Re: disable ivopts by default for 4.0?

Re: disable ivopts by default for 4.0?

[Nathanael Nerode]

>I think I fixed all wrong-code regressions for ivopts
...
>Under these circumstances, I do not see how disabling ivopts could
>somehow dramatically improve situation.

Quite right.  Suggestion withdrawn.  :-)

-- 
This space intentionally left blank.

disable ivopts by default for 4.0?

[Nathanael Nerode]

It seems to be the cause of a lot of the remaining 4.0 regressions, including
many of the wrong-code regressions.

Disabling it by default (rendering the ivopts bugs non-regressions) would cut
the regression list down to a reasonable and more manageable length.  Would
it generate any new regressions?  I doubt it.

-- 
This space intentionally left blank.

Re: disable ivopts by default for 4.0?

[Giovanni Bajo]

Nathanael Nerode <neroden@fastmail.fm> wrote:

> Disabling it by default (rendering the ivopts bugs non-regressions)
> would cut the regression list down to a reasonable and more
> manageable length.  Would it generate any new regressions?  I doubt
> it.

Only if you prove that it does not generate any codegen regression from older
versions of GCC, on any major platform. I do not think that starting disabling
problematic passes in GCC is a good shortcut to get the release out.

> It seems to be the cause of a lot of the remaining 4.0 regressions,
> including many of the wrong-code regressions.

I would debate this as well. There are around 200 regressions now, 80 of which
are already 3.4 or 3.3 regressions so they are not surely related of ivopts. Of
the other 120, I would be surprised if more than 15 are caused by ivopts. You
can setup a ivopts meta-bug, link all the regressions ivopts causes to this
meta-bug, and get back to us with some numbers.

Giovanni Bajo

Re: disable ivopts by default for 4.0?

[Zdenek Dvorak]

Hello,

> It seems to be the cause of a lot of the remaining 4.0 regressions, including
> many of the wrong-code regressions.

I think I fixed all wrong-code regressions for ivopts (in particular,
none of the PR's you have gathered in PR18693 is a misscompilation).  If
you have pointers for some not fixed (and not with pending patch),
please let me know.

In particular, disabling ivopts would "fix" exactly two of the PR's gathered under
the meta-PR:

PR 18595: serious, but not critical, compile-time problems with ivopts.
PR 18048: 10% performance regression on mgrid

And partially help with

PR 18549: Code size increase.  It would probably make sense to disable
  ivopts at -0s for now.

For the remaining PR's under the meta-bug, the part of problems in ivopts
is already fixed.

Under these circumstances, I do not see how disabling ivopts could
somehow dramatically improve situation.

Zdenek