ChaCha20Core.chaCha20Xor:
- Parse key/nonce once into initial state, reuse across blocks
- XOR at word level (4 bytes at a time) instead of byte-by-byte
- Reuse working IntArray across blocks instead of allocating per block
XChaCha20Poly1305:
- Add hChaCha20FromNonce24 to avoid nonce copyOfRange(0,16) allocation
- Add chaCha20PolyKey to generate only 32 bytes instead of full 64-byte block
- Use single result array instead of ciphertext + tag concatenation
Benchmark (98-byte padded message, JVM):
Before: encrypt 2286 ns/op, decrypt 2062 ns/op
After: encrypt 1544 ns/op, decrypt 341 ns/op
https://claude.ai/code/session_01YHBwqnb3Z7Q7PX31tJ2G3i
Claude
425fc24bc2