Dynamically tagged expression not allowed. Why?
Marc A. Criley
using the binding generated by g++ -fdump-ada-spec as a starting point.
While the binding will compile cleanly, the client code gets a
compilation error, "dynamically tagged expression not allowed", when
trying to set up the declarations to utilize it.
Stripping out a test case, I see that it has nothing to with the C++
binding per se, but it's an Ada issue that's flummoxing me. Here's the
test code:
procedure Dyty_Test is
type Class_Type is tagged limited record
null;
end record;
function New_Class_Instance return Class_Type'Class;
Conn : Class_Type := New_Class_Instance;
function New_Class_Instance return Class_Type'Class is
begin
-- Just a stub
return New_Class_Instance;
end New_Class_Instance;
begin
null;
end Dyty_Test;
The Class_Type definition MUST be a tagged limited record to correspond
to the C++ class, and the New_Class_Instance function (the constructor)
MUST return a class-wide instance of that type. (There are
corresponding C++ pragmas as well for these, but the error is the same
whether they're present or not.)
The error is on the declaration of "Conn", with the error occurring on
the invocation of the initializing New_Class_Instance function. I'm just
not understanding something here.
I've tried a number of variations, but these have more or less become
guesswork, and so it's time to ask for help :-)
Thanks.
Marc A. Criley
McKae Technologies
www.mckae.com
Dmitry A. Kazakov
> Stripping out a test case, I see that it has nothing to with the C++
> binding per se, but it's an Ada issue that's flummoxing me. Here's the
> test code:
>
> procedure Dyty_Test is
>
> type Class_Type is tagged limited record
> null;
> end record;
>
> function New_Class_Instance return Class_Type'Class;
>
> Conn : Class_Type := New_Class_Instance;
Should Conn be mutable? Otherwise you could use a function instead.
> function New_Class_Instance return Class_Type'Class is
> begin
> -- Just a stub
> return New_Class_Instance;
Infinite recursion. I suppose it is
return X : Class_Type;
> end New_Class_Instance;
>
> The error is on the declaration of "Conn", with the error occurring on
> the invocation of the initializing New_Class_Instance function. I'm just
> not understanding something here.
You are "returning" a class-wide limited object, which is to "initialize" a
specific object.
A more difficult problem is that to initialize Conn, GNAT wants to know the
body of the initializer.
--------------------------------------------------------------
Variant 1. Two packages (to have the body):
package P is
type Class_Type is tagged limited record
null;
end record;
function New_Class_Instance return Class_Type'Class;
function New_Class_Object return Class_Type;
end P;
package P.Instances is
Conn : Class_Type := New_Class_Object;
end P.Instances;
package body P is
function New_Class_Object return Class_Type is
begin
return X : Class_Type;
end New_Class_Object;
function New_Class_Instance return Class_Type'Class is
begin
return New_Class_Object;
end New_Class_Instance;
end P;
----------------------------------------------------------------------------------------------------
Variant 2. Old good Ada initialization (in 80's we knew how to do it right)
package P is
type Class_Type is tagged limited record
null;
end record;
function New_Class_Instance return Class_Type'Class;
Conn : Class_Type; -- It will be OK upon the body elaboration!
end P;
package body P is
procedure Construct (Object : in out Class_Type) is
begin
null;
end Construct;
function New_Class_Instance return Class_Type'Class is
begin
return X : Class_Type do
Construct (X);
end return;
end New_Class_Instance;
begin
Construct (Conn); -- Fix it now
end P;
--
Regards,
Dmitry A. Kazakov
http://www.dmitry-kazakov.de
Randy Brukardt
news:c2ead$4c0150a0$4ca4a823$11...@API-DIGITAL.COM...
...
> Stripping out a test case, I see that it has nothing to with the C++
> binding per se, but it's an Ada issue that's flummoxing me. Here's the
> test code:
>
> procedure Dyty_Test is
>
> type Class_Type is tagged limited record
> null;
> end record;
>
> function New_Class_Instance return Class_Type'Class;
>
> Conn : Class_Type := New_Class_Instance;
This is illegal because the function returns a class-wide object and you are
assigning it into an object of a specific type. You have to explicitly use a
type conversion here:
Conn : Class_Type := Class_Type (New_Class_Instance);
to specify that you want to truncate the returned object, or make the object
classwide:
Conn : Class_Type'Class := New_Class_Instance;
But notice that while both of these are legal, they'll both raise
Program_Error because an access-before-elaboration error (the body of
New_Class_Instance hasn't been elaborated at the point of this call).
Randy.
Yannick Duchêne (Hibou57)
a écrit:
> You have to explicitly use a
> type conversion here:
>
> Conn : Class_Type := Class_Type (New_Class_Instance);
>
allowed on limited type due to their copy based semantic.
