[glangdale]

User hairtuq gave a good summary. There seems to be a tension between pushing more work into the solver vs breaking up the problem into smaller sets, so there are many different designs. For example, I can pass thousands of candidate programs to the solver with opcodes specified ("please go get me some constants that make this program work for this task") or one completely unspecified program ("find me all the opcodes and all the programs") or some continuum between. My current experience is that partial specialization works well.

I have a lot of optimizations I'd like to apply. I'm doing the slightly brain-melting task of trying to get the super-optimizer to cough up its own optimization rules to speed the process, but that's a long topic and I'll be a lot chattier about if I see it ever work properly.

I'm also trying to figure out a way to make this pay in some fashion, being between jobs. Love to work on it more but I suspect reality will kick in soon.