Whoever implemented the RNM method, made it in an unefficient way for instructions count.
It converts normal to 0,1 then make some modifications and reconverts them to -1,1 which causes 2 more instuctions. Also it doesn’t use component masking which causes 1 more instructions.
https://s32.postimg.org/z8lnmp5tf/image.png
It should’ve implemented like this which causes 3 less instructions in base pass shader.
https://s32.postimg.org/uci0v05o3/image.png
The result is completely same as a proof.