Re: fwprop patch testing

[Paolo Bonzini <paolo dot bonzini at lu dot unisi dot ch>]

The only big regression for fwprop on PPC is bzip2.  I've distilled it 
to this small testcase:

   int f(int *);
   int verbosity;
   int *arr;
   int last;

   void g ()
   {
      int i;
      if (last < 4000) {
         if (verbosity >= 4) f(&verbosity);
         for (i = 0; i <= last; i++) arr[i] = i;
      }
   }

where CSE previously was performing load PRE.  So, actually this 
regression is caused by -fno-cse-skip-blocks rather than deleting parts 
of fwprop.  Note that this can be performed on the GIMPLE-level only to 
some extent.  What we are looking for is the following pseudo-C:

   void g ()
   {
      int i;
      int *plast = &last;
      int mylast = *plast;
      if (mylast < 4000) {
        if (verbosity >= 4) f(&verbosity);
        i = 0;
        mylast = *plast;
        while (i <= mylast)
          {
             arr[i] = i;
             /* Note the unfortunate aliasing between arr and plast */
             mylast = *plast;
          }
      }
   }

And tree-based load PRE can only get to

   void g ()
   {
      int i;
      int *plast = &last;
      int mylast = *plast;
      if (mylast < 4000) {
        if (verbosity >= 4) f(&verbosity);
        i = 0;
        plast = &last;      /* redundant */
        mylast = *plast;
        while (i <= mylast)
          {
             arr[i] = i;
             plast = &last;    /* redundant, and very expensive! */
             mylast = *plast;
          }
      }
}

since the "plast" lowering is obviously happening during expand.

I have not yet looked at whether it is possible to teach fwprop about 
this case, but I am not sure it is possible (I'm more inclined to think 
that it is simply not fwprop's job).  Right now, I am thinking more 
about shuffling the pass order.  One possibility that comes to mind is 
GCSE+fwprop+CSE, where GCSE could work out the common code for loading 
the address, and fwprop/CSE could do the addressing mode selection properly.

On one hand, this may mean increasing the number of GCSE passes from 1 
to 2.  On the other hand, it may mean that the CSE pass at the hand of 
rest_of_handle_gcse could become redundant.

Paolo