}
}
}
-
-
+
+ // 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);
+ }
+ }
+ }
+
// If the input is a shl/ashr pair of a same constant, then this is a sign
// extension from a smaller value. If we could trust arbitrary bitwidth
// integers, we could turn this into a truncate to the smaller bit and then
Arg->getOperand(0)->getType()->isFloatTy()) {
Function *Callee = Call->getCalledFunction();
Module *M = CI.getParent()->getParent()->getParent();
- Constant* SqrtfFunc = M->getOrInsertFunction("sqrtf",
+ Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf",
Callee->getAttributes(),
Builder->getFloatTy(),
Builder->getFloatTy(),
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 Value % Ty->getPrimitiveSizeInBits() == 0;
}
-static bool getTypeSizeIndex(unsigned Value, const Type *Ty) {
+static unsigned getTypeSizeIndex(unsigned Value, const Type *Ty) {
return Value / Ty->getPrimitiveSizeInBits();
}
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;
return true;
}
- //if (Constant *C = dyn_cast<Constant>(V)) {
+ 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;