Use of one in generator causing world age issues for ForwardDiff

[jrevels]

ref JuliaDiff/ForwardDiff.jl#277

MWE (expected output is an empty SVector):

julia> using ForwardDiff, StaticArrays

julia> StaticArrays.diff(SVector(ForwardDiff.Dual(1)))
ERROR: MethodError: Cannot `convert` an object of type Int64 to an object of type ForwardDiff.Dual{Void,Int64,0}
The applicable method may be too new: running in world age 21813, while current world is 21817.
This may have arisen from a call to the constructor ForwardDiff.Dual{Void,Int64,0}(...),
since type constructors fall back to convert methods.
Stacktrace:
 [1] one(::Type{ForwardDiff.Dual{Void,Int64,0}}) at ./number.jl:163
 [2] _diff(...) at /Users/jarrettrevels/.julia/v0.6/StaticArrays/src/mapreduce.jl:245
 [3] diff(::SVector{1,ForwardDiff.Dual{Void,Int64,0}}) at /Users/jarrettrevels/.julia/v0.6/StaticArrays/src/mapreduce.jl:238

AFAICT this is caused by the use of one (which ForwardDiff overloads for Dual numbers) within the generator here.

[mschauer]

Probably you should explain how the resulting eltype T = typeof(one(eltype(a)) - one(eltype(a))) can be computed for Dual numbers.

[timholy]

Is there clear evidence that this line couldn't be moved into the "body" (the quoted expression)? Inference is getting better and better at static evaluation...

[c42f]

Yes, moving the type computation into the body should fix this up. We do the same thing elsewhere in StaticArrays.

[c42f]

Does #317 do it for you? I found quite a lot of other places where this same problem exists.

[mschauer]

Addresses the MWE. Please bump if there are remaining issues.

[jrevels]

This fixes the broken ForwardDiff tests. Can you tag a new release containing #317? Thanks!