Hangs with `div` in indices?

[mcabbott]

Here's a stranger bug, this seems to run forever?

function act!(R, A, B, ax_i=1:4, ax_j=Base.OneTo(4))
    for j in ax_j
        for i in ax_i
            R[i, j] = A[2i + (j - 1) ÷ 2] + 0 * B[j]
        end
    end
    R
end
function act_avx!(R, A, B, ax_i=1:4, ax_j=Base.OneTo(4))
    @avx for j in ax_j
        for i in ax_i
            R[i, j] = A[2i + (j - 1) ÷ 2] + 0 * B[j]
        end
    end
    R
end

A, B = rand(10), rand(4)
act!(zeros(4,4), A, B)
act_avx!(zeros(4,4), A, B)  # needs ^C to exit?

Also found by CI in mcabbott/Tullio.jl#79, no urgent need... possibly I will remove the handling of ÷ in indices entirely. (This has bugs, e.g. mcabbott/Tullio.jl#76)

But I presume @avx should either reject this at once, or work.

[chriselrod]

Just fixed it on master.

[chriselrod]

Arbitrary functions that return integers should work; the clamp fix in the commit addressed that.
But I also realized div was missing from the cost table, so I added it.

It may be worth adding an explicit check for literal constants to replace the div with a cheaper shift.
LLVM will do this too, but it only gets the chance to after LoopVectorization made its modeling decisions using the more expensive div cost.

[chriselrod]

Fixed and releasing shortly.