Nice work. I suspect a lot of the time in mapping tests written in pike is spent outside the hashing and bucket lookup functions themselves. I.e. a total increase of 10% implies that the core parts are quite a lot faster.
That's just my guess though. To really make a good comparison one would have to write tests in C, but that's not necessary to convince me. I think your patch looks good for 7.9. Just a couple of notes: find_next_power maybe fits better in stuff.c, and the new bit fiddling at the end of hash_svalue could use a comment on the origin and rationale behind it.
