Re: [commit] Testsuite updates

[Daniel Jacobowitz <drow at false dot org>]

Sorry for batching updates like that; I'll be more verbose.

On Sun, Dec 31, 2006 at 09:35:00PM +0100, Mark Kettenis wrote:
> > I was using a HEAD build of glibc rather than the system's glibc,
> > in order to pick up Jan's fix for unwinding information in signal
> > handlers.  Having debugging information for the C library caused the
> > annota1.exp and annota3.exp failures.
> 
> So continueing to printf didn't actually get is to printf?  What's
> going on here?

We got to printf just fine.  But, printf is really the exported name of
__printf.  If we have only the dynamic symbol table, backtrace will
display printf; if we have debugging info, it will display __printf.
There was already some accomodation for _printf in the tests.

I think this is the work of strong_alias or attribute((alias)).

> > The selftest.exp patch adds an XFAIL for a GCC bug in my system
> > compiler, which I reported several months ago to the GCC bugzilla.
> > I'm not optimistic about it getting fixed soon.
> 
> A bug in *your* system compiler.  How widespread is this bug?  We
> really should be careful with complicating tests to work around
> problems in particular GCC versions.

It's in every version of GCC after 4.0, for every platform.  It's a
pretty serious problem, too.  Here's more information, if you're
interested:

  http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26475

We have an initialization of some variable to a constant.  The internal
representation does not keep track of the location of constants, so
when it is optimized away on entry to SSA, there's no way to recover
the location.  And when it recreates the constant, it picks a most
peculiar location for it (based on where the PHI nodes are).

> > And finally, for linux-dp.exp, an omission in Nathan's recent change
> > generated UNSUPPORTEDs on native GNU/Linux systems, where the test
> > should pass.
> 
> Can you elaborate?  I mean, what output is genererated that should be
> ignored?

Output from the test program, rather than GDB.  I believe Nathan was
working on our ColdFire toolchains at the time, which means he was
testing gdbserver - in which case the console output of linux-dp
doesn't show up.  But in a native configuration it's quite a noisy
program.  The pattern matches the status of the philosophers for each
thread.

-- 
Daniel Jacobowitz
CodeSourcery