Unexpected broadcast materialization

[chrstphrbrns]

Does f need to materialize the broadcast here?

julia> VERSION
v"1.9.0-DEV.990"

julia> f(a) = length(sin.(a))
f (generic function with 1 method)

julia> g(a) = length(Broadcast.Broadcasted(sin, (a,)))
g (generic function with 1 method)

julia> a = 1:10000000
1:10000000

julia> @btime f(a)
  204.279 ms (3 allocations: 76.29 MiB)
10000000

julia> @btime g(a)
  14.118 ns (1 allocation: 16 bytes)
10000000

julia> @code_lowered f(a)
CodeInfo(
1 ─ %1 = Base.broadcasted(Main.sin, a)
│   %2 = Base.materialize(%1)
│   %3 = Main.length(%2)
└──      return %3
)

[antoine-levitt]

Dot expressions will always end up materializing at the boundary (because it has to work with arbitrary functions that expect Arrays). Same thing with (the more useful) sum, etc. Changing that is breaking, but would be a cool 2.0 addition, I think there are already open issues for that (search lazy broadcast). Or I guess there could be a mechanism by which the effective return type of a broadcast could change depending on the methods of the caller, but I'd imagine that would cause more trouble than it's worth...