Bug 79199 – Zero out CopiedBlocks on initialization

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Mark Hahnenberg]

We should do this in preparation for refactoring op_new_array. This turns out to be a performance win, even if we still manually clear the arrays.

[Mark Hahnenberg]

Created attachment 128139 [details]
Patch

[Mark Hahnenberg]

Created attachment 128140 [details]
Bencher results

~1.5% win on v8, neutral on Kraken and SunSpider

[Philippe Normand]

Comment on attachment 128139 [details]
Patch

Attachment 128139 [details] did not pass gtk-ews (gtk):
Output: http://queues.webkit.org/results/11557399

[Early Warning System Bot]

Comment on attachment 128139 [details]
Patch

Attachment 128139 [details] did not pass qt-ews (qt):
Output: http://queues.webkit.org/results/11561397

[Gyuyoung Kim]

Comment on attachment 128139 [details]
Patch

Attachment 128139 [details] did not pass efl-ews (efl):
Output: http://queues.webkit.org/results/11558397

[Filip Pizlo]

Comment on attachment 128139 [details]
Patch

Why not just have a thing that blitzes empty JS values on 32_64, instead of making 32_64 pay the price twice?

[Filip Pizlo]

Ahhh, the wonders of bzero.  You have two options here:

- Use memset instead, which is generally a good idea since bzero is somewhat deprecated.  Be sure to retest performance in case there's some weirdness with the compiler doing smarter things to bzero than it did to memset.

- Include all the headers.  Different platforms seem to put it in different places.  It might be in strings.h, or string.h, or stdlib.h.  But then strings.h might not exist on some platforms, ...  Yuck.

Moral of the story: just use memset if you can.

[Mark Hahnenberg]

(In reply to comment #6)
> (From update of attachment 128139 [details])
> Why not just have a thing that blitzes empty JS values on 32_64, instead of making 32_64 pay the price twice?

The problem with this strategy is that on 32-bit, we alternate 32-bits of zero and 32-bits of non-zero. If one of the allocations into the block is size n where n mod 8 = 4, we get out of sync with this pattern, thus requiring us to re-clear the arrays anyways. I guess we could potentially keep track of whether we're in sync or not inside each CopiedBlock and just check if we need to manually clear or not?

[Mark Hahnenberg]

Created attachment 128394 [details]
Patch

[Filip Pizlo]

Comment on attachment 128394 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=128394&action=review

> Source/JavaScriptCore/heap/CopiedBlock.h:54
> +        ASSERT(is8ByteAligned(static_cast<void*>(m_payload)));
> +        JSValue emptyValue;
> +#if OS(MAC_OS_X) || OS(IOS)
> +        memset_pattern8(static_cast<void*>(m_payload), static_cast<void*>(&emptyValue), static_cast<size_t>((reinterpret_cast<char*>(this) + allocation.size()) - m_payload));
> +#else
> +        JSValue* limit = reinterpret_cast<JSValue*>(reinterpret_cast<char*>(this) + allocation.size());
> +        for (JSValue* currentValue = reinterpret_cast<JSValue*>(m_payload); currentValue < limit; currentValue++)
> +            *currentValue = emptyValue;
> +#endif

For good measure, I'd have a case for USE(JSVALUE64) in which I call memset.  I don't trust memset_pattern8 to be as fast as memset if the pattern is zero.

Also, did you verify that there is any speed-up with memset_pattern8 at all on JSVALUE32_64?

[Mark Hahnenberg]

> Also, did you verify that there is any speed-up with memset_pattern8 at all on JSVALUE32_64?

Yep. The speedup is less dramatic, but it's still there.

[Mark Hahnenberg]

Created attachment 128409 [details]
32-bit bencher results

[Filip Pizlo]

(In reply to comment #11)
> > Also, did you verify that there is any speed-up with memset_pattern8 at all on JSVALUE32_64?
> 
> Yep. The speedup is less dramatic, but it's still there.

I mean, is there any speed-up over writing your own loop?  Like the one that you have when not on IOS or MAC_OS_X?

(And btw you can use test both of them as DARWIN.  I don't remember is it's OS(DARWIN) or PLATFORM(DARWIN).)

[Mark Hahnenberg]

Created attachment 128507 [details]
Patch

[Mark Hahnenberg]

Created attachment 128508 [details]
32-bit results with just copying JSValues (no memset_pattern)

There wasn't much of a difference (and the JSValue version might even be a tiny bit faster) so I just got rid of the memset_pattern part altogether to reduce the amount of preprocessor cruft.

[WebKit Review Bot]

Comment on attachment 128507 [details]
Patch

Clearing flags on attachment: 128507

Committed r108716: <http://trac.webkit.org/changeset/108716>

[WebKit Review Bot]

All reviewed patches have been landed.  Closing bug.

[Csaba Osztrogonác]

Reopen, because it made all tests crash on 32 bit Qt.
Maybe on other 32 bit platforms too, but I don't know,
because build.webkit.org is unavailable nowadays.

http://build.webkit.sed.hu/results/x86-32%20Linux%20Qt%20Release%20NRWT/r108716%20%286626%29/results.html


