simdxml

SIMD-accelerated XML parser with full XPath 1.0 support.

Uses a two-pass structural indexing architecture (adapted from simdjson) to parse XML into flat arrays instead of a DOM tree, then evaluates XPath expressions against those arrays with cache-friendly access patterns.

Performance

Benchmarked on Apple M4 Max (NEON) and AMD Ryzen 9 3950X (AVX2) against pugixml (C++), libxml2 (C), and others.

Where simdxml wins — large files and attribute-dense XML:

Where it's close — text-heavy XML, simple queries:

pugixml's single-pass DOM build has lower overhead for queries where parse time dominates. simdxml's advantage grows with file size and attribute density.

Library parse throughput vs Rust parsers (M4 Max, criterion):

simdxml is consistently faster than quick-xml (a streaming pull parser) and significantly faster than roxmltree (a DOM builder). The gap widens on attribute-dense XML where SIMD quote masking avoids per-character scanning. Unlike quick-xml, simdxml produces a queryable structural index — not just an event stream.

Quick Start

use simdxml::{parse, CompiledXPath};

let xml = b"<library><book><title>Rust</title></book></library>";
let index = parse(xml).unwrap();

// One-shot query
let titles = index.xpath_text("//title").unwrap();
assert_eq!(titles, vec!["Rust"]);

// Compiled query (reusable across documents)
let query = CompiledXPath::compile("//title").unwrap();
let titles = query.eval_text(&index).unwrap();
assert_eq!(titles, vec!["Rust"]);

Features

• SIMD structural indexing — NEON (aarch64), SSE4.2/AVX2 (x86_64), scalar fallback
• Full XPath 1.0 — all 13 axes, predicates, functions, operators. 327/327 libxml2 conformance, 1015/1023 pugixml conformance
• Flat-array index — ~16 bytes/tag vs ~35 for a DOM tree. Cache-friendly for repeated queries
• Parallel parsing — split large files across cores with parallel::parse_parallel
• Lazy indexing — only build structures needed by the query
• Bloom filter prescan — skip files that can't match before parsing
• Persistent indices — serialize to .sxi files, mmap for instant reload
• Zero-copy — borrows input bytes, no allocation for tag names or text content

XPath Support

All XPath 1.0 expressions work, including:

// Location paths
index.xpath("//book/title")?;
index.xpath("//book[@lang='en']/title")?;
index.xpath("//book[position() < 3]")?;

// All 13 axes
index.xpath("//title/ancestor::library")?;
index.xpath("//book/following-sibling::book")?;

// Functions
index.eval("count(//book)")?;          // -> Number(42)
index.eval("string(//title)")?;        // -> String("Rust")
index.eval("contains(//title, 'Ru')")?; // -> Boolean(true)

// Unions
index.xpath("//title | //author")?;

Batch Processing

Process many documents with compiled queries and automatic parallelism:

use simdxml::{parse, batch, CompiledXPath};

let docs: Vec<&[u8]> = vec![
    b"<r><claim>First</claim></r>",
    b"<r><claim>Second</claim></r>",
];
let query = CompiledXPath::compile("//claim").unwrap();

// Parallel batch with bloom filter prescan
let results = batch::eval_batch_text_bloom(&docs, &query).unwrap();

CLI

The sxq command-line tool provides fast XML/XPath queries from the terminal:

sxq '//title' book.xml
sxq -c '//claim' patents/*.xml       # count matches
sxq -r '//path' drawing.svg          # raw XML output
sxq -j '//author' papers/*.xml       # JSON output
cat data.xml | sxq '//record/@id'    # stdin
sxq 'count(//article)' dblp.xml      # scalar expressions
sxq info large.xml                   # structural statistics

Install: cargo install simdxml-cli

Platform Support

License

MIT OR Apache-2.0