Re: Sending Variable Length Messages with GNAT.Sockets

[Simon Wright <simon@pushface.org>]

"Marc A. Criley" <mcNOSPAM@mckae.com> writes:

> Simon Wright wrote:
>
>> If the C side says
>>   struct rec {
>>     char flag;
>>     int count;
>>   };
>> there will typically be 3 bytes of padding between flag & count, for
>> 8
>> bytes on the wire.
>> If the Ada side says
>>   type Rec is record
>>      Flag : Interfaces.Unsigned_8;
>>      Count : Interfaces.Integer_32;
>>   end record;
>>   pragma Convention (C, Rec);
>> the 'Write will output *5* bytes; but an unchecked conversion to
>> Stream_Element_Array (1 .. 8) will succeed & do The Right Thing
>> (endianness permitting).
>
> This is why trying to use Ada.Streams "out-of-the-box" to communicate
> between Ada applications and those written in other languages is
> problematic at best.  You pretty much have to write your own stream
> I/O routines for each type to be transferred to get the data into an
> interoperable layout, especially for structured types like those in
> Simon's example.
>
> I've pretty much given up on doing that, only using streams for
> intra-system Ada-to-Ada communication, and doing Unchecked_Conversions
> to Stream_Element_Array and using GNAT.Sockets for all external comms.

Our project has lots of Ada (our bit) and lots of C (the other
bits). The interfaces are defined using only 32-bit quantities (after
a packed 20-byte header which initially caused us a pile of grief,
especially since we wanted to be interoperable between PC and
PowerPC).

On our side there's quite a lot of generated code to cope with mapping
Booleans to packed arrays of bits, driven by UML tags in the models,
but once the data is in the 32-bit form we do indeed use Streams.

I suppose we could have achieved the same end by writing our own
'Read, 'Write, but that might have been a bridge too far; and this way
at least the main code generator isn't intertwingled with the network
i/o part. I can't remember whether that argument occurred to me at the
time!