gccgo produces much slower code than gc
805 views
Skip to first unread message
Tamir Duberstein
Feb 14, 2018, 4:20:11 PM2/14/18
to golan...@googlegroups.com
Running the benchmarks in github.com/cockroachdb/cockroach/pkg/roachpb:
I chose this package because it doesn't depend on any of the fancy Makefile magic in the CockroachDB repo; you should be able to reproduce these results using just the go tool.
PATH=$HOME/local/go1.10/bin:$PATH
go test -i ./pkg/roachpb && go test -run - -bench . ./pkg/roachpb -count 5 -benchmem > gc.txt
PATH=$GCCROOT/gcc/bin:$PATH
go test -i ./pkg/roachpb && go test -run - -bench . ./pkg/roachpb -count 5 -benchmem > gccgo.txt
benchstat gc.txt gccgo.txt
name old time/op new time/op delta
ValueSetBytes-16 38.7ns ± 1% 188.8ns ±44% +388.36% (p=0.008 n=5+5)
ValueSetFloat-16 27.6ns ± 1% 112.4ns ± 4% +306.95% (p=0.008 n=5+5)
ValueSetBool-16 29.5ns ± 0% 69.5ns ± 7% +135.59% (p=0.008 n=5+5)
ValueSetInt-16 35.9ns ± 1% 89.0ns ± 5% +147.83% (p=0.008 n=5+5)
ValueSetProto-16 45.5ns ± 0% 127.4ns ± 0% +180.00% (p=0.008 n=5+5)
ValueSetTime-16 52.4ns ± 1% 136.4ns ± 0% +160.40% (p=0.008 n=5+5)
ValueSetDecimal-16 95.5ns ± 1% 255.0ns ± 0% +166.96% (p=0.008 n=5+5)
ValueSetTuple-16 38.7ns ± 1% 116.0ns ± 0% +200.05% (p=0.016 n=5+4)
ValueGetBytes-16 9.22ns ± 0% 31.60ns ± 0% +242.66% (p=0.008 n=5+5)
ValueGetFloat-16 12.0ns ± 0% 49.9ns ± 0% +315.83% (p=0.016 n=4+5)
ValueGetBool-16 14.4ns ± 0% 39.3ns ± 0% +172.92% (p=0.029 n=4+4)
ValueGetInt-16 13.7ns ± 0% 37.3ns ± 0% +172.12% (p=0.016 n=4+5)
ValueGetProto-16 26.1ns ± 0% 60.5ns ± 0% +131.72% (p=0.016 n=4+5)
ValueGetTime-16 39.6ns ± 0% 172.0ns ± 0% +334.34% (p=0.008 n=5+5)
ValueGetDecimal-16 95.1ns ± 0% 264.0ns ± 0% +177.49% (p=0.008 n=5+5)
ValueGetTuple-16 9.84ns ± 0% 31.50ns ± 0% +220.25% (p=0.008 n=5+5)
name old alloc/op new alloc/op delta
ValueSetBytes-16 32.0B ± 0% 32.0B ± 0% ~ (all equal)
ValueSetFloat-16 16.0B ± 0% 16.0B ± 0% ~ (all equal)
ValueSetBool-16 8.00B ± 0% 8.00B ± 0% ~ (all equal)
ValueSetInt-16 16.0B ± 0% 16.0B ± 0% ~ (all equal)
ValueSetProto-16 8.00B ± 0% 8.00B ± 0% ~ (all equal)
ValueSetTime-16 16.0B ± 0% 16.0B ± 0% ~ (all equal)
ValueSetDecimal-16 32.0B ± 0% 32.0B ± 0% ~ (all equal)
ValueSetTuple-16 32.0B ± 0% 32.0B ± 0% ~ (all equal)
ValueGetBytes-16 0.00B 0.00B ~ (all equal)
ValueGetFloat-16 0.00B 0.00B ~ (all equal)
ValueGetBool-16 0.00B 0.00B ~ (all equal)
ValueGetInt-16 0.00B 0.00B ~ (all equal)
ValueGetProto-16 0.00B 0.00B ~ (all equal)
ValueGetTime-16 0.00B 0.00B ~ (all equal)
ValueGetDecimal-16 48.0B ± 0% 48.0B ± 0% ~ (all equal)
ValueGetTuple-16 0.00B 0.00B ~ (all equal)
