[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20160713075346.GC28723@dhcp22.suse.cz>
Date: Wed, 13 Jul 2016 09:53:46 +0200
From: Michal Hocko <mhocko@...nel.org>
To: Chen Gang <chengang@...ndsoft.com.cn>
Cc: Minchan Kim <minchan@...nel.org>, akpm@...ux-foundation.org,
vbabka@...e.cz, mgorman@...hsingularity.net,
gi-oh.kim@...fitbricks.com, iamjoonsoo.kim@....com,
hillf.zj@...baba-inc.com, rientjes@...gle.com, linux-mm@...ck.org,
linux-kernel@...r.kernel.org, Chen Gang <gang.chen.5i5j@...il.com>
Subject: Re: [PATCH] mm: migrate: Use bool instead of int for the return
value of PageMovable
On Wed 13-07-16 00:50:12, Chen Gang wrote:
>
>
> On 7/12/16 15:48, Michal Hocko wrote:
> > On Tue 12-07-16 03:47:42, Chen Gang wrote:
> > [...]
> >> In our case, the 2 output size are same, but under x86_64, the insns are
> >> different. After uses bool, it uses push/pop instead of branch, for me,
> >> it should be a little better for catching.
> >
> > The code generated for bool version looks much worse. Look at the fast
> > path. Gcc tries to reuse the retq from the fast path in the bool case
> > and so it has to push rbp and rbx on the stack.
> >
> > That being said, gcc doesn't seem to generate a better code for bool so
> > I do not think this is really worth it.
> >
>
> The code below also merge 3 statements into 1 return statement, although
> for me, it is a little more readable, it will generate a little bad code.
> That is the reason why the output looks a little bad.
>
> In our case, for gcc 6.0, using bool instead of int for bool function
> will get the same output under x86_64.
If the output is same then there is no reason to change it.
> In our case, for gcc 4.8, using bool instead of int for bool function
> will get a little better output under x86_64.
I had a different impression and the fast path code had more
instructions. But anyway, is there really a strong reason to change
those return values in the first place? Isn't that just a pointless code
churn?
--
Michal Hocko
SUSE Labs
Powered by blists - more mailing lists