fourthTier: DFG should have an SSA form for use by FTL
https://bugs.webkit.org/show_bug.cgi?id=118338
Source/JavaScriptCore:
Reviewed by Mark Hahnenberg.
Adds an SSA form to the DFG. We can convert ThreadedCPS form into SSA form
after breaking critical edges. The conversion algorithm follows Aycock and
Horspool, and the SSA form itself follows something I've done before, where
instead of having Phi functions specify input nodes corresponding to block
predecessors, we instead have Upsilon functions in the predecessors that
specify which value in that block goes into which subsequent Phi. Upsilons
don't have to dominate Phis (usually they don't) and they correspond to a
non-SSA "mov" into the Phi's "variable". This gives all of the good
properties of SSA, while ensuring that a bunch of CFG transformations don't
have to be SSA-aware.
So far the only DFG phases that are SSA-aware are DCE and CFA. CFG
simplification is probably SSA-aware by default, though I haven't tried it.
Constant folding probably needs a few tweaks, but is likely ready. Ditto
for CSE, though it's not clear that we'd want to use block-local CSE when
we could be doing GVN.
Currently only the FTL can generate code from the SSA form, and there is no
way to convert from SSA to ThreadedCPS or LoadStore. There probably will
never be such a capability.
In order to handle OSR exit state in the SSA, we place MovHints at Phi
points. Other than that, you can reconstruct state-at-exit by forward
propagating MovHints. Note that MovHint is the new SetLocal in SSA.
SetLocal and GetLocal only survive into SSA if they are on captured
variables, or in the case of flushes. A "live SetLocal" will be
NodeMustGenerate and will always correspond to a flush. Computing the
state-at-exit requires running SSA liveness analysis, OSR availability
analysis, and flush liveness analysis. The FTL runs all of these prior to
generating code. While OSR exit continues to be tricky, much of the logic
is now factored into separate phases and the backend has to do less work
to reason about what happened outside of the basic block that is being
lowered.
Conversion from DFG SSA to LLVM SSA is done by ensuring that we generate
code in depth-first order, thus guaranteeing that a node will always be
lowered (and hence have a LValue) before any of the blocks dominated by
that node's block have code generated. For Upsilon/Phi, we just use
alloca's. We could do something more clever there, but it's probably not
worth it, at least not now.
Finally, while the SSA form is currently only being converted to LLVM IR,
there is nothing that prevents us from considering other backends in the
future - with the caveat that this form is designed to be first lowered to
a lower-level SSA before actual machine code generation commences. So we
ought to either use LLVM (the intended path) or we will have to write our
own SSA low-level backend.
This runs all of the code that the FTL was known to run previously. No
change in performance for now. But it does open some exciting
possibilities!
* JavaScriptCore.xcodeproj/project.pbxproj:
* bytecode/Operands.h:
(JSC::OperandValueTraits::dump):
(JSC::Operands::fill):
(Operands):
(JSC::Operands::clear):
(JSC::Operands::operator==):
* dfg/DFGAbstractState.cpp:
(JSC::DFG::AbstractState::beginBasicBlock):
(JSC::DFG::setLiveValues):
(DFG):
(JSC::DFG::AbstractState::initialize):
(JSC::DFG::AbstractState::endBasicBlock):
(JSC::DFG::AbstractState::executeEffects):
(JSC::DFG::AbstractState::mergeStateAtTail):
(JSC::DFG::AbstractState::merge):
* dfg/DFGAbstractState.h:
(AbstractState):
* dfg/DFGAdjacencyList.h:
(JSC::DFG::AdjacencyList::justOneChild):
(AdjacencyList):
* dfg/DFGBasicBlock.cpp: Added.
(DFG):
(JSC::DFG::BasicBlock::BasicBlock):
(JSC::DFG::BasicBlock::~BasicBlock):
(JSC::DFG::BasicBlock::ensureLocals):
(JSC::DFG::BasicBlock::isInPhis):
(JSC::DFG::BasicBlock::isInBlock):
(JSC::DFG::BasicBlock::removePredecessor):
(JSC::DFG::BasicBlock::replacePredecessor):
(JSC::DFG::BasicBlock::dump):
(JSC::DFG::BasicBlock::SSAData::SSAData):
(JSC::DFG::BasicBlock::SSAData::~SSAData):
* dfg/DFGBasicBlock.h:
(BasicBlock):
(JSC::DFG::BasicBlock::operator[]):
(JSC::DFG::BasicBlock::successor):
(JSC::DFG::BasicBlock::successorForCondition):
(SSAData):
* dfg/DFGBasicBlockInlines.h:
(DFG):
* dfg/DFGBlockInsertionSet.cpp: Added.
