Inference regression in 1.6

[marius311]

I ran into this starting to upgrade my code to 1.6. A MWE with 1.6.0-DEV.1722 (599d329)

julia> @code_warntype (xs -> map((F,x)->F(x), (Float32,Float64), xs))((1,2))
Variables
  #self#::Core.Const(var"#23#25"())
  xs::Tuple{Int64, Int64}
  #24::var"#24#26"

Body::Tuple{Any, Any}
1 ─      (#24 = %new(Main.:(var"#24#26")))
│   %2 = #24::Core.Const(var"#24#26"())
│   %3 = Core.tuple(Main.Float32, Main.Float64)::Core.Const((Float32, Float64))
│   %4 = Main.map(%2, %3, xs)::Tuple{Any, Any}
└──      return %4

whereas on 1.5.3:

julia> @code_warntype (xs -> map((F,x)->F(x), (Float32,Float64), xs))((1,2))
Variables
  #self#::Core.Compiler.Const(var"#21#23"(), false)
  xs::Tuple{Int64,Int64}
  #22::var"#22#24"

Body::Tuple{Float32,Float64}
1 ─      (#22 = %new(Main.:(var"#22#24")))
│   %2 = #22::Core.Compiler.Const(var"#22#24"(), false)
│   %3 = Core.tuple(Main.Float32, Main.Float64)::Core.Compiler.Const((Float32, Float64), false)
│   %4 = Main.map(%2, %3, xs)::Tuple{Float32,Float64}
└──      return %4

However, on 1.6, its fine if you wrap it in another function call (presumably due to constant prop?)

julia> @code_warntype (() -> (xs -> map((F,x)->F(x), (Float32,Float64), xs))((1,2)))()
Variables
  #self#::Core.Const(var"#37#40"())
  #38::var"#38#41"

Body::Tuple{Float32, Float64}
1 ─      (#38 = %new(Main.:(var"#38#41")))
│   %2 = #38::Core.Const(var"#38#41"())
│   %3 = Core.tuple(1, 2)::Core.Const((1, 2))
│   %4 = (%2)(%3)::Core.Const((1.0f0, 2.0))
└──      return %4

So perhaps not really an issue? Not sure, but thought I'd report it here.

[marius311]

Seems like #38240 could be related, featuring mapreduce and also a regression since 1.5.

[JeffBezanson]

Some of the map methods for tuples are declared inline, and some are not. Adding inline declarations fixes this, but then I'm not sure why it worked on 1.5. Maybe changed inlining heuristics?