Re: packages requiring Intel's libmkl_rt library

[Brook Milligan <brook%nmsu.edu@localhost>]

> On Feb 8, 2018, at 3:15 PM, Joerg Sonnenberger <joerg%bec.de@localhost> wrote:
> 
> On Tue, Feb 06, 2018 at 10:50:16AM -0700, Brook Milligan wrote:
>> I am installing math/py-numpy on a system with Intel compilers.  It ends up requiring the libmkl_rt library, which is in /opt/intel/..., which is exactly where the path in the relevant numpy library points.  However, the checks in mk/check/check-shlibs.mk indicate that it is missing.
>> 
>> How is this supposed to be handled so that libraries like this can be found?
> 
> Do you have that directory on the BUILDLINK_PASSTHRU_RPATHDIRS list?

Yes, I have tried both that and BUILDLINK_PASSTHRU_DIRS in both the package Makefile and etc/mk.conf.  That does not help.

My understanding (correct me if I am wrong) is that BUILDLINK_PASSTHRU_DIRS AND BUILDLINK_PASSTHRU_RPATHDIRS signal that the wrappers should not filter those directories out and so whatever is being built can in a sense reach out into those parts of the system rather than be limited to the buildlink restrictions.

My interpretation (based upon this Makefile comment # This package bypasses the wrappers and calls the compilers directly when linking) is that py-numpy is avoiding the wrappers altogether.  If that is true, then I expect BUILDLINK_PASSTHRU_* to have no impact, which I what I am seeing.

That is why I have opted to use CHECK_SHLIBS_SKIP.  I believe the libraries (now narrowed down to three) are being built just fine; just the checks are over ambitious.

If it is ok to skip these checks unconditionally, there is a simple fix to the package Makefile.  However, I suspect a preference that these checks be performed under "normal" circumstances and thus am seeking a way to make the skipping conditional.  Please suggest suitable make variables that could be used in a conditional expression to only skip the checks on systems that are using an Intel compiler with libmkl_rt.

Perhaps this is too complicated, though, and we should just always skip the checks on the three offending libraries and be confident that no other odd rpaths will creep in.

Please advise.

Thank you very much.

Cheers,
Brook