(*CS.getInstruction());
}
// If this cast does not effect the value passed through the varargs
// area, we can eliminate the use of the cast.
Value *Op = CI->getOperand(0);
- if (CI->getType()->isLosslesslyConvertibleTo(Op->getType())) {
+ if (CI->isLosslessCast()) {
*I = Op;
Changed = true;
}
bool InstCombiner::transformConstExprCastCall(CallSite CS) {
if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
- if (CE->getOpcode() != Instruction::Cast || !isa<Function>(CE->getOperand(0)))
+ if (CE->getOpcode() != Instruction::BitCast ||
+ !isa<Function>(CE->getOperand(0)))
return false;
Function *Callee = cast<Function>(CE->getOperand(0));
Instruction *Caller = CS.getInstruction();
// Check to see if we are changing the return type...
if (OldRetTy != FT->getReturnType()) {
if (Callee->isExternal() &&
- !(OldRetTy->isLosslesslyConvertibleTo(FT->getReturnType()) ||
+ !Caller->use_empty() &&
+ !(OldRetTy->canLosslesslyBitCastTo(FT->getReturnType()) ||
(isa<PointerType>(FT->getReturnType()) &&
- TD->getIntPtrType()->isLosslesslyConvertibleTo(OldRetTy)))
- && !Caller->use_empty())
+ TD->getIntPtrType()->canLosslesslyBitCastTo(OldRetTy)))
+ )
return false; // Cannot transform this return value...
// If the callsite is an invoke instruction, and the return value is used by
CallSite::arg_iterator AI = CS.arg_begin();
for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
const Type *ParamTy = FT->getParamType(i);
- bool isConvertible = (*AI)->getType()->isLosslesslyConvertibleTo(ParamTy);
+ const Type *ActTy = (*AI)->getType();
+ ConstantInt *c = dyn_cast<ConstantInt>(*AI);
+ //Either we can cast directly, or we can upconvert the argument
+ bool isConvertible = ActTy->canLosslesslyBitCastTo(ParamTy) ||
+ (ParamTy->isIntegral() && ActTy->isIntegral() &&
+ ParamTy->isSigned() == ActTy->isSigned() &&
+ ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize()) ||
+ (c && ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize() &&
+ c->getSExtValue() > 0);
if (Callee->isExternal() && !isConvertible) return false;
}
if ((*AI)->getType() == ParamTy) {
Args.push_back(*AI);
} else {
- Args.push_back(InsertNewInstBefore(new CastInst(*AI, ParamTy, "tmp"),
- *Caller));
+ Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
+ (*AI)->getType()->isSigned(), ParamTy, ParamTy->isSigned());
+ CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
+ Args.push_back(InsertNewInstBefore(NewCast, *Caller));
}
}
// If we are removing arguments to the function, emit an obnoxious warning...
if (FT->getNumParams() < NumActualArgs)
if (!FT->isVarArg()) {
- std::cerr << "WARNING: While resolving call to function '"
- << Callee->getName() << "' arguments were dropped!\n";
+ cerr << "WARNING: While resolving call to function '"
+ << Callee->getName() << "' arguments were dropped!\n";
} else {
// Add all of the arguments in their promoted form to the arg list...
for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
const Type *PTy = getPromotedType((*AI)->getType());
if (PTy != (*AI)->getType()) {
// Must promote to pass through va_arg area!
- Instruction *Cast = new CastInst(*AI, PTy, "tmp");
+ Instruction::CastOps opcode = CastInst::getCastOpcode(
+ *AI, (*AI)->getType()->isSigned(), PTy, PTy->isSigned());
+ Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
InsertNewInstBefore(Cast, *Caller);
Args.push_back(Cast);
} else {
Value *NV = NC;
if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
if (NV->getType() != Type::VoidTy) {
- NV = NC = new CastInst(NC, Caller->getType(), "tmp");
+ const Type *CallerTy = Caller->getType();
+ Instruction::CastOps opcode = CastInst::getCastOpcode(
+ NC, NC->getType()->isSigned(), CallerTy, CallerTy->isSigned());
+ NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
// If this is an invoke instruction, we should insert it after the first
// non-phi, instruction in the normal successor block.
return true;
}
+/// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
+/// and if a/b/c/d and the add's all have a single use, turn this into two phi's
+/// and a single binop.
+Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
+ Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
+ assert(isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst) ||
+ isa<GetElementPtrInst>(FirstInst) || isa<CmpInst>(FirstInst));
+ unsigned Opc = FirstInst->getOpcode();
+ Value *LHSVal = FirstInst->getOperand(0);
+ Value *RHSVal = FirstInst->getOperand(1);
+
+ const Type *LHSType = LHSVal->getType();
+ const Type *RHSType = RHSVal->getType();
+
+ // Scan to see if all operands are the same opcode, all have one use, and all
+ // kill their operands (i.e. the operands have one use).
