Performance regression between 3/1 and 3/2

[ishitatsuyuki]

Per #47861 (comment).

[nikomatsakis]

This was caused by #47861, it seems. cc @sgrif. Rather surprising since we did measurements beforehand though!

[nikomatsakis]

I'm inclined to roll it back and try landing again after doing more investigation, though I'd not be opposed to doing some quick measurements to see if we can identify the problem.

[sgrif]

ccd92c2 was meant to be removed from that PR, but somehow I accidentally re-added it. I'll see if I can cleanly just revert that one commit.