pulley: Make canonicalize-nan.wast pass

[eagr]

Helps #9783

[eagr]

cargo run --features pulley -- wast --target pulley64 ./tests/misc_testsuite/simd/canonicalize-nan.wast -W nan-canonicalization

Seems passing. Is it how the test supposed to run?

About the fn naming, I'm just kinda going along with veq32x4 and friends. If you'd prefer, say, vlef32x4 over vlteqf32x4, or vueqf64x2 over vunoeqf64x2, I'm happy to change them.

And I haven't implemented all the float condcodes, just enough to get the tests pass. I did skim through the docs, but I don't believe I understand half of the implications. It'd great to hear your inputs before I proceed implementing others.

cc @alexcrichton

[github-actions]

Subscribe to Label Action

cc @fitzgen

This issue or pull request has been labeled: "cranelift", "pulley"

Thus the following users have been cc'd because of the following labels:

• fitzgen: pulley

To subscribe or unsubscribe from this label, edit the .github/subscribe-to-label.json configuration file.

Learn more.

[alexcrichton]

Thanks for this! I think it might be best to cut down on some of these opcodes in Pulley though since wasm doesn't actually use many of the FloatCC::Unordered* orderings for example. I'd recommend having a pulley opcode for wasm-based orderings but those needed only for nan canonicalization for example I think should be implemented like other lowering rules on other platforms (e.g. using combinations of opcodes instead of one). If this ever becomes performance sensitive we can of course revisit.

I'd recommend taking a look at how the riscv64 backend does this for example

[alexcrichton]

Oh I'll also note that another example is in #9884 (which I'm happy to merge your PR before mind and rebase around yours), which has another possible strategy for implementing some comparison-related opcodes here.

[eagr]

I didn't notice your PR. I couldn't see this one is contributing anything on top of yours, so I guess it's faster to just close this one to get out of the way. :)