Techniques to beat Arrays.hashCode(byte[]) using Java's own means

10 months ago

Java's Arrays.hashCode(byte[]) computes a 32-bit hash value for byte arrays.
Default implementation in OpenJDK up to version 20 uses a simple loop, which is not highly optimized.
Loop unrolling and vectorization techniques (SWAR and SIMD) can significantly speed up hash computation.
SWAR-based implementation processes 8 bytes at once using 64-bit long integers and bit manipulation.
SIMD-based implementation uses Java's Vector API to process multiple bytes in parallel, outperforming intrinsic implementations.
Benchmark results show SWAR is up to 2.9x faster than default, and SIMD is faster than OpenJDK's intrinsic implementation.
Both approaches handle remaining bytes separately if the array length is not a multiple of the processing chunk size.
The findings suggest OpenJDK's current implementations could be improved with these techniques.

Hasty Briefsbeta