Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Wed, 19 Jul 2017 21:45:56 -0400
From: Rich Felker <dalias@...c.org>
To: musl@...ts.openwall.com
Subject: Re: [PATCH] prefer Oz optimization if the compiler supports it

On Wed, Jul 19, 2017 at 04:48:49PM +0300, Dmitry Golovin wrote:
> In theory Oz should produce slimmer binary than Os, but I want first
> to perform tests to see how optimization level really affects size
> and performance.

My impression is that clang's -O2 and -Os are comparable in goals to
gcc's -Os, and that -Oz just makes some minor size-improvement
tradeoffs that could be bad for performance.

I've been wanting for a while to actually drop use of -Os with
selective -O3 for some dirs, and instead use -O2 for everything by
default. With recent compilers, -Os seems to do gratuitously stupid
things, and doesn't make much of a size difference vs -O2 (and even
less so if you use -fno-align-*). The main thing that's kept me from
actually pursuing such a change is not having had time to test the
effects myself or work with someone else to get it done.

Rich


> 19.07.2017, 15:52, "Rich Felker" <dalias@...c.org>:
> > On Wed, Jul 19, 2017 at 01:44:17PM +0300, Dmitry Golovin wrote:
> >>  Hi all!
> >>
> >>  This is a very small and simple patch, it adds support for Oz level
> >>  of optimization (clang supports it). Since the kernel switched to
> >>  Oz, I think it's a good idea to use it as long as the compiler
> >>  supports it, on the other hand I didn't test if it has any
> >>  advantages in case of musl.
> >>
> >>  The patch is attached.
> >
> > I'm not clear what the goal is. Why would Oz be preferred to Os? A
> > quick search shows that all it seems to do is disabling some
> > vectorization, which sounds undesirable.
> >
> > Rich

Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.