+ for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
+ Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
+ if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
+ // Verify type of the LHS matches so we don't fold cmp's of different
+ // types or GEP's with different index types.
+ I->getOperand(0)->getType() != LHSType ||
+ I->getOperand(1)->getType() != RHSType)
+ return 0;
+
+ // If they are CmpInst instructions, check their predicates
+ if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
+ if (cast<CmpInst>(I)->getPredicate() !=
+ cast<CmpInst>(FirstInst)->getPredicate())
+ return 0;
+
+ // Keep track of which operand needs a phi node.
+ if (I->getOperand(0) != LHSVal) LHSVal = 0;
+ if (I->getOperand(1) != RHSVal) RHSVal = 0;
+ }
+
+ // Otherwise, this is safe to transform, determine if it is profitable.
+
+ // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
+ // Indexes are often folded into load/store instructions, so we don't want to
+ // hide them behind a phi.
+ if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
+ return 0;
+
+ Value *InLHS = FirstInst->getOperand(0);
+ Value *InRHS = FirstInst->getOperand(1);
+ PHINode *NewLHS = 0, *NewRHS = 0;
+ if (LHSVal == 0) {
+ NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
+ NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
+ NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
+ InsertNewInstBefore(NewLHS, PN);
+ LHSVal = NewLHS;
+ }
+
+ if (RHSVal == 0) {
+ NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
+ NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
+ NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
+ InsertNewInstBefore(NewRHS, PN);
+ RHSVal = NewRHS;
+ }
+
+ // Add all operands to the new PHIs.
+ for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
+ if (NewLHS) {
+ Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
+ NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
+ }
+ if (NewRHS) {
+ Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
+ NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
+ }
+ }
+
+ if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
+ return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
+ else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
+ return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
+ RHSVal);
+ else if (ShiftInst *SI = dyn_cast<ShiftInst>(FirstInst))
+ return new ShiftInst(SI->getOpcode(), LHSVal, RHSVal);
+ else {
+ assert(isa<GetElementPtrInst>(FirstInst));
+ return new GetElementPtrInst(LHSVal, RHSVal);
+ }
+}
+
+/// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
+/// of the block that defines it. This means that it must be obvious the value
+/// of the load is not changed from the point of the load to the end of the
+/// block it is in.
+static bool isSafeToSinkLoad(LoadInst *L) {
+ BasicBlock::iterator BBI = L, E = L->getParent()->end();
+
+ for (++BBI; BBI != E; ++BBI)
+ if (BBI->mayWriteToMemory())
+ return false;
+ return true;
+}
+
// FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
// operator and they all are only used by the PHI, PHI together their
// code size and simplifying code.
Constant *ConstantOp = 0;
const Type *CastSrcTy = 0;
+ bool isVolatile = false;
if (isa<CastInst>(FirstInst)) {
CastSrcTy = FirstInst->getOperand(0)->getType();
- } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst)) {
- // Can fold binop or shift if the RHS is a constant.
+ } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst) ||
+ isa<CmpInst>(FirstInst)) {
+ // Can fold binop, compare or shift here if the RHS is a constant,
+ // otherwise call FoldPHIArgBinOpIntoPHI.
ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
- if (ConstantOp == 0) return 0;
+ if (ConstantOp == 0)
+ return FoldPHIArgBinOpIntoPHI(PN);
+ } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
+ isVolatile = LI->isVolatile();
+ // We can't sink the load if the loaded value could be modified between the
+ // load and the PHI.
+ if (LI->getParent() != PN.getIncomingBlock(0) ||
+ !isSafeToSinkLoad(LI))
+ return 0;
+ } else if (isa<GetElementPtrInst>(FirstInst)) {
+ if (FirstInst->getNumOperands() == 2)
+ return FoldPHIArgBinOpIntoPHI(PN);
+ // Can't handle general GEPs yet.
+ return 0;
} else {
return 0; // Cannot fold this operation.
}
for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
- if (!I->hasOneUse() || I->getOpcode() != FirstInst->getOpcode())
+ if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
return 0;
if (CastSrcTy) {
if (I->getOperand(0)->getType() != CastSrcTy)
return 0; // Cast operation must match.
+ } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
+ // We can't sink the load if the loaded value could be modified between
+ // the load and the PHI.
