https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97389
--- Comment #2 from CVS Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Jan Hubicka <hubi...@gcc.gnu.org>: https://gcc.gnu.org/g:56cb815ba22dd2ec00fee7a38f0862bc21d1c2a9 commit r11-3842-g56cb815ba22dd2ec00fee7a38f0862bc21d1c2a9 Author: Jan Hubicka <j...@suse.cz> Date: Tue Oct 13 09:19:54 2020 +0200 Fix tramp3d PGO misoptimization this patch fixes tramp3d ICE with PGO. It has turned out to be by a misupdate in ignore_edge I introduced in previous patch that made us to not compute SCCs correctly with -fno-lto. While looking for problem I proofread the sources and also fortified the srouces for situation where we insert a summary for no good reason and noticed a problem that early ipa-modref disabled itself in some cases. I also noticed that param_index is treamed as uhwi while it is signed (that wastes file space). Bootstrapping/regtesting x86_64-linux, will commit it tomorrow if that passes. gcc/ChangeLog: 2020-10-13 Jan Hubicka <hubi...@ucw.cz> PR ipa/97389 * ipa-modref.c (dump_lto_records): Fix formating of dump file. (modref_summary::dump): Do not check loads to be non-null. (modref_summary_lto::dump): Do not check loads to be non-null. (merge_call_side_effects): Improve debug output. (analyze_call): Crash when cur_summary->loads is NULL. (analyze_function): Update. (modref_summaries::insert): Insert only into summaries, not optimization_summaries. (modref_summaries::duplicate): Likewise; crash when load or sotres are NULL. (modref_summaries_lto::duplicate): Crash when loads or stores are NULL. (write_modref_records): param_index is signed. (read_modref_records): param_index is signed. (modref_write): Crash when loads or stores are NULL. (read_section): Compensate previous change. (pass_modref::execute): Do not check optimization_summaries t be non-NULL. (ignore_edge): Fix. (compute_parm_map): Fix formating. (modref_propagate_in_scc): Do not expect loads/stores to be NULL.
