formalize component value definitions

[rvolosatovs]

I used

component-model/design/mvp/Binary.md

Lines 363 to 367 in 145752a

	namesubsection_N(B) ::= N:<byte> size:<u32> B (if size == |B|)
	componentnamesubsec ::= namesubsection_0(<name>)
	sortnamesubsec ::= namesubsection_1(<sortnames>)
	sortnames ::= sort:<sort> names:<namemap>

,

component-model/design/mvp/Binary.md

Line 265 in 145752a

canon ::= 0x00 0x00 f:<core:funcidx> opts:<opts> ft:<typeidx> => (canon lift f opts type-index-space[ft])

as main points of reference for the initial approach.

Tried to stay as close as possible to core Wasm spec and the canonical ABI spec.

• flags are encoded as variable-length uN bitmask - this does favor cases where flags with lowest indexes are set, however I feel like this is a reasonable approach to take at least for consistency with all other numerical values
• enum and variant discriminants are encoded as variable-length u32-s, same argument applies as above. I do feel like for most use cases the discriminant should fit in a single byte
• results and options match canonical ABI despecialization rules (https://github.com/WebAssembly/component-model/blob/main/design/mvp/CanonicalABI.md#despecialization)
  • 0x00 is used to identify result::ok
  • 0x01 for result:error
  • option::none is 0x00
  • option::some is 0x01
• tuples and records are encoded as just the sequences of their field/element values, in order
• valtype is specified once in the top-level (value t:<valtype> v:<val(t)>), because, although type inference should be possible in most use cases by looking at the usages of the value index, values may potentially be unused by the component directly (for example, exported). I feel like it's a very little price to pay for simpler implementation and type checking.

val encoding reference implementation is available at https://github.com/rvolosatovs/wrpc/blob/4fd7488caa4cff19079f851cd7d0efda8f613a64/crates/transport/src/lib.rs#L2557-L3223

Refs #15

[lukewagner]

Great start, this is looking good.

[lukewagner]

Nice work! This looks good to me, but I think @sunfishcode has also thought a lot about how to express component-model values in a similar context where we are not inferring the types from the value, but rather checking that they match a given target type, so I'd appreciate his review too.

[lukewagner]

Great, thanks a bunch for all the changes (again). Just a few nits:

[primoly]

Should those core:u8 and core:s8 be core:byte instead?

[rvolosatovs]

I think this is ready for the next review round

[lukewagner]

Awesome, thanks again! I think that addresses all the feedback I'm aware of. I'll merge at the end of the week if nothing else pops up.