+ if (LI->isVolatile() != isVolatile ||
+ LI->getParent() != PN.getIncomingBlock(i) ||
+ !isSafeToSinkLoad(LI))
+ return 0;
} else if (I->getOperand(1) != ConstantOp) {
return 0;
}
}
// Insert and return the new operation.
- if (isa<CastInst>(FirstInst))
- return new CastInst(PhiVal, PN.getType());
+ if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
+ return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
+ else if (isa<LoadInst>(FirstInst))
+ return new LoadInst(PhiVal, "", isVolatile);
else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
+ else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
+ return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
+ PhiVal, ConstantOp);
else
return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
PhiVal, ConstantOp);
// PHINode simplification
//
Instruction *InstCombiner::visitPHINode(PHINode &PN) {
+ // If LCSSA is around, don't mess with Phi nodes
+ if (mustPreserveAnalysisID(LCSSAID)) return 0;
+
if (Value *V = PN.hasConstantValue())
return ReplaceInstUsesWith(PN, V);
- // If the only user of this instruction is a cast instruction, and all of the
- // incoming values are constants, change this PHI to merge together the casted
- // constants.
- if (PN.hasOneUse())
- if (CastInst *CI = dyn_cast<CastInst>(PN.use_back()))
- if (CI->getType() != PN.getType()) { // noop casts will be folded
- bool AllConstant = true;
- for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
- if (!isa<Constant>(PN.getIncomingValue(i))) {
- AllConstant = false;
- break;
- }
- if (AllConstant) {
- // Make a new PHI with all casted values.
- PHINode *New = new PHINode(CI->getType(), PN.getName(), &PN);
- for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
- Constant *OldArg = cast<Constant>(PN.getIncomingValue(i));
- New->addIncoming(ConstantExpr::getCast(OldArg, New->getType()),
- PN.getIncomingBlock(i));
- }
-
- // Update the cast instruction.
- CI->setOperand(0, New);
- WorkList.push_back(CI); // revisit the cast instruction to fold.
- WorkList.push_back(New); // Make sure to revisit the new Phi
- return &PN; // PN is now dead!
- }
- }
-
// If all PHI operands are the same operation, pull them through the PHI,
// reducing code size.
if (isa<Instruction>(PN.getIncomingValue(0)) &&
return 0;
}
-static Value *InsertSignExtendToPtrTy(Value *V, const Type *DTy,
- Instruction *InsertPoint,
- InstCombiner *IC) {
- unsigned PS = IC->getTargetData().getPointerSize();
- const Type *VTy = V->getType();
- if (!VTy->isSigned() && VTy->getPrimitiveSize() < PS)
- // We must insert a cast to ensure we sign-extend.
- V = IC->InsertNewInstBefore(new CastInst(V, VTy->getSignedVersion(),
- V->getName()), *InsertPoint);
- return IC->InsertNewInstBefore(new CastInst(V, DTy, V->getName()),
- *InsertPoint);
+static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
+ Instruction *InsertPoint,
+ InstCombiner *IC) {
+ unsigned PtrSize = DTy->getPrimitiveSize();
+ unsigned VTySize = V->getType()->getPrimitiveSize();
+ // We must cast correctly to the pointer type. Ensure that we
+ // sign extend the integer value if it is smaller as this is
+ // used for address computation.
+ Instruction::CastOps opcode =
+ (VTySize < PtrSize ? Instruction::SExt :
+ (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
+ return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
}
}
} else if (SrcTy->getPrimitiveSize() < DestTy->getPrimitiveSize() &&
SrcTy->getPrimitiveSize() == 4) {
- // We can always eliminate a cast from int to [u]long. We can
- // eliminate a cast from uint to [u]long iff the target is a 32-bit
- // pointer target.
- if (SrcTy->isSigned() ||
- SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
+ // We can eliminate a cast from [u]int to [u]long iff the target
+ // is a 32-bit pointer target.
+ if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
MadeChange = true;
GEP.setOperand(i, Src);
}
Value *Op = GEP.getOperand(i);
if (Op->getType()->getPrimitiveSize() > TD->getPointerSize())
if (Constant *C = dyn_cast<Constant>(Op)) {
- GEP.setOperand(i, ConstantExpr::getCast(C,
- TD->getIntPtrType()->getSignedVersion()));
+ GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
MadeChange = true;
} else {
- Op = InsertNewInstBefore(new CastInst(Op, TD->getIntPtrType(),
- Op->getName()), GEP);
+ Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
+ GEP);
GEP.setOperand(i, Op);
MadeChange = true;
}
-
// If this is a constant idx, make sure to canonicalize it to be a signed
// operand, otherwise CSE and other optimizations are pessimized.
