mirror of
https://github.com/vercel-labs/scriptc.git
synced 2026-11-03 07:54:40 +08:00
Malte Ubl
363fb69f07
An async function whose body can never park its execution context completes before its call returns, exactly like a fiber that never suspends. The LLVM emitter now proves that from the emitted IR: a body qualifies when neither its own text nor that of any module function it mentions (transitively) contains a suspension entry point (scr_await_*, scr_module_await, scr_promise_await_settled, scr_test_sub, generator yields, async-generator operations, fiber accessors). wasm32, programs embedding the dynamic engine, module-evaluation bodies, dyn results, and programs using node:test subtests (t.test awaits on the current fiber even from sync code) keep fibers. Qualifying spawn wrappers call the body on the caller's stack between scr_async_inline_enter/leave, which save and restore what a fresh fiber isolates (the Error.stack frame chain and the AsyncLocalStorage context), then settle a fresh promise from the outcome through the same fulfill/reject paths the fiber trampoline uses. A park or hop while a fiberless frame is on a fiber's stack traps as an internal error rather than reordering. In async-pipeline, loadUser and authorize become fiberless: one fiber per request instead of five. New corpus programs pin microtask ordering, rejections (handled and unhandled), finally, nested suspending spawns, closures, methods, Error.stack, AsyncLocalStorage, and a fiber-switch stress test. Final benchmarks for the branch (three commits) vs perf/harness, runs=15. macOS arm64 (local): | workload | baseline ms | candidate ms | change | 95% CI | verdict | size | node ms | candidate/node | | --- | ---: | ---: | ---: | --- | --- | ---: | ---: | ---: | | json-records | 279.2 | 280.7 | +0.5% | -1.8% .. +3.0% | neutral | +0.0% | 230.93 | 1.22x | | regex-logs | 305.8 | 306.0 | +0.1% | -0.8% .. +2.0% | neutral | +0.0% | 132.42 | 2.31x | | ast-interp | 235.3 | 235.1 | -0.1% | -0.6% .. +0.4% | neutral | +0.0% | 133.55 | 1.76x | | records-sort | 621.6 | 619.0 | -0.4% | -1.1% .. +0.6% | neutral | +0.0% | 131.56 | 4.71x | | template-render | 202.8 | 201.9 | -0.4% | -1.7% .. +2.6% | neutral | +0.0% | 99.28 | 2.03x | | functional-pipeline | 226.0 | 226.4 | +0.2% | -0.4% .. +0.6% | neutral | +0.0% | 120.24 | 1.88x | | async-pipeline | 267.0 | 21.6 | -91.9% | -92.2% .. -91.4% | faster | +0.0% | 58.58 | 0.37x | | word-graph | 417.6 | 566.8 | +35.7% | -49.6% .. +171.5% | neutral | +0.0% | 944.76 | 0.60x | | alloc-trees | 311.1 | 311.3 | +0.1% | -4.8% .. +2.3% | neutral | +0.0% | 87.37 | 3.56x | | numeric-kernels | 79.0 | 80.5 | +1.9% | -4.1% .. +8.4% | neutral | +0.0% | 94.05 | 0.86x | | validate-errors | 169.9 | 169.4 | -0.3% | -12.0% .. +12.3% | neutral | +0.0% | 517.13 | 0.33x | | csv-numbers | 379.3 | 372.8 | -1.7% | -27.8% .. +12.5% | neutral | +0.0% | 360.69 | 1.03x | | log-summary | 180.7 | 184.0 | +1.8% | -14.4% .. +9.1% | neutral | +0.0% | 181.51 | 1.01x | | inventory-report | 231.7 | 239.0 | +3.2% | -11.1% .. +10.9% | neutral | +0.0% | 245.91 | 0.97x | (word-graph was host noise from concurrent agents; the immediate recheck measured -2.0%, neutral.) Linux x86_64 (sandbox): | workload | baseline ms | candidate ms | change | 95% CI | verdict | size | node ms | candidate/node | | --- | ---: | ---: | ---: | --- | --- | ---: | ---: | ---: | | json-records | 1031.8 | 1019.7 | -1.2% | -5.5% .. +3.2% | neutral | +0.0% | 627.39 | 1.63x | | regex-logs | 534.1 | 533.6 | -0.1% | -5.5% .. +1.9% | neutral | +0.0% | 336.44 | 1.59x | | ast-interp | 538.1 | 536.8 | -0.2% | -1.7% .. +2.3% | neutral | +0.0% | 324.66 | 1.65x | | records-sort | 1798.6 | 1749.9 | -2.7% | -4.8% .. +0.5% | neutral | +0.0% | 326.49 | 5.36x | | template-render | 379.2 | 371.0 | -2.1% | -9.3% .. +1.0% | neutral | +0.0% | 287.51 | 1.29x | | functional-pipeline | 608.5 | 596.7 | -1.9% | -3.9% .. +3.2% | neutral | +0.0% | 390.16 | 1.53x | | async-pipeline | 250.3 | 35.9 | -85.7% | -86.2% .. -85.3% | faster | -0.0% | 123.43 | 0.29x | | word-graph | 462.1 | 461.7 | -0.1% | -3.3% .. +2.4% | neutral | +0.0% | 690.81 | 0.67x | | alloc-trees | 690.4 | 699.1 | +1.3% | -0.9% .. +4.1% | neutral | +0.0% | 174.67 | 4.00x | | numeric-kernels | 176.0 | 172.8 | -1.8% | -3.4% .. +0.6% | neutral | +0.0% | 176.9 | 0.98x | | validate-errors | 304.4 | 301.0 | -1.1% | -4.9% .. +1.6% | neutral | +0.0% | 945.13 | 0.32x | | csv-numbers | 602.8 | 605.0 | +0.4% | -4.7% .. +1.9% | neutral | +0.0% | 696.66 | 0.87x | | log-summary | 332.9 | 332.6 | -0.1% | -4.0% .. +1.4% | neutral | +0.0% | 384.79 | 0.86x | | inventory-report | 601.0 | 583.5 | -2.9% | -6.5% .. +5.6% | neutral | +0.0% | 585.31 | 1.00x |