The most popular reflection users these days are then Object Relational Mappers (ORMs). They need to manufacture classes at run time, for this they need the language's reflection system.

Yes. I picked out ORMs (along with Plug-in architectures and frameworks) as "system programmer use cases" in Re: i want to learn programming language. I definitely see the case for these uses of run-time reflection in (potentially) simplifying and clarifying the code required to achieve them.

However, I believe even these can be achieved through compile-time code generation. In the case of ORMs, if you query the schema of the tables to be mapped, these are defined in terms of a limited set of datatypes. Ie. the particular RDBMSs own native type set. It therefore becomes possible to construct a base library of mappings between that set of datatypes and those of the host language. Then, at compile time, you can query the schema of the tables and build a class that maps the fields of the table to attributes with accessors that do whatever translation is required between the two sets of datatypes.

In this way, you end up with a wholy generated, concrete custom class (alliteration unintentional), for each table. This can be done relatively easily in static or dynamic languages. (Even Java! :)

And of course, given sufficient information available from the schema, it can even handle cross table dependancies (foreign key constraints etc.) far more easily and efficiently than using run-time reflection. It is quite easy to see how you can perform the 'schema to class generation process' offline, once-only when the schema changes, rather than everytime at run-time, whether it has changed or not.

In the archetypal scenario for ORM use, that of servers fielding high speed, connectionless protocols, this removal of run-time effort by pre-compiling can become very critical. Of course, dynamic languages tend to benefit less from pre-compilation than statics, but even they can benefit hugely.

Pre-loading a suite of pre-generated class definitions at mod_perl/Fastcgi "load time" and then just using them, is far better than having to generate the mappings, reflection-wise on-the-fly, for every page load.