.abs() and .signum() cause 64-bit integer instruction generation

[hrydgard]

This is arguably not rust-gpu's fault, but num-traits, but it'll happen again, so we need a record of it.

Using vector abs ends up generating 64-bit ops due to https://github.com/rust-num/num-traits/blob/47d69223ce7707ebbf8a91c3e6b3997e48653feb/src/float.rs#L762 .

Also see the following image:

Ideally of course, abs should just generate a SPIR-V abs instruction instead of doing bithacks.

[khyperia]

Just adding here that the error reporting system for this is absolutely horrible right now, and should be improved (the way I got that screenshot was hacking in random stuff to the compiler to add error points to get a backtrace to figure out what was wrong, see the nyaah error message in the screenshot, haha).

The way that would be done is implementing the capability computation system @XAMPPRocky came up with, tracked in #42 (I'll add the system to a comment there)

[hrydgard]

Another affected function is signum, likely also copysign.

[khyperia]

Ah, never actually wrote this down, .abs() is fixed by rust-num/num-traits#196, and I don't believe copysign exists in num-traits for some reason. signum is still broken, though, going to see if I can PR a fix now. However, rust-num/num-traits#196 isn't in any num-traits release yet, so it's still broken for us.

[XAMPPRocky]

…and I don't believe copysign exists in num-traits for some reason.

@khyperia I opened a PR for it awhile ago rust-num/num-traits#207. I've sent them a friendly ping to look at it again.

[khyperia]

@XAMPPRocky are you still planning on getting rust-num/num-traits#207 in?

[XAMPPRocky]

@khyperia Just made the final needed changes there, but need to wait for the CI run to be approved first to check.

[oisyn]

num-traits issue was merged into main on May 3. Everything seems to work as expected now.