(DFG):
(JSC::DFG::BlockInsertionSet::BlockInsertionSet):
(JSC::DFG::BlockInsertionSet::~BlockInsertionSet):
(JSC::DFG::BlockInsertionSet::insert):
(JSC::DFG::BlockInsertionSet::insertBefore):
(JSC::DFG::BlockInsertionSet::execute):
* dfg/DFGBlockInsertionSet.h: Added.
(DFG):
(BlockInsertionSet):
* dfg/DFGCFAPhase.cpp:
(JSC::DFG::CFAPhase::run):
* dfg/DFGCFGSimplificationPhase.cpp:
* dfg/DFGCPSRethreadingPhase.cpp:
(JSC::DFG::CPSRethreadingPhase::canonicalizeLocalsInBlock):
* dfg/DFGCommon.cpp:
(WTF::printInternal):
* dfg/DFGCommon.h:
(JSC::DFG::doesKill):
(DFG):
(JSC::DFG::killStatusForDoesKill):
* dfg/DFGConstantFoldingPhase.cpp:
(JSC::DFG::ConstantFoldingPhase::foldConstants):
(JSC::DFG::ConstantFoldingPhase::isCapturedAtOrAfter):
* dfg/DFGCriticalEdgeBreakingPhase.cpp: Added.
(DFG):
(CriticalEdgeBreakingPhase):
(JSC::DFG::CriticalEdgeBreakingPhase::CriticalEdgeBreakingPhase):
(JSC::DFG::CriticalEdgeBreakingPhase::run):
(JSC::DFG::CriticalEdgeBreakingPhase::breakCriticalEdge):
(JSC::DFG::performCriticalEdgeBreaking):
* dfg/DFGCriticalEdgeBreakingPhase.h: Added.
(DFG):
* dfg/DFGDCEPhase.cpp:
(JSC::DFG::DCEPhase::run):
(JSC::DFG::DCEPhase::findTypeCheckRoot):
(JSC::DFG::DCEPhase::countNode):
(DCEPhase):
(JSC::DFG::DCEPhase::countEdge):
(JSC::DFG::DCEPhase::eliminateIrrelevantPhantomChildren):
* dfg/DFGDriver.cpp:
(JSC::DFG::compile):
* dfg/DFGEdge.cpp:
(JSC::DFG::Edge::dump):
* dfg/DFGEdge.h:
(JSC::DFG::Edge::Edge):
(JSC::DFG::Edge::setNode):
(JSC::DFG::Edge::useKindUnchecked):
(JSC::DFG::Edge::setUseKind):
(JSC::DFG::Edge::setProofStatus):
(JSC::DFG::Edge::willNotHaveCheck):
(JSC::DFG::Edge::willHaveCheck):
(Edge):
(JSC::DFG::Edge::killStatusUnchecked):
(JSC::DFG::Edge::killStatus):
(JSC::DFG::Edge::setKillStatus):
(JSC::DFG::Edge::doesKill):
(JSC::DFG::Edge::doesNotKill):
(JSC::DFG::Edge::shift):
(JSC::DFG::Edge::makeWord):
* dfg/DFGFixupPhase.cpp:
(JSC::DFG::FixupPhase::fixupNode):
* dfg/DFGFlushFormat.cpp: Added.
(WTF):
(WTF::printInternal):
* dfg/DFGFlushFormat.h: Added.
(DFG):
(JSC::DFG::resultFor):
(JSC::DFG::useKindFor):
(WTF):
* dfg/DFGFlushLivenessAnalysisPhase.cpp: Added.
(DFG):
(FlushLivenessAnalysisPhase):
(JSC::DFG::FlushLivenessAnalysisPhase::FlushLivenessAnalysisPhase):
(JSC::DFG::FlushLivenessAnalysisPhase::run):
(JSC::DFG::FlushLivenessAnalysisPhase::process):
(JSC::DFG::FlushLivenessAnalysisPhase::setForNode):
(JSC::DFG::FlushLivenessAnalysisPhase::flushFormat):
(JSC::DFG::performFlushLivenessAnalysis):
* dfg/DFGFlushLivenessAnalysisPhase.h: Added.
