Re: MPS codegen

[Gerd Möllmann]

Helmut Eller <eller.helmut@gmail.com> writes:

>> I think the thing I don't like is that I believe there should be only
>> one description of an object's layout. Be it C (struct), or be it
>> somethings else, from which C is generated.
>
> Sure, but that's easier said than done.

Yeah :-). Too bad that GCC doesn't have something like Clang's Python
API. Somehow I don't really believe in tree-sitter for this specific
purpose. But what do I know...

> BTW, the MPS documentation says that mps_reserve allocates a block.  In
> MPS, a block is not necessarily a single object.  One possible
> interpretation of this could be that mps_reserve can be used to allocate
> a block that contains many objects.  Perhaps we could use a single block
> for the entire dump?  E.g. implement a version of dump_mmap_contiguous
> that uses mps_reserve.  What's your interpretation?

The same. I think I've read somewhere that anything goes in such a block
as long as the object format can cope with it.

OTOH, on my machine, the whole copying of the dump to MPS, mirroring and
redirecting roots takes ca. 38ms ATM. Where the copying and mirroring
are the two main contributors and take about the same time. Maybe
there's not much to win.