Clarification in the docs for possible bugs with the Alpha port

[Chris Devers <cdevers at pobox dot com>]

(Apologies if anyone gets this twice, accidently hit send with no subject
line & got a bounce back from <MAILER-DAEMON@sources.redhat.com>.  Oops.)


Hello everyone

I have been working with the FSF to copyedit the final draft of the _Using
GCC_ manual, and a small number of technial questions remain. I have been
working with Joseph Myers & Nathan Sidwell to resolve most of the
questions I've encountered, but there are still a couple of others that I
would like some clarification with.

The copy of the manual I'm working with is a little bit out of step with
the official one at the moment; patches will be submitted once I finish
putting things together. In the meantime, in case the material I am
describing here varies from the standard documentation, I have put a
snapshot of the manual on my home server at:

    http://devers.homeip.net:8080/gccbook/edited/gcc_chd.pdf

Here's the question.

In this version of the manual, at the end of section 10.3 on page 309, the
text reads as follows:

  * On the Alpha, you may get assembler errors about invalid syntax as a
    result of floating point constants. This is due to a bug in the C
    library functions ecvt, fcvt and gcvt. Given valid floating point
    numbers, they sometimes print 'NaN'.

Many of the bullet points in this chapter describe a general class of
problem such as this, and then offer a suggested workaround. No such
workaround is offered here. Is there a way the user can try to avoid
floating point numbers being interpreted as NaN on the Alpha, or are they
just stuck?

I brought this question up with Joseph & Nathan, and got these replies:


  On Thu, 31 Jul 2003, Joseph S. Myers wrote:

  > On Thu, 31 Jul 2003, Nathan Sidwell wrote:
  >
  > > [....] I don't know. I suspect it means 'valid but very small' (or
  > > possibly denormal -- a special type of very small number).
  >
  > I think this would only have been a problem when we used sometimes to
  > output floating point numbers as floating point in assembler.  At
  > least on mainline we determine the correct bytes to output and output
  > them as such; if that's the case on 3.3 branch as well, this point can
  > simply be removed as obsolete.


It was then suggested that I bring this up on the list, and cc: the
maintainer of the Alpha port, Richard Henderson. Here we are :)

Is Joseph's description of the situation reasonable, and if so can we go
with his recommendation to delete the warning (or perhaps prefix it with
"on versions earlier than 3.3, you may have...")? I'd like to make sure
this is okay before committing the change to the documentation.


Thank you for any advice you may be able to offer. There may be a few more
questions along these lines over the next few days, but this appears to be
the only one specific to the Alpha platform.



-- 
Chris Devers    cdevers@pobox.com

come from, n.
An instruction proposed by R. Lawrence Clark (1973) to resolve the GOTO
controversy. The industry is still (1995) bristling with acute
disbelief.

    -- from _The Computer Contradictionary_, Stan Kelly-Bootle, 1995