https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125305

--- Comment #9 from chenglulu <chenglulu at loongson dot cn> ---
(In reply to Xi Ruoyao from comment #5)
> I can still observe performance gain from scheduling on LA664 for e.g.
> https://git.kernel.org/torvalds/c/9805f39d423a.
> 
> AArch64 is even using 10 for
> TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD (when tuning for a core
> with such an issue rate) but they disable early scheduling at -O2 and only
> enable it at -O3 by default (they once disabled it at all but then lbm
> regressed for 6% on neoverse, again demonstrating the usefulness of
> scheduling even on OoO core).
> 
> x86 disables early scheduling by default for any optimization level.
> 
> IMO we should try disabling early scheduling for OoO cores like AArch64 or
> x86 instead of modifying the lookahead, so the people requiring "top
> performance" can still bring it back with -fschedule-insns.  And I think we
> need some benchmark results to make the decision.

I see what you mean now. However, I found that passing -fno-schedule-insns
during compilation still does not reduce the compilation time. According to the
-ftime-report, the prolonged compilation time is due to the fact that the vast
majority of the time is being spent in the sched2 pass.

Reply via email to