This is the mail archive of the fortran@gcc.gnu.org mailing list for the GNU Fortran project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Testsuite failures with multi-file tests in gfortran.dg


On Wed, Feb 20, 2013 at 11:37 AM, Tobias Burnus <burnus@net-b.de> wrote:
> Diego Novillo wrote:
>>
>> I am running into some failures in gfortran.dg with tests that rely on
>> dejagnu's compilation order.  This problem occurs in my environment
>> because I split testing across several nodes in a cluster, each of
>> which executes a subset of the testsuite.
>
>
> I didn't know that such a splitting was possible - I only know Jakub's
> manual splitting for "make -j...", which are not affected.

It is.  DejaGNU can be told to execute a subset of tests inside a
directory.  That's what makes distribution so effective in reducing
test times.  The main trick is to replicate the testsuite trees so
that auxiliary files can be found.  But if the .exp file relies on the
order of which tests are executed then this trick can break down (as
it does in this case).

Jakub's parallelization works because it does not split directories.
This makes it less balanced (some directories are very large).

> See http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56408 for the tracking bug
> and my attempts to fix this issue. :-(
>
> Suggestions how to fix this are welcome.

Thanks for opening a bug report.  My approach would be to replicate
the scheme that lto.exp and pch.exp implement.  They explicitly
compile the dependencies (say, all the files named '*_0.c') first.
But I have not looked at how gfortran.exp is organized.  Perhaps split
the module tests into a separate directory and have them controlled by
a separate .exp driver?


Diego.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]