Re: master b0ba0d42b0f: * src/lisp.h (EQ): Improve generated code.

[Pip Cet via "Emacs development discussions." <emacs-devel@gnu.org>]

On Thursday, November 28th, 2024 at 17:53, Mattias Engdegård <mattias.engdegard@gmail.com> wrote:
> 28 nov. 2024 kl. 14.53 skrev Pip Cet via Emacs development discussions. emacs-devel@gnu.org:
> 
> > Maybe we should use __builtin_expect_with_probability instead, in those rare cases when we are certain we're making a correct prediction? Or, my preference, avoid using __builtin_expect entirely, so our scarce resources can be spent on more important issues?
> 
> Actually __builtin_expect can definitely provide a measurable performance improvement,

If we want that improvement, let's use a profile-guided build. No need to reduce the legibility of our C code further by using a GCC extension which is largely obsoleted by PGO and __builtin_expect_with_probability.

> mainly for BB ordering and cold-path moving as suggested by Andrea's commit note. I've been thinking about using it in other cases.

I think we should heed the warning in the GCC manual: it is very hard to get a good grasp of whether __builtin_expect provides any benefit.

> We should probably define some variant of likely/unlikely because those cover most needs of __builtin_expect, rather than using it directly. Then portability wouldn't be a problem.

That's what the Android code already does, IIUC. I don't think that solves any problems, because the compiler uses a 90% probability for __builtin_expect, and "likely" is simply misleading for that.

> > I also don't think the assumption you're telling GCC to make in this specific case (more than 90% of calls to EQ happen while syms_with_pos_enabled == false) is obviously correct.
> 
> On this point I agree -- this particular change may have been made in haste, although I applaud the spirit.

So let's discuss this if and when a real usecase for __builtin_expect arises, and we conclude that use case couldn't be addressed by PGO?

Pip