------- Comment #8 from rguenther at suse dot de 2008-07-03 11:49 -------
Subject: Re: [4.4 regression] r137252 breaks -O2 optimization
on x86_64-unknown-linux-gnu
On Thu, 3 Jul 2008, dfranke at gcc dot gnu dot org wrote:
> ------- Comment #7 from dfranke at gcc dot gnu dot org 2008-07-03 11:42
> -------
> Differences in *.optimized dumps seem to be variable renaming only, e.g.
> - temp.1140 = tdata->p;
> + temp.1137 = tdata->p;
>
> Differences in *.alias seems to be mostly of the kind:
> -PARM_NOALIAS.895 = &ANYTHING
> +PARM_NOALIAS.895 = &ESCAPED
>
> However, I also found:
> -PARM_NOALIAS.895 = { ANYTHING ESCAPED NONLOCAL }
> -D.3410_6 = same as thread_data
> -thread_data = { ANYTHING }
> -D.3411_7 = same as thread_data
> +PARM_NOALIAS.895 = { ESCAPED NONLOCAL }
> +D.3410_6 = same as derefaddrtmp.893
> +thread_data = same as derefaddrtmp.893 <---- ?!?
> +D.3411_7 = same as derefaddrtmp.893
Well, it's not really different - only thread_data now doesn't
point to anything but to the same as derefaddrtmp.893 (I guess
that points to { ESCAPED NONLOCAL }?).
> which might hint at the problem: the variable thread_data is an allocated
> POINTER to a vector of structures (which hold more pointers), public within a
> module. USE-ing it, makes it (sort of) local to the function. Most (>80% of
> all) changes in *.optimized where in the one source file that uses this module
> ...
>
> Richard, would this fit the bill?
Well, if there is no difference in the optimized dumps then this
doesn't explain it (but the points-to differences also look good).
Richard.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36713
