[petsc-dev] adding -fopenmp three times

[Barry Smith]

  Test it and if it works on Apple it is probably ok. I don’t remember why I needed to add it but presumably otherwise it didn’t work :-)


On Nov 12, 2013, at 7:16 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:

> Matthew Knepley <knepley at gmail.com> writes:
> 
>> I am a little worried. I think Satish put this in there because something
>> failed. This is a big
>> problem with configure. We really need to reference a certain test (even if
>> we cannot
>> replicate it) when these changes go in.
>> 
>> The central problem is that there is no internal logic to this. It is just
>> a set of jumbled facts
>> and scenarios. And we cannot fix it, we just deal. This is also the web
>> browser problem.
> 
> This is why configure is difficult, though I think a lot of the kludgy
> code that goes into BuildSystem comes from not understanding the precise
> semantics of the framework.
> 
> Barry, do you know what was the rationale for adding the addLinkerFlag
> here?
> 
> https://bitbucket.org/petsc/petsc/commits/504565928e62d2a531f1584c908a5465d42e9a14
> 
> As Matt says (but is easily forgotten), compiler flags are passed to the
> linker unless they are excluded.
> 
> 
> The issue here is that with the code in the repository now, -fopenmp
> gets passed three times on the command line.  It would be less cluttered
> to pass it just once, which the diff below achieves.
> 
>> 
>>  Matt
>> 
>> 
>> On Tue, Nov 12, 2013 at 6:31 PM, Jed Brown <jed at jedbrown.org> wrote:
>> 
>>> Matthew Knepley <knepley at gmail.com> writes:
>>> 
>>>> On Tue, Nov 12, 2013 at 5:36 PM, Jed Brown <jed at jedbrown.org> wrote:
>>>> 
>>>>> How do we make fopenmp only appear once?
>>>>> 
>>>> 
>>>> I can already explain two of these I think. The compiler gets -openmp,
>>> but
>>>> so does the linker,
>>>> and so the link gets both (by our convention that the linker gets any
>>>> compiler flag no explicitly
>>>> excluded)
>>> 
>>> Okay to make the change below?  (With this patch, the flag appears only
>>> once.)
>>> 
>>> diff --git i/config/PETSc/packages/openmp.py
>>> w/config/PETSc/packages/openmp.py
>>> index c788d8a..e93217f 100644
>>> --- i/config/PETSc/packages/openmp.py
>>> +++ w/config/PETSc/packages/openmp.py
>>> @@ -39,8 +39,6 @@ class Configure(PETSc.package.NewPackage):
>>>         ompflag = flag
>>>         break
>>>     self.setCompilers.addCompilerFlag(ompflag)
>>> -    if self.setCompilers.checkLinkerFlag(ompflag):
>>> -      self.setCompilers.addLinkerFlag(ompflag)
>>>     self.setCompilers.popLanguage()
>>>     if hasattr(self.compilers, 'FC'):
>>>       self.setCompilers.pushLanguage('FC')
>>> @@ -49,8 +47,6 @@ class Configure(PETSc.package.NewPackage):
>>>     if hasattr(self.compilers, 'CXX'):
>>>       self.setCompilers.pushLanguage('Cxx')
>>>       self.setCompilers.addCompilerFlag(ompflag)
>>> -      if self.setCompilers.checkLinkerFlag(ompflag):
>>> -        self.setCompilers.addLinkerFlag(ompflag)
>>>       self.setCompilers.popLanguage()
>>>     # OpenMP threadprivate variables are not supported on all platforms
>>> (for e.g on MacOS).
>>>     # Hence forcing to configure additionally with --with-pthreadclasses
>>> so that pthread
>>> 
>>> 
>> 
>> 
>> -- 
>> What most experimenters take for granted before they begin their
>> experiments is infinitely more interesting than any results to which their
>> experiments lead.
>> -- Norbert Wiener