- if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op)) {
- GEP.setOperand(i, ConstantExpr::getCast(CUI,
- CUI->getType()->getSignedVersion()));
- MadeChange = true;
- }
+ if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op))
+ if (CUI->getType()->isUnsigned()) {
+ GEP.setOperand(i,
+ ConstantExpr::getBitCast(CUI, CUI->getType()->getSignedVersion()));
+ MadeChange = true;
+ }
}
if (MadeChange) return &GEP;
// target's pointer size.
if (SO1->getType() != GO1->getType()) {
if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
- SO1 = ConstantExpr::getCast(SO1C, GO1->getType());
+ SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
} else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
- GO1 = ConstantExpr::getCast(GO1C, SO1->getType());
+ GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
} else {
unsigned PS = TD->getPointerSize();
if (SO1->getType()->getPrimitiveSize() == PS) {
// Convert GO1 to SO1's type.
- GO1 = InsertSignExtendToPtrTy(GO1, SO1->getType(), &GEP, this);
+ GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
} else if (GO1->getType()->getPrimitiveSize() == PS) {
// Convert SO1 to GO1's type.
- SO1 = InsertSignExtendToPtrTy(SO1, GO1->getType(), &GEP, this);
+ SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
} else {
const Type *PT = TD->getIntPtrType();
- SO1 = InsertSignExtendToPtrTy(SO1, PT, &GEP, this);
- GO1 = InsertSignExtendToPtrTy(GO1, PT, &GEP, this);
+ SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
+ GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
}
}
}
// Replace all uses of the GEP with the new constexpr...
return ReplaceInstUsesWith(GEP, CE);
}
- } else if (Value *X = isCast(PtrOp)) { // Is the operand a cast?
+ } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
if (!isa<PointerType>(X->getType())) {
// Not interesting. Source pointer must be a cast from pointer.
} else if (HasZeroPointerIndex) {
Value *V = InsertNewInstBefore(
new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
GEP.getOperand(1), GEP.getName()), GEP);
- return new CastInst(V, GEP.getType());
+ // V and GEP are both pointer types --> BitCast
+ return new BitCastInst(V, GEP.getType());
}
// Transform things like:
} else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
if (Inst->getOpcode() == Instruction::Shl &&
isa<ConstantInt>(Inst->getOperand(1))) {
- unsigned ShAmt =cast<ConstantUInt>(Inst->getOperand(1))->getValue();
- if (Inst->getType()->isSigned())
- Scale = ConstantSInt::get(Inst->getType(), 1ULL << ShAmt);
- else
- Scale = ConstantUInt::get(Inst->getType(), 1ULL << ShAmt);
+ unsigned ShAmt =
+ cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
+ Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
NewIdx = Inst->getOperand(0);
} else if (Inst->getOpcode() == Instruction::Mul &&
isa<ConstantInt>(Inst->getOperand(1))) {
// If the index will be to exactly the right offset with the scale taken
// out, perform the transformation.
- if (Scale && Scale->getRawValue() % ArrayEltSize == 0) {
- if (ConstantSInt *C = dyn_cast<ConstantSInt>(Scale))
- Scale = ConstantSInt::get(C->getType(),
- (int64_t)C->getRawValue() /
- (int64_t)ArrayEltSize);
- else
- Scale = ConstantUInt::get(Scale->getType(),
- Scale->getRawValue() / ArrayEltSize);
- if (Scale->getRawValue() != 1) {
- Constant *C = ConstantExpr::getCast(Scale, NewIdx->getType());
+ if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
+ if (isa<ConstantInt>(Scale))
+ Scale = ConstantInt::get(Scale->getType(),
+ Scale->getZExtValue() / ArrayEltSize);
+ if (Scale->getZExtValue() != 1) {
+ Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
+ true /*SExt*/);
Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
NewIdx = InsertNewInstBefore(Sc, GEP);
}
// Insert the new GEP instruction.
- Instruction *Idx =
+ Instruction *NewGEP =
new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
NewIdx, GEP.getName());
- Idx = InsertNewInstBefore(Idx, GEP);
- return new CastInst(Idx, GEP.getType());
+ NewGEP = InsertNewInstBefore(NewGEP, GEP);
+ // The NewGEP must be pointer typed, so must the old one -> BitCast
+ return new BitCastInst(NewGEP, GEP.getType());
}
}
}
Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
// Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
if (AI.isArrayAllocation()) // Check C != 1
- if (const ConstantUInt *C = dyn_cast<ConstantUInt>(AI.getArraySize())) {
- const Type *NewTy = ArrayType::get(AI.getAllocatedType(), C->getValue());
+ if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
+ const Type *NewTy =
+ ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
AllocationInst *New = 0;
// Create and insert the replacement instruction...
