lang-parsing-substrate

A shared language-parsing substrate for static-analysis tools: tree-sitter grammar dispatch, language detection, and a growing set of language-agnostic analysis primitives (import/call graphs, control-flow graphs, structural fingerprinting, suppression comments) built on top of a unified LanguageInfo registry across 16 languages — compiled in at build time via Cargo feature flags.

What's in the substrate

Everything below the registry is deliberately per-file: a module extracts what one parse tree contains, and leaves assembling a corpus-wide graph, dedup report, or coupling metric to the caller. This keeps the substrate's job narrow (one authoritative, correct answer per file) and lets each consumer choose its own storage model (in-memory, SQLite, whatever) without the substrate needing to know about it.

Language coverage varies by module — the registry knows about all 16 languages, but modules like cfg only model the languages they've been built out for. A module never fabricates a result for a language it doesn't support; it returns None (or an empty result) instead of guessing.

Feature flags

Not every consumer needs all 16 languages. Each language is an optional Cargo feature; the all-languages convenience feature (enabled by default) pulls in the full set.

A consumer that only cares about C/C++, for example, would declare:

lang-parsing-substrate = { version = "0.5", default-features = false, features = ["lang-c", "lang-cpp"] }

Usage

# Cargo.toml — full language set (default)
lang-parsing-substrate = "0.5"

# Cargo.toml — C/C++ only
lang-parsing-substrate = { version = "0.5", default-features = false, features = ["lang-c", "lang-cpp"] }

use lang_parsing_substrate::{language_for_file, languages, supported_languages_report};
use std::path::Path;

// Detect language for a file
if let Some(lang) = language_for_file(Path::new("main.c")) {
    let mut parser = tree_sitter::Parser::new();
    parser.set_language(&lang).unwrap();
    // parse...
}

// Enumerate compiled-in languages (reflects feature flags)
for info in languages() {
    println!("{}: {:?}", info.name, info.extensions);
}

// Human-readable summary (for --supported-languages flags)
print!("{}", supported_languages_report());

Grammar crates are re-exported so consumers reach them transitively:

// No direct tree-sitter-rust dependency needed in your Cargo.toml
use lang_parsing_substrate::tree_sitter_rust;

Early exit before parsing

language_for_file trusts the extension. To skip archives, executables, images and oversized files before they reach tree-sitter, classify first — it reads only a bounded prefix, never the whole file:

use lang_parsing_substrate::{classify_file, ClassifyLimits, FileClass};

// The size limit is the caller's policy; the default is no limit.
let limits = ClassifyLimits { max_size: Some(16 << 20), ..Default::default() };
match classify_file(path, &limits)? {
    FileClass::SourceText(_) => { /* parse it */ }
    FileClass::Binary { kind, mime } => { /* skip-and-report, or try anyway */ }
    FileClass::Oversize { size, limit } => { /* caller's call */ }
    _ => { /* Empty, or a future variant */ }
}

The result is a heuristic and can be wrong in both directions (see the classify module docs). Magic numbers come from the dependency-free infer crate. Short ASCII-looking signatures such as MZ or BM only count when the bytes also look binary. classify(path, prefix, size, &limits) is the I/O-free core, for callers that have already read the bytes. The thresholds are calibrated in docs/classify-calibration.md; re-run that check on a new corpus with cargo run --release --example classify_calibrate -- ROOT....

Analysis primitives

use lang_parsing_substrate::{call_edges, import_sources, build_function_cfg, structural_hash};

// Call-graph edges for every named function/macro in a parsed file
let edges = call_edges(tree.root_node(), source);

// Import/use-statement sources, for Ce/Ca coupling metrics
let imports = import_sources(&tree, source.as_bytes(), "rust");

// Control-flow graph for a single function body (c/cpp/rust)
if let Some(cfg) = build_function_cfg(func_node, source, "rust") {
    println!("{} basic blocks", cfg.block_count());
}

