Re: [PATCH] tracing/histogram: Optimize division by constants

[Kalesh Singh]

On Thu, Oct 28, 2021 at 7:01 PM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>
> On Wed, 27 Oct 2021 20:28:54 -0700
> Kalesh Singh <kaleshsingh@xxxxxxxxxx> wrote:
>
> > +/*
> > + * Returns the specific division function to use if the divisor
> > + * is constant. This avoids extra branches when the trigger is hit.
> > + */
> > +static hist_field_fn_t hist_field_get_div_fn(struct hist_field *divisor)
> > +{
> > +     u64 div;
> > +
> > +     if (divisor->flags & HIST_FIELD_FL_VAR_REF) {
> > +             struct hist_field *var;
> > +
> > +             var = find_var_field(divisor->var.hist_data, divisor->name);
> > +             div = var->constant;
> > +     } else
> > +             div = divisor->constant;
> > +
> > +     if (!div)
> > +             return div_by_zero;
>
> Do we really need a div_by_zero constant function? What about just
> erroring here and perhaps return -EDOM?

Good point. We can detect this if the divisor is a constant and print
an error message instead. It'll also warrant a small update to the
tests and documentation. I'll send out a new version addressing these.

Thanks,
Kalesh

>
> -- Steve
>
> > +
> > +     if (!(div & (div - 1)))
> > +             return div_by_power_of_two;
> > +
> > +     /* If the divisor is too large, do a regular division */
> > +     if (div > (1 << HIST_DIV_SHIFT))
> > +             return div_by_not_power_of_two;
> > +
> > +     divisor->div_multiplier = div64_u64((u64)(1 << HIST_DIV_SHIFT), div);
> > +     return div_by_mult_and_shift;
> > +}
> > +