This is inspired by
https://github.com/llvm/llvm-project/pull/77342#pullrequestreview-1814673242,
and is split off of same with some differences in style.
A select is a vmerge.vv with the additional cost of materializing the
bitmask vector in a vreg. All masks fit within a single vector register
(e8 + m8 is the worst case), and thus our worst case cost should be
roughly 3 (2 scalar to produce the address, one vector load op). Given
most shuffles are small, and the mask will be instead produced by
LUI/ADDI + vmv.s.x or ADDI + vmv.s.x, using 2 as the default seems quite
reasonable. At worst, we're not going to be off by much.
The prior lowering scaled the cost of the bitmask with LMUL, which I
don't understand. At m1 it did use the same base cost of 2. (@lukel97
You wrote the original code here, anything I'm missing here?)
2663d2cb9c