The spec has the following sentence:

The _d discriminator modifier can only be used to set the discriminant to a value within the same union member.

This is not really clear and should be improved, maybe

The _d discriminator modifier can only be used to set the discriminant to a value that has the same union member as currently selected.

The question is about 99-07-41.pdf, as far as I know the latest C++ mapping specification. "1.41.5 Any Class" prescribes "Boolean operator>>=(const Any&, const char*&);" for unbounded strings, and "1.16.3 Extraction from any" says that unbounded strings are extracted by value. This seems to be contradictory: I seem to remember that const char* cannot be delete'd, though I don't find it in ISO/IEC 14882:1998(E). But, regardless of anything imposed by C++, the current mapping precludes a safe technique like (safe as in exception-proof): >> String_var theString; if(!(theAny>>=theString.out()))

<< I therefore propose that the mapping be changed to "Boolean operator>>=(const Any&, char*&);". This seems to work all right on my implementation (RayORB), for gcc 3.2 (MS VC++ 6.0 doesn't seem to care one way or the other, but I suppose that is wrong, although I'm not entirely sure).

Summary: If I want to use the DII to send out multiple simultaneous requests, I don"t see a practical way to associate any client specific context that is C++ compliant to those requests.

Althrough I report my issue as a C++ issue, actually it is a common issue for other languages also. In current CORBA standard, a activated CORBA Servant object in the POA is shared by all the clients. But in many applications, a client connects to the server and calls some CORBA object methods to create a lot of connection related CORBA objects. Normally, the client will call "destroy" method of these CORBA object to destroy them before existing. But if the client exists abnormally, these created CORBA object will never be destroyed and never be used by any client. We can call this kind of CORBA object as "floating" CORBA object. The CORBA server application runs longer, the "floating" CORBA objects becomes more. And eventually, these "floating" CORBA objects will consume all the memory resources/CPU resources, then the CORBA server crashes and has to be restarted. The "floating" CORBA objects problem at the server side is not easy to be solved. For this problem, if the CORBA standard provides a method to solve it, the CORBA server will run more robust. Actually, the underlying network connection broken can be detected very effectively and rapidly. The CORBA server just simply ignores the network connection lost event and does not pass it to activated CORBA object even if the CORBA object is connection related object because current CORBA standard thinks that all the CORBA objects in the server are shared to all clients. If the server can notify the network connection lost event to the connection related CORBA objects, the developper can do the clean job on it. A simple method can solve the "floating" CORBA object problem. It is not required to add any new interface or data structure to current orb.idl. It is just required to add some new language mapping methods at server side only. I will give an example in C++ to show this simple method. A IShared,IConnSpec and IOtherConnectionSpec interfaces are defined below: interface IOtherConnectionSpec

interface IConnectionSpec

interface IShared

; These interfaces will be used in this example. 1) the IShared will be shared by all the clients, it is activated like this: SharedImpl sharedImpl( ... ); ... sharedImpl._this(); 2) add new mapping method for the "POA_IShared::createConnSpec()" like this: IConnectionSpec_ptr POA_IShared::createConnSpec( char* param, const ConnectionContext& connCtx ); In this method, a parameter named ConnectionContext that represents the connection between the client and server is added. This method has a default implementation like this: IConnectionSpec_ptr POA_IShared::createConnSpec( const char* param, const ConnectionContext& connCtx )

3) If the user wants to create a connection related object, he/she must overload the "IConnectionSpec_ptr createConnSpec( char* param, const ConnectionContext& connCtx )" method and active the object with "this_( const ConnectionContext& connCtx )" method: IConnectionSpec_ptr POA_IShared::createConnSpec( const char* param, const ConnectionContext& connCtx )

in this method, the user uses the "this_( const ConnectionContext& connCtx)" method instead of the old "this()"" method to active the CORBA object. This activated CORBA object becomes a connection related object instead of shared object. Note: User can use the "this_( const ConnectionContext& connCtx)" method to active a shared object also. At this time, the ConnectionContext is a global variable "NO_CONNECTION_CONTEXT". 4) add a new mapping method named "getConnectionContext()" for every CORBA object. const ConnectionContext& IShared::getConnectionContext(); const ConnectionContext& POA_IConnectionSpec::getConnectionContext(); In this way, the IConnectionSpec can create and activate any other connection related CORBA object like this: IOtherConnectionSpec_ptr POA_IConnectionSpec::someMethod()

5) add a new mapping method named "connectionLost()" for every CORBA object. When the network connection is lost, the CORBA server should find all the CORBA object associated with this network connection and call their "connectionLost()" method. The "connectionLost()" and its default implementation is listed below: void POA_IConnectionSpec::connectionLost()

Because the SharedImpl object that created in the 1) step does not assicate with any connection, its "connectionLost()" method will not be called for ever.

The second-to-last paragraph of section 1.17.10.3 says

"For valuetypes that have no operations or initializers, a concrete type-specific factory class is generated whose implementation of the create_for_unmarshal function simply constructs an instance of the OBV_ class for the valuetype using new and the default constructor."

As specified, that requires the generation of invalid C++. The OBV_ class is abstract since it does not have implementations of the ValueBase reference counting functions.

Perhaps the intention is that the OBV_ classes in such cases should derive from DefaultValueRefCountBase. However, the wording and explanation in section 1.17.6 explicitly forbids this:

"Note that it is the application-supplied concrete valuetype classes that must derive from these mix-in classes, not the valuetype classes generated by the IDL compiler."

One solution that avoids the problem, and avoids restricting the application's use of the OBV_ classes is to generate yet another class that derives from both the OBV_ class and DefaultValueRefCountBase, for instantiation by the _init class's create_for_unmarshal function.

1.9.1 says:

"The copy constructor deep-copies any data pointed to by the T_var constructor
parameter. This copy will be destroyed when the T_var is destroyed or when a new
value is assigned to it. Compliant implementations may, but are not required to, utilize
some form of reference counting to avoid such copies."

and

"The normal assignment operator deep-copies any data pointed to by the T_var
assignment parameter. This copy will be destroyed when the T_var is destroyed or
when a new value is assigned to it. Assigning a null pointer to a T_var is legal and
results in deallocation of the data pointed to by the T_var."

