Does it make any sense to use a template allocator parameter, like the standard container types, to control creation and destruction of my instance, or are new and delete sufficient?

Singleton creation is an interesting issue - Loki provides a number of different creation policies, including creating using new, malloc, or creating statically. Implementing them with a policy template.

Is there any other functionality that should be added? Specifically, is it sufficient to enforce that singletons are created in a specific order before main and destroyed in the reverse order after main (using dependency instances), or should I provide some way to support alternate lifetimes (ex: create on first reference to instance)?

Ordering of Singletons is, in my opinion, one of the most important issues with Singletons. It is almost certainly insufficient in most applications to create in a specific order, and delete in reverse order. One reason for this is start-up performance - you don't want every Singleton to have to create at startup (necessarily). Another reason can be memory efficiency. Loki again provides a number of policies including a Phoenix Singleton and a Longevity managed singleton. I personally tend to use Longevity managed singletons linked to the at_exit function. Nevertheless, there are some additional things that I would like to see - for example, a phoenix singleton that could be created and destroyed more than once during a program run could be useful - either explicitly or even if not used for a while. Also, allowing thread-safety as an option is also crucial.

Is there any interest in adding this library to Boost?

I would love to see a robust implementation of a Singleton in Boost. Whenever I start a software project, almost without exception I tend to add both Boost and Loki to the project - it would be great if I could add only Boost. I have to say however, that a more general library which implemented a variety of design patterns (like Loki) would be even better. For example: 1) Singleton 2) Visitor - like Loki implementing both a conventional cyclic visitor, and an acyclic visitor. 3) Composite 4) Abstract Factory 5) etc. etc. I've written my own templatised composite (available on codeproject.com), implementing a number of different iteration strategies (deep iteration, shallow iteration, typed iteration, etc.) and would be more than willing to throw that into the melting pot if there was some interest in putting together a library of design patterns for Boost. I've also written an acyclic visitor, like Loki but with a key change in the way that a failed visit is handled - in Loki if the visitor cannot visit a particular type of node, the default action is handled by a template in the base class of the visitable hierarchy. This in my view is flawed, and the default action should actually be handled by the visitor if required - this keeps the concept of action and data separated in accordance with the original visitor design pattern. This would probably need some work if it was to go into any attempt at a design patterns library. Dave Handley