Investigate perf difference between multiple `SetAllVector128` and a `Load + Permute` sequence

[tannergooding]

As per the comment here: dotnet/corefx#31779 (comment)

I would expect that an explicit Load + four Permute operations (for four sequential memory addresses) would be faster than (or at least as fast as) four SetAllVector128 (which should be equivalent to four loads and four permutes).

Investigate the codegen between the two to see if there is some bug blocking this optimization.

[tannergooding]

CC. @fiigii, @eerhardt, @CarolEidt

Also CC. @EgorBo

[tannergooding]

I've self assigned and should get to this on Monday.

[EgorBo]

@tannergooding turns out I was benchmarking that test on some old 2.1 runtime, with the latest 3.0 Shuffle is 20% faster.
JIT output:
Multiply-SetAllVector.cs: https://gist.github.com/EgorBo/84ae4be8f4a8024548615e486e603962
Multiply-Shuffle.cs: https://gist.github.com/EgorBo/c44417038ad56e8ec145ea1945257a36
C#: https://gist.github.com/EgorBo/e2717606fdc4c3a14a7fa16617f87ad8

[tannergooding]

Thanks for the update @EgorBo!