`u128::midpoint`'s implementation could be better

[scottmcm]

Today it uses the hackers delight trick

rust/library/core/src/num/mod.rs

Lines 127 to 131 in 96f1da8

	pub const fn midpoint(self, rhs: $SelfT) -> $SelfT {
	// Use the well known branchless algorithm from Hacker's Delight to compute
	// `(a + b) / 2` without overflowing: `((a ^ b) >> 1) + (a & b)`.
	((self ^ rhs) >> 1) + (self & rhs)
	}

but

#![feature(num_midpoint)]
#[no_mangle] pub fn demo(x: u128, y: u128) -> u128 { x.midpoint(y) }

gives

demo:
        mov     rax, rdx
        xor     rax, rdi
        mov     r8, rcx
        xor     r8, rsi
        shrd    rax, r8, 1
        shr     r8
        and     rcx, rsi
        and     rdx, rdi
        add     rax, rdx
        adc     r8, rcx
        mov     rdx, r8
        ret

Whereas if it was instead

#[no_mangle] pub fn demo(x: u128, y: u128) -> u128 {
    let (z, c) = x.overflowing_add(y);
    (z >> 1) | (u128::from(c) << 127)
}

that compiles to

demo:
        mov     rax, rdi
        add     rax, rdx
        adc     rsi, rcx
        setb    cl
        shrd    rax, rsi, 1
        movzx   edx, cl
        shld    rdx, rsi, 63
        ret

which seems strictly better.

(Alternatively you could try making it an intrinsic with a fallback body, since the obvious LLVM -- https://llvm.godbolt.org/z/c56YrMz8c -- seems to work decently.)