LGTM

Last commit also eliminates makeIsFalseish which was used only once when compiling !x, that is essentially "not is trueish" anyway.

Last commit makes isNaN and isFinite builtins again. Leads to slightly more code but also guarantees that these precompute at some places, which might be beneficial in static branch elimination. cc @MaxGraey wdyt?

I think this problem need solve via enhancing inlining limits. Yes we could improve isNaN and isFinite but this not solve problem with user space which still required manual inlining. Also I wonder how this better than simply add @inline decorators?

It's slightly more efficient than @inlineing since it doesn't create a block with aliased locals, but functionally equivalent and should yield about the same optimized code. Both would precompute.

Found one missing part. When we have casting from float to boolean like:

function cast(x: f64): bool {
   return bool(x);
}

it becomes:

return x != 0;

without isNaN check. So perhaps better implicitly transform to !!x for floats. WDYT?

Hmm, yeah, that's odd. Not sure what you mean with !!x, though. Looks like this needs to be handled on codegen level?

Not sure what you mean with !!x

Oh, mean it just should generate same code (x != 0.0) & (x == x)

Oh no, I spot a temporary local there. But yeah, guess this can't be avoided :)