> Scheduling should never move very expensive instructions to places they > are executed more frequently. This patch fixes that, reducing the > execution time of c-ray by over 40% (I tested on a BE Power7 system). > > This introduces a new target hook sched.can_speculate_insn which returns > whether the scheduler is allowed to speculate a given instruction. The > rs6000 implementation disallows all divide and square root instructions. > > > PR rtl-optimization/68664 > * target.def (can_speculate_insn): New hook. > * doc/tm.texi.in (TARGET_SCHED_CAN_SPECULATE_INSN): New hook. > * doc/tm.texi: Regenerate. > * sched-rgn.c (can_schedule_ready_p): Use the new hook. > * config/rs6000/rs6000.c (TARGET_SCHED_CAN_SPECULATE_INSN): New macro. > (rs6000_sched_can_speculate_insn): New function.
There was also regression on cray for x86-64 https://gcc.opensuse.org/c++bench-czerny/c-ray/ Is it the same issue? Also with profile feedback perhaps you have enough info to tell that the speculative path is almost as likely as the original placement. I wonder if we can provie resonable default by RTX cost... Honza