So my question is, is it legal to use ValueBase::_add_ref() and _remove_ref() instead of _copy_value() in this instance when T_var is representing a valuetype?

C++ programmers will often want to use strings as
object identifiers, the C++ mapping provides several conversion
functions that convert strings to ObjectId and vice-versa:"

The purpose is so the programmer can pick an arbitrary natural language
string and use it as an ObjectId, not so that the programmer can
generate a randomly unreadable string out of a binary ObjectId.
[...] the C++ mapping provides several conversion functions
that convert strings to ObjectId and viceversa:
[...]
If conversion of an ObjectId to a string would result in
illegal characters in the string (such as a NUL), the first two
functions throw the CORBA::BAD_PARAM exception.

The conversion algorithm is not specified, and the ORB is free to
choose whatever encoding it likes for its Object IDs. (Object IDs
in stringified form need not be moved across address spaces (or,
at least, not across ORB boundaries), so having a proprietary
encoding is perfectly OK.)

The ORB interface destroy() operation defined in the core CORBA 2.3 spec seems to be missing from the associated C++ language mapping.

Add the destroy operation to the ORB definition in 4.35.1 and 4.45.21

Summary: We recently added _ptr_type and _var_type typedefs into interface classes
to make writing templates for interface types easier. We should add the
same typedefs for valuetype classes.

Add the missing typedefs

Suppose we have:

typedef fixed<4,3> FT;

interface I

;

Suppose the server does:

FT
I_impl::
op() throw(CORBA::SystemException)

There are lots more digits in the return value than what is expected by the
client. What should be returned to the client. The rounded value? The
truncated value?

Similarly, what if we have:

double d = 10000000;
return CORBA::Fixed(d);

Do we return 9.999 to the client (which is the best we can do in this case)?

Of course, it is the responsibility of the programmer to make sure that
nonsense such as the second case doesn't happen. But the spec has to say
what happens if it does happen

Also, the first case will be very common – what should happen in this case?

The Object::get_domain_managers() operation is a transmissible operation
with CORBA 2.3. Does ServantBase need anything added to handle this or
is this managed internally by the ORB/POA?

Footnote 11 on page 1-48 of the 1.1 version of the C++ language mapping states that sequence _var classes don't have a const version of "operator []". The justification in the footnote is incorrect.

This footnote should be removed, and the operator provided, since it otherwise prevents accessing sequence members through a const reference to a sequence _var.

Since the Any constructor from type code, void* value and boolean release flag has been eliminated, there is no longer any portable way to create an OUT argument for a DII request, without also assigning a value to it.

I propose either a constructor from type code or changing the behavior of the type code set method to always succeed if the Any's TCKind is tk_null.

Just as we prohibited direct extraction from any any into a _var type
due to the perils of memory management problems, we ought to prohibit
extracting into _out types of variable sized structured types for the
same reason

The addition amounts to a set of typedefs that would facilitate template
programming, added to each C++ skeleton class created for an IDL interface,
and analogous to the typedefs that already exist in the mapping for the stub
side. Say we have an IDL file

module foo
{
interface bar {};
};

Then in generated code we're talking about something like:

namespace POA_foo
{
class bar

;
};

need an _unchecked_narrow static operation that is similar to the _narrow. It will take in an object pointer but unlike _narrow will not check for compatibility before creating a new object reference.

In IDL it gives:

// IDL
valuetype V

;

But it should be

// IDL
valuetype V

;

implementing support for wchar/wstring I ran into some potential
problems with the UTF-8 encoding for the IDL 'char' type.

Lets suppose we have a C++ server with a native single-byte code set
like ISO 8859-1.
The Code Set Conversion specification states that UTF-8 is the fallback
code set for 'char'.
-> a client could decide to send characters in UTF-8 encoding.

What happens on the server side with UTF-8 encoded characters that use
more than 1 byte and thus don't fit into the single byte character as
specified by the C++ mapping for IDL type 'char'?

the C++ mapping should explicitly list operator == and operator Unable to render embedded object: File (= in the examples and text. If it is not allowed to use ==/) not found.= the methods should be declared private

The spec defines: // C++ class Seq_var; class Seq

; We propose to extend this to the following to make it possible to write more generic C++ code. // C++ class Seq_var; class Seq_out; class Seq

;

In section 4.11 add to the example at the end Seq_out and a _out_type typedef
as in the revised text below.
The _size_type typedef will make it possible in the user code to differentiate
between the index of a sequence type and a user defined ULong value. To
section 4.15 add the sentence
To facilitate template-based programming, a nested public typedef _size_type is delivered as the type
representing the length of the sequence.
And to the sequence class examples in section 4.15 the following
typedef ULong _size_type;

I was looking at the OMG IDL to C++ mapping specification, specifically
page 123 and came across the below. It looks to me that their
implementation in the spec is NOT correct. Shouldn't the swap function
take the "tmp" Servant_var?

Servant_var& operator=(Servant* p)
{
if (_ptr != p)

return *this;
}

Servant_var&
operator=(const Servant_var& b)
{
if (_ptr != b._ptr)

return *this;
}

Update assignment operator as proposed

To really support template meta programming the valuetype classes should deliver a typedef _out_type besides the _ptr_type and _var_type which are proposed to be added as part of issue 1794

Add a _out_type typedef to the valuetype class

To really support template meta programming the inteface should deliver a typedef _out_type besides the already defined _ptr_type and _var_type

Add a _out_type typedef to the interface class

The C++ keywords in section 4.47 should be updated and extended with the addition of C++0x

Add the following keywords to the table on page 154 as part of section 4.47

In section 4.5.4 it says: Unlike _duplicate, the parameter to _narrow is a reference of an object of any interface type (Object_ptr). If the actual (runtime) type of the parameter object can be widened to the requested interface’s type, then _narrow will return a valid object reference; it should say (widened to narrowed) Unlike _duplicate, the parameter to _narrow is a reference of an object of any interface type (Object_ptr). If the actual (runtime) type of the parameter object can be narrowed to the requested interface’s type, then _narrow will return a valid object reference;

In section 4.5.4 change widened to narrowed

Section 1.31.3 says:

• In the C mapping, set_one_value used strings, while
  others used NamedValues containg any. Even though implementations
  need only support strings as values, the signatures now uniformly
  allow alternatives.

