Discussed with @aclements. Open an issue for tracking.

@martisch and I also discussed this recently. Having cache misses and branch mispredictions would be super helpful.

I think this would be great. Probably we would want to build this on #16110. Probably we would want to put this behind a flag to go test since it will increase its chattiness (then it could also fail if the flag was specified but couldn't be supported).

What is the proposal to evaluate here? Everyone (including me) seems to think this is a good idea but there are no actual details.

Here's a more concrete proposal as a starting point.

Add a -test.benchperf flag to the flags exported by the standard testing package. When this flag is specified, benchmarks report additional per-operation metrics based on performance counters exposed by the hardware performance monitoring unit. If this flag is passed but the platform does not support hardware performance counters, it is a no-op. Exact counters would depend on the platform, but a good set to collect on x86 is: cycles, LLC-misses, branch-misses. These would be reported as new metrics cycles/op, etc.

On Linux, this would be built on the kernel's perf_event_open API, which takes an event description and returns an FD from which the event counter's current value can be read at any time.

Questions:

• I don't like -test.benchperf. What should it be called?
• Should the flag accept the names of counters to collect? At least on Linux there's a standard way of naming events and a way to enumerate supported events (see perf list).

-test.counters

I want a retired instruction count. That's even more stable than cycles (although not as important).

Sure, -test.counters seems fine (or something else if you decide that's not accurate enough). Marking proposal accepted.

Please do give some thought to having a few -counters= names that work portably across architectures. If there's a cycle count for x86 and one for arm it would be nice if there is a single name that enables either one, for example. (I'm assuming this is a comma-separated list and that unknown things are just ignored, or something like that.)

FWIW, Linux perf has a list of portable names already, printed by perf list (specifically the "Hardware event" and "Hardware cache event" categories). I'm pretty sure the "Hardware events" are supported basically everywhere. I'm less sure about the "Hardware cache events". These event types are baked into the perf ABI.

For reference, here are the "Hardware events": branch-instructions (alias branches), branch-misses, bus-cycles, cache-misses, cache-references, cpu-cycles (alias cycles), instructions, ref-cycles.

Has there been any progress on this one?