On Mon, Mar 25, 2019 at 11:38:41AM +0100, SZEDER Gábor wrote:
> diff --git a/progress.c b/progress.c
> index 02a20e7d58..b57c0dae16 100644
> --- a/progress.c
> +++ b/progress.c
> @@ -86,28 +86,30 @@ static void display(struct progress *progress, uint64_t
> n, const char *done)
> return;
>
> progress->last_value = n;
> +
> + if (!is_foreground_fd(fileno(stderr)) && !done) {
> + progress_update = 0;
> + return;
> + }
> +
Moving it here causes a measurable slowdown for:
git rev-list --progress=foo --objects --all
This function gets called for every single increment of the progress
counter. Whereas in its old location:
> tp = (progress->throughput) ? progress->throughput->display.buf : "";
> eol = done ? done : " \r";
> if (progress->total) {
> unsigned percent = n * 100 / progress->total;
> if (percent != progress->last_percent || progress_update) {
> progress->last_percent = percent;
> - if (is_foreground_fd(fileno(stderr)) || done) {
> - fprintf(stderr, "%s: %3u%%
> (%"PRIuMAX"/%"PRIuMAX")%s%s",
> - progress->title, percent,
> - (uintmax_t)n,
> (uintmax_t)progress->total,
> - tp, eol);
> - fflush(stderr);
> - }
It was only triggered when we accumulated enough increments to print. So
we save a few instructions in the backgrounded case, but it costs us a
lot of extra syscalls in every other case.
According to "strace -c", the number of ioctls for that rev-list on
git.git went from 6 to 373,461. But more importantly, my best-of-five
timings went from 3.340s from 3.407s. That's only 2%, but it would be
nice not to pay it if we don't need to.
-Peff
