Uncommenting unused variable assignment (or other various things) gives different results

[KristofferC]

Using the code

using ForwardDiff: Dual
using StaticArrays

v = ((1.,2.), (2.,3.))

function f(x::Vector{T}) where {T}
    x1 = SVector((x[1],)) # comment this to get correct result
    la1 = SVector((x[1],))
    f1 = zero(SVector{1}) # can also initialize with zero(SVector{1, T}) to get correct result 
    for _ in v
        f1 += la1
        @show f1
    end
    return f1
end

x = [Dual(1.0,1.0)]
@show f(x)

gives the erroneous output:

f1 = Dual{Nothing,Float64,1}[Dual{Nothing}(1.0,1.0)]
f1 = Dual{Nothing,Float64,1}[Dual{Nothing}(1.0,1.0)]
f(x) = Dual{Nothing,Float64,1}[Dual{Nothing}(1.0,1.0)]

Now, we comment out the line x1 = SVector((x[1],))

f1 = Dual{Nothing,Float64,1}[Dual{Nothing}(1.0,1.0)]
f1 = Dual{Nothing,Float64,1}[Dual{Nothing}(2.0,2.0)]
f(x) = Dual{Nothing,Float64,1}[Dual{Nothing}(2.0,2.0)]

and get the correct result.

Executing the original code in the REPL also gives the correct result

julia> let
           x1 = SVector((x[1],))
           la1 = SVector((x[1],))
           f1 = zero(SVector{1})
           for _ in v
               f1 += la1
               @show f1
           end
           f1
       end
f1 = Dual{Nothing,Float64,1}[Dual{Nothing}(1.0,1.0)]
f1 = Dual{Nothing,Float64,1}[Dual{Nothing}(2.0,2.0)]
1-element SArray{Tuple{1},Dual{Nothing,Float64,1},1,1}:
 Dual{Nothing}(2.0,2.0)

If we initialize f1 to be of type T we also get correct result.

My guess it is something with the type instability in combination with broadcasting in combination with optimization but not sure...

[KristofferC]

Bump, anything I can do here to help?

[StefanKarpinski]

@Keno, I speculatively assigned this to you, assuming this was an optimizer bug.

[Keno]

Looks like an SROA bug that then gives LLVM license to do the wrong thing