Am 22.06.2019 um 20:09 schrieb Jeremie Courreges-Anglas:
> On Thu, Jun 06 2019, Martin Reindl <martin@catai.org> wrote:
>> On Sat, Jun 01, 2019 at 08:03:17PM +0100, Stuart Henderson wrote:
>>> That needs fixing then..
>>> --
>>> Sent from a phone, apologies for poor formatting.
>>>
>>> On 1 June 2019 17:15:19 Jeremie Courreges-Anglas <jca@wxcvbn.org> wrote:
>>>
>>>> On Sat, Jun 01 2019, Stuart Henderson <stu@spacehopper.org> wrote:
>>>>> Please don't do the huge bump for SHARED_LIBS, just a standard major
>>>>> bump for the existing libraries and start the new ones at 0.0
>>>>
>>>> IIRC the problem is that SHARED_LIBS isn't respected.
>>>>
>>
>> Here is a polished version of the diff which should pass strict inspection
>> by the ports@ team:
>>
>> - respects SHARED_LIBS by using base libtool, m4 from devel/libtool and -ltdl
>> - c++ interface is deprecated and not enabled in contrast to 1.0.4
>
> Maybe this should use COMPILER_LANGS = c, then. make
> port-lib-depends-check complains about 'Extra: estdc++.19' on sparc64.
>
>> - now uses SEPERATE-BUILD
>> - passes make test on macppc, arm64 and amd64 and schedules jobs on virtual
>> amd64 mini-mpi-cluster
>> - now uses egfortran from ports-gcc instead of f77
>
>> - tidy up Makefile a bit so things look more in order for me
>
> Well, that's a bit of churn and it still doesn't match the canonical
> order recommended by Makefile.template. SHARED_LIBS and MASTER_SITES
> look completely out of place.
>
> Except for this point, this update looks good. make package fails on
> sparc64 because lib/openmpi/mca_patcher_overwrite.* are arch-specific.
> Here's an updated diff that uses the order of Makefile.template and
> @comments out the arch-specific files.
>
> Additional tweaks:
> - use https HOMEPAGE/MASTER_SITES
> - PERMIT_PACKAGE_CDROM -> PERMIT_PACKAGE
> - strip "base-gcc" from COMPILER, it's not needed/useful
> - use V=1 in MAKE_FLAGS, looks like --disable-silent-rules isn't passed
> down to all configure scripts.
>
> The remaining patch is still a bit broken, comparisons like
>
> "if (fsbuf.f_fstypename == MOUNT_NFS) {"
>
> can't yield true since MOUNT_NFS is string literal "nfs" and
> fsbuf.f_fstypename is a char array in a struct. I'm not sure how much
> we care, though, at least it builds...
>
> What do you think? Yay, nay? :)
Well, yes, this is an improvement, thank you for the additional work.
Things like the NFS problem remain (but openmpi can just run with ssh keys).
And there is a pthread_mutexattr_setpshared() (which we don't have)
problem deep down in pmix which needs to be investigated at some point.
What is the status with
pthread_mutexattr_setpshared()/pthread_mutexattr_getpshared these days?
-m
No comments:
Post a Comment