Re: Getting GNAT to issue ARM error messages

[steve quinlan <steven.quinlan@lmco.com>]

Actually, the "statically matching" error message you gave was the one in
GNAT at the time I was doing the trade study. My questions to you on that
point   led to your changing that message, I believe, as well as to the
exchange we had at the time about RM references in error messages. Do you
recall this?  Is your new message better? Yes. I don't maintain that ONLY
RM sections should be given in error messages. Take your best shot at
giving the user a way to fix the problem. Then, in parens, you could give
the RM reference for non-trivial errors.

The problem with adding such references onesy-twosy as you find them to be
useful is that no one is using the FUTURE version of the compiler, where
that reference will be added after you've decided it would be useful to add
it. It's always today's version, where that reference is not there, which
is in use.

Where I miss such references most is not when I've made some stupid syntax
error, or a trivial semantic error, but when I've violated some limitation
of Ada, as in the "statically matching" example. Certainly I like to be
told what to do to fix my error, but I also would like to understand why I
can't do what I thought I could. Maybe there are related issues to  that
particular error that a  reading of the RM on that topic will raise. The
rule about statically matching subtypes is such an arcane rule. It's so
bizare, that besides upgrading your text error message, I would have argued
for putting in an RM reference as well on that one.  As long as you were in
there changing the message, what could it have hurt to put "RM 3.10.2(27)"
in parentheses?

I was taken to task by several for claiming that the overwhelming number of
users would want full, or nearly full, RM references. Of course, I have no
survey to go on. I would love to take one. Anyone who wants to send me
their opinion can do so. The question would be : "would you like to see RM
section references in compiler error messages (in addition to the BEST,
MOST INSTRUCTIVE text message) for non-trivial errors? Trivial would be
syntax errors and semantic errors that are almost of the same class as dumb
syntax mistakes -- trying to assign a string variable to an  integer, stuff
like that. I will tablulate and post the results. Of course, it would be an
unscientific, self-selecting sample so it will have only an entertainment
value, right Robert?  (;-)