On Mon, Dec 18, 2017 at 04:58:53PM -0500, Shih-Wei Li wrote:
> On Mon, Dec 18, 2017 at 2:10 PM, Andrew Jones <drjones@xxxxxxxxxx> wrote:
> >> +static void loop_test(struct exit_test *test)
> >> +{
> >> + unsigned long i, iterations = 32;
> >> + unsigned long sample, cycles;
> >> + unsigned long long min = 0, max = 0;
> >> + const unsigned long long goal = (1ULL << 29);
> >> +
> >> + do {
> >> + iterations *= 2;
> >> + cycles = 0;
> >> + for (i = 0; i < iterations; i++) {
> >> + sample = test->test_fn();
> >> + if (sample == 0) {
> >> + /*
> >> + * If something went wrong or we had an
> >> + * overflow, don't count that sample.
> >> + */
> >> + iterations--;
> >> + i--;
> >
> > I'd prefer incrementing a different counter after the sample==0 test /
> > continue, than to decrement both i and iterations.
>
> I'm not sure about the comment above. Could you please substantiate
> how a new counter should work here?
It doesn't really matter, but how about
do {
iterations *= 2;
cycles = 0;
i = 0;
while (i < iterations) {
sample = test->test_fn();
if (sample == 0)
continue;
cycles += sample;
...
++i;
}
} while (cycles < goal);
>
> >
> >> + debug("cycle count overflow: %lu\n", sample);
> >> + continue;
> >> + }
> >> + cycles += sample;
> >> + if (min == 0 || min > sample)
> >> + min = sample;
> >> + if (max < sample)
> >> + max = sample;
> >> + }
> >> + } while (cycles < goal);
> >> + printf("%s:\t avg %lu\t min %llu\t max %llu\n",
> >> + test->name, cycles / (iterations), min, max);
> >
Thanks,
drew
