Re: g77 -ieee crash in 1.0.3a

[Craig Burley <burley at gnu dot org>]

>>>>>> "Craig" == Craig Burley <burley@gnu.org> writes:
>
> Craig> In particular, -mieee should be one of the options that gets
> Craig> exercised for *all* tests on Alphas, at least for the g77
> Craig> tests (but why stop there?!).  The commentary already mentions
> Craig> it, so that's taken care of.
>
>What commentary?  I can't find anything to that effect in the
>testsuite, if that's what you mean.  The comments in what you sent
>don't make sense to me as they are.  It seems to be referring to an
>article I haven't seen (or maybe it's expired).

Hmm, maybe we're talking about different things.  Here is the case
I thought we were talking about, as submitted to this list by me:

* I get the same error message using a 19980726-CVS'd egcs-1.1  
* prerelease snapshot, using the following, somewhat simpler example:
*
      subroutine a(b,c)
      b = max(b,c)
      end
*
* The only important option in the list above is -mieee:  With it,  
* you'll get the error, without it there's no problem (and the  
* expected floating conditional move is used; probably the above insn  
* also tries that, but perhaps it gets confused by the float_extends -  
* I've no idea why they're necessary, because all variables are  
* single precision).

I have no problems with your editing the commentary to better
highlight the importance of -mieee if you like -- I only very
quickly edited the example when I submitted it.

> Craig> And if we have to simply stuff a .x file in for it someday to
> Craig> just take care of that case, we can do that, but I'd rather
> Craig> get the test in to the mainline (just so I can stop worrying
> Craig> about it) and deal with -mieee later.
>
>I'm confused.  Is -mieee not necessary to get it to fail, then?  Is
>someone else dealing with the test stuff, i.e. _I_ don't have to worry
>about it, for which I'd certainly be grateful?

-mieee *is*, I assume, necessary to get it to fail.  What I mean is
that -mieee should be *automatically* specified in a set of test
runs on the g77 test suite on machines that support that option --
just like it spins through -O0, -O1, -Os, and so on.  So this case
should not be the only one to be compiled and perhaps executed with
-mieee in effect -- they *all* should, on an Alpha for example,
along with the current tests (that is, no -mieee).

I don't know whether this can be done, though.  In the meantime, at
least plugging in the test case itself is better than not doing so;
we can add a .x later as we learn how to do it so it adds -mieee
when suitable (and maybe doesn't even actually compile the code
otherwise, though it's so tiny that shouldn't matter).

> Craig> even if it doesn't, -mieee surely should do nothing worse to
> Craig> any code than to slow it down.  Crashing isn't nice;
> Craig> especially on a three-line test case.  :)
>
>Sorry, I thought I'd be credited with realizing that.  It just sounded
>as though background info could be usefully noted for the future,
>especially as it was being assumed we should have tested it like that.

I kind of made my statement "for the record", since you and I aren't
the only ones in this conversation....

And, IMO, Toon's three-liner seems suitably short and sweet that,
combined with the explicit note about -mieee being needed to trigger
any bug we know about, there's no need for further explanation.
But if you want to say more in the commentary, feel free!  (Long
rambling notes about how we should all be using Cobol will likely
be edited post haste, however.  :)

        tq vm, (burley)