Petr Penzin

> Interpreter code that I've seen does seem to rely on native targets preserving NaN payloads, for add/sub/mul/div/sqrtl, and most native targets do propagate NaNs for those in practice. For...

> That would add overhead though. The motivation for the feature was to avoid this overhead. I think it adds overhead either way - currently it adds it to (x86)...

> I kind of barely understand "A canonical NaN is a floating-point value ±nan(canonₙ) where canonₙ is a payload whose most significant bit is 1 while all others are 0"...

> The rules documented for Rust currently are not the rules we want, they are the rules we have because LLVM and wasm don't let us implement the rules we...