-
Wei Xiao authored
Optimize arm64 CALLFN wrapper with LDP/STP instructions. This provides a significant speedup for big argument copy. Benchmark results for reflect: name old time/op new time/op delta Call-8 79.0ns ± 4% 73.6ns ± 4% -6.78% (p=0.000 n=10+10) CallArgCopy/size=128-8 80.5ns ± 0% 60.3ns ± 0% -25.06% (p=0.000 n=10+9) CallArgCopy/size=256-8 119ns ± 2% 67ns ± 1% -43.59% (p=0.000 n=8+10) CallArgCopy/size=1024-8 524ns ± 1% 99ns ± 1% -81.03% (p=0.000 n=10+10) CallArgCopy/size=4096-8 837ns ± 0% 231ns ± 1% -72.42% (p=0.000 n=9+9) CallArgCopy/size=65536-8 13.6µs ± 6% 3.1µs ± 1% -77.38% (p=0.000 n=10+10) PtrTo-8 12.9ns ± 0% 13.1ns ± 3% +1.86% (p=0.000 n=10+10) FieldByName1-8 28.7ns ± 2% 28.6ns ± 2% ~ (p=0.408 n=9+10) FieldByName2-8 928ns ± 4% 946ns ± 8% ~ (p=0.326 n=9+10) FieldByName3-8 5.35µs ± 5% 5.32µs ± 5% ~ (p=0.755 n=10+10) InterfaceBig-8 2.57ns ± 0% 2.57ns ± 0% ~ (all equal) InterfaceSmall-8 2.57ns ± 0% 2.57ns ± 0% ~ (all equal) New-8 9.09ns ± 1% 8.83ns ± 1% -2.81% (p=0.000 n=10+9) name old alloc/op new alloc/op delta Call-8 0.00B 0.00B ~ (all equal) name old allocs/op new allocs/op delta Call-8 0.00 0.00 ~ (all equal) name old speed new speed delta CallArgCopy/size=128-8 1.59GB/s ± 0% 2.12GB/s ± 1% +33.46% (p=0.000 n=10+9) CallArgCopy/size=256-8 2.14GB/s ± 2% 3.81GB/s ± 1% +78.02% (p=0.000 n=8+10) CallArgCopy/size=1024-8 1.95GB/s ± 1% 10.30GB/s ± 0% +427.99% (p=0.000 n=10+9) CallArgCopy/size=4096-8 4.89GB/s ± 0% 17.69GB/s ± 1% +261.87% (p=0.000 n=9+9) CallArgCopy/size=65536-8 4.84GB/s ± 6% 21.36GB/s ± 1% +341.67% (p=0.000 n=10+10) Change-Id: I775d88b30c43cb2eda1d0612ac15e6d283e70beb Reviewed-on: https://go-review.googlesource.com/70570Reviewed-by:
Cherry Zhang <cherryyz@google.com> Run-TryBot: Cherry Zhang <cherryyz@google.com> TryBot-Result: Gobot Gobot <gobot@golang.org>
18508740
