https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120614
--- Comment #21 from kugan at gcc dot gnu.org --- I looked into 531.deepsjeng_r. For deepsjeng_r we see similar performance for AutoFDO as without it. Still looks like we have a missed opportunity there as srearch() now accounts for higher time in the profile. IPA inliner is now inlining them more making function search() slower. Using guessed profile seems to make them hot compared to PGO. Inlining into search() _ZL12history_goodP7state_tii (history_good) was inlined in the AutoFDO build _ZL11history_badP7state_tii (history_bad) was inlined in the AutoFDO build _ZL13history_scoreP7state_ti (history_score) was inlined in the AutoFDO build _ZL15remove_one_fastPiS_S_i (remove_one_fast) was inlined multiple times in the AutoFDO build4 Inlining into qsearch() _ZL16fast_order_movesP7state_tPiS1_ij (fast_order_moves) was inlined in the AutoFDO build _ZL15history_pre_cutP7state_tii (history_pre_cut) was inlined in the AutoFDO build.
