Feature request: unsafe_free!

Internals & Design
21

try/finally blocks are not the answer for performance critical code

Yeah, I did actually think about using try/finally here but it was a bit too costly to build into @no_escape, though it is actually cheaper than one might think, e.g.

julia> function foo1(buf=default_buffer())
           @no_escape buf begin
               z = @alloc(Int, 10)
               z .= 1
               sum(z)
           end
       end;

julia> function foo2(buf=default_buffer())
           b = default_buffer()
           cp = Bumper.checkpoint_save(b)
           try
               foo1(buf)
           finally
               Bumper.checkpoint_restore!(cp)
           end
       end;

and then

julia> @btime foo1();
  12.717 ns (0 allocations: 0 bytes)

julia> @btime foo2();
  34.566 ns (0 allocations: 0 bytes)

julia> @btime foo1($(default_buffer()));
  5.610 ns (0 allocations: 0 bytes)

julia> @btime foo2($(default_buffer()));
  26.099 ns (0 allocations: 0 bytes)

so if you never actually hit the finally branch it’s not that expensive to have around so long as you’re not using it in the tightest of tight loops.

I could also make a varitant of @no_escape that does the finally branch automatically, e.g. it could be called @no_escape_finally or something, would that help?

Good idea.

3 Likes