faster-hex

Releases

1. ▸▾1.0.023 Sept 2026

   The 1.0 API decodes the complete input, returns the written output slice, and leaves destinations unchanged on checked errors. It removes the public u

   The 1.0 API decodes the complete input, returns the written output slice, and leaves destinations unchanged on checked errors. It removes the public unchecked and backend functions and reports precise length and invalid-byte errors.

   This release adds array and Vec decoding, borrowed Hex formatting, String reuse, optional heapless support, Serde array and bounded adapters, AArch64 NEON encoding/decoding, and AVX-512BW validation/decoding. The 0.10.1 safety fixes are included.

   Requires Rust 1.95.0. See MIGRATION.md for the API changes. CI covers the supported feature combinations and platforms, Miri, libFuzzer/AFL, native x86 backends, and the 1.0 API baseline.

   Open source →

   Additional notes

   Changes since 0.10.1. See MIGRATION.md for upgrading existing code.

   Breaking changes

   • Require Rust 1.95.0. Edition 2018 and the default std, serde features are retained.
   • Decode the complete input and return the written prefix as Result<&mut [u8], Error>. A short destination returns an error; extra capacity remains untouched. Odd input is rejected, including when the destination is empty.
   • Make Error non-exhaustive. Replace InvalidLength(n) with OutputTooSmall { required, .. }; add OddLength and byte positions to InvalidChar { index, byte, .. }. Error formatting changes.
   • Remove public backend, fallback and unchecked functions, and the deprecated hex_to.
   • Make heapless support opt-in through heapless-08 and the heapless_08 module. hex_string and hex_string_upper require alloc and always return String.

   Added

   • Export CheckCase and hex_decode_with_case for case-specific decoding.
   • Add hex_append and hex_append_upper for reusing a String allocation.
   • Add exact hex_decode_array::<N> and owned hex_decode_vec decoding, each with a case-policy variant. Array length errors use LengthMismatch { expected, actual, .. }.
   • Add borrowed Hex formatting through core::fmt, including integer-style flags.
   • Add Serde Option policies, fixed-array adapters and deserialize_bounded helpers.
   • Add AArch64 NEON encoding, validation and decoding.
   • Add AVX-512BW validation and decoding with runtime CPU/OS detection.
   • Add defmt-03 support and implement Hash for Error and CheckCase.

   Fixed

   • Preserve the entire destination on checked slice conversion errors.
   • Keep SIMD accesses within their slices at short lengths and independent alignments.
   • Check supported CPUID leaves and OS AVX/ZMM state before selecting AVX2 or AVX-512.
   • Preserve Serde's Some tag in binary formats and read each serialized AsRef<[u8]> view once. Existing JSON prefix and case rules remain unchanged.
   • Support dependency-free core builds and Serde with alloc but without std. Implement core::error::Error with every feature set.

   Performance and validation

   • Refine short-input SIMD decoding, SSE4.1 conversion, AVX2 validation and scalar paths.
   • Reuse the existing NEON block for 4-byte array decoding.
   • Keep 65–96-byte x86 slice decoding in registers until validation succeeds.
   • Decode large owned arrays and vectors in one checked pass using existing SIMD blocks; discard private output on error and retain exact error positions.
   • Encode strings and Serde output directly into uninitialized storage; borrow Serde input and use stack storage for serialized fixed values through 65 bytes, including their prefix.
   • Avoid clearing unused Hex scratch storage, stream long values in larger chunks, and reduce short-value formatting and padding overhead.
   • Add differential tests, guard pages, feature consumers, Miri checks and fuzz targets.
   • Exercise independent decode lengths at guard pages and run 32-bit x86 tests.
   • Add explicit native AVX-512 fuzz acceptance for libFuzzer and AFL, requiring execution inside every AVX-512 kernel after corpus minimization.
   • Check structured valid and damaged inputs, allocation-free contracts, concurrent dispatch, fixed-capacity boundaries and partial writer failures.
   • Expand CI across operating systems, architectures and feature combinations, with Wasm/embedded/32-bit builds and native x86 backend checks.
   • Keep focused Criterion suites for slice and owned conversion, validation, capacity reuse, borrowed formatting and Serde. Development aliases support saved baselines and competitor comparisons; CI executes the cases without collecting timings.