I have been poking around at at https://github.com/JuliaTime/TimeZones.jl/tree/cv/zdt-isbits-poc
The only reason creating a ZonedDateTime allocates is because of the lines in Interpret where a vector of interpetations is created and pushed to.

I did a benchmark of replacing just the UTC case with a tuple.
This case is easy, we know there is only going to be one result.
Allocations all gone, time for creation goes from:

julia> @btime ZonedDateTime(Date(2011, 6, 1), tz"America/Winnipeg", from_utc=true)
  113.625 ns (1 allocation: 112 bytes)
2011-05-31T19:00:00-05:00

down to

julia> @btime ZonedDateTime(Date(2011, 6, 1), tz"America/Winnipeg", from_utc=true)
  85.499 ns (0 allocations: 0 bytes)
2011-05-31T19:00:00-05:00

This case is particularly important is it shows up when you do things like adding 1 hour to a zoned date time.
So just changing that the a Tuple might be nice.

The other case has 2 allocations.

julia> @btime ZonedDateTime(Date(2011, 6, 1), tz"America/Winnipeg")
  144.806 ns (2 allocations: 192 bytes)
2011-06-01T00:00:00-05:00

Because we alocate a vector and push to it.
So we can give a good error message if there is 0 or 2 results.
Probably the easiest way to remove this is to change the logic for the constructor to not use interpret at all,
or to make a interpret1_or_error function, ether way to avoid a vector and just give one result or throw the error then and there