runtime: improve memclr on ppc64x/power10
Rewrite memclr asm function to use the new power10 instruction stxvl
or the store vector with length which can specify the number of bytes
to be stored in a register, thereby avoiding loops to store the tail
end bytes.
On power9 and power8 the code remains unchanged.
The performance for all sizes<16 improve on power10 with this change.
name old time/op new time/op delta
Memclr/1 2.59ns ± 1% 2.80ns ± 9% ~
Memclr/2 2.77ns ± 0% 2.90ns ± 8% ~
Memclr/3 3.70ns ± 1% 3.00ns ± 1% -19.02%
Memclr/4 4.56ns ± 5% 2.98ns ± 1% -34.56%
Memclr/5 10.1ns ± 8% 2.9ns ± 4% -70.82%
Memclr/6 6.99ns ± 4% 2.84ns ± 9% -59.40%
Memclr/7 8.66ns ± 5% 2.92ns ± 4% -66.27%
Memclr/8 2.75ns ± 1% 2.75ns ± 0% ~
Memclr/9 3.43ns ± 1% 2.75ns ± 0% -19.78%
Memclr/10 3.30ns ± 0% 2.77ns ± 0% -15.81%
Memclr/12 5.09ns ± 0% 2.78ns ± 0% -45.49%
Memclr/15 9.79ns ± 0% 2.78ns ± 0% -71.64%
Memclr/16 2.65ns ± 1% 2.77ns ± 0% +4.77%
GoMemclr/1 2.46ns ± 3% 2.46ns ± 3% ~
GoMemclr/2 2.91ns ± 0% 2.46ns ± 1% -15.38%
GoMemclr/3 4.85ns ± 6% 2.45ns ± 1% -49.50%
GoMemclr/4 5.44ns ± 5% 2.47ns ± 1% -54.63%
GoMemclr/5 6.48ns ± 0% 2.45ns ± 1% -62.20%
GoMemclr/6 9.33ns ±20% 2.45ns ± 1% -73.76%
GoMemclr/7 8.79ns ± 2% 2.46ns ± 1% -72.03%
GoMemclr/8 2.91ns ± 0% 2.84ns ± 0% -2.35%
GoMemclr/9 5.04ns ± 0% 2.84ns ± 0% -43.67%
GoMemclr/10 5.17ns ± 1% 2.84ns ± 0% -44.99%
GoMemclr/12 6.22ns ± 0% 2.84ns ± 0% -54.41%
GoMemclr/15 10.1ns ± 0% 2.8ns ± 0% -72.04%
GoMemclr/16 2.90ns ± 0% 2.92ns ± 0% +0.83%
Change-Id: I2b7e1a9ceafba11c57c04a2e8cbd12b5ac87ec9b
Reviewed-on: https://go-review.googlesource.com/c/go/+/467635
Reviewed-by: Lynn Boger <laboger@linux.vnet.ibm.com>
Reviewed-by: Than McIntosh <thanm@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Archana Ravindar <aravind5@in.ibm.com>