configury: revamp ucx detection

[ggouaillardet]

• when --with-ucx=DIR is not set, try the default path and fallback to /opt/ucx
• when --with-ucx-libdir is not set, try lib64 and then lib directories
• do not handle --with-ucx-libdir (this is a user mistake, no need to over-complicate our logic)

Signed-off-by: Gilles Gouaillardet [email protected]

[ggouaillardet]

:bot:aws:retest

[ggouaillardet]

:bot:aws:retest

[bwbarrett]

bot:ompi:retest

@ggouaillardet, the OMPI Jenkins sitting in AWS didn't understand bot:aws:retest (only ompi: or lanl:). It seemed like a reasonable guess, so I added aws: to the list as well.

[ggouaillardet]

Thanks !
so that means there is a bigger mystery here. i issued the aws command on #4384 and it instantly triggered a rebuilt ! then i issued the same aws command twice on this PR, but nothing happened ... and still nothing happened with the ompi command you issued.

[bwbarrett]

there was a typo in my update to make bot:aws:retest work. It never would have worked before (but there are some other updates, like a new commit, that can cause a retest.

bot:aws:retest

[ggouaillardet]

thanks !

[jsquyres]

I'm confused by the commit message on this commit:

 - when --with-ucx-libdir is not set, try lib64 and then lib directories
 - do not handle --with-ucx-libdir (this is a user mistake, no need to over-complicate our logic)

The 1st line says it will handle --with-ucx-libdir. But the 2nd line says it will not handle --with-ucx-libdir. Which is it?

[ggouaillardet]

configure --with-ucx=/foo will look for lib in /foo/lib64 and /foo/lib
configure --with-ucx=/foo --with-ucx-libdir=/foo/bar will look for lib in /foo/bar
configure --with-ucx=/foo --with-ucx-libdir will look for lib in yes (this is a dumb thing to do, and we do not protect end users against themselves any more in this case)

[jsquyres]

Ooohhhh... I see. I'm sorry to be type-A, but can you add something to that 3rd line in the commit to explain the yes value?

...and/or, what do we do elsewhere if the user specifies --with-FOO-libdir (with no value) -- do we ignore, or do we abort? We should be consistent.

[ggouaillardet]

ok, i ll reword this.
I checked here and there, and in other places, if --with-FOO-libdir is used, then the libdir is yes (e.g. we do nothing special), i will check again to be sure.

[jsquyres]

Is it worth making a support m4 macro that we can use in multiple places (for --with-FOO and --with-FOO-libdir handling) to guarantee consistent handling of these CLI args across the code base?

[ggouaillardet]

it turns we treat --with-FOO-libdir=yes as a no-op almost everywhere. (and a notable exception is opal_check_pmi.m4 and this is what i used as a reference ...)
fwiw, if we gently ignore --with-FOO-libdir, should we also ignore --without-FOO-libdir
and yes, a macro could be helpful to achieve consistency