[Patch debug] Do not emit debug sections until they are required.

[IainS]

Hi Richard,
On 6 Nov 2010, at 00:08, Richard Henderson wrote:

> On 11/03/2010 06:33 AM, IainS wrote:
>>> This is the only hunk that concerns me.  I don't believe that this
>>> is valid dwarf.  In particular, a non-empty pubnames implies that
>>> we must have had an info section from which to pull the names.
>>>
>>> Did you have a test case that caused you to make this change?
>>> If not, ok with gcc_assert (info_section_emitted).
>>
>> not a test-case -- but Java achieves it ...
>> ... and with the gcc_assert (), fails to bootstrap on i686-darwin.
>
> Ok, got it.

:-)

> We generate pubtypes before -feliminate-unused-debug-types does its
> thing.  So we're putting stuff in pubtypes that are no longer in the
> compilation unit.  Which is a Bad Thing.  I was right about us
> emitting invalid dwarf2.

it looks like we have possibly only been doing it when the table was  
empty...

> We should generate pubnames/pubtypes from the final debug_info tree,
> rather than trying to emit them earlier.

That looks quite involved (but maybe I miss something).

==

It seems that checking the the Vec has size > 0  is not enough for  
pubtypes (because of the possible elimination).

Would it be sufficient (although another pass through the table :( )  
to check that at least one type will be emitted before trying?

==

Also I note that the

   size_of_pubnames()
   output_pubnames ()

routines check p->die->die_offset != 0 as the condition.

I wonder if that is not supposed to be p->die->die_mark?

thanks
Iain