Make the Bytes struct the size of 3 pointers rather than 4? #799
Description
Currently the Bytes struct is the size of 4 pointers:
Line 101 in 4b53a29
However, we could reduce it to 3 pointers by replacing the vtable pointer with an index into a fixed list of vtables, and packing that index into unused bits of the data (owner) pointer. That would actually simplify the code in some ways, because the KIND_ARC vs KIND_VEC tagging would not be necessary since the pointer and vtable could be atomically replaced together.
As there are fewer than 8 distinct vtables, only 3 bits are needed.
It would be easiest and most portable to use the low bits, which would mean requiring 8-byte alignment. That is easy to require for everything except the Vec backing buffer allocation. However, I believe in practice the default allocator on all or most systems will always return 8-byte aligned buffers, and therefore it may be fine to require that. In the (assumed to be rare) case that a non-8-byte-aligned Vec buffer is received, it can just be promoted to the Shared representation.
Alternatively, on 64-bit systems there are plenty of high bits to use, and the size optimization could be disabled for 32-bit systems (where already it is half the size). This would require keeping the KIND_ARC and KIND_VEC tagging for 32-bit systems, though.
If the maintainers are open to this, I can create a PR.