I wrote:
the compile-time/run-time dichotomy is only relevant when a value depends on data only available at run-time.
Something I've wanted to experiment with for a long time and never got round to is writing CAFs back to the load module at the end of a run (if they're small enough or took a long time to evaluate). Has anyone tried this? (It would have a jolly entertaining effect on benchmark pages!). The logical extension of this would be that compiling a programme did the typechecking and then just wrote the binary equivalent of 'evaluate $ code-generate "...lambda expressions from programme text..."' into the load-module. If you never run the programme, this would be quicker. If you only run the programme once, it would take about the same time, and running it several times would be quicker -- very much so if it didn't depend on any run-time data. -- Jón Fairbairn Jon.Fairbairn@cl.cam.ac.uk