Bug 21500 – Qt-port on Linux should be built with g++ -O3 switch

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Csaba Osztrogonác]

WebKit Qt-port on Linux is built with g++ -O2 switch because of Qt default config (Qt-4.4.1/mkspecs/common/g++.conf) Building with -O3 switch instead of -O2, JSC is faster. (I used gcc version 4.1.2, Qt version 4.4.1 and WebKit rev37424)

runtime and code size of JSC:
- 3.51% speedup on SunSpider (+/-0.8%)
- 1.30% speedup on V8 (+/-0.65%)
- 3.81% speedup on WindScorpion (+/-1.0%)
- Code size of JSC .text section increased by 9.96%

I'll send patch soon.

[Csaba Osztrogonác]

Layout tests result - the whole WebKit built with -O3:
489.63s total testing time
7272 test cases (89%) succeeded
836 test cases (10%) had incorrect layout
3 test cases (<1%) crashed

Layout tests result - only JSC built with -O3:
513.98s total testing time
7277 test cases (89%) succeeded
831 test cases (10%) had incorrect layout
3 test cases (<1%) crashed

Layout tests result - original 37424 revision (with default -O2): 
531.12s total testing time
7272 test cases (89%) succeeded
836 test cases (10%) had incorrect layout
3 test cases (<1%) crashed

[Csaba Osztrogonác]

Created attachment 24230 [details]
proposed patch

[Alp Toker]

(In reply to comment #2)
> Created an attachment (id=24230) [edit]
> proposed patch
> 

It's incorrect to build WebCore with -O3 I believe (only JavaScriptCore should be -03).

[Csaba Osztrogonác]

(In reply to comment #3)
> It's incorrect to build WebCore with -O3 I believe (only JavaScriptCore should
> be -03).

Why is it incorrect? Does building WebCore with -O3 cause any problem? It seems to work fine (and faster) here with -O3. Maybe on other platforms have a problem with -O3. A little code size growth can be bad for embedded systems, but not for PC's.

[Simon Hausmann]

I agree with Alp here, only JavaScriptCore should be build with O3. Can you patch JavaScriptCore/JavaScriptCore.pri instead?

[Csaba Osztrogonác]

(In reply to comment #5)
I will create patch for ToT and measure its speedup soon.

[Csaba Osztrogonác]

I've created new patch for JavaScriptCore/JavaScriptCore.pri at r37894.

runtime and code size of JSC:
- 2.02% speedup on SunSpider (+/-0.8%)
- 1.10% speedup on V8 (+/-0.7%)
- 3.45% speedup on WindScorpion (+/-1.1%)
- Code size of JSC .text section increased by 9.70%

[Csaba Osztrogonác]

Created attachment 24686 [details]
proposed patch

[Csaba Osztrogonác]

new results for r38367:
- 1.41% speedup on SunSpider (+/-0.5%)
- 0.12% speedup on V8 (+/-0.3%) - not significant
- 6.33% speedup on WindScorpion (+/-0.3%)
- Code size of JSC .text section increased by 9.70%

There isn't activity a long time. Is anybody interested in landing this patch?

[Simon Hausmann]

Comment on attachment 24686 [details]
proposed patch

Looks good. I'll also add -fstrict-aliasing when landing like we do in WebCore/WebCore.pro, so that we are on the safe side.

[Simon Hausmann]

Nevermind the comment about aliasing, that was nonsense :-)

Landed your patch in r38398. Thanks!