I would suggest to delete the entire bullet point. In particular, the notion
of allowing alternative data types as propery values for contexts doesn't
work because the receiver expects a sequence of strings with an even number
of elements; if anything but a string sequence is sent, the receiver
has no chance of unmarshaling it.

Remove bullet as suggested

The Object::get_policy returns a Policy in IDL, not a Policy_ptr, that is the C++ return value

Replace in 4.36.1 the text Policy_ptr to Policy

On page 1-119, the spec says the following about _create_request() and
_create_request2():

[...] the ORB requires:

• a target object reference
• an operation name
• a list of arguments (optional)
  ^^^^^^^^^^
• a place to put the result (optional)
  ^^^^^^^^^^
• a place to put any returned exceptions
• a Context (optional)
  ^^^^^^^^^^
• a list of the user-defined exceptions that can be
  thrown (optional)
  ^^^^^^^^^^
• a list of Context strings that must be sent with the
  operation (optional)
  ^^^^^^^^^^

Note all the "optional" remarks.

It's not clear what "optional" actually means. We have two cases for
these parameters:

• Arguments, user exceptions, and IDL contexts are sequences.

• Result and context are object references.

Two questions:

• What does it mean for a sequence parameter to be "optional"?
  That I can pass a null pointer or that I can pass an empty
  sequence? I assume that an empty sequence is meant, but the spec
  doesn't say that.

• What does it mean for a reference parameter to be "optional"?
  That I can pass a nil reference or that I must pass a reference
  to some dummy object that indicates that the value really isn't
  there?

Where this is particularly annoying is for the return value. (The
"result" parameter to _create_request()):

• If I can pass a nil reference, no problem. This could
  be interpreted to mean the same thing as a void
  return type.

• If I can't pass a nil reference, what should I pass?
  Obviously, it would have to be a reference to a valid
  NamedValue object. But how do I create that NamedValue
  object?

I can call ORB::create_named_value(), but what should
I do then?

CORBA::NamedValue_var result;
orb->create_named_value(result);

At this point, I have a NamedValue containing an Any with
tk_null (because that's what the Any default constructor
creates). However, to correctly indicate that the operation
I want to call has a void return type, I have to make a
NamedValue that contains an Any with tk_void. But, how do
I achieve that? I can't use one of the Any inserters to
turn the TypeCode in the Any into tk_void, and I can't
use the type() member of the Any because I can't change the
type of an Any if it isn't consistent with the TypeCode that's
already there...

Even worse, the NamedValue doesn't seem to make sense as
the return value at all. For one, it has a name attribute.

• What is the value of that name for a return value?

• How would I set that string after having create the NamedValue
  by calling create_named_value()?

• What is the value of that string once I have called
  create_named_value()?

Second, the NamedValue contains a flags attribute.

• What is the value of that flags attribute for a return value?
  None of ARG_IN, ARG_INOUT, or ARG_OUT make sense. (One could
  argue that ARG_OUT could be used, but I think that sucks...)

• How would I set that flag on the NamedValue I have just created?
  The mapping for NamedValue only offers accessor but no
  modifiers, so I can't set the value of the flag.

• What is the value of the flag once I have called
  create_named_value()?

It seems that the easiest way to fix the problem is to state that, if a
parameter isn't needed, this is indicated by an empty sequence for lists,
and by a nil reference for objects.

However, the problems around create_named_value() and appear to be more
serious. How should we fix those?

The June 99 version of the C++ mapping doesn't mention ORB::destroy().
It should.

The second version of _create_request() accepts two context lists:

void Object::_create_request(
Context_ptr ctx, // <===
const char * operation,
NVList_ptr arg_list,
NamedValue_ptr result,
ExceptionList_ptr,
ContextList_ptr, // <===
Request_out request,
Flags req_flags
);

The spec then says:

The ContextList differs from the Context in that the former supplies
only the context strings whose values are to be looked up and sent
with the request invocation (if applicable), while the latter
is where those values are obtained.

So, I think (but I'm not sure), this means to say that:

• The Context parameter points at a tree of context objects.

• The ContextList pointer points at an object that contains
  a sequence of strings.

• The ContextList determines which context values are to be
  picked out of the tree pointed at by the Context parameter and
  to be sent with the invocation.

If that is the intended interpretation, it's very difficult to discern
from the words in the spec now. I think this needs clarification.

Questions:

• What happens if the ContextList contains the name of a context
  variable that isn't available in the context tree?

• What happens if I have a non-empty context list but a nil
  context tree?

Also, looking at the ContextList interface:

pseudo interface ContextList

;

There is no further documentation on this. Some questions:

• As far as I can see, this interface is meant to maintain a
  simple sequence of strings. So, why not simply use a string
  sequence?

• At what position does add() add the item? Presumably at the tail,
  but that isn't stated.

• How can I replace the value of string at a certain position?
  It looks like I can't do that at all because there is no
  random-access modifier.

This seems insanely complex.

Suggestion: We should

• add words to make it clear how the context parameters work

• consider overloading _create_request() yet again to use
  an ordinary string sequence

• deprecate the ContextList pseudo interface

Or, we could drop support for IDL context altogether (but I suspect we
can't get away with that

with 2.3, we got rid of all the C/C++ pseudo code for the DII and replaced
it with IDL. Unfortunately, this has broken something in the C++ mapping...

