This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Why alias set 0 ?
- To: Richard Kenner <kenner at vlsi1 dot ultra dot nyu dot edu>
- Subject: Re: Why alias set 0 ?
- From: Paolo Carlini <pcarlini at unitus dot it>
- Date: Tue, 30 Oct 2001 17:40:44 +0100
- CC: gcc at gcc dot gnu dot org, mark at codesourcery dot com, jh at suse dot cz
- References: <10110301636.AA07310@vlsi1.ultra.nyu.edu>
Richard Kenner wrote:
> 2- Besides any interesting problem with the C++ aliasing treatment,
> which I would like to understand but I can't, unfortunately, right
> now, my question is: *why* no more than 10 days ago (in some utterly
> misterious way) the compiler managed to deal much better with that
> testcase????
>
> It had a bug in that it sometimes used an alias set of a component even
> when the outside record was alias set 0.
Oh!
Thanks. Now, finally, everything is clear.
Very sad situation, anyway... It looks like in two different tests of that
renowned testsuite, first "Complex Matrix" and now "Real Matrix", the
compiler performed much better than now by pure (unsafe) chance... Very,
very sad situation...
Anyone volunteering for improving C++ alias analysis out there???? :-)
Some time ago a read a beautiful popularization article by Mark Mitchell
(http://www.ddj.com/articles/2000/0010/0010d/0010d.htm) where he explained
very clearly that the optmizer of GNU's C++, thanks to type based alias
analysis may in fact *beat* that of other languages (such as C) , usually
considered more efficient... Now, if I understand well, it seems that some
steps are still to be taken in order to fully accomplish the project, right?
Cheers,
Paolo Carlini.