Add Crystal::Tracing for runtime tracing

[ysbaddaden]

Implements tracing of the garbage collector and the scheduler as per #14618

Tracing is enabled by compiling with -Dtracing then individual tracing must be enabled at runtime with the CRYSTAL_TRACE environment variable that is a comma separated list of sections to enable, for example:

• none to disable everything (by default)
• gc
• sched
• gc,sched
• all to enable everything

The traces are printed to the standard error by default, but this can be changed at runtime with the CRYSTAL_TRACE_FILE environment variable. For example trace.log. You can also redirect the standard error to a file (e.g. 2> trace.log on UNIX shell).

Technical note: tracing happens before the stdlib is initialized, so the implementation must rely on some LibC methods directly (i.e. read environment variable, write to file descriptor) and can't use the core/stdlib abstractions.

Note: the Thread#current_thread? commit could be extracted into a distinct PR. Done: #14660

[straight-shoota]

the Thread#current_thread? commit could be extracted into a distinct PR.

Then let's do that, please 🙏

[straight-shoota]

polish: Standardize format for calls to Crystal.trace.
The section argument is always a symbol, the operation argument sometimes a symbol, sometimes a string.

Crystal.trace :sched, :spawn
Crystal.trace :sched, "event_loop"

It would be nice using the same format, maybe the operation could always be a string?

[ysbaddaden]

@straight-shoota following your review, I made the following changes:

1. I replaced the Crystal.trace macro with a couple def methods (one that yield and one that don't), which led to:

• the operation arg is now always a String 👍
• the method still yields twice, but this is no longer injecting the code twice, so it should be fine?

2. I replaced the LibC.WideCharToMultiByte call with a pure Crystal alternative that leverages String.each_utf16_char(slice, &) and loops to avoid a buffer overflow and splitting a multibyte char across multiple prints.

[ysbaddaden]

Thanks @straight-shoota. I applied all your suggestions!

[ysbaddaden]

I noticed that one macro used constant.downcase while the other used lowerscore in a completely different place, so I grouped the logic into the Tracing::Section enum as a couple methods: .from_id(slice) and #to_id (instead of #id) and now both are using lowerscore.