bump slop for twiceprecision division

[oscardssmith]

division (unlike the other intrinsics) can have table-maker dilema problems since the true result has infinite bits. Bumping the slop for it by 1 reduces failures from 80 in 10^10 to <1 in 10^11.

I think accepting an extra bit of inaccuracy is reasonable here (the alternative would be to make the division do an extra loop and do some extra extended precision, which seems unlikely to be worth it.

fixes #23497

[oscardssmith]

tests were done with the following code: (using QuadMath rather than BigFloat to make the tests run faster)

using Quadmath

asbits(x) = reinterpret(Unsigned, x)

asww(x) = Float128(x.hi) + Float128(x.lo)
astuple(x) = (x.hi, x.lo)

function cmp_sn2(w, hi, lo, slopbits=0)
    if !isfinite(hi)
        if abs(w) > floatmax(typeof(hi))
            return isinf(hi) && sign(w) == sign(hi)
        end
        if isnan(w) && isnan(hi)
            return true
        end
        return w == hi
    end
    if abs(w) < nextfloat(zero(hi))
        return (hi == zero(hi) || abs(w - widen(hi)) < abs(w)) && lo == zero(hi)
    end
    z = widen(hi) + widen(lo)
    w == z && return true
    zu, wu = asbits(z), asbits(w)
    tz = trailing_zeros(zu)
    zu >>= tz
    wu >>= tz
    zu = zu >> slopbits
    wu = wu >> slopbits
    return wu - 1 <= zu <= wu + 1
end

function test_twice(n, slopbits=5)
    Threads.@threads for i = 1:n
        T = Float32
        Tw = widen(T)
        x = Base.TwicePrecision{T}(rand())
        y = Base.TwicePrecision{T}(rand())
        xw, yw = asww(x), asww(y)
        if !cmp_sn2(Float64(xw/yw), astuple(x/y)..., slopbits)
            println((x,y))
        end
    end
    return
end

[giordano]

Why is CI not starting on this PR?

[Keno]

In general, I really dislike vague-posting comments like this. The PR description is much more informative. Of course it's possible to find the PR with git blame, but let's please be precise where we can.