Fix broadcast with RowVectors of matrices

Fix #20979. Amusingly, this bug is a direct of `transpose` being recursive.

Author: mbauman

base/linalg/rowvector.jl

@@ -144,12 +144,12 @@ end
 @inline to_vec(x::Number) = x
 @inline to_vecs(rowvecs...) = (map(to_vec, rowvecs)...)
 
-# map
-@inline map(f, rowvecs::RowVector...) = RowVector(map(f, to_vecs(rowvecs...)...))
+# map: Preserve the RowVector, but note that `f` expects transposed elements
+@inline map(f, rowvecs::RowVector...) = RowVector(map(transpose∘f∘transpose, to_vecs(rowvecs...)...))
 
 # broacast (other combinations default to higher-dimensional array)
 @inline broadcast(f, rowvecs::Union{Number,RowVector}...) =
-    RowVector(broadcast(f, to_vecs(rowvecs...)...))
+    RowVector(broadcast(transpose∘f∘transpose, to_vecs(rowvecs...)...))
 
 # Horizontal concatenation #
 

test/linalg/rowvector.jl

@@ -249,3 +249,13 @@ end
         @test A'*x' == A'*y == B*x' == B*y == C'
     end
 end
+
+@testset "issue #20979" begin
+    f20979(z::Complex) = [z.re -z.im; z.im z.re]
+    v = [1+2im]'
+    @test (f20979.(v))[1] == f20979(v[1])
+    @test f20979.(v) == f20979.(collect(v))
+    
+    w = rand(Complex128, 3)
+    @test f20979.(v') == f20979.(collect(v')) == (f20979.(v))'
+end