GitLab

This was a long-standing TODO.

No reviewer.

was broken in r6273, prevents CPPFLAGS from being honored
(but MPICH2LIB_CPPFLAGS is a workaround).

No reviewer.

is needed for suncc.

Reviewed by gropp.

appropriate variables are declared for it to work in emulation mode as
well. This showed up as an error with the suncc compiler on linux
which is not natively supported by OPA.

error out at configure time, instead of just throwing a warning and
failing at make time.

These channels are old and have been deprecated for a while.  Their
presence is hampering PMI API development, so I have removed them.
These were the only channels seriously using the "process locks" code,
so that mess has also been deleted.  The only remaining useful
functionality (MPIDU_Yield) has been moved to the OS wrappers.

or automatically picked by the compiler before looking for it in
librt.

Reported as a failure on FreeBSD by Aleksej Saushev. This commit also
cleans up some of the checks and adds everything to LIBS, instead of
LDFLAGS. Fixes ticket #1013.

Reviewed by goodell.

It's still possible for some platforms to not like this format (perhaps
they don't have symlinks, etc).  But this is at least an improvement.

No reviewer.

Reviewed by goodell.

Previously flags to AR were fixed by simplemake to be "cr".  With this
change this can be controlled by configure/make via the new AR_FLAGS
precious variable.  On darwin using "Scr" shaves ~10% off the normal
build time.

This is a minor, nibbling around the edges fix.  The real fix is to
rework all of the linking in MPICH2 so that parallel builds are possible
and only a single invocation of AR/RANLIB occurs for each library that
we produce.

Reviewed by balaji@.

[svn-r6268] Added checks for multiple aliases support from C compiler.  setbot.c.in is a real item in svn instead of generated from buildiface.  Modified mpif.h.in and setbotf.f.in to allow (better) multiple aliases support.

In the compiler wrappers we force the linker to use the "flat namespace"
model instead of the default "two-level namespace" model.  If we add
dynamic library support for MPL and OPA then we can probably drop this
option again.

Reviewed by balaji@.

No reviewer.

broken, requiring lots of hacks to get it working correctly). We now
allow each channel to pick what all common directories it wants
configured instead of blindly configuring everything. But we still
retain the hack where the setup of the common directories has to run
before the device configuration while the actual configuration needs
to run after, because the build seems to fail without this.

configured. The problem is that users who are used to the mpich2-1.2.1
behavior continue to use:

% mpdboot ...

% mpiexec -n 4 ./foo

% mpdallexit

This is buggy since the mpiexec now belongs to Hydra which completely
ignores where the mpd daemons are started and launches everything
locally. Disabling mpd at least gets users doing this to notice that
something has changed.

overwritten. Thanks to Seung for reporting the bug.

duplicated.

are eventually dumped into CPPFLAGS.

reimplementing the same code.

autoconf variables to be passed in as well.

environment variable propagation cleaner.

If the <complex> header was unavailable it was still possible to end up
with HAVE_CXX_COMPLEX defined.  This causes problems for the various
predefined MPI reduction operators, which the presence/absence of
HAVE_CXX_COMPLEX to conditionally compile case statements.  Incorrect
values can result in "duplicate case value" errors from the compiler.

No reviewer.

src/openpa location, but rather from the standard lib directory.

Reviewed by goodell.

of the example programs is successful. Fixes ticket #993.

This commit moves the existing tracing memory allocation and valgrind
code utilities to MPL.  This permits other code like gforker to use the
trmem code while allowing MPICH2 to remain thread-safe when using
tracing.  gforker is not updated to the MPL routines by this commit.

Reviewed by buntinas@.

Reviewed by buntinas@.

Necessary for moving other code to MPL.

Reviewed by buntinas@.

WRAPPER_LIBS.

LIBS are all libraries detected and added by MPICH2. These are
maintained internally, so appropriate autoconf checks can use them
correctly. WRAPPER_LIBS are only used within mpicc and friends; they
are a combination of LIBS and other libraries such as opa and mpl
which are only generated at make time. These are maintained separately
so autoconf doesn't try to use them for link or run tests.

else. Otherwise, the detection for other functionality might be
incorrect.

MPICH2 now.

macro. It seems to use device-specific code, so the device needs to be
built before the locks directory can be built.

of CFLAGS. make dependencies fails without this.