Oil 0.12.4 - User Feedback

2022-08-24

This release addresses great feedback and bug reports on OSH and Oil. It's also a checkpoint for the work "under the hood" on the garbage collector, which is now being funded by NLnet.

Oil version 0.12.4 - Source tarballs and documentation.

Table of Contents

Summary and Acknowledgements

Appendix: Metrics for the 0.12.4 Release

Spec Tests

Benchmarks

Native Code Metrics

Summary and Acknowledgements

Here are notable changes, with credits to users who tried Oil and sent feedback.

OSH

Fixed a crash in the printf builtin with NUL bytes. Reported by Peter Debelak.
- This exposed implementation differences between Oil-CPython and oil-native, which we still need to fix. Most CPython functions raise errors when strings contain NUL bytes, but Oil shouldn't.
A change from the previous release: The -i flag to assignment builtins, like declare -i, is now a no-op. Docs:
- Known Differences > Values are Tagged With Types, Not Locations
- Shell Idioms > Removing Dynamic Parsing

Oil Language

Remove bare $foo from expressions, in favor of foo (a typed variable) or "$foo" (a string).
- In particular, this removes confusion between $dict['key'] and dict['key']. The former may be more familiar to Perl and PHP users, but Oil uses the latter, which is more like Python and JavaScript!
- Documented this at Oil language FAQ > What's the difference myvar, $myvar, and "$myvar"?
  - In Oil, it's best to think of $ as the substitution operator. So foo is a variable, and $foo substitutes it into a string.
- Thanks to Samuel Hierholzer for feedback that motivated this. Your feedback can influence the Oil language!
Arithmetic and comparison operators now convert strings to integers.
- For example, '2' + '3' evaluates to 5 because 2 + 3 is 5.
- And '22' < '3' is false because 22 < 3 is false. That is, comparison operators are always numeric; they don't do lexicographic comparison, as in Python.
- I believe this behavior is more appropriate for a shell-like language, e.g. for processing TSV files.
- Thanks to terinjokes for feedback on this.
- Update: More rationale on lobste.rs.
Fixed a few misleading error locations. They will look more like this:

  if (len(42)) {
  ^~
'_tmp/a':1: fatal: Type error in expression: object of type 'int' has no len()

Please file a bug if you get bad or confusing errors. (This error should be more precise, but it now points to the right general location.)

Eggex

Fixed translation of eggexes with literal [ or ]. Reported by Samuel Hierholzer.
That fix led to more fixes in class literal translation. Oil is now smart enough to translate / [ 'a' ']' ] / to []a], which is a very confusing aspect of regex syntax!
- It also handles these special characters correctly: ^ - \.
Fixed incorrect caching of eggex evaluation, also found by Samuel Hierholzer. The code was too clever, and now it's a straightforward AST evaluator.

Under the Hood

Jesse Hughes and I have been working on the C++ translation. We're merging two proofs of concept:

The "leaky" translation I did a couple years ago, which validated the performance and feasibility of translating Python to C++.
The Cheney garbage collector.
- For reasons I should write about later, we're probably using mark-and-sweep instead: #oil-dev > Simple Mark and Sweep Design for a Shell

I also did more work on the CI and devtools:

mycpp runs with Python 3.10, which fixes a problem for contributors, and will let us use Python's new match statement.
- Thanks to Brandon for walking through this with me.
Our CI containers are versioned and tagged. I hope to fold some of this experience with OCI/Docker into Oil.

Closed Issues

Here is more detail on the issues above, and more.

#1281	Build error for x86_64: fanos.c needs -std=c99
#1278	Eggex with literal [ or ] translated incorrectly
#1271	Regex in a loop exits FATAL on second iteration
#1262	Run mycpp with a consistent Python version
#1258	Remove bare $foo in expression mode; it should be foo or "$foo"
#1257	printf crashes with NUL bytes, e.g. string `$'\U0'`
#1247	Get rid of old Str* layout (mylib_leaky)
#1243	CI should run with container version built from the repo
#1221	leaky bindings: get rid of Str0 and IsNulTerminated() before porting to GC
#1118	Missing location info
#815	Implement / document string <-> integer coercion rules

Commit Log

Here are the commits from other contributors. You can also view the full changelog.

