096. Specialize Generated Go Work-Control Methods

• Status: Superseded by G050: a generated parser has one form of each function, so the branch this record measures is not in the output
• Date: 2026-10-14

#Problem and Mechanism

Work-control checks add measurable cost to generated Go ordinary parsing and recognition. A broader ordinary-method specialization reduced part of that cost, but retained a sizable gap to the pre-work-control baseline. The selected variant specializes emitted ordinary methods and clones only the newNode, newMatch, setCapture and expect helpers. It hardens the generated-method registry and adds a generation-only ablation control. The internal emitter uses synchronous callbacks and preserves parser/call-local state rather than introducing escaping aliases or unrelated shadow state. It does not change the public API or matching semantics.

The specialization is adopted with the work-control implementation. A broader runtime-helper variant was rejected: it gained about 1–2% in recognition but made controlled XML parsing about 3% slower. A separate ordinary-method split experiment was also rejected after three paired 500 ms samples: JSON/XML recognition was about 2% slower and Outline about 4% slower against the selected variant; other profiles showed mixed results without a consistent gain. Correctness tests for that experiment passed, but it was not adopted.

#Applicability

#Reproduction

The 131-profile comparison used Go 1.27.1 on an Apple M3 Max (darwin/arm64), with GOMAXPROCS=16, GOGC=100 and GOMEMLIMIT=off. Three alternating 500 ms samples compared main 77f12f2 with the source-pinned candidate in this implementation, based on work-control baseline 186e1b0. The candidate included a source-format fix and regression tests. Samples were serialized. The same-source ablation disabled only the specializer while leaving work controls enabled. The fixed-allocation check uses 100 iterations and five alternating pairs. Build benchmark binaries before timing and alternate processes under the same environment; the existing active-control harnesses do not pre-parse before timing.

The internal transform accepts only emitted execution methods. The registry identifies the receiver and method together; missing or duplicate anchors, receiver replacement/shadowing, and direct control mutation/address escapes fail generation. Invocation control stays immutable, generated p/c names refer to the current parser/context, and built-in action callbacks run synchronously. This is not a transform for arbitrary Go code. Clone statements use their own source positions; insertion into original text preserves controlled statements and comments. Labels are renamed within ordinary clones. Runtime helpers outside the four families retain their existing nil checks.

#Results

The short focused run used three alternating 500 ms pairs against the stable default-off work-control candidate. On the sampled JSON and XML profiles, generated recognition was 20–24% faster and typed parsing was 17–18% faster. That comparison did not establish parity with the pre-work-control baseline.

The subsequent 131-profile comparison used three paired 500 ms samples against the exact main baseline. In the 24 generated Go Node parse/recognition profiles, candidate/main median ratios ranged from 1.039× to 1.332×; in eight typed-parse profiles, they ranged from 1.082× to 1.234×. Examples include JSON/XML recognition at 1.083×/1.183×, Minilang recognition at 1.259×, JSON/XML Node parsing at 1.161–1.184×/1.163–1.172×, and JSON/XML typed parsing at 1.082×/1.191×. These are descriptive paired medians, not a general estimate for other workloads. The candidate includes the source-format fix and regression tests. This comparison measures total work-control feature cost; the separate enabled/disabled comparison isolates this mechanism.

The same-source ablation comparison used three alternating 500 ms pairs over 32 ordinary and eight active profiles. With the specializer enabled, ordinary generated Node parse/recognition medians were 0.760–0.969× and typed-parse medians were 0.815–0.936× the disabled variant. Active-control medians were 0.985–1.005×; individual sample ranges remained noisy, including an XML case with a 9.7% slower sample. These harnesses do not explicitly pre-parse, so they do not isolate warmed active calls. In a separate fixed-100-iteration allocation comparison of Minilang ParseAST, the time ratio was 0.922× [0.908–0.926], bytes/op 1.000× [0.959–1.071], and allocations/op 1.000× [0.970–1.053]. No persistent allocation increase was identified in that control.

The selected binaries for JSON, XML and Minilang each exercised Node parsing, typed AST parsing and recognition. Their generated parser sources match the corresponding files in the 131-profile candidate manifest. Executable sizes were:

The specialized generated sources were 38.75% (JSON), 58.71% (XML) and 37.61% (Minilang) larger than their same-source ablations. Builds used go build -trimpath; both variants used the same caller exercising all three entry points. These size results are limited to the listed consumers.

#Decision and Limits

Adopted with the measured residual costs. Against pre-feature main, the selected work-control implementation still has generated Go ordinary- path overhead; the selected specialization reduces part of it, at the measured code-size cost. Further ordinary-path improvement is separate follow-up work and does not change the cancellation or budget contract. The ordinary-method split and broader runtime-helper alternatives were measured and rejected. Do not infer broader gains or parity outside the measured profiles and three consumer sizes.

#Validation

Local verification used Go 1.27.1, Node.js 24.19.0 and TypeScript 5.9.3. go test -timeout=30m ./... passed all 18 root-module packages; the engine package took 851.406 seconds. The two skipped engine tests are opt-in artifact writers, not semantic coverage. go generate ./parsers ./bench, every parser module in parsers/test.sh, site tests and the actual site build/link check passed (229 pages, 257 files). Independent source and evidence review found no unresolved material findings. These checks describe local output, not remote CI or deployed behavior.