v0.3.9 | Performance: 20+ Micro-Optimizations — 40% Faster Text Extraction

Performance

• O(n²) string concat fix (#135) — Replaced String::push_str() accumulation in span merging with pre-allocated Vec<&str> joined at the end. Eliminates quadratic growth on pages with thousands of merged spans.

• Image-only content stream parser (#113) — New parse_content_stream_images_only() fast path that only extracts image operators (Do, BI), skipping text and graphics entirely. Used by extract_images() for 3-5× faster image extraction.

• Fingerprint-based font cache (#130) — Font identity determined by hashing encoding+widths+flags instead of comparing full FontInfo structs. Cache hits avoid re-parsing font programs entirely.

• Streaming parser with font identity cache (#136) — Content stream parsing now streams operators instead of collecting into Vec<Operator>. Font identity cache skips redundant font switches when the same physical font is referenced under different names.

• Name-based font set cache (#136) — Cache font set lookups by resource name to avoid repeated dictionary traversal on pages that reference many fonts.

• Fast inline parser for BT/ET (#136) — Specialized inline parser for text blocks that handles the most common operators (Tj, TJ, Tf, Td, Tm, T*) with direct byte matching instead of full nom parsing.

• Font cache trust, fast parser (#137) — Trust cached font data without re-validation on cache hits. Fast parser combines operator dispatch with operand parsing in a single pass.

• Skip Image XObject loading (#138) — Do operator for Image XObjects skips loading the stream entirely during text extraction. Only Form XObjects are loaded and processed.

• Fast-path literal string parsing — Literal strings (...) parsed with direct byte scanning and in-place unescaping instead of nom combinators.

• Byte-to-char lookup table (#144) — 256-entry lookup table for single-byte character mapping replaces HashMap lookups in the hot path.

• Cache font Arc in TjBuffer (#145) — TjBuffer holds an Arc<FontInfo> reference, avoiding repeated cache lookups within the same text block.

• Streaming XObject parser, span flush (#146) — Form XObjects parsed with streaming parser. Span flush moved out of hot loop.

• Width lookup table, XObject text cache (#148) — Fixed-size array for glyph width lookups replaces HashMap. XObject text content cached to avoid re-parsing shared Form XObjects.

• Merged single-pass Unicode decode + width (#149) — Character-to-Unicode conversion and width lookup merged into a single pass, eliminating redundant character code processing.

• Shrink Operator enum 112→40 bytes (#150) — Boxed large variants (String, HashMap, Vec) in the Operator enum to reduce memory footprint by 64%. Improves cache locality during content stream iteration.

• Fast-path single-char operator dispatch (#151) — Single-character operators (q, Q, f, W, n, etc.) dispatched via direct byte match instead of string comparison.

• Switch flate2 to zlib-rs backend (#152) — Replaced default miniz_oxide backend with zlib-rs (Rust port of zlib-ng). 15-25% faster stream decompression.

• Smart Tm batching, TjBuffer precomputation (#153) — Batch consecutive Tm operators (keep only the last one before text). Precompute TjBuffer text metrics to avoid redundant calculations.

• Single-pass TJ array processing (#154) — TJ arrays processed in a single pass, combining string extraction and numeric advance handling without intermediate allocation.

Bug Fixes

• Compilation bug (#134) — Fixed unused variable _page_num → page_num in page renderer (reported by @mpannu03)

• Font encoding with embedded programs (#119) — When /BaseEncoding is absent in a font's encoding dictionary and the font has an embedded program, the font program's built-in encoding is now used as the base encoding per PDF spec ISO 32000-1:2008 Section 9.6.6.1. Previously defaulted to StandardEncoding, causing ligature and glyph mapping failures.

• Supplementary Unicode (U+10000+) (#124) — fallback_char_to_unicode() changed from u16 to u32 parameter, fixing truncation of supplementary Unicode code points (e.g., CJK Extension B characters). Also added 5-6 digit hex support in CMap bfrange array format.

