tree 418d8640bd12c56f0a1014000c43ed09f554b60f
parent 119ebdf7496e51d583789d4e5a5d6964bb4aca7a
author bnc <bnc@google.com> 1601994157 -0700
committer Copybara-Service <copybara-worker@google.com> 1601994198 -0700

Improve http2::HuffmanEncodeFast() by skipping rest of iteration if possible.

While I originally thought that a loop without branches results in fastest
encoding, benchmarks show that skipping to the next iteration if the Huffman
code has ended actually speeds up encoding.  This is probably due to the fact
that most codes are short (by design of the Huffman code).

Of course this encoding algorithm would be a lot simpler if it was only for
64-bit big endian platforms, as a single |= instruction would do the job.

PiperOrigin-RevId: 335634321
Change-Id: Ia95c52b15c3d50600e40cd0beef914773bf2c068