Re: [PATCH v1 1/2] x86: Implement sched_yield syscall for x86 only.

["Zack Weinberg" <zack@owlfolio.org>]

On Thu, Jun 8, 2023, at 1:38 PM, Noah Goldstein via Libc-alpha wrote:
> On Thu, Jun 8, 2023 at 3:23 PM Zack Weinberg via Libc-alpha
> <libc-alpha@sourceware.org> wrote:
>>
>> On Thu, Jun 8, 2023, at 12:53 PM, Noah Goldstein via Libc-alpha wrote:
>> > On Thu, Jun 8, 2023 at 2:41 PM Florian Weimer <fw@deneb.enyo.de> wrote:
>> >>
>> >> * Zack Weinberg via Libc-alpha:
>> >>
>> >> > If these registers are all call-clobbered then maybe it makes sense
>> >> > to do this unconditionally in the syscall entry path, kernel
>> >> > side.
>> >>
>> >> This is not a backwards-compatible change and probably breaks glibc
>> >> itself because the asm constraints clearly indicate that vector
>> >> registers are NOT clobbered.
>> >
>> > we only need a few select functions.
>>
>> If the vector regs aren't call clobbered (and I really mean *call* clobbered here, not syscall clobbered) then this isn't a safe change *at all*, ne?
>>
>> I see why compatibility precludes doing this kernel-side, but then it seems to me the proper place is in the syscall stub macros.
>>
> We are taking advantage of the fact that call ABI clobbers all
> vectors. macro doesn't imply any clobbers.

OK, so then why *not* alter the syscall stub macros to do this uniformly for all syscalls, or for all but a handful of things which are unlikely to cause a context switch and the extra cost of the clear instruction itself is significant (e.g. get*id, sigprocmask).

zw