Determine value in moving type construction to compiler

[stevehalliwell]

Presently the compiler emits bytecode that the vm uses to construct classes at runtime. Would it be possible and beneficial to move that responsibility prior to runtime. This would free up a number of opcodes and simplify the vm, mostly it might push us towards a more typed feature set.

A few opcodes are used for types, Type, Field, Method, Mixin. But these are only used once in most cases, to generate the runtime class object, not in the creation on instances.

[stevehalliwell]

Having bytecode to create classes rather than compiler doing all of it, allows for bytecode serialisation being the only requirement for preventing recompile at every run. Presently, I'd prefer tiny recompile times over cached bytecode chunks.

[stevehalliwell]

class elements

• vars
  • init chains
  • field names
• methods
  • funcs with a this
  • method names
• mixin
  • would require the other type is already compiled, slightly more order restrictive than current
• signs
  • requires the other type is already compiled, is now a compile time check not runtime
• init
  • just a special method

The path forward to trying this is get us to a stage where we could emit type info after compile but before vm run. Once that worked replace.

[stevehalliwell]

Having started some of this the next hurdle is closures. init chains etc. were using closures they run in not just chunk and label/location.

From testing they can just use chunk, nothing we have with test coverage or in current projects uses upvals in methods or init chains. They should be using labelIDs though, but appear to use instruction index at the moment