[Bug tree-optimization/81633] [7/8 Regression] Incorrect floating point result with tree vectoriser

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81633

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I'm afraid the tree-vect-slp.c change is completely wrong, the original change
in
https://gcc.gnu.org/ml/gcc-patches/2016-08/msg01090.html if the oprnd == NULL
has child_index++; looks reasonable to me.  It changed a fragile code into much
more fragile code which will be always wrong if there are stmts with duplicated
(or more) operands.
Trying to recompute child_index from just oprnd is wrong, there can be multiple
arguments equal to the same SSA_NAME and which exact one it is should really be
derived from the index (dunno why we need child_index var, wouldn't it be
sufficient to use i for that)?
Are there (after the tree-vect-loop.c caller has been fixed in the same
revision) any vect_get_slp_defs callers which pass arguments with gaps or where
the oprnd i to child_index mapping shouldn't be identity?
I wonder about COND_EXPR with non-COMPARISON_CLASS_P first argument, but am not
sure.
E.g. vect_create_new_slp_node counts it as 2 regardless of what the first
argument is:
  else if (is_gimple_assign (stmt))
    {
      nops = gimple_num_ops (stmt) - 1;
      if (gimple_assign_rhs_code (stmt) == COND_EXPR)
        nops++;
    }
while vect_get_and_check_slp_defs has:
      if (gimple_assign_rhs_code (stmt) == COND_EXPR
          && COMPARISON_CLASS_P (gimple_assign_rhs1 (stmt)))
        {
          first_op_cond = true;
          number_of_oprnds++;
        }
which then matches e.g. masked handling in tree-vect-stmts.c
(vectorizable_condition).