Ãyvind Harboe <[EMAIL PROTECTED]> writes: > - Many backends do not support trampolines. Are trampolines > something that is ultimately being added to the backends?
As and if the port maintainers care, I think. > - Do (theoretical?) alternatives to trampolines exist? > I.e. something that does not generate code runtime. So, first off, let me point out that trampolines are only used if you take a pointer to a nested function. Many uses of nested functions (including those in languages where they are natural, e.g. Pascal, Ada) don't need to do that. Now, if the routine that calls the function pointer is aware that it is calling a nested function, the trampoline can be avoided. The typical technique is to make the pointer to the nested function be a 2-tuple of the actual code address and the value to load into the static chain register. The problem with this is, the caller generally doesn't know that, so you would have to make all function pointers be like that. And then you run up against the problem that in C there are strong practical constraints against making function pointers be bigger than data pointers (tho the language would permit it). It's not at all out of the question for Ada, though, and I recall seeing one of the Ada maintainers mention that they were interested in moving to that technique. Some architecture ABIs also allow clever tricks. For example, on ia64 the on-stack trampoline is just four pointers, no executable code (there is also a helper routine in libgcc). A similar technique is usable with some PowerPC OSes. zw
