On 27/05/2024 09:52, Roger Pau Monné wrote: >> My intent here was to prevent the compiler omitting the write (though in >> practice it didn't). I think the barrier is still required to prevent >> reordering according to the spec. > > Yes, my understating is that you added the ACCESS_ONCE() to possibly > prevent the compiler from re-ordering the CPU_TO_X2APICID[ap_cpuid]) = > read_apic_id() after the 'hlt' loop?
Yes, but not only that. Also preventing the compiler from eliding the write altogether on the basis that it's not read after within the function. I have suffered that particular behaviour on older versions of GCC and it stung very much. > > AFAICT the expressions in the `for` statement are considered sequence > points, and the calling `read_apic_id()` could have side-effects, so > the compiler won't be able to elide or move the setting of > CPU_TO_X2APICID[] due to all side-effects of previous evaluations must > be complete at sequence points. > > I'm fine with leaving the wmb() + ACCESS_ONCE(). > > Thanks, Roger. Cheers, Alejandro
