setindex! no longer inlined for rank N>6

[Jutho]

I just noticed this today on latest master (Version 0.4.0-dev+2498), I guess this must be regression?

f6(A)=(A[1,1,1,1,1,1]=1.)
f7(A)=(A[1,1,1,1,1,1,1]=1.)
code_typed(f6,(Array{Float64,6},))
code_typed(f7,(Array{Float64,7},))

returns

1-element Array{Any,1}:
 :($(Expr(:lambda, Any[:A], Any[Any[],Any[Any[:A,Array{Float64,6},0]],Any[]], :(begin  # none, line 1:
        (top(arrayset))(A::Array{Float64,6},1.0,1,1,1,1,1,1)::Array{Float64,6}
        return 1.0
    end::Float64))))

versus

1-element Array{Any,1}:
 :($(Expr(:lambda, Any[:A], Any[Any[],Any[Any[:A,Array{Float64,7},0]],Any[]], :(begin  # none, line 1:
        setindex!(A::Array{Float64,7},1.0,1,1,1,1,1,1,1)::Any
        return 1.0
    end::Float64))))

Consequently, indexing into an array of rank>6 becomes slow. Indexing with a CartesianIndex could be fixed by letting the stagedfunction generate the arrayset call automatically, but I guess this also needs to be fixed for normal indexing?

[Jutho]

For getindex, inlining stops for N>7.

[timholy]

Dupe of #5393. Jeff closed that, but I'm not sure it's redundant with anything else, so perhaps we should leave this open.

As @simonster said there,

I think performance drops because MAX_TUPLETYPE_LEN in inference.jl is 8, which appears to prevent getindex from getting inlined when it is passed >8 arguments.

[Jutho]

I thought I had seen something like this before, but couldn't find it. Thanks. Ok then, I hope you agree that we should at least change indexing with CartesianIndex to immediately return the arrayref and arrayset instruction, so as to have at least one way of fast indexing?

[Jutho]

Although with the @inline macro and/or stagedfunctions it should also be possible to make getindex and setindex! fast, no?

[timholy]

I certainly support doing something (whatever seems best) to make this fast.

[mbauman]

Hm, this got worse after #10525. Inlining now stops for rank N>4. I'll try to look into this one a bit more.

[mbauman]

Ah, found it. We're still one-worse than we were since I dispatch to an internal function that has one more argument (to specify the LinearIndexing trait), and so we hit MAX_TUPLETYPE_LEN one dimension sooner.

[matthieugomez]

I encountered this issue. If it's hard to solve now, a warning would be nice - it was hard to pinpoint the source of the problem in my code.

[KristofferC]

This seems fixed now,

julia> code_typed(f10,(Array{Float64,10},))
1-element Array{Any,1}:
 CodeInfo(:(begin 
        (Base.arrayset)(A,(Core.typeassert)(1.0,Float64)::Float64,1,1,1,1,1,1,1,1,1,1)::Array{Float64,10}
        return 1.0
    end))=>Float64

Please repoen if this is still a problem.