VRP? (was Re: PATCH: Re: gcj non-optimization curiosity)

[John Wehle <john at feith dot com>]

> Yeah. I wonder if this is something that John Wehle's value range 
> propagation pass can optimize?

Unfortunately not at the moment.  Currently VRP can optimize:

struct z
{
  const int length;
  int ary[0];
};

extern struct z *foo ();


int main ()
{
  struct z *l = foo ();
  int i;
  int x = 0;
  int j = 10;

  for (i = 0; i < j; ++i)
    {
      if ((unsigned) i >= (unsigned) (j))
       throw 1;
      x += l->ary[i];
    }

  return x;
}

Two problems handling your example:

  1) VRP currently tracks ranges for pseudo registers.  l->length
     involves a mem.

  2) Assigning l->length to a pseudo (i.e. j) yields:

...
BB 3
    reg 62
        min <integer_cst 850382c 0>        max <integer_cst 85038a0 2147483646>
    reg 64
        min <integer_cst 8503d04 1>        max <integer_cst 8503d78 2147483647>
...

     VRP determines the correct ranges for i (reg 62) and j (reg 64) however
     fails to realize that i will always be less than any given j when the
     if is evaluated.  It would probably need to use something other than
     absolute ranges in order to handle this.

-- John
-------------------------------------------------------------------------
|   Feith Systems  |   Voice: 1-215-646-8000  |  Email: john@feith.com  |
|    John Wehle    |     Fax: 1-215-540-5495  |                         |
-------------------------------------------------------------------------