Performance regression in #master

[ufechner7]

I see a performance regression in #master of 12 to 15%. Part of it is probably due to the new implementation of norm, but that cannot be the only reason for this regression.

For my benchmarks in KiteModels.jl I see:

StaticArrays 1.5.12   540ns (3point model) and 1256 ns (4point model)
StaticArrays master   621ns                and 1394 ns

I use the following setting in my code:

set_zero_subnormals(true)           # required to avoid drastic slow down on Intel CPUs when numbers become very small

[hyrodium]

This issue has been well discussed in #975 (comment) and there seems to be nothing new to discuss (at least there were no responses for a week).
Feel free to comment on this issue if you have any additional comments.