Changeset 102743 in webkit for trunk/Source/JavaScriptCore/bytecode

Timestamp:

Dec 13, 2011, 11:16:36 PM (13 years ago)

Author:

[email protected]

Message:

DFG should infer when local variables are doubles
​https://bugs.webkit.org/show_bug.cgi?id=74480

Reviewed by Oliver Hunt.

Introduced the notion that a local variable (though not an argument, yet!) can
be stored as a double, and will be guaranteed to always contain a double. This
requires more magic in the OSR (conversion in both entry and exit). The inference
is quite unorthodox: all uses of a variable vote on whether they think it should
be a double or a JSValue, based on how they use it. If they use it in an integer
or boxed value context, they vote JSValue. If they use it in a double context,
they vote double. This voting is interleaved in the propagator's fixpoint, so
that variables voted double then have a double prediction propagated from them.
This interleaving is needed because a variable that actually always contains an
integer that always gets used in arithmetic that involves doubles may end up
being voted double, which then means that all uses of the variable will see a
double rather than an integer.

This is worth 18% to SunSpider/3d-cube, 7% to Kraken/audio-beat-detection, 7%
to Kraken/audio-fft, 6% to Kraken/imaging-darkroom, 20% to
Kraken/imaging-gaussian-blur, and just over 1% to Kraken/json-parse-financial.
It results in a 1% speed-up on SunSpider and a 4% speed-up in Kraken. Similar
results on JSVALUE32_64, though with a bigger win on Kraken (5%) and no overall
win on SunSpider.

• bytecode/ValueRecovery.h:

(JSC::ValueRecovery::alreadyInRegisterFileAsUnboxedDouble):
(JSC::ValueRecovery::dump):

• dfg/DFGAbstractState.cpp:

(JSC::DFG::AbstractState::execute):

• dfg/DFGAssemblyHelpers.h:

(JSC::DFG::AssemblyHelpers::boxDouble):

• dfg/DFGGraph.cpp:

(JSC::DFG::Graph::dump):

• dfg/DFGJITCompiler.h:

(JSC::DFG::JITCompiler::noticeOSREntry):

• dfg/DFGOSREntry.cpp:

(JSC::DFG::prepareOSREntry):

• dfg/DFGOSREntry.h:
• dfg/DFGOSRExitCompiler64.cpp:

(JSC::DFG::OSRExitCompiler::compileExit):

• dfg/DFGPropagator.cpp:

(JSC::DFG::Propagator::vote):
(JSC::DFG::Propagator::doRoundOfDoubleVoting):
(JSC::DFG::Propagator::propagatePredictions):
(JSC::DFG::Propagator::fixupNode):

• dfg/DFGSpeculativeJIT.cpp:

(JSC::DFG::ValueSource::dump):
(JSC::DFG::SpeculativeJIT::compile):
(JSC::DFG::SpeculativeJIT::computeValueRecoveryFor):

• dfg/DFGSpeculativeJIT.h:
• dfg/DFGSpeculativeJIT32_64.cpp:

(JSC::DFG::SpeculativeJIT::compile):

• dfg/DFGSpeculativeJIT64.cpp:

(JSC::DFG::SpeculativeJIT::compile):

• dfg/DFGVariableAccessData.h:

(JSC::DFG::VariableAccessData::VariableAccessData):
(JSC::DFG::VariableAccessData::clearVotes):
(JSC::DFG::VariableAccessData::vote):
(JSC::DFG::VariableAccessData::doubleVoteRatio):
(JSC::DFG::VariableAccessData::shouldUseDoubleFormatAccordingToVote):
(JSC::DFG::VariableAccessData::shouldUseDoubleFormat):
(JSC::DFG::VariableAccessData::tallyVotesForShouldUseDoubleFormat):

• runtime/Arguments.cpp:

(JSC::Arguments::tearOff):

• runtime/Heuristics.cpp:

(JSC::Heuristics::initializeHeuristics):

• runtime/Heuristics.h:

File:

• trunk/Source/JavaScriptCore/bytecode/ValueRecovery.h (modified) (3 diffs)

trunk/Source/JavaScriptCore/bytecode/ValueRecovery.h

@@ -48 +48 @@
     AlreadyInRegisterFileAsUnboxedCell,
     AlreadyInRegisterFileAsUnboxedBoolean,
+    AlreadyInRegisterFileAsUnboxedDouble,
     // It's in a register.
     InGPR,
@@ -103 +104 @@
     }
     
+    static ValueRecovery alreadyInRegisterFileAsUnboxedDouble()
+    {
+        ValueRecovery result;
+        result.m_technique = AlreadyInRegisterFileAsUnboxedDouble;
+        return result;
+    }
+    
     static ValueRecovery inGPR(MacroAssembler::RegisterID gpr, DataFormat dataFormat)
     {
@@ -248 +256 @@
             fprintf(out, "(bool)");
             break;
+        case AlreadyInRegisterFileAsUnboxedDouble:
+            fprintf(out, "(double)");
+            break;
         case InGPR:
             fprintf(out, "%%r%d", gpr());