In the 2.0 spec, we had:

pseudo interface ORB

;

With 2.3, we changed this to:

module CORBA

;

Unfortunately, the C++ mapping still shows the (now incorrect) definition
from CORBA 2.0 in Section 1.33.1.

In addition, the C++ mapping shows in Section 1.33.2:

class ORB {
public:
class RequestSeq

;
// ...
};

But then, in Section 1.41.21:

class ORB

;

The latter definition isn't C++...

So, we have several issues here:

1) How can we fix the C++ mapping to be in line with the core?

I'm toying with the idea of saying that RequestSeq is defined
in the CORBA namespace, with a typedef for backward compatibility
in the ORB interface. But I'm not sure what will break with
this kind of aliasing (repository IDs might change unexpectedly?)

2) Section 1.41.21 should be changed to show legal C++.

3) Depending on the resolution to this issue, both 1.33.2 and
1.41.21 will probably need updating to reflect the resolution.

what should _default_POA() return if the default ORB was never initialized?
The spec doesn't say...

I have the following questions to IDL - C++ mapping CORBA 2.3 (concerning valueboxes).
Can somebody give me an answer?

1. I assume that valueboxes T as a special kind of valuetypes also
need in C++ the T_var and the T_out types and the T_out types will be
used in function signatures for IDL out-parameters. Is it true?

2. The mapping for strings and wstrings requires a definition of C++
operators << and >> to allow a use of strings (wstrings) with the c++ iostreams.
What about valueboxes of strings (wstrings) - chapter 1.17.7.4?
Is it required to define (in C++) the operators <<, >> to use T_var and T_out
with C++ iostreams, where T is a type for a valuebox T of string (wstring)?

Given

valuetype WStringValue wstring;

is there any requirement to have stream inserters and extractors for the
boxed value type itself? The spec is currently silent on this issue.

Should the following work?

WStringValue ws;
cin >> ws;
cout << ws;

Incorporate Messaging changes relevant to the C++ mapping, as shown in
orbos/98-05-05 pages 115 and 116 together with any changes made to them
by the Messaging RTF, into the IDL-C++ mapping chapter.

quite some time ago, we added the _var_type and _ptr_type definitions
to proxies to make it easier to write templates. Similarly, we have
the _whatever_seq typedefs for recursive structs and unions, to avoid
problems with anonymous types.

What's missing at the moment is a similar feature for _var types.
When I'm writing a template that does it's job in terms of _var types,
I also quite often want to do something to the underlying target type
of the _var. However, I can't do that unless I pass in an extra template
parameter (which, in turn, doesn't always work if I also want to use
STL standard binders and such...)

So, why not add a typedef for the target type to every _var type?

the spec is currently not terribly clear about the server's responsibilities
when throwing an exception from an operation that has a variable-length
out param.

The intent of the spec is that the server is responsible for deleting
the memory it allocated to an out param before it throws an exception:

// Correct implementation
void
Foo_impl::
op(CORBA::String_out out_p) throw(CORBA::SystemException)

// Incorrect implementation
void
Foo_impl::
op(CORBA::String_out out_p) throw(CORBA::SystemException)

However, the spec never states this clearly. In fact, it sort of says
the opposite. On page 1-110, table item 3:

To maintain local/remote transparency, the caller must always
release the returned storage, regardless of whether the callee
is located in the same address space as the caller or is located
in a different address space.

There is no mention here of what should happen in the presence of exceptions.

I think it would be nice to clarify that the skeleton will never look
at an out param in the presence of exceptions and that the operation
implementation is responsible for deallocating memory in this case.

in the resolution to issue 863, we decided to get rid of the read-only
restriction for parameters. However, we forgot to remove a few snippets.

Page 1-110, table, case 3:

Remove the following text:

Following the completion of a request, the caller is not allowed
to modify any values in the returned storage--to do so, the
caller must first copy the returned instance into a new instance,
then modify the new instance.

Page 1-110, table, case 6:

Remove the following text:

Following completion of a request, the caller is not allowed
to modify any values in the returned storage--to do so, the
caller must first copy the returned array instance into a new
array instance, then modify the new instance.

Section 1.17.11 of ptc/2000-01-02 says that DataInputStream &
DataOutputStream are mapped according to the normal valuetype mappings.
This mapping results in operations that are at best cumbersome to use in
marshalling sequences of primitive types:

Operations read_xxx_array and write_xxx_array are defined, taking argments
of type CORBA::xxxSeq. These are obviously intended to be sufficient for
marshalling all defined sequences of the corresponding primitive type xxx.

However, the c++ mapping for sequences requires that each distinctly
declared sequence type be unique and separately overloadable. This
guarantees that a marshaller attempting to marshal an sequence of primitive
other than CORBA::xxxSeq will not be able to pass that sequence to the
corresponding read_xxx_array and write_xxx_array operations. Instead, code
must be written to bypass strict typing.

To fix this, either the mappings for DataInputStream and DataOutputStream
need to be non-standard, the mapping of sequences needs to change, or the
mapping of CORBA::xxxSeq needs to change to something special.

the text for set_servant doesn't say what should happen if I call
set_servant(0). I would suggest to throw BAD_PARAM in this case.

For servant reference counting, the words "at least once" appear
in a number of places when it comes to servant activation. set_servant(),
activate_object(), activate_object_with_id(), servant_to_id(),
servant_to_reference(), and _this() all use this language (pages 1-137 and
1-138):

... will invoke _add_ref at least once on the Servant argument...

Problem: suppose my ORB calls _add_ref() twice in each of these operations
for some reason. I now have a problem. That's because, for servant
activators, responsibility for calling _remove_ref() passes to
etherialize(). However, if the ORB is permitted to call _add_ref() as
many times as it likes, I have no idea how many times I have to call
_remove_ref() from within etherealize(). I think that the spec should say
that _add_ref() is called exactly once for these operations if the
corresponding servant is not in the AOM already.

