Array{T, N} is a non-concrete type and thus an inference failure when a
concrete type could have been inferred. Most of the concatenation inference
failures existed in 0.3 but there are also some regressions, most
annoyingly for some constructions referred to by deprecation messages.
Den måndag 19 oktober 2015 kl. 17:17:29 UTC+2 skrev Andras Niedermayer:
>
> Yes, https://github.com/JuliaLang/julia/issues/13254 and
> https://github.com/JuliaLang/julia/issues/13665 is what I meant. I'm not
> sure any more whether "type instability" is the right expression, since the
> type inferencer indeed consistently gives you Array{T,N} for [[1,2];3] for
> all values. But you'd get more efficient code if the type inferencer gave
> you Array{Int64,1} instead of Array{T,N} for vcat.
>
> It doesn't look like a regression to me, I get this already on Julia 0.3:
>
> julia> g() = [[1,2];3]
> g (generic function with 1 method)
>
>
> julia> @code_typed g()
> 1-element Array{Any,1}:
> :($(Expr(:lambda, {}, {{:_var0,:_var1},{{:_var0,(Array{Int64,1},Int64),0
> },{:_var1,Array{Int64,1},18}},{}}, :(begin # none, line 1:
> _var1 = vcat(1,2)::Array{Int64,1}
> return cat(1,_var1::Array{Int64,1},3)::Array{T,N}
> end::Array{T,N}))))
>
>
>
>
> On Monday, October 19, 2015 at 5:02:57 PM UTC+2, Kristoffer Carlsson wrote:
>>
>> [a;b] is a bit regressed now when it comes to type stability, see:
>> https://github.com/JuliaLang/julia/issues/13254
>>
>> On Monday, October 19, 2015 at 4:56:40 PM UTC+2, Josh Langsfeld wrote:
>>>
>>> [a; b] / vcat(...) is type stable. It may produce different output
>>> depending on the types of a and b but it won't change behavior depending on
>>> their values.
>>>
>>> On Monday, October 19, 2015 at 10:20:06 AM UTC-4, Andras Niedermayer
>>> wrote:
>>>>
>>>> In light of the recent discussions (
>>>> https://groups.google.com/forum/#!topic/julia-users/xJ7GpKAa16E and
>>>> https://groups.google.com/forum/#!topic/julia-users/_lIVpV0e_WI) I got
>>>> curious, whether there is a non-modifying version of push!, since
>>>> push!(copy(a),b) doesn't feel right (I try to avoid modifying functions,
>>>> except if I really need performance) and [a; b] is not type stable:
>>>>
>>>> f(a) = [a; 3]
>>>> f2(a) = push!(copy(a),3)
>>>>
>>>>
>>>> julia> Base.return_types(f,(Array{Int64,1},))
>>>> 1-element Array{Any,1}:
>>>> Array{T,N}
>>>>
>>>>
>>>> julia> Base.return_types(f2,(Array{Int64,1},))
>>>> 1-element Array{Any,1}:
>>>> Array{Int64,1}
>>>>
>>>>
>>>> I couldn't find anything in the Julia docs. Of course, I could just
>>>> define my own
>>>>
>>>> push(a,vars...) = push!(copy(a),vars...)
>>>>
>>>>
>>>> but if there is a standard way to do that, I'd prefer that. Maybe
>>>> there's also some clever way to avoid making a copy if `a` is a literal
>>>> (e.g.
>>>> push([1,2])
>>>> ).
>>>>
>>>> The same applies to unshift, etc.
>>>>
>>>> (Non-modifying `shift!` and `pop!` already exist with the names `first`
>>>> and `last`. I'd find it easier to remember `shift` and `pop` -- it would
>>>> also be more similar to `merge` vs `merge!` for dictionaries, but I guess
>>>> that's just my taste.)
>>>>
>>>
