TODO

Use std::shared_ptr with a custom allocator for our SmartPtr. This will ensure
thread safety.

In all codecs/negotiators:
  peek()
  rdbuf().in_avail() and cache result
  rdbuf().sgetc() for each char
  rdbuf().xsgetn() for many (in negotiators)

Need to add to the codec interface so that we can verify codec is at
stream level before and after building a datagram.

Logic in StreamFoo::poll() is bogus. can_read does not mean that sockets
can be read from many times, just once, so can_read should be checked outside
the do loop, then inside a while check should be used to copy all bytes
from the stream into m_buf at once. This makes the do..while loop redundant.
We can read all the bytes into the buffer straight away. The neogitaters
are not byte based, but expect all available input to have been copied into
the buffer.

Add IHEAR to negotiation, so that it is possible to negotiate various
alternate transports include UDP, SCTP etc..

Get rid of overridden removeAttr, as the base class handles this now.
See if copyAttr and setAttr can be better handled this way.
copyAttr returns 0 if attribute is hard coded, but not set.

Add template member to Objects for list types, so that value can be queried
into anything that implements operator[]
Something like:
    template<typename VectorT>
    int getPos(VectorT & v) {
        v[0] = m_pos[0];
        ...
    }
How much does this really help? Not sure.

Make sure objects deep copy really copies arguments of ops.

hasAttr currently returns true if a hardcoded attribute is not really set.
Needs fixing - very counterintuitive.

Don't use explicit Element constructor when assigning to another Element.

Note that Atlas::Message::Element segfaults of a map is set to be a member
of itself.

Don't pass Bridge to StreamFoo at construction time. Its not required until
getCodec() time. Make it an argument of getCodec(). Bridge passed to Codec
should be a reference rather than a pointer.

Fix all use of if (..) .. to always use { } in Root.h etc.

Make sure parents and objtype are not sent if they are empty.

If we move get|set(...)AsList() methods from Root.h and other headers,
into body, we will remove the need for Element.h from any header.
This is also the key to removing SmartPtr.h from Root*.h

There is some inefficient copying going on in the way args are encoded.

Note that most people are not interested in the Objects::Decoder also
handling dispatch. Just decode and pass to the application.

Invstigate removing methods which only exist to prevent hiding methods in the
base class. This should be possible now that the methods do not have the same
name.

Ensure that the Decoder method registration code is type safe. Probably good
to use templates.

Items below are probably still valid, but this is where i reconsider the
order and priority of tasks, and make my own notes.

Write a very fast binary codec, with a name derived from our architecture.

Make the code for adding new decoder methods templaty so we can ovoid nasty casting.

Write codecs, connectors and acceptors for datagram based communication,
specialising Codec<> template using an abstract datagram class interface.

Write acceptor/negotiator for file based atlas, to load from a file
of unknown codec.

Modify the objects API so we can track which attributes have been set, rather
than relying on their value. Then we can get orientation, and perhaps even
bbox into the spec, and support that well.

Add cout operators for Message::Object at least. Its a pain having to mess about
with them.

TODO list for Atlas-C++:

Items are written up by those who they are assigned to, unless otherwise noted
(using " -- ").

Notes:

Stefanus Du Toit is no longer participating largely.
Michael Day is no longer participating largely.
Al Riddoch is no longer participating largely.
Erik Ogenvik has now taken over ownership this code.