[mono] fix --trace

[pavelsavara]

Fixes regression caused by #112352 that breaks --trace on mono VM.
Root cause is stack allocation inside of a loop for emit_fill_call_ctx of new SAMPLEPOINT_CONTEXT profiler event.

That event is not necessary for any of the other scenarios. And also samplepoint is not interesting trace event.

Fix

• don't emit samplepoint for traces
• remove bad implementation of SAMPLEPOINT_CONTEXT from mini_profiler_emit_samplepoint

[Copilot]

Pull Request Overview

This PR addresses a regression in the --trace option by adjusting the profiler instrumentation flags set during JIT compilation.

• Removed samplepoint-related flags from cfg->prof_flags when --trace is enabled.
• Ensures only enter/leave instrumentation is active under --trace.

Comments suppressed due to low confidence (2)

src/mono/mono/mini/mini.c:3456

• Add a unit or integration test that verifies enabling --trace produces exactly the expected prof_flags (ENTER, ENTER_CONTEXT, LEAVE, LEAVE_CONTEXT) to guard against future regressions in flag composition.

	if (trace)

src/mono/mono/mini/mini.c:3458

• [nitpick] Add or update a comment explaining why the samplepoint instrumentation flags were removed under --trace to clarify the intent for future maintainers.

			MONO_PROFILER_CALL_INSTRUMENTATION_ENTER | MONO_PROFILER_CALL_INSTRUMENTATION_ENTER_CONTEXT |

[BrzVlad]

Could you add a small explanation for this ? Is tracing simply incompatible with samplepoint instrumentation ?

[pavelsavara]

Could you add a small explanation for this ? Is tracing simply incompatible with samplepoint instrumentation ?

We discovered that the root cause it that MONO_PROFILER_CALL_INSTRUMENTATION_SAMPLEPOINT_CONTEXT allocates on the via emit_fill_call_ctx stack via localloc inside of the loop. Implementing it properly would take some effort.

Because the SAMPLEPOINT_CONTEXT feature is only new in Net10 previews and it's not used by the browser profiler, we could remove that instruction type.

[BrzVlad]

Does MONO_PROFILER_CALL_INSTRUMENTATION_SAMPLEPOINT_CONTEXT still need to exist ? How exactly is it superior to the standard samplepoint instrumentation ?

[pavelsavara]

Does MONO_PROFILER_CALL_INSTRUMENTATION_SAMPLEPOINT_CONTEXT still need to exist ? How exactly is it superior to the standard samplepoint instrumentation ?

It only needs to exist for the INTERP_PROFILER_RAISE macro to stay simple. Alternatively we could copy&paste that macro and make version without _CONTEXT concatenation. Do you prefer that ?

[BrzVlad]

Yes, I would prefer to have it removed if it's an option that will not be used

[pavelsavara]

[pavelsavara]

/ba-g CI timeouts