Skip site navigation (1)Skip section navigation (2)
Date:      Mon, 12 Sep 2011 07:43:01 -0400
From:      John Baldwin <jhb@freebsd.org>
To:        Andriy Gapon <avg@freebsd.org>
Cc:        freebsd-fs@freebsd.org, FreeBSD-Current <freebsd-current@freebsd.org>
Subject:   Re: archaic/useless CFLAGS options for x86 boot blocks
Message-ID:  <201109120743.02181.jhb@freebsd.org>
In-Reply-To: <4E6DB696.1080608@FreeBSD.org>
References:  <4E6DB696.1080608@FreeBSD.org>

next in thread | previous in thread | raw e-mail | index | archive | help
On Monday, September 12, 2011 3:36:54 am Andriy Gapon wrote:
> 
> This email is in part inspired by the following problem:
> http://article.gmane.org/gmane.os.freebsd.current/135292
> So "harmful" could also be added to the subject line.
> 
> So here is my proposal.
> 
> Part I.  ZFS and GPT bootblocks.
> 
> I believe that we do not need here any extra optimizations and happy dances that
> seem to be carried over from boot2.
> I think that just the -Os should be sufficient of optimization flags.  Maybe
> even that is not really required.
> Rationale:
> - these boot blocks are not as nearly space-constrained as boot2
> - using untypical flags increases chances of hitting compiler bugs,
>   especially for those compilers where we are stuck with
>   unsupported / locally-maintained versions or where a compiler is maturing yet
> - assembly / machine code and debugging may become easier
> 
> Additionally, the '/align/d' '/nop/d' filtering of the intermediate assembly
> file seems to be not needed for zfsboot.
> 
> Part II.  The original boot2.
> 
> My testing shows that -Os -fomit-frame-pointer are sufficient to produce a small
> enough boot2 image (~300 bytes remain available with gcc, 51 bytes for clang).
> -mrtd -mregparm=3 do not change size with gcc, but with clang they increase
> _available_ space to 79 bytes.
> 
> The '/align/d' '/nop/d' filtering seems to shave off only 7 bytes here.
> Not suggesting anything, just an observation...
> 
> Part III.  History.
> 
> It seems that all those optimization related options were introduced very long
> time ago when the compiler(s) were quite different from what they are now.
> So, some re-evaluation may be (long over)due.
> For example, -fno-unit-at-a-time is definitely an anti-optmization option and it
> was introduced to fight some gcc bugs way back in 2004 (r132870).  Its merits
> have never been re-evaluated after switch to gcc 4.2, it seems.
> -fno-guess-branch-probability and -mno-align-long-strings are even less obvious
> options (see e.g. r108149).
> 
> 
> Finally, here is a diff:
> http://people.freebsd.org/~avg/boot-cflags.diff
> All the boot blocks are boot tested in qemu.
> boot2 is also tested with -mrtd -mregparm removed.

I suspect some of the recent changes to shave space down for Clang have made
some of the optimization options no longer necessary.  I think the patch is
fine, and I'd even prefer to go ahead and drop the extra cruft (like removing
nops and aligns as well as -mrtd and -mregparm) from the UFS boot2 as well.

-- 
John Baldwin



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201109120743.02181.jhb>