Re: [PATCH 3/3] Mark various cold functions as __COLD

[Sergey Bugaev <bugaevc@gmail.com>]

Hello,

On Tue, May 2, 2023 at 12:54 AM Samuel Thibault <samuel.thibault@gnu.org> wrote:
> Cristian Rodríguez, le lun. 01 mai 2023 17:52:17 -0400, a ecrit:
> > abort() and all the function wrappers called by assert on failure are also
> > cold.
>
> Indeed, most __noreturn__ functions a probably worth having a look at.

I was under the impression that noreturn functions automatically get
treated similarly to ones with an explicit attribute((cold)). (But I
still marked some, but not all -- fairly arbitrarily -- noreturn
functions __COLD in this patch...)

From taking a closer look now, I see that while the generated assembly
instructions may be similar/identical, it's not strictly the same
thing: with attribute((cold)) the cold path code goes into
.text.unlikely, and without it's still in the regular .text. So
applying __COLD should be beneficial indeed. I will send a follow-up
patch; thank you for pointing this out and making me double-check :)

Is there some sort of a benchmark where we could see whether this
makes any difference? (Not calling the cold function in a loop, of
course -- that is likely to regress, in fact -- but rather calling a
hot function in a loop and benefitting from the cold path not getting
in the way, improved code locality etc.)

Sergey