Improve math_builder_methods.

[nnethercote]

It currently produces silly verbose code like this:

fn add(&mut self, a: &'ll Value, b: &'ll Value) -> &'ll Value {
    #[allow(unused_variables)]
    let needs_i128_emulation = false || self.is_i128(a) || self.is_i128(b);
    if needs_i128_emulation {
        let args_vec: Vec<&'ll Value> = <[_]>::into_vec(
            ::alloc::boxed::box_new([a, b]),
        );
        match "add" {
            "add" | "unchecked_uadd" | "unchecked_sadd" => {
                assert_eq!(args_vec.len(), 2);
                return self.emulate_i128_add(args_vec[0], args_vec[1]);
            }
            "sub" | "unchecked_usub" | "unchecked_ssub" => { ... }
            "mul" | "unchecked_umul" | "unchecked_smul" => { ... }
            "and" => { ... }
            ...
            _ => { self.cx.fatal(...); }
        }
    }
    unsafe { ... }
}

and relies on constant folding to eliminate most of the code.

This commit changes the macro to produce code like this:

fn add(&mut self, a: &'ll Value, b: &'ll Value) -> &'ll Value {
    if self.is_i128(a) || self.is_i128(b) {
        self.emulate_i128_add(a, b)
    } else {
        unsafe { ... }
    }

and float ops are even shorter, because they don't need a u128 check and are now done with a new, separate macro:

fn fadd(&mut self, a: &'ll Value, b: &'ll Value) -> &'ll Value {
    unsafe { ... }
}

The generated code is much shorter: cargo expand's length has reduced from 34,636 lines to 26,356 lines, a 24% reduction.

The original code is also a little shorter. arg_vecs is gone because it isn't needed. The asserts are gone because they were very low value. There is now a tiny bit of code duplication, but it's trivial and not worth trying to avoid.

[LegNeato]

Awesome! 🍻