Re: [tuples][patch] Convert pass_object_sizes to tuples

[Diego Novillo <dnovillo at google dot com>]

[ Jakub, I CCd you because there is a question about seemingly dead code 
in compute_builtin_object_size.  See below. ]


On 3/18/08 1:34 AM, Bill Maddox wrote:

> This patch enables pass_object_sizes, and corrects mishandling of
> unary NOPs in get_maxval_strlen.
>
>         * builtins.c (fold_builtin_object_size):
> 	Enable call to compute_builtin_object_size, previously
> 	stubbed out awaiting tuplification.
> 	* tree-ssa-ccp.c (valid_gimple_rhs_p, valid_gimple_call_p,
> 	move_ssa_defining_stmt_for_defs, update_call_from_tree):
> 	Deleted, moved to tree-ssa-propagate.c.
> 	(get_maxval_strlen): Use gimple_assign_single_p.
> 	Handle assignment with unary NOP correctly.
> 	* tree-ssa-propagate.c (valid_gimple_rhs_p, valid_gimple_call_p,
> 	move_ssa_defining_stmt_for_defs, update_call_from_tree):
> 	Moved here from tree-ssa-ccp.c.
> 	* tree-ssa-propagate.h (valid_gimple_rhs_p, valid_gimple_call_p,
> 	move_ssa_defining_stmt_for_defs, update_call_from_tree): Declared.
> 	* gimple-dummy.c (compute_builtin_object_size): Removed dummy.
> 	* tree_object_size.c (addr_object_size, alloc_object_size)
> 	Tuplified.
> 	(pass_through_call, compute_builtin_object_size): Tuplified.
> 	(expr_object_size): Tuplified.  Some cases broken out.
> 	(plus_expr_object_size): Deleted.
> 	(call_object_size, unknown_object_size, plus_stmt_object_size):
> 	New functions.  Handle cases broken out from expr_object_size.
> 	(cond_expr_object_size): Fix comment.
> 	(collect_object_sizes_for): Tuplify.
> 	(check_for_plus_in_loops_1, check_for_plus_in_loops): Tuplify.
> 	(compute_object_sizes): Tuplify.
> 	* gimple.c (gimple_assign_single_p, gimple_assign_unary_nop_p):
> 	New functions.
> 	* gimple.h (gimple_assign_single_p, gimple_assign_unary_nop_p):
> 	Declare.
> 	* passes.c (init_optimization_passes): Enable pass_object_sizes.

OK with a couple of changes below:

> @@ -330,21 +328,16 @@ compute_builtin_object_size (tree ptr, i
>  {
>    gcc_assert (object_size_type >= 0 && object_size_type <= 3);
>
> +  /* FIXME tuples.  Paranoia.  */
> +  gcc_assert (TREE_CODE (ptr) != CALL_EXPR);
> +

No need to have a 'FIXME tuples' here.  I agree that PTR should never be 
a CALL_EXPR.  I wonder why Jakub added it to begin with.

Jakub, why are we handling CALL_EXPRs in compute_builtin_object_size? 
It looks dead.

> @@ -490,10 +481,72 @@ expr_object_size (struct object_size_inf
>
>    if (TREE_CODE (value) == ADDR_EXPR)
>      bytes = addr_object_size (value, object_size_type);
> -  else if (TREE_CODE (value) == CALL_EXPR)
> -    bytes = alloc_object_size (value, object_size_type);
>    else
> -    bytes = unknown[object_size_type];
> +    {
> +      /* FIXME tuples.  Paranoia.  */
> +      gcc_assert (TREE_CODE (value) != CALL_EXPR
> +                  && TREE_CODE (value) != POINTER_PLUS_EXPR);

No need to have 'FIXME tuples'.  Jakub, same question here re. CALL_EXPRs.


Thanks.  Diego.