Re: trampolines

Stefan Monnier (monnier+/news/lists/linux/kernel@TEQUILA.SYSTEMSZ.CS.YALE.EDU)
15 Apr 1997 20:05:03 -0400

Matei Conovici <> writes:
> What are they exactly used for (and possibly why)?

It doesn't have much to do with linux-kernel, but here it goes:

say you have the following code:

some_fun (List vals)
int limit = 2;
Bool my_filter (int val) { return val > limit; }

filter_out(vals, my_filter);

the function my_filter refers to "limit" which is not a global value, but
rather ta value that will sit on the stack (and maybe multiple times, maybe
filter_out will even call recursively some_fun, who knows), so my_filter cannot
just be a "pointer to code" but needs to encode two things: the code and the
location of the "limit" variable it refers to. This is called a closure in
usual parlance. Now C doesn't expect function-pointers to be such "complex"
datastructure, so in order for filter_out to be able to call that closure as if
it really were just a pointer-to-code, gcc creates (by copying from a template)
a piece of code which wraps the my_filter closure so that when this piece of
code is called from filter_out, it will call my_filter and pass it the location
of the "limit" variable. This piece of code has to be created at run time
and the memory space for it will naturally be allocated on the stack.

There are alternative ways to deal with this problem, but none of them are
clearly superior (and this one isn't any more clearly superior to the others).

I'd (personally) prefer to have functgion-pointers be actually represented by a
structure containing the address of the code and the address of the
corresponding environment (in which "limit" will be found) and I believe that
the PowerPC standard API uses such a scheme, but it's a pain to change such a
thing (just like changing the size of "long" tends to break many things)