Reconstruct Comment#string lazily from its location

[ksss]

Summary

RBS::AST::Comment#string is now computed on first access from location.source instead of being materialized by the parser for every comment. The parser used to build a Ruby String for each comment, duplicating bytes the Buffer already holds, while consumers rarely need all of them at once: Environment and DefinitionBuilder only carry them through, and Steep reads a handful on demand (hover, completion). Materializing every comment up front pays for text that is mostly never looked at, and the lazy version makes the first access of each comment pay instead.

To stop the C translation from building the string, config.yml gains a per-field option expose_to_ruby: false: the field stays in the C struct and in the serialized form, but ast_translation.c does not pass it to the Ruby constructor. The WASM deserializer honors the same flag by reading and discarding the field, so both construction paths produce identical objects. The C library, the Rust crate and the serialization format are unchanged.

Measurements

Both numbers below are from the benchmarks in this repository, run on master and on this branch with the same Ruby (4.0.6) and the same rbs collection.

benchmark/memory_new_env.rb (core)

memory_profiler	master	this branch	change
total retained	10.12 MB / 81,580 objects	8.00 MB / 78,617 objects	−2.12 MB (−20.9%)
total allocated	15.59 MB / 163,525 objects	12.75 MB / 158,742 objects	−2.84 MB (−18.2%)
retained String	4.42 MB	2.30 MB	−2.12 MB
retained RBS::Location	2.39 MB	2.39 MB	0
retained Array / Types::Function / Hash / MethodType	0.62 / 0.56 / 0.34 / 0.25 MB	same	0

benchmark/memory_new_rails_env.rb (Rails rbs collection, 93 gems)

memory_profiler	master	this branch	change
total retained	61.86 MB / 656,840 objects	55.88 MB / 645,551 objects	−5.98 MB (−9.7%)
total allocated	99.75 MB / 1,233,463 objects	92.40 MB / 1,219,074 objects	−7.35 MB (−7.4%)
retained String	15.13 MB	9.15 MB	−5.98 MB
retained RBS::Location	19.48 MB	19.48 MB	0
retained Array / Types::Function / Hash / MethodType	5.19 / 4.55 / 2.93 / 2.15 MB	same	0

The whole difference is in String; every other class is byte-identical. The smaller ratio for the Rails collection reflects that gem signatures carry far less documentation than core.

Process RSS and wall time (without memory_profiler)

new_env	master	this branch
RSS delta, core (mean of 3)	12.4 MB	11.0 MB (−11%)
RSS delta, Rails collection (mean of 3)	105.0 MB	99.2 MB (−5.5%)
wall time, core (median of 7)	39.8 ms	38.9 ms
wall time, Rails collection (medians of 5 interleaved processes)	296–340 ms	306–335 ms (within noise)

No Comment#string is called while building an Environment, so the lazy path adds no work there.

Compatibility

• Comment.new(string:, location:) keeps working; string: is now optional, and a comment built without either raises on string.
• The reconstruction mirrors parse_comment_lines in src/parser.c, including CRLF (the CR is not part of the comment token) and bare CR (kept). Verified by comparing the lazily reconstructed string against the previously materialized one for all 7,607 comments in core + stdlib, plus new tests for CRLF, bare CR, indentation and # without a following space.
• The slice is taken directly from the buffer rather than through Location#source, so no second copy of the text is memoized on the Location after the first access.
• @string is nil until first access; code inspecting the ivar directly (the WASM round-trip test did) sees that change.

🤖 Generated with Claude Code