// free undef -> unreachable.
if (isa<UndefValue>(Op)) {
// Insert a new store to null because we cannot modify the CFG here.
- new StoreInst(ConstantBool::True,
+ new StoreInst(ConstantBool::getTrue(),
UndefValue::get(PointerType::get(Type::BoolTy)), &FI);
return EraseInstFromFunction(FI);
}
CI->getName(),
LI.isVolatile()),LI);
// Now cast the result of the load.
- return new CastInst(NewLoad, LI.getType());
+ return new BitCastInst(NewLoad, LI.getType());
}
}
}
Value *Op = LI.getOperand(0);
// load (cast X) --> cast (load X) iff safe
- if (CastInst *CI = dyn_cast<CastInst>(Op))
+ if (isa<CastInst>(Op))
if (Instruction *Res = InstCombineLoadCast(*this, LI))
return Res;
return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
}
- } else if (CE->getOpcode() == Instruction::Cast) {
+ } else if (CE->isCast()) {
if (Instruction *Res = InstCombineLoadCast(*this, LI))
return Res;
}
LI.setOperand(0, SI->getOperand(1));
return &LI;
}
-
- } else if (PHINode *PN = dyn_cast<PHINode>(Op)) {
- // load (phi (&V1, &V2, &V3)) --> phi(load &V1, load &V2, load &V3)
- bool Safe = PN->getParent() == LI.getParent();
-
- // Scan all of the instructions between the PHI and the load to make
- // sure there are no instructions that might possibly alter the value
- // loaded from the PHI.
- if (Safe) {
- BasicBlock::iterator I = &LI;
- for (--I; !isa<PHINode>(I); --I)
- if (isa<StoreInst>(I) || isa<CallInst>(I)) {
- Safe = false;
- break;
- }
- }
-
- for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e && Safe; ++i)
- if (!isSafeToLoadUnconditionally(PN->getIncomingValue(i),
- PN->getIncomingBlock(i)->getTerminator()))
- Safe = false;
-
- if (Safe) {
- // Create the PHI.
- PHINode *NewPN = new PHINode(LI.getType(), PN->getName());
- InsertNewInstBefore(NewPN, *PN);
- std::map<BasicBlock*,Value*> LoadMap; // Don't insert duplicate loads
-
- for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
- BasicBlock *BB = PN->getIncomingBlock(i);
- Value *&TheLoad = LoadMap[BB];
- if (TheLoad == 0) {
- Value *InVal = PN->getIncomingValue(i);
- TheLoad = InsertNewInstBefore(new LoadInst(InVal,
- InVal->getName()+".val"),
- *BB->getTerminator());
- }
- NewPN->addIncoming(TheLoad, BB);
- }
- return ReplaceInstUsesWith(LI, NewPN);
- }
}
}
return 0;
// the same size. Instead of casting the pointer before the store, cast
// the value to be stored.
Value *NewCast;
- if (Constant *C = dyn_cast<Constant>(SI.getOperand(0)))
- NewCast = ConstantExpr::getCast(C, SrcPTy);
+ Instruction::CastOps opcode = Instruction::BitCast;
+ Value *SIOp0 = SI.getOperand(0);
+ if (isa<PointerType>(SrcPTy)) {
+ if (SIOp0->getType()->isIntegral())
+ opcode = Instruction::IntToPtr;
+ } else if (SrcPTy->isIntegral()) {
+ if (isa<PointerType>(SIOp0->getType()))
+ opcode = Instruction::PtrToInt;
+ }
+ if (Constant *C = dyn_cast<Constant>(SIOp0))
+ NewCast = ConstantExpr::getCast(opcode, C, SrcPTy);
else
- NewCast = IC.InsertNewInstBefore(new CastInst(SI.getOperand(0),
- SrcPTy,
- SI.getOperand(0)->getName()+".c"), SI);
-
+ NewCast = IC.InsertNewInstBefore(
+ CastInst::create(opcode, SIOp0, SrcPTy, SIOp0->getName()+".c"), SI);
return new StoreInst(NewCast, CastOp);
}
}
// If the pointer destination is a cast, see if we can fold the cast into the
// source instead.
