opaque_closure_test

assign_test.jl

#%%
using BenchmarkTools
BenchmarkTools.DEFAULT_PARAMETERS.seconds = 2.50
using Base: @opaque

add_opaq = @opaque (a::Vector, b::Vector) -> a .+ b
@noinline add(a::Vector, b::Vector) = a .+ b
@noinline add!(c::Vector, a::Vector, b::Vector) = c .= a .+ b
@inline add_inline(a::Vector, b::Vector) = a .+ b
#%%

test() = begin
	a = randn(1024)
	b = randn(1024)
	c = randn(1024)
	@btime $c .= $add($a, $b)
	@btime $c .= $add!($c, $a, $b)
	@btime $c .= $add_inline($a, $b)
	@btime $c .= $add_opaq($a, $b)
end
test()
;
# 816.818 ns (1 allocation: 8.12 KiB)
# 126.728 ns (0 allocations: 0 bytes)
# 857.111 ns (1 allocation: 8.12 KiB)
# 4.096 μs (11 allocations: 8.48 KiB)
#%%
using Base: @opaque
test2 = @opaque () -> begin
	a = randn(1024)
	b = randn(1024)
	c = randn(1024)
	@time c .= add(a, b)
	@time c .= add!(c, a, b)
	@time c .= add_inline(a, b)
	@time c .= add_opaq(a, b)
end
test2()
;
# 0.000037 seconds (9 allocations: 8.344 KiB)
# 0.000037 seconds (8 allocations: 224 bytes)
# 0.000043 seconds (9 allocations: 8.344 KiB)
# 0.000074 seconds (16 allocations: 8.672 KiB)

maybe you need @btime $c .= $add_opaq($a, $b) since add_opaq is a non-constant global?

You are right, it is worth a try, so:

@btime $c .= $add_opaq($a, $b)
#  4.064 μs (11 allocations: 8.53 KiB)
@btime $c .= add_opaq($a, $b)
# 4.096 μs (10 allocations: 8.48 KiB)

Interestingly it increased allocations by 1.

One of my friend told me, that in C++ we have "move semantics" and "return value optimization" developed to prevent these allocations. What we need to know why julia can't control the LLVM to apply the optimizations in such situations.