Disabling DomainError (in sqrt())?

[Elrod]

@bremez

Also interesting! Is it indeed a 128 bits limitation, or a 2 operands limitation? i.e. would it be able to operate on four Float32s?

Yes, four Float32s.

@bremez

Just out of curiosity, how do the check implementations differ so that one is optimizable and one is not?

You can use Cthulhu.@descend to look at the typed Julia code, LLVM IR, and assembly.

@bremez

Indeed, the issue was that I am calling sqrt on Int64s. I was not aware that this is considered a “weird type” - are there any other pathologies I should be aware of? With root-taking, perhaps the correct way to go about this is to manually convert to a floating type and then call @fastmath sqrt()?

Ints are for things like indexing arrays, not numerical computation IMO.
The integer has to be converted to a floating point number first before getting the root, anyway. It must also be converted before operating on it with other floating point numbers.
Other possible issues are things like overflow: The speed of light is an integer. Why should we care?