- if (CastInst *CI = dyn_cast<CastInst>(Ptr))
+ if (isa<CastInst>(Ptr))
if (Instruction *Res = InstCombineStoreToCast(*this, SI))
return Res;
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
- if (CE->getOpcode() == Instruction::Cast)
+ if (CE->isCast())
if (Instruction *Res = InstCombineStoreToCast(*this, SI))
return Res;
return &BI;
}
- // Cannonicalize setne -> seteq
- Instruction::BinaryOps Op; Value *Y;
- if (match(&BI, m_Br(m_SetCond(Op, m_Value(X), m_Value(Y)),
+ // Cannonicalize fcmp_one -> fcmp_oeq
+ FCmpInst::Predicate FPred; Value *Y;
+ if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
+ TrueDest, FalseDest)))
+ if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
+ FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
+ FCmpInst *I = cast<FCmpInst>(BI.getCondition());
+ std::string Name = I->getName(); I->setName("");
+ FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
+ Value *NewSCC = new FCmpInst(NewPred, X, Y, Name, I);
+ // Swap Destinations and condition...
+ BI.setCondition(NewSCC);
+ BI.setSuccessor(0, FalseDest);
+ BI.setSuccessor(1, TrueDest);
+ removeFromWorkList(I);
+ I->getParent()->getInstList().erase(I);
+ WorkList.push_back(cast<Instruction>(NewSCC));
+ return &BI;
+ }
+
+ // Cannonicalize icmp_ne -> icmp_eq
+ ICmpInst::Predicate IPred;
+ if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
TrueDest, FalseDest)))
- if ((Op == Instruction::SetNE || Op == Instruction::SetLE ||
- Op == Instruction::SetGE) && BI.getCondition()->hasOneUse()) {
- SetCondInst *I = cast<SetCondInst>(BI.getCondition());
+ if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
+ IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
+ IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
+ ICmpInst *I = cast<ICmpInst>(BI.getCondition());
std::string Name = I->getName(); I->setName("");
- Instruction::BinaryOps NewOpcode = SetCondInst::getInverseCondition(Op);
- Value *NewSCC = BinaryOperator::create(NewOpcode, X, Y, Name, I);
+ ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
+ Value *NewSCC = new ICmpInst(NewPred, X, Y, Name, I);
// Swap Destinations and condition...
BI.setCondition(NewSCC);
BI.setSuccessor(0, FalseDest);
(CheapToScalarize(BO->getOperand(0), isConstant) ||
CheapToScalarize(BO->getOperand(1), isConstant)))
return true;
+ if (CmpInst *CI = dyn_cast<CmpInst>(I))
+ if (CI->hasOneUse() &&
+ (CheapToScalarize(CI->getOperand(0), isConstant) ||
+ CheapToScalarize(CI->getOperand(1), isConstant)))
+ return true;
return false;
}
if (isa<UndefValue>(CP->getOperand(i)))
Result.push_back(NElts*2); // undef -> 8
else
- Result.push_back(cast<ConstantUInt>(CP->getOperand(i))->getValue());
+ Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
return Result;
}
return CP->getOperand(EltNo);
else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
// If this is an insert to a variable element, we don't know what it is.
- if (!isa<ConstantUInt>(III->getOperand(2))) return 0;
- unsigned IIElt = cast<ConstantUInt>(III->getOperand(2))->getValue();
+ if (!isa<ConstantInt>(III->getOperand(2)))
+ return 0;
+ unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
// If this is an insert to the element we are looking for, return the
// inserted value.
- if (EltNo == IIElt) return III->getOperand(1);
+ if (EltNo == IIElt)
+ return III->getOperand(1);
// Otherwise, the insertelement doesn't modify the value, recurse on its
// vector input.
// If extracting a specified index from the vector, see if we can recursively
// find a previously computed scalar that was inserted into the vector.
- if (ConstantUInt *IdxC = dyn_cast<ConstantUInt>(EI.getOperand(1))) {
- if (Value *Elt = FindScalarElement(EI.getOperand(0), IdxC->getValue()))
+ if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
+ // This instruction only demands the single element from the input vector.
+ // If the input vector has a single use, simplify it based on this use
+ // property.
+ uint64_t IndexVal = IdxC->getZExtValue();
+ if (EI.getOperand(0)->hasOneUse()) {
+ uint64_t UndefElts;
+ if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
+ 1 << IndexVal,
+ UndefElts)) {
+ EI.setOperand(0, V);
+ return &EI;
+ }
+ }
+
+ if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
return ReplaceInstUsesWith(EI, Elt);
}
InsertNewInstBefore(newEI1, EI);
return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
}
- } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
- Value *Ptr = InsertCastBefore(I->getOperand(0),
+ } else if (isa<LoadInst>(I)) {
+ Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
PointerType::get(EI.getType()), EI);
GetElementPtrInst *GEP =
- new GetElementPtrInst(Ptr, EI.getOperand(1),
- I->getName() + ".gep");
+ new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
InsertNewInstBefore(GEP, EI);
return new LoadInst(GEP);
}
} else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
// If this is extracting an element from a shufflevector, figure out where
// it came from and extract from the appropriate input element instead.
