tail calls in const functions?

tail calls in const functions?

[Jan Hubicka]

Hi
The tail call optimization makes function to modify the stack, that
belongs to the caller.  Is this valid for const function?
IMO gcc has right to optimize out:
int test()
{
return c(1) + d(1);
}
the second store of 1, in case c is const function (it don't do that currently).
When tail call optimization is done for c, the stack is clobbered.

So perhaps it will be neccesary to disable tail call for const functions.

Honza

Re: tail calls in const functions?

[Jakub Jelinek]

On Thu, Mar 23, 2000 at 03:24:31PM +0100, Jan Hubicka wrote:
> Hi
> The tail call optimization makes function to modify the stack, that
> belongs to the caller.  Is this valid for const function?
> IMO gcc has right to optimize out:
> int test()
> {
> return c(1) + d(1);
> }
> the second store of 1, in case c is const function (it don't do that currently).

This is actually a wrong example, because none of these functions can be
a tail call (the parent routine has to sum the two returned values up).

Cheers,
    Jakub
___________________________________________________________________
Jakub Jelinek | jakub@redhat.com | http://sunsite.mff.cuni.cz/~jj
Linux version 2.3.99-pre2 on a sparc64 machine (1343.49 BogoMips)
___________________________________________________________________

Re: tail calls in const functions?

[Jan Hubicka]

> On Thu, Mar 23, 2000 at 03:24:31PM +0100, Jan Hubicka wrote:
> > Hi
> > The tail call optimization makes function to modify the stack, that
> > belongs to the caller.  Is this valid for const function?
> > IMO gcc has right to optimize out:
> > int test()
> > {
> > return c(1) + d(1);
> > }
> > the second store of 1, in case c is const function (it don't do that currently).
> 
> This is actually a wrong example, because none of these functions can be
> a tail call (the parent routine has to sum the two returned values up).
The problem is not in function test, but in the function c.
Imagine function c like this one:

int
__attribute__ ((const))
c(int a)
{
  return e(a+1);
}

The c (after tail-call optimizations done) will clobber the stack
created by function test.

Honza
> 
> Cheers,
>     Jakub
> ___________________________________________________________________
> Jakub Jelinek | jakub@redhat.com | http://sunsite.mff.cuni.cz/~jj
> Linux version 2.3.99-pre2 on a sparc64 machine (1343.49 BogoMips)
> ___________________________________________________________________

Re: tail calls in const functions?

[Jeffrey A Law]

  In message < 20000323154300.D2277@atrey.karlin.mff.cuni.cz >you write:
  > > On Thu, Mar 23, 2000 at 03:24:31PM +0100, Jan Hubicka wrote:
  > > > Hi
  > > > The tail call optimization makes function to modify the stack, that
  > > > belongs to the caller.  Is this valid for const function?
I would think so -- it's only going to diddle in the parameter area.

One way to think about a const function is does it read/write *state* that
is needed across invocations of the function or which are significant for
the behavior of other code in the program.

In this case we're modifying stack slots that will not be read again anyway;
think of those slots as pass-by-value parameters.  We can stomp on the contents
of those slots all we want since we know their value will never be used again.

jeff

Re: tail calls in const functions?

[Jan Hubicka]

> 
>   In message < 20000323154300.D2277@atrey.karlin.mff.cuni.cz >you write:
>   > > On Thu, Mar 23, 2000 at 03:24:31PM +0100, Jan Hubicka wrote:
>   > > > Hi
>   > > > The tail call optimization makes function to modify the stack, that
>   > > > belongs to the caller.  Is this valid for const function?
> I would think so -- it's only going to diddle in the parameter area.
> 
> One way to think about a const function is does it read/write *state* that
> is needed across invocations of the function or which are significant for
> the behavior of other code in the program.
> 
> In this case we're modifying stack slots that will not be read again anyway;
> think of those slots as pass-by-value parameters.  We can stomp on the contents
> of those slots all we want since we know their value will never be used again.
OK.
I was bringing this issue mainly because I was unsure whther we want to read it
again or not.  In the example I gave the caller was storing value "1" twice to
the same position once for first function, later for the second.  I was unsure
whther we would not want to implement later pass that will eliminate the second
sture.

With const fuctions modifying their arguemnt area we will have to hack
this future pass for such special case.

Honza

> jeff

Re: tail calls in const functions?

[Jeffrey A Law]

  In message < 20000323232003.B5316@atrey.karlin.mff.cuni.cz >you write:
  > OK.
  > I was bringing this issue mainly because I was unsure whther we want to rea
  > d it
  > again or not.  In the example I gave the caller was storing value "1" twice
  >  to
  > the same position once for first function, later for the second.  I was uns
  > ure
  > whther we would not want to implement later pass that will eliminate the se
  > cond
  > sture.
  > 
  > With const fuctions modifying their arguemnt area we will have to hack
  > this future pass for such special case.
We must always consider the argument area clobbered by a call, even to a
const function.

jeff

Re: tail calls in const functions?

[Richard Henderson]

On Thu, Mar 23, 2000 at 03:24:31PM +0100, Jan Hubicka wrote:
> The tail call optimization makes function to modify the stack, that
> belongs to the caller.  Is this valid for const function?

It doesn't belong to the caller.  Not any more than
call-clobbered registers do.

> IMO gcc has right to optimize out:
> int test()
> {
> return c(1) + d(1);
> }
> the second store of 1, in case c is const function

I don't agree.


r~