On Sun, Jul 04, 2021 at 11:42:03PM +0100, Matthew Wilcox wrote:
> On Sun, Jul 04, 2021 at 11:20:07PM +0100, Gary Guo wrote:
> > This is big endian.
>
> Fundamentally, it doesn't matter whether it's encoded as top-7 +
> bottom-8 or bottom-7 + top-8. It could just as well be:
>
> if (len >= 128) {
> len -= 128;
> len += *data * 256;
> data++;
> }
>
> It doesn't matter whether it's compatible with some other encoding.
> This encoding has one producer and one consumer. As long as they agree,
> it's fine. If you want to make an argument about extensibiity, then
> I'm going to suggest that wanting a symbol name more than 32kB in size
> is a sign you've done something else very, very wrong.
>
> At that point, you should probably switch to comparing hashes of the
> symbol instead of the symbol. Indeed, I think we're already there at
> 300 byte symbols; we should probably SipHash the full, unmangled symbol
> [1]. At 33k symbols in the current kernel, the risk of a collision of
> a 64-bit value is negligible, and almost every kernel symbol is longer
> than 7 bytes (thankfully).
We really should have a better standard varint encoding - open coding varint
encodings in 2021 is offensive, and LEB128 is retarded due to using the high bit
of _every_ byte. Here's the encoding I did for bcachefs, which I nominate for a
standard varint encoding, unless someone knows of a way to do better:
https://evilpiepirate.org/git/bcachefs.git/tree/fs/bcachefs/varint.c
