More work on xtrace

[andychu]

continuing from #583

• special first and last line
  • first: PID. Maybe also: hostname, username, and absolute start time (unlike timestamp which is relative?)
  • last: status
• Trace across xargs $0
  • maybe AUTOMATICALLY export SHX_indent, SHX_pid_str, SHX, etc.
  • not included: SHX_punct
• BASH_XTRACEFD or SHX_file_desc
  • then all processes at once can trace here?
  • { BASH_XTRACEFD=$fd $0 -x myscript.sh } {fd}>out.txt
• $SECONDS for COARSE profiling
  • you can remove all the "spans" that are smaller than that
• fix SHELLOPTS bug? Sometimes it says "no such option"
• PARSE traces and make a diagram
  • flame graph? That only works for a single thread
• It would be nice to trace redirects. This came up while trying to figure out what's going on with OSH running sharness. It does a lot of redirects to descriptor 2, 3, and 4, and it uses BASH_XTRACE_FD.
• Idea from Better tracing for Oil #583: custom proc hooks? Or is PS4 good enough?

[andychu]

ShellSpec could use BASH_XTRACEFD.

[andychu]

Actually xtrace_rich did help in debugging sharness! It traced it to a subshell exiting 1.

Also sharness uses BASH_XTRACE_FD, so we should probably implement it.

[andychu]

I'm confused: I see forkwait of the same process TWICE here. I think this is a buffering issue with fork(). I think we need a flush() somewhere.

Yeah I think if the tracer is shared across fork, it needs to flush.... gah.

842  osh -x -O xtrace_rich --debug-file _tmp/debug --xtrace-to-debug-file ./fail.t    -x