glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

[Carlos O'Donell]

Community,

The next release of glibc 2.28 is planned for August 1st.
We normally enter a global ABI/API slushy freeze on July 1st.

Both IBM and Red Hat have a strong desire to see the IBM ppc64le
transition to the new long double format without having to wait an
additional 6 months until we have another release.

For example I want these changes in Fedora 29 so we can showcase the 
latest IBM POWER9 hardware.

I'm willing to extend the ABI/API freeze to July 13th for the ppc64le
float128 changes, to allow those to go into the glibc 2.28 release.

This would leave us with 2 weeks of testing for most arches, and I
can make sure we get good testing coverage during those two weeks from
the set of hardware Red Hat has access to, and the Red Hat glibc team
doing builds and tests on x864_64, i686, ppc64be, ppc64le, s390, s390x,
armv7hl, and aarch64.

I'm not suggesting we move the August 1st release, but just decrease
the window of the ABI/API freeze, and restrict it to only the changes
we accept as such e.g. ppc64le float128 changes.

Does anyone object to this?

Cheers,
Carlos.

Re: glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

[Florian Weimer]

* Carlos O'Donell:

> I'm willing to extend the ABI/API freeze to July 13th for the ppc64le
> float128 changes, to allow those to go into the glibc 2.28 release.

Do you expect these changes to be specific to the ppc64le ABI, or
might there be impact on other ABIs?

Re: glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

[Tulio Magno Quites Machado Filho]

Florian Weimer <fw@deneb.enyo.de> writes:

> * Carlos O'Donell:
>
>> I'm willing to extend the ABI/API freeze to July 13th for the ppc64le
>> float128 changes, to allow those to go into the glibc 2.28 release.
>
> Do you expect these changes to be specific to the ppc64le ABI, or
> might there be impact on other ABIs?

They're all specific to the ppc64le ABI.

-- 
Tulio Magno

Re: glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

[Joseph Myers]

On Tue, 26 Jun 2018, Tulio Magno Quites Machado Filho wrote:

> Florian Weimer <fw@deneb.enyo.de> writes:
> 
> > * Carlos O'Donell:
> >
> >> I'm willing to extend the ABI/API freeze to July 13th for the ppc64le
> >> float128 changes, to allow those to go into the glibc 2.28 release.
> >
> > Do you expect these changes to be specific to the ppc64le ABI, or
> > might there be impact on other ABIs?
> 
> They're all specific to the ppc64le ABI.

Handling float128 for argp.h, err.h and error.h functions, while keeping 
using the same machinery for header redirections for both float128 long 
double and -mlong-double-64 and keeping the machinery as consistent as 
possible between different headers, means that all ldbl-opt platforms will 
get new symbols __nldbl_argp_error, __nldbl_argp_failure, __nldbl_warn, 
__nldbl_vwarn, __nldbl_warnx, __nldbl_vwarnx, __nldbl_err, __nldbl_verr, 
__nldbl_errx, __nldbl_verrx, __nldbl_error, __nldbl_error_at_line.  
(These are the functions I know about that were missed when the ldbl-opt 
machinery was originally set up, but I don't guarantee that list is 
complete.)  Thus, as well as powerpc64le, all other powerpc systems are 
affected, as are alpha, s390 and 32-bit sparc.

(See what I said in 
<https://sourceware.org/ml/libc-alpha/2018-05/msg00803.html> regarding the 
general strategy of sharing bits/*ldbl.h headers between the different 
cases of long double redirections, with new macro variants being added if 
necessary to handle cases that need to be distinguished for float128 long 
double but not for the present -mlong-double-64 redirections.)

