RFR: 8320500: [vectorapi] RISC-V: Optimize vector math operations with SLEEF [v6]

[Fei Yang]

On Wed, 25 Sep 2024 20:10:41 GMT, Hamlin Li <mli at openjdk.org> wrote:

>> Good question!
>> Let me do some further investigation and get back later.
>
> I see sleef code only set frm to RNE, but I'm not quite sure.
> Even if we can make sure current sleef only set frm to RNE, seems to me we can not depends on current implement, it could change. Although good news is we don't update sleef regularly.
> Maybe we should take similar action as call-to-java and return-from-jni?

Just a bit worried about the fact that manipunating CSR could be very costly on RISC-V. Another choice would be adding an assertion about FP rounding mode expecting RNE when returning back from the SLEEF routine. I also checked floating-point intrinsics with `_rm` suffix in the function name in SLEEF src code and only witnessed use of `__RISCV_FRM_RNE`. I didn't see uses of other rounding modes as specified by the rvv-intrinsic-spec [1].

[1] https://github.com/riscv-non-isa/rvv-intrinsic-doc/blob/main/doc/rvv-intrinsic-spec.adoc

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/21083#discussion_r1776213231