https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89829
--- Comment #2 from Serge Belyshev <belyshev at depni dot sinp.msu.ru> ---
On the other hand, benchmarking shows that better training brings no advantage.
Or rather, slight measurable regression is apparent:
option | training dataset | benchmark | compiler binary
| | time, s | size, MB
-------------------------------------------------------------------------------
bootstrap | N/A | 11.203 | 32.2
profiledbootstrap | none | 11.518 | 38.0
profiledbootstrap | libgcc | 10.091 | 30.1
profiledbootstrap | stagetrain | 10.130 | 30.8
profiledbootstrap | libgcc+stagetrain | 10.116 | 30.9
profiledbootstrap | as above + combined tree | 10.128 | 30.9
--------------------------------------------------------------------------------
Thus, some tuning is needed, I guess.
Notes: BOOT_CFLAGS=-Ofast was used for the above. Benchmark is cc1files from
gcc 3.4, compiled by compiler under test with -Ofast. Time values are averages
of 100 runs, standard error of mean is about 0.001 s, sample standard deviaton
is 0.011 s.
