Re: [PATCH v3 1/5] AArch64: Improve A64FX memset

[Wilco Dijkstra via Libc-alpha <libc-alpha@sourceware.org>]

Hi Naohiro,

> Would you update the commit title so as not to be the same among 5
> patches?
> Because we need to ask distro to backport these patches.
> If all commit titles are the same, it will increase the room to happen
> confusion and mistake.
> 
> How about "AArch64: Improve A64FX memset for less than 512B" ?

Generally the commit title in a patch series would include the series number,
however it's also easy to add something to the title as suggested. As for
backporting, one uses the hash of the patch in the cherry-pick rather than
the title, so once you have the right hashes, there should be no possibility
of confusion.
 
> > -#define ZF_DIST                (CACHE_LINE_SIZE * 21)  // Zerofill distance
> 
> This caused compile error.

Sorry, that should be part of the 2nd patch.

> > -ENTRY (MEMSET)
> > +#undef BTI_C
> > +#define BTI_C
>
> We discussed how should be defined BTI_C macro before, at that time conclusion
> was "NOP" rather than empty unless HAVE_AARCH64_BTI.
> Now the above code defines BTI_C as empty unconditionally.
> A64FX doesn't support BTI, so this code is OK.
> But I'm just interested in the reason why it is changed.

We changed to NOP in the generic code, so that works for all string functions.
In this specific case removing the initial NOP as well allows all performance critical
code for <= 512 bytes to be perfectly aligned to 16-byte fetch blocks.

Cheers,
Wilco