On Oct 15, 2008, at 4:26 PM, Sylvestre Ledru wrote:
We just need a precision on a specific point.
When we use the option --enable-static, a dynamic library is no longer
available (libmca_common_sm.so.0.0.0). The Makefile.am says that:
# 2. libmca_common_sm.la is a static library. In this case, it will
# be rolled up into the top-level libmpi.la. It will also be rolled
# into each component, but then the component will also be rolled up
# into the upper-level libmpi.la.
but we don't really understand why this becomes mandatory when
compiling
in static.
Does it make sense to keep this dynamic library (despite the
enable-static option) and to produce the static too or will it cause
bugs ?
Sorry -- I forgot to address this point.
Short version: this is mainly a reminder/note to us developers, and
should not be a problem for you.
Longer version:
The way we have our Makefile.am's setup, this library is linked
against a few different OMPI plugins. That is, both libmpi and each
of the plugins have:
... -lmca_common_sm ...
in their link step.
When libmca_common_sm.so.0.0.0 is built dynamically, it's all good
because each of those plugins simply has a reference to the dynamic
libmca_common_sm.so.0.0.0 library, and the run-time linker figures it
all out nicely. Specifically: the run-time linker will only import
the libmca_common_sm.so.0.0.0 library *once*.
When building statically, then each plugin (and libmpi) will contain a
full copy of whatever symbols it pulls out of libmca_common_sm.a. So
there could be a possibility of duplicated symbols. E.g., say symbol
"foo" is in libmca_common_sm.a and is used by all the plugins that use
this library and libmpi. Then a full copy of the foo symbol
definition will be in each. Then when libmpi slurps in all the
plugins via a link step, will we get duplicate "foo" symbol errors?
Thankfully, no. We have not yet found a linker that doesn't just
"figure it out" and know that the duplicate foo's are actually the
*same* foo, and therefore only include *one* foo in the resulting
executable.
Make sense?
--
Jeff Squyres
Cisco Systems
--
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
