[130954] trunk/dports/python/py-numpy/Portfile

[Joshua Root]

On 2015-1-5 07:26 , Sean Farley wrote:
> 
> Joshua Root writes:
> 
>>> Revision: 130954
>>>           https://trac.macports.org/changeset/130954
>>> Author:   sean at macports.org
>>> Date:     2015-01-03 22:01:06 -0800 (Sat, 03 Jan 2015)
>>> Log Message:
>>> -----------
>>> py-numpy: refactor port to use compilers portgroup
>>>
>>> This is a big refactor that cleans up a lot of old and now incorrect logic,
>>> mostly with ATLAS. This changeset is accomplishes:
>>>
>>> - ensure that numpy has a fortran compiler
>>> - ensure that compilers match the same ones as ATLAS
>>> - remove python 3.2
>>>
>>> Modified Paths:
>>> --------------
>>>     trunk/dports/python/py-numpy/Portfile
>>>
>>> Modified: trunk/dports/python/py-numpy/Portfile
>>> ===================================================================
>>> --- trunk/dports/python/py-numpy/Portfile	2015-01-04 06:01:04 UTC (rev 130953)
>>> +++ trunk/dports/python/py-numpy/Portfile	2015-01-04 06:01:06 UTC (rev 130954)
>>> @@ -4,9 +4,11 @@
>>>  PortSystem              1.0
>>>  PortGroup               python 1.0
>>>  PortGroup               github 1.0
>>> +PortGroup               compilers 1.0
>>>  
>>>  github.setup            numpy numpy 1.9.1 v
>>>  name                    py-numpy
>>> +revision                1
>>>  categories-append       math
>>>  license                 BSD
>>>  platforms               darwin
>>> @@ -17,8 +19,10 @@
>>>  checksums               rmd160  c043ce0045166762e36c513458abc499d36bbcf5 \
>>>                          sha256  504b7cbc3b47df592d683e0f3b992cfa4d29338f8d23fca07f040164018b607f
>>>  
>>> -python.versions         26 27 32 33 34
>>> +python.versions         26 27 33 34
>>>  
>>> +compilers.setup         require_fortran -clang -dragonegg -gcc44 -gcc45 -gcc46 -g95
>>
>> Is it really necessary to make all numpy users install gcc? Anyone who
>> is just using numpy arrays for something like py-opengl is going to find
>> it a large and useless extra dependency.
> 
> As far as I can tell, numpy requires a fortran compiler. Not installing
> one is what leads to errors in #43321. Any port touching computation
> (i.e. blas/lapack) almost certainly will have to install some kind of
> fortran compiler.

Surely the active_variants checks in py-scipy would prevent that now?

> I think this kind of dependency is unavoidable if we want to maintain
> consistency and reproducibility. :-/

I think not being able to compile fortran when you explicitly specify
-gfortran is consistent and reproducible. :-)

- Josh