[tree-ssa] Out of SSA status and issues

[Michael Matz]

Hi,

On Tue, 13 May 2003, Richard Kenner wrote:

>     I see nothing wrong in replacing 'i + 9' with '*p + 9'.  It would
>     probably not be efficient, but I can't see it being wrong.
>
> If an optimizer pessimizes the code, I'd consider that "wrong".
>
> This isn't a machine-dependent issue: with CPU speeds the way they are,
> a memory reference is *always* many times more expensive than an addition.

You apply a much too narrow view to optimization, or better to
transformation.  For instance in this situation:

  p1 = *p2;
  ... nothing changing *p2 or using p1...
  p3 = p1 + 1;

Suppose there is register pressure and we are in the register allocator
and decide that p1 doesn't get a hardreg.  If you now don't allow the use
of p1 to be replaced with *p2, you are forced to emit such code:

  p1 = *p2;
  stack[place] = p1;
  ... nothing changing *p2 or using p1...
  p3 = stack[place] + 1;

whereas it would have been better to use *p2 directly:

  (p1 = ... removed, because dead)
  ... nothing changing ....
  p3 = *p2 + 1;


I.e. it depends on the context if you want to replace register references
with memory accesses or not.  You can't just say it pessimizes code in all
cases.


Ciao,
Michael.