(DFG):
* dfg/DFGGraph.cpp:
(JSC::DFG::Graph::dump):
(JSC::DFG::Graph::dumpBlockHeader):
(DFG):
(JSC::DFG::Graph::addForDepthFirstSort):
(JSC::DFG::Graph::getBlocksInDepthFirstOrder):
* dfg/DFGGraph.h:
(JSC::DFG::Graph::convertToConstant):
(JSC::DFG::Graph::valueProfileFor):
(Graph):
* dfg/DFGInsertionSet.h:
(DFG):
(JSC::DFG::InsertionSet::execute):
* dfg/DFGLivenessAnalysisPhase.cpp: Added.
(DFG):
(LivenessAnalysisPhase):
(JSC::DFG::LivenessAnalysisPhase::LivenessAnalysisPhase):
(JSC::DFG::LivenessAnalysisPhase::run):
(JSC::DFG::LivenessAnalysisPhase::process):
(JSC::DFG::LivenessAnalysisPhase::addChildUse):
(JSC::DFG::performLivenessAnalysis):
* dfg/DFGLivenessAnalysisPhase.h: Added.
(DFG):
* dfg/DFGNode.cpp:
(JSC::DFG::Node::hasVariableAccessData):
(DFG):
* dfg/DFGNode.h:
(DFG):
(Node):
(JSC::DFG::Node::hasLocal):
(JSC::DFG::Node::variableAccessData):
(JSC::DFG::Node::hasPhi):
(JSC::DFG::Node::phi):
(JSC::DFG::Node::takenBlock):
(JSC::DFG::Node::notTakenBlock):
(JSC::DFG::Node::successor):
(JSC::DFG::Node::successorForCondition):
(JSC::DFG::nodeComparator):
(JSC::DFG::nodeListDump):
(JSC::DFG::nodeMapDump):
* dfg/DFGNodeFlags.cpp:
(JSC::DFG::dumpNodeFlags):
* dfg/DFGNodeType.h:
(DFG):
* dfg/DFGOSRAvailabilityAnalysisPhase.cpp: Added.
(DFG):
(OSRAvailabilityAnalysisPhase):
(JSC::DFG::OSRAvailabilityAnalysisPhase::OSRAvailabilityAnalysisPhase):
(JSC::DFG::OSRAvailabilityAnalysisPhase::run):
(JSC::DFG::performOSRAvailabilityAnalysis):
* dfg/DFGOSRAvailabilityAnalysisPhase.h: Added.
(DFG):
* dfg/DFGPlan.cpp:
(JSC::DFG::Plan::compileInThreadImpl):
* dfg/DFGPredictionInjectionPhase.cpp:
(JSC::DFG::PredictionInjectionPhase::run):
* dfg/DFGPredictionPropagationPhase.cpp:
(JSC::DFG::PredictionPropagationPhase::propagate):
* dfg/DFGSSAConversionPhase.cpp: Added.
(DFG):
(SSAConversionPhase):
(JSC::DFG::SSAConversionPhase::SSAConversionPhase):
(JSC::DFG::SSAConversionPhase::run):
(JSC::DFG::SSAConversionPhase::forwardPhiChildren):
(JSC::DFG::SSAConversionPhase::forwardPhi):
(JSC::DFG::SSAConversionPhase::forwardPhiEdge):
(JSC::DFG::SSAConversionPhase::deduplicateChildren):
(JSC::DFG::SSAConversionPhase::addFlushedLocalOp):
(JSC::DFG::SSAConversionPhase::addFlushedLocalEdge):
(JSC::DFG::performSSAConversion):
* dfg/DFGSSAConversionPhase.h: Added.
