http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57235
Richard Biener <rguenth at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Status|UNCONFIRMED |ASSIGNED Last reconfirmed| |2013-05-10 Component|tree-optimization |middle-end Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot gnu.org Target Milestone|--- |4.9.0 Summary|ICE verify_ssa failied |[4.9 Regression] ICE | |verify_ssa failied Ever confirmed|0 |1 --- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> --- Confirmed. This is because GIMPLE_RESX does not have virtual operands and thus there doesn't need to be a PHI node to merge the virtual operand chain into the successor we move the clobber to. GIMPLE_RETURNs now always have a virtual use to ease updating SSA form. Possible fix that disables the transform in these cases: Index: gcc/tree-eh.c =================================================================== --- gcc/tree-eh.c (revision 198768) +++ gcc/tree-eh.c (working copy) @@ -3360,6 +3360,11 @@ sink_clobbers (basic_block bb) } } + /* ??? If we have multiple predecessors but no virtual PHI we cannot + easily update virtual SSA form. */ + if (!vphi && !single_pred_p (succbb)) + return 0; + dgsi = gsi_after_labels (succbb); gsi = gsi_last_bb (bb); for (gsi_prev (&gsi); !gsi_end_p (gsi); gsi_prev (&gsi)) thus not sink the clobber. Another possibility is to remove such clobbers instead (if the block is reachable from elsewhere we may increase the objects lifetime by the sinking).