Re: [PATCH v2 0/9] jevents/pmu-events improvements

[Ian Rogers]

On Mon, Jan 23, 2023 at 5:26 AM John Garry <john.g.garry@xxxxxxxxxx> wrote:
>
> On 21/12/2022 22:34, Ian Rogers wrote:
> > Add an optimization to jevents using the metric code, rewrite metrics
> > in terms of each other in order to minimize size and improve
> > readability. For example, on Power8
> > other_stall_cpi is rewritten from:
> > "PM_CMPLU_STALL / PM_RUN_INST_CMPL - PM_CMPLU_STALL_BRU_CRU / PM_RUN_INST_CMPL - PM_CMPLU_STALL_FXU / PM_RUN_INST_CMPL - PM_CMPLU_STALL_VSU / PM_RUN_INST_CMPL - PM_CMPLU_STALL_LSU / PM_RUN_INST_CMPL - PM_CMPLU_STALL_NTCG_FLUSH / PM_RUN_INST_CMPL - PM_CMPLU_STALL_NO_NTF / PM_RUN_INST_CMPL"
> > to:
> > "stall_cpi - bru_cru_stall_cpi - fxu_stall_cpi - vsu_stall_cpi - lsu_stall_cpi - ntcg_flush_cpi - no_ntf_stall_cpi"
> > Which more closely matches the definition on Power9.
> >
> > A limitation of the substitutions are that they depend on strict
> > equality and the shape of the tree. This means that for "a + b + c"
> > then a substitution of "a + b" will succeed while "b + c" will fail
> > (the LHS for "+ c" is "a + b" not just "b").
> >
> > Separate out the events and metrics in the pmu-events tables saving
> > 14.8% in the table size while making it that metrics no longer need to
> > iterate over all events and vice versa. These changes remove evsel's
> > direct metric support as the pmu_event no longer has a metric to
> > populate it. This is a minor issue as the code wasn't working
> > properly, metrics for this are rare and can still be properly ran
> > using '-M'.
> >
> > Add an ability to just build certain models into the jevents generated
> > pmu-metrics.c code. This functionality is appropriate for operating
> > systems like ChromeOS, that aim to minimize binary size and know all
> > the target CPU models.
>
>  From a glance, this does not look like it would work for arm64. As I
> see in the code, we check the model in the arch folder for the test to
> see if built. For arm64, as it uses arch/implementator/model folder org,
> and not just arch/model (like x86)
>
> So on the assumption that it does not work for arm64 (or just any arch
> which uses arch/implementator/model folder org), it would be nice to
> have that feature also. Or maybe also support not just specifying model
> but also implementator.

Hmm.. this is tricky as x86 isn't following the implementor pattern. I
will tweak the comment for the ARM64 case where --model will select an
implementor.

> >
> > v2. Rebase. Modify the code that skips rewriting a metric with the
> >      same name with itself, to make the name check case insensitive.
> >
>
>
> Unfortunately you might need another rebase as this does not apply to
> acme perf/core (if that is what you want), now for me at:
>
> 5670ebf54bd2 (HEAD, origin/tmp.perf/core, origin/perf/core, perf/core)
> perf cs-etm: Ensure that Coresight timestamps don't go backwards

Will do, thanks!
Ian

> > Ian Rogers (9):
> >    perf jevents metric: Correct Function equality
> >    perf jevents metric: Add ability to rewrite metrics in terms of others
> >    perf jevents: Rewrite metrics in the same file with each other
> >    perf pmu-events: Separate metric out of pmu_event
> >    perf stat: Remove evsel metric_name/expr
> >    perf jevents: Combine table prefix and suffix writing
> >    perf pmu-events: Introduce pmu_metrics_table
> >    perf jevents: Generate metrics and events as separate tables
> >    perf jevents: Add model list option
>
> Thanks,
> John