Is linkonce underused?

[Matt Austern]

On Tuesday, December 17, 2002, at 02:06  PM, H. J. Lu wrote:

> On Tue, Dec 17, 2002 at 01:34:50PM -0800, Matt Austern wrote:
>> Consider the following C++ source file:
>>   void check(const void*);
>>   template <class T> T* identity(T* p) { check(p); return p; }
>>   void foo() { identity((double*) 0); }
>>
>> Using the top of the 3.3 release branch to compile this
>> on Linus, and using the options -O0 -g -S, it's easy to
>> see what gets placed in which sections.
>>
>> _Z8identityIdEPt_S1_ gets placed in a linkonce section,
>> as it ought to be.  However, it's the only thing in that
>> linkonce section.  This is puzzling, because that
>> template instantiation has two other parts: debug
>> information, and EH information.  One would think that
>> these should all be retained or discarded as a group.
>> It's nice that we'll only have one copy of the actual
>> code for identity<double> even if it's instantiated in
>> a hundred translation units, but it's a shame that we'll
>> still get a hundred copies of the EH information and the
>> debug information.
>>
>> Am I misunderstanding something?  If not, is there some
>> reason we aren't using linkonce for EH and debug
>> information?
>
> GRP_COMDAT can be used for that. It is available in ELF. I'd like to
> see gcc support it.

Has anyone done any preliminary design work on this?  Apple
is doing some work that's relevant to it, and we'd hate to
duplicate anyone else's effort.

			--Matt