Re: SSA implementation

Re: SSA implementation

[Richard Kenner]

    We will be contributing a dead-code elimination pass in the very near
    future that operates on the SSA form.  One big improvement is that
    this algorithm is a) very fast and b) can eliminate loops in things
    like:

  void f () { 
    int i;
    for (i = 0; i < 100; ++i)
      ;
  }

We could always have eliminated such loops, but as I understood it, we
chose not to under the assumption they were there for timing delay
purposes.  Has this policy chaged?

Re: SSA implementation

[Bruce Korb]

Richard Kenner wrote:
> 
>     We will be contributing a dead-code elimination pass in the very near
>     future that operates on the SSA form.  One big improvement is that
>     this algorithm is a) very fast and b) can eliminate loops in things
>     like:
> 
>   void f () {
>     int i;
>     for (i = 0; i < 100; ++i)
>       ;
>   }
> 
> We could always have eliminated such loops, but as I understood it, we
> chose not to under the assumption they were there for timing delay
> purposes.  Has this policy chaged?

This won't work for timing delay on advanced processors anyway.
You would have to make "i" be "volatile" or the processor itself
will optimize it away (essentially).

Re: SSA implementation

[Mark Mitchell]

>>>>> "Richard" == Richard Kenner <kenner@vlsi1.ultra.nyu.edu> writes:

    Richard>     We will be contributing a dead-code elimination pass
    Richard> in the very near future that operates on the SSA form.
    Richard> One big improvement is that this algorithm is a) very
    Richard> fast and b) can eliminate loops in things like:

    Richard>   void f () { int i; for (i = 0; i < 100; ++i) ; }

    Richard> We could always have eliminated such loops, but as I
    Richard> understood it, we chose not to under the assumption they
    Richard> were there for timing delay purposes.  Has this policy
    Richard> chaged?

Yes.  This was debated to death a year or two ago.  The manual has
said:

  Historically, GCC has not deleted ``empty'' loops under the
  assumption that the most likely reason you would put one in a program is
  to have a delay, so deleting them will not make real programs run any
  faster.

  However, the rationale here is that optimization of a nonempty loop
  cannot produce an empty one, which holds for C but is not always the
  case for C++.

  Moreover, with @samp{-funroll-loops} small ``empty'' loops are already
  removed, so the current behavior is both sub-optimal and inconsistent
  and will change in the future.

for quite some time.

--
Mark Mitchell                   mark@codesourcery.com
CodeSourcery, LLC               http://www.codesourcery.com

Re: SSA implementation

[Gerald Pfeifer]

On Thu, 29 Jun 2000, Mark Mitchell wrote:
> We will be contributing a dead-code elimination pass in the very near
> future that operates on the SSA form.  One big improvement is that
> this algorithm is a) very fast and b) can eliminate loops in things
> like:
> 
>   void f () { 
>     int i;
>     for (i = 0; i < 100; ++i)
>       ;
>   }

When you do so, please don't also update the "Deleting ``empty'' loops"
section in the GCC manual.

Some time ago I added a note that in the future the behavior described
there will change which apparently is going to happen now. :-)

Gerald
-- 
Gerald "Jerry" pfeifer@dbai.tuwien.ac.at http://www.dbai.tuwien.ac.at/~pfeifer/
Have a look at http://petition.eurolinux.org -- it's not about Linux, btw!

Re: SSA implementation

[Mark Mitchell]

>>>>> "Geoff" == Geoff Keating <geoffk@cygnus.com> writes:

    Geoff> It does turn the code into SSA form.  That's close to
    Geoff> everything we want it to do.  In the future there will be
    Geoff> other things done while the code is in SSA form, but they
    Geoff> will be separate passes.

We will be contributing a dead-code elimination pass in the very near
future that operates on the SSA form.  One big improvement is that
this algorithm is a) very fast and b) can eliminate loops in things
like:

  void f () { 
    int i;
    for (i = 0; i < 100; ++i)
      ;
  }

Jeffrey Oldham has a mostly working implementation of this algorithm.

--
Mark Mitchell                   mark@codesourcery.com
CodeSourcery, LLC               http://www.codesourcery.com

Re: SSA implementation

[Geoff Keating]

"David Dolan" <ddolan@andrew.cmu.edu> writes:

> Has the ssa pass been fully implemented in the most current snapshot?

It does turn the code into SSA form.  That's close to everything we
want it to do.  In the future there will be other things done while
the code is in SSA form, but they will be separate passes.
-- 
- Geoffrey Keating <geoffk@cygnus.com>

SSA implementation

[David Dolan]

Has the ssa pass been fully implemented in the most current snapshot?

Thanks!!

Dave Dolan
ddolan@andrew.cmu.edu