Custom types (RFC)

[skrah]

There have been various questions about custom types. This is the current state and an opportunity for concrete feedback.

Custom types are created as follows, an example is the graph custom type:

1. You need to know at least the abstract typedef that describes the underlying memory. Here we define a weighted graph as an adjacency list. We use a typedef to declare a new custom type:

https://github.com/plures/gumath/blob/daa7bb94730efe163e6129de5998a28d016e538d/libgumath/extending/graph.c#L346

2. To validate graph invariants, the constraint function graph_constraint is automatically called upon initialization.

3. Functions that operate on the graph obviously still need to be written in C.

4. We define a Python class that inherits from xnd, with type checked initialization of values:

https://github.com/plures/gumath/blob/daa7bb94730efe163e6129de5998a28d016e538d/python/extending.py#L11

5. We use this xnd subclass just like any other Python type:

https://github.com/plures/gumath/blob/daa7bb94730efe163e6129de5998a28d016e538d/python/test_gumath.py#L241

6. If desired, type hints could look like this:

def f(g : Graph):
    pass

[skrah]

Naturally, the above is syntax sugar and dispatching on types can still look like this:

def f(x):
    if ndt("graph").match(x.type):
        do_this()
    elif ndt("2 * 2 * int64").match(x.type):
        do_that()
    ...

[sklam]

For numba usecases, we will need to construct types dynamically. I suppose we can invoke ndt_typedef_from_string directly.

[skrah]

Yes, all that matters is that the typedef exists in the typedef table before ndt_nominal() is called.

[dhirschfeld]

A dual number type would be very interesting for forward mode AD!