[Bug middle-end/84200] New: r256888 causes 30% performance regression of 519.lbm_r at -Ofast generic tuning on Zen

[jamborm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84200

            Bug ID: 84200
           Summary: r256888 causes 30% performance regression of 519.lbm_r
                    at -Ofast generic tuning on Zen
           Product: gcc
           Version: 8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jamborm at gcc dot gnu.org
  Target Milestone: ---

Revision 256888 caused 30% run-time regression of 519.lbm_r when
compiled with -Ofast (with generic march and tuning) and when run on a
Zen CPU.  I have observed it on both a Ryzen and an EPYC.
Surprisingly, the revision does not seem to have any effect on this
particular benchmark on my Intel Sandy Bridge desktop.

The changelog for the revision is below, I believe it is the first
listed change that is the culprit.

  2018-01-19  Martin Liska  <mliska@suse.cz>

            * predict.def (PRED_LOOP_EXIT): Change from 85 to 89.
            (PRED_LOOP_EXIT_WITH_RECURSION): Change from 72 to 78.
            (PRED_LOOP_EXTRA_EXIT): Change from 83 to 67.
            (PRED_OPCODE_POSITIVE): Change from 64 to 59.
            (PRED_TREE_OPCODE_POSITIVE): Change from 64 to 59.
            (PRED_CONST_RETURN): Change from 69 to 65.
            (PRED_NULL_RETURN): Change from 91 to 71.
            (PRED_LOOP_IV_COMPARE_GUESS): Change from 98 to 64.
            (PRED_LOOP_GUARD): Change from 66 to 73.