- if (ConstantUInt *Elt = dyn_cast<ConstantUInt>(EI.getOperand(1))) {
- unsigned SrcIdx = getShuffleMask(SVI)[Elt->getValue()];
+ if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
+ unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
Value *Src;
if (SrcIdx < SVI->getType()->getNumElements())
Src = SVI->getOperand(0);
} else {
return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
}
- return new ExtractElementInst(Src,
- ConstantUInt::get(Type::UIntTy, SrcIdx));
+ return new ExtractElementInst(Src, SrcIdx);
}
}
}
return true;
} else if (V == LHS) {
for (unsigned i = 0; i != NumElts; ++i)
- Mask.push_back(ConstantUInt::get(Type::UIntTy, i));
+ Mask.push_back(ConstantInt::get(Type::UIntTy, i));
return true;
} else if (V == RHS) {
for (unsigned i = 0; i != NumElts; ++i)
- Mask.push_back(ConstantUInt::get(Type::UIntTy, i+NumElts));
+ Mask.push_back(ConstantInt::get(Type::UIntTy, i+NumElts));
return true;
} else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
// If this is an insert of an extract from some other vector, include it.
if (!isa<ConstantInt>(IdxOp))
return false;
- unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getRawValue();
+ unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
// Okay, we can handle this if the vector we are insertinting into is
if (isa<ConstantInt>(EI->getOperand(1)) &&
EI->getOperand(0)->getType() == V->getType()) {
unsigned ExtractedIdx =
- cast<ConstantInt>(EI->getOperand(1))->getRawValue();
+ cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
// This must be extracting from either LHS or RHS.
if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
// If so, update the mask to reflect the inserted value.
if (EI->getOperand(0) == LHS) {
Mask[InsertedIdx & (NumElts-1)] =
- ConstantUInt::get(Type::UIntTy, ExtractedIdx);
+ ConstantInt::get(Type::UIntTy, ExtractedIdx);
} else {
assert(EI->getOperand(0) == RHS);
Mask[InsertedIdx & (NumElts-1)] =
- ConstantUInt::get(Type::UIntTy, ExtractedIdx+NumElts);
+ ConstantInt::get(Type::UIntTy, ExtractedIdx+NumElts);
}
return true;
Mask.assign(NumElts, UndefValue::get(Type::UIntTy));
return V;
} else if (isa<ConstantAggregateZero>(V)) {
- Mask.assign(NumElts, ConstantUInt::get(Type::UIntTy, 0));
+ Mask.assign(NumElts, ConstantInt::get(Type::UIntTy, 0));
return V;
} else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
// If this is an insert of an extract from some other vector, include it.
if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
EI->getOperand(0)->getType() == V->getType()) {
unsigned ExtractedIdx =
- cast<ConstantInt>(EI->getOperand(1))->getRawValue();
- unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getRawValue();
+ cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
+ unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
// Either the extracted from or inserted into vector must be RHSVec,
// otherwise we'd end up with a shuffle of three inputs.
RHS = EI->getOperand(0);
Value *V = CollectShuffleElements(VecOp, Mask, RHS);
Mask[InsertedIdx & (NumElts-1)] =
- ConstantUInt::get(Type::UIntTy, NumElts+ExtractedIdx);
+ ConstantInt::get(Type::UIntTy, NumElts+ExtractedIdx);
return V;
}
// Everything but the extracted element is replaced with the RHS.
for (unsigned i = 0; i != NumElts; ++i) {
if (i != InsertedIdx)
- Mask[i] = ConstantUInt::get(Type::UIntTy, NumElts+i);
+ Mask[i] = ConstantInt::get(Type::UIntTy, NumElts+i);
}
return V;
}
// Otherwise, can't do anything fancy. Return an identity vector.
for (unsigned i = 0; i != NumElts; ++i)
- Mask.push_back(ConstantUInt::get(Type::UIntTy, i));
+ Mask.push_back(ConstantInt::get(Type::UIntTy, i));
return V;
}
if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
EI->getOperand(0)->getType() == IE.getType()) {
unsigned NumVectorElts = IE.getType()->getNumElements();
- unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getRawValue();
- unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getRawValue();
+ unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
+ unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
if (ExtractedIdx >= NumVectorElts) // Out of range extract.
