[uecker]

See my other comment in this thread with my preliminary benchmarking results. I have a patched GCC with two new builtins __builtin_static_chain and __builtin_nested_code (needs a better name), that give you the static chain pointer and the code pointer without creating a trampoline. Then I put both into a structure to simulate a wide pointer. Later I call it with __builtin_call_with_static_chain.

A trick one can do is to let it create the trampoline and then read off the two pointer from the position in the code where it is stored. Not portable and you still have the overhead for creating the trampoline, but you do not need the executable stack anymore.