Type inference issues on Julia master (v1.6)

[jipolanco]

The following is not inferred on Julia master (but works correctly on 1.5):

x = rand(2, 5)
@code_warntype ApplyArray(vcat, x)

# Output:
Variables
  #self#::Type{ApplyArray}
  f::Core.Const(vcat)
  factors::Tuple{Matrix{Float64}}

Body::Vcat{Float64, _A, Tuple{Matrix{Float64}}} where _A
1 ─ %1 = Core.tuple(f)::Core.Const((vcat,))
│   %2 = Core._apply_iterate(Base.iterate, LazyArrays.applied, %1, factors)::Applied{LazyArrays.DefaultArrayApplyStyle, typeof(vcat), Tuple{Matrix{Float64}}}
│   %3 = LazyArrays.ApplyArray(%2)::Vcat{Float64, _A, Tuple{Matrix{Float64}}} where _A
└──      return %3

The critical step seems to be around here:

M = applied(vcat, x)
@inferred ApplyArray(M)                       # fails
@inferred ApplyArray{eltype(M)}(M)            # fails
@inferred ApplyArray{eltype(M), ndims(M)}(M)  # works

I'll make a PR with a workaround that seems to fix this issue.

[dlfivefifty]

A PR is great. These type-inferrence issues are a bit delicate though in the past they got better so it may be a regression in master.

I tend not to support master until release candidate stage.

[jipolanco]

I'm closing this because inference regression is fixed on Julia master (JuliaLang/julia#38240).