/// X*Scale+Offset.
///
static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
- int &Offset) {
- assert(Val->getType()->isInteger(32) && "Unexpected allocation size type!");
+ uint64_t &Offset) {
if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
Offset = CI->getZExtValue();
Scale = 0;
- return ConstantInt::get(Type::getInt32Ty(Val->getContext()), 0);
+ return ConstantInt::get(Val->getType(), 0);
}
if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
if (I->getOpcode() == Instruction::Shl) {
// This is a value scaled by '1 << the shift amt'.
- Scale = 1U << RHS->getZExtValue();
+ Scale = UINT64_C(1) << RHS->getZExtValue();
Offset = 0;
return I->getOperand(0);
}
// See if we can satisfy the modulus by pulling a scale out of the array
// size argument.
unsigned ArraySizeScale;
- int ArrayOffset;
+ uint64_t ArrayOffset;
Value *NumElements = // See if the array size is a decomposable linear expr.
DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
if (Scale == 1) {
Amt = NumElements;
} else {
- Amt = ConstantInt::get(Type::getInt32Ty(CI.getContext()), Scale);
+ Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
// Insert before the alloca, not before the cast.
Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
}
- if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
- Value *Off = ConstantInt::get(Type::getInt32Ty(CI.getContext()),
+ if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
+ Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
Offset, true);
Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
}
}
-/// CanEvaluateInDifferentType - Return true if we can take the specified value
-/// and return it as type Ty without inserting any new casts and without
-/// changing the computed value. This is used by code that tries to decide
-/// whether promoting or shrinking integer operations to wider or smaller types
-/// will allow us to eliminate a truncate or extend.
-///
-/// This is a truncation operation if Ty is smaller than V->getType(), or an
-/// extension operation if Ty is larger.
-///
-/// If CastOpc is a truncation, then Ty will be a type smaller than V. We
-/// should return true if trunc(V) can be computed by computing V in the smaller
-/// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
-/// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
-/// efficiently truncated.
-///
-/// If CastOpc is a sext or zext, we are asking if the low bits of the value can
-/// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
-/// the final result.
-bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
- unsigned CastOpc,
- int &NumCastsRemoved){
- // We can always evaluate constants in another type.
- if (isa<Constant>(V))
- return true;
-
- Instruction *I = dyn_cast<Instruction>(V);
- if (!I) return false;
-
- const Type *OrigTy = V->getType();
-
- // If this is an extension or truncate, we can often eliminate it.
- if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
- // 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)) && I->hasOneUse())
- ++NumCastsRemoved;
- return true;
- }
- }
-
- // We can't extend or shrink something that has multiple uses: doing so would
- // require duplicating the instruction in general, which isn't profitable.
- if (!I->hasOneUse()) return false;
-
- unsigned Opc = I->getOpcode();
- switch (Opc) {
- case Instruction::Add:
- case Instruction::Sub:
- case Instruction::Mul:
- case Instruction::And:
- case Instruction::Or:
- case Instruction::Xor:
- // These operators can all arbitrarily be extended or truncated.
- return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
- NumCastsRemoved) &&
- CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
- NumCastsRemoved);
-
- case Instruction::UDiv:
- case Instruction::URem: {
- // UDiv and URem can be truncated if all the truncated bits are zero.
- uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
- uint32_t BitWidth = Ty->getScalarSizeInBits();
- if (BitWidth < OrigBitWidth) {
- APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
- if (MaskedValueIsZero(I->getOperand(0), Mask) &&
- MaskedValueIsZero(I->getOperand(1), Mask)) {
- return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
- NumCastsRemoved) &&
- CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
- NumCastsRemoved);
- }
- }
- break;
- }
- case Instruction::Shl:
- // If we are truncating the result of this SHL, and if it's a shift of a
- // constant amount, we can always perform a SHL in a smaller type.
- if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
- uint32_t BitWidth = Ty->getScalarSizeInBits();
- if (BitWidth < OrigTy->getScalarSizeInBits() &&
- CI->getLimitedValue(BitWidth) < BitWidth)
- return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
- NumCastsRemoved);
- }
- break;
- case Instruction::LShr:
- // If this is a truncate of a logical shr, we can truncate it to a smaller
- // lshr iff we know that the bits we would otherwise be shifting in are
- // already zeros.
- if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
- uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
- uint32_t BitWidth = Ty->getScalarSizeInBits();
- if (BitWidth < OrigBitWidth &&
- MaskedValueIsZero(I->getOperand(0),
- APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
- CI->getLimitedValue(BitWidth) < BitWidth) {
- return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
- NumCastsRemoved);
- }
- }
- break;
- case Instruction::ZExt:
- case Instruction::SExt:
- case Instruction::Trunc:
- // If this is the same kind of case as our original (e.g. zext+zext), we
- // can safely replace it. Note that replacing it does not reduce the number
- // of casts in the input.
- if (Opc == CastOpc)
- return true;
-
- // sext (zext ty1), ty2 -> zext ty2
- if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
- return true;
- break;
- case Instruction::Select: {
- SelectInst *SI = cast<SelectInst>(I);
- return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
- NumCastsRemoved) &&
- CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
- NumCastsRemoved);
- }
- case Instruction::PHI: {
- // We can change a phi if we can change all operands.
- PHINode *PN = cast<PHINode>(I);
- for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
- if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
- NumCastsRemoved))
- return false;
- return true;
- }
- default:
- // TODO: Can handle more cases here.
- break;
- }
-
- return false;
-}
/// EvaluateInDifferentType - Given an expression that
-/// CanEvaluateInDifferentType returns true for, actually insert the code to
-/// evaluate the expression.
+/// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually
+/// insert the code to evaluate the expression.
Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
bool isSigned) {
- if (Constant *C = dyn_cast<Constant>(V))
- return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
+ if (Constant *C = dyn_cast<Constant>(V)) {
+ C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
+ // If we got a constantexpr back, try to simplify it with TD info.
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
+ C = ConstantFoldConstantExpression(CE, TD);
+ return C;
+ }
// Otherwise, it must be an instruction.
Instruction *I = cast<Instruction>(V);
return I->getOperand(0);
// Otherwise, must be the same type of cast, so just reinsert a new one.
- Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),Ty);
+ // This also handles the case of zext(trunc(x)) -> zext(x).
+ Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
+ Opc == Instruction::SExt);
break;
case Instruction::Select: {
Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
return Instruction::CastOps(Res);
}
-/// 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.
-bool InstCombiner::ValueRequiresCast(Instruction::CastOps opcode,const Value *V,
- const Type *Ty) {
+/// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
+/// results in any code being generated and is interesting to optimize out. If
+/// the cast can be eliminated by some other simple transformation, we prefer
+/// to do the simplification first.
+bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
+ const Type *Ty) {
+ // Noop casts and casts of constants should be eliminated trivially.
if (V->getType() == Ty || isa<Constant>(V)) return false;
- // If this is another cast that can be eliminated, it isn't codegen either.
+ // If this is another cast that can be eliminated, we prefer to have it
+ // eliminated.
if (const CastInst *CI = dyn_cast<CastInst>(V))
- if (isEliminableCastPair(CI, opcode, Ty, TD))
+ if (isEliminableCastPair(CI, opc, Ty, TD))
return false;
+
+ // If this is a vector sext from a compare, then we don't want to break the
+ // idiom where each element of the extended vector is either zero or all ones.
+ if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
+ return false;
+
return true;
}
if (isa<PHINode>(Src)) {
// We don't do this if this would create a PHI node with an illegal type if
// it is currently legal.
- if (!isa<IntegerType>(Src->getType()) ||
- !isa<IntegerType>(CI.getType()) ||
+ if (!Src->getType()->isIntegerTy() ||
+ !CI.getType()->isIntegerTy() ||
ShouldChangeType(CI.getType(), Src->getType()))
if (Instruction *NV = FoldOpIntoPhi(CI))
return NV;
return 0;
}
-/// commonIntCastTransforms - This function implements the common transforms
-/// for trunc, zext, and sext.
-Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
- if (Instruction *Result = commonCastTransforms(CI))
- return Result;
-
- // See if we can simplify any instructions used by the LHS whose sole
- // purpose is to compute bits we don't care about.
- if (SimplifyDemandedInstructionBits(CI))
- return &CI;
+/// CanEvaluateTruncated - Return true if we can evaluate the specified
+/// expression tree as type Ty instead of its larger type, and arrive with the
+/// same value. This is used by code that tries to eliminate truncates.
+///
+/// Ty will always be a type smaller than V. We should return true if trunc(V)
+/// can be computed by computing V in the smaller type. If V is an instruction,
+/// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
+/// makes sense if x and y can be efficiently truncated.
+///
+/// This function works on both vectors and scalars.
+///
+static bool CanEvaluateTruncated(Value *V, const Type *Ty) {
+ // We can always evaluate constants in another type.
+ if (isa<Constant>(V))
+ return true;
- // If the source isn't an instruction or has more than one use then we
- // can't do anything more.
- Instruction *Src = dyn_cast<Instruction>(CI.getOperand(0));
- if (!Src || !Src->hasOneUse())
- return 0;
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (!I) return false;
- const Type *SrcTy = Src->getType();
- const Type *DestTy = CI.getType();
- uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
- uint32_t DestBitSize = DestTy->getScalarSizeInBits();
+ const Type *OrigTy = V->getType();
+
+ // If this is an extension from the dest type, we can eliminate it, even if it
+ // has multiple uses.
+ if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
+ I->getOperand(0)->getType() == Ty)
+ return true;
- // Attempt to propagate the cast into the instruction for int->int casts.
- int NumCastsRemoved = 0;
- // Only do this if the dest type is a simple type, don't convert the
- // expression tree to something weird like i93 unless the source is also
- // strange.
- if ((isa<VectorType>(DestTy) ||
- ShouldChangeType(Src->getType(), DestTy)) &&
- CanEvaluateInDifferentType(Src, DestTy,
- CI.getOpcode(), 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 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 = false;
- bool JustReplace = false;
- switch (CI.getOpcode()) {
- default:
- // All the others use floating point so we shouldn't actually
- // get here because of the check above.
- llvm_unreachable("Unknown cast type");
- case Instruction::Trunc:
- DoXForm = true;
- break;
- case Instruction::ZExt: {
- DoXForm = NumCastsRemoved >= 1;
-
- if (!DoXForm && 0) {
- // If it's unnecessary to issue an AND to clear the high bits, it's
- // always profitable to do this xform.
- Value *TryRes = EvaluateInDifferentType(Src, DestTy, false);
- APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
- if (MaskedValueIsZero(TryRes, Mask))
- return ReplaceInstUsesWith(CI, TryRes);
-
- if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
- if (TryI->use_empty())
- EraseInstFromFunction(*TryI);
- }
- break;
- }
- case Instruction::SExt: {
- DoXForm = NumCastsRemoved >= 2;
- if (!DoXForm && !isa<TruncInst>(Src) && 0) {
- // If we do not have to emit the truncate + sext pair, then it's always
- // profitable to do this xform.
- //
- // It's not safe to eliminate the trunc + sext pair if one of the
- // eliminated cast is a truncate. e.g.
- // t2 = trunc i32 t1 to i16
- // t3 = sext i16 t2 to i32
- // !=
- // i32 t1
- Value *TryRes = EvaluateInDifferentType(Src, DestTy, true);
- unsigned NumSignBits = ComputeNumSignBits(TryRes);
- if (NumSignBits > (DestBitSize - SrcBitSize))
- return ReplaceInstUsesWith(CI, TryRes);
-
- if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
- if (TryI->use_empty())
- EraseInstFromFunction(*TryI);
- }
- break;
- }
- }
-
- if (DoXForm) {
- DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type"
- " to avoid cast: " << CI);
- Value *Res = EvaluateInDifferentType(Src, DestTy,
- CI.getOpcode() == Instruction::SExt);
- if (JustReplace)
- // Just replace this cast with the result.
- return ReplaceInstUsesWith(CI, Res);
-
- assert(Res->getType() == DestTy);
- switch (CI.getOpcode()) {
- default: llvm_unreachable("Unknown cast type!");
- case Instruction::Trunc:
- // Just replace this cast with the result.