I vaguely remember the discussion about optimization of the calls
to _add_ref() and _remove_ref(). I think the idea was to permit the ORB
to avoid redundant calls. However, it seems that the language in the spec
isn't precise enough. Under one interpretation, the refcount counts
the number of entries in the AOM. Under another interpretation, it counts
the number of calls in progress as well (because an ORB could call _add_ref()
when a call is dispatched and call _remove_ref() when it finishes).
Under yet a third interpretation, the refcount counts the number of
object references in my address space. That interpretation is happens
if every call to _this() calls _add_ref()...

The language is not precise enough, I think...

I believe the specification for insertion of object references to Anys is
somewhat ambiguous. And, if it is intended to be as I think, it may also be
less than ideal.

Consider the following idl:

interface B

;

And then consider the following code:

D_ptr d; // initialize this to non-null value somehow
B_ptr b = B::_narrow(d);
Any ab;
ab <<= b;
Any ad;
ad <<= d;
// ...
B_ptr b_val;
if (ab>>=b_val)

; // 2
D_ptr d_val;
if (ab>>=d_val)

; // 4

>From what I can see of the spec, it is a bit unclear about whether then
insertion of an object should use the static type or the dynamic type of the
object to initialize the typecode. Simplicity and consistency with other
related operations suggests that it should use the static type. That is what
we currently do, and a quick test of ORBIX 2000 seems to tell me it does it
that way too.

With that interpretation, 1&4 will work, while 2&3 will fail.
Nobody should be surprised that 3 fails, but it is inconvenient
that 2 doesn't work.

If insertion used the dynamic type of its argument, then 1&2
would work, while 3&4 would fail.

To get reasonable behavior when derived types might be present,
(and how often can you be certain they cannot?)
it seems that one should almost never use type specific object
extraction. Instead, one must do something like:

Object_var o;
B_var bv;
if (ad>>=to_object(o._out()) &&
!CORBA::is_nil(bv = B::_narrow(o)))

;

This is unfortunately a bit inconvenient.

So, is there any text in the spec that says, when inserting an object
reference type into an any, if the repository id in the typecode should be
the static type of the inserted argument or the dynamic type of the value of
the argument?

If not, I think we need to add some text.

The C++ "Mapping of PortableServer Dynamic Implementation Routine" states
that "If DynamicImplementation::_this() is invoked outside of the context
of a request invocation on a target object being served by the DSI servant,
it raises the PortableServer::WrongPolicy exception".

This conflicts with the behaviour of _this() in static skeletons as de-
fined in "Skeleton Operations".

In particular, this means that DSI servants cannot be implicitly acti-
vated, and therefore, the choice of DSI vs. static skeleton is not trans-
parent to the application integrator.

Is there any rationale behind this?

I seem to remember that we decided to deprecate the void * functions
on type Any. However, looking at the latest C++ draft, they are not
marked as deprecated.

Can anyone remember what we decided to do? Is this a bug in the spec
or a bug in my memory?

C++ issue #1: The C++ specification should state that valuetype
parameters which are copied by the ORB for collocated invocations are
done using an ORB internal mechanism, not _copy_value().

The CORBA 2.3 C++ mapping is clear on what the use of
ValueBase::_copy_value is, but is unclear as to who is responsible for
providing an overriding definition of this pure virtual function. Is
it the IDL compiler, generating an overriding _copy_value() function for
each valuetype C++ class, or is the user, when he provides a valuetype
implementation class?

C++ issue #2: The ValueBase::_copy_value() function should be
deprecated in favor of a new ValueBase::_clone_value() operation:

// IDL
module CORBA {
abstract valuetype CloneContext { };
};

// C++
namespace CORBA {
...
class ValueBase

;
...
};

The _clone_value() function provides an independant copy of the
valuetype it is invoked on. Any valuetypes reachable via the state of
the original valuetype are also copied, and relationships between
original valuetype(s) will be preserved in the cloned copies. The
CloneContext argument provides the necessary state information for the
ORB to properly maintain relationships between copied valuetypes. If
_clone_value() is called with a null CloneContext, a new CloneContext
will be generated and returned by the ORB as a result of the call.

The CORBA 2.3 C++ specification, section 1.22 states that valuetypes
passed as in parameters to an invocation of an object are copied, even
if the object is collocated with the caller. It does not make this
statement for inout or out parameters (or return results), which
strongly suggests that valuetype copying is not necessary. In fact, the
text for valuetype inout parameters strongly suggests that copying is
not performed.

I think this is wrong and inout & out valuetypes should be copied as
well (inout parameters should be copied before and after the invocation,
while out and return values should be copied after the invocation
completes.) Without the copies, call transparency will be broken and
the client can distinguish between a local and a remote call.

The mapping for user exceptions and structures is identical, except
for one thing: user exceptions have an additional constructor with
one parameter for each member, so I can construct and throw the exception
with a single throw statement.

However, structures are second-class citizens: I can't instantiate and
initialize a structure at the same time. (Well, at least not in general,
because static initialization only works for aggregates and, at any rate,
I can only instantiate and initialize with compile-time constants.)

So, why don't we add the same constructor to the mapping for structures?
It seems inconsistent to have one mapping for structures and a slightly
different one for exceptions, when in fact they both could be the same.

Summary: This may be a naive question, but it seems like the C++ mapping
for value boxes is incomplete. As far as I can tell, the specification
defines the mapping for the following types:

basic types, enum, objref, string, wstring, struct, union, sequence, array, fixed, any

But what about the types that have not been mentioned?

value, value box, TypeCode, Principal, native, except

If these types aren"t allowed, does the specification say that somewhere?

Summary: We all know that the names of the C++ #include files generated from IDL are
not standardized, and are thus the one remaining major source portability
issue. One way to fix this would be to agree on some standard filenames,
but we"ve tried this before and have never succeeded.

Just thinking out loud here, but another way to fix it would be to agree on
some standard macro names that applications could use to portably include
the appropriate files. For example, define one macro for a client-side
include and one macro for a server-side include, both taking the basename
of the IDL file as an argument:

#ifndef CORBA_CXX_CLIENT_INCLUDE
#define CORBA_CXX_CLIENT_INCLUDE(base) <base ## Client.hh>
#endif

#ifndef CORBA_CXX_SERVER_INCLUDE
#define CORBA_CXX_SERVER_INCLUDE(base) <base ## Server.hh>
#endif

Obviously, the exact definition of the macro would depend on the names of
the generated files.

