Panic-free ops via masking: {div,rem}_or

[Ben-Lichtman]

Hi, I was wondering if this team is considering the addition of masking arbitrary operations rather than just the gather select?

This is something already available in AVX512 through the kmask registers, though would probably require emulation on other platforms, but it seems like a very useful thing to have if at all possible.

One of the more awesome use cases I've seen for this is with vectorised emulation: https://gamozolabs.github.io/fuzzing/2018/10/14/vectorized_emulation.html which makes heavy use of masking to run / fuzz 8 virtual machines in parallel per core per thread.

[calebzulawski]

The recommended way to do this is to simply perform the operation and mask as desired--when producing the asm, the backend will recognize that the two operations can be performed with a single instruction.

[Ben-Lichtman]

Ah, I see. That's reasonable then - something like

let a = ...;
let b = ...;
let mask = ...;

let operated = operate(a, b);

mask.select(operated, a)

[programmerjake]

this could still be reopened, cuz many operations still need masked versions, such as masked integer division, masked shift, etc.

basically everything that could panic or cause UB or has other side-effects that can't be masked just by ignoring lanes in the result.

[workingjubilee]

re: shifts, I believe we should simply default our SIMD shift to the "wrapping" semantic the same way we default other operations to wrapping, also see rust-lang/rust#91237

I would be surprised if shl_or yielded better perf than wrapping_shl. The wrap should if anything help a compiler to simply serialize the RHS operand into an immediate operand, which would dominate perf next to dynamic operands on most arches.
If shl_or did tend to offer better perf, I would be sooner to consider shl_unchecked as a first step.

[workingjubilee]

This is a duplicate of #118,
but I am arbitrarily closing that one instead because this has more actual content.

[workingjubilee]

As of the sync getting merged, the shift ops on Simd types will wrap, like most of our ops do.