If a client uses AMI to make an asynchronous call to a collocated server, with
collocation optimisations enabled, the call is in fact made synchronously (ie
within the call to sendc_foo, the servant's foo method is called).

There are a number of issues with this:

1. It's Just Wrong. The semantics of AMI require asynchronous invocation, and
any optimisation for collocated objects isn't allowed to change that. The
handler function is not called as required.

2. It can't work, because the signatures of foo and sendc_foo are usually
different (sendc_foo always returns void and omits out parameters). In my
testing with a function returning long I saw stack corruption, because the long
result was written to the void retval of the sendc function.

3. Exceptions don't work either - if the server throws a user exception then it
is thrown directly from sendc_foo (which is forbidden by the spec) and the
_excep function is not invoked.

A test program to demonstrate this is fairly easy, but I don't currently have a
standalone one. I'll see if I can produce a stripped down test harness.

Note that disabling collocation (-ORBcollocation no) is not a full workaround
for this, since it breaks AMI exception handling - see bug 2350.