Remove final uses of Inference.return_type

With this approach based on the eltype of tuple of computed elements, we
get the correct dynamic result independently of whether inference can
infer the return type.

This fixes some of the residual issues from #198.

src/broadcast.jl

@@ -102,12 +102,10 @@ end
         end
     end
 
-    eltype_exprs = [t <: AbstractArray ? :($(eltype(t))) : :($t) for t ∈ a]
-    newtype_expr = :(Core.Inference.return_type(f, Tuple{$(eltype_exprs...)}))
-
     return quote
         @_inline_meta
-        @inbounds return similar_type($first_staticarray, $newtype_expr, Size($newsize))(tuple($(exprs...)))
+        elements = tuple($(exprs...))
+        @inbounds return similar_type($first_staticarray, eltype(elements), Size($newsize))(elements)
     end
 end
 

src/mapreduce.jl

@@ -22,11 +22,10 @@ end
         tmp = [:(a[$j][$i]) for j ∈ 1:length(a)]
         exprs[i] = :(f($(tmp...)))
     end
-    eltypes = [eltype(a[j]) for j ∈ 1:length(a)] # presumably, `eltype` is "hyperpure"?
-    newT = :(Core.Inference.return_type(f, Tuple{$(eltypes...)}))
     return quote
         @_inline_meta
-        @inbounds return similar_type(typeof(_first(a...)), $newT, Size(S))(tuple($(exprs...)))
+        elements = tuple($(exprs...))
+        @inbounds return similar_type(typeof(_first(a...)), eltype(elements), Size(S))(elements)
     end
 end
 
@@ -110,8 +109,6 @@ end
 @generated function _mapreducedim(f, op, ::Size{S}, a::StaticArray, ::Type{Val{D}}) where {S,D}
     N = length(S)
     Snew = ([n==D ? 1 : S[n] for n = 1:N]...)
-    T0 = eltype(a)
-    T = :((T1 = Core.Inference.return_type(f, Tuple{$T0}); Core.Inference.return_type(op, Tuple{T1,T1})))
 
     exprs = Array{Expr}(Snew)
     itr = [1:n for n ∈ Snew]
@@ -128,7 +125,8 @@ end
 
     return quote
         @_inline_meta
-        @inbounds return similar_type(a, $T, Size($Snew))(tuple($(exprs...)))
+        elements = tuple($(exprs...))
+        @inbounds return similar_type(a, eltype(elements), Size($Snew))(elements)
     end
 end
 

test/broadcast.jl

@@ -134,22 +134,22 @@ end
     @testset "eltype after broadcast" begin
         # test cases issue #198
         let a = SVector{4, Number}(2, 2.0, 4//2, 2+0im)
-            @test_broken eltype(a + 2) == Number
-            @test_broken eltype(a - 2) == Number
-            @test_broken eltype(a * 2) == Number
-            @test_broken eltype(a / 2) == Number
+            @test eltype(a + 2) == Number
+            @test eltype(a - 2) == Number
+            @test eltype(a * 2) == Number
+            @test eltype(a / 2) == Number
         end
         let a = SVector{3, Real}(2, 2.0, 4//2)
-            @test_broken eltype(a + 2) == Real
-            @test_broken eltype(a - 2) == Real
-            @test_broken eltype(a * 2) == Real
-            @test_broken eltype(a / 2) == Real
+            @test eltype(a + 2) == Real
+            @test eltype(a - 2) == Real
+            @test eltype(a * 2) == Real
+            @test eltype(a / 2) == Real
         end
         let a = SVector{3, Real}(2, 2.0, 4//2)
-            @test_broken eltype(a + 2.0) == Float64
-            @test_broken eltype(a - 2.0) == Float64
-            @test_broken eltype(a * 2.0) == Float64
-            @test_broken eltype(a / 2.0) == Float64
+            @test eltype(a + 2.0) == Float64
+            @test eltype(a - 2.0) == Float64
+            @test eltype(a * 2.0) == Float64
+            @test eltype(a / 2.0) == Float64
         end
         let a = broadcast(Float32, SVector(3, 4, 5))
             @test eltype(a) == Float32