This isn't really a bugfix, right? I can see the argument for backporting to 1.12 (since that's the current release and when Array became alloc-opt eligible), but it's unclear to me why we would backport further.

Is this expected to yield a performance improvement?

so this is a bugfix since dropping metadata can cause issues (both performance and otherwise). Julia, for example already uses tbaa extensively to describe loads being invariant within a module. Admittedly this is less important for alloca than load/store's, but still an issue.

This is probably not legal. The metadata says it's a heap object when at this point it no longer is one. Also it's an alloca it's probably quite easy to see that it doesn't alias other things most of the time
This doesn't actually copy any of the metadata I thought it would (turns out gc alloc doesn't really have much metadata)
LGTM