Wrong gradients due to global without shadow

[vchuravy]

const data = zeros(3)

function h(x, i)
    data[1] = x
    y = data[i]
   x*y
end

julia> Enzyme.autodiff(Reverse, h, Active, Active(1.0), Const(1))
((1.0, nothing),)
# Expected ((2.0, nothing), )

Amusingly if Enzyme can statically prove that the index is the same, the answer is correct.

julia> function g(x, i)
         data[i] = x
         y = data[i]
         y*x
       end
g (generic function with 2 methods)

julia> Enzyme.autodiff(Reverse, g, Active, Active(1.0), Const(1))
((2.0, nothing),)

@tgymnich For this case I would like to have a warning. Would also be helpful for arguments where enzyme deduces const, e.g. having a MaybeConst.

[tgymnich]

C example also wrong in a different way: https://fwd.gymni.ch/doh4qV

[vchuravy]

Shouldn't it be something like https://fwd.gymni.ch/QgKyvm

With a global const in C it can do store load forwarding.

[wsmoses]

@tgymnich how is your case wrong, that looks fine to me.

[vchuravy]

In discussion with Billy the right answer is to treat inttoptr here equivalent to a extern global in C.