- return ReplaceInstUsesWith(CI, Res);
- case Instruction::ZExt: {
- assert(SrcBitSize < DestBitSize && "Not a zext?");
-
- // If the high bits are already zero, just replace this cast with the
- // result.
- APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
- if (MaskedValueIsZero(Res, Mask))
- return ReplaceInstUsesWith(CI, Res);
+ // We can't extend or shrink something that has multiple uses: doing so would
+ // require duplicating the instruction in general, which isn't profitable.
+ if (!I->hasOneUse()) return false;
- // We need to emit an AND to clear the high bits.
- Constant *C = ConstantInt::get(CI.getContext(),
- APInt::getLowBitsSet(DestBitSize, SrcBitSize));
- return BinaryOperator::CreateAnd(Res, C);
- }
- case Instruction::SExt: {
- // If the high bits are already filled with sign bit, just replace this
- // cast with the result.
- unsigned NumSignBits = ComputeNumSignBits(Res);
- if (NumSignBits > (DestBitSize - SrcBitSize))
- return ReplaceInstUsesWith(CI, Res);
+ unsigned Opc = I->getOpcode();
+ switch (Opc) {
+ case Instruction::Add:
+ case Instruction::Sub:
+ case Instruction::Mul:
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor:
+ // These operators can all arbitrarily be extended or truncated.
+ return CanEvaluateTruncated(I->getOperand(0), Ty) &&
+ CanEvaluateTruncated(I->getOperand(1), Ty);
- // We need to emit a cast to truncate, then a cast to sext.
- return new SExtInst(Builder->CreateTrunc(Res, Src->getType()), DestTy);
+ case Instruction::UDiv:
+ case Instruction::URem: {
+ // UDiv and URem can be truncated if all the truncated bits are zero.
+ uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
+ uint32_t BitWidth = Ty->getScalarSizeInBits();
+ if (BitWidth < OrigBitWidth) {
+ APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
+ if (MaskedValueIsZero(I->getOperand(0), Mask) &&
+ MaskedValueIsZero(I->getOperand(1), Mask)) {
+ return CanEvaluateTruncated(I->getOperand(0), Ty) &&
+ CanEvaluateTruncated(I->getOperand(1), Ty);
}
+ }
+ break;
+ }
+ case Instruction::Shl:
+ // If we are truncating the result of this SHL, and if it's a shift of a
+ // constant amount, we can always perform a SHL in a smaller type.
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ uint32_t BitWidth = Ty->getScalarSizeInBits();
+ if (CI->getLimitedValue(BitWidth) < BitWidth)
+ return CanEvaluateTruncated(I->getOperand(0), Ty);
+ }
+ break;
+ case Instruction::LShr:
+ // If this is a truncate of a logical shr, we can truncate it to a smaller
+ // lshr iff we know that the bits we would otherwise be shifting in are
+ // already zeros.
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
+ uint32_t BitWidth = Ty->getScalarSizeInBits();
+ if (MaskedValueIsZero(I->getOperand(0),
+ APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
+ CI->getLimitedValue(BitWidth) < BitWidth) {
+ return CanEvaluateTruncated(I->getOperand(0), Ty);
}
}
+ break;
+ case Instruction::Trunc:
+ // trunc(trunc(x)) -> trunc(x)
+ return true;
+ case Instruction::ZExt:
+ case Instruction::SExt:
+ // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
+ // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
+ return true;
+ case Instruction::Select: {
+ SelectInst *SI = cast<SelectInst>(I);
+ return CanEvaluateTruncated(SI->getTrueValue(), Ty) &&
+ CanEvaluateTruncated(SI->getFalseValue(), Ty);
+ }
+ case Instruction::PHI: {
+ // We can change a phi if we can change all operands. Note that we never
+ // get into trouble with cyclic PHIs here because we only consider
+ // instructions with a single use.
+ PHINode *PN = cast<PHINode>(I);
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
+ if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty))
+ return false;
+ return true;
+ }
+ default:
+ // TODO: Can handle more cases here.
+ break;
}
- return 0;
+ return false;
}
Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
- if (Instruction *Result = commonIntCastTransforms(CI))
+ if (Instruction *Result = commonCastTransforms(CI))
return Result;
+ // See if we can simplify any instructions used by the input whose sole
+ // purpose is to compute bits we don't care about.
+ if (SimplifyDemandedInstructionBits(CI))
+ return &CI;
+
Value *Src = CI.getOperand(0);
- const Type *DestTy = CI.getType();
+ const Type *DestTy = CI.getType(), *SrcTy = Src->getType();
+
+ // Attempt to truncate the entire input expression tree to the destination
+ // type. Only do this if the dest type is a simple type, don't convert the
+ // expression tree to something weird like i93 unless the source is also
+ // strange.
+ if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
+ CanEvaluateTruncated(Src, DestTy)) {
+
+ // If this cast is a truncate, evaluting in a different type always
+ // eliminates the cast, so it is always a win.
+ DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
+ " to avoid cast: " << CI << '\n');
+ Value *Res = EvaluateInDifferentType(Src, DestTy, false);
+ assert(Res->getType() == DestTy);
+ return ReplaceInstUsesWith(CI, Res);
+ }
// Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
if (DestTy->getScalarSizeInBits() == 1) {
Value *Zero = Constant::getNullValue(Src->getType());
return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
}
+
+ // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
+ Value *A = 0; ConstantInt *Cst = 0;
+ if (match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst))) &&
+ Src->hasOneUse()) {
+ // We have three types to worry about here, the type of A, the source of
+ // the truncate (MidSize), and the destination of the truncate. We know that
+ // ASize < MidSize and MidSize > ResultSize, but don't know the relation
+ // between ASize and ResultSize.
+ unsigned ASize = A->getType()->getPrimitiveSizeInBits();
+
+ // If the shift amount is larger than the size of A, then the result is
+ // known to be zero because all the input bits got shifted out.
+ if (Cst->getZExtValue() >= ASize)
+ return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
+
+ // Since we're doing an lshr and a zero extend, and know that the shift
+ // amount is smaller than ASize, it is always safe to do the shift in A's
+ // type, then zero extend or truncate to the result.
+ Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
+ Shift->takeName(Src);
+ return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
+ }
return 0;
}
if (CI.getType() == In->getType())
return ReplaceInstUsesWith(CI, In);
- else
- return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
+ return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
}
}
}
return 0;
}
+/// CanEvaluateZExtd - Determine if the specified value can be computed in the
+/// specified wider type and produce the same low bits. If not, return false.
+///
+/// If this function returns true, it can also return a non-zero number of bits
+/// (in BitsToClear) which indicates that the value it computes is correct for
+/// the zero extend, but that the additional BitsToClear bits need to be zero'd
+/// out. For example, to promote something like:
+///
+/// %B = trunc i64 %A to i32
+/// %C = lshr i32 %B, 8
+/// %E = zext i32 %C to i64
+///
+/// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
+/// set to 8 to indicate that the promoted value needs to have bits 24-31
+/// cleared in addition to bits 32-63. Since an 'and' will be generated to
+/// clear the top bits anyway, doing this has no extra cost.
+///
+/// This function works on both vectors and scalars.
+static bool CanEvaluateZExtd(Value *V, const Type *Ty, unsigned &BitsToClear) {
+ BitsToClear = 0;
+ if (isa<Constant>(V))
+ return true;
+
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (!I) return false;
+
+ // If the input is a truncate from the destination type, we can trivially
+ // eliminate it, even if it has multiple uses.
+ // FIXME: This is currently disabled until codegen can handle this without
+ // pessimizing code, PR5997.
+ if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
+ return true;
+
+ // We can't extend or shrink something that has multiple uses: doing so would
+ // require duplicating the instruction in general, which isn't profitable.
+ if (!I->hasOneUse()) return false;
+
+ unsigned Opc = I->getOpcode(), Tmp;
+ switch (Opc) {
+ case Instruction::ZExt: // zext(zext(x)) -> zext(x).
+ case Instruction::SExt: // zext(sext(x)) -> sext(x).
+ case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
+ return true;
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor:
+ case Instruction::Add:
+ case Instruction::Sub:
+ case Instruction::Mul:
+ case Instruction::Shl:
+ if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) ||
+ !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp))
+ return false;
+ // These can all be promoted if neither operand has 'bits to clear'.
+ if (BitsToClear == 0 && Tmp == 0)
+ return true;
+
+ // If the operation is an AND/OR/XOR and the bits to clear are zero in the
+ // other side, BitsToClear is ok.
+ if (Tmp == 0 &&
+ (Opc == Instruction::And || Opc == Instruction::Or ||
+ Opc == Instruction::Xor)) {
+ // We use MaskedValueIsZero here for generality, but the case we care
+ // about the most is constant RHS.
+ unsigned VSize = V->getType()->getScalarSizeInBits();
+ if (MaskedValueIsZero(I->getOperand(1),
+ APInt::getHighBitsSet(VSize, BitsToClear)))
+ return true;
+ }
+
+ // Otherwise, we don't know how to analyze this BitsToClear case yet.
+ return false;
+
+ case Instruction::LShr:
+ // We can promote lshr(x, cst) if we can promote x. This requires the
+ // ultimate 'and' to clear out the high zero bits we're clearing out though.
+ if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear))
+ return false;
+ BitsToClear += Amt->getZExtValue();
+ if (BitsToClear > V->getType()->getScalarSizeInBits())
+ BitsToClear = V->getType()->getScalarSizeInBits();
+ return true;
+ }
+ // Cannot promote variable LSHR.
+ return false;
+ case Instruction::Select:
+ if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) ||
+ !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) ||
+ // TODO: If important, we could handle the case when the BitsToClear are
+ // known zero in the disagreeing side.
+ Tmp != BitsToClear)
+ return false;
+ return true;
+
+ case Instruction::PHI: {
+ // We can change a phi if we can change all operands. Note that we never
+ // get into trouble with cyclic PHIs here because we only consider
+ // instructions with a single use.
+ PHINode *PN = cast<PHINode>(I);
+ if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear))
+ return false;
+ for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
+ if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) ||
+ // TODO: If important, we could handle the case when the BitsToClear
+ // are known zero in the disagreeing input.
+ Tmp != BitsToClear)
+ return false;
+ return true;
+ }
+ default:
+ // TODO: Can handle more cases here.
+ return false;
+ }
+}
+
Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
+ // If this zero extend is only used by a truncate, let the truncate by
+ // eliminated before we try to optimize this zext.
+ if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
+ return 0;
+
// If one of the common conversion will work, do it.
- if (Instruction *Result = commonIntCastTransforms(CI))
+ if (Instruction *Result = commonCastTransforms(CI))
return Result;
+ // See if we can simplify any instructions used by the input whose sole
+ // purpose is to compute bits we don't care about.
+ if (SimplifyDemandedInstructionBits(CI))
+ return &CI;
+
Value *Src = CI.getOperand(0);
+ const Type *SrcTy = Src->getType(), *DestTy = CI.getType();
+
+ // Attempt to extend the entire input expression tree to the destination
+ // type. Only do this if the dest type is a simple type, don't convert the
+ // expression tree to something weird like i93 unless the source is also
+ // strange.
+ unsigned BitsToClear;
+ if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
+ CanEvaluateZExtd(Src, DestTy, BitsToClear)) {
+ assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
+ "Unreasonable BitsToClear");
+
+ // Okay, we can transform this! Insert the new expression now.
+ DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
+ " to avoid zero extend: " << CI);
+ Value *Res = EvaluateInDifferentType(Src, DestTy, false);
+ assert(Res->getType() == DestTy);
+
+ uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
+ uint32_t DestBitSize = DestTy->getScalarSizeInBits();
+
+ // If the high bits are already filled with zeros, just replace this
+ // cast with the result.
+ if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize,
+ DestBitSize-SrcBitsKept)))
+ return ReplaceInstUsesWith(CI, Res);
+
+ // We need to emit an AND to clear the high bits.
+ Constant *C = ConstantInt::get(Res->getType(),
+ APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
+ return BinaryOperator::CreateAnd(Res, C);
+ }
// 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 (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
+ // TODO: Subsume this into EvaluateInDifferentType.
