Hey, all.
I'm writing some code that performs commits via the Subversion Python
bindings, and I'm struggling to understand some things I see there.
In the svn_fs.i interface file, there's this block of code:
/* -----------------------------------------------------------------------
Fix the return value for svn_fs_commit_txn(). If the conflict result is
NULL, then %append_output() is passed Py_None, but that goofs up
because that is *also* the marker for "I haven't started assembling a
multi-valued return yet" which means the second return value (new_rev)
will not cause a 2-tuple to be manufactured.
The answer is to explicitly create a 2-tuple return value.
FIXME: Do the Perl and Ruby bindings need to do something similar?
*/
#ifdef SWIGPYTHON
%typemap(argout) (const char **conflict_p, svn_revnum_t *new_rev) {
/* this is always Py_None */
Py_DECREF($result);
/* build the result tuple */
$result = Py_BuildValue("zi", *$1, (long)*$2);
}
#endif
This reads and claims to behave exactly as I'd expect, given the dual
return values of svn_fs_commit_txn (and svn_repos_fs_commit_txn which wraps
it). And since this interface file is included from svn_repos.i, I would
expect those typemaps to apply also to svn_repos_fs_commit_txn, which has
matching parameter types and names.
But this isn't how the code appears to work in practice. A successful
commit gets back from svn.repos.fs_commit_txn not a 2-tuple, but just the
newly created revision number. Moreover, if my commit succeeds but the
post-commit hook fails, svn.repos.fs_commit_txn() raises an Exception,
masking the return of the newly created revision number altogether. Now,
the masked revision number thing I can understand -- it's hard to do in
Python what we promise to do in C (which is to return an svn_error_t but
still set the new_rev return value). But the 2-tuple thing I have no
explanation for. Any ideas?
-- Mike
