[mpich-discuss] Errors with alloca

[Eric A. Borisch]

On Tue, Oct 23, 2012 at 2:05 PM, Dave Goodell <goodell at mcs.anl.gov> wrote:
> On Oct 23, 2012, at 1:48 PM CDT, Eric A. Borisch wrote:
>
>> I filed this (https://trac.mcs.anl.gov/projects/mpich2/ticket/1704 --
>> "MPI_Reduce breaks with --enable-alloca at some sizes") and I was just
>> wondering if anyone has had a chance to look at it.
>
> I took a quick peek at it, but decided it was of very low priority.  Just don't configure with "--enable-alloca" for now.  Have you seen substantial speed improvements from using it that would justify development effort to remedy this in the near term?
>
> I wasn't aware that anyone was even using this switch and assumed it was just leftover from some past research effort...
>
> -Dave

To be honest I haven't benchmarked it recently; I thought I had
observed some (likely tiny) improvement at some point, and it is
available as a configure (presumably performance-related) option, so I
went for it. (We are using MPI to go fast, right?)

It's one of those nasty bugs that works most of the time, so long as
the buffers aren't too big, and the remaining stack space too small.

I won't shed a tear if the solution is to remove the configure option
(or at least marked unsupported.) Just trying to save the next user
some debug time.

 - Eric