[tuples] Tuplify some passes

[Zdenek Dvorak]

Hi,

> >> > the simplest solution seems to be to make get_val_for return &buf + 2 in this
> >> > case.
> >>
> >> Yes, but it'd have to be an old-style tree, not a tuple.  I'm not sure
> >> if we want that.  Do we build this kind of trees elsewhere on the
> >> tuples branch?
> >
> > yes -- certainly in scev analysis and all the related analyses,
> 
> OK, so you mean something like the following patch?

more or less:

> -  FOR_EACH_SSA_USE_OPERAND (op, stmt, iter, SSA_OP_USE)
> -    {
> -      /* FIXME -- rewriting the statement this way in order to fold its rhs
> -	 is an ugly hack.  */
> -      nx = USE_FROM_PTR (op);
> -      val = get_val_for (nx, base);
> -      SET_USE (op, val);
> -      val = gimple_fold (stmt);
> -      SET_USE (op, nx);
> -      /* only iterate loop once.  */
> -      return val;
> +  op1 = gimple_assign_rhs1 (stmt);
> +  op2 = gimple_assign_rhs2 (stmt);

Changing the code this way will require also changing
chain_of_csts_start to check that
gimple_assign_rhs1 (stmt) or gimple_assign_rhs2 (stmt) is an ssa name
(just now statements like

... = &a[ssa_name]

are allowed, which would not be handled correctly by the changed code)

> +  code = gimple_assign_rhs_code (stmt);
> +  if (TREE_CODE (op1) != SSA_NAME)
> +    {
> +      gcc_assert (TREE_CODE (op2) == SSA_NAME);
> +      tmp = op1;
> +      op1 = op2;
> +      op2 = tmp;
>      }

This should be

if (TREE_CODE (op1) == SSA_NAME)
  op1 = get_val_for (op1, base)
else
  op2 = get_val_for (op2, base)

return fold_build2 (code, TREE_TYPE (x), op1, op2);

(code does not have to be commutative)

Zdenek

> 
> -  /* Should never reach here.  */
> -  gcc_unreachable ();
> +  val = get_val_for (op1, base);
> +  return fold_build2 (code, TREE_TYPE (op1), val, op2);
>  }
> 
>  /* Tries to count the number of iterations of LOOP till it exits by EXIT