overreads

[graphitemaster]

it's documented behavior so not strictly a bug, however you could avoid overreads by by computing the offsets from a table, there's only a few.

[skeeto]

Ah, interesting idea. So if I understand correctly, there will still be four reads, but the overreads could be directed somewhere safe, making the interface a bit nicer. I see two downsides: 1. Those reads are now each dependent on an offset table lookup, which depends on the length. Currently the reads can be performed even before the length is known. 2. In theory, on alignment-relaxed architectures like x86, the compiler could currently coalesce those four reads into a single 4-byte read. Using dynamic offsets would prevent this. However, neither GCC nor Clang currently seem to take this approach anyway.

[graphitemaster]

there's no such thing as a read that is smaller than a cache line anyways, so consuming the first byte to calculate length then using a table (or some masking technique) would be fine, though it's hard to say how it'll perform in practice.