State of gfortran regressions

State of gfortran regressions

[Thomas Koenig]

Hello world,

even if the gcc 8 release is still some time away, we should consider
spending some effort on fixing the regressions.

Here is a current list of the gcc-8 regressions, with some comments from
my side.

PR 82976 - non-trivial conversion at assignment since r254526
This one is a bit of a mystery to debug - it happens, it is
reproducible, but the *.original tree dump shows nothing.

PR 83999 - an ICE-on-valid with a class pointer assignment.
Has not yet been bisected to a revision.

PR 84135 - a coarray issue, maybe fairly easy to fix, a missing
NULL check.

PR 84615 - it seems that, with the size_t string lengths, -m32 got
broken for character arguments:-(  This is actually a fairly serious
one. If we cannot fix that one in time for the gcc-8 release, we should
consider reverting the new string length patch. I'm going to take a
look, but I won't be offended if anybody else beats me to it :-)

PR 82972 and 83088 - two -finit-derived bugs. Fritz, you have assigned
the first one to yourself. What are your plans with this?

PR 84487. Another mystery - rodata increases a lot for a test case.
I have compared the assembly listings, and I don't understand
why that happens.

PR 84848 - a HP-PA failure. So far, no gfortran maintainer
has reproduced this on a system he has access to.

PR 83705, a corner case where REPEAT doesn't work for incredibly
large values when simplifying. Probably the least important of
all the bugs.

There are also 21 7/8 regressions, which I will also look through
to see if there is anything that I could get a handle on.

Comments?

Regards

	Thomas

Re: State of gfortran regressions

[Paul Richard Thomas]

Hi Thomas,

I am waiting for a flight home. I'll get back in touch tomorrow sometime.

Cheers

Paul


On 16 March 2018 at 14:57, Thomas Koenig <tkoenig@netcologne.de> wrote:
> Hello world,
>
> even if the gcc 8 release is still some time away, we should consider
> spending some effort on fixing the regressions.
>
> Here is a current list of the gcc-8 regressions, with some comments from
> my side.
>
> PR 82976 - non-trivial conversion at assignment since r254526
> This one is a bit of a mystery to debug - it happens, it is
> reproducible, but the *.original tree dump shows nothing.
>
> PR 83999 - an ICE-on-valid with a class pointer assignment.
> Has not yet been bisected to a revision.
>
> PR 84135 - a coarray issue, maybe fairly easy to fix, a missing
> NULL check.
>
> PR 84615 - it seems that, with the size_t string lengths, -m32 got
> broken for character arguments:-(  This is actually a fairly serious
> one. If we cannot fix that one in time for the gcc-8 release, we should
> consider reverting the new string length patch. I'm going to take a
> look, but I won't be offended if anybody else beats me to it :-)
>
> PR 82972 and 83088 - two -finit-derived bugs. Fritz, you have assigned
> the first one to yourself. What are your plans with this?
>
> PR 84487. Another mystery - rodata increases a lot for a test case.
> I have compared the assembly listings, and I don't understand
> why that happens.
>
> PR 84848 - a HP-PA failure. So far, no gfortran maintainer
> has reproduced this on a system he has access to.
>
> PR 83705, a corner case where REPEAT doesn't work for incredibly
> large values when simplifying. Probably the least important of
> all the bugs.
>
> There are also 21 7/8 regressions, which I will also look through
> to see if there is anything that I could get a handle on.
>
> Comments?
>
> Regards
>
>         Thomas



-- 
"If you can't explain it simply, you don't understand it well enough"
- Albert Einstein

Re: State of gfortran regressions

[Damian Rouson]

 

Hi Paul,  

I’m hoping an OpenCoarrays regression caused by the descriptor changes in gfortran  
commit r257065 can also be addressed before GCC 8 releases. With guidance from Alessandro,  
I was able to adjust OpenCoarrays to get all of the resulting test failures eliminated
except for the one identified in OpenCoarrays issue #518. I would greatly appreciate if  
you could assist with eliminating this one remaining failure.  Alessandro started a new job  
last month and will not be able to contribute to gfortran or OpenCoarrays for the foreseeable  
future.  Soren might be able to look into this, but he’s working on another issue first.
I’m thinking you might be the best person to help because you made the descriptor changes:

Here’s the test that fails:  

https://github.com/sourceryinstitute/OpenCoarrays/blob/master/src/tests/regression/reported/issue-422-send-get.F90  

and here's the issue set up to track work on this failure:  

https://github.com/sourceryinstitute/OpenCoarrays/issues/518  

Damian  



On March 16, 2018 at 1:34:16 PM, Paul Richard Thomas (paul.richard.thomas@gmail.com(mailto:paul.richard.thomas@gmail.com)) wrote:

> Hi Thomas,
>  
> I am waiting for a flight home. I'll get back in touch tomorrow sometime.
>  
> Cheers
>  
> Paul
>  
>  
> On 16 March 2018 at 14:57, Thomas Koenig wrote:
> > Hello world,
> >
> > even if the gcc 8 release is still some time away, we should consider
> > spending some effort on fixing the regressions.
> >
> > Here is a current list of the gcc-8 regressions, with some comments from
> > my side.
> >
> > PR 82976 - non-trivial conversion at assignment since r254526
> > This one is a bit of a mystery to debug - it happens, it is
> > reproducible, but the *.original tree dump shows nothing.
> >
> > PR 83999 - an ICE-on-valid with a class pointer assignment.
> > Has not yet been bisected to a revision.
> >
> > PR 84135 - a coarray issue, maybe fairly easy to fix, a missing
> > NULL check.
> >
> > PR 84615 - it seems that, with the size_t string lengths, -m32 got
> > broken for character arguments:-( This is actually a fairly serious
> > one. If we cannot fix that one in time for the gcc-8 release, we should
> > consider reverting the new string length patch. I'm going to take a
> > look, but I won't be offended if anybody else beats me to it :-)
> >
> > PR 82972 and 83088 - two -finit-derived bugs. Fritz, you have assigned
> > the first one to yourself. What are your plans with this?
> >
> > PR 84487. Another mystery - rodata increases a lot for a test case.
> > I have compared the assembly listings, and I don't understand
> > why that happens.
> >
> > PR 84848 - a HP-PA failure. So far, no gfortran maintainer
> > has reproduced this on a system he has access to.
> >
> > PR 83705, a corner case where REPEAT doesn't work for incredibly
> > large values when simplifying. Probably the least important of
> > all the bugs.
> >
> > There are also 21 7/8 regressions, which I will also look through
> > to see if there is anything that I could get a handle on.
> >
> > Comments?
> >
> > Regards
> >
> > Thomas
>  
>  
>  
> --
> "If you can't explain it simply, you don't understand it well enough"
> - Albert Einstein