hbs: yeah, we could shuffle the opcode assignments dynamically but it only requires a simple lookup table to deal with that

And opcode shuffling will be dealt with by VLIW compiler anyway, if it's a VLIW architecture CPU. If it's ARM based, opcode shuffling won't do anything

Is this CFROUND tweak a response to someone using it already? Or just discovered / requires a fork and +10% for ryzen 0.1% for intel is fair? Or do ryzens already have a huge advantage?

CFROUND is alreafy fast on Intel, but slow on Ryzen. I discovered it long time ago.