(DFG):
* dfg/DFGSpeculativeJIT32_64.cpp:
(JSC::DFG::SpeculativeJIT::compile):
* dfg/DFGSpeculativeJIT64.cpp:
(JSC::DFG::SpeculativeJIT::compile):
* dfg/DFGValidate.cpp:
(JSC::DFG::Validate::validate):
(Validate):
(JSC::DFG::Validate::validateCPS):
* dfg/DFGVariableAccessData.h:
(JSC::DFG::VariableAccessData::flushFormat):
(VariableAccessData):
* ftl/FTLCapabilities.cpp:
(JSC::FTL::canCompile):
* ftl/FTLLowerDFGToLLVM.cpp:
(JSC::FTL::LowerDFGToLLVM::LowerDFGToLLVM):
(JSC::FTL::LowerDFGToLLVM::lower):
(JSC::FTL::LowerDFGToLLVM::createPhiVariables):
(JSC::FTL::LowerDFGToLLVM::compileBlock):
(JSC::FTL::LowerDFGToLLVM::compileNode):
(JSC::FTL::LowerDFGToLLVM::compileUpsilon):
(LowerDFGToLLVM):
(JSC::FTL::LowerDFGToLLVM::compilePhi):
(JSC::FTL::LowerDFGToLLVM::compileJSConstant):
(JSC::FTL::LowerDFGToLLVM::compileWeakJSConstant):
(JSC::FTL::LowerDFGToLLVM::compileGetArgument):
(JSC::FTL::LowerDFGToLLVM::compileGetLocal):
(JSC::FTL::LowerDFGToLLVM::compileSetLocal):
(JSC::FTL::LowerDFGToLLVM::compileAdd):
(JSC::FTL::LowerDFGToLLVM::compileArithSub):
(JSC::FTL::LowerDFGToLLVM::compileArithMul):
(JSC::FTL::LowerDFGToLLVM::compileArithDiv):
(JSC::FTL::LowerDFGToLLVM::compileArithMod):
(JSC::FTL::LowerDFGToLLVM::compileArithMinOrMax):
(JSC::FTL::LowerDFGToLLVM::compileArithAbs):
(JSC::FTL::LowerDFGToLLVM::compileArithNegate):
(JSC::FTL::LowerDFGToLLVM::compileBitAnd):
(JSC::FTL::LowerDFGToLLVM::compileBitOr):
(JSC::FTL::LowerDFGToLLVM::compileBitXor):
(JSC::FTL::LowerDFGToLLVM::compileBitRShift):
(JSC::FTL::LowerDFGToLLVM::compileBitLShift):
(JSC::FTL::LowerDFGToLLVM::compileBitURShift):
(JSC::FTL::LowerDFGToLLVM::compileUInt32ToNumber):
(JSC::FTL::LowerDFGToLLVM::compileInt32ToDouble):
(JSC::FTL::LowerDFGToLLVM::compileGetButterfly):
(JSC::FTL::LowerDFGToLLVM::compileGetArrayLength):
(JSC::FTL::LowerDFGToLLVM::compileGetByVal):
(JSC::FTL::LowerDFGToLLVM::compileGetByOffset):
(JSC::FTL::LowerDFGToLLVM::compileGetGlobalVar):
(JSC::FTL::LowerDFGToLLVM::compileCompareEqConstant):
(JSC::FTL::LowerDFGToLLVM::compileCompareStrictEq):
(JSC::FTL::LowerDFGToLLVM::compileCompareStrictEqConstant):
(JSC::FTL::LowerDFGToLLVM::compileCompareLess):
(JSC::FTL::LowerDFGToLLVM::compileCompareLessEq):
(JSC::FTL::LowerDFGToLLVM::compileCompareGreater):
(JSC::FTL::LowerDFGToLLVM::compileCompareGreaterEq):
(JSC::FTL::LowerDFGToLLVM::compileLogicalNot):
(JSC::FTL::LowerDFGToLLVM::speculateBackward):
(JSC::FTL::LowerDFGToLLVM::lowInt32):
(JSC::FTL::LowerDFGToLLVM::lowCell):
(JSC::FTL::LowerDFGToLLVM::lowBoolean):
(JSC::FTL::LowerDFGToLLVM::lowDouble):
(JSC::FTL::LowerDFGToLLVM::lowJSValue):
(JSC::FTL::LowerDFGToLLVM::lowStorage):
(JSC::FTL::LowerDFGToLLVM::speculate):
(JSC::FTL::LowerDFGToLLVM::speculateBoolean):
(JSC::FTL::LowerDFGToLLVM::isLive):
(JSC::FTL::LowerDFGToLLVM::use):
(JSC::FTL::LowerDFGToLLVM::initializeOSRExitStateForBlock):
(JSC::FTL::LowerDFGToLLVM::appendOSRExit):
(JSC::FTL::LowerDFGToLLVM::emitOSRExitCall):
(JSC::FTL::LowerDFGToLLVM::addExitArgumentForNode):
(JSC::FTL::LowerDFGToLLVM::linkOSRExitsAndCompleteInitializationBlocks):
(JSC::FTL::LowerDFGToLLVM::setInt32):
(JSC::FTL::LowerDFGToLLVM::setJSValue):
(JSC::FTL::LowerDFGToLLVM::setBoolean):
(JSC::FTL::LowerDFGToLLVM::setStorage):
(JSC::FTL::LowerDFGToLLVM::setDouble):
(JSC::FTL::LowerDFGToLLVM::isValid):
* ftl/FTLLoweredNodeValue.h: Added.
