The example was taken from Issue3681.hx which failed to compile for C# with -D fast-cast if I fix the redundant __hx_cast generation.

I don"t even know where to start looking here. I"m assuming it"s not a regression?

Is that actually unexpected? You call new Node(1, null), so T gets bound to Int.

Yes, but the problem is that next compiler argument receives a different T (Int instead of Null<Int> in this case).

Yes, because they unify. It"s clearer if you split up the call:

var i:Null<Int> = 1;
var next = new Node(1, null);
new Node(i, next);

The only problem here is that the second argument is silently cast from Node<Int> to Node<Null<Int>> because Null<T> is a typedef. That"s not a type-inference issue though.

You could maybe make the argument that in the merged call, top-down inference should cause the inner type to be inferred to Null<T> straight away. I"m not quite sure about that, but it wouldn"t solve the problem in the split-up version anyway.

You could maybe make the argument that in the merged call, top-down inference should cause the inner type to be inferred to Null straight away.

Yeah that what I was expecting here originally. The split-up version is more explicit and should cause an error...