Re: [PATCH 1/4] mm, page_alloc: add likely macro to help compileroptimization

[Johannes Weiner]

On Mon, Aug 05, 2013 at 05:50:41PM +0900, Joonsoo Kim wrote:
> On Mon, Aug 05, 2013 at 05:10:08PM +0900, Joonsoo Kim wrote:
> > Hello, Michal.
> > 
> > On Fri, Aug 02, 2013 at 11:36:07PM +0200, Michal Hocko wrote:
> > > On Fri 02-08-13 16:47:10, Johannes Weiner wrote:
> > > > On Fri, Aug 02, 2013 at 06:27:22PM +0200, Michal Hocko wrote:
> > > > > On Fri 02-08-13 11:07:56, Joonsoo Kim wrote:
> > > > > > We rarely allocate a page with ALLOC_NO_WATERMARKS and it is used
> > > > > > in slow path. For making fast path more faster, add likely macro to
> > > > > > help compiler optimization.
> > > > > 
> > > > > The code is different in mmotm tree (see mm: page_alloc: rearrange
> > > > > watermark checking in get_page_from_freelist)
> > > > 
> > > > Yes, please rebase this on top.
> > > > 
> > > > > Besides that, make sure you provide numbers which prove your claims
> > > > > about performance optimizations.
> > > > 
> > > > Isn't that a bit overkill?  We know it's a likely path (we would
> > > > deadlock constantly if a sizable portion of allocations were to ignore
> > > > the watermarks).  Does he have to justify that likely in general makes
> > > > sense?
> > > 
> > > That was more a generic comment. If there is a claim that something
> > > would be faster it would be nice to back that claim by some numbers
> > > (e.g. smaller hot path).
> > > 
> > > In this particular case, unlikely(alloc_flags & ALLOC_NO_WATERMARKS)
> > > doesn't make any change to the generated code with gcc 4.8.1 resp.
> > > 4.3.4 I have here.
> > > Maybe other versions of gcc would benefit from the hint but changelog
> > > didn't tell us. I wouldn't add the anotation if it doesn't make any
> > > difference for the resulting code.
> > 
> > Hmm, Is there no change with gcc 4.8.1 and 4.3.4?
> > 
> > I found a change with gcc 4.6.3 and v3.10 kernel.
> 
> Ah... I did a test on mmotm (Johannes's git) and found that this patch
> doesn't make any effect. I guess, a change from Johannes ('rearrange
> watermark checking in get_page_from_freelist') already makes better code
> for !ALLOC_NO_WATERMARKS case. IMHO, although there is no effect, it is
> better to add likely macro, because arrangement can be changed from time
> to time without any consideration of assembly code generation. How about
> your opinion, Johannes and Michal?

I'm not against it.  It does not really matter if gcc gets it right by
accident right now and the annotation does not have an immediate
effect.  It's a compiler hint and we want gcc to know it so it doesn't
ever assume otherwise in the future.  Yes, nobody will re-evaluate if
the conditional still generates the jumps correctly after shifting the
code around.

Thanks!
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/