// Best-effort lower bound on the C standard a file requires
if let Some(standard) = detect_min_c_standard(&tree, source.as_bytes()) {
    println!("requires at least {standard:?}");
}

API

• languages() -> &'static [LanguageInfo] — compiled-in language set
• language_for_file(path: &Path) -> Option<Language> — grammar dispatch by extension
• language_for_key(key: &str) -> Option<Language> — grammar dispatch by registry key
• language_info_for_file(path: &Path) -> Option<&'static LanguageInfo>
• classify / classify_file / FileClass / BinaryKind / TextEncoding / SourceText / ClassifyLimits — heuristic pre-parse file classification (classify)
• is_source_extension / is_parseable_extension(ext: &OsStr) -> bool — recursive-discovery gates
• supported_languages_report() -> String — human-readable language summary
• LanguageInfo / SlocMode — registry metadata and comment-style enum (drives SLOC calculation)
• find_descendants / find_first_descendant / find_ancestor / node_text and friends — traversal helpers (query)
• import_sources / distinct_import_count — import extraction (imports)
• call_edges / CallEdge / is_function_kind / get_function_name — call-graph extraction (calls)
• build_function_cfg / FunctionCfg / BasicBlock / CfgEdge — control-flow graphs (cfg)
• detect_min_c_standard / CStandard — C standard lower-bound detection (c_standard)
• function_fingerprints / duplicate_groups / Fingerprint / CorpusFingerprint — structural hashing (fingerprint)
• ignored_regions / IgnoredRegion — tools:off/tools:on markers (regions)
• suppressions / Suppression — tools:suppress comments (suppressions)
• dead_code_ranges / DeadCodeRegion / DeadCodeReason — preprocessor dead-code regions, C/C++ only (dead_code)
• swift_dead_code_regions / SwiftDeadCodeRegion — Swift #if/#elseif/#else dead-code regions (dead_code_swift)
• csharp_dead_code_regions / CSharpDeadCodeRegion — C# #if/#elif/#else dead-code regions (dead_code_csharp)
• PathIgnore — compiled glob ignore sets (path_ignore)

Building

cargo build                                          # all languages (default)
cargo build --no-default-features --features lang-c,lang-cpp  # subset
cargo test

Requires no C compiler — tree-sitter grammar crates ship pre-generated C sources and compile via the cc crate.

Python bindings

The pyo3 Cargo feature (off by default) exposes the substrate's language-agnostic analysis primitives as a Python extension module, published to PyPI as prebuilt abi3 wheels (CPython 3.10+, one wheel per platform — see docs/releasing.md):

pip install lang-parsing-substrate

Building it yourself from source uses maturin:

pip install maturin
maturin build --release          # writes target/wheels/lang_parsing_substrate-*.whl
pip install target/wheels/lang_parsing_substrate-*.whl

import lang_parsing_substrate as lps

src = "fn helper(x: i32) -> i32 { x + 1 }\nfn main() { helper(41); }\n"
edges = lps.call_edges("rust", src)          # [CallEdge(caller='main', callee='helper', ...)]
cfg = lps.function_cfg("rust", src, "main")  # FunctionCfg | None
fps = lps.function_fingerprints("rust", src, min_nodes=1)

Python can't hand this crate a tree_sitter::Node/Tree directly — this crate's tree-sitter version has no ABI relationship to tree-sitter's own, separate Python bindings — so every bound function takes (language_key, source), parses internally, and returns owned data (CallEdge, FunctionCfg, Fingerprint, Suppression, IgnoredRegion, LanguageInfo, all plain attribute-holding classes). A consumer that also needs to walk the tree itself (e.g. for domain-specific semantics this crate doesn't model) still parses separately with a language-specific tree-sitter Python package; the bindings here only cover the substrate's own primitives.

invoke build-wheel builds the wheel locally for testing. The actual PyPI release happens in CI on a vX.Y.Z tag push, via Trusted Publishing (OIDC, no API token) — see docs/releasing.md.

License

MIT — see LICENSE.