[PATCH] D139644: [InlineAdvisor] Allow loading advisors as plugins

Fri Dec 23 10:21:05 PST 2022

IBricchi added a comment.

In D139644#4015314 <https://reviews.llvm.org/D139644#4015314>, @thakis wrote:

> In D139644#4003692 <https://reviews.llvm.org/D139644#4003692>, @thakis wrote:
>
>>> Adds the ability to load InlineAdvisors as plugins. This allows developing and distributing inlining heuristics outside of tree.
>>
>> By the way, was there an RFC somewhere that discussed why this is a good thing for the project? (I have no opinion on this myself, but it clearly comes with a cost, so it'd be good if people working in this domain overall agreed that this is a good direction.)
>
> BTW, was this ever answered?

Usefulness: In our group (and this  applies to others as well) we have been working on inlining for a while. However our work is experimental/research, and it makes little sense to try to upstream it directly on LLVM. his creates a barrier for sharing it with the community so that it can be used and evaluated.  LLVM already had the correct abstraction to solve our problem, the InliningAdvisor; all that was missing was what this patch to enable the ability to externally provide one. This is greatly reduces the barrier for sharing inlining (heuristic) related work, not only for our group, but anyone who's working on that.

Cost: the only runtime cost is a boolean check, whether a PluginAdvisor has been registered, which should not even be measurable.

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D139644/new/

https://reviews.llvm.org/D139644