Re: Allocators and exceptions

["Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de>]

On Wed, 12 Sep 2007 17:45:07 -0500, Randy Brukardt wrote:

> "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> wrote in message
> news:1bi8swgas8pjl$.1tr7pfyemom8q.dlg@40tude.net...

> As it stands, a compiler that does the right thing (in the sense of avoiding
> a memory leak when it is certain that no reference to the object remain) is
> actually wrong vis-a-vis the language definition.

This need to be fixed.

>> Unchecked_Deallocation may not be called on a pointer to non-constructed
>> objects. Semantically it is not a pointer at all. It is a raw address on
>> which Deallocate has to be called. This is indeed a problem. If Ptr is
>> invalid, then Ptr.all'Address is erroneous.
>>
>> A related issue, why on earth "new" is allowed to propagate anything but
>> Storage_Error or else Program_Error?
> 
> "new" doesn't propagate anything other than Storage_Error or Program_Error.
> The initialization expression, OTOH, can propagate anything it wants. Those
> are separate things from a language perspective; the problem is that you
> can't *write* them separately.

Because there is no constructors. One can reject Maciej's idea of rolling
everything back, but then the language shall be consistent with
un-/partially constructed objects, hence constructors made visible, storage
pools visible, etc, in order to manually roll a failed initialization back
and then to deallocate. Otherwise, it shall not allow such objects. The
second choice *implies* that any exceptions propagating through "new" were
illegal, hence converted to Program_Error. The current status is obviously
unsatisfactory.

As for your concern about efficiency of construction with exceptions. I
think that contracted exceptions could fix that nicely. With them Maciej's
example:

   type T (Init : Integer) is record
      C : Positive := Positive (Init);
   end record;

would become illegal because the default contract of T's constructor will
be exception-free. The compiler will reject Positive applied to Init for
the component C as potentially raising an unlisted exception. No any
roll-back code need to be generated. In rare cases when Constraint_Error
exception were indeed the programmer's intention, he could add it to the
contract and so the compiler would say - aha, this is what you want, then I
will generate roll-back stuff for this.

-- 
Regards,
Dmitry A. Kazakov
http://www.dmitry-kazakov.de