[Bug c/46076] [4.6 regression] constant propagation and compile-time math no longer happening versus 4.4 and 4.5

["rguenther at suse dot de" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46076

--- Comment #12 from rguenther at suse dot de <rguenther at suse dot de> 2010-11-18 11:31:59 UTC ---
On Wed, 17 Nov 2010, matt at use dot net wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46076
> 
> --- Comment #11 from Matt Hargett <matt at use dot net> 2010-11-17 21:37:43 UTC ---
> Given that "int foo()" and "int foo(void)" are not varargs functions, shouldn't
> GCC mainline be able to optimize this?

The types are not treated as compatible.

The following would change that:

Index: tree-ssa.c
===================================================================
--- tree-ssa.c  (revision 166900)
+++ tree-ssa.c  (working copy)
@@ -1410,6 +1410,11 @@ useless_type_conversion_p (tree outer_ty
       if (!TYPE_ARG_TYPES (outer_type))
        return true;

+      /* A conversion between unprototyped and empty argument list is ok.  
*/
+      if (TYPE_ARG_TYPES (outer_type) == void_list_node
+         && !TYPE_ARG_TYPES (inner_type))
+       return true;
+
       /* If the unqualified argument types are compatible the conversion
         is useless.  */
       if (TYPE_ARG_TYPES (outer_type) == TYPE_ARG_TYPES (inner_type))

But for example our type-generic builtins for isinf and friends also
use a NULL TYPE_ARG_TYPES list but they _are_ variadic and certainly
not compatible with isinf (void).  I'm also not sure if other languages
allow variadic functions without a first arg (the backends would
certainly handle that just fine I think).

That said - the middle-end doesn't really have a notion of "unprototyped".
That's a concept the frontend should lower (at call-sites, but that
requires some middle-end infrastructure change).

Richard.