As per https://github.com/dotnet/runtime/pull/53578/files#r645282551, we are currently emitting a pslddq followed by a psrldq to zero the 4th element.

This is decent codegen for the SSE/SSE2 baseline but is inefficient compared to other patterns we could generate for modern hardware.

We should likely replace this with the relevant logic from vector.WithElement(3, 0.0f), which can then generate a insertps xmm, xmm, 0b00_00_1000 which will preserve all existing values in xmm and zero the third element.

category:cq
theme:codegen
skill-level:expert
cost:small
impact:small