StaticArrays seems to apply promotion well for Base number types but not dual numbers

[jrevels]

julia> using ForwardDiff, StaticArrays

julia> x = ForwardDiff.Dual(1.0)
Dual{Nothing}(1.0)

julia> SMatrix{2,2}(x, 2, 3.0, 4.0)
2×2 SArray{Tuple{2,2},Real,2,4}:
 Dual{Nothing}(1.0)  3.0
               2     4.0

julia> promote(x, 2, 3.0, 4.0)
(Dual{Nothing}(1.0), Dual{Nothing}(2.0), Dual{Nothing}(3.0), Dual{Nothing}(4.0))

julia> SMatrix{2,2}(Int32(1), 2, 3.0, 4.0)
2×2 SArray{Tuple{2,2},Float64,2,4}:
 1.0  3.0
 2.0  4.0

[chethega]

Ok, so this bug is super weird and I have the nagging feeling that it is upstream (in inference? Include/compile-order dependent?).

julia> using ForwardDiff, StaticArrays

julia> x = ForwardDiff.Dual(1.0); tup=(x, 1);

julia> SMatrix{2,1}(tup)
2×1 SArray{Tuple{2,1},Real,2,2}:
 Dual{Nothing}(1.0)
               1   

julia> @which SMatrix{2,1}(tup)
(::Type{SArray{Tuple{S1,S2},T,2,L} where L where T})(x::Tuple{Vararg{Any,L}}) where {S1, S2, L} in StaticArrays at ~/.julia/dev/StaticArrays/src/SMatrix.jl:33

Let's look into the code and copy-paste it into the REPL

julia> import StaticArrays.SMatrix, StaticArrays.promote_tuple_eltype

julia> @generated function (::Type{SMatrix{S1,S2}})(x::NTuple{L,Any}) where {S1,S2,L}
           T = promote_tuple_eltype(x)

           return quote
               $(Expr(:meta, :inline))
               SMatrix{S1, S2, $T, L}(x)
           end
       end

julia> @which SMatrix{2,1}(tup)
(::Type{SArray{Tuple{S1,S2},T,2,L} where L where T})(x::Tuple{Vararg{Any,L}}) where {S1, S2, L} in Main at REPL[6]:2

julia> SMatrix{2,1}(tup)
2×1 SArray{Tuple{2,1},ForwardDiff.Dual{Nothing,Float64,0},2,2}:
 Dual{Nothing}(1.0)
 Dual{Nothing}(1.0)

[andyferris]

Yeah, no, that's definitely our fault!

We should not be calling promote_tuple_eltype from the generator. It needs to be put into the generated code. There will be no overhead for this.

(For some history, a lot of this package was written before the world-age-consistency system was completed, and particulary before it was understood (by myself). So there's some warts like this left and we should really audit the code for stuff like this, since it behaves so inconsistently)

[c42f]

I think was fixed a while back in #503