SystemError: module dl requires sizeof(int) == sizeof(long) == sizeof(char*)

That's just unspeakably lame. I realize this is longstanding, and it's
even documented, but just... FAIL.

Moreover, eliding the check and subsequent error from the code yields a
module that appears to work in (very) brief testing on x86_64.

The limitation is genuine; it's not possible to implement dl_call
correctly if sizeof(long) != sizeof(void*), say. Otherwise, it would
have been fixed long ago. Use ctypes instead, if that is supported for
your hardware. (More generally, it's not possible to implement this at
all in portable C, that's why ctypes uses assembler code. That dl works
on most "pure" 32-bit systems is by sheer luck)

sizeof(long) != sizeof(void *) isn't the epic fail here. It's the
complaint about sizeof(int) != sizeof(long).

> sizeof(long) != sizeof(void *) isn't the epic fail here. It's the
> complaint about sizeof(int) != sizeof(long).

Same story.

If it's not possible to implement at all, why does it work without the
test?  What specific cases do you think will fail?

Heck, if you don't think it's functional on 90% of hardware sold these
days, just remove it, it makes python look less silly. (Portability is
hard, let's punt.)

It fails if the function being called has parameters which are not
sizeof(long), and the platform has a calling convention where such
parameters are passed on the stack, and consume a size different from
sizeof(long) on the stack.

I disagree that it won't run on 90% of the hardware sold today - you
just have to operate that hardware properly to make it work. In any
case, removal of the dl module in favor of ctypes seems reasonable.