name old allocs/op new allocs/op delta
ValueSetBytes-16 1.00 ± 0% 1.00 ± 0% ~ (all equal)
ValueSetFloat-16 1.00 ± 0% 1.00 ± 0% ~ (all equal)
ValueSetBool-16 1.00 ± 0% 1.00 ± 0% ~ (all equal)
ValueSetInt-16 1.00 ± 0% 1.00 ± 0% ~ (all equal)
ValueSetProto-16 1.00 ± 0% 1.00 ± 0% ~ (all equal)
ValueSetTime-16 1.00 ± 0% 1.00 ± 0% ~ (all equal)
ValueSetDecimal-16 1.00 ± 0% 1.00 ± 0% ~ (all equal)
ValueSetTuple-16 1.00 ± 0% 1.00 ± 0% ~ (all equal)
ValueGetBytes-16 0.00 0.00 ~ (all equal)
ValueGetFloat-16 0.00 0.00 ~ (all equal)
ValueGetBool-16 0.00 0.00 ~ (all equal)
ValueGetInt-16 0.00 0.00 ~ (all equal)
ValueGetProto-16 0.00 0.00 ~ (all equal)
ValueGetTime-16 0.00 0.00 ~ (all equal)
ValueGetDecimal-16 1.00 ± 0% 1.00 ± 0% ~ (all equal)
ValueGetTuple-16 0.00 0.00 ~ (all equal)
I chose this package because it doesn't depend on any of the fancy Makefile magic in the CockroachDB repo; you should be able to reproduce these results using just the go tool.
Are these results expected? I did minimal digging using pprof and perf but nothing obvious jumps out - things are just slower across the board. These results are on linux amd64.
Ian Lance Taylor
Feb 14, 2018, 7:44:55 PM2/14/18
to Tamir Duberstein, golang-nuts
Ian
Tamir Duberstein
Feb 15, 2018, 9:43:33 AM2/15/18
to Ian Lance Taylor, golang-nuts
Built at this revision: https://github.com/gcc-mirror/gcc/commit/a82f431e184a9ac922ad43df73cdcc702ab0f279
Ian Lance Taylor
Feb 15, 2018, 1:59:32 PM2/15/18
to Tamir Duberstein, golang-nuts
On Thu, Feb 15, 2018 at 6:42 AM, Tamir Duberstein <tam...@gmail.com> wrote:
> Built at this revision:
> https://github.com/gcc-mirror/gcc/commit/a82f431e184a9ac922ad43df73cdcc702ab0f279
Thanks. What do you see from
> Built at this revision:
> https://github.com/gcc-mirror/gcc/commit/a82f431e184a9ac922ad43df73cdcc702ab0f279
go test -gccgoflags="-g -O2"
?
Ian
Ian Lance Taylor
Feb 15, 2018, 2:00:34 PM2/15/18
to Tamir Duberstein, golang-nuts
On Thu, Feb 15, 2018 at 10:59 AM, Ian Lance Taylor <ia...@golang.org> wrote:
> On Thu, Feb 15, 2018 at 6:42 AM, Tamir Duberstein <tam...@gmail.com> wrote:
>> Built at this revision:
>> https://github.com/gcc-mirror/gcc/commit/a82f431e184a9ac922ad43df73cdcc702ab0f279
>
> Thanks. What do you see from
>
> go test -gccgoflags="-g -O2"
>
> ?
Sorry, make that
> On Thu, Feb 15, 2018 at 6:42 AM, Tamir Duberstein <tam...@gmail.com> wrote:
>> Built at this revision:
>> https://github.com/gcc-mirror/gcc/commit/a82f431e184a9ac922ad43df73cdcc702ab0f279
>
> Thanks. What do you see from
>
> go test -gccgoflags="-g -O2"
>
> ?