-- 
Joseph S. Myers
joseph@codesourcery.com

Re: glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

[Tulio Magno Quites Machado Filho]

Joseph Myers <joseph@codesourcery.com> writes:

> On Tue, 26 Jun 2018, Tulio Magno Quites Machado Filho wrote:
>
>> Florian Weimer <fw@deneb.enyo.de> writes:
>> 
>> > * Carlos O'Donell:
>> >
>> >> I'm willing to extend the ABI/API freeze to July 13th for the ppc64le
>> >> float128 changes, to allow those to go into the glibc 2.28 release.
>> >
>> > Do you expect these changes to be specific to the ppc64le ABI, or
>> > might there be impact on other ABIs?
>> 
>> They're all specific to the ppc64le ABI.
>
> Handling float128 for argp.h, err.h and error.h functions, while keeping 
> using the same machinery for header redirections for both float128 long 
> double and -mlong-double-64 and keeping the machinery as consistent as 
> possible between different headers, means that all ldbl-opt platforms will 
> get new symbols __nldbl_argp_error, __nldbl_argp_failure, __nldbl_warn, 
> __nldbl_vwarn, __nldbl_warnx, __nldbl_vwarnx, __nldbl_err, __nldbl_verr, 
> __nldbl_errx, __nldbl_verrx, __nldbl_error, __nldbl_error_at_line.  
> (These are the functions I know about that were missed when the ldbl-opt 
> machinery was originally set up, but I don't guarantee that list is 
> complete.)  Thus, as well as powerpc64le, all other powerpc systems are 
> affected, as are alpha, s390 and 32-bit sparc.

Indeed.  I didn't take that into account.

Thanks!

-- 
Tulio Magno

Re: glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

[Szabolcs Nagy]

On 26/06/18 20:32, Carlos O'Donell wrote:
> I'm willing to extend the ABI/API freeze to July 13th for the ppc64le
> float128 changes, to allow those to go into the glibc 2.28 release.
> 

i just posted math patches which would be nice to have in 2.28

there are two parts:
- the first does not change abi and provides new implementations
with significant performance improvements,
- the second changes abi by adding new symbol versions that don't
do SVID error handling and further improves performance a bit.

these require non-trivial amount of review and testing and i'd
like to extend the review period a bit before the freeze for
these too.  i think the first part has a good chance to get
into an acceptable state.  the second part is independent but
would be nice to get it in 2.28 too.

thoughts?

Re: glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

[Florian Weimer]

On 06/29/2018 04:33 PM, Szabolcs Nagy wrote:
> On 26/06/18 20:32, Carlos O'Donell wrote:
>> I'm willing to extend the ABI/API freeze to July 13th for the ppc64le
>> float128 changes, to allow those to go into the glibc 2.28 release.
>>
> 
> i just posted math patches which would be nice to have in 2.28
> 
> there are two parts:
> - the first does not change abi and provides new implementations
> with significant performance improvements,
> - the second changes abi by adding new symbol versions that don't
> do SVID error handling and further improves performance a bit.
> 
> these require non-trivial amount of review and testing and i'd
> like to extend the review period a bit before the freeze for
> these too.  i think the first part has a good chance to get
> into an acceptable state.  the second part is independent but
> would be nice to get it in 2.28 too.
> 
> thoughts?

I'm generally sympathetic to this, but we have a lot of unreviewed 
math-related patches that should go into 2.28.  Would you be able to 
help with review of the other patches?

Thanks,
Florian

Re: glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

[Szabolcs Nagy]

On 29/06/18 15:35, Florian Weimer wrote:
> On 06/29/2018 04:33 PM, Szabolcs Nagy wrote:
>> On 26/06/18 20:32, Carlos O'Donell wrote:
>>> I'm willing to extend the ABI/API freeze to July 13th for the ppc64le
>>> float128 changes, to allow those to go into the glibc 2.28 release.
>>>
>>
>> i just posted math patches which would be nice to have in 2.28
>>
>> there are two parts:
>> - the first does not change abi and provides new implementations
>> with significant performance improvements,
>> - the second changes abi by adding new symbol versions that don't
>> do SVID error handling and further improves performance a bit.
>>
>> these require non-trivial amount of review and testing and i'd
>> like to extend the review period a bit before the freeze for
>> these too.  i think the first part has a good chance to get
>> into an acceptable state.  the second part is independent but
>> would be nice to get it in 2.28 too.
>>
>> thoughts?
> 
> I'm generally sympathetic to this, but we have a lot of unreviewed math-related patches that should go into 2.28.  Would you be able to help 
> with review of the other patches?
> 

yes, i can help with reviews next week.

