Re: nested functions

On 2006-08-30, Tommy Thorn <tommy.thorn@gmail.com> wrote:
> > I realise that you cannot return a nested func pointer and pass it
> > around since theres no way GCC associates state with the ptr.
>
> Exactly! Trampolines are a sad compromise to implement closure-like
> structures in a way that can co-exist with ordinary C (which expects
> function pointers to point to code, not data structures). For language
> implementations with direct support for closures, it's generally
> better to use a structure with free variables and a function pointer
> (very much like a C++ object and one virtual function).
>
> There are exceptions to this which are hard to classify. Check
> "Optimization for Lazy Functional Languages" by Urban Boquist for a
> really all-out approach (and google "jhc haskell").


IIRC GPC (Gnu Pascal compiler) uses trampolines too. I can't remember
the exact details, but afaik the main reason had to do with passing
nested procs to other procedures. Some code to explain:


procedure xxx(proctype : someproctype);


begin
  proctype; // can access sharedstate
end;


procedure outer;


var sharedstate : integer;


procedure inner;


begin
    sharedstate:=sharedstate+1; // r/w
end;


begin
    xxx(inner);
end;


procedure otherproc;


begin
    xxx(outer);
end;


I believe they used trampolines to make sure the call chains were valid both
when passing inner and outer (top level) procs.


Some simple gpc -S 'ing might reveal some more info. (probably there are
other ways too, but gcc probably already supported trampolines)

Return to the comp.compilers page.
Search the comp.compilers archives again.