More correct at the LLVM level?

[ChrisRackauckas]

Does this implementation give false positives? I would think you would need to run this after optimizations, such as DCE, in order for this to not have false positives?

[MilesCranmer]

It relies on Base.promote_op to do the inference. So whatever optimizations are used for Base.promote_op will apply here.

I haven't found a false positive but I am sure they exist. For example I ran into #27 but it turned out to be a real typeof instability due to how Core.kwcall works.