[Bug tree-optimization/69400] [5/6 Regression] wrong code with -O and int128

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69400

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
The dividend is val = {-1, 0}, len = 2  (bah, we need a debug_wide_int for
gdb use!)

We do

  n = divisor_blocks_needed;
  while (n > 1 && b_divisor[n - 1] == 0)
    n--;

stripping the leading zeros and feed that n to

  if (remainder)
    {
      wi_pack ((unsigned HOST_WIDE_INT *) remainder, b_remainder, n);
      *remainder_len = canonize (remainder, (n + 1) / 2, dividend_prec);
      /* The remainder is always the same sign as the dividend.  */
      if (dividend_neg)
        *remainder_len = wi::sub_large (remainder, zeros, 1, remainder,
                                        *remainder_len, dividend_prec,
                                        UNSIGNED, 0);

so in int_const_binop we end up with val = {-2 }, len = 1 (that's still correct
I think given wide-int rules).

Interestingly things go downhill in wide_int_to_tree where we compute
ext_len as 3 and go the build_new_int_cst path which does

  if (len < ext_len)
    {
      --ext_len;
      TREE_INT_CST_ELT (nt, ext_len)
        = zext_hwi (-1, cst.get_precision () % HOST_BITS_PER_WIDE_INT);
      for (unsigned int i = len; i < ext_len; ++i)
        TREE_INT_CST_ELT (nt, i) = -1;

thus puts { -2, -1, 0 } in the INTEGER_CST storage (instead of { -2, 0, 0 }).

I think it needs to put 0 there for TYPE_UNSIGNED?  But then not sure
why it computes ext_len to 3 instead of 2 for this case anyway...  it does

  if (TYPE_UNSIGNED (type) && wi::neg_p (cst))
    return cst.get_precision () / HOST_BITS_PER_WIDE_INT + 1;

but I'd have expected

    return cst.get_len () + 1;

well, the three kinds of lengths are somewhat odd, so I simply assume 3
is correct for now.

Index: gcc/tree.c
===================================================================
--- gcc/tree.c  (revision 232666)
+++ gcc/tree.c  (working copy)
@@ -1267,7 +1267,7 @@ build_new_int_cst (tree type, const wide
       TREE_INT_CST_ELT (nt, ext_len)
        = zext_hwi (-1, cst.get_precision () % HOST_BITS_PER_WIDE_INT);
       for (unsigned int i = len; i < ext_len; ++i)
-       TREE_INT_CST_ELT (nt, i) = -1;
+       TREE_INT_CST_ELT (nt, i) = TYPE_UNSIGNED (type) ? 0 : -1;
     }
   else if (TYPE_UNSIGNED (type)
           && cst.get_precision () < len * HOST_BITS_PER_WIDE_INT)

doesn't fix this unfortunately, so the bug must be elsewhere (eventually
-2 and len = 1 is not correct as result from divmod).