[ARM 2005 4.6] says about type conversions
“Neither the target type nor the operand type shall be limited”
> to specify that you want to truncate the returned object, or make the
> object
> classwide:
>
> Conn : Class_Type'Class := New_Class_Instance;
That is a beautiful one (to my eyes)
> But notice that while both of these are legal, they'll both raise
> Program_Error because an access-before-elaboration error (the body of
> New_Class_Instance hasn't been elaborated at the point of this call).
Not necessarily, as he can move the body of the function returning
Class_Type before the declaration of the variable initialized from this
function.
Instead of
function New_Class_Instance return Class_Type'Class;
Conn : Class_Type := New_Class_Instance;
function New_Class_Instance return Class_Type'Class is
begin
-- Just a stub
return New_Class_Instance;
end New_Class_Instance;
Simply do
function New_Class_Instance return Class_Type'Class;
function New_Class_Instance return Class_Type'Class is
begin
-- Just a stub
return New_Class_Instance;
end New_Class_Instance;
Conn : Class_Type := New_Class_Instance;
and everything will go fine.
--
There is even better than a pragma Assert: a SPARK --# check.
--# check C and WhoKnowWhat and YouKnowWho;
--# assert Ada;
-- i.e. forget about previous premises which leads to conclusion
-- and start with new conclusion as premise.
Yannick Duchêne (Hibou57)
<mai...@dmitry-kazakov.de> a écrit:
>> Conn : Class_Type := New_Class_Instance;
>
> Should Conn be mutable? Otherwise you could use a function instead.
>
May be the reason was a matter of efficiency ?
Dmitry A. Kazakov
> Le Sat, 29 May 2010 20:23:36 +0200, Dmitry A. Kazakov
> <mai...@dmitry-kazakov.de> a écrit:
>>> Conn : Class_Type := New_Class_Instance;
>>
>> Should Conn be mutable? Otherwise you could use a function instead.
>>
> Good note (formally talking).
>
> May be the reason was a matter of efficiency ?
A renaming could also go, but it would not because of the "unseen body"
problem.
Dmitry A. Kazakov
> Not necessarily, as he can move the body of the function returning
> Class_Type before the declaration of the variable initialized from this
> function.
You cannot have bodies in the package specifications. This is a difficult
problem, since present construction model does not support deferred
initialization. In short, you cannot have this in a single package.
Yannick Duchêne (Hibou57)
<mai...@dmitry-kazakov.de> a écrit:
> You cannot have bodies in the package specifications. This is a difficult
> problem, since present construction model does not support deferred
> initialization. In short, you cannot have this in a single package.
So indeed, if the variable is to be in a spec, it needs two packages, one
containing the function, and one containing the variable, so that one
package -- the one containing the function -- can be elaborated before the
other -- containing the variable. A child package would be a good choice
to hold that variable.
Adam Beneschan
<yannick_duch...@yahoo.fr> wrote:
> Le Wed, 02 Jun 2010 01:36:17 +0200, Randy Brukardt <ra...@rrsoftware.com>
> a écrit:> You have to explicitly use a
> > type conversion here:
>
> > Conn : Class_Type := Class_Type (New_Class_Instance);
>
> This one is not legal, as Class_Type is limited and conversion are not
> allowed on limited type due to their copy based semantic.
> [ARM 2005 4.6] says about type conversions
> “Neither the target type nor the operand type shall be limited”
Umm, no, that sentence (4.6(24.7)) applies only to array types. (And,
I think, only in cases where there is no common ancestor of the
operand and target type.) However, 7.5(2.1-2.2) makes it illegal.
Type conversions on limited types are allowed in some cases; e.g. if
T1 is a limited tagged type, and T2 is a type extension, and X2 is an
object of type T2, you can use T1(X2) as an actual parameter to a
procedure that takes an IN OUT parameter of type T1. This is a view
conversion and doesn't involve a copy.
-- Adam
Marc A. Criley
Actually, stripping out the C++ interface pragmas was probably a bad
idea on my part. (Though the thread *was* educational :-)
I thought that to create an instance of the C++ class I had to invoke
the New_Class_Instance function myself, from which arose this issue.
Turns out I don't. GNAT's handling of the C++ pragmas does it all for me
automagically.
The documentation on creation an Ada binding to C++ is somewhat...terse
when it comes to describing the mechanics of what is going on "under the
hood", from whence came my misunderstanding.
So this part is all better now, I can step in and verify that my
constructors are being executed. I'm having other problems, but that's
just the usual banging the head against the wall stuff.
Marc A. Criley
www.mckae.com