Absoft MTT 32 bit build failure on master

[jsquyres]

Absoft's MTT is showing a 32 bit build failure on master -- linking opal_wrapper seems to be looking for a non-existent opal_cmpset_64 symbol.

See https://mtt.open-mpi.org/index.php?do_redir=2478.

@cagoelz sent the attached full logs, but they apparently don't show much beyond what we see in MTT generated by:

$ ../openmpi-master-201708290239-a84af67/configure FC=af90 FCFLAGS=-Z790 "CFLAGS=-g -pipe" --enable-picky --enable-debug --enable-mpirun-prefix-by-default 2>&1 | tee config_output.txt
$ make 2>&1 | tee make_output.txt

make_output.txt
config_output.txt

[jsquyres]

From @cagoelz:

$ find . -name "*.o" -exec grep opal_atomic_cmpset_64 {} \;
Binary file ./class/.libs/opal_list.o matches
Binary file ./mca/mpool/hugepage/.libs/mpool_hugepage_module.o matches

$ nm ./class/.libs/opal_list.o | grep opal_atomic_cmpset_64
        U opal_atomic_cmpset_64

$ nm ./mca/mpool/hugepage/.libs/mpool_hugepage_module.o | grep opal_atomic_cmpset_64
        U opal_atomic_cmpset_64

[jsquyres]

@cagoelz Can we get the config.log file as well? @hjelmn would like to see it.

[cagoelz]

Here it is.
config.log.zip

[hjelmn]

Hmm, seems this is a side-effect of deleting the old hand-coded ASM. This commit is why it is failing: b5ea5e09 . @bwbarrett Do you remember any of the details? I wonder if gcc still has this limitation.

[bwbarrett]

Wow; talk about bad flashbacks. I love this part:

For the beta, just use the non-inline version

Apparently it took 12 years for the hack to catch up with us.

Shockingly, I do remember this one. cmpxchg8b requires ebx to be used, but some versions of gcc were throwing an error trying to protect ebx, because shared libraries totally break if you change ebx (which we handle by saving/restoring the register). It was a hard compile-time error. We should be able to try a couple modern versions of gcc and if they compile, we're good.

[hjelmn]

Should be fixed now. We will re-open if MTT continues to show failures.