linearize aliased kernel args

[jumerckx]

Can't test end-to-end but the HLO contains the stores.

using Reactant, CUDA

A = ConcreteRArray(fill(1))

struct MyStruct{T, B}
    b::B
    A::T
end

function f!(tup)
    tup[1].A[] = 2
    return nothing
end

s = MyStruct(10, A)

f(s) = @cuda threads=1 f!((s, s))
@code_hlo optimize=false f(s);

[jumerckx]

Since I moved argidx++ out of the loop, I'm not sure whether this is still correct because I don't really know what's happening here.

[wsmoses]

@jumerckx to make review easier, can you rebase this on main?

[jumerckx]

@jumerckx to make review easier, can you rebase this on main?

Done

[wsmoses]

can you add a test with the repeated arg, otherwise lgtm

[wsmoses]

@jumerckx can you add that as a test, then let's merge!

[jumerckx]

I'm afk til tomorrow but will add it then.

[wsmoses]

LGTM but I think there's a conflict that needs to be fixed to merge