Re: Event reordering regression for software events

[Andi Kleen <ak@linux.intel.com>]

<sorry for the late reply I was on vacation>

On Mon, Jul 17, 2023 at 10:37:35AM -0300, Arnaldo Carvalho de Melo wrote:
> Don't we have a JSON output? My expectation is that after the work that
> Ian has been doing we will settle down, but even then downstreamers
> should try to use JSON as input as that would avoid the problems we are
> having?


In classic perf CSV was our machine readable output, and that is what got broken
here. I hope nobody is advocating for abolishing CSV output as a machine
readable format.

For JSON it got broken too because of:

> > I haven't tried to use it, but if you have something like
> > perf stat --json -e '{cycles,branches},{cycles,cache-misses}' how would you
> > distinguish it without order?

> 
> You mentioned that it was difficult to add entries to 'perf test' but

I can add some tests for basic reordering not occurring. That should
be possible.

> that we could try running toplev's regression test, right? I can add it
> to my set of tests but perhaps the best thing would be to wire it up to
> one of the CIs out there?

Yes that would be a good idea. Note that there are sometimes problems
which are on the toplev side, not perf side, but I guess can sort that
out when that happens.


-And