On Tue, Jun 17 2025, Pip Cet wrote:

> "Helmut Eller" <eller.helmut@gmail.com> writes:
>
>> This is a proposal to make the variable lface_id_to_name an exact root.
>
> Thank you!
>
> I'll apply this as it is, but I'd suggest one change, which I'd like to
> quickly run by you:
>
> Some of the code in igc_xpalloc_lisp_objs_exact handles item_size !=
> word_size, and some doesn't.  Would you prefer to remove the parameter
> or change the code to handle multi-word items, which might be useful for
> arrays of arrays of Lisp_Object?

Let's remove it.  The item_size is redundant here.  See the amended
patch.  I also added the missing xfree :-).

> I think we should add label arguments to all igc*alloc* functions.  With
> your patch, we don't see the lface-id-to-name root until the first time
> it's been resized, and it would make more sense to see it right away.

I added that as a separate patch.

> There's the general problem of memcpy atomicity, but I don't have a
> solution for that; the complicated code in xpalloc_exact is still
> insufficient, so we might as well remove it.
>
> While the ordinary memcpy function in glibc will always copy large
> chunks, some debug builds use byte-based replacements, and then we have
> a potential problem.  This is particularly annoying because
> stress-testing such builds by triggering collections from a background
> thread might see false positive segfaults.  I haven't seen one in
> practice, but I use such tools very rarely.

Fixing all problematic uses of memcpy would be quite tricky.  Two
obvious cases are larger_vector and copy_hash_table.  I suppose that
memcpy is also a reason why software write barriers would be a lot of
work.