_mm_movemask_epi8 regression

[Jackson Newhouse]

The aarch64 code path for _mm_movemask_epi8 introduced in #50 looks to be a regression when you actually compile it. The default behavior compiles to 7 instructions with no constants, while the "fast path" is 14 instructions plus a constant. Should it be reverted?

fast path: https://godbolt.org/z/41s54d
default: https://godbolt.org/z/xsYfz8

[marktwtn]

I'll make a simple time experiment of _mm_movemask_epi8.

[marktwtn]

I do the experiment on the ARMv8-A CPU, which is an ARM 64-bit architecture with optimization level 0.

It turns out that the aarch64 code path does behave worse.
We should revert it for performance consideration.

[marktwtn]

The performance of optimization level 3 is measured as well.

The performance does not have too much difference.
@jserv I think we need to decide which optimization level we should focus on for the future performance improvement.