ideas.txt

Liveness checks:
- a, x, y, flags ; 4 bytes

- always live: sp, memory, ip ; 4 byte hash

8 bytes: instructions (up to 4 instructions)

Algorithm:
1. Write in 0.out the maybe-live registers and
   hash of the always live parts resulting from
   just the initial random machines with no instructions
   added.
2. Load the smallest file yet to be read.
   Sort the file by hash and then cost using a radix sort.
   (Requires ~48 passes of the file assuming a binary radix, but can be done
   completely outside of memory).

   For each sequence in the file:
   - Add each possible instruction to the end of the sequence.
   - If the sequence can be optimized at all, then ignore it and continue.
   - Calculate the new registers and hash.
   - Output the results to the appropriate files based on cost.
   
   Use a merge-sort-like algorithm to find all sequences with the same
   hash from this file and all files with lower cost. Compare them to
   discover equivalences depending on liveness. If a sequence is unique
   in some liveness criteria, then output it to a new file for the cost.
   If a sequence has an unconditional optimization, then don't output it there.
   Send all optimization patterns to the optimizations file.
3. Swap the new file of unique sequences for the original output file.
4. Go to 2.

Memory models:
- Model stack effects -- memory decays once it's off the stack. An interrupt
  can trigger at any time, causing some decay of the stacks. The harware stack
  contents can decay at any time because interrupts push 3 bytes to the stack.
  The software stack is under our control. We can provide a red zone of a few
  bytes