TODO

• Code cleanup
  • Fork/move source-level stuff (e.g. code-walker? AST-to-source? maybe even AST stuff generally)?
  • Chopping block: BIR-to-CL. Flow? Graph? Liveness? Def-use-chains, dominance, loops, and reaching-definitions at least require more thought. Not sure I know how to write them without an actual use case.
• Tests
  • AST should be reasonably stable now: Fix up cst-to-ast tests
  • Tests for ast-to-bir, more
  • Write up a github workflow to automate
• BIR cleanup
  • Remove typeq (replace w/ understanding of typep calls from client)
  • Remove eq-test if possible (harder because of important (if (eq x nil) y z) => (if x z y) transform - could be done as (if (if x nil t) y z) followed by if-if elimination, but I don't like relying on if-if elimination)
  • come-from redesign - direct succession, make catch links something other than successors
  • "sea of nodes" or weaker control flow. good for LICM and otherwise
  • Come up with a design less susceptible to cache invalidation bugs. Ideally, effortlessly thread safe too
• Backend
  • Port representation selection, primops, deftransform from clasp
  • Provide defaults/etc library for use by clients (or is Typo it already?)
  • Improve representation selection algorithm (it's not very good right now)
  • register-allocation: Check on what nodefunallowed/SICL's been doing
• Abstract interpretation
  • Ideally, replace meta-evaluate with AI
  • Performance evaluation
• Optimizations
  • LICM
  • CSE/GVN
  • DX inference
  • Control-flow analyses w/AI generally (type inference w/variables)
  • SSA? Probably easier than control-flow sensitive AI, but doesn't cover closure variables, but how often can we really do anything with closure variables that wouldn't be covered by local-call info?
  • Non-cartesian AI? (e.g. multivariable type constraints)