Rob, I applied the patch and the analysis doesn't report any scale-dependent overflow now. Thanks! This fixes it. If you are interested, I could try the analysis on other parts of mpich that perhaps are not frequently used -- I only tested these functions: broadcast, scatter, gather, reduce, alltoall, gatherv, isend, irecv, and wait. I didn't try I/O functions for example. We could coordinate this in this forum or in a separate email. Just let me know. Unfortunately, until we release it, it can only be tested on my side. Thanks for your help! Ignacio On 10/2/15 11:50 AM, Rob Latham wrote:

On 10/01/2015 12:35 PM, Ignacio Laguna wrote:

...

Once you have a version I can download with the patches, I can run the analysis again and see if I can catch more things. Just let me know and I would be happy to test it.

I tried to decipher the LLVM output: I see a few more i64 values, but not certain I'm reading it correctly.

Does this patch do the trick?

Promoting the declaration of recvcount would work too, but sets off a larger cascade of type promotions.

==rob

...

In general, overflows can also occur in intermediate computations, depending on how code is optimized, I believe.

Thanks!

Ignacio

...

==rob

_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss