Wrong result when the same array appears in both sides of =

[maphdze]

When the same array appears in both side of =, the result from Tullio is strange

using Tullio
a = [1,2]
b = [1 2;3 4]
a = b * a
println(a) # [5,11]
a = [1,2]
@tullio a[x] = b[x,y] * a[y]
println(a) # [5,23]

It seems only the first component is right. When calculate the second componet, the first component is already updated to 5, and the second component is calculated by 3*5+4*2=23. When there are some complicated operations which have to be seprated into several steps, to avoid extra allocation, I prefer to use the same array repeatedly, and sometimes wrong results may occur like above.

[mcabbott]

This writes the following function:

# @tullio a[x] = b[x,y] * a[y] verbose=2

function 𝒜𝒸𝓉!(::Type, ℛ::AbstractArray{𝒯}, b, a, 𝒶𝓍x, 𝒶𝓍y, ♻️ = nothing, 💀 = true) where 𝒯
    for x in 𝒶𝓍x
        𝒜𝒸𝒸 = if ♻️ === nothing
                zero(𝒯)
            else
                ℛ[x]
            end
        for y in 𝒶𝓍y
            𝒜𝒸𝒸 = 𝒜𝒸𝒸 + b[x, y] * a[y]
        end
        ℛ[x] = 𝒜𝒸𝒸
    end
end

This is called like 𝒜𝒸𝓉!(typeof(a), a, b, a, axes(a,1), axes(b,2)), and so the loop overwrites values of a that it needs later.

I guess it would be nice to make this an error, by checking for aliasing.

It should still be safe to re-use an array, as long as you are reading and writing at the same point, not different ones. (E.g. things like a .= a .+ 1 are safe, of course.)

[maphdze]

Thanks for your response! I think it would be nice to throw an error/warning when reading and writing the same array at different point.