Strange behavior of `Const` annotation

[gdalle]

Here's an MWE involving differentiation of a function with closed-over cache storage.

using Enzyme

struct FuncWithCacheSum{C}
    cache::C
end

struct FuncWithCacheFirst{C}
    cache::C
end

function (f::FuncWithCacheSum)(x)
    (; cache) = f
    cache[1] = x[1]
    y = sum(abs2, cache)  # only difference is here
    cache[1] = zero(eltype(cache))
    return y
end

function (f::FuncWithCacheFirst)(x)
    (; cache) = f
    cache[1] = x[1]
    y = abs2(cache[1])  # only difference is here
    cache[1] = zero(eltype(cache))
    return y
end

apply(x, f::F) where {F} = f(x)

x = [3.0]
fs = FuncWithCacheSum([0.0])
ff = FuncWithCacheFirst([0.0])

julia> ff(x) == fs(x)
true

julia> gradient(Reverse, apply, x, Const(fs))[1]  # zero as expected
1-element Vector{Float64}:
 0.0

julia> gradient(Reverse, apply, x, Const(ff))[1]  # non-zero???
1-element Vector{Float64}:
 6.0

Is it expected that the Const annotation doesn't seem to work for the second function?

[vchuravy]

LLVM can sometimes perform "store-forwarding" and so when it sees a store followed directly by a load, it will forward the value of the store directly, this then cause the activity to not be truncated, but rather follow the value.

We would likely need an "activity blocking" intrinsic instead of relaying on the semantics of Const