Eyevinn / mp4ff / 36988755743

82%

master: 83%

Pull #605

github

perf(mp4): decode fragments with fewer allocations

Every box header made a string of its four-byte type, and decoding then
looked the string up in the decoder map. A table keyed by the type bytes as a
big-endian uint32 now returns both the type string and the decoder in one
lookup, so a box type with a decoder costs no allocation and one cheaper hash.
The table is built from decodersSR and kept in step by SetBoxDecoder and
RemoveBoxDecoder. DecodeHeader, DecodeHeaderSR and BoxSeekReader.PeekBoxHeader
all use it, and PeekBoxHeader reads headers into scratch arrays instead of new
slices.

moof and traf decoding copied their decoded children into a second slice;
AddChild appends each child once, in order, so the first slice is refilled in
place. A fragment's children slice is sized for moof and mdat up front.

On testdata/1.m4s cut into 60 one-sample fragments, decoding a fragment makes
14 instead of 24 allocations with DecodeFileSR and 16 instead of 28 when
streaming, and is 13-24% faster.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

41 of 46 new or added lines in 6 files covered. (89.13%)

22962 of 27835 relevant lines covered (82.49%)

0.83 hits per line