[JSC] Improve the memory locality of DFG Node's AbstractValues
https://bugs.webkit.org/show_bug.cgi?id=160443
Patch by Benjamin Poulain <[email protected]> on 2016-08-03
Reviewed by Mark Lam.
The AbstractInterpreter spends a lot of time on memory operations
for AbstractValues. This patch attempts to improve the situation
by putting the values closer together in memory.
First, AbstractValue is moved out of DFG::Node and it kept in
a vector addressed by node indices.
I initially moved them to InPlaceAbstractState but I quickly discovered
initializing the values in the vector was costly.
I moved the vector to Graph as a cache shared by every instantiation of
InPlaceAbstractState. It is mainly there to avoid constructors and destructors
of AbstractValue. The patch of https://bugs.webkit.org/show_bug.cgi?id=160370
should also help eventually.
I instrumented CFA to find how packed is SparseCollection.
The answer is it can be very sparse, which is bad for CFA.
I added packIndices() to repack the collection before running
liveness since that's where we start using the memory intensively.
This is a measurable improvement but it implies we can no longer
keep indices on a side channel between phases since they may change.
(JSC::B3::SparseCollection::packIndices):
(JSC::DFG::Graph::packNodeIndices):
(JSC::DFG::Graph::abstractValuesCache):
- dfg/DFGInPlaceAbstractState.cpp:
(JSC::DFG::InPlaceAbstractState::InPlaceAbstractState):
- dfg/DFGInPlaceAbstractState.h:
(JSC::DFG::InPlaceAbstractState::forNode):
- dfg/DFGLivenessAnalysisPhase.cpp:
(JSC::DFG::performLivenessAnalysis):
