There should be two allocations, one as a scratch space for elements and one as a count vector. There are three:

julia> using BenchmarkTools

julia> v = rand(Int, 1000);

julia> @btime sort!(rand!($v));
21.259 μs (3 allocations: 10.16 KiB)

I tested this on 5dfd6c9 and current master and results are the same.

VSCode's @profview_allocs tracks the third allocation to the iteration in this line: