Seems to be a built-in, extremely specific use case: https://github.com/llvm-mirror/llvm/blob/ae889d36724efb174e8...

I suppose it makes sense, given: 1) popcount is very common in bit-manipulating programs, 2) there is no standard popcount in ISO C, 3) it potentially saves 100+ instructions.

(Edit: I see a comment on Lemire's blog already points to that.)