121. Inactive Work-Control Checks in the Generated Go Runtime

• Status: Implemented
• Date: 2026-10-25

#Problem and Mechanism

A generated parser that is used without a work control still runs the runtime helpers that can charge one. Each such helper tests the control, and the compiler can inline it only when its body is small. The helpers are therefore split: the test, the call of a ...Controlled function that holds the charging body, and the ordinary body stay in the helper; the controlled body is a separate unexported function. This covers appendWork, copyWork, clearWork, growWork, the typed memoGet and memoPut, valueField, setValueField, the typed and node action contexts (And, Or, Unary, Fold, ListItems, asNode), stringBinary, compareText, equalExpected, sortExpected and evalErrorf.

FirstCall runs the ordinary body itself after controlledFirstCallPrepare, which does the controlled charges and growth, so a controlled parse executes the same steps. The operator apply of both runtimes replays recovery errors only when there are some (an empty append changes nothing), through an out-of-line function, so apply inlines.

In the runtime build only appendWork, copyWork, clearWork and the two apply functions inline; the other helpers lose a call layer or a large frame but stay calls (the typed memoPut costs 152 and Parser.memoPut 81 against a budget of 80 even without the check). Charges, cancellation points, limits and results of controlled parses are as in the unsplit helpers.

#Results

Reference control: the same source with every work-control check deleted at build time (the ceiling, not a supported build). Generated parsers of parsers/json, duckdb, cel and golang (Node, recognizer and typed, generated without work controls, embedded runtime), inputs of the benchmark suite. Ten interleaved rounds over the build before, the build after, the control and a copy of the build before, -benchtime=500ms; medians as the change against the build before (the copy of the build before stayed within ±1.2%):

Bytes and allocations per operation are identical. The generated parser.go files are unchanged (the change is in the runtime). A test binary linking all eight parsers grows from 28,715,058 to 28,775,698 bytes (+0.2%).

#Limits

Of the gap to the control, between 28% (golang typed) and 69% (json node) is recovered; on most rows 65-72% remains. The json recognize control is slower than the build before, and cel recognize has no gap. What remains is the checks in Expect, Text, memoEnv, bind and the other helpers that do more than one thing, the helpers that stay calls (memoGet, memoPut, valueField), and code layout. Splitting those further needs a second call that costs as much as the check.