+
// Get the sizes of the types involved. We know that the intermediate type
// will be smaller than A or C, but don't know the relation between A and C.
Value *A = CSrc->getOperand(0);
APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
return BinaryOperator::CreateAnd(Trunc,
ConstantInt::get(Trunc->getType(),
- AndValue));
+ AndValue));
}
}
// zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
Value *X;
- if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isInteger(1) &&
+ if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) &&
match(SrcI, m_Not(m_Value(X))) &&
(!X->hasOneUse() || !isa<CmpInst>(X))) {
Value *New = Builder->CreateZExt(X, CI.getType());
return 0;
}
+/// CanEvaluateSExtd - Return true if we can take the specified value
+/// and return it as type Ty without inserting any new casts and without
+/// changing the value of the common low bits. This is used by code that tries
+/// to promote integer operations to a wider types will allow us to eliminate
+/// the extension.
+///
+/// This function works on both vectors and scalars.
+///
+static bool CanEvaluateSExtd(Value *V, const Type *Ty) {
+ assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
+ "Can't sign extend type to a smaller type");
+ // If this is a constant, it can be trivially promoted.
+ if (isa<Constant>(V))
+ return true;
+
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (!I) return false;
+
+ // If this is a truncate from the dest type, we can trivially eliminate it,
+ // even if it has multiple uses.
+ // FIXME: This is currently disabled until codegen can handle this without
+ // pessimizing code, PR5997.
+ if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
+ return true;
+
+ // We can't extend or shrink something that has multiple uses: doing so would
+ // require duplicating the instruction in general, which isn't profitable.
+ if (!I->hasOneUse()) return false;
+
+ switch (I->getOpcode()) {
+ case Instruction::SExt: // sext(sext(x)) -> sext(x)
+ case Instruction::ZExt: // sext(zext(x)) -> zext(x)
+ case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
+ return true;
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor:
+ case Instruction::Add:
+ case Instruction::Sub:
+ case Instruction::Mul:
+ // These operators can all arbitrarily be extended if their inputs can.
+ return CanEvaluateSExtd(I->getOperand(0), Ty) &&
+ CanEvaluateSExtd(I->getOperand(1), Ty);
+
+ //case Instruction::Shl: TODO
+ //case Instruction::LShr: TODO
+
+ case Instruction::Select:
+ return CanEvaluateSExtd(I->getOperand(1), Ty) &&
+ CanEvaluateSExtd(I->getOperand(2), Ty);
+
+ case Instruction::PHI: {
+ // We can change a phi if we can change all operands. Note that we never
+ // get into trouble with cyclic PHIs here because we only consider
+ // instructions with a single use.
+ PHINode *PN = cast<PHINode>(I);
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
+ if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false;
+ return true;
+ }
+ default:
+ // TODO: Can handle more cases here.
+ break;
+ }
+
+ return false;
+}
+
Instruction *InstCombiner::visitSExt(SExtInst &CI) {
- if (Instruction *I = commonIntCastTransforms(CI))
+ // If this sign extend is only used by a truncate, let the truncate by
+ // eliminated before we try to optimize this zext.
+ if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
+ return 0;
+
+ if (Instruction *I = commonCastTransforms(CI))
return I;
+ // See if we can simplify any instructions used by the input whose sole
+ // purpose is to compute bits we don't care about.
+ if (SimplifyDemandedInstructionBits(CI))
+ return &CI;
+
Value *Src = CI.getOperand(0);
+ const Type *SrcTy = Src->getType(), *DestTy = CI.getType();
+
+ // Attempt to extend the entire input expression tree to the destination
+ // type. Only do this if the dest type is a simple type, don't convert the
+ // expression tree to something weird like i93 unless the source is also
+ // strange.
+ if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
+ CanEvaluateSExtd(Src, DestTy)) {
+ // Okay, we can transform this! Insert the new expression now.
+ DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
+ " to avoid sign extend: " << CI);
+ Value *Res = EvaluateInDifferentType(Src, DestTy, true);
+ assert(Res->getType() == DestTy);
+
+ uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
+ uint32_t DestBitSize = DestTy->getScalarSizeInBits();
+
+ // If the high bits are already filled with sign bit, just replace this
+ // cast with the result.
+ if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize)
+ return ReplaceInstUsesWith(CI, Res);
+
+ // We need to emit a shl + ashr to do the sign extend.
+ Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
+ return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
+ ShAmt);
+ }
+
+ // If this input is a trunc from our destination, then turn sext(trunc(x))
+ // into shifts.
+ if (TruncInst *TI = dyn_cast<TruncInst>(Src))
+ if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
+ uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
+ uint32_t DestBitSize = DestTy->getScalarSizeInBits();
+
+ // We need to emit a shl + ashr to do the sign extend.
+ Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
+ Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
+ return BinaryOperator::CreateAShr(Res, ShAmt);
+ }
- // Canonicalize sign-extend from i1 to a select.
- if (Src->getType()->isInteger(1))
- return SelectInst::Create(Src,
- Constant::getAllOnesValue(CI.getType()),
- Constant::getNullValue(CI.getType()));
-
- // See if the value being truncated is already sign extended. If so, just
- // eliminate the trunc/sext pair.
- if (Operator::getOpcode(Src) == Instruction::Trunc) {
- Value *Op = cast<User>(Src)->getOperand(0);
- unsigned OpBits = Op->getType()->getScalarSizeInBits();
- unsigned MidBits = Src->getType()->getScalarSizeInBits();
- unsigned DestBits = CI.getType()->getScalarSizeInBits();
- unsigned NumSignBits = ComputeNumSignBits(Op);
-
- if (OpBits == DestBits) {
- // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
- // bits, it is already ready.
- if (NumSignBits > DestBits-MidBits)
- return ReplaceInstUsesWith(CI, Op);
- } else if (OpBits < DestBits) {
- // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
- // bits, just sext from i32.
- if (NumSignBits > OpBits-MidBits)
- return new SExtInst(Op, CI.getType(), "tmp");
- } else {
- // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
- // bits, just truncate to i32.
