core: skip tracer on ext/heredoc thunks

[melvinw]

The tracer overhead is guaranteed to be a waste in both of these cases. Skipping it decreases benchmarks/osh-runtime/bin_true wall time by 5-7%.

[andychu]

Hmmm, surprised the effect is so big here

I was hoping to add more shell-level tracing, e.g. with TSV/JSON output and so forth

There was a Zulip thread about "zero overhead tracing as a motivation for performance"

I guess our language is kinda slow for that, and ebpf does a lot better ... but I still think there are some use cases for shell-level tracing (e.g. enhance set -x) and OILS_TRACE_DIR ...

---

Let me look into this a bit

[melvinw]

Sounds good. We could tweak SkipTrace to accept a tracer to make the decision a little smarter, too.

If I remember correctly the main thing this skips is an allocation in tracer.OnNewProcess(). I had a more targeted patch that deferred that to later that might help in lieu of this too d9cb6a2

Let me know what you decided and I'll adjust.

[melvinw]

One more clarification just in case it's not clear: we only skip the tracer in the child after forking (and only for ExternalThunk and HereDocWriterThunk). We still call OnProcessStart() from the parent unconditionally