Re: [RFA 2/4] dwarf2_physname

[Daniel Jacobowitz <drow@false.org>]

On Mon, Nov 23, 2009 at 08:51:16AM -0800, Keith Seitz wrote:
> On 11/20/2009 02:09 PM, Daniel Jacobowitz wrote:
> 
> >I am generally opposed to committing known regressions.  If there are
> >supporting patches we need to get in first, let's do that; if there
> >are tests we decide to break, let's XFAIL or KFAIL them.  That's the
> >only way we can make the testsuite more useful.
> 
> Sami has a follow-on patch that he could submit to fix all of these
> tests (they all pass on archer-keiths-expr-cumulative). Perhaps it
> would be acceptable for Sami to submit that patchset when/if this
> patch is accepted? [His patches rely on this patchset.]

If it applies on top of this, could he post it now?  Then we can treat
them as a unit for review and testing purposes.

> >The related regressions, just with arm-none-eabi GCC and the default
> >multilib:
> 
> I'm building an arm-elf toolchain now, and I will run it through testing.

Might need to be arm-eabi for some cases, I'm not sure.  I haven't
tried arm-elf in a while.

> >I'm going to skip the RealView regressions; those I'm willing to
> >handle in followups.  Most template tests still fail with RealView.
> >Some tests from namespace.exp improved, others regressed.
> 
> I think we discussed it earlier, but to be clear: RealView will
> demonstrate problems with templates, since it relies on
> DW_TAG_template_{value,type}_parameter, which gdb does not yet
> understand. [I've played with it a bit, but value parameters are
> *tough* on dwarf2_physname.]

Right.  I figured this out after the fact.

Did I send you my local implementation of
DW_TAG_template_value_parameter?  It's... not pretty.

> >Any idea what the above failures might be?  I can send you logs
> >offlist, they're large.
> 
> Yes, send the to me at this account.

Will do.  Rerunning with cpexprs.exp included this time.

> If we can define a suitable test procedure for this, I would be happy
> to produce some comparisons.

I'd like to leave Tom's work - exciting as it is - out of the
discussion for the moment.  I suggest picking a couple of C++
programs, preferably with different degrees of template-ness.

The interesting numbers are, IMO:

* psymtab creation ("time gdb -batch foo").
* full symbol creation ("time gdb -batch -readnow foo").
* some vaguely realistic operation, e.g. set a breakpoint near the
start of the application and run it through loading shared libraries
and print a backtrace.

-readnow numbers are interesting, but not vital.  That is, a big
slowdown there is not necessarily a big problem.

Thoughts?

-- 
Daniel Jacobowitz
CodeSourcery