Gunnar> dhawkins(a)rmas.com writes: Gunnar> [snip] Gunnar> I can reproduce that and have investigated a little bit. I have a Gunnar> pretty good idea what is causing it, but am too afraid to touch the Gunnar> marzilla stuff. :-) Gunnar> first of all a smaller recipe for the crash: Gunnar> - create a file test.el containing: (defun foo (bar) bar) Gunnar> - byte-compile and load this file Gunnar> - in *scratch* eval: (foo 7) Gunnar> OK, '7' is printed Gunnar> - make bar buffer-local: (make-local-variable 'bar) Gunnar> - eval (foo 7) again Gunnar> CRASH Gunnar> The problem is that funcall_compiled_function() correctly creates the Gunnar> specbinding for the local variable bar: Gunnar> $27 = {symbol = 139220936, Gunnar> old_value = 138893388, Gunnar> func = 0x80bdbec <specbind_unwind_local>} Gunnar> but at eval.c:532 UNBIND_TO_GCPRO_VARIABLES_ONLY is used which assumes Gunnar> that all bindings are 'normal' variables (i.e. func==0). Gunnar> I'm pretty sure that all this can be cured by using UNBIND_TO_GCPRO Gunnar> instead but maybe there is a smarter way. Martin? Thanks, Gunnar. I think your analysis is excellent. I don't see offhand how to keep the UNBIND_TO_GCPRO_VARIABLES_ONLY optimization, and see nothing better than to use UNBIND_TO_GCPRO. My logic error was thinking that func was only used for unwind-protects and such things. This is kind of annoying. We have to give up this optimization for the case of magic variables being used as function arguments, which *almost* never happens, and is probably never actually useful. Martin