Runtime event trace visualization

[kpamnany]

This adds a bpftrace script containing handlers for many of the probes in the Julia runtime, and also adds some probes. The various handlers produce JSON output that can be read by Perfetto.

Here's what this looks like for parallel Fibonacci with 4 threads:

BUILD INSTRUCTIONS:

• Create Make.user containing WITH_DTRACE = 1.

TODO:

• Determine why there are misplaced end events and fix.
• Cannot use task pointer as task ID as tasks can be reused.
• Add more probes (start+finish inference, start+finish codegen, others?).

Cc: @vtjnash, @JeffBezanson

[vilterp]

May be worth trying to feed this info into Catapult, the tracing UI built into Chrome, so we don't have to write our own. The format is documented here: https://docs.google.com/document/d/1CvAClvFfyA5R-PhYUmn5OOQtYMH4h6I0nSsKchNAySU/preview

The Go runtime uses this to visualize scheduling and GC events. Here's an example debugging session: https://about.sourcegraph.com/go/an-introduction-to-go-tool-trace-rhys-hiltner/

[kpamnany]

With the current set of jl_trace() calls, the trace file generated from julia --threads=4 --trace-file=julia_startup.xml looks like:

[kpamnany]

After some discussions, we decided that this will be reworked on top of #43453 as it will be possible to completely avoid overhead when not interested in tracing (without recompiling).

[kpamnany]

For the added probes:

• Type inference begin and end seem to properly demarcate inference but no contextual information is available at these probe points (i.e. what is being inferred).
• Compilation begin and end seem to include codegen (begin and end) and LLVM optimization (begin and end) but from looking at the visualization, there is more than these two things happening during compilation.

I've also switched to using jl_object_id(task) for a unique task identifier (since the task pointer itself is not unique).

Hope to get suggestions for additional useful probe points, new/more useful probe parameters or any other changes. Otherwise, this is good to go.

[kpamnany]

Here's a zoomed-in section of the Perfetto visualization for parallel fib with 4 threads, showing the newly added sections: