I’m not so sure. Returning non-bitstype tuples is currently pretty hit-and-miss with respect to avoiding the allocation for the tuple.
Once this https://discourse.julialang.org/t/immutables-with-reference-fields-why-boxed/7706 is fixed (if ever) I totally agree that sentinel values are stupid-- as far as I understand, we should have several free (already invalidated) registers where we can place our bool, and the actual return value can go via sret / other register / whatever the current ABI wants.
Maybe an easier clean fix would be to special case that Tuple{atype, btype} with inferred bitstype never allocates. Of course, unless this is packed into a tuple / array / etc, or this rule would need to be applied recursively. In other words: I am asking for a special ABI for returning pairs. I would even be happy if I was forced to annotate that I want @pair_return_ABI or there was a magic built-in __unpacked_pair that lazily materializes the tuple.