sync subarray and darray behaviour

[gpapaure]

Basic usage works again for both subarray and darray by using create_hindexed instead of create_resized
Test case cited in http://lists.mpich.org/pipermail/discuss/2015-January/003608.html still works

[jsquyres]

@bosilca can you comment? Thanks.

[ggouaillardet]

is this related to #965 ?

[gpapaure]

I haven't seen this issue before posting the pull request, thanks.
Yes i think it is related.
As you said in your last comment create_resized doesn't produce the right datatype layout.
The solution in this patch is to use create_hindexed with only one element and the same offset than previous create_resized.
I've tested darray-alltoallw-test and darray-alltoallw-test2 and both of them works now.
The fix may apply on release branch.

[ggouaillardet]

Thanks Guillaume,

bottom line, is MPI_Type_create_resized broken or not ?

in my test case, both mpich and ompi have the same behavior (well, ompi also issues a warning), and this is not the behavior I expected (could be a bug, or my incorrect expectation)

[jsquyres]

Add @bosilca

[hjelmn]

We might want to make this issue a blocker on 2.0.0. I remember fixing an issue with subarray but didn't realize darray was also broken.

[ggouaillardet]

IMHO #965 is the blocker and the milestone should be v1.10.1

[bosilca]

Due to the emergency of the next release, after talking with Jeff we propose to get the patch proposed here in the 1.10, and create an issue for the trunk. This will give me a little bit more time to dig into the internals of the type_resize and understand what is not going as expected there.

[bosilca]

The corresponding issue for the devel version is #1063.

[jsquyres]

@gpapaure Is this PR obsolete due to #1063?

[jsquyres]

I assume this issue is now closed (due to #1063 and #1072). Closing the ticket.