Chris Fallin

Ah, and I guess using the egraph pass in this way would also force the question of side-effecting ops -- we need to be able to create new ones, and...

CFG modification would be difficult at best in the egraph rewrite infra, because that's not really the paradigm the thing is built around -- the skeleton remains fixed and we...

(And on a little further thought, with some more autoconversions, the `return_two` bit could go away too)

@afonso360 I agree, but I think that's pretty fundamental here: if we start making backends require a specific subset of total CLIF after opts, then we'll have to be aware...

Re: the "scoping" approach, an efficient implementation (I think!) occurred to me a few days ago. It would probably get at what I think y'all are trying to get at...

That all sounds basically correct to me! A few thoughts: - SSA dominance implies a *partial* order on args, at least with respect to block-level dominance; the greatest-lower-bound (where "top"...

@jameysharp most of this sounds basically right to me and agrees with my understanding; thank you all for looking into this more deeply! One thought on the "trying to revert..."...

(Actually, that gives me an idea: maybe a chaos-mode knob, or just an `OptLevel`, could be "as pessimized as possible"; this inverts cost comparisons so we always pick the *most...

@jameysharp this aligns with my understanding from our discussion yesterday; thanks for putting together the concrete test-case! Just to ensure I understand your path from here (and to have it...

Thanks for writing this all out! One question: > When `subsume` is disabled and the cost model picks the worst alternative, elaboration tries to duplicate the load into `block2`. That...