[alex]

This comes out of the rustls provider analysis in #991. That provider seals TLS records with ChaCha20-Poly1305, and against aws-lc-rs on a Cascade Lake–class Xeon it is 0.58–0.84× as fast below 1 KiB. encrypt from two pieces (a TLS 1.3 record's shape) is also 3–7% slower than copying the pieces and calling encrypt_in_place. Neither shows in the benchmarks as they are: they measure 64 B, 1 KiB and 16 KiB only, and encrypt from a single piece. This PR changes only the benchmarks, so the PRs that follow are measured where it matters.

• Every ChaCha20-Poly1305 group measures 16, 64, 256, 512, 768, 1024, 4096 and 16384 bytes.
• chacha20poly1305-encrypt-out-of-place takes its plaintext in two pieces: size - 1 bytes of payload and the content type. aws-lc-rs runs seal_out_of_place_scatter with the content type as its extra_in; OpenSSL encrypts one buffer, as before.
• New group chacha20poly1305-copy-encrypt-in-place: this library copying the two pieces into the output, then encrypt_in_place. It is what the provider does today, and encrypt should be no slower.

On a Sapphire Rapids–class machine (AVX-512 backend), encrypt is 5–9% slower than copy + encrypt_in_place at every size:

bytes	encrypt (2 pieces)	copy + encrypt_in_place
16	229 ns	212 ns
1024	872 ns	830 ns
16384	7.51 µs	6.86 µs

(Minimum of 41 interleaved samples each.)

🤖 Generated with Claude Code

https://claude.ai/code/session_01Ma2fA9rUTPXEEQjYyCMCRw

---

Generated by Claude Code