RS FNT: improve performance in NonSIMD mode

[lamphamsy]

The purpose is to let compiler optimize/vectorize operations
It's expected to obtain a perf comparable to that in SIMD mode