Missed optimization for flooring division of signed by power of two

[tspiteri]

If a signed integer is divided by a power of two it is not a simple shift like for unsigned because division rounds to zero while the shift would round down. If flooring division is implemented, then flooring division by a power of two is equivalent to an arithmetic shift.

Here, floor_div4_a and floor_div4_b are equivalent, but the code generated for _a is not as good as _c.

inline int floor_div(int n, int d)
{
    int q = n / d;
    int r = n % d;
    if ((r < 0 && d > 0) || (r > 0 && d < 0))
        q -= 1;
    return q;
}

int floor_div4_a(int i)
{
    return floor_div(i, 4);
}

int floor_div4_b(int i)
{
    return i >> 2;
}

[Compiler explorer]

(This is similar to rustc issue rust-lang/rust#71096.)

[scottmcm]

Alive2 proof: https://alive2.llvm.org/ce/z/vpidT9

EDIT: Updated to 16-bit because the 32-bit seems to time out often, https://alive2.llvm.org/ce/z/vgj3BD

[rotateright]

Generalized pattern and requirements:
https://alive2.llvm.org/ce/z/MJzlhl

I was hoping there was some sub-pattern we could squash within there, but I don't see it, so I'll try to add the large match.

[rotateright]

I typo'd the bug number in the commit message, but this should be fixed with:
86b4a23