asmjson

A fast JSON parser that classifies 64 bytes at a time using SIMD or portable SWAR (SIMD-Within-A-Register) bit tricks, enabling entire whitespace runs and string bodies to be skipped in a single operation.

Quick start

use asmjson::{parse_to_tape, choose_classifier, JsonRef};

let classify = choose_classifier(); // picks best for the current CPU
let tape = parse_to_tape(r#"{"name":"Alice","age":30}"#, classify).unwrap();

assert_eq!(tape.root().get("name").as_str(), Some("Alice"));
assert_eq!(tape.root().get("age").as_i64(), Some(30));

For repeated parses, store the result of choose_classifier in a static once cell or pass it through your application rather than calling it on every parse.

Optimisation tips

TapeRef is a plain Copy cursor — two usizes — so it is cheap to store and reuse. Holding on to a TapeRef you have already located lets you skip re-scanning work on subsequent accesses.

Cache field refs from a one-pass object scan

get(key) walks the object from the start every time it is called. If you need several fields from the same object, iterate once with object_iter and keep the values you care about:

use asmjson::{parse_to_tape, choose_classifier, JsonRef, TapeRef};

let classify = choose_classifier();
let src = r#"{"items":[1,2,3],"meta":{"count":3}}"#;
let tape = parse_to_tape(src, classify).unwrap();
let root = tape.root().unwrap();

// Single pass — O(n_keys) regardless of how many fields we need.
let mut items_ref: Option<TapeRef> = None;
let mut meta_ref:  Option<TapeRef> = None;
for (key, val) in root.object_iter().unwrap() {
    match key {
        "items" => items_ref = Some(val),
        "meta"  => meta_ref  = Some(val),
        _ => {}
    }
}

// Subsequent accesses go straight to the cached position — no re-scan.
let count = meta_ref.unwrap().get("count").unwrap().as_i64();
assert_eq!(count, Some(3));

Collect array elements for indexed or multi-pass access

array_iter yields each element once in document order. Collecting the results into a Vec<TapeRef> gives you random access and any number of further passes at zero additional parsing cost:

use asmjson::{parse_to_tape, choose_classifier, JsonRef, TapeRef};

let classify = choose_classifier();
let src = r#"[{"name":"Alice","score":91},{"name":"Bob","score":78},{"name":"Carol","score":85}]"#;
let tape = parse_to_tape(src, classify).unwrap();
let root = tape.root().unwrap();

// Collect once — O(n) scan.
let rows: Vec<TapeRef> = root.array_iter().unwrap().collect();

// Random access is now O(1) — no re-scanning.
assert_eq!(rows[1].get("name").unwrap().as_str(), Some("Bob"));

// Multiple passes over the same rows are free.
let total: i64 = rows.iter()
    .filter_map(|r| r.get("score").and_then(|s| s.as_i64()))
    .sum();
assert_eq!(total, 91 + 78 + 85);

Output formats

• parse_to_tape — allocates a flat Tape of tokens with O(1) structural skips.
• parse_with — drives a custom JsonWriter sink; zero extra allocation.

Classifiers

The classifier is a plain function pointer that labels 64 bytes at a time. Three are provided:

Use choose_classifier to select automatically at runtime.

Benchmarks

Measured on a single core with cargo bench against 10 MiB of synthetic JSON. Comparison point is sonic-rs (lazy Value, AVX2).

Each benchmark measures parse + full traversal: after parsing, every string value and object key is visited and its length accumulated. This is necessary for a fair comparison because sonic-rs defers decoding string content until the value is accessed (lazy evaluation); a parse-only measurement would undercount its work relative to any real use-case where the parsed data is actually read.

asmjson zmm tape leads across all three workloads. It writes a flat TapeEntry array in the assembly parser itself — one pointer-sized entry per value — so structural traversal is a single linear scan with no pointer chasing. The baseline asmjson zmm parser also leads on string-dominated workloads; the portable u64 SWAR classifier is neck-and-neck with sonic-rs on string arrays despite using no SIMD instructions, and beats it on string objects. sonic-rs narrows the gap on mixed JSON through its lazy string decoding, but zmm tape still leads by 90 %.

Conformance note

asmjson is slightly permissive: its classifier treats any byte with value < 0x20 (i.e. all C0 control characters) as whitespace, rather than strictly the four characters the JSON specification allows (0x09 HT, 0x0A LF, 0x0D CR, 0x20 SP). Well-formed JSON is parsed identically; input that embeds bare control characters other than the four legal ones will be accepted where a strict parser would reject it.

License

MIT — see LICENSE.

For internals documentation (state machine annotation, register allocation, design decisions) see doc/dev.md.