evcholesky / moebetter Goto Github PK

An experimental sketch version of the Moe compiler. (I'm trying to take the patient apart, rearrange it and put it back together without killing it)

• Recover from errors rather than aborting on first one
• Query based compilation
• Multi-threaded

• String -> Using interred strings, just check pointer rather than HV, threads will need to use mutex when adding to the hash

• SLexerLocation -> LexerSpan; should always be passed as const &

• All types should be unique from the start, don't set an AST node's type until it can be uniquely determined

• Collapse TypeInfoInt and TypeInfoFloat to TypeInfoNumeric

• No leading S/C for structs or classes
• EWC macros/namespaces changed to MOE
• No more pCoz/pChz, all string code should handle UTF8

• Alloc -> not using arena based loading - this would be nicer, but I don't think it will work with caching query results

• Removing acast
• Removing for_each
• Using unnamed types require a : before the type

• Each thread has write access to the AST in it's working set (parser=file, typecheck=WorkspaceEntry) and any cross set searches (symbol lookup, registry) needs to be guarded [ ] change the type creation routines to go through a typeFactory that will guarantee uniqueness and control thread access

[ ] symbols should be defined during parse and then symbol tables should be locked

• Not all files will parse before we start type checking, this means you may need to wait for a symbol to be defined (not just waiting for defined symbols to parse) This is especially bad for function overloading

• "array references" -> array slice
• unify Proc/Procedure naming - everything should be proc (I think it's just TypeInfoProcedure)
• fix error recovery stack to synch with multiple error outcomes ( ie. failure to parse expression list should recover to ',' or recover to '}'
• track nested constructs in error recovery?
• nested block comments
• Get rid of errep passing, (just add a hide error count to ErrorManager?)

• add module support