Re: curcpu() and curlwp()

[christos%astron.com@localhost (Christos Zoulas)]

In article <20110605171617.GD2029%snowdrop.l8s.co.uk@localhost>,
David Laight  <david%l8s.co.uk@localhost> wrote:
>I've just been looking at the definitions of curcpu() etc on
>i386 and amd64 (they differ only in the segment register used).
>
>in sys/arch/i386/include/cpu.h we have:
>
>__inline static lwp_t * __attribute__ ((const))
>x86_curlwp(void)
>{
>        lwp_t *l;
>
>        __asm volatile("movl %%fs:%1, %0" :
>            "=r" (l) :
>            "m"
>            (*(struct cpu_info * const *)offsetof(struct cpu_info, 
> ci_curlwp)));
>        return l;
>}
>
>I think it should be:
>       "m" (((const struct cpu_info *)0)->ci_curlwp)
>
>The actual code seems to be a reference to a pointer to 'struct cpu_info'.
>The generated code is probably identical, but the inferences are different.
>
>(cpu_set_curpri() is broken in a different, but similar way.)
>
>I also think that the curcpu() and culwp() functions need not be volatile.

I think that if offsetof() produces suboptimal code, it should be fixed,
not expanded in place differently.

christos