• StandardEncoding + ligature mapping (#125) — Correct glyph-to-Unicode mapping for fi, fl, ff, ffi, ffl ligatures via Adobe Glyph List lookup.

• Kangxi Radical normalization (#123) — Full normalization table (U+2F00–U+2FD5) mapping Kangxi Radicals to CJK Unified Ideographs for searchable text.

• Tolerate misaligned startxref offsets — XRef scanning with 32+64 byte tolerance for PDFs with slightly incorrect startxref pointers.

• Character repetition blowup guard — 32K character cap on text strings to prevent pathological memory growth from malformed content streams.

• Cap Tj/TJ string length — Maximum string length enforced in text operators to prevent memory exhaustion.

• Filter leaked PDF metadata — Strip leaked WhitePoint, BlackPoint, and CalRGB metadata strings from extracted text output.

• TrueType cmap fallback (#127) — Improved cmap table selection for TrueType fonts with multiple subtables. Falls back through format 4 → format 6 → format 0 when primary subtable fails.

• TrueType cmap for subset fonts (#94) — Correct cmap lookup for subset fonts where glyph IDs don't match Unicode code points.

• MacRomanEncoding table (#93) — Fixed MacRomanEncoding character mapping table to match PDF spec Table D.2.

• Indirect refs in /Encoding /Differences (#119) — Resolve indirect object references within encoding dictionaries and /Differences arrays.

• CJK CID-to-Unicode tables (#120) — Expanded CID-to-Unicode mapping tables for Adobe-CNS1, Adobe-GB1, Adobe-Japan1, Adobe-Korea1 character collections.

• RTL text character order (#121) — Right-to-left text (Arabic, Hebrew) now extracted in logical reading order instead of visual glyph order.

• Arabic Presentation Forms normalization (#96) — Arabic Presentation Forms (U+FB50–U+FDFF, U+FE70–U+FEFF) normalized to base Arabic characters for searchable text.

• Multi-column text separation (#122) — Improved column detection using gap analysis to prevent text from adjacent columns being merged.

• Supplementary Unicode in CMap (#124) — CMap bfchar and bfrange entries with 5-8 hex digit targets now correctly produce supplementary Unicode characters.

Features

• extract_all_text() (#128) — New convenience method that extracts text from all pages and joins with newlines. Available in both Rust and Python APIs. (Requested by @SeanPedersen)

• source_role for StructElem (#118) — Structure tree elements now expose source_role preserving the original PDF role name before role mapping. (Requested by @QuickWrite)

CI/CD

• Fixed cargo fmt formatting across all source files
• Resolved 5 clippy warnings: empty doc comments, doc indentation, Box<Collection> on intentional enum shrink, collapsible else-if, needless borrow

Community Contributors

Thanks to everyone who reported issues, sent test PDFs, and requested features for v0.3.9:

• @SeanPedersen — Feature request for extract_all_text() (#128) and continued contributions of real-world test PDFs that drove performance profiling and optimization work
• @mpannu03 — Reported compilation bug (#134) in page renderer
• @QuickWrite — Feature request for source_role on StructElem (#118)

---

📥 Installation

Rust (crates.io)

cargo add pdf_oxide

Python (PyPI)

pip install pdf_oxide

Pre-built Binaries
Download archives for Linux, macOS, and Windows from the assets below.

Platform Support

Platform	Architecture	Archive
Linux	x86_64 (glibc)	pdf_oxide-linux-x86_64-*.tar.gz
Linux	x86_64 (musl)	pdf_oxide-linux-x86_64-musl-*.tar.gz
Linux	ARM64	pdf_oxide-linux-aarch64-*.tar.gz
macOS	x86_64 (Intel)	pdf_oxide-macos-x86_64-*.tar.gz
macOS	ARM64 (Apple Silicon)	pdf_oxide-macos-aarch64-*.tar.gz
Windows	x86_64	pdf_oxide-windows-x86_64-*.zip

Changelog

See CHANGELOG.md for full details.