I believe these could then be used portably in the client and server source
code like this:

#include CORBA_CXX_CLIENT_INCLUDE(file)

With this approach, nobody has to change the names of the files they
currently generate.

Summary: Consider the following IDL:

----------------------------------------------------------------------
// IDL
typedef long MyArray[1000000];

interface I

;
----------------------------------------------------------------------

Now let"s assume that I want to implement this using the DSI:

----------------------------------------------------------------------
// C++
void
I_impl::invoke(ServerRequest_ptr request) throw()
{
String_var name = request -> op_name();

if(strcmp(name, "set_array") == 0)

else

}
----------------------------------------------------------------------

At the line I marked with XXX, I have to set the TypeCode of the
Any.

Summary: Value Boxes inherit from CORBA::ValueBase and
must therefore implement get_value_def(), but
cannot return a sensible value.

Summary: Michi Henning & Steve Vinoski have previously challenged people to come
up with a modification to the C++ language mapping that would allow for
type safe widening of object reference _var types for assignment or copy
construction. I believe I have come up with the solution, and Michi
agrees with me:

Proposal:

For object reference _var types, replace the copy and assignment
operators:

class T_var

;

with:

class T_var {
public:
...
template <class C_var>
T_var(const C_var &cv) : ptr_(T::_duplicate(cv.in()) {
}

template <class C_var>
T_var &operator = (const C_var &cv) {
if ((void *)this != (void *)cv)

return *this;
}
...
private:
T_ptr ptr_;
};

Summary: There seems to be nothing in the spec that specifies the name of the include files for things in the CORBA module (e.g. type definitions). Add such a requirement to each language mapping

Summary: In a number of places in the sample code for the C++ binding, the symbol
_ptr is used as a member variable in _var types. In all of these the
actual member is shown as private, so it is clear that actual
implementations could use something else.

The one exception to this is in section 20.10 on the Mapping for Struct
Types. This discusses (without sample code) the mapping for string and
object reference members of structs.

Summary: The current C++ chapter for custom marshalled values on p. 20-94 lacks
information about how one implements a custom value.
The "Value Type Semantics" chapter in "ptc/98-10-06, 15 Oct. 98[REVIEW]"
p. 5-11 says,
" The implementer of a custom value type shall provide an
implementation of the CustomMarshaller operations. The manner
in which this [sic] done shall be specified for each language mapping..."

Summary:
The C++ mapping does not permit extraction of a Fixed from an Any
in a generic way – I must always specify matching digits and scale in
order to call Any::to_fixed(). This is inconvenient if an application
wants to deal with Fixed values generically (because Fixed is a generic
type in C++ anyway).

Proposal:

Add an overloaded >>= operator for extraction of Fixed from an Any.
The operator sets the Fixed value to whatever scale and digits
are present in the type code.

Cheers,

Michi.

Summary: The mapping right now offers Any::to_fixed to get a Fixed value out of an Any:

to_fixed(Fixed &f, UShort d, UShort s);

The spec doesn"t state what should happen if the digits and scale do
not match what is in the type code. I believe extraction should fail in
this case.

Summary: Section 20.9, page 20-28 of orbos/98-07-12 describes what types are
considered variable-length. Since the new Fixed class has non-trivial
constructors, it should also be a considered a variable-length type. Note
that any fixed-length struct containing one cannot be statically initialized.

Summary: There is no mention of PortableServer::ValueRefCountBase after this
page. It is not clear why values that also implement interfaces
do not use the same reference counting scheme as other values.

Summary: The C++ mapping document (98-09-03, p. 108) states that "... the callee
shall receive a copy of each valuetype argument passed to it even if the
caller and callee are collocated in the same process."

In the collocated case, should the ORB invoke _copy_value() to produce
the copy?

Since the user could implement _copy_value() to return a nil value, it
seems unlikely that the ORB could rely on this mechanism. However, a
properly implemented _copy_value() would likely provide a significant
speed improvement over marshalling and unmarshalling.

Summary: Section 20.21, "Argument Passing Considerations," says that for valuetypes:
"The caller shall eventually invoked _remove_ref on the valuetype
instance it receives back as either an inout, out, or return value."

For memory management purposes, this is not sufficient in some cases.

Summary: What happens if I write

CORBA::Any a = ...;
const char *p;

a >>= p;

and the type code in the Any indicates a bounded string? Does the extraction
succeed or fail? The mapping doesn"t say.

The semantics for insertion of a valuetype into an Any are unclear.
(Note, this is related to issue 2531 in the IDL-to-Java RTF.
It is also related to orb_revision issue 3205.)

In section 1.16.2 of ptc/2000-01-02, two forms of insertion are defined:
copying and non-copying. The non-copying form is described as:

"The noncopying valuetype insertion consumes the valuetype pointed to by the
pointer that T** points to. After insertion, the caller may not access the
valuetype instance pointed to by the pointer that T* points to. The caller
maintains ownership of the storage for the pointed-to-T* itself."

There is no specific description of the copying form specific to valuetypes,
so the generic description must apply:

"For the copying version of operator<<=, the lifetime of the value in the
any is independent of the lifetime of the value passed to operator<<=. The
implementation of the any may not store its value as a reference or pointer
to the value passed to operator<<=."

One possible interpretation (1) is that the copying form should be
implemented via a call to the _copy_value virtual function, while the
non-copying form should simply retain the provided pointer (without calling
_add_ref) and eventually call _remove_ref when done with it.

If so, what is the significance of the rule about the caller not continuing
to use the pointer? It it only that it has lost a reference count, and may
continue using the pointer if it has another reference count? Or does this
imply that continued access to the value is forbidden regardless of
reference count?

