[browser] samplepoint instrumentation into Mono profiler

[pavelsavara]

Motivation

Collect sample stack traces in the single-threaded browser.
Single threaded environment can't have dedicated thread for "stop-the-world + collect stack traces" implementation.
So we need to collect samples by instrumenting the code instead.

Previously we implemented browser based profiler which collects "sample" on every method entry/leave. In realistic application, that generates too much traffic for the browser profiler.

This enhancement could also be used to implement sampling profiler with ST diagnostic server.

Goal

• add new samplepoint instrumentation into Mono profiler, use the existing enter/leave too.
• inject the event/call into same locations as GC safepoint.
  • we can skip emitting it at the method start, because there is already method_enter event/instrumentation.
  • This is also same location where cooperative MT would stop-the-world to take a sample.
• take sample on each Nth visit to the samplepoint (because measuring current time is more expensive)
• keep updating the size of N to match expected sample time interval

Contributes to #76316

Non-goal: the browser profiler is not compatible with MT.

Changes

Interop

• new interp IR instruction MINT_PROF_SAMPLEPOINT
• new profiler event method_samplepoint
• new interp macro INTERP_PROFILER_RAISE

Jiterp

• new jiterp helper append_profiler_event calling new mono_jiterp_prof_enter, mono_jiterp_prof_samplepoint, mono_jiterp_prof_leave

AOT

• new Mono JIT method mini_profiler_emit_samplepoint
• example of code emitted by AOT with call $mono_profiler_raise_method_samplepoint

browser profiler

• new config option sampleIntervalMs for browser profiler. 1ms by default. All samples when set to 0.
• changes JavaScript called only to get time mono_wasm_profiler_now and mono_wasm_profiler_record
• all sampling logic in browser.c new should_record_frame
• stack of start times profiler_stack_frames and skip counter sample_skip_counter
• default sample interval 10ms
• callspec parametr to profiler allows to filter the instrumentation, for example by namespace callspec=N:Sample.FooNs

How to use

Interpreter

• dotnet workload install wasm-tools

In the project file

<WasmProfilers>browser;</WasmProfilers>

In your main.js

import { dotnet } from './dotnet.js'
dotnet.withConfig({
        browserProfilerOptions: {
            sampleIntervalMs: 10, // optional
        }
    })
    .create();

or in your Blazor app

Blazor.start({
            configureRuntime: function (builder) {
                builder.withConfig({
                        browserProfilerOptions: {
                            sampleIntervalMs: 10, // optional
                        }
                });
            }
        });

AOT

You can add following, so that the native code produced by AOT has symbols, which are visible in the browser devtools profiler.

<RunAOTCompilation>true</RunAOTCompilation>
<RunAOTCompilationAfterBuild>true</RunAOTCompilationAfterBuild>
<WasmNativeStrip>false</WasmNativeStrip>
<WasmProfilers>browser</WasmProfilers>
<WasmNativeDebugSymbols>true</WasmNativeDebugSymbols>
<WasmBuildNative>true</WasmBuildNative>

callspec

If you want to filter profiled methods, you can use callspec

Callspec is documented here https://github.com/dotnet/runtime/blob/main/docs/design/mono/diagnostics-tracing.md#trace-monovm-profiler-events-during-startup

browserProfilerOptions: {
    callSpec: "N:SampleNamespace" // optional
}

Your for AOT needs to match callspec in the browserProfilerOptions in JS

<WasmProfilers>browser:callspec=N:Sample;</WasmProfilers>

[pavelsavara]

Interpreter: 1ms, skip rate between 200-500 sample points.

[hakenr]

Hi @pavelsavara,
I'm in! I'll try to reach you via Teams.

[pavelsavara]

My testing shows that enter/leave events are unbalanced.
I think my adding samplepoint in instrumentation didn't change that.
But the design of measuring the difference between method entry and method end requires that the events are balanced.

I will test it further to see if this could be fixed in the instrumentation or if this must be detected and survived by the receiving side.

[kg]

I think you can't randomly record only a % of enter/leave events, or if you do you need to ensure that they are balanced - for example if you record an entry for method x the matching return also needs to be recorded. So you can only pick a % of enters to record but you need to do the check for every LEAVE as to whether this current stack frame previously recorded an enter.

[BrzVlad]

LGTM

[pavelsavara]

/ba-g CI timeouts and known issues

[lateralusX]

Looks like I was late to the party (was OOF last week), but add my comments/questions here anyways, either for future follow ups or PR's.

[lateralusX]

Do we need to heap allocate this? Looks like stack allocation should be enough, unless we are afraid consuming stack space since the MonoProfilerCallContext that includes MonoContext.

[pavelsavara]

This is the place that possibly deals with MonoProfilerCallContext -> context

runtime/src/mono/mono/mini/mini-profiler.c

Line 19 in a015e48

emit_fill_call_ctx (MonoCompile *cfg, MonoInst *method, MonoInst *ret)

What would happen if we removed context field completely ?

[lateralusX]

If you hand out ownership of this pointer and expect others to free it, it shouldn't be const.