Hacker News Reader

Performance aside, is it feasible for those extensions to trap and emulate in the kernel? Like old ARM Softfloat or linux/arch/x86/math-emu ?

Sure of course, but that's going to cost something like 500 cycles [1] to emulate something that only exists to save 2 or 3 cycles.

Also, not in the kernel but in SBI -- in Machine mode not Supervisor mode.

[1] estimate based on how long it takes to trap and emulate misaligned accesses on VisionFive 2.