[llvm] r207719 - Revert "Emit DW_AT_object_pointer once, on the declaration, for each function."

[Eric Christopher]

On Thu, May 1, 2014 at 11:22 AM, Robinson, Paul
<Paul_Robinson at playstation.sony.com> wrote:
>> -----Original Message-----
>> From: Eric Christopher [mailto:echristo at gmail.com]
>> On Thu, May 1, 2014 at 10:27 AM, David Blaikie <dblaikie at gmail.com>
>> wrote:
>> > On Thu, May 1, 2014 at 8:50 AM, Robinson, Paul
>> > <Paul_Robinson at playstation.sony.com> wrote:
>> >>> -----Original Message-----
>> >>> From: David Blaikie [mailto:dblaikie at gmail.com]
>> >>> On Wed, Apr 30, 2014 at 7:40 PM, Robinson, Paul
>> >>> <Paul_Robinson at playstation.sony.com> wrote:
>> >>> > Yeah, I was wondering.  Attributes like object_pointer seem more
>> >>> > appropriate to a definition than a declaration, as the declaration
>> >>> > typically wouldn't provide the actual location.
>> >>>
>> >>> I'm not sure the actual location is a prerequisite for making use of
>> >>> object_pointer - I figured callers might like to know about it (but
>> I
>> >>> suppose you can't /call/ a function from just the DWARF declaration?
>> >>> Never really thought about whether debuggers would manage that -
>> they
>> >>> could /maybe/ build the mangled name, etc). Useful for rendering
>> >>> functions in pretty printers (well, the first parameter is flagged
>> as
>> >>> artifiical in the declaration, perhaps that's sufficient for
>> printing
>> >>> purposes - another case where it'd be nice if we didn't have to
>> >>> duplicate that information in the definition, though).
>> >>>
>> >>> But, sure, it'd be great if there was no reason to put
>> object_pointer
>> >>> on declarations - not insignificant space savings when emitting
>> types
>> >>> all over the place.
>> >>>
>> >>> - David
>> >>>
>> >>> > --paulr
>> >>
>> >> If you have a definition, then the location is definitely relevant;
>> >> a location expression might want to push the object address, and for
>> >> that you need to know where the object is!
>> >
>> > I'm not suggesting it's not helpful at the definition, just that it
>> > shouldn't need to be duplicated there.
>> >
>> >> But if you're inheriting
>> >> the object_pointer from the declaration, which only points to the
>> >> declaration's formal_parameter, which has no location... not very
>> >> helpful.
>> >
>> > I don't know why an implementation wouldn't derive it from the
>> > declaration - they must have the same parameters in the same order,
>> > right?
>> >
>>
>> Exactly.
>
> The declaration and definition have distinct sets of DIEs that describe
> the same parameters in the same order; that by itself *intuitively*
> establishes a correspondence, but there's no direct connection from a
> DWARF perspective.
>

This is why we have abstract origin though yes? :)

That said, it seems that uses of abstract origin to say "hey, all of
my children are roughly the same as well" has been called out as
verboten since I last looked.

Shame that.

-eric

>>
>> >> That's why I'd expect the definition to have one, and if
>> >> the definition has one, then having one on the declaration is not
>> >> necessary.
>> >>
>> >> If you have a declaration with no definition... I suppose it would
>> >> be "for completeness" to have object_pointer point to the 'this'
>> >> parameter (in C++) although some debuggers I've worked with are
>> >> happy to infer that from the first formal_parameter being named
>> >> 'this' and marked artificial.
>> >
>> > Right - and I assume they'd be able to infer that in the definition
>> > too. Evidently GCC doesn't infer it in the definition, nor derive it
>> > from the declaration.
>>
>> They should be able to, but probably hasn't come up and so it hasn't
>> been thought about.
>>
>> -eric
>
> The definition's formal_parameter has a location attribute, but the
> declaration's object_pointer points to the declaration's formal_parameter
> and so won't find the location attribute.  Unless you persuade your
> debugger to troll for the corresponding formal_parameter DIEs in the
> definition, after finding the object_pointer link to the declaration's
> formal_parameter...  This degree of indirect association is not very
> DWARF-ish.  The appropriate solution is just to put the object_pointer
> on the definition, and have it directly point to the formal_parameter
> that has the location.  The size cost is small and the links all do
> the right thing.
> --paulr
>