GCC quality (was Re: AMD-64 optimizations, some (new) code)
Richard B. Kreckel
kreckel at ginac.de
Tue Sep 27 22:48:23 CEST 2005
On 27 Sep 2005, Torbjorn Granlund wrote:
> I am not aware of having failed giving feedback. Please point me
> to the bug reports in question.
> I admit to not report all GCC bugs that I find. I work around
> all of them in GMP, typically in configure, thereby ensuring GMP
> works correctly with all compilers I test. Does this hurt GCC?
> Hardly, they have a very large number of bugs to work on already.
This entire discussion started off with the picture of wrong-code bugs in
our minds (the issue of the reliability of the speed program). And, no,
the number of *known* wrong-code bugs in GCC is not very large. Adding
some of these would be very much appreciated by the GCC community.
> The GCC quality problems aren't new. There have been periods
> where it has been better, and periods where it has been worse.
> One of my early contributions to the GCC project was the
> c-torture tests; before that there was no test suite at all and
> releases where not subject to any organized testing. Obviously,
> the releases back then varied in quality...
That effort cannot be underestimated. However, the times with no test
sute are long past.
> Don't get me wrong: If your motto is "pauca sed matura" you are entitled
> to call a product with four bugs crappy, all right.(*) But publicly
> whining about bugs that haven't surfaced except in an unpublished
> testsuite does have an odor of spreading FUD.
> You're just guessing here. And you're guessing wrong.
I'm not aware of having done any guessing here. Are you implying the bugs
you are talking about on this list are already known? Then, please, point
us to the corresponding bug reports.
> Also, I
> cannot believe that you seriously believe there are just 4 bugs
> in any GCC release.
I didn't say that. I said *finding* four bugs (as in per person).
> You seem to be aware of the bugs data base,
> and I am pretty sure it has thousands of open and closed bug
> reports for every release.
No. I'm pretty sure when it comes to wrong-code bugs in the C programming
language, that count is much, much less. And this entire discussion
clearly started off with bugs of the wrong-code category.
> You should try an experiment. Disable the various compiler bug
> worksounds from GMP 4.1.4 configure. Then try a couple of GCC
> releases on a coupld of platforms and see it blow. FUD?
Yes, FUD. At least until you define "blow". Internal compiler errors?
The harm that can be done by a compiler that fails to compile a valid
translation unit is not great. The harm that can be done by a compiler
that silently emits wrong code can hardly be underestimated. If you are
aware of such cases, by all means: don't delay and submit them!
Remember to be careful with what you say. There are other stakeholders in
GCC out there who become quite concerned when the quality of the generated
code is disputed.
Richard B. Kreckel
More information about the gmp-discuss