Closed
Description
We could possibly save some time over our structural hashing and equality code if we revamped ty::t to become a packed u64 value in something approaching this form:
- nil, bot, bool, ints, uints, floats, str, self, type, opaque box become constant values
- estr, box, uniq, vec, ptr, rptr become (type tag: u8, mutability flags: u8, vstore/region ID: u16, enclosed type: u32)
- enum, iface, class, res become (type tag: u8, crate ID: u8, substs ID: u16, node ID: u32)
- tup becomes (type tag: u8, tup ID: u32)
- rec becomes (type tag: u8, fieldset ID: u16, tup ID: u32)
- fn becomes (type tag: u8, purity/proto/retstyle flags: u8, constrs ID: u16, signature ID: u32)
- param becomes (type tag: u8, crate ID: u8, parameter ID: u16, node ID: u32)
- var becomes (type tag: u8, vid: u32)
- constr becomes (type tag: u8, constrs ID: u16)
- opaque closure ptr becomes (type tag: u8, closure kind: u8)
We need several side tables:
- A hash table of regions found in the program.
- A table of linearized substs. The table contains subsituted types; at each node is a hash table (or an association list) of self_r and self_ty to ID.
- A fieldset table.
- A table of vstores and regions.
- A table of linearized tuples.
- A table of linearized function types. The return value comes first, followed by the arguments.
- A table of constraints. Doesn't need to be fast; we rarely if ever use constraints.
By "linearized" I mean that the type in question is converted into a list and concatenated into one large string of types encoded in this way. So, for example, a list of 4 types would be 16 bytes long (but note that there are some complications with function parameter modes). This byte string forms the key in the table; thus there is never any need to traverse pointers when doing comparisons for equality, as we simply compare byte strings.
This is very rough, but I think it might speed things up.