Missed optimization in clang only on powers of two

Question

When compiling with -Ofast, clang correctly deduces that the following function will always return 0.

int zero(bool b) {
    const int x = 5;
    return (x * b) + (-x * b);
}

compiles to

zero(bool):                               # @zero(bool)
        xor     eax, eax
        ret

However, if I change the constant to be any power of two (except 1 or 0), clang no longer makes the same deduction

int zero(bool b) {
    const int x = 8;
    return (x * b) + (-x * b);
}

compiles to

zero(bool):                               # @zero(bool)
        mov     eax, edi
        shl     eax, 3
        xor     dil, 1
        movzx   ecx, dil
        lea     eax, [rax + 8*rcx]
        add     eax, -8
        ret

The code compiled with compiler explorer.

If I change the function parameter to be anything bigger (short, int, long) the optimization is correctly made.

What causes this weird edge case?

Answer 1

It is an optimization problem in Clang, which was fixed after version 10.0.0. We can reproduce this problem in version 9.0.1.

Compared versions 9.0.1 and 10.0.0 in compiler explorer.