code review 6258067: runtime: add arm soft division and modulo benchmarks

runtime: add arm soft division and modulo benchmarks

arm soft div and mod performance plays an important part
in runtime performance.

The currently implementation is very stable, but we believe
we can best it. This benchmark provides a reference point.

linux/arm omap4 pandaboard

BenchmarkUint32Div7     10000000               298 ns/op
BenchmarkUint32Div37    10000000               298 ns/op
BenchmarkUint32Div123   10000000               298 ns/op
BenchmarkUint32Div763   10000000               298 ns/op
BenchmarkUint32Div1247  10000000               299 ns/op
BenchmarkUint32Div9305  10000000               298 ns/op
BenchmarkUint32Div13307 10000000               298 ns/op
BenchmarkUint32Div52513 10000000               298 ns/op
BenchmarkUint32Div60978747      10000000               298 ns/op
BenchmarkUint32Div106956295     10000000               297 ns/op
BenchmarkUint32Mod7     10000000               280 ns/op
BenchmarkUint32Mod37    10000000               280 ns/op
BenchmarkUint32Mod123   10000000               280 ns/op
BenchmarkUint32Mod763   10000000               280 ns/op
BenchmarkUint32Mod1247  10000000               280 ns/op
BenchmarkUint32Mod9305  10000000               280 ns/op
BenchmarkUint32Mod13307 10000000               280 ns/op
BenchmarkUint32Mod52513 10000000               280 ns/op
BenchmarkUint32Mod60978747      10000000               280 ns/op
BenchmarkUint32Mod106956295     10000000               280 ns/op

[dave_cheney.net, 2012-06-14 03:57:38]

Hello minux.ma@gmail.com, rsc@golang.org (cc: golang-dev@googlegroups.com),

I'd like you to review this change to
https://code.google.com/p/go

[rsc, 2012-06-14 04:00:40]

Do you really mean runtime? It is true that package runtime's
performance is affected by C division, but this is benchmarking Go
division, and they are not guaranteed to be related. Also this is
benchmarking division by variables, when the issue in the runtime is
(I believe) division by constants, which can (and probably don't but
should) be handled differently.

Russ

[dave_cheney.net, 2012-06-14 04:03:53]

There are a few areas, mainly related to modulo that are still a
problem for arm in the runtime. itab (iface.c#71) is one, scanblock
(mgc0.c#249) is another.

On Thu, Jun 14, 2012 at 2:00 PM, Russ Cox <rsc@golang.org> wrote:
> Do you really mean runtime? It is true that package runtime's
> performance is affected by C division, but this is benchmarking Go
> division, and they are not guaranteed to be related. Also this is
> benchmarking division by variables, when the issue in the runtime is
> (I believe) division by constants, which can (and probably don't but
> should) be handled differently.
>
> Russ

[minux1, 2012-06-15 18:19:01]

On 2012/06/14 04:00:40, rsc wrote:
> Do you really mean runtime? It is true that package runtime's
> performance is affected by C division, but this is benchmarking Go
> division, and they are not guaranteed to be related. Also this is
As ARM cores generally don't have hardware div/mod operations,
its div/mod operation is implemented in runtime, so IMO, add a
benchmark for it is worthwhile (I've already got a CL for it).
> benchmarking division by variables, when the issue in the runtime is
> (I believe) division by constants, which can (and probably don't but
> should) be handled differently.
yeah, 5g and 8g don't even special case power-of-2 divisors; 5c/8c only
turn simple power-of-2 division into shifts, but no magic multiply
optimization.
also note that div/mod usage in runtime are almost all division by variables.

[dave_cheney.net, 2012-06-19 05:20:07]

<ping />

[rsc, 2012-06-25 21:37:53]

LGTM

in the description s/Arm/arm/

http://codereview.appspot.com/6258067/diff/7002/src/pkg/runtime/vlop_arm_test.go
File src/pkg/runtime/vlop_arm_test.go (right):

http://codereview.appspot.com/6258067/diff/7002/src/pkg/runtime/vlop_arm_test...
src/pkg/runtime/vlop_arm_test.go:10: //
http://ridiculousfish.com/files/division_benchmarks.tar.gz
FWIW ridiculousfish suggests those benchmarks because they are difficult for
compilers to optimize when dividing by a constant. It is less clear that they're
interesting for non-constant division, but they're still fine.

[dave_cheney.net, 2012-06-25 22:59:18]

*** Submitted as http://code.google.com/p/go/source/detail?r=6f2a9ecdccf9 ***

runtime: add arm soft division and modulo benchmarks

arm soft div and mod performance plays an important part
in runtime performance.

The currently implementation is very stable, but we believe
we can best it. This benchmark provides a reference point.

linux/arm omap4 pandaboard

BenchmarkUint32Div7     10000000               298 ns/op
BenchmarkUint32Div37    10000000               298 ns/op
BenchmarkUint32Div123   10000000               298 ns/op
BenchmarkUint32Div763   10000000               298 ns/op
BenchmarkUint32Div1247  10000000               299 ns/op
BenchmarkUint32Div9305  10000000               298 ns/op
BenchmarkUint32Div13307 10000000               298 ns/op
BenchmarkUint32Div52513 10000000               298 ns/op
BenchmarkUint32Div60978747      10000000               298 ns/op
BenchmarkUint32Div106956295     10000000               297 ns/op
BenchmarkUint32Mod7     10000000               280 ns/op
BenchmarkUint32Mod37    10000000               280 ns/op
BenchmarkUint32Mod123   10000000               280 ns/op
BenchmarkUint32Mod763   10000000               280 ns/op
BenchmarkUint32Mod1247  10000000               280 ns/op
BenchmarkUint32Mod9305  10000000               280 ns/op
BenchmarkUint32Mod13307 10000000               280 ns/op
BenchmarkUint32Mod52513 10000000               280 ns/op
BenchmarkUint32Mod60978747      10000000               280 ns/op
BenchmarkUint32Mod106956295     10000000               280 ns/op

R=minux.ma, rsc
CC=golang-dev
http://codereview.appspot.com/6258067