[Bug bootstrap/80838] PGO/LTO bootstrapped compiler 5% slower than pure PGO bootstrapped one

["hubicka at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80838

--- Comment #5 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
OK, found the problem.  LTO binary is built with -fPIC because of the way we
merge options.  I saw the same problem previously on Firefox. It is easy
mistake to do.  I think we should
 1) warn in this case
 2) downgrade options,  i.e. linking non-PIC onject to PIC leads to non-PIC
binary.

I don't think it makes sense to support mixed binaries because I see no benefit
in that.

With hack to lto-wrapper:
Index: lto-wrapper.c
===================================================================
--- lto-wrapper.c       (revision 249928)
+++ lto-wrapper.c       (working copy)
@@ -238,10 +238,12 @@ merge_and_complain (struct cl_decoded_op
        case OPT_fdiagnostics_show_option:
        case OPT_fdiagnostics_show_location_:
        case OPT_fshow_column:
+#if 0
        case OPT_fPIC:
        case OPT_fpic:
        case OPT_fPIE:
        case OPT_fpie:
+#endif
        case OPT_fcommon:
        case OPT_fexceptions:
        case OPT_fnon_call_exceptions:
@@ -501,10 +503,12 @@ append_compiler_options (obstack *argv_o
        case OPT_fdiagnostics_show_option:
        case OPT_fdiagnostics_show_location_:
        case OPT_fshow_column:
+#if 0
        case OPT_fPIC:
        case OPT_fpic:
        case OPT_fPIE:
        case OPT_fpie:
+#endif
        case OPT_fcommon:
        case OPT_fexceptions:
        case OPT_fnon_call_exceptions:


I now get:
$ perf_4.3 stat ./xgcc -B ./ -O2 -S ~/combine.ii

 Performance counter stats for './xgcc -B ./ -O2 -S /aux/hubicka/combine.ii':

       5664.950671      task-clock (msec)         #    0.998 CPUs utilized
               581      context-switches          #    0.103 K/sec
                 5      cpu-migrations            #    0.001 K/sec
            45,008      page-faults               #    0.008 M/sec
    16,882,004,890      cycles                    #    2.980 GHz
     2,818,057,410      stalled-cycles-frontend   #   16.69% frontend cycles
idle
     5,462,286,464      stalled-cycles-backend    #   32.36% backend  cycles
idle
    13,279,784,844      instructions              #    0.79  insns per cycle
                                                  #    0.41  stalled cycles per
insn
     2,919,823,112      branches                  #  515.419 M/sec
       141,311,318      branch-misses             #    4.84% of all branches

       5.674328322 seconds time elapsed

Which is nice improvement over non-LTO binary.