[PATCH] D35081: [ThinLTO] Allow multiple summary entries.

[Teresa Johnson via Phabricator via cfe-commits]

tejohnson added a comment.

In https://reviews.llvm.org/D35081#808517, @mehdi_amini wrote:

> In https://reviews.llvm.org/D35081#808207, @tejohnson wrote:
>
> > My first option was if any copy is under the threshold, simply pick the prevailing copy (it may be over threshold, but assume we want that one anyway). Another possibility is to only allow importing of the prevailing copy, if it is over the inst limit, too bad. The main reason I think that could be better is when we have PGO - the prevailing copy would for sure have the matched PGO data, but the non-prevailing copies may not.
>
>
> Ah makes sense, both way are fine with me.
>
> I'd rather limit to the prevailing copy though and too bad if it is over the limit, to avoid the possibility of too large importing just because there was a small non-prevailing somewhere, but likely a rare case.


SGTM. That's also very simple to implement.  I've just been to swamped with other things so far to do it. Just chatted with Yunlian, he is not blocked.

> Limiting to the prevailing copy means that we could prune the index early after symbol resolution to keep only the prevailing summary, which could speed-up all the process (and possibly get more accurate dead-stripping if we don't use the prevailing information there already?).

I think we want to keep the other copies so we can mark their new linkage types (e.g. demoted to available_externally) for use in the backends. Dead stripping should use prevailing info, think there is a todo there.


https://reviews.llvm.org/D35081