Mysterious 32 bit MTT build failures on Absort

[jsquyres]

We're seeing odd MPI Install failures on 32 bit Absoft failures, but the log messages are not showing what the actual error is.

@cagoelz Can you shed any light, perchance, on what is failing in these builds?

https://mtt.open-mpi.org/index.php?do_redir=2446

[cagoelz]

The failures occur building ompi_datatype_external.lo:

  CC       ompi_datatype_external.lo
In file included from ompi_datatype_external.c:29:
../../ompi/communicator/communicator.h:265: error: size of array ‘padding’ is too large
make: *** [ompi_datatype_external.lo] Error 1

The code around ../../ompi/communicator/communicator.h:265 is

#define PREDEFINED_COMMUNICATOR_PAD (sizeof(void*) * 64)

struct ompi_predefined_communicator_t {
    struct ompi_communicator_t comm;
    char padding[PREDEFINED_COMMUNICATOR_PAD - sizeof(ompi_communicator_t)];
};

The size of ompi_communicator_t is larger than 256 bytes?

[bosilca]

git blame 50aa143. On 32 bits architectures, as the void* size is 32 bits, the opal_hash_table_t becomes too large for the PREDEFINED_COMMUNICATOR_PAD. We might have an issue with our ABI ...

[jsquyres]

Ah! Good point.

I guess we intended to allow 64 pointers worth of padding. But in hindsight, that's a poor measure. Should we just listed a number of bytes, instead? (e.g., 64*8=512)

Sidenote: I'm not worried about our ABI between 32 and 64 bit builds.

[bosilca]

I was not worried about ABI 32 vs. 64 bit builds (it has no sense), but we will now have to increase the size of the PREDEFINED_COMMUNICATOR_PAD, which will translate into a ABI breakage.

[jsquyres]

Yes, we will break ABI for 32 bit builds. But if we get this into 3.0 (and beyond), I'm not worried about it.

Do you agree?

[bosilca]

I see what you're saying: if we change to a fixed length, equal to the size of 64*8=512, we will not change the size on 64 bits builds, so we will only break the AI on 32 bits builds. Sounds good. 👍

[jsquyres]

@cagoelz I didn't say it earlier: thanks for the detailed compilation output! 😄