Compiler-input stress benchmark

python/benchmarks/stress_instance_benchmark.py generates deterministic MLIR modeled on a production compiler dump. The input has forward-referenced locations, a 100 KiB block-argument line, shallow fan-out, varied symbols, representative stir.* forms, both type trailers, and hexadecimal floats.

The benchmark times two registries against the same input. The lambda-only registry recognizes stir.lambda. The full registry recognizes every generated stir.* form. Each measurement gets one warm-up run, then three timed runs. Input generation happens before timing starts.

The Python traversal measurement parses the file and visits every operation through File.operation(index), counting syntax errors and checking that all operations were visited.

uv run maturin develop --release
uv run python python/benchmarks/stress_instance_benchmark.py --smoke
uv run python python/benchmarks/stress_instance_benchmark.py
uv run python python/benchmarks/stress_instance_benchmark.py --full

The default is 16 MiB; --smoke uses 2 MiB and --full uses 268 MiB. Pass --output PATH to keep the input. JSON results on stdout compare parser scaling and registry costs.

Python operation access

File.operation(index) uses a lazy index of syntax operation nodes. The first operation count or lookup builds it; parsing alone does not. The traversal measurement includes that construction cost.