mpi/cxx: isolate internal headers from C++ bindings

[hjelmn]

This commit adds some glue code to support the C++ bindings and
updates the bindings to use the new glue code. This protects our
internal headers (which are C99) from C++. This is done as a quick
workaround to compilation errors when the legacy C++ bindings are
requested.

Fixes #2055

Signed-off-by: Nathan Hjelm [email protected]

[hjelmn]

@jsquyres This should be adequate to allow the C++ bindings to be compiled without having to mess with internal headers.

[hjelmn]

@jsquyres Any comments on this? It needs to go into 2.0.2 and 2.1.0 so I plan to merge it later today if there are no objections.

[jsquyres]

I get compile fails:

  CXX      intercepts.lo
intercepts.cc:29:21: fatal error: cc_glue.h: No such file or directory

[jsquyres]

After fixing intercepts.cc to include cxx_glue.h, I still get moar compile failz:

  CXX      comm.lo
comm.cc: In static member function ‘static int MPI::Comm::do_create_keyval(int (*)(MPI_Comm, int, void*, void*, void*, int*), int (*)(MPI_Comm, int, void*, void*), int (*)(const MPI::Comm&, int, void*, void*, void*, bool&), int (*)(MPI::Comm&, int, void*, void*), void*, int&)’:
comm.cc:91:47: error: ‘malloc’ was not declared in this scope
         malloc(sizeof(keyval_intercept_data_t));
                                               ^
comm.cc:116:29: error: ‘free’ was not declared in this scope
         free(cxx_extra_state);
                             ^
make: *** [comm.lo] Error 1
  CXX      datatype.lo
datatype.cc: In static member function ‘static int MPI::Datatype::do_create_keyval(int (*)(MPI_Datatype, int, void*, void*, void*, int*), int (*)(MPI_Datatype, int, void*, void*), int (*)(const MPI::Datatype&, int, void*, const void*, void*, bool&), int (*)(MPI::Datatype&, int, void*, void*), void*, int&)’:
datatype.cc:55:47: error: ‘malloc’ was not declared in this scope
         malloc(sizeof(keyval_intercept_data_t));
                                               ^
datatype.cc:80:29: error: ‘free’ was not declared in this scope
         free(cxx_extra_state);
                             ^
make: *** [datatype.lo] Error 1
  CXX      win.lo
win.cc: In static member function ‘static int MPI::Win::do_create_keyval(int (*)(MPI_Win, int, void*, void*, void*, int*), int (*)(MPI_Win, int, void*, void*), int (*)(const MPI::Win&, int, void*, void*, void*, bool&), int (*)(MPI::Win&, int, void*, void*), void*, int&)’:
win.cc:64:47: error: ‘malloc’ was not declared in this scope
         malloc(sizeof(keyval_intercept_data_t));
                                               ^
win.cc:89:29: error: ‘free’ was not declared in this scope
         free(cxx_extra_state);
                             ^
make: *** [win.lo] Error 1
make: Target `all' not remade because of errors.

[jsquyres]

Compile errors now fixed, but when I run the cxx test suite in ompi-tests, it segvs in a type test. Here's a bt from the corefile:

(gdb) bt
#0  0x00007fffffffc318 in ?? ()
#1  0x00002aaaaaac4c8c in ompi_mpi_cxx_type_copy_attr_intercept (
    oldtype=0x64f4a0 <ompi_mpi_byte>, keyval=13, extra_state=0x7fffffffc2f0, 
    attribute_val_in=0x7fffffffc318, attribute_val_out=0x7fffffffbf90, 
    flag=0x7fffffffc00c) at intercepts.cc:404
#2  0x00002aaaaacf64a9 in ompi_attr_copy_all (type=TYPE_ATTR, 
    old_object=0x64f4a0 <ompi_mpi_byte>, new_object=0x828d90, oldattr_hash=0x829410, 
    newattr_hash=0x8292f0) at attribute/attribute.c:877
#3  0x00002aaaaad8b94e in PMPI_Type_dup (type=0x64f4a0 <ompi_mpi_byte>, 
    newtype=0x7fffffffc0a8) at ptype_dup.c:74
#4  0x000000000040db2f in MPI::Datatype::Dup (this=0x7fffffffc280)
    at /home/jsquyres/bogus/include/openmpi/ompi/mpi/cxx/datatype_inln.h:266
#5  0x000000000042bcc0 in typekeyval () at typekeyval.cc:154
#6  0x000000000040c2fe in main (argc=1, argv=0x7fffffffc548) at mpi2c++_test.cc:221

It failed on this line:

403       ret = kid->cxx_copy_fn(cxx_datatype, keyval, kid->extra_state,
404 =>                           attribute_val_in, attribute_val_out, bflag);

It looks like the kid extra state has a bogus value for several of the callbacks:

(gdb) p kid
$1 = (MPI::Datatype::keyval_intercept_data_t *) 0x7fffffffc2f0
(gdb) p *kid
$2 = {c_copy_fn = 0x0, c_delete_fn = 0x1600000016, cxx_copy_fn = 0x7fffffffc318, cxx_delete_fn = 0x200000001, extra_state = 0xfffffff000000002}

This problem does not occur on master HEAD; only on this PR. So something must have changed w.r.t. the C++ bindings besides just isolation...?

[jsquyres]

@hppritcha Good to go when CI finishes.