Fix performance regression in broadcasting with CartesianIndices (#39333)

* Fix performance regression in broadcasting with CartesianIndices

This avoids the boundary check due to a change in the implementation
of iteration using `CartecianIndices` in PR #37829.
This is a workaround on the caller side and does not change
the iteration mechanism itself.

Co-authored-by: Matt Bauman <mbauman@gmail.com>
Co-authored-by: thofma <thofma@gmail.com>
(cherry picked from commit a4cd68c)

Author: 3 people

base/broadcast.jl

@@ -978,8 +978,10 @@ preprocess_args(dest, args::Tuple{}) = ()
         end
     end
     bc′ = preprocess(dest, bc)
-    @simd for I in eachindex(bc′)
-        @inbounds dest[I] = bc′[I]
+    # Performance may vary depending on whether `@inbounds` is placed outside the
+    # for loop or not. (cf. https://github.com/JuliaLang/julia/issues/38086)
+    @inbounds @simd for I in eachindex(bc′)
+        dest[I] = bc′[I]
     end
     return dest
 end