Performance

GaleforceCSS is faster than Tailwind v3 by a meaningful margin. Numbers below are what to expect on real workloads.

Real-world baseline

horizon-tailwind-react (2,551 candidates → 706 rules):

	Tailwind 3 CLI	GaleforceCSS CLI	Speedup
Cold build	246 ms	10 ms	23x
Warm import	15 ms	2 ms	7.5x

Synthetic smoke (734 candidates):

	Tailwind 3 CLI	GaleforceCSS CLI	Speedup
Cold build	195 ms	5 ms	38x

Pure compute

The compile loop — no spawn, no I/O — runs in 1.24 ms on the real corpus.

Candidate	Cost
flex (static)	0.34 µs
Unknown utility	0.21 µs
bg-red-500 (value lookup)	0.41 µs
dark:hover:bg-blue-500/50	0.41 µs

Where the gains come from

1. default_theme() cached in OnceLock — was rebuilding ~12k JSON-tree allocations per build. 245 ms → 16.5 ms (15x).
2. Prefix-indexed find_value_utilities — replaces a linear walk over ~150 entries with HashMap<&str, Vec<&'static ValueUtility>>. 16.5 ms → 2.4 ms (6.8x).
3. Rayon parallelization — par_iter over candidates when n ≥ 256. 2.4 ms → 1.2 ms (1.9x).
4. Scanner fast path for explicit file paths — bypass WalkBuilder instances when the Vite plugin pre-expands globs. ~16k files: 2.0 s of overhead removed.
5. Parallel content scan in the CLI — par_iter across 8 cores. 3.9 s → 460 ms on a 16k-file repo.

The hot loop is now ~0.4 µs/candidate for color resolution. Further gains need lower-level work (skipping String allocations in the rule emitter, tighter rayon chunking).

HMR latency

For small edits, the compile loop is rarely the bottleneck:

Phase	Typical
File watch debounce	200 ms (fixed)
Scanner re-walk	2–10 ms
Compile	1–5 ms
IPC round-trip (CLI bridge)	1–3 ms
Vite HMR send	5–20 ms (browser-side)

Levers if you need more:

• Lower watch debounce. Default 200 ms in galeforcecss watch.
• Narrow content globs. Fewer files = faster scan.
• Use createCompileStream() in long-running tools. Amortizes process startup.

Benchmarking

pnpm --filter @galeforcecss/conformance bench                                 # synthetic
pnpm --filter @galeforcecss/conformance bench -- --project /path/to/your/app  # real
cargo bench -p galeforce-compiler                                             # pure compute

Memory

The CLI binary uses mimalloc — ~25% faster on warm-stream compile vs the system allocator. Peak resident on the real corpus is well under 50 MB.