On 21/04/23 06:27, Sergey Bugaev wrote:
> Hello,
>
> On Thu, Apr 20, 2023 at 11:38 PM Adhemerval Zanella Netto
> <adhemerval.zane...@linaro.org> wrote:
>> Can't you use a similar strategy done by
>> 5355f9ca7b10183ce06e8a18003ba30f43774858 ?
>
> Do I understand it right that that is the moral equivalent of
>
> #define memcpy __memcpy_sse2_unaligned
>
> except that it works aat assembly level and so will catch implicit
> calls to memcpy that the compiler may insert?
Yes, that's the idea.
>
> Assuming for a second that we don't care about implicit memcpys (but
> we should) and only about the explicit one made by the MIG runtime, we
> could maybe even just do
>
> void *
> __mig_memcpy (void *dst, const void *src, vm_size_t len)
> {
> - return memcpy (dst, src, len);
> + return __memcpy_sse2_unaligned (dst, src, len);
> }
>
> but that (as well as your proposal) would make *all* calls to memcpy
> in this place go through the baseline version, even after the early
> startup is done. Whereas my proposal attempted to avoid that -- unless
> of course H.J. is right and this prevents the indirect relocation from
> replacing the function pointer later, in which case it's even worse
> because it would sabotage memcpy for the whole program and not just a
> couple of files.
It might work if you don't care about a different architecture than x86,
and that's why I added the alias (so each architecture is free to name
its ifunc variant). And the patch was exactly to handle the implicit
created mem* call from compiler, so I think you should take it in
consideration.
And I trying to make reason why you need __mig_memcpy indirection for
MIG.
