Chris Fallin

@fitzgen and I just talked about this at some length and I also agree that rewriting from the "wasm load" semantic unit-of-work directly to lower-level instructions is a good approach....

Hmm -- adding a side-table to hold some instruction data out-of-band, or picking just a single bit of trapcode to store, both seem somewhat suboptimal to me. We could do...

> why we'd prefer the status quo. Two reasons, one positive and one negative: - Positive: its definition is simple, and it doesn't really add any complexity to backends: its...

I like this! In conjunction with a "no path-sensitive deductions" rule that we hold throughout optimizations, I think the constant-folding you're describing would be safe too. To elaborate on that...

I second the table-driven approach above; we want to do it eventually anyway and it solves this problem, so it seems to be the cheapest and simplest way forward. One...

@tmngo we can definitely look into this; we have some long-running efforts to find ways to speed things up. One quick thing I would note though is that you have...

FreeBSD isn't an officially-supported platform but we can take patches on a best-effort basis. I'm happy to take a look if someone wants to make a PR (but I don't...

> Thanks for the review! I also fixed an issue where `alloc_affine_and_clear_affinity` wasn't working for anything other than `ReuseAffinity` which was also one of the motivations for the refactoring here....

Adding myself as a reviewer in addition to Trevor as I definitely want to look this over... this is really exciting!

One high-level thought: if legalization rewrites were egraph rules, that would allow us to use the cost mechanism to push away from "illegal" outputs for a given ISA: the op...