(FTL):
(LoweredNodeValue):
(JSC::FTL::LoweredNodeValue::LoweredNodeValue):
(JSC::FTL::LoweredNodeValue::isSet):
(JSC::FTL::LoweredNodeValue::operator!):
(JSC::FTL::LoweredNodeValue::value):
(JSC::FTL::LoweredNodeValue::block):
* ftl/FTLValueFromBlock.h:
(JSC::FTL::ValueFromBlock::ValueFromBlock):
(ValueFromBlock):
* ftl/FTLValueSource.cpp:
(JSC::FTL::ValueSource::dump):
* ftl/FTLValueSource.h:
Source/WTF:
Reviewed by Mark Hahnenberg.
- Extend variadicity of PrintStream and dataLog.
- Give HashSet the ability to add a span of things.
- Give HashSet the ability to == another HashSet.
- Note FIXME's in HashTable concerning copying performance, that affects
the way that the DFG now uses HashSets and HashMaps.
- Factor out the bulk-insertion logic of JSC::DFG::InsertionSet into
WTF::Insertion, so that it can be used in more places.
- Create a dumper for lists and maps.
* WTF.xcodeproj/project.pbxproj:
* wtf/DataLog.h:
(WTF):
(WTF::dataLog):
* wtf/HashSet.h:
(HashSet):
(WTF):
(WTF::::add):
(WTF::=):
* wtf/HashTable.h:
(WTF::::HashTable):
(WTF::=):
* wtf/Insertion.h: Added.
(WTF):
(Insertion):
(WTF::Insertion::Insertion):
(WTF::Insertion::index):
(WTF::Insertion::element):
(WTF::Insertion::operator<):
(WTF::executeInsertions):
* wtf/ListDump.h: Added.
(WTF):
(ListDump):
(WTF::ListDump::ListDump):
(WTF::ListDump::dump):
(MapDump):
(WTF::MapDump::MapDump):
(WTF::MapDump::dump):
(WTF::listDump):
(WTF::sortedListDump):
(WTF::lessThan):
(WTF::mapDump):
(WTF::sortedMapDump):
* wtf/PrintStream.h:
(PrintStream):
(WTF::PrintStream::print):
Conflicts:
Source/JavaScriptCore/JavaScriptCore.xcodeproj/project.pbxproj
git-svn-id: http://svn.webkit.org/repository/webkit/trunk@153274 268f45cc-cd09-0410-ab3c-d52691b4dbfc
diff --git a/Source/JavaScriptCore/dfg/DFGFixupPhase.cpp b/Source/JavaScriptCore/dfg/DFGFixupPhase.cpp
index eccec9b..24e29f4 100644
--- a/Source/JavaScriptCore/dfg/DFGFixupPhase.cpp
+++ b/Source/JavaScriptCore/dfg/DFGFixupPhase.cpp
@@ -120,6 +120,7 @@
case ValueToInt32: {
if (node->child1()->shouldSpeculateInteger()) {
setUseKindAndUnboxIfProfitable<Int32Use>(node->child1());
+ node->setOpAndDefaultFlags(Identity);
break;
}
@@ -868,6 +869,8 @@
case GetArrayLength:
case Phi:
+ case Upsilon:
+ case GetArgument:
case ForwardInt32ToDouble:
case PhantomPutStructure:
case GetIndexedPropertyStorage:
