https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92357
--- Comment #10 from Jakub Jelinek <jakub at gcc dot gnu.org> --- As for the other ICE, I'm still lost, I don't see similar bug to the above in other functions that write summaries. In output_symtab I see that all cgraph_nodes are streamed out, just those that fail lto_symtab_encoder_in_partition_p (i.e. the non-offloaded ones) will not have the callees and summaries and various other stuff streamed. lto_output_node sets boundary_p for the other partition nodes, and that sometimes leads to in_other_partition being set. Now that certainly isn't set when such nodes are streamed in again and e.g. FOR_EACH_DEFINED_FUNCTION then happily iterates over those in the offloading lto1. Note, modifying some flags on nodes during the offloading LTO streaming is dangerous, as if we also stream normal LTO, where e.g. in_other_partition shouldn't be set when for offloading LTO it should be, we could be in trouble.
