Ideas

Fix infinite loop bug in Python
- is it due to lastlist?
convert postfix to prefix?
- so we can try derivatives code?
- alternative:
  - parse to variadic prefix
    - (++ a b c)
    - (| (++ a b))
    - this is easier to read and write, and is the style we can use with Yaks
  - convert variadic prefix to binary postfix
Port that to tiny code from Darius, which also uses prefix
- Is it exponential time? The derivatives approach
- Also needs char class, dot, etc. Maybe try first with "(a|b)+"
How to do linear time derivatives approach?
- Look at smart constructors / hash consing
- Does that make all the difference in the SAME algorithm?
How to build a full DFA like re2c?
- with any optimization?
Compare SIZE of state machine, for derivatives, and Dragon book DFA -> optimized DFA
- I think the RSC articles never give the full DFA approach.
Advanced: compile UTF-8 into the DFA
- A tree of ranges?

Semantic Actions

For yaks lexer, we need either
- semantic actions, I guess attached to the final state?
- submatches, which we use in yaks/lex.ts
  - nfa-perl.y is 639 lines? Looks significantly different
    - has LeftmostBiased and LeftmostLongest
    - RepeatMinimal, RepeatLikePerl
    - it adds the idea of "Thread" I think?
QUESTION: are submatches and semantic actions the same thing?
- I don't think so, because submatches are more general
- re2c just has TOP-LEVEL alternation
- The 1994 re2c paper says "In the longer term, in-line actions will be added to re2c" Example: decoding integers But Figure 13 already has semantic actions, with if statements and printf error messages?
- it has it re2c-0.9.1/code.cc line 575 prints the actions

Backtracking: this isn't linear time! Because you do an unanchored match over and over again. You can test this out with re2c.

"a" { *tok = A }
"b" { *tok = B }
.* Z { *tok = Z }

Literal(c=97, out=Split(out1=..., out2=Literal(c=98, out=Match())))

patch() function is longer - still need to understand this
- seemed to just work
reminds me of Pratt parsing -- transcribing code, and fixing bugs is the easiest way to "feel" the algorithm