Re: glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

[Maciej W. Rozycki]

On Tue, 26 Jun 2018, Carlos O'Donell wrote:

> I'm not suggesting we move the August 1st release, but just decrease
> the window of the ABI/API freeze, and restrict it to only the changes
> we accept as such e.g. ppc64le float128 changes.

 Can I have:

<https://patchwork.sourceware.org/patch/27910/>
<https://sourceware.org/ml/libc-alpha/2018-06/msg00511.html>

decided for 2.28?  Unfortunately I only discovered it so late in the 
development cycle, because I have only recently dived into addressing 
binutils PR ld/21375 and discovered BZ #23307 in the course.

 The patch blocks the fix for PR ld/21375, and it will probably make no 
sense once we have a release with absolute symbols working (BZ #19818 + BZ 
#23307), but no ABI marking for that.  So I think it's a 2.28-or-never 
change.  The linker-side fix is scheduled for Jan 2019 binutils release.

 Shall I file a separate BZ just for this patch, with the milestone set 
to 2.28?

  Maciej

Re: glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

[Joseph Myers]

On Mon, 2 Jul 2018, Maciej W. Rozycki wrote:

>  Shall I file a separate BZ just for this patch, with the milestone set 
> to 2.28?

Milestones should only be set on FIXED bugs; they aren't how we plan for 
what we want in what release.  See point M1 at 
<https://sourceware.org/glibc/wiki/Bugzilla%20Procedures>.

-- 
Joseph S. Myers
joseph@codesourcery.com

Re: glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

[Maciej W. Rozycki]

On Mon, 2 Jul 2018, Joseph Myers wrote:

> >  Shall I file a separate BZ just for this patch, with the milestone set 
> > to 2.28?
> 
> Milestones should only be set on FIXED bugs; they aren't how we plan for 
> what we want in what release.  See point M1 at 
> <https://sourceware.org/glibc/wiki/Bugzilla%20Procedures>.

 Ack.

  Maciej

Re: glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

[Carlos O'Donell]

On 07/02/2018 09:59 AM, Maciej W. Rozycki wrote:
> On Tue, 26 Jun 2018, Carlos O'Donell wrote:
> 
>> I'm not suggesting we move the August 1st release, but just decrease
>> the window of the ABI/API freeze, and restrict it to only the changes
>> we accept as such e.g. ppc64le float128 changes.
> 
>  Can I have:
> 
> <https://patchwork.sourceware.org/patch/27910/>
> <https://sourceware.org/ml/libc-alpha/2018-06/msg00511.html>
> 
> decided for 2.28?  Unfortunately I only discovered it so late in the 
> development cycle, because I have only recently dived into addressing 
> binutils PR ld/21375 and discovered BZ #23307 in the course.

This is OK for 2.28.

>  The patch blocks the fix for PR ld/21375, and it will probably make no 
> sense once we have a release with absolute symbols working (BZ #19818 + BZ 
> #23307), but no ABI marking for that.  So I think it's a 2.28-or-never 
> change.  The linker-side fix is scheduled for Jan 2019 binutils release.

I have reviewed the thread between you and Florian and I agree that this
change can go into 2.28.

>  Shall I file a separate BZ just for this patch, with the milestone set 
> to 2.28?

Answered by Joseph.

-- 
Cheers,
Carlos.

Re: glibc 2.28 release on August 1st and extended ABI freeze for ppc64le float128.

[Tulio Magno Quites Machado Filho]

Carlos O'Donell <carlos@redhat.com> writes:

> The next release of glibc 2.28 is planned for August 1st.
> We normally enter a global ABI/API slushy freeze on July 1st.
>
> Both IBM and Red Hat have a strong desire to see the IBM ppc64le
> transition to the new long double format without having to wait an
> additional 6 months until we have another release.
>
> For example I want these changes in Fedora 29 so we can showcase the 
> latest IBM POWER9 hardware.
>
> I'm willing to extend the ABI/API freeze to July 13th for the ppc64le
> float128 changes, to allow those to go into the glibc 2.28 release.

We've decided to delay this transition to glibc 2.29 in order to give more
time for tests.

So, I'm removing this item from the release blockers list.

Thank you all!

-- 
Tulio Magno