gfortran/mpich2 uninitialized variables on non-root processes

[W Spector]

Janne,

This morning I updated my gfortran trunk snapshot (to v4.10.0 20140524) 
and tested your mod with mpich2 mpirun again.  It appears that valgrind 
is happy now.  Thank you!

Walter

On 05/14/2014 04:42 PM, Janne Blomqvist wrote:
> On Wed, May 14, 2014 at 5:55 PM, W Spector <w6ws@earthlink.net> wrote:
>> Janne,
>>
>> Thank you for verifying this, isolating the issue, and creating the PR.  I
>> tried the test with OpenMPI and all was fine.  So I figured I should dig a
>> little deeper before informing the mpich2 folks.  Just haven't had time
>> lately.  Based on your analysis, maybe OpenMPI does not close stdin like
>> mpich2/mvapch2 do?
>
> Based on some quick googling, it seems openmpi redirects stdin to
> /dev/null on all but one rank (default 0, see -stdin option to
> mpirun). Arguably the MPICH behavior is in fact better, as that should
> generate an error immediately if one tries to read stdin on the wrong
> rank, rather than reading a bunch of nulls and the application failing
> mysteriously later.
>
>> The project I work on has recently started running our test base under
>> valgrind on a regular basis.  Although benign, these warnings add a lot of
>> noise/false positives to the output.
>
> Thanks for the testing, and the report. I have posted a patch which I
> believe fixes this issue.
>
>
>
>