Re: [PATCH] x86: Optimize variable_test_bit()

[Ingo Molnar <mingo@kernel.org>]

* Peter Zijlstra <peterz@infradead.org> wrote:

> On Fri, May 01, 2015 at 06:33:29PM +0200, Jakub Jelinek wrote:
> > On Fri, May 01, 2015 at 09:03:32AM -0700, Linus Torvalds wrote:
> > > > PPS. Jakub, I see gcc5.1 still hasn't got output operands for asm goto;
> > > >      is this something we can get 'fixed' ?
> > 
> > CCing Richard as author of asm goto and Vlad as register allocator 
> > maintainer.  There are a few enhancement requests to support this, 
> > like http://gcc.gnu.org/PR59615 and http://gcc.gnu.org/PR52381 , 
> > but indeed the reason why no outputs are allowed is the register 
> > allocation issue. Don't know if LRA would be better suited to 
> > handle that case, but it would indeed be pretty hard.
> 
> So it would b awesome if we could use these freshly modeled flags as 
> output for regular asm stmts; that would obviate much of the asm 
> goto hackery we now do/have and allow gcc to pick the right branch 
> for likely/unlikely.

If I may hijack the discussion a bit: it would also be awesome if 
there was a GCC flag that would allow us to use __builtin_expect() 
hints even when automatic branch heuristics are disabled:

I.e. very similar to -fno-guess-branch-probability, just that explicit 
__builtin_expect() hints would not be ignored (like 
-fno-guess-branch-probability does it today).

We could use this to compress the kernel instruction cache footprint 
by about 5% on x86-64, while still having all the hand-made 
optimizations that __builtin_expect() allows us.

It would be a perfect solution if -fno-guess-branch-probability just 
stopped ignoring __builtin_expect().

Thanks,

	Ingo