Re: Don't we already have a 'Valid? (was Re: Unchecked_Conversion...)

[eachus@mitre-bedford.arpa (Robert I. Eachus)]

In article <CD59oG.Mru@yktnews.watson.ibm.com> ncohen@watson.ibm.com (Norman H.
 Cohen) writes:

   > It seemes to me that the 9X RM actually needs a little more
   > tinkering to make it work...

   I follow Norm's logic, but I can't agree with the conclusion.  As
Norm reads the manual right now (and I agree) the following sequence
should either assign true to B or raise PROGRAM_ERROR:

    declare
      I: Integer := Signalling_NaN;
      B: Boolean := False;
      S: Short_Float; -- change as necessary for your implementation...
      function UC is new Unchecked_Conversion(From => Integer, To => Float);
    begin
      S := UC(I);
      B := Short_Float'VALID(S);
    end;

    changing the order to:

    declare...
    begin
      B := Short_Float'VALID(UC(I));
      S := UC(I);
    end;

    may change things but doesn't really help.  (In fact it is
technically slightly worse.  There can be cases where B is assigned
True AND PROGRAM_ERROR is then raised. But I can't imagine a real
instance of this program where the call to UC produced different
results.)

    What you really want to be able to say is:

    declare...
    begin
      B := Short_Float'VALID(I);
      S := UC(I);
    end;

    Which would test whether the bits of I are would be a valid value
of type short_float after the Unchecked_Conversion.  I'm not
recommending this feature be added to the language, I'm just pointing
out that it probably has a much lower (implementation and run-time)
cost than requiring that B never be stored in a floating point
register until its value is formally accessed.

--

					Robert I. Eachus

with Standard_Disclaimer;
use  Standard_Disclaimer;
function Message (Text: in Clever_Ideas) return Better_Ideas is...