[DicksonWu654]

HuffmanEncoder.encode() appends each symbol to a single Python integer (final_num <<= length; final_num |= code). Every step copies the whole integer, so encoding is quadratic in the value length, and long values such as cookies dominate encode time.

For bytes values of 256 bytes or more, this builds a small map from each distinct input byte to its code written as a string of 0/1 digits, joins those strings, pads with EOS (1) bits to a whole byte and converts once with int(bits, 2).to_bytes(...). The map is rebuilt on every call from the encoder's own huffman_code_list / huffman_code_list_lengths, so nothing is cached between calls and changes to those tables are still picked up.

Everything else keeps the existing loop: values shorter than 256 bytes, bytes subclasses, HuffmanEncoder subclasses, and code tables that are not plain lists of ints or contain a code with a length outside 1..30 or a value that does not fit its length (the loop's masking/overlap behaviour for such tables is not reproduced, so they fall back). The output is identical in all cases.

Benchmarks

Each workload runs header blocks through Encoder.encode() and, for *_roundtrip, Decoder.decode(). browser_mixed alternates 16 browser-like requests and 16 responses; repeated_requests sends one request 16 times; long_cookie is 16 requests whose cookie is about 1.2 KB; no_huffman_encode is browser_mixed with huffman=False; fixture_stories_roundtrip replays the first four stories of the hpack-test-case fixtures in tests/.

master and this branch were imported side by side in the same interpreter and timed in alternating ~20 ms CPU-time blocks (40 rounds in each of 6 processes, pinned to one core). The table shows the median per-round speedup with a 95% bootstrap interval. The machine was shared with other work, so the absolute times are only indicative.

Workload	CPython 3.12.3 master → branch (µs)	CPython 3.12.3 speedup	CPython 3.14.8 speedup
browser_mixed_encode	1,780.1 → 1,775.1	1.00x [0.98, 1.03]	1.00x [0.98, 1.03]
browser_mixed_roundtrip	3,755.1 → 3,640.0	1.02x [1.00, 1.05]	1.00x [0.98, 1.01]
repeated_requests_encode	437.5 → 428.4	1.01x [0.99, 1.03]	0.98x [0.96, 1.01]
repeated_requests_roundtrip	987.8 → 1,002.5	0.99x [0.97, 1.01]	1.03x [1.01, 1.05]
responses_decode	685.0 → 686.8	1.01x [0.99, 1.03]	1.01x [0.98, 1.02]
long_cookie_encode	7,502.8 → 2,401.4	3.14x [3.09, 3.20]	3.22x [3.17, 3.28]
long_cookie_roundtrip	13,269.7 → 7,809.5	1.70x [1.67, 1.73]	1.68x [1.66, 1.70]
no_huffman_encode	1,071.6 → 1,041.1	1.01x [1.00, 1.04]	1.00x [0.99, 1.02]
fixture_stories_roundtrip	2,103.3 → 2,123.9	0.98x [0.96, 1.00]	1.00x [0.98, 1.02]

Only values of 256 bytes or more take the new path, so the other workloads are unchanged within noise.

Behaviour

• New tests: mutating the code tables between calls; fallback for over-wide, zero-length, int-subclass, list-subclass and short tables, with the exact bytes the existing loop produces.
• pytest passes with 100% statement and branch coverage on CPython 3.11, 3.12, 3.13 and 3.14; ruff check src/ and mypy src/ (the lint tox env) pass.
• I also compared master and this branch on 1,500 random encoder/decoder sessions per interpreter (CPython 3.11 and 3.14): header lists with values of 0 to 20,000 bytes (text, random bytes and repeated bytes), sensitive headers, str input, table size changes, Huffman on and off, plus direct HuffmanEncoder calls. The encoded bytes, decoded headers, exceptions and, with DEBUG logging enabled, every log record were identical.

🤖 Generated with Claude Code