6198e29a	Jesse Hughes	[mycpp] Refactor/move mylib_old Str code into leaky_types (#1261)
b9d58e13	Jesse Hughes	[mycpp] Continue porting old Str layout to GC layout (#1263)
019d835a	Jesse Hughes	[mycpp] Big refactoring to get rid of old Str layout (#1264)
23bf527e	Jesse Hughes	[mycpp] Move Str out of gc_heap namespace (#1265)
b16d5ddf	Jesse Hughes	[refactor] Big refactoring away from OLDSTL types, towards GC types (#1268)
e06b7928	Jesse Hughes	[mycpp] Remove ALLOCATE macro hack (#1269)
bcaef989	Jesse Hughes	[refactor] Remove duplicate test_str_replace (#1279)
1e10dcbe	Jesse Hughes	[mycpp runtime] Remove OLDSTL List and consolidate code (#1282)

What's Next?

We have more people interested in working on Oil! I'll be helping them get up to speed.
- Feel free to post a message on #new members if you want to get involved!
The C++ translation is the highest priority.
- #oil-dev > Plan to translate oil-native
I'm working on the Oil language in parallel. Note that it is not yet translated to C++ — the first step is to translate OSH. But this hasn't stopped users from sending great feedback!

Please try Oil and let me know what happens on Zulip or Github!

Appendix: Metrics for the 0.12.4 Release

These metrics help me keep track of the project. Let's compare this release with the last one announced, version 0.12.0.

Spec Tests

We've been working under the hood, so OSH didn't change that much.

OSH spec tests for 0.12.0: 2014 tests, 1786 passing, 89 failing
OSH spec tests for 0.12.4: 2023 tests, 1789 passing, 91 failing

I wrote new tests to cover Oil's integer arithmetic and comparison behavior, and eggex.

Oil spec tests for 0.12.0: 483 tests, 444 passing, 41 failing
Oil spec tests for 0.12.4: 502 tests, 464 passing, 38 failing

Benchmarks

The performance of the oil-native parser reflects current work "under the hood". Context: #oil-dev > Two Changes That Will Decrease Perf

Parser Performance for 0.11.0: 8.9 thousand irefs per line. Our "baseline" for several months.
Parser Performance for 0.12.0: 6.3 thousand irefs per line. I mistakenly removed GC rooting here, which led to an artificial speedup!
Parser Performance for 0.12.3: 14.0 thousand irefs per line. I restored GC rooting, and started using slow malloc() instead of a bump allocator! We're preparing for mark-and-sweep.
Parser Performance for 0.12.4: 16.4 thousand irefs per line. What caused this regression? Probably some of the work on consolidating the "leaky" and GC runtime.

Wall time measurements for 0.12.4:

bash: 349.8 and 626.3 lines per ms
oil-native: 234.8 and 433.6 lines per ms

I want to get back to bash parity, but we should get a basic garbage collector working first.

Native Code Metrics

OSH C++ spec tests for 0.12.0: 1789 green, 1720 green in osh_eval.py, 1680 green in C++
OSH C++ spec tests for 0.12.4: 1792 green, 1723 green in osh_eval.py, 1684 green in C++

The size of the translated code remained the same:

oil-cpp for 0.12.0: 89,043 lines of C++
oil-cpp for 0.12.4: 89,167 lines of C++

The binary got unexpectedly smaller! This could be due to removing std::vector<T> in favor of our garbage-collected List<T>.

ovm-build for 0.12.0: 959,120 bytes of native code (under GCC)
ovm-build for 0.12.4: 828,088 bytes of native code (under GCC)