Another possible interpretation (2) is that the description is nonsense, and
that the non-copying form should use _add_ref and the copying form should
use _copy_value. In this interpretation the caller would be free to continue
using the original pointer and would be obligated to _remove_ref it
eventually. This seems like a more practical interpretation, but is
inconsistent with usage for other non-copying insertions.

Suggested Resolution:

Replace the paragraph on non-copying insertion of valuetypes (quoted above)
with:

"The noncopying valuetype insertion takes ownership of one reference count
to the valuetype pointed to by the pointer that T** points to. After
insertion, the caller should treat the pointer as if _remove_ref had been
called on it. The caller maintains ownership of the storage for the
pointed-to-T* itself."

"For copying valuetype insertion, the lifetime of the value in the any is
independent of the lifetime of the value provided. The implementation of the
any shall duplicate the value using the virtual function _copy_value or an
equivalent mechanism. The caller retains ownership of the T* pointer and
remains obliged to call _remove_ref on it."

The mapping is currently a bit ambiguous about the insertion/extraction
operators for Any. It says:

Assuming that boolean, char, and octet all map the C++ type
unsigned char, the private and unimplemented operator<<= and
operator>>= functions for unsigned char will cause a compile-time
error if straight insertion or extraction of any of the
boolean, char, or octet types is attempted.

This is ambiguous. It is not clear what is qualified by the first part
of the sentence "Assuming that...". It could qualify all of the paragraph,
in which case the interpretation is:

Only on platform where these three types indeed map to the same
C++ type will it be necessary to have these unimplemented operators.

// C++ Octet
oct = 040;
Any any;
any <<= oct; // this line will not compile
any <<= Any::from_octet(oct); // but this one will

This is unambiguous, but it doesn't make it clear whether this will be
the case for all mapping implementations, or only those where the
IDL types map to ambiguous C++ types.

It is important to note that the previous example is only one
possible implementation for these helpers, not a mandated one.
Other compliant implementations are possible, such as providing
them via in-lined static any member functions if boolean, char,
and octet are in fact mapped to distinct C++ types. All
compliant C++ mapping implementations must provide these helpers,
however, for purposes of portability.

Again, this is slightly ambiguous because, even though it requires the
presence of the helpers, it doesn't make any statement about whether the
prohibition of the direct insertion operators is mandatory for all
implementations.

I would suggest to clarify the text to state that direct insertion/extraction
of Bool, Octet, and Char must fail with a compile-time error, regardless
of how these types are mapped.

Question: Is it legal to do the following if I use a C++ mapping that
uses C++ exceptions instead of using CORBA::Environment to
handle errors?

CORBA::Environment my_env;

The spec says (page 114):

The C++-PIDL specification differs from the C-PIDL specification
as follows:
[...]
Supports a default constructor that initializes it to hold no
exception information.

However, the class definition that follows does not show the
default constructor.

So, the text disagrees with the class definition that is shown because
"supports a default constructor" does not have a "may" or "might", so
the text would appear to make the default constructor mandatory for
both EH and non-EH mappings.

The C++ language mapping does not specify what criterion should be used to
determine the validity of a typesafe extraction from an Any.

The closest it ever comes is the statement in 1.16.3:
"In this case, the version of operator>>= for type Long must be able to
determine whether the Any truly does contain a value of type Long...".

There are two obvious candidates: the equal and equivalent operations of
TypeCode.

Proposed resolution:

Replace the first sentence of 1.16.3:

"To allow type-safe retrieval of a value from
an any, the mapping provides the following
operators for each OMG IDL type T:"

with:

To allow type-safe retrieval of a value
from an Any, the mapping provides an
operator>>= for each OMG IDL type T.
Each such operator returns a boolean
indicating success or failure, and if
successful, makes the value of the any
available via a user supplied argument.
The success of the operation is determined
by applying the

needs to be discussed. I believe there
are valid arguments for each.

I believe that the CORBA::release and CORBA::is_nil
functions that take a POA_ptr argument do not
reference the proper scope:

1.41.11 release and is_nil
// C++
namespace CORBA

Should be:

1.41.11 release and is_nil
// C++
namespace CORBA

I don't see in the specification where the scope
of POA_ptr is explicitly defined. But, I believe
that the correct definition of POA_ptr is in the
PortableServer namespace (i.e. the enclosing scope
of the POA class).

Then again I can't find anything in the specification
that asserts that any Foo_ptr type must go in the
immediately enclosing class or namespace containing
the Foo type. :-/

Also, if POA_ptr is in PortableServer when an ORB is
mapping of modules to classes the definition of the
above release and is_nil functions in the CORBA class
will be impossible.

So I feel compelled to ask:

Do we really need to have release and is_nil functions
for types outside of the CORBA module?

the new mapping for _var types for fixed-length underlying types shows:

T_var &operator=(T *t) {
if (t != m_ptr)

return *this;
}

This guards against self assignment when a pointer is assigned to the _var.

I don't think this is right:

• Assigning a pointer to a _var that already owns what the pointer
  points at is almost certainly an error:

MyStruct * p = ...;
MyStruct_var v = p; // OK

// ...

v = p; // Almost certainly an error

• We don't do the same thing elsewhere. On page 1-13:

A_var &operator=(const A_var& a)

This is inconsistent: assignment of a _ptr to a _var reference
is not safe against self assignment, so assignment of a pointer
to the _var for a fixed-length type shouldn't be either.

Note that the other assignment operators are just fine – I'm only objecting
to testing for self-assignment for operator= with a pointer as the RHS.
(A nice compiler could even insert an assertions if a _var is assigned the
same thing that it already points at.)

the spec currently says (page 1-101):

Request invocations made through the DII may result in
user-defined exceptions that cannot be fully represented
in the calling program because the specific exception type
was not known at compile-time. The mapping provides the
UnknownUserException so that such exceptions can be represented
in the calling process: [...]

