Possible type inference regression in mapreduce

[jipolanco]

On Julia master, there seems to be a type inference issue when reducing over values known at compile time. This looks similar to #35800, but I'm not sure if it's related.

For instance,

x = rand(3, 4)
y = rand(20)
ndims_max(arrays...) = Val(maximum(map(ndims, arrays)))
@code_warntype ndims_max(x, y)  # returns Val(ndims(x)) = Val(2)

On Julia master:

@code_warntype ndims_max(x, y)
Variables
  #self#::Core.Const(ndims_max)
  arrays::Tuple{Matrix{Float64}, Vector{Float64}}

Body::Val{_A} where _A
1 ─ %1 = Main.map(Main.ndims, arrays)::Core.Const((2, 1))
│   %2 = Main.maximum(%1)::Int64
│   %3 = Main.Val(%2)::Val{_A} where _A
└──      return %3

On Julia 1.5 the result Val(2) is correctly inferred:

Variables
  #self#::Core.Compiler.Const(ndims_max, false)
  arrays::Tuple{Array{Float64,2},Array{Float64,1}}

Body::Val{2}
1 ─ %1 = Main.map(Main.ndims, arrays)::Core.Compiler.Const((2, 1), false)
│   %2 = Main.maximum(%1)::Core.Compiler.Const(2, false)
│   %3 = Main.Val(%2)::Core.Compiler.Const(Val{2}(), true)
└──      return %3

Note that in this specific case, a workaround is to replace maximum(map(ndims, arrays)) by max(map(ndims, arrays)...) (see LazyArrays PR).

It seems like the issue can be tracked down to mapreduce. The following also fails on Julia master, while it works on 1.5 (note that reduce calls mapreduce):

freduce(::Val{n}) where {n} = Val(reduce(identity, (n, )))
@code_warntype freduce(Val(3))  # returns Val(3)

The same happens with foldl and foldr.

Also note that the following is correctly inferred:

fidentity(::Val{n}) where {n} = Val(identity(n))

so the issue really seems to come from the reduction step.

---

Version info:

julia> versioninfo()
Julia Version 1.6.0-DEV.1381
Commit 0097bddf90* (2020-10-30 08:20 UTC)
Platform Info:
  OS: Linux (x86_64-redhat-linux)
  CPU: Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-11.0.0 (ORCJIT, skylake)

[JeffBezanson]

Fixed on latest master.

[oscardssmith]

Does/should this have a test?