[llvm-branch-commits] AlwaysInliner: A new inlining algorithm to interleave alloca promotion with inlines. (PR #145613)

[Amara Emerson via llvm-branch-commits]

aemerson wrote:

> > Final ping. Does anyone have _objections_ to this patch?
> 
> It seems very odd to tackle this in the always inliner, as it adds a new concern to a fairly scoped pass. Are you sure there are no alternatives, for instance, having a pre-AlwaysInliner pass to mark off callsites that shouldn't be inlined, and then re-enabling them afterwards?
> 
> Maybe a RFC would help gather more feedback and consensus?

I can see where you're coming from, we'd ideally like to not couple optimizations together. I still don't see an alternative that's any more elegant. The root cause is that we don't have an interleaved inline/optimize pattern like the main inliner.

I don't think an RFC is going to yield any more than the reviewers here, but I can re-consider in future if we see more cases of this issue occur.

https://github.com/llvm/llvm-project/pull/145613