Here is a code snippet for the DII:

req->invoke();
if (req->env()->exception())

The para on page 1-101, by implication, says that:

• If there are concrete C++ types available in the caller that
  can represent the user exception, a real user exception is
  instantiated and the pointer returned by exception() points
  at the concrete user exception instance.

• If there is no concrete C++ type available in the caller for
  a user exception, the pointer returned by exception() points
  at an instance of UnknownUserException.

It's not as clearly spelled out as this, but it can be implied from the
words on page 1-101.

This is bad. For one, it implies "action at a distance". For example,
linking the stubs for a user exception into a completely unrelated
part of the same binary (possibly via a library) would change
the behavior of the above DII call. Further, to make this behavior
happen would require some form of global initialization data structure.
In effect, there would have to be something that would let the ORB
know (globally) for which user exceptions stub code is linked in.

We rejected the need for global data recently in another context (for
the proposed extraction operator for user exceptions). For the same reason,
we should reject this here and mandate that, if I use the DII, all user
exceptions are always returned as UnknownUserException.

I think we have a defect/omission in the C++ mapping with respect
to exception safety. Consider a servant class with a constructor:

class FooImpl : public virtual POA_Foo

;

Consider what happens if FooImpl() throws an exception for some reason.
By the time FooImpl() runs, the base class constructor POA_Foo() has run
already. So, when the compiler deals with the exception, it invokes
~POA_Foo().

The problem arises because, in our implementation at least, ~POA_Foo()
checks if the reference count is zero and asserts if not.

There doens't appear to be any portable way to implement an object that
inherits from an abstract interface using the DSI in C++ without compile
time knowledge of the abstract interface. The basic problem is that I
can create an object reference from a POA, but there is no way to
convert the reference into an abstract interface reference so that I can
send it out on the wire.

We need some mechanism to coerce an object reference into an abstract
interface reference (with a runtime check) to make this work.

what should _default_POA() return if no default ORB exists in the server
(that is, ORB_init() with an empty ORB ID was never called)?

Summary: The ValueBase::_copy_value() function is included in the discussion
of the "reference counting interface" in section 1.17.5 of 99-07-41.
Later, the description of the reference counting mix-in classes
says:

"Each of these classes shall be fully concrete and shall completely
fulfill the ValueBase reference counting interface..."

However, I do not believe that the intent was to require the mix-in
classes to implement _copy_value().

Therefore I suggest one of two clarifications:

1) Move the discussion of _copy_value() out of the reference-counting
section, or

2) Specify which functions the mix-in classes are actually expected to
implement, e.g.,

"Each of these classes shall be fully concrete and shall completely
fulfill the ValueBase reference counting interface (_add_ref,
_remove_ref, and _refcount_value), ..."

the spec says (on page 1-23):

The T* constructor creates a T_var that, when destroyed, will
delete the storage pointed to by the T* parameter. The parameter
to this constructor should never be a null pointer. Compliant
implementations are not required to detect null pointers passed
to this constructor.

This seems broken for two reasons:

• In an environment without real exceptions, null pointers must
  be returned for variable-length types in the presence of an
  exception. So if I write

Foo_var fv(some_ref->op(my_corba_environment));

and op() raises an exception (which will be returned in the
environment), I'm hosed.

• The assignment operator permits assignment of null, but the
  constructor doesn't. This is inconsistent, if nothing else.

It seems that disallowing initialization from null pointer is some
historical hangover? I think the restriction should be removed.

In section 1.17.5, the C++ class for Valuebase does not show the
get_value_def operation.

Section 1.17.7.5 states: "Value boxes for these types [including Fixed]
map to classes that have exactly the same public interfaces as the
underlying boxed types...".

Does this also include overloaded operators that are defined in the
Fixed class?

It sure seems weird to allow some operators to work but not others:

// IDL
valuetype FixedVal fixed<5,2>;

// C++
FixedVal *fv = ...;

++fv; // legal?

CORBA::Fixed f = fv * 2; // illegal?

In the CORBA 2.3 spec, section 6.2, CORBA::AbstractBase is defined as:

module CORBA

;

This implies that the C++ mapping for AbstractBase when used as a
parameter is like this:

class DataOutputStream

;

But section 1.18.1 of the CORBA 2.3 C++ mapping makes it clear that the
signature should be:

class DataOutputStream

;

Now I know that DataInputStream & DataOutputStream can be special cased
to handle this, but if we need to add additional operations that use
AbstractBase in the future, it would be nice if this could be fixed to
behave consistently with the other native type mappings to C++.

Summary: The C++ language mapping chapter contains many blocks of IDL like stuff
with the comment

// IDL

in the first line, but the stuff in the block is not valid IDL for
various reasons:

Uses "pseudo" as an apparent keyword.

(ii) Contains declarations like

attribute exception exception;

I suggest that the comment "// IDL" be replaced by "// Augmented IDL
(see "Usage" on page x-y)" cross-referencing to section 20.23 Usage, so
that people know for sure that this is not IDL.

Furthermore, to make the claim in section 20.23 true, the declaration:

attribute exception exception;

should be fixed to be something else, or alternatively, the exceptional
use of exception should be called out as a specific augmentation of IDL
in section 20.23.

consider:

// IDL

struct VariableStruct

;

interface I

;

Then:

// C++
void
MyImplForI::foo(VariableStruct_out s)

void
MyImplForI::bar(VariableStruct_out s)

The freeing of memory for out params relies on the default conversion
by the _out constructor from a pointer to the _out type which, as a
side effect, frees the memory return by the previous call. However,
in this case, and _out param is passed to another call, so the
assignment operator runs, not the constructor:

T_out& operator=(T* p)

The assignment operator doesn't free the previous memory, so we get
the leak.

Should the assignment operator be changed?

The mapping requires that it must be possible to insert CORBA::Exception
into an any (section 1.19.2).

Question: Is it possible to insert UnknownUserException into an Any?

If the answer is yes, what are the members of UnknownUserException, what
is its CDR representation, and what is its TypeCode?

If the answer is no, we should clearly state this and specify what happens
if I attempt to insert UnknownUserException into an Any.