Improve runtime feature detection (and performance)

[onethumb]

The Problem

By relying heavily on compile-time feature detection, rather than runtime feature detection, the library was more fragile (leading to bugs like #14) and unable to gracefully degrade across CPU architecture variants with a single build.

The Solution

Rely on runtime feature detection (out of the hot path) to determine which hardware acceleration target to use, enabling graceful degradation across CPU types with a single build, and minimizing the risk of a SIGILL or similar bug sneaking in.

As a side benefit, AWS Graviton targets are ~36% faster and peak at ~53GiB/s (Graviton4).

Changes

• Adds a feature detection mechanism at instantiation time to determine which acceleration target is ideal.
• Removes much of the compile-time detection.
• Enables a single binary build to gracefully degrade for older CPUs in the same family.
• Improves Graviton4 performance by ~36% to ~53GiB/s.
• Improves file structure and directory layout to isolate each acceleration target further.
• Adds a benchmarking flag to the checksum utility to enable easier benchmarking using a single binary, rather than a source checkout, across platforms.
• Updates cargo packages to latest supported.

Planned version bump

• Which: MINOR
• Why: non-breaking new functionality

Links

• SIGILL on 1.4 with machine that doesn't support avx512 #14

[Copilot]

Pull Request Overview

This PR adds comprehensive x86 architecture support (32-bit) and reorganizes CRC-32 fusion implementations to improve code maintainability across multiple architectures (x86, x86_64, and aarch64). The main changes involve:

• Adding 32-bit x86 support for CRC calculations with native CRC32C and PCLMULQDQ instructions
• Restructuring the codebase to separate architecture-specific implementations into dedicated modules
• Adding extensive integration tests for benchmark functionality
• Implementing a feature detection system for optimal hardware acceleration selection
• Adding future-proof tests for CRC key storage functionality

Reviewed Changes

Copilot reviewed 38 out of 44 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/checksum_integration_tests.rs	New integration tests for benchmark flag parsing and various input scenarios
src/test/mod.rs	Renamed module reference from future_proof to future_proof_tests
src/test/future_proof_tests.rs	Extensive new tests for CRC key storage bounds checking and backwards compatibility
src/lib.rs	Added x86 support to conditional compilation, new feature_detection module, enhanced documentation
src/feature_detection.rs	New comprehensive feature detection system with performance tier selection
src/crc32/mod.rs	Extended fusion support to include x86 (32-bit) architecture
src/crc32/fusion/mod.rs	Simplified architecture dispatching with cleaner conditional compilation
src/crc32/fusion/x86/mod.rs	New x86-specific CRC implementation with AVX512 and SSE support
src/crc32/fusion/x86/iscsi/*.rs	Multiple new files implementing SSE, AVX512 PCLMULQDQ, and VPCLMULQDQ variants
src/crc32/fusion/aarch64/mod.rs	Reorganized aarch64 implementation into separate sub-modules
src/crc32/fusion/aarch64/iscsi/*.rs	Split iSCSI implementations into PMULL and PMULL+SHA3 variants
src/crc32/fusion/aarch64/iso_hdlc/*.rs	Split ISO-HDLC implementations into PMULL and PMULL+SHA3 variants

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.