improve constant propagation of named tuples

[kcajf]

This might already be known - there seem to be a few issues around keyword arguments and constant propagation / type inference, but I'm not really in the loop and it wasn't sure how out of date the issues are.

An attempt at a MWE:

f(; x) = g(x)
g(x) = x

z1() = g(3)
x1() = f(x=3)

z2() = g(Int64)
x2() = f(x=Int64)

julia> @code_warntype z1()
Variables
  #self#::Core.Compiler.Const(z1, false)

Body::Int64
1 ─ %1 = Main.g(3)::Core.Compiler.Const(3, false)
└──      return %1

julia> @code_warntype x1()
Variables
  #self#::Core.Compiler.Const(x1, false)

Body::Int64
1 ─ %1 = (:x,)::Core.Compiler.Const((:x,), false)
│   %2 = Core.apply_type(Core.NamedTuple, %1)::Core.Compiler.Const(NamedTuple{(:x,),T} where T<:Tuple, false)
│   %3 = Core.tuple(3)::Core.Compiler.Const((3,), false)
│   %4 = (%2)(%3)::NamedTuple{(:x,),Tuple{Int64}}
│   %5 = Core.kwfunc(Main.f)::Core.Compiler.Const(getfield(Main, Symbol("#kw##f"))(), false)
│   %6 = (%5)(%4, Main.f)::Int64
└──      return %6

julia> 

julia> @code_warntype z2()
Variables
  #self#::Core.Compiler.Const(z2, false)

Body::Type{Int64}
1 ─ %1 = Main.g(Main.Int64)::Core.Compiler.Const(Int64, false)
└──      return %1

julia> @code_warntype x2()
Variables
  #self#::Core.Compiler.Const(x2, false)

Body::DataType
1 ─ %1 = (:x,)::Core.Compiler.Const((:x,), false)
│   %2 = Core.apply_type(Core.NamedTuple, %1)::Core.Compiler.Const(NamedTuple{(:x,),T} where T<:Tuple, false)
│   %3 = Core.tuple(Main.Int64)::Core.Compiler.Const((Int64,), false)
│   %4 = (%2)(%3)::NamedTuple{(:x,),Tuple{DataType}}
│   %5 = Core.kwfunc(Main.f)::Core.Compiler.Const(getfield(Main, Symbol("#kw##f"))(), false)
│   %6 = (%5)(%4, Main.f)::DataType

For z1 and x1, inference seems OK. For x2, it seems to break (can only get DataType out), despite the fact that the non-kwargs version z2 infers something more precise (Type{Int64}).

The real usecase I met this in was trying to implement something like Base.dropdims(a::MyType; dims) = _dropdims(a, dims). Inference was completely failing, despite _dropdims(my_object, 3) inferring perfectly.

Ran on master:

julia> versioninfo()
Julia Version 1.4.0-DEV.103
Commit cb76f2a0e7 (2019-09-06 22:43 UTC)
Platform Info:
  OS: Linux (x86_64-redhat-linux)
  CPU: Intel(R) Core(TM) i7-6700 CPU @ 3.40GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, skylake)

[JeffBezanson]

Fixed by #35976