Re: [PING][PATCH][REVISED] Fix PR middle-end/PR28690, modify swap_commutative_operands_p

["H.J. Lu" <hjl at lucon dot org>]

On Fri, Jul 13, 2007 at 11:59:43PM +0200, Richard Guenther wrote:
> On 7/13/07, H.J. Lu <hjl@lucon.org> wrote:
> >On Fri, Jul 13, 2007 at 11:43:22PM +0200, Paolo Bonzini wrote:
> >>
> >> >I don't know even know which changes should be guarded by a target
> >> >macro since there are quite a few independent changes in that
> >> >patch.
> >>
> >> Two hunks are bugfixes.  The other two (in rtlanal.c) are the only
> >> targets that could affect performance but, again, it's just luck or bad
> >> luck.
> >
> >Then bug fixes should be submitted with testcases separately from
> >the performance change for Power 6.
> >
> >>
> >> I don't see the addressing modes in x86 that well tuned.  Surely not the
> >> tuning that Pat and Peter are doing, and not a kind of tuning that could
> >> be incompatible.
> >>
> >
> >We are working hard to make sure that gcc won't getting worse on
> >x86 :-(.
> 
> There is more volatileness in x86 performance numbers already than
> the predicted change by this.  For example I see both leslie3d and
> GemmsFDTD dropped 30% / 20% around Jul 5th.  I would rather
> wait for a quieter period to analyze the cummulative effect of all
> the changes we piled up.  Fixing them afterwards is probably better
> than trying to block changes that cannot definitely explain the
> seen change.

The cummulative effects of all changes can add up. The proposed
patch has many indepedent changes. I believe they should be broken
up and submitted separately.


H.J.