On Thu, Jul 13, 2017 at 14:18:11 -1000, Richard Henderson wrote:
> The new title holder for perf top is helper_lookup_tb_ptr.
> Those targets that have a complicated cpu_get_tb_cpu_state
> function are going to regret that.
>
>
> This cleans up the Alpha version of that function such that it is
> just two loads and one mask. Which is one practically-free mask
> away from being as minimal as one can get.
Tested-by: Emilio G. Cota <cota@braap.org>
for the series.
I tried to get some perf numbers but really booting linux
doesn't spend much time in lookup_tb_ptr, nor does dbt-bench; so
I get very similar before/after numbers (slight perf decrease for
booting, tiny perf increase for dbt-bench). Numbers are below, FWIW.
Emilio
* I modified the gentoo-alpha image I'm using [1] to shut down once
it has fully booted. Results before/after this patchset:
Performance counter stats for 'taskset -c 0 alpha-softmmu/qemu-system-alpha \
-m 512 -drive \
file=../img/alpha/die-on-boot.img,media=disk,format=raw,index=0 \
-kernel ../img/alpha/vmlinux -append root=/dev/sda2 \
-accel accel=tcg,thread=single -smp 1 -nographic' (10 runs):
Before:
30586.631281 task-clock (msec) # 0.883 CPUs utilized ( +- 0.56% )
16,373 context-switches # 0.535 K/sec ( +- 1.16% )
1 cpu-migrations # 0.000 K/sec
10,269 page-faults # 0.336 K/sec ( +- 1.39% )
128,287,167,139 cycles # 4.194 GHz ( +- 0.55% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
244,179,137,606 instructions # 1.90 insns per cycle ( +- 0.66% )
45,088,775,217 branches # 1474.133 M/sec ( +- 0.61% )
267,065,722 branch-misses # 0.59% of all branches ( +- 0.84% )
34.639115913 seconds time elapsed ( +- 0.50% )
After:
31358.851235 task-clock (msec) # 0.892 CPUs utilized ( +- 1.07% )
16,352 context-switches # 0.521 K/sec ( +- 1.59% )
1 cpu-migrations # 0.000 K/sec
10,643 page-faults # 0.339 K/sec ( +- 1.18% )
131,620,007,449 cycles # 4.197 GHz ( +- 1.07% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
249,714,336,126 instructions # 1.90 insns per cycle ( +- 1.35% )
46,259,663,064 branches # 1475.171 M/sec ( +- 1.27% )
269,500,888 branch-misses # 0.58% of all branches ( +- 0.71% )
35.136529309 seconds time elapsed ( +- 0.99% )
perf diff doesn't show anything interesting (all differences, <1%, are due to kernel code)
* DBT-bench before/after:
NBench score, higher is better
100 +-+---+-----+-----+----+-----+-----+-----+-----+-----+----+-----+---+-+
| ***## ***## |
90 +-+..................*+*.#.......*.*.#.................before +-+
| * * # * * # after |
| ***# * * # +++++ * * # |
80 +-+.......***##.*.*#.*.*.#.***##.*.*.#..............................+-+
| * * # * *# * * # * * # * * # |
70 +-+.......*.*.#.*.*#.*.*.#.*.*.#.*.*.#..............................+-+
| * * # * *# * * # * * # * * # |
| * * # * *# * * # * * # * * # |
60 +-+.......*.*.#.*.*#.*.*.#.*.*.#.*.*.#..............................+-+
| * * # * *# * * # * * # * * # ***## |
50 +-+.......*.*.#.*.*#.*.*.#.*.*.#.*.*.#.*.*.#........................+-+
| * * # * *# * * # * * # * * # * * # |
| * * # * *# * * # * * # * * # * * # |
40 +-+.......*.*.#.*.*#.*.*.#.*.*.#.*.*.#.*.*.#........................+-+
| ***## * * # * *# * * # * * # * * # * * # |
30 +-+.*.*.#.*.*.#.*.*#.*.*.#.*.*.#.*.*.#.*.*.#........................+-+
| * * # * * # * *# * * # * * # * * # * * # ***## |
| * * # * * # * *# * * # * * # * * # * * # * * # |
20 +-+.*.*.#.*.*.#.*.*#.*.*.#.*.*.#.*.*.#.*.*.#..................*.*.#.+-+
| * * # * * # * *# * * # * * # * * # * * # * * # |
10 +-+.*.*.#.*.*.#.*.*#.*.*.#.*.*.#.*.*.#.*.*.#..................*.*.#.+-+
| * * # * * # * *# * * # * * # * * # * * # * * # |
| * * # * * # * *# * * # * * # * * # * * # ***# ***## * * # |
0 +-+-***##-***##-***#-***##-***##-***##-***##-***##-***#-***##-***##-+-+
STRING SOBFP EMULAASSIGNMENT IDEHUFFMAFOLU DECOMPOSITION gmean
png: http://imgur.com/oFFYSKd
[1] https://lists.gnu.org/archive/html/qemu-devel/2017-05/msg00630.html