Congrats to the team, looks like some more fun things to play with!
> [floating point math is faster] thanks to optimal management of NaNs
Curious what's involved here if someone from the team can chime in.
> [floating point math is faster] thanks to optimal management of NaNs
Curious what's involved here if someone from the team can chime in.
We had a custom NaN "canonicalizer" that was always canonicalizing NaNs (that means: making sure the NaN bits were always the same in any chipset) regardless of use case and need. We took a look on how other runtimes were doing it more optimally and realized we could use LLVM intrinsic calls for math operations so LLVM would handle it directly.
If you are curious about the tech details, here's the PR! https://github.com/wasmerio/wasmer/pull/2161
Huh? This sounds totally incorrect to me.