Ingo Molnar <mingo@xxxxxxx> wrote:
>
>
> * Davide Libenzi <davidel@xxxxxxxxxxxxxxx> wrote:
>
> > > why is it wrong?
> >
> > For HZ == 1000 it's fine, even if it'd better to explicitly make it HZ
> > dependent and let the compiler to discard them.
>
> the compiler cannot discard the multiplication and the division from the
> following:
>
> x * 1000 / 1000
>
> due to overflows. But we know that HZ is 1000 in the arch-dependent
> param.h, and in sched.c we use the HZ dependent variant:
>
> #ifndef JIFFIES_TO_MSEC
> # define JIFFIES_TO_MSEC(x) ((x) * 1000 / HZ)
> #endif
> #ifndef MSEC_TO_JIFFIES
> # define MSEC_TO_JIFFIES(x) ((x) * HZ / 1000)
> #endif
>
Yes, that's a correct optimisation. This is simply a namespace clash.
How about we do:
#if HZ=1000
#define MSEC_TO_JIFFIES(msec) (msec)
#define JIFFIES_TO_MESC(jiffies) (jiffies)
#elif HZ=100
#define MSEC_TO_JIFFIES(msec) (msec * 10)
#define JIFFIES_TO_MESC(jiffies) (jiffies / 10)
#else
#define MSEC_TO_JIFFIES(msec) ((HZ * (msec) + 999) / 1000)
#define JIFFIES_TO_MSEC(jiffies) ...
#endif
in some kernel-wide header then kill off all the private implementations?
|