Bug 21463 – 11.8% SunSpider regression in non-CTI from r37324

• Status: RESOLVED
• Resolution: WORKSFORME
• Component: JavaScriptCore

[Gabor Loki]

Hi,

we have noticed a 11.8% slow down on SunSpider from 37324 revision on Qt-linux port (it was 15.7% regression on V8).

Do you know this? Or does it appear only on linux?

You can also check our results from http://www.sed.hu/webkit/?page=timeline or http://www.sed.hu/webkit/?page=revisions&order=ASC&num=30&pg=26 pages.

[Cameron Zwarich (cpst)]

This is in bytecode, not CTI, right?

[Gabor Loki]

You are right. This happened in bytecode (CTI disabled).

As far as I know CTI is only enabled for Windows and Mac on x86 target.
Platform.h says: #if !defined(ENABLE_CTI) && PLATFORM(X86) && (PLATFORM(MAC) || PLATFORM(WIN))

Probably, I can check if it exists on qt-linux with CTI.

[Cameron Zwarich (cpst)]

Some of the Qt people have a branch based on the patch to enable CTI on Linux. I was wondering if you were using that.

Since this is my patch, I will check what happens on the Mac tomorrow.

[Gabor Loki]

Do you mean code.staikos.net? No, I did not use it yet, and I have not found any CTI related modification in that tree.

We have a separate patch to enable CTI for Qt port. I will send it as soon as possible.

[Gabor Loki]

(In reply to comment #3)
> Since this is my patch, I will check what happens on the Mac tomorrow.

Cameron, is there any progress in this issue?

You can check our performance results at:
http://www.sed.hu/webkit/?page=timelinerange&s_ss=on&s_ws=on&s_v8=on&from=37301&till=37396

[Cameron Zwarich (cpst)]

I'll try to take a look at it today on my machine. I suspect that it is something about your compiler version or flags, because no one tracking PPC performance has mentioned such a dramatic drop.

[Cameron Zwarich (cpst)]

I see no slowdown between those two revisions on my machine, and I actually see a slight speedup.

[Gabor Loki]

No regression? On Mac, am I right?
It would be nice if someone else can confirm or reject it on linux!

Anyway, this is not the only one regression over 10%.
Between 38153 and 38187: http://www.sed.hu/webkit/?page=timelinerange&s_ss=on&s_ws=on&s_v8=on&from=38114&till=38201

I will open another bug report if I find the concert changeset number.

[Cameron Zwarich (cpst)]

I am on Mac. Have you tried using the same compiler flags for optimizations as the Mac port, and using an AllInOne.cpp build?

[Gabor Loki]

Not yet.
We are using the 'WebKitTools/Scripts/build-webkit --qt' script which uses
DEFINES = -DBUILDING_QT__ -DNDEBUG -DUSE_SYSTEM_MALLOC -DBUILDING_QT__=1 -DUSE_SYSTEM_MALLOC -DNDEBUG -DBUILDING_QT__ -DQT_NO_DEBUG -DQT_CORE_LIB -DQT_SHARED
and
CFLAGS = -pipe -O2 -Wall -W -D_REENTRANT $(DEFINES)

without AllInOneFile.cpp method

As I see in the Makefiles neither GTK nor QT-port uses AllInOneFile.cpp feature.
Well, GTK uses AllInOneFile.cpp, but only in Debug mode.

[Alexey Proskuryakov]

This bug seems too old to be useful. Please feel free to re-open if this is still an issue.