Performance Benchmarks

SuperJS compiler throughput on real hardware. Numbers below are measured, not estimated.

Environment: macOS arm64, Node v24.14.1, 10 CPU cores — 2026-06-24.

Reproduce locally:

cd superjs && pnpm nx build compiler && pnpm nx build lsp
cd .. && node scripts/gen-bench-corpus.mjs
node scripts/bench.mjs              # compile metrics → benchmarks/results.json
node --expose-gc scripts/bench-lsp.mjs   # LSP metrics (merge into results.json)

Stage 6 targets

Metric	Target	Measured (corpus)	Status
Cold compile (~10k LOC)	≤ 2s	81.5 ms	✓
Warm rebuild (avg 5×)	≤ 100 ms	56.5 ms	✓
LSP idle memory (100k LOC)	≤ 250 MB	195.8 MB	✓
LSP P99 hover	≤ 200 ms	13.2 ms	✓

Compile time

Synthetic corpus: ~14k LOC of functions, sum types, and match arms (benchmarks/corpus-10k.sjs). Stdlib corpus: all superjs/libs/stdlib/src/modules/*.sjs (382 LOC, 11 files).

Input	LOC	Cold compile	Warm (avg 5×)
stdlib	382	19.9 ms	3.7 ms
synthetic corpus	14,000	81.5 ms	56.5 ms

vs TypeScript tooling (same synthetic TS corpus)

Comparisons use an equivalent TypeScript file (benchmarks/corpus-10k.ts) with the same function count and control flow (discriminated unions instead of SJS match). This is not apples-to-apples: SuperJS runs parse + typecheck + IR + codegen; tsc is typecheck-only; esbuild is transpile-only.

Tool	Wall time (cold)	Relative to SJS cold
SuperJS `compile()`	81.5 ms	1.00×
tsc `--noEmit`	720.7 ms	8.8× slower than SJS
esbuild transpile	431.0 ms	5.3× slower than SJS

SJS cold compile is ~0.11× the wall time of tsc and ~0.19× esbuild on this corpus. Your mileage will vary with module graph size, import resolution, and source maps.

LSP

Single-file workspace: ~140k LOC synthetic corpus (genSjsCorpus(100_000)), one textDocument/didOpen, memory budget 512 MB.

Metric	Value	Target	Status
Open + index (cold)	735.5 ms	—	—
Idle heap after open	195.8 MB	≤ 250 MB	✓
P50 hover latency	5.6 ms	—	—
P99 hover latency	13.2 ms	≤ 200 ms	✓

Harness: scripts/bench-lsp.mjs (200 hover samples at a fixed identifier position). Run with --expose-gc for stable heap readings.

See also: LSP Memory Audit — didClose release and LRU eviction checks.

Methodology

Cold: first compile() call after process start, no persistent cache.
Warm: average of 5 subsequent compile() calls on identical sources (in-memory incremental session).
Harness: scripts/bench.mjs (compile), scripts/bench-lsp.mjs (LSP)
Corpus generator: scripts/gen-bench-corpus.mjs
Committed results: benchmarks/results.json

Per-phase lexer/parser/checker/codegen breakdown is not yet instrumented in the compiler pipeline.

Last updated: 2026-06-24.