http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54200
Richard Guenther <rguenth at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Status|UNCONFIRMED |NEW Last reconfirmed| |2012-08-08 Ever Confirmed|0 |1 --- Comment #1 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-08-08 11:37:19 UTC --- Apart from generic brokeness of copyrename (as it leaks DECL_NAMEs relevant to debug info to places where another same DECL_NAME could be still live) this specific issue is here: for (i = 0; i < gimple_phi_num_args (phi); i++) { tree arg = gimple_phi_arg (phi, i)->def; if (TREE_CODE (arg) == SSA_NAME) updated |= copy_rename_partition_coalesce (map, res, arg, debug); } copy_rename_partition_coalesce is symmetric, but we only can assign arg the partition of res this way and at the end, if all args then have the same partition assign that to res.