[Cbe-oss-dev] [PATCH 1/3] Fix Unlikely(x) == y

[Nick Piggin]

On Tuesday 19 February 2008 16:58, Willy Tarreau wrote:
> On Tue, Feb 19, 2008 at 01:33:53PM +1100, Nick Piggin wrote:
> > > Note in particular the last predictors; assuming branch ending
> > > with goto, including call, causing early function return or
> > > returning negative constant are not taken. Just these alone
> > > are likely 95+% of the unlikelies in the kernel.
> >
> > Yes, gcc should be able to do pretty good heuristics, considering
> > the quite good numbers that cold CPU predictors can attain. However
> > for really performance critical code (or really "never" executed
> > code), then I think it is OK to have the hints and not have to rely
> > on gcc heuristics.
>
> in my experience, the real problem is that gcc does what *it* wants and not
> what *you* want. I've been annoyed a lot by the way it coded some loops
> that could really be blazingly fast, but which resulted in a ton of
> branches due to its predictors. And using unlikely() there was a real mess,
> because instead of just hinting the compiler with probabilities to write
> some linear code for the *most* common case, it ended up with awful
> branches everywhere with code sent far away and even duplicated for some
> branches.
>
> Sometimes, for performance critical paths, I would like gcc to be dumb and
> follow *my* code and not its hard-coded probabilities. For instance, in a
> tree traversal, you really know how you want to build your loop. And these
> days, it seems like the single method of getting it your way is doing asm,
> which obviously is not portable :-(

Probably all true.


> Maybe one thing we would need would be the ability to assign probabilities
> to each branch based on what we expect, so that gcc could build a better
> tree keeping most frequently used code tight.

I don't know if that would *directly* lead to gcc being smarter. I
think perhaps they probably don't benchmark on code bases that have
much explicit annotation (I'm sure they wouldn't seriously benchmark
any parts of Linux as part of daily development). I think the key is
to continue to use annotations _properly_, and eventually gcc should
go in the right direction if enough code uses it.

And if you have really good examples like it sounds like above, then
I guess that should be reported to gcc?