6 Dec
2023
6 Dec
'23
1:24 a.m.
Hello Tamar,
Thank you for your clarifications. They're very helpful.
Tamar Christina Tamar.Christina@arm.com writes:
tcwg_bmk-code_size-cpu2017fast
- gnu-aarch64-master-O2: grew in size by 3% - 508.namd_r
- gnu-aarch64-master-O2_LTO: grew in size by 2% - 508.namd_r
tcwg_bmk-code_size-spec2k6
- gnu-aarch64-master-O2: grew in size by 2% - 447.dealII
- gnu-aarch64-master-Os: grew in size by 2% - 456.hmmer
Size gains are expected. The cost model turns on throughput based optimizations. Since will grow because among other things it unrolls.
Ok, that is good to know.
- tcwg_bmk-code_speed-cpu2017rate
- gnu-aarch64-master-O2: slowed down by 5% - 519.lbm_r
- gnu-aarch64-master-O3: slowed down by 5% - 519.lbm_r
As the commit message says, there are outstanding costing issues that will be resolved at stage 4 when the remaining infrastructure are in place.
Nice.
- tcwg_bmk-fujitsu_speed-cpu2017speed
- gnu-aarch64-master-O2: slowed down by 19% - 644.nab_s:[.]
exp@@GLIBC_2.29
This is suspect, it says the slowdown is in exp in glibc, that would be unrelated to my patch.
Indeed, thanks for pointing out.
--
Thiago