Re: About gcc builtin atomics
On Mon, 25 Jun 2012 22:14:03 +0000 (UTC), Thorsten Glaser <tg@mirbsd.de> wrote:
> Mikael Pettersson dixit:
>
> >Now we know that aranym had a broken
> >CAS emulation :-(
>
> Those problems are gone now though?
I hope so. I've updated my previous sync implementations for m68k
(both the linux-atomic.c one and the sync.md one) and I'm testing
them again with aranym-0.9.13 + a few backports from cvs head.
> >I've always relied on gcc's c, c++, and libgomp test suites when
> >testing my attempts to implement atomics in gcc.
>
> *sigh* I guess I=E2=80=99ll port the patches to doko=E2=80=99s latest uploa=
> d and
> build that withOUT nocheck, but the results won=E2=80=99t be in time for
> the freeze then. libdrm and mesa built from unchanged sources now,
> though. Should I submit the patches to doko right now, or do we
> want to chance me maintaining gcc-4.6 in unreleased until after
> the unfreeze?
>
> What, in particular, should I look out for in the testsuite?
> (I imagine that many things would be broken, independent of
> this change.)
Many things are indeed failing in the test suite on m68k. What you
should do is to make two full bootstraps and regression test runs,
first with a baseline version (I use FSF vanilla, you might want to
use Debian's common gcc without your patch), and then with the
baseline plus your patch. Then you 'diff' the test suite summary
files from the two runs and analyze any new failures.
> Maybe I should just mail the entire build log
> to the list? (Unless someone can figure out how I can upload
> it to debian-ports.org=E2=80=A6 and I whether my MTA can deal with
> that =E2=80=93 probably on a good-connected machine.)
Build logs aren't interesting unless they show build failures.
Proper test suite results are produced as follows:
0. prep sources, mkdir objdir, cd objdir, /path/to/source/configure ...
as usual for any gcc build
1. make bootstrap
2. make -k check (don't forget the -k!)
3. make mail-report.log
The last step scans the detailed test suite logs and produces a
summary that can be diffed against a previous one, or mailed to
gcc-testresults (but don't do that for a heavily-modified vendor
gcc, perhaps Debian has its own place for posting test suite
results?)
On the Core i7-970 (3.2GHz, 6 cores, Westmere) that hosts my
aranym VMs, a VM can do a gcc 4.6 bootstrap plus test suite run
in slightly over 2.5 days, but that's with --enable-languages=c,c++
and with two patches to reduce 'genattrtab' overheads; with all
languages and without the genattrtab patches it might take 4-5 days.
/Mikael
Reply to: