Why Sponsor Oils? | blog | oilshell.org

Release of OSH 0.5.alpha3

This is the latest version of OSH, a bash-compatible shell:

• https://www.oilshell.org/download/oil-0.5.alpha3.tar.xz (gzip version)

Please try it on your shell scripts and report bugs! To build and run it, follow the instructions in INSTALL.txt.

If you're new to the project, see Why Create a New Shell?. OSH can run unmodified shell scripts that are thousands of lines long, as described in the announcement for OSH 0.4.

Why an 'alpha' Release?

I'm calling it 0.5.alpha3 because most changes are under the hood, rather than user-facing.

I return from the Recurse Center in August, and I expect to release version 0.5 a few weeks thereafter. If you want to keep up with the project over the summer, please join oilshell.zulipchat.com, which I described in the last post.

It's more of a "pre-release" than an "alpha", but I didn't want to disrupt the naming scheme (note 1). In other words, it's still a high-quality release that passes a large suite of tests. The results are on the /release/$VERSION/ page.

It's also now live on the home page, so new users will download and test it.

Summary of Changes

This section summarizes the raw changelog.

• Thanks to timetoplatypus for reporting bugs that led to fixes:

  • #82: Crash in the [ builtin
  • #87: Problem with exit in the interactive shell
  • I'm working on a fix for #90: &> isn't implemented, which unfortunately isn't in this release.

  I'm still open to more bug reports! I'll respond to them while I'm away for the summer.

• Not visible to end users: Yorwba fixed a subtle byterun bug after I spent a couple hours trying to track it down.

  • Context: I played with byterun as an exercise for OVM. Good news: the OSH unit tests and spec tests now pass under byterun.
  • I now realize I can use byterun in another way: to generate type signatures for Oil's Python code, which is a step toward making it smaller and faster. I tried pyannotate as an alternative, but I'm not sure it's the best tool for Oil (e.g. due to limitations of the sys.setprofile() hook).

• In addition to osh, there's now an oshc command, which you can read as osh command or osh compiler. It has subcommands for tools, like oshc deps and oshc translate.

  • There will be an analogous oil / oilc pair
  • As well as an opy / opyc pair. This one is only relevant to Oil developers, not users. OPy is an implementation detail.

• Most other changes were "under the hood". For context, see Building Oil with the OPy bytecode compiler.

  • My last status update described opy/callgraph.py, as well as the work to make OPy's output deterministic. The drastic refactoring I mentioned made OPy simpler and slightly faster (TODO: publish benchmarks).
  • Since that update, I also got OSH tests passing under byterun, as mentioned above.

Metrics

Compared to the previous release, the benchmarks and metrics haven't changed much. This is good, because it means that replacing the bytecode compiler and then refactoring it didn't break anything.

For example, the bytecode isn't noticeably bigger:

• pyc-bytes for 0.5.alpha2: 996,913 bytes
• pyc-bytes for 0.5.alpha3: 1,005,443 bytes

The one anomaly I noticed is that the native code size is significantly different:

• ovm-build for 0.5.alpha2: native code size is 1,596,576 bytes
• ovm-build for 0.5.alpha2: native code size is 1,762,384 bytes

I still need to look into it, but it's probably caused by one of:

• -fomit-no-frame-pointer, which I added to make profiling with flame graphs work.
• I used the strip command rather than generating a stripped binary with gcc -s. Again, this was to support profiling on the optimized binary.

Lesson: It's surprising how many times simple metrics reveal problems. I don't know what the cause of this problem is yet, but I'm sure I'll learn something when I figure it out.

What's Next?

• Fix a few issues that timetoplatypus brought up on #oil-dev.
• I hope to summarize the blog backlog before I leave for the east coast. I've done this in previous posts tagged #blog-topics.
• OSH needs a "carrot" — something that makes it better than bash. I'm still working on a shell trace tool, which I also hope to release before I leave. It will be a rough first cut.
• As mentioned, the 0.5 release likely won't come until August or September.

Feel free to send me feedback in the comments!

Footnotes

[1] Recall that I released 0.5.alpha2 in early March, when I built Oil with the OPy bytecode compiler.