DFG should keep captured variables alive until the (inline) return.

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Filip Pizlo]

Otherwise OSR exit may decided to set them to Undefined.

[Filip Pizlo]

Currently a variable may appear to be dead at a basic block boundary even though it is captured. As a hack, we can peek at the first node of the successor block(s) and check what variables are captured at that node's code origin. But that's just disgusting and wrong.

The right solution is to change how captured variables are flushed. Currently we flush SetLocals by planting a Flush node right after them:

a: SetLocal(thingy)
b: Flush(@a)

But we could instead have a flush right before each SetLocal:

a: Flush(...)
b: SetLocal(thingy)

As well as a Flush for all captured variables at the return site. That will effectively keep variables alive through the whole span of code in which they are captured.

[Filip Pizlo]

Created attachment 144458 [details]
work in progress

[Filip Pizlo]

Created attachment 144464 [details]
it's starting to work

[Filip Pizlo]

Created attachment 144616 [details]
more

Still making sure that I've dotted all of my t's.

[Filip Pizlo]

Created attachment 144629 [details]
even more

It's starting to pass tests.

[Filip Pizlo]

Created attachment 144638 [details]
the patch

[Filip Pizlo]

Landed in http://trac.webkit.org/changeset/118858

[Kangil Han]

Just for your information, I've got a build warning on my laptop as below. :-)

[  5%] Building CXX object Source/JavaScriptCore/CMakeFiles/javascriptcore_efl.dir/dfg/DFGCapabilities.cpp.o
/home/kangilhan/dev/WebKit/Source/JavaScriptCore/dfg/DFGByteCodeParser.cpp: In member function ‘JSC::DFG::ArgumentPosition* JSC::DFG::ByteCodeParser::findArgumentPositionForLocal(int)’:
/home/kangilhan/dev/WebKit/Source/JavaScriptCore/dfg/DFGByteCodeParser.cpp:361:73: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]