1079126 - 7.5% regression in kraken-parse-financial on MacOSX 32bit

Status: RESOLVED WONTFIX

(Core :: JavaScript Engine: JIT, defect)

Description

[Hannes Verschore [:h4writer]]

There is a 7.5% regression on the benchmark kraken-parse-financial on MacOSX 32bit:
http://arewefastyet.com/#machine=11&view=single&suite=kraken&subtest=parse-financial&start=1412330000&end=1412342618

The regression range points to:
http://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=2f7bdbc8beb2&tochange=e25f541006dc

Comment 1

[Nicolas B. Pierron [:nbp]]

parse-financial is just a noisy benchmark.  For example, I do not think that SIMD is used by this benchmark, but this still improve the score.

http://arewefastyet.com/#machine=11&view=single&suite=kraken&subtest=parse-financial&start=1412746813&end=1412785695

http://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=1b195cb0d08d&tochange=2ae98c304c5e

Some benchmark seems to be highly dependent on code location and we have no control on that.

Comment 2

[Hannes Verschore [:h4writer]]

(In reply to Nicolas B. Pierron [:nbp] from comment #1)
> parse-financial is just a noisy benchmark.  For example, I do not think that
> SIMD is used by this benchmark, but this still improve the score.
> 
> http://arewefastyet.com/#machine=11&view=single&suite=kraken&subtest=parse-
> financial&start=1412746813&end=1412785695
> 
> http://hg.mozilla.org/integration/mozilla-inbound/
> pushloghtml?fromchange=1b195cb0d08d&tochange=2ae98c304c5e
> 
> Some benchmark seems to be highly dependent on code location and we have no
> control on that.

That change can be explained by bug 1076091, which is not related to simd

Comment 3

[Hannes Verschore [:h4writer]]

(In reply to Nicolas B. Pierron [:nbp] from comment #1)
> parse-financial is just a noisy benchmark.  For example, I do not think that
> SIMD is used by this benchmark, but this still improve the score.

Could you reproduce it? And have you taken a look what changed with your patches?
I know we have some benchmarks which are noisy and depend on code location. Didn't know this was one of them...

Comment 4

[Nicolas B. Pierron [:nbp]]

(In reply to Hannes Verschore [:h4writer] from comment #3)
> (In reply to Nicolas B. Pierron [:nbp] from comment #1)
> > parse-financial is just a noisy benchmark.  For example, I do not think that
> > SIMD is used by this benchmark, but this still improve the score.
> 
> Could you reproduce it?

No,


* 014e7fd - (HEAD, bugzil.la/1079126/after) Bug 1072188 - Flag resume point operands of branches removed by UCE. r=sunfish (7 days ago) <Nicolas B. Pierron>
* 8ea2421 - Bug 1072188 - Recover truncated instruction. r=sunfish (7 days ago) <Nicolas
B. Pierron>
* f50a307 - Bug 1072188 - IonMonkey: Split truncate function. r=sunfish (7 days ago) <Nicolas B. Pierron>
* aae0cf9 - (bugzil.la/1079126/before) Bug 1076941: … <Stephen Pohl>

exec: bash -c perl ./sunspider --suite=kraken-1.1 --shell="/home/nicolas/mozilla/alternate-dev/js/src/_build/js-bugzil.la-1079126-before-opt-x64-clang33" --runs=50 --tests=finan
| sed -n '/====/ { :b; n; p; b b; }'

    parse-financial:  35.8ms +/- 0.5%

exec: bash -c perl ./sunspider --suite=kraken-1.1 --shell="/home/nicolas/mozilla/alternate-dev/js/src/_build/js-bugzil.la-1079126-after-opt-x64-clang33" --runs=50 --tests=finan | sed -n '/====/ { :b; n; p; b b; }'

    parse-financial:  35.8ms +/- 0.6%

Comment 5

[Hannes Verschore [:h4writer]]

I had a look what was going wrong. So apparently the extra virtual function introduced in the first patch caused this regression. So something not really actionable.