[Caml-list] Fwd: Proposal: re-design of ocaml headers

[Yotam Barnoy <yotambarnoy@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 1748 bytes --]

>
>   Can you please share your experience of writing bindings to some C
> libraries for <your preferable language>?  How it compares to  writing
> bindings for OCaml?  (this is a thread about runtime values
> representation, I suppose.)
>
>
This isn't really relevant to this topic, since this discussion is just
about ocaml headers, rather than the ocaml C FFI. The FFI would remain
largely the same.



>   Why will anyone ever need more than 200 constructors of a sum type?
> (also note the presence of polymorphic variant types.)
>

I thought so too, but as mentioned in the previous discussion thread
(entitled 'Expanding the Float Array Tag'), code generators can easily run
out of 200 constructors. However, it's possible that 65,000 constructors is
really excessive, in which case some bits can be shifted around. I'd be
happy to get some feedback on this.


> > random access to this data is never needed
>
>   mkay...  :[
>

This is only talking about automatic ocaml functions, which work without
any type information. Proper ocaml code is typed, so it doesn't need this
information. But internal mechanisms such as the GC, marshal module, and
polymorphic comparison are 'dumb' in the sense that they're missing all of
the crucial type information. For these functions, there needs to be some
basic type information within the data structure itself. Currently, floats
are stored in their own independent data structure with a 'double_tag'
header, or otherwise they're in a 'double_array'. Having some information
in the header about internal floats and pointers is much more efficient,
but it doesn't need to have random access -- the internal ocaml functions
that use this information process the entire data structure at once.

Yotam

[-- Attachment #2: Type: text/html, Size: 2693 bytes --]