go test -gccgoflags=all="-g -O2"
Tamir Duberstein
Feb 15, 2018, 2:32:43 PM2/15/18
to Ian Lance Taylor, golang-nuts
What does all do? Anyway, the results are better, but still not "good":
name old time/op new time/op delta
ValueSetBytes-16 38.5ns ± 0% 105.8ns ± 4% +174.81% (p=0.008 n=5+5)
ValueSetFloat-16 27.5ns ± 1% 73.2ns ± 1% +166.38% (p=0.008 n=5+5)
ValueSetBool-16 29.4ns ± 0% 52.2ns ± 5% +77.77% (p=0.016 n=4+5)
ValueSetInt-16 34.0ns ± 1% 74.8ns ± 1% +119.62% (p=0.008 n=5+5)
ValueSetProto-16 45.4ns ± 0% 87.8ns ± 1% +93.57% (p=0.008 n=5+5)
ValueSetTime-16 52.9ns ± 1% 111.4ns ±18% +110.67% (p=0.008 n=5+5)
ValueSetDecimal-16 94.6ns ± 0% 214.2ns ±36% +126.43% (p=0.008 n=5+5)
ValueSetTuple-16 38.7ns ± 0% 105.6ns ± 3% +172.87% (p=0.008 n=5+5)
ValueGetBytes-16 9.22ns ± 0% 11.60ns ± 0% +25.84% (p=0.008 n=5+5)
ValueGetFloat-16 12.0ns ± 0% 23.8ns ± 0% +97.67% (p=0.016 n=5+4)
ValueGetBool-16 14.4ns ± 0% 15.2ns ± 0% +5.56% (p=0.029 n=4+4)
ValueGetInt-16 13.7ns ± 0% 14.6ns ± 0% +6.57% (p=0.016 n=5+4)
ValueGetProto-16 26.1ns ± 0% 22.6ns ± 0% -13.41% (p=0.008 n=5+5)
ValueGetTime-16 41.0ns ± 4% 78.9ns ± 0% +92.68% (p=0.008 n=5+5)
ValueGetDecimal-16 130ns ±24% 183ns ± 1% +40.29% (p=0.008 n=5+5)
ValueGetTuple-16 9.87ns ± 1% 11.60ns ± 0% +17.58% (p=0.008 n=5+5)
Ian Lance Taylor
Feb 15, 2018, 7:38:21 PM2/15/18
to Tamir Duberstein, golang-nuts
On Thu, Feb 15, 2018 at 11:31 AM, Tamir Duberstein <tam...@gmail.com> wrote:
> What does all do? Anyway, the results are better, but still not "good":
Using "all" applies the options to all packages, not just the one being built.
> What does all do? Anyway, the results are better, but still not "good":
Thanks for the benchmarks, it's something to look at.
Ian
Tamir Duberstein
May 3, 2018, 3:54:12 PM5/3/18
to Ian Lance Taylor, golang-nuts
Looks like performance is virtually identical in GCC 8.1:
ValueSetBytes-16 39.1ns ± 1% 104.5ns ± 0% +167.26% (p=0.029 n=4+4)
ValueSetFloat-16 25.9ns ± 1% 67.8ns ± 0% +161.78% (p=0.029 n=4+4)
ValueSetBool-16 27.5ns ± 0% 54.0ns ± 1% +96.18% (p=0.029 n=4+4)
ValueSetInt-16 35.1ns ± 4% 75.0ns ± 0% +113.51% (p=0.029 n=4+4)
ValueSetProto-16 45.9ns ± 0% 86.7ns ± 0% +88.89% (p=0.029 n=4+4)
ValueSetTime-16 52.7ns ± 1% 100.2ns ± 1% +90.32% (p=0.029 n=4+4)
ValueSetDecimal-16 88.9ns ± 1% 177.2ns ± 0% +99.33% (p=0.029 n=4+4)
ValueSetTuple-16 39.4ns ± 1% 104.5ns ± 0% +165.23% (p=0.029 n=4+4)
ValueGetBytes-16 9.48ns ± 0% 11.62ns ± 1% +22.66% (p=0.029 n=4+4)
ValueGetFloat-16 13.3ns ± 0% 23.9ns ± 1% +79.70% (p=0.029 n=4+4)
ValueGetBool-16 14.8ns ± 1% 15.2ns ± 0% +2.36% (p=0.029 n=4+4)
ValueGetInt-16 14.0ns ± 0% 14.6ns ± 0% +4.29% (p=0.029 n=4+4)
ValueGetProto-16 26.4ns ± 0% 22.7ns ± 0% -14.08% (p=0.029 n=4+4)
ValueGetTime-16 39.6ns ± 0% 79.0ns ± 0% +99.24% (p=0.029 n=4+4)
ValueGetDecimal-16 101ns ± 3% 182ns ± 0% +80.04% (p=0.029 n=4+4)
ValueGetTuple-16 9.14ns ± 0% 11.60ns ± 0% +26.85% (p=0.029 n=4+4)
Reply all
Reply to author
Forward
0 new messages