[PATCH] Add new debug kind LocTrackingOnly.

[David Blaikie]

On Mon, Jun 23, 2014 at 2:02 PM, Eric Christopher <echristo at gmail.com> wrote:
>>> Hm, perhaps. I would be penalizing -Rpass, however. But I guess it doesn't
>>> matter much. I'll amend the documentation to state that -gcolumn-inf is
>>> needed with -Rpass.
>>
>> Well it's not worse - currently -Rpass doesn't enable debug info at
>> all , but I suppose it at least suggests the user enable both -gmlt
>> and -gcolumn-info. But I assume if they enabled -g(mlt or otherwise)
>> they wouldn't be told they should use -gcolumn-info.
>>
>> So, committing this patch without the column info would be sort of a
>> regression, but sort of not... Given this is a developer feature only,
>> I'm OK with it being a little rough around the edges. Up to you
>> though, we can discuss the issues with -gcolumn-info and -Rpass in
>> this review if you prefer. Just figured it'd be better to separate
>> them if practical.
>>
>
> Separately from the rest of the patch we might want to look into
> turning column info back on for -gmlt in particular. It's the use case
> for the sanitizers etc and so having slightly more accurate back
> traces is a fairly good thing there. Especially as we get the size of
> -gmlt down we can look into using that room for things that are
> explicitly useful.
>
> Thoughts?

Sure, sounds plausible. I have this vague recollection that one of the
reasons it was turned off was performance implications in clang,
querying for the column info for everything wasn't terribly cheap -
but I'm probably misremembering. Maybe it's just creating the extra
MDNodes that gets costly.

In any case, whatever the cost was, if any (maybe it was just "GDB
gets weird when you add column info" - in which case that's less of
(not at all?) a problem for -gmlt).

- David