------- Comment #31 from pcarlini at suse dot de 2007-05-12 20:26 -------
(In reply to comment #30)
> At the moment, I'm investigating a fix along the lines suggested in
> comment #26. The implementation of strtold in HP-UX 11.11 and earlier
> is broken in at least four ways:
>
> 1) Return of plus or minus _LDBL_MAX instead of HUGE_VALL if correct
> value would cause overflow,
> 2) locale support is worse than for sscanf,
> 3) it is listed by HP as not being thread safe,
> 4) it doesn't support "inf", "infinite" and "nan" conversions.
>
> As a result of 1, we would have to check errno if we are going to
> use strtold. This is likely not thread safe. Thus, it looks better
> to use sscanf. I checked. It returns 0 if the result would cause
> overflow. Is there any reason not to sscanf other than performance?
Indeed, I was also considering this kind of solution. Excellent.
> I sent mail to Steve Ellcey asking him to check the behavior of
> strtold in HP-UX 11.23 and 11.31. If there aren't good technical
> reasons to favor strtold over sscanf, I think we had best go for
> consistency on all versions of HP-UX. In that case, we wouldn't
> need the configure check.
Fine. Actually, it would have to be an AC_TRY_RUN (we are trying to keep those
to a minimum), and I'm just learning to you would have to be rather complex, to
be safe. Then, how would we deal with the issue, in practice?
_GLIBCXX_HAVE_STRTOLD would still be true on HP-UX, because the function is in
any case available at compile time. By way of the _GLIBCXX_HAVE_BROKEN_STRTOLD
which I mentioned before? Like:
#if defined(_GLIBCXX_HAVE_STRTOLD) && !defined(_GLIBCXX_HAVE_BROKEN_STRTOLD)
> I see that at one time we used _GLIBCXX_USE_C99 instead of
> _GLIBCXX_HAVE_STRTOLD. However, this caused a problem on mips.
I was also considering this option too ;) Too bad we had such kind of problem
on mips, because, in principle, it also looks like a conservative solution:
that macro results only from TRY_COMPILE tests, but very extensive. I would say
it's very-very unlikely that a target passes all the _GLIBCXX_USE_C99 test
still such basic functionality of the strto* functions is missing...
Anyway, Dave, since you are working hard on the issue (thanks!), I can't
consider myself responsible for fixing the problem, for better or worse, I'm
unassigning myself from the PR, still in CC and helping of course.
--
pcarlini at suse dot de changed:
What |Removed |Added
----------------------------------------------------------------------------
AssignedTo|pcarlini at suse dot de |unassigned at gcc dot gnu
| |dot org
Status|ASSIGNED |NEW
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31836
