RFC: PATCH: PR gas/12049: Unnecessary relaxation

[H.J. Lu]

On Tue, Oct 19, 2010 at 4:59 AM, Alan Modra <amodra@gmail.com> wrote:
> On Mon, Oct 18, 2010 at 10:51:32AM -0700, H.J. Lu wrote:
>> I tried it on gcc trunk at revision 165633 with i686-linux:
>
> Timing gcc bootstrap and regression testing isn't particularly
> convincing, since the assembler time is likely only a small part of
> the total.
>
> If you take a look at the history of this code, you'll find
> http://sourceware.org/ml/binutils/2001-02/msg00320.html.  In that
> patch, Ian considerably reduced the time in gas by minimizing (a
> little too enthusiastically) the cases we did a scan over frags for
> each rs_machine_dependent frag.  You might understand why I'd rather
> not go back to doing that again.  In fact, the more I look at your
> patch, the less I like it.  You're using code that initializes frag
> addresses to update addresses??  That would reset branches, leb128
> etc. to their initial sizes!  Also, you only need to traverse frags
> until you find the symbol frag, not update all of them.
>
> I think the following is a much better solution.
>
>        PR gas/12049
>        * frags.h (struct frag): Add "region" field.
>        * write.c (relax_frag): Don't add "stretch" to forward reference
>        target if there is an intervening org or align.
>        (relax_segment): Set region.
>

Hi Alan,

Your fix doesn't handle the new testcase in

http://sourceware.org/bugzilla/show_bug.cgi?id=12049


H.J.