Re: [PATCH] x86: Optimize variable_test_bit()

[Peter Zijlstra <peterz@infradead.org>]

On Fri, May 01, 2015 at 06:16:54PM +0200, Peter Zijlstra wrote:
> On Fri, May 01, 2015 at 09:03:32AM -0700, Linus Torvalds wrote:
> > On Fri, May 1, 2015 at 8:16 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> > >
> > > Since test_bit() doesn't actually have any output variables, we can use
> > > asm goto without having to add a memory clobber. This reduces the code
> > > to something sensible:
> > 
> > Yes, looks good, except if we have anything that actually wants to use
> > the value rather than branch on it. But a quick grep seems to show
> > that the vast majority of them are all about just directly testing the
> > result.
> > 
> > It worries me a bit that gcc now cannot pick the likely branch any
> > more. It will always branch out for the bit being set. So code like
> > this:
> > 
> >     net/core/dev.c:         if
> > (likely(!test_bit(__QDISC_STATE_DEACTIVATED, &q->state)))
> > 
> > wouldn't work, but almost all of those seem to be the constant case
> > that doesn't get to this anyway.
> 
> Ah yes, that's another thing we've previously discussed with the GCC
> people (IIRC). The GCC manual states you can use hot and cold attributes
> on the labels (although when we tested that it didn't actually work, it
> might now). But that's no good if the hint is one (or more) layer up
> from the asm goto.
> 
> If would indeed be very good if the likely/unlikely thing would work as
> expected.

Ah, I see what you meant. Yes we're stuck with the 'jc', the compiler
cannot flip that into a jnc.

The best it can do is add unconditional jumps to re-arrange the blocks,
and that's somewhat ugly indeed, although better than nothing at all.
And from experiments back when we did the static_branch stuff that all
didn't work at all.