Benchmarks¶

This report records release-mode latency for the public lowering and extraction paths. It is a reproducible local snapshot, not a cross-machine performance guarantee.

Contents¶

• Environment
• Methodology
• Transpilation results
• Extraction results
• Interpretation
• Reproduction

Environment¶

Methodology¶

• Durations use std::time::Instant around one public API call.
• Five warm-up calls run before distribution sampling.
• Percentiles use sorted nearest-rank samples at indexes derived from (n - 1) * percentile / 100.
• p50 and median are mathematically identical; both columns are retained to match requested release-report terminology.
• Source inputs are wrapped in std::hint::black_box.
• Transpilation uses the committed monty-v0.0.19.json manifest.
• Extraction scans the same exact Monty 0.0.19 source in two forms.
• In-memory ZIP timing excludes the initial archive read from disk and includes validation, decompression, source loading, and scanning.
• Network release resolution/download is not measured because remote latency is not a stable property of the extractor.

Transpilation results¶

Both generated workloads are approximately 100 KiB.

The first cache population calls measured 4.210708 ms for the supported workload and 30.831666 ms for the match-heavy workload. Those are single observations and therefore are not reported as p50/p99 distributions.

Median exact-hit speedups in this run:

An exact cache hit still hashes/looks up the complete source string and clones an Arc; its cost is therefore not constant with respect to source size.

Extraction results¶

The ZIP archive is 2,356,205 bytes. Both operations produce the complete static capability graph.

Local extraction includes filesystem reads. ZIP extraction reads the archive once before sampling, then includes bounded in-memory ZIP validation, decompression, UTF-8 decoding, and static scanning in every timed iteration.

Interpretation¶

• Ordinary supported 100 KiB source lowers in roughly 4.2 ms at p50 on this machine.
• A rule-dense source costs more because match lowering performs multiple evidence-gated edit passes and reinserts/parses generated source.
• Exact-source caching reduces repeated calls to tens of microseconds.
• Static extraction is a release/discovery operation around 0.11–0.12 seconds, not a request-path operation.
• Python transpiler(...) embeds the manifest and performs no extraction.

OS scheduling, thermal state, allocator behavior, source composition, compiler version, and target CPU affect these numbers. Compare results only when the benchmark environment and workload definitions match.

Reproduction¶

Transpilation and cache:

CARGO_INCREMENTAL=0 \
cargo +1.95.0 run --release --locked \
  -p monty-compat --example cache_bench

Extraction from an exact checkout and its GitHub-style ZIP:

CARGO_INCREMENTAL=0 \
cargo +1.95.0 run --release --locked \
  -p monty-compat-extract --example extraction_bench -- \
  /path/to/monty-0.0.19 \
  /path/to/monty-v0.0.19.zip

Benchmark source is versioned in:

• crates/monty-lower/examples/cache_bench.rs
• crates/monty-extract/examples/extraction_bench.rs