Re: [PATCH][TUPLES] Fix for a number of ICE's in g++ regression test.

[Diego Novillo <dnovillo at google dot com>]

On 5/9/08 4:23 PM, Doug Kwan (éæå) wrote:
>  	     }
> +	  /* Sometimes the C++ front-end generates a POINTER_PLUS_EXPR whose
> +	     operand 0 has type not trivailly convertible to the expression
> +	     type.  We need to do a conversion.  */
> +	  if (!useless_type_conversion_p (TREE_TYPE (*expr_p),
> +					  TREE_TYPE (TREE_OPERAND (*expr_p,
> +								   0))))
> +	    {
> +	      tmp = fold_convert (TREE_TYPE (*expr_p),
> +				  TREE_OPERAND (*expr_p, 0));
> +	      *expr_p = build2 (POINTER_PLUS_EXPR, TREE_TYPE (*expr_p),
> +				tmp, TREE_OPERAND (*expr_p, 1));
> +	      break;
> +	    }
> +	

Andrew suggests changing the C++ FE here, but I actually would rather 
allow the FEs more latitude.  The gimplifier is here to do all these 
conversions when needed.  Andrew, was your intention to make 
POINTER_PLUS_EXPR stricter even for the FEs?


>            /* FALLTHRU */
>  
>  	default:
> Index: gcc/gcc/gimple.h
> ===================================================================
> --- gcc/gcc/gimple.h	(revision 135125)
> +++ gcc/gcc/gimple.h	(working copy)
> @@ -1468,8 +1468,8 @@ gimple_call_return_type (const_gimple gs
>    tree fn = gimple_call_fn (gs);
>    tree type = TREE_TYPE (fn);
>  
> -  /* See through pointer to functions.  */
> -  if (TREE_CODE (type) == POINTER_TYPE)
> +  /* See through pointer or reference to functions.  */
> +  if (TREE_CODE (type) == POINTER_TYPE || TREE_CODE (type) == REFERENCE_TYPE)

Use POINTER_TYPE_P here as Andrew suggested.


OK otherwise.


Diego.