Ignore:
Timestamp:
Apr 18, 2020, 7:20:59 PM (5 years ago)
Author:
[email protected]
Message:

Support an inlined representation in JSValue of small BigInts ("BigInt32")
https://bugs.webkit.org/show_bug.cgi?id=206182

Reviewed by Yusuke Suzuki.

JSTests:

I improved several of the tests to give more informative error messages in the process of fixing them.

More interestingly I had to modify "missing-exception-check-in-string-compare" because it relied on "s1 == s1" resolving ropes, and we now just return true.

  • stress/big-int-division.js:

(testDiv):

  • stress/big-int-left-shift-wrapped-value.js:

(assert.sameValue):

  • stress/big-int-logical-not.js:

(assert):

  • stress/big-int-mod-jit.js:
  • stress/big-int-right-shift-general.js:

(testRightShift):

  • stress/big-int-type-of-proven-type.js:

(assert):

  • stress/compare-strict-eq-on-various-types.js:

(testAllTypesCall):

  • stress/ftl-string-strict-equality.js:
  • stress/missing-exception-check-in-string-compare.js:

Source/JavaScriptCore:

This patch attempts to optimize the performance of BigInts, when they are small (32 bit or less).
It works by inlining them into JSValue on 64-bit platforms, avoiding the allocation of a JSBigInt.
The bit pattern we use is 0000:XXXX:XXXX:0012
This representation works because of the following things:

  • It cannot be confused with a Double or Integer thanks to the top bits
  • It cannot be confused with a pointer to a Cell, thanks to bit 1 which is set to true
  • It cannot be confused with a pointer to wasm thanks to bit 0 which is set to false
  • It cannot be confused with true/false because