return ReplaceInstUsesWith(IE, VecOp);
Mask.assign(NumVectorElts, UndefValue::get(Type::UIntTy));
else {
assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
- Mask.assign(NumVectorElts, ConstantUInt::get(Type::UIntTy,
+ Mask.assign(NumVectorElts, ConstantInt::get(Type::UIntTy,
NumVectorElts));
}
- Mask[InsertedIdx] = ConstantUInt::get(Type::UIntTy, ExtractedIdx);
+ Mask[InsertedIdx] = ConstantInt::get(Type::UIntTy, ExtractedIdx);
return new ShuffleVectorInst(EI->getOperand(0), VecOp,
ConstantPacked::get(Mask));
}
bool MadeChange = false;
+ // Undefined shuffle mask -> undefined value.
if (isa<UndefValue>(SVI.getOperand(2)))
return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
Mask[i] = 2*e; // Turn into undef.
else
Mask[i] &= (e-1); // Force to LHS.
- Elts.push_back(ConstantUInt::get(Type::UIntTy, Mask[i]));
+ Elts.push_back(ConstantInt::get(Type::UIntTy, Mask[i]));
}
}
SVI.setOperand(0, SVI.getOperand(1));
if (NewMask[i] >= e*2) {
Elts.push_back(UndefValue::get(Type::UIntTy));
} else {
- Elts.push_back(ConstantUInt::get(Type::UIntTy, NewMask[i]));
+ Elts.push_back(ConstantInt::get(Type::UIntTy, NewMask[i]));
}
}
return new ShuffleVectorInst(LHSSVI->getOperand(0),
}
/// OptimizeConstantExpr - Given a constant expression and target data layout
-/// information, symbolically evaluation the constant expr to something simpler
+/// information, symbolically evaluate the constant expr to something simpler
/// if possible.
static Constant *OptimizeConstantExpr(ConstantExpr *CE, const TargetData *TD) {
if (!TD) return CE;
if (isFoldableGEP) {
std::vector<Value*> Ops(CE->op_begin()+1, CE->op_end());
uint64_t Offset = TD->getIndexedOffset(Ptr->getType(), Ops);
- Constant *C = ConstantUInt::get(Type::ULongTy, Offset);
- C = ConstantExpr::getCast(C, TD->getIntPtrType());
- return ConstantExpr::getCast(C, CE->getType());
+ Constant *C = ConstantInt::get(TD->getIntPtrType(), Offset);
+ return ConstantExpr::getIntToPtr(C, CE->getType());
}
}
// DCE instruction if trivially dead.
if (isInstructionTriviallyDead(Inst)) {
++NumDeadInst;
- DEBUG(std::cerr << "IC: DCE: " << *Inst);
+ DOUT << "IC: DCE: " << *Inst;
Inst->eraseFromParent();
continue;
}
if (Constant *C = ConstantFoldInstruction(Inst)) {
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
C = OptimizeConstantExpr(CE, TD);
- DEBUG(std::cerr << "IC: ConstFold to: " << *C << " from: " << *Inst);
+ DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
Inst->replaceAllUsesWith(C);
++NumConstProp;
Inst->eraseFromParent();
while (Term != BB->begin()) { // Remove instrs bottom-up
BasicBlock::iterator I = Term; --I;
- DEBUG(std::cerr << "IC: DCE: " << *I);
+ DOUT << "IC: DCE: " << *I;
++NumDeadInst;
if (!I->use_empty())
AddUsesToWorkList(*I);
++NumDeadInst;
- DEBUG(std::cerr << "IC: DCE: " << *I);
+ DOUT << "IC: DCE: " << *I;
I->eraseFromParent();
removeFromWorkList(I);
if (Constant *C = ConstantFoldInstruction(I)) {
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
C = OptimizeConstantExpr(CE, TD);
- DEBUG(std::cerr << "IC: ConstFold to: " << *C << " from: " << *I);
+ DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
// Add operands to the worklist.
AddUsesToWorkList(*I);
++NumCombined;
// Should we replace the old instruction with a new one?
if (Result != I) {
- DEBUG(std::cerr << "IC: Old = " << *I
- << " New = " << *Result);
+ DOUT << "IC: Old = " << *I
+ << " New = " << *Result;
// Everything uses the new instruction now.
I->replaceAllUsesWith(Result);
// Erase the old instruction.
InstParent->getInstList().erase(I);
} else {
- DEBUG(std::cerr << "IC: MOD = " << *I);
+ DOUT << "IC: MOD = " << *I;
// If the instruction was modified, it's possible that it is now dead.
// if so, remove it.
projects / oota-llvm.git / blobdiff