Checking for fixed counter should probably use numGPCounters+1 events instead of just numGPCounters

[echiugoog]

PerfSpect/cmd/metrics/metadata.go

Line 311 in 8661c5f

for range numGPCounters {

To determine if a fixed counter is available, should probably add in numGPCounters+1 events, instead of just numGPCounters

e.g. if there are 4 numGPCounters, we want to do
perf stat -e "event,event,event,event,event" to see if there's a fixed counter for event

Alternatively, maybe use rdmsr?

[harp-intel]

As these metadata items aren't used, I plan to remove them altogether.

[echiugoog]

I use it in an upcoming PR: https://github.com/echiugoog/PerfSpect/blob/f9dae79cc4da7b3d130712c4865b13e341ff1346/cmd/metrics/perf.go#L112

[harp-intel]

That's a good idea, but I would consider doing it where the event groups are loaded instead of when the perf command args are formed.

[echiugoog]

That's a good idea, but I would consider doing it where the event groups are loaded instead of when the perf command args are formed.

For ARM, because I use the events and metrics from perf, I regroup them at a later stage. So doing the check later works in both cases. Ideally I can identify the groupings ahead of time, but I need ConfigureMetrics to parse the expression to identify all the events.