Chris Fallin

Ah and re: > The question is, why is an "available block" precise enough, rather than an "available instruction"? is it enough that we do a forward scan through the...

I really like this idea! It would definitely simplify things overall. A few thoughts to add: - The main sticking point is going to be unwind-info, I think. Alongside the...

This is implemented in the [`machinst::abi` module](https://github.com/bytecodealliance/wasmtime/blob/main/cranelift/codegen/src/machinst/abi.rs) and the corresponding trait impls for each ISA; grep for `gen_clobber_save`. Could you say more about your idea?

Ah, that could work. My main concerns would be: - This creates a bunch of liveranges that span the whole function; I'd be worried about nontrivial impact on compile time,...

> (a) Isn't the argument that RA2 is smart enough to handle those? It certainly can handle function-long liveranges, sure. My main concern described above is performance (of the compiler,...

> > Another aspect of generated-code perf that just occurred to me: on aarch64, we can do more efficient "bulk saves" with STP/LDP (store/load pair instructions). RA2 isn't able to...

For sure, I'd be interested in the possible performance upside of this too -- @ghostway0, would you be able to do some compile+runtime tests with your prototype? We're happy to...

> @cfallin, could you take a think at the output in the filetests in this PR? I'm not sure what's going on, to be honest -- I stared at the...

I had some thoughts I was saving for the Wasmtime agenda item you had scheduled on Feb 29 (are we still discussing it then?) but to add some here: -...

It's possible that #6960 would fix this issue for you (it's not yet in a release -- it looks like it just missed the cutoff for 13.0 in September, but...