Given that this introduces a genuinely new class of optimizations, I think the size of the regressions is acceptable.
Something I was surprised about is that we don’t seem to see much in terms of benefits from the pass – usually we would see some minor improvements on check builds (due to elimination of bounds checks in the bootstrapped rustc compiler). Possibly this is less useful for rust code than one might naively think (e.g. due to prevalence of iterators rather than manual indexing).