VFRCZ* are AMD XOP instructions.
That instruction set is only supported by bdver[1-4] cpu models.
These instructions are unrelated to the instructions grouped by the WriteFRnd sched class.
On bdver[1-3], they have latency of 10, as per agner's tables,
and as per my llvm-exegesis measurements for bdver2.
Which is different from what WriteFRnd specifies (5).
Now, the test coverage is saddening. We currently don't have any
dedicated sched models for any of these 4 cpu's, so the test coverage decreases.
I'm not sure this can be improved, until the said sched model is ready for integration...
This comment explains that SNB must support all models as its used as the generic model.