$ cat block-wrappers-necessary-crash-log.txt | c++filt
1   0x8066e9b /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/bin/DumpRenderTree() [0x8066e9b]
2   0xf775e400 [0xf775e400]
3   0xf7023981 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WTF::fastRealloc(void*, unsigned int)+0xa21) [0xf7023981]
4   0xf6e81981 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(JSC::JIT::privateCompile(JSC::MacroAssemblerCodePtr*)+0xd21) [0xf6e81981]
5   0xf6f6f767 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(JSC::JIT::compile(JSC::JSGlobalData*, JSC::CodeBlock*, JSC::MacroAssemblerCodePtr*)+0x77) [0xf6f6f767]
6   0xf6f7134d /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(JSC::jitCompileFunctionIfAppropriate(JSC::JSGlobalData&, WTF::OwnPtr<JSC::FunctionCodeBlock>&, JSC::JITCode&, JSC::MacroAssemblerCodePtr&, JSC::SharedSymbolTable*&, JSC::JITCode::JITType)+0x12d) [0xf6f7134d]
7   0xf6f6cc0f /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(JSC::FunctionExecutable::compileForCallInternal(JSC::ExecState*, JSC::ScopeChainNode*, JSC::JITCode::JITType)+0x13f) [0xf6f6cc0f]
8   0xf6e06b68 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(JSC::FunctionExecutable::compileFor(JSC::ExecState*, JSC::ScopeChainNode*, JSC::CodeSpecializationKind)+0x68) [0xf6e06b68]
9   0xf6ea6715 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(cti_vm_lazyLinkCall+0x185) [0xf6ea6715]
10  0xf0d050c8 [0xf0d050c8]
11  0xf6e5b4f1 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(JSC::Interpreter::executeCall(JSC::ExecState*, JSC::JSObject*, JSC::CallType, JSC::CallData const&, JSC::JSValue, JSC::ArgList const&)+0x401) [0xf6e5b4f1]
12  0xf6f56ad6 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(JSC::call(JSC::ExecState*, JSC::JSValue, JSC::CallType, JSC::CallData const&, JSC::JSValue, JSC::ArgList const&)+0x66) [0xf6f56ad6]
13  0xf5e3f783 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::JSEventListener::handleEvent(WebCore::ScriptExecutionContext*, WebCore::Event*)+0x453) [0xf5e3f783]
14  0xf6021ddd /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::EventTarget::fireEventListeners(WebCore::Event*, WebCore::EventTargetData*, WTF::Vector<WebCore::RegisteredEventListener, 1u>&)+0x12d) [0xf6021ddd]
15  0xf6021f9d /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::EventTarget::fireEventListeners(WebCore::Event*)+0x5d) [0xf6021f9d]
16  0xf63960da /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::DOMWindow::dispatchEvent(WTF::PassRefPtr<WebCore::Event>, WTF::PassRefPtr<WebCore::EventTarget>)+0xca) [0xf63960da]
17  0xf639d137 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::DOMWindow::dispatchLoadEvent()+0x117) [0xf639d137]
18  0xf5fd8c3e /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::Document::dispatchWindowLoadEvent()+0x2e) [0xf5fd8c3e]
19  0xf5fe2762 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::Document::implicitClose()+0x122) [0xf5fe2762]
20  0xf631da34 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::FrameLoader::checkCallImplicitClose()+0x74) [0xf631da34]
21  0xf6326b11 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::FrameLoader::checkCompleted()+0xc1) [0xf6326b11]
22  0xf6326c7d /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::FrameLoader::loadDone()+0x1d) [0xf6326c7d]
23  0xf630bdeb /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::CachedResourceLoader::loadDone()+0x4b) [0xf630bdeb]
24  0xf635bc8f /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::SubresourceLoader::releaseResources()+0x5f) [0xf635bc8f]
25  0xf6351a7a /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::ResourceLoader::didFinishLoading(double)+0x3a) [0xf6351a7a]
26  0xf635c551 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::SubresourceLoader::didFinishLoading(double)+0xe1) [0xf635c551]
27  0xf6351541 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::ResourceLoader::didFinishLoading(WebCore::ResourceHandle*, double)+0x71) [0xf6351541]
28  0xf666aa2d /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::QNetworkReplyHandler::finish()+0x1bd) [0xf666aa2d]
29  0xf6667394 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::QNetworkReplyHandlerCallQueue::flush()+0x64) [0xf6667394]
30  0xf6667934 /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::QNetworkReplyHandlerCallQueue::push(void (WebCore::QNetworkReplyHandler::*)())+0x34) [0xf6667934]
31  0xf666797f /home/webkitbuildbot/slaves/release32bit-NRWT/buildslave/qt-linux-32-release-NRWT/build/WebKitBuild/Release/lib/libQtWebKit.so.4(WebCore::QNetworkReplyWrapper::didReceiveFinished()+0x3f) [0xf666797f]

Zoltán, Gábor, Péter, could you check it?

[Csaba Osztrogonác]

Additionally there are crashes on Qt ARM too. :(

[Zoltan Herczeg]

I summarize what I know so far.

This is the new code:

JSValue emptyValue;
JSValue* limit = reinterpret_cast<JSValue*>(reinterpret_cast<char*>(this) + allocation.size());
for (JSValue* currentValue = reinterpret_cast<JSValue*>(m_payload); currentValue < limit; currentValue++)
    *currentValue = emptyValue; 

In DRT this function runs only once for dom/svg/level3/xpath/XPathEvaluator_evaluate_NAMESPACE_ERR.svg

the 'this' (0xf02f0000) and 'limit' (0xf0300000) variables seems ok, but m_payload contains 0xf02f002c which is not divisible by 8! I feel this is the root of the issues. I continue debugging.

[Zoltan Herczeg]

I suspect this cause the alignment error:

    void* m_offset;
    uintptr_t m_isPinned;
    uintptr_t m_padding;
    char m_payload[1];

This is 3*4 = 12 bytes on 32 bit, while it is 8*3 = 24 bytes on 64 bit. 24 bytes are divisible by 8, but 12 bytes are not!

[Zoltan Herczeg]

Temporary fix for http://trac.webkit.org/changeset/108779

[Mark Hahnenberg]

Bug 79556 should have taken care of any cross-platform alignment issues. Closing bug.