Now that the option has probably been extensively tested, it would be nice to enable computed gotos by default on systems that support them.Perhaps this needs a dedicated test in the configure script.

Antoine Pitrou wrote:> > New submission from Antoine Pitrou <pitrou@free.fr>:> > Now that the option has probably been extensively tested, it would be nice to enable computed gotos by default on systems that support them.> Perhaps this needs a dedicated test in the configure script.I doubt that it has been tested enough, but enabling itfor those platform/compiler where we know it works, should befine.Have you tried testing this using the buildbots ?

Marc-Andre Lemburg wrote:> > Marc-Andre Lemburg <mal@egenix.com> added the comment:> > Antoine Pitrou wrote:>>>> New submission from Antoine Pitrou <pitrou@free.fr>:>>>> Now that the option has probably been extensively tested, it would be nice to enable computed gotos by default on systems that support them.>> Perhaps this needs a dedicated test in the configure script.> > I doubt that it has been tested enough, but enabling it> for those platform/compiler where we know it works, should be> fine.Here's a compiler/platform combination where it should not be usedper default:gcc 4.4.2 on AMD processors ...http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42621The fix will go into gcc 4.5, the report doesn't say whether it wasalso backported to 4.4.x.

> gcc 4.4.2 on AMD processors ...>http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42621So? Compilers have all kinds of optimization bugs, it's not our job tohave a compiler matrix with optimal flags per compiler.(well, you can of course maintain such a matrix if you want)For the record, I had an Athlon 64 machine until very recently, used itwith gcc 4.4.x and computed gotos were better than without.

Antoine Pitrou wrote:> > Antoine Pitrou <pitrou@free.fr> added the comment:> >> gcc 4.4.2 on AMD processors ...>>http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42621> > So? Compilers have all kinds of optimization bugs, it's not our job to> have a compiler matrix with optimal flags per compiler.No, but we do need to make sure that the casual user does notrun into such issues by using the default compiler on a typicalUnix system with Python default settings.An expert user can always use the right compiler... but then anexpert can also use the configure option :-)gcc 4.4.2 is not exactly an old version of gcc. The perspectiveof (apparently) not having the patch in in Ubuntu 10.4 LTS'sdefault compiler suggests that we do need be more careful aboutenabling the switch per default.> (well, you can of course maintain such a matrix if you want)> > For the record, I had an Athlon 64 machine until very recently, used it> with gcc 4.4.x and computed gotos were better than without.Perhaps someone can report whether this affects Ubuntu 10.4 LTS ?!

> No, but we do need to make sure that the casual user does not> run into such issues by using the default compiler on a typical> Unix system with Python default settings.The point is that we cannot make it sure. There is nothing specificabout computed gotos here: any kind of compiler bug could threatenPython performance without us being able to do anything about it. Evenif we restrict ourselves to standard C features.(also, I'll point out that it's not demonstrated that the problem aboveaffects Python, rather than simply the particular piece of code shown inthe bugzilla report)

> No, but we do need to make sure that the casual user does not> run into such issues by using the default compiler on a typical> Unix system with Python default settings.I think there is no risk here. The casual Python user on an AMD64 systemwhere gcc is the default compiler will use the Python installation thatcomes with the operating system.

Martin v. Löwis wrote:> > Martin v. Löwis <martin@v.loewis.de> added the comment:> >> No, but we do need to make sure that the casual user does not>> run into such issues by using the default compiler on a typical>> Unix system with Python default settings.> > I think there is no risk here. The casual Python user on an AMD64 system> where gcc is the default compiler will use the Python installation that> comes with the operating system.I'd be fine with enabling the switch on those platforms as well,if we could get feedback from someone testing the behavior of gccon an AMD platform running Ubuntu 10.4.If the problem is really only with the code snippet listed in thebug report, there shouldn't be a problem and we can go ahead withenabling computed gotos on that platform/compiler combination as well.

I've committed a patch inr83986. I will watch the buildbots and close if everything's fine.