- if (NumSignBits > OpBits-MidBits)
- return new TruncInst(Op, CI.getType(), "tmp");
+
+ // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
+ // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
+ {
+ ICmpInst::Predicate Pred; Value *CmpLHS; ConstantInt *CmpRHS;
+ if (match(Src, m_ICmp(Pred, m_Value(CmpLHS), m_ConstantInt(CmpRHS)))) {
+ // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
+ // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
+ if ((Pred == ICmpInst::ICMP_SLT && CmpRHS->isZero()) ||
+ (Pred == ICmpInst::ICMP_SGT && CmpRHS->isAllOnesValue())) {
+ Value *Sh = ConstantInt::get(CmpLHS->getType(),
+ CmpLHS->getType()->getScalarSizeInBits()-1);
+ Value *In = Builder->CreateAShr(CmpLHS, Sh, CmpLHS->getName()+".lobit");
+ if (In->getType() != CI.getType())
+ In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/, "tmp");
+
+ if (Pred == ICmpInst::ICMP_SGT)
+ In = Builder->CreateNot(In, In->getName()+".not");
+ return ReplaceInstUsesWith(CI, In);
+ }
+ }
+ }
+
+ // vector (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed.
+ if (const VectorType *VTy = dyn_cast<VectorType>(DestTy)) {
+ ICmpInst::Predicate Pred; Value *CmpLHS;
+ if (match(Src, m_ICmp(Pred, m_Value(CmpLHS), m_Zero()))) {
+ if (Pred == ICmpInst::ICMP_SLT && CmpLHS->getType() == DestTy) {
+ const Type *EltTy = VTy->getElementType();
+
+ // splat the shift constant to a cosntant vector
+ Constant *Sh = ConstantInt::get(EltTy, EltTy->getScalarSizeInBits()-1);
+ std::vector<Constant *> Elts(VTy->getNumElements(), Sh);
+ Constant *VSh = ConstantVector::get(Elts);
+
+ Value *In = Builder->CreateAShr(CmpLHS, VSh,CmpLHS->getName()+".lobit");
+ return ReplaceInstUsesWith(CI, In);
+ }
}
}
// %a = shl i32 %i, 30
// %d = ashr i32 %a, 30
Value *A = 0;
+ // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
ConstantInt *BA = 0, *CA = 0;
- if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
+ if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
m_ConstantInt(CA))) &&
- BA == CA && isa<TruncInst>(A)) {
- Value *I = cast<TruncInst>(A)->getOperand(0);
- if (I->getType() == CI.getType()) {
- unsigned MidSize = Src->getType()->getScalarSizeInBits();
- unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
- unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
- Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
- I = Builder->CreateShl(I, ShAmtV, CI.getName());
- return BinaryOperator::CreateAShr(I, ShAmtV);
- }
+ BA == CA && A->getType() == CI.getType()) {
+ unsigned MidSize = Src->getType()->getScalarSizeInBits();
+ unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
+ unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
+ Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
+ A = Builder->CreateShl(A, ShAmtV, CI.getName());
+ return BinaryOperator::CreateAShr(A, ShAmtV);
}
return 0;
break;
}
}
+
+ // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x)
+ // NOTE: This should be disabled by -fno-builtin-sqrt if we ever support it.
+ CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0));
+ if (Call && Call->getCalledFunction() &&
+ Call->getCalledFunction()->getName() == "sqrt" &&
+ Call->getNumArgOperands() == 1) {
+ CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0));
+ if (Arg && Arg->getOpcode() == Instruction::FPExt &&
+ CI.getType()->isFloatTy() &&
+ Call->getType()->isDoubleTy() &&
+ Arg->getType()->isDoubleTy() &&
+ Arg->getOperand(0)->getType()->isFloatTy()) {
+ Function *Callee = Call->getCalledFunction();
+ Module *M = CI.getParent()->getParent()->getParent();
+ Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf",
+ Callee->getAttributes(),
+ Builder->getFloatTy(),
+ Builder->getFloatTy(),
+ NULL);
+ CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0),
+ "sqrtfcall");
+ ret->setAttributes(Callee->getAttributes());
+
+
+ // Remove the old Call. With -fmath-errno, it won't get marked readnone.
+ Call->replaceAllUsesWith(UndefValue::get(Call->getType()));
+ EraseInstFromFunction(*Call);
+ return ret;
+ }
+ }
+
return 0;
}
}
Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
- // If the source integer type is larger than the intptr_t type for
- // this target, do a trunc to the intptr_t type, then inttoptr of it. This
- // allows the trunc to be exposed to other transforms. Don't do this for
- // extending inttoptr's, because we don't know if the target sign or zero
- // extends to pointers.
- if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() >
- TD->getPointerSizeInBits()) {
- Value *P = Builder->CreateTrunc(CI.getOperand(0),
- TD->getIntPtrType(CI.getContext()), "tmp");
- return new IntToPtrInst(P, CI.getType());
+ // If the source integer type is not the intptr_t type for this target, do a
+ // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
+ // cast to be exposed to other transforms.
+ if (TD) {
+ if (CI.getOperand(0)->getType()->getScalarSizeInBits() >
+ TD->getPointerSizeInBits()) {
+ Value *P = Builder->CreateTrunc(CI.getOperand(0),
+ TD->getIntPtrType(CI.getContext()), "tmp");
+ return new IntToPtrInst(P, CI.getType());
+ }
+ if (CI.getOperand(0)->getType()->getScalarSizeInBits() <
+ TD->getPointerSizeInBits()) {
+ Value *P = Builder->CreateZExt(CI.getOperand(0),
+ TD->getIntPtrType(CI.getContext()), "tmp");
+ return new IntToPtrInst(P, CI.getType());
+ }
}
if (Instruction *I = commonCastTransforms(CI))
}
Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
- // If the destination integer type is smaller than the intptr_t type for
- // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
- // trunc to be exposed to other transforms. Don't do this for extending
- // ptrtoint's, because we don't know if the target sign or zero extends its
- // pointers.
- if (TD &&
- CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
- Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
- TD->getIntPtrType(CI.getContext()),
- "tmp");
- return new TruncInst(P, CI.getType());
+ // If the destination integer type is not the intptr_t type for this target,
+ // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
+ // to be exposed to other transforms.
