feat(build): disable semantic interposition

[aws-nslick]

Description of changes:

feat(build): disable semantic interposition

Tell the compiler to treat symbols marked explicitly as
NCCL_OFI_EXPORT_SYMBOL as preemptible, but not others. This gives the
compiler the freedom to assume that no external LD_PRELOAD'd library can
inject itself as an implementation of any given function in the library,
which means the compiler has the freedom to optimize across function.

Signed-off-by: Nicholas Sielicki <[email protected]>

1 file changed, 10 insertions(+), 1 deletion(-)
configure.ac | 11 ++++++++++-

fix(build): add missing AC_PROG_RANLIB

Automake manual:

> This is required if any libraries are built in the package. See
> Particular Program Checks in The Autoconf Manual.

add this for correctness.

Signed-off-by: Nicholas Sielicki <[email protected]>

1 file changed, 1 insertion(+)
configure.ac | 1 +

fix(build): ensure -pthread is passed

Linking against pthread and adding the compiler option for pthread are
two different things. Without the compiler option, various builtin
'#define's are not set. Notably, this should change the behavior around
errno in a multithreaded context within glibc; _REENTRANT will now be
defined and errno is now thread-local.

Signed-off-by: Nicholas Sielicki <[email protected]>

1 file changed, 3 insertions(+)
configure.ac | 3 +++

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[aws-nslick]

I was hoping to avoid putting this into the stack, however, it's also hitting the issue fixed in #589 so no such luck.

edit: actually, this reintroduces the bug that #589 fixed -- unclear to me whether or not we can just move this line below the dependency resolution or not.

[bwbarrett]

I think all the flags are safe (some useless, but safe) for executable, but do be careful because these are all going to end up in the compile and link commands for the tests.

[aws-nslick]

@bwbarrett I know you hate this and I am sorry to split a review, again, but once we land this change, it gives the compiler visibility into the fact that there are uninitialized stack variables that are not guarded with NULL checks, and so the build fails due to new compiler warnings.

I didn't notice it previously, because it was fixed with #594

So I've rebased this commit on that PR, which means that means that the source branch for this PR needs to move from my fork into the github stack suffix on the main repo. I'll repost your comments on the previous PR, but they've been addressed.

Closing this PR and moving over to #624