Re: demonstration of proper tail calls

On Tue, 4 May 2004, Fergus Henderson wrote:

> On 29-Apr-2004, Chris Lattner <[email protected]> wrote:
>> If there is a real need for it we can always add a bit to the
>> call instruction to allow the front-end to indicate that it is always safe
>> to do a tail call (even if there are stack allocations), but I'm not sure
>> this is really desirable or safe.
>
> IMHO this is indeed desirable, since there are situations in which
> it is easy for the front-end to provide that the call is safe, but much
> harder for the back-end to do the same proof.

Agreed.

> <snip>
> then after the call to bar(), the front-end can be sure that once bar()
> has returned, there are no more references to the parameter X, because
> this is enforced by the front-end's language rules which guarantee that
> references are only passed downwards.  So the front-end knows that foo's
> stack frame can be deallocated.  But generally the back-end can't easily
> determine this.  Having the front-end generate code in such a way that
> the back-end can easily recognize this fact becomes quite awkward.

Great point.

FWIW, we just released LLVM 1.5:
http://mail.cs.uiuc.edu/pipermail/llvm-announce/2005-May/000016.html

LLVM 1.5 includes the ability for front-ends to indicate that function 
calls to not access their caller stack frames (the tail marker) to handle 
cases like the above.  In addition, it supports general proper tail calls: 
http://llvm.cs.uiuc.edu/releases/1.5/docs/ReleaseNotes.html#tailcalls

... as was requested on this list.

Cheers,

-Chris

-- 
http://nondot.org/sabre/
http://llvm.cs.uiuc.edu/