tree a1c85a1a98938a2eee519fd6e7a8f25fdcfa0466
parent 88c0ec71086d4060687afb7c75c64f6ae24502f0
author birenroy <birenroy@google.com> 1786483081 -0700
committer Copybara-Service <copybara-worker@google.com> 1786483102 -0700

Refactors `HuffmanBitBuffer::AppendBytes` with a 64-bit word read.

Uses a 64-bit unaligned load (`quiche::QuicheEndian::HostToNet64`) when at least
8 input bytes are available to refill `HuffmanBitBuffer::accumulator_` in a
single 64-bit operation rather than a scalar byte-by-byte loop.

Based on benchmark runs against the previous state, here is how this optimization compares:
- **Overall Performance (`geomean`)**: Significant overall throughput improvement of **+22.71%** (from `100.4 MiB/s` to `123.2 MiB/s`).
- **Broad Improvements**: Every single benchmark demonstrated a positive throughput gain, with the vast majority showing improvements between **+23%** and **+28%**.
- **Top Performers**:
  - `BM_DecodeFastPath7Bit/64k`: **+27.93%**
  - `BM_DecodeFastPath7Bit/1k`: **+27.63%**
  - `BM_DecodeSlowPath/1k`: **+26.23%**
  - `BM_DecodeSlowPath/64k`: **+25.96%**
  - `BM_DecodeSmallStrings`: **+25.69%**

This optimization was generated by Gemini.

Protected by FLAGS_gfe2_reloadable_flag_hpack_huffman_decoder_optimizations.

PiperOrigin-RevId: 962996201