+ if (TD) {
+ if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
+ Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
+ TD->getIntPtrType(CI.getContext()),
+ "tmp");
+ return new TruncInst(P, CI.getType());
+ }
+ if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) {
+ Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
+ TD->getIntPtrType(CI.getContext()),
+ "tmp");
+ return new ZExtInst(P, CI.getType());
+ }
}
return commonPointerCastTransforms(CI);
}
+/// OptimizeVectorResize - This input value (which is known to have vector type)
+/// is being zero extended or truncated to the specified vector type. Try to
+/// replace it with a shuffle (and vector/vector bitcast) if possible.
+///
+/// The source and destination vector types may have different element types.
+static Instruction *OptimizeVectorResize(Value *InVal, const VectorType *DestTy,
+ InstCombiner &IC) {
+ // We can only do this optimization if the output is a multiple of the input
+ // element size, or the input is a multiple of the output element size.
+ // Convert the input type to have the same element type as the output.
+ const VectorType *SrcTy = cast<VectorType>(InVal->getType());
+
+ if (SrcTy->getElementType() != DestTy->getElementType()) {
+ // The input types don't need to be identical, but for now they must be the
+ // same size. There is no specific reason we couldn't handle things like
+ // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
+ // there yet.
+ if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
+ DestTy->getElementType()->getPrimitiveSizeInBits())
+ return 0;
+
+ SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
+ InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
+ }
+
+ // Now that the element types match, get the shuffle mask and RHS of the
+ // shuffle to use, which depends on whether we're increasing or decreasing the
+ // size of the input.
+ SmallVector<Constant*, 16> ShuffleMask;
+ Value *V2;
+ const IntegerType *Int32Ty = Type::getInt32Ty(SrcTy->getContext());
+
+ if (SrcTy->getNumElements() > DestTy->getNumElements()) {
+ // If we're shrinking the number of elements, just shuffle in the low
+ // elements from the input and use undef as the second shuffle input.
+ V2 = UndefValue::get(SrcTy);
+ for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
+ ShuffleMask.push_back(ConstantInt::get(Int32Ty, i));
+
+ } else {
+ // If we're increasing the number of elements, shuffle in all of the
+ // elements from InVal and fill the rest of the result elements with zeros
+ // from a constant zero.
+ V2 = Constant::getNullValue(SrcTy);
+ unsigned SrcElts = SrcTy->getNumElements();
+ for (unsigned i = 0, e = SrcElts; i != e; ++i)
+ ShuffleMask.push_back(ConstantInt::get(Int32Ty, i));
+
+ // The excess elements reference the first element of the zero input.
+ ShuffleMask.append(DestTy->getNumElements()-SrcElts,
+ ConstantInt::get(Int32Ty, SrcElts));
+ }
+
+ Constant *Mask = ConstantVector::get(ShuffleMask.data(), ShuffleMask.size());
+ return new ShuffleVectorInst(InVal, V2, Mask);
+}
+
+static bool isMultipleOfTypeSize(unsigned Value, const Type *Ty) {
+ return Value % Ty->getPrimitiveSizeInBits() == 0;
+}
+
+static unsigned getTypeSizeIndex(unsigned Value, const Type *Ty) {
+ return Value / Ty->getPrimitiveSizeInBits();
+}
+
+/// CollectInsertionElements - V is a value which is inserted into a vector of
+/// VecEltTy. Look through the value to see if we can decompose it into
+/// insertions into the vector. See the example in the comment for
+/// OptimizeIntegerToVectorInsertions for the pattern this handles.
+/// The type of V is always a non-zero multiple of VecEltTy's size.
+///
+/// This returns false if the pattern can't be matched or true if it can,
+/// filling in Elements with the elements found here.
+static bool CollectInsertionElements(Value *V, unsigned ElementIndex,
+ SmallVectorImpl<Value*> &Elements,
+ const Type *VecEltTy) {
+ // Undef values never contribute useful bits to the result.
+ if (isa<UndefValue>(V)) return true;
+
+ // If we got down to a value of the right type, we win, try inserting into the
+ // right element.
+ if (V->getType() == VecEltTy) {
+ // Inserting null doesn't actually insert any elements.
+ if (Constant *C = dyn_cast<Constant>(V))
+ if (C->isNullValue())
+ return true;
+
+ // Fail if multiple elements are inserted into this slot.
+ if (ElementIndex >= Elements.size() || Elements[ElementIndex] != 0)
+ return false;
+
+ Elements[ElementIndex] = V;
+ return true;
+ }
+
+ if (Constant *C = dyn_cast<Constant>(V)) {
+ // Figure out the # elements this provides, and bitcast it or slice it up
+ // as required.
+ unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
+ VecEltTy);
+ // If the constant is the size of a vector element, we just need to bitcast
+ // it to the right type so it gets properly inserted.
+ if (NumElts == 1)
+ return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
+ ElementIndex, Elements, VecEltTy);
+
+ // Okay, this is a constant that covers multiple elements. Slice it up into
+ // pieces and insert each element-sized piece into the vector.
+ if (!isa<IntegerType>(C->getType()))
+ C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
+ C->getType()->getPrimitiveSizeInBits()));
+ unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
+ const Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
+
+ for (unsigned i = 0; i != NumElts; ++i) {
+ Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
+ i*ElementSize));
+ Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
+ if (!CollectInsertionElements(Piece, ElementIndex+i, Elements, VecEltTy))
+ return false;
+ }
+ return true;
+ }
+
+ if (!V->hasOneUse()) return false;
+
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (I == 0) return false;
+ switch (I->getOpcode()) {
+ default: return false; // Unhandled case.
+ case Instruction::BitCast:
+ return CollectInsertionElements(I->getOperand(0), ElementIndex,
+ Elements, VecEltTy);
+ case Instruction::ZExt:
+ if (!isMultipleOfTypeSize(
+ I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
+ VecEltTy))
+ return false;
+ return CollectInsertionElements(I->getOperand(0), ElementIndex,
+ Elements, VecEltTy);
+ case Instruction::Or:
+ return CollectInsertionElements(I->getOperand(0), ElementIndex,
+ Elements, VecEltTy) &&
+ CollectInsertionElements(I->getOperand(1), ElementIndex,
+ Elements, VecEltTy);
+ case Instruction::Shl: {
+ // Must be shifting by a constant that is a multiple of the element size.
+ ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
+ if (CI == 0) return false;
+ if (!isMultipleOfTypeSize(CI->getZExtValue(), VecEltTy)) return false;
+ unsigned IndexShift = getTypeSizeIndex(CI->getZExtValue(), VecEltTy);
+
+ return CollectInsertionElements(I->getOperand(0), ElementIndex+IndexShift,
+ Elements, VecEltTy);
+ }
+
+ }
+}
+
+
+/// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we
+/// may be doing shifts and ors to assemble the elements of the vector manually.
+/// Try to rip the code out and replace it with insertelements. This is to
+/// optimize code like this:
+///
+/// %tmp37 = bitcast float %inc to i32
+/// %tmp38 = zext i32 %tmp37 to i64
+/// %tmp31 = bitcast float %inc5 to i32
+/// %tmp32 = zext i32 %tmp31 to i64
+/// %tmp33 = shl i64 %tmp32, 32
+/// %ins35 = or i64 %tmp33, %tmp38
+/// %tmp43 = bitcast i64 %ins35 to <2 x float>
+///
+/// Into two insertelements that do "buildvector{%inc, %inc5}".
+static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI,
+ InstCombiner &IC) {
+ const VectorType *DestVecTy = cast<VectorType>(CI.getType());
+ Value *IntInput = CI.getOperand(0);
+
+ SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
+ if (!CollectInsertionElements(IntInput, 0, Elements,
+ DestVecTy->getElementType()))
+ return 0;
+
+ // If we succeeded, we know that all of the element are specified by Elements
+ // or are zero if Elements has a null entry. Recast this as a set of
+ // insertions.
+ Value *Result = Constant::getNullValue(CI.getType());
+ for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
+ if (Elements[i] == 0) continue; // Unset element.
+
+ Result = IC.Builder->CreateInsertElement(Result, Elements[i],
+ IC.Builder->getInt32(i));
+ }
+
+ return Result;
+}
+
+
+/// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double
+/// bitcast. The various long double bitcasts can't get in here.
+static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){
+ Value *Src = CI.getOperand(0);
+ const Type *DestTy = CI.getType();
+
+ // If this is a bitcast from int to float, check to see if the int is an
+ // extraction from a vector.
+ Value *VecInput = 0;
+ // bitcast(trunc(bitcast(somevector)))
+ if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) &&
+ isa<VectorType>(VecInput->getType())) {
+ const VectorType *VecTy = cast<VectorType>(VecInput->getType());
+ unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
+
+ if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) {
+ // If the element type of the vector doesn't match the result type,
+ // bitcast it to be a vector type we can extract from.
+ if (VecTy->getElementType() != DestTy) {
+ VecTy = VectorType::get(DestTy,
+ VecTy->getPrimitiveSizeInBits() / DestWidth);
+ VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
+ }
+
+ return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(0));
+ }
+ }
+
+ // bitcast(trunc(lshr(bitcast(somevector), cst))
+ ConstantInt *ShAmt = 0;
+ if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)),
+ m_ConstantInt(ShAmt)))) &&
+ isa<VectorType>(VecInput->getType())) {
+ const VectorType *VecTy = cast<VectorType>(VecInput->getType());
+ unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
+ if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 &&
+ ShAmt->getZExtValue() % DestWidth == 0) {
+ // If the element type of the vector doesn't match the result type,
+ // bitcast it to be a vector type we can extract from.
+ if (VecTy->getElementType() != DestTy) {
+ VecTy = VectorType::get(DestTy,
+ VecTy->getPrimitiveSizeInBits() / DestWidth);
+ VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
+ }
+
+ unsigned Elt = ShAmt->getZExtValue() / DestWidth;
+ return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
+ }
+ }
+ return 0;
+}
+
Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
// If the operands are integer typed then apply the integer transforms,
// otherwise just apply the common ones.
Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
unsigned NumZeros = 0;
while (SrcElTy != DstElTy &&
- isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
+ isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
SrcElTy->getNumContainedTypes() /* not "{}" */) {
SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
++NumZeros;
((Instruction*)NULL));
}
}
+
+ // Try to optimize int -> float bitcasts.
+ if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
+ if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this))
+ return I;
if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
- if (DestVTy->getNumElements() == 1 && !isa<VectorType>(SrcTy)) {
+ if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
// FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
}
+
+ if (isa<IntegerType>(SrcTy)) {
+ // If this is a cast from an integer to vector, check to see if the input
+ // is a trunc or zext of a bitcast from vector. If so, we can replace all
+ // the casts with a shuffle and (potentially) a bitcast.
+ if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
+ CastInst *SrcCast = cast<CastInst>(Src);
+ if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
+ if (isa<VectorType>(BCIn->getOperand(0)->getType()))
+ if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0),
+ cast<VectorType>(DestTy), *this))
+ return I;
+ }
+
+ // If the input is an 'or' instruction, we may be doing shifts and ors to
+ // assemble the elements of the vector manually. Try to rip the code out
+ // and replace it with insertelements.
+ if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this))
+ return ReplaceInstUsesWith(CI, V);
+ }
}
if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
- if (SrcVTy->getNumElements() == 1 && !isa<VectorType>(DestTy)) {
+ if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) {
Value *Elem =
Builder->CreateExtractElement(Src,
Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
// Okay, we have (bitcast (shuffle ..)). Check to see if this is
- // a bitconvert to a vector with the same # elts.
- if (SVI->hasOneUse() && isa<VectorType>(DestTy) &&
+ // a bitcast to a vector with the same # elts.
+ if (SVI->hasOneUse() && DestTy->isVectorTy() &&
cast<VectorType>(DestTy)->getNumElements() ==
SVI->getType()->getNumElements() &&
SVI->getType()->getNumElements() ==
}
}
- if (isa<PointerType>(SrcTy))
+ if (SrcTy->isPointerTy())
return commonPointerCastTransforms(CI);
return commonCastTransforms(CI);
}