Re: us_delay() improvement

[Nick Garnett <nickg at ecoscentric dot com>]

Jonathan Larmour <jifl@eCosCentric.com> writes:

> Nick Garnett wrote:
> > +            if( usecs > usec_per_period/2 )
> > +                slice = CYGNUM_KERNEL_COUNTERS_RTC_PERIOD/2;
> > +            else if( CYGNUM_KERNEL_COUNTERS_RTC_PERIOD >= 0x7FFFFFFF/usec_per_period )
> > +                slice = usecs * ticks_per_usec;
> > +            else
> > +            {
> > +                slice = usecs*CYGNUM_KERNEL_COUNTERS_RTC_PERIOD;
> > +                slice /= usec_per_period;
> > +            }
> 
> I've been thinking about this.... we get overflow if
> usecs*CYGNUM_KERNEL_COUNTERS_RTC_PERIOD overflows, which means that
> since usecs should be at most usec_per_period/2, we have to "play
> safe" if plugging usecs == usec_per_period/2 in to the multiplication
> would overflow, because that's the worst case.
> 
> So we get the expression to test as:
> 
> (usec_per_period/2)*CYGNUM_KERNEL_COUNTERS_RTC_PERIOD >= 0x7FFFFFFF
> 
> which we reorganise to prevent overflow of this expression to become:
> 
> CYGNUM_KERNEL_COUNTERS_RTC_PERIOD/2 >= 0x7FFFFFFF/usec_per_period
> 
> which should be sufficient. The code as written will unnecessarily
> catch certain configurations I'd think.

Good point. I'd missed that implication. I'll make the change.


-- 
Nick Garnett                    eCos Kernel Architect
http://www.ecoscentric.com/     The eCos and RedBoot experts