Re: 32 and 64 bit floats

Re: 32 and 64 bit floats

[Keith Thompson]

In <dewar.820957348@schonberg> dewar@cs.nyu.edu (Robert Dewar) writes:
> Keith says
> 
> "I think we both agree that the sentence is poorly worded.  In any case,
> it's obsolete.  AI-00459, which was approved by the ARG and WG9 and
> applies to Ada 83 implementations, states that predefined integer types
> may have the same range, and RM95-3.5.4(25) says, "Different predefined
> integer types are allowed to have the same base range.""
> 
> Oops, no! AI's do not make anything in the RM obsolete, they merely
> point you to the right interpretation, so the existence of AI-00459 just
> goes to prove that my interpretatoin was right, and yours wrong (it's
> sort of like papal councils telling you after the fact whether you have
> sinned :-)

Well, that's the theory behind the AI's, but in practice some of them
do contradict (and effectively make obsolete) clear wording in the
reference manual.  I won't try to extend the papal council analogy.  8-)}

For example, the Ada 83 RM clearly states that integer division by zero
raises Numeric_Error (assuming the check is not suppressed), yet AI-00387
allows and encourages implementations to raise Constraint_Error instead.
Similarly for Ada 95, AI-00101 (not yet formally approved) says that
Ada.Task_Identification.Abort_Task's parameter is of mode "in", even
though the RM clearly shows its mode as "in out".

Furthermore, the wording of AI-00459 (the one that allows different
predefined integer types to have the same range) does not attempt to claim
that the original wording in the Ada 83 RM allows this interpretation.
It merely justifies the change on the basis that the restriction serves
no useful purpose.

I'm not saying that this is bad, merely that not all AIs are really just
"interpretations" of the original text, or even of the original intent.

I once heard that either Algol or Algol 68 had a similar mechanism for
issuing official interpretations of the language standard.  The committee
once issued an "interpretation" stating that, in a particular sentence
in the reference manual, the word "is" shall be interpreted to mean
"is not".  (I don't have a source for this story, and I can't guarantee
its accuracy.)

In the unlikely event that anyone else is interested in this, the text
of AI-00459 is at
<ftp://sw-eng.falls-church.va.us/public/AdaIC/standards/83com/ai-00459-bi.wa>.

-- 
Keith Thompson (The_Other_Keith)  kst@thomsoft.com (kst@alsys.com still works)
TeleSoft^H^H^H^H^H^H^H^H Alsys^H^H^H^H^H Thomson Software Products
10251 Vista Sorrento Parkway, Suite 300, San Diego, CA, USA, 92121-2718
Et tu, Barada Nikto?

Re: 32 and 64 bit floats

[Robert Dewar]

I said

"> Oops, no! AI's do not make anything in the RM obsolete, they merely
> point you to the right interpretation, so the existence of AI-00459 just
> goes to prove that my interpretatoin was right, and yours wrong (it's
> sort of like papal councils telling you after the fact whether you have
> sinned :-)"

to which Keith went into a long explanation of how some AI's are not
really interpretations but changes.

Keith, how many smileys do you need in a message :-) :-) :-)

By the way, the story about the insertion of not is entirely apocryphal,
I was familiar with both the efforts Keith mentions (I was chair of the
Algol 60/68 maintenance committee for many years). We must used this
mythical happening as a way of emphasizing just how much interpretation
freedom we had, but in formal terms, WG9 did not change the standard,
they only interpreted it (here are some more smileys just to make things
clear :-) :-) :-)

Historically, the argument that the RM did not in any case require 
long_integer to be longer than integer did play a part (I was there!)
What it did was to weaken an otherwise strong argument that the RM
was decisive. The ARG was always a little reluctant to make interpretations
that clearly changed the wording and intent of the RM, so it makes it easier
to get agreement to a "change" if it is not so clear that it is a change!