https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92924
--- Comment #17 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
I further hacked the script to record only values that are useful, where useful
means with greater count then all / TOPN_VALUES / 2. I use same test in GCC
itself (that was bug in original luxou's patch that it required count to be at
least 50% so we did more then two speculations only when the number of targets
was precisely 2 and probability precisely 50%).
I also added special case for those calls which have only one target.
Now I get:
== Stats for all ==
stats for indirect_call:
total: 160441 freq: 833372200
not executed at all: 134836
invalid: 0 (0.00%) freq:0 (0.00%)
only one target: 20036 (12.49%) freq:108329394 (13.00%)
tracked values:
0 values: 111 times (0.07%) freq: 14162014 (1.70%)
1 values: 2963 times (1.85%) freq: 594520090 (71.34%)
2 values: 2284 times (1.42%) freq: 111578254 (13.39%)
3 values: 185 times (0.12%) freq: 2525320 (0.30%)
4 values: 26 times (0.02%) freq: 2257128 (0.27%)
stats for topn:
total: 6620 freq: 16471134
not executed at all: 6236
invalid: 0 (0.00%) freq:0 (0.00%)
only one target: 196 (2.96%) freq:3535103 (21.46%)
tracked values:
0 values: 39 times (0.59%) freq: 5340350 (32.42%)
1 values: 84 times (1.27%) freq: 6006897 (36.47%)
2 values: 59 times (0.89%) freq: 1547519 (9.40%)
3 values: 6 times (0.09%) freq: 41265 (0.25%)
4 values: 0 times (0.00%) freq: 0 (0.00%)
== Stats for mainline ==
stats for indirect_call:
total: 160441 freq: 865715132
not executed at all: 134862
invalid: 156 (0.10%) freq:64839775 (7.49%)
only one target: 1255 (0.78%) freq:604290 (0.07%)
tracked values:
0 values: 14939 times (9.31%) freq: 356592480 (41.19%)
1 values: 8535 times (5.32%) freq: 377193150 (43.57%)
2 values: 660 times (0.41%) freq: 66082791 (7.63%)
3 values: 15 times (0.01%) freq: 133242 (0.02%)
4 values: 19 times (0.01%) freq: 269404 (0.03%)
stats for topn:
total: 6620 freq: 16793729
not executed at all: 6235
invalid: 10 (0.15%) freq:1307937 (7.79%)
only one target: 10 (0.15%) freq:1300 (0.01%)
tracked values:
0 values: 182 times (2.75%) freq: 9195915 (54.76%)
1 values: 156 times (2.36%) freq: 6111713 (36.39%)
2 values: 26 times (0.39%) freq: 173054 (1.03%)
3 values: 1 times (0.02%) freq: 3810 (0.02%)
4 values: 0 times (0.00%) freq: 0 (0.00%)
So with 4 targets we seem to be able to track 98.3% of trained branhces while
with reproducible merging this drops to 45%
I will send you the hacked script
