[Bug tree-optimization/103266] [12 regression] llvm-13 miscompilation: __builtin_assume_aligned causes over-aggressive dce

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103266

Jan Hubicka <hubicka at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenther at suse dot de

--- Comment #2 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
Thanks for a nice testcase!

We use "1cX " fnspec string_for assume aligned which indeed is problematic
since 'X' means that parameter is unused and thus we thus that it is not
returned.
"1" however specifies that function returns first parameter which is bit in a
contradiction with parameter being unused.  PTA codes takes precedence to "1"
and thus works, while modref sees 'X' and ignore the parameters and thus
propagation breaks.

In general we may have a specifier for pointer parameter that is used only as a
scalar value and never read/written from.  Perhaps 'S'? However I can also
teach modref to ignore unused for return handling.

Following untesed patch should fix the issue
diff --git a/gcc/builtins.c b/gcc/builtins.c
index 7d0f61fc98b..baeaff6eb1f 100644
--- a/gcc/builtins.c
+++ b/gcc/builtins.c
@@ -11101,7 +11101,7 @@ builtin_fnspec (tree callee)
       case BUILT_IN_STACK_SAVE:
        return ".c";
       case BUILT_IN_ASSUME_ALIGNED:
-       return "1cX ";
+       return "1cW ";
       /* But posix_memalign stores a pointer into the memory pointed to
         by its first argument.  */
       case BUILT_IN_POSIX_MEMALIGN: