Trim duplicate CSR swaps from user exceptions.

[jhb]

Summary

The stack pointer is swapped with the sscratch CSR just before the
jump to cpu_exception_handler_user where the first instruction swaps
it again. The two swaps together are a no-op, but the csr swap
instructions can be expensive (e.g. on Bluespec RISC-V cores csr swap
instructions force a full pipeline stall).

Test Plan

• booted in qemu

[br]

nice catch!

[mhorne]

sys/riscv/riscv/exception.S

209 ↗

(On Diff #67394)

So we check that the sscratch register is empty as a means of determining which mode we came from?

I think this would be better as a check of the SPP bit in sstatus, because it looks like you could eliminate another double sscratch swap in the supervisor case.

[jhb]

sys/riscv/riscv/exception.S

209 ↗

(On Diff #67394)

Yes, checking sscratch in this way is pretty typical. You can't check sstatus because you haven't yet saved any GPRs anywhere, so you can't do a 'csrr` to get the value of sstatus and check it. The scheme is that the kernel clears sscratch so that nested kernel exceptions see a 0 value. bbl does the same thing with mscratch FWIW, and I suspect it's probably what Linux does as well. There aren't really any other options when you only have the single scratch register.