unsigned combine losage

[Richard Henderson <rth at cygnus dot com>]

Your March 25 unsigned spree broke get_pos_from_mask, since it
no longer reports failure of a contiguous bit range.  This
results in rtl like

(insn 757 755 759 (set (zero_extract:DI (reg:DI 671)
            (const_int -1 [0xffffffffffffffff])
            (const_int 8 [0x8]))
        (subreg:DI (reg:CC 668) 0)) 30 {*insv_internal} (nil)
    (expr_list:REG_DEAD (reg:CC 668)
        (nil)))

which is clearly nonsensical.

That aside, I'll take this opportunity to re-iterate my objection to the
indiscriminate unsignedness changing that's been going on in the recent
months.  Just because the normal range of a particular value is non-negative
doesn't mean we never have call for out-of-band values, nor that we'll
never do subtractions.


r~


        * combine.c (get_pos_from_mask): Test exact_log2 result as signed.

Index: combine.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/combine.c,v
retrieving revision 1.133
diff -c -p -d -r1.133 combine.c
*** combine.c	2000/05/29 07:40:51	1.133
--- combine.c	2000/05/31 00:42:01
*************** get_pos_from_mask (m, plen)
*** 6588,6604 ****
  {
    /* Get the bit number of the first 1 bit from the right, -1 if none.  */
    int pos = exact_log2 (m & - m);
  
    if (pos < 0)
      return -1;
  
    /* Now shift off the low-order zero bits and see if we have a power of
       two minus 1.  */
!   *plen = exact_log2 ((m >> pos) + 1);
  
!   if (*plen <= 0)
      return -1;
  
    return pos;
  }
  
--- 6588,6606 ----
  {
    /* Get the bit number of the first 1 bit from the right, -1 if none.  */
    int pos = exact_log2 (m & - m);
+   int len;
  
    if (pos < 0)
      return -1;
  
    /* Now shift off the low-order zero bits and see if we have a power of
       two minus 1.  */
!   len = exact_log2 ((m >> pos) + 1);
  
!   if (len <= 0)
      return -1;
  
+   *plen = len;
    return pos;
  }