Chris Fallin

This is an interesting requirement, and it would certainly be useful, I agree! It seems technically feasible-ish, and the general solution would be to (i) modify the `Requirement` lattice to...

In the end it's possible we might need to grow `Operand` slightly, though we should be careful doing so. It might be possible to experiment with e.g. 40 bits by...

Interesting -- thank you for thinking about this problem at least! I do agree that the current approach is a little ugly. This solves the problem for now, but I...

That's certainly an option. At least in Cranelift's use-case, though, we wouldn't accept a 3% regression; and also, 10 bits of index-space for custom locations would not be enough (we...

As someone who's spent a good amount of time staring at a sightglass invocation waiting for it to finish... some indication, even just "1% ... 2% ... 3% ... "...

Even simpler, just newlines? Perhaps with a `--verbose` flag we could emit a line after every run "Running compilation 1/N...", "Running instantiation 1/N...", "Running compilation 2/N...", etc?

As described by @jlb6740 just now over in #138, it would probably be even better to check for a mode that explicitly keeps a constant CPU frequency. The `userspace` governor...

Allowing a second `u32`, even, tagged with an instruction index, seems like it could be a reasonable approach to allow more information to be provided in "rare" cases (stackslots, fixed...

Ah, I found it -- I did some measurements [here](https://github.com/bytecodealliance/regalloc2/pull/1#discussion_r654835927) and found a 3% overhead when using a 64-bit `Operand`. That's not great but it does upper-bound things (it's the...

@Amanieu I'm a bit swamped at the moment but you might be able to run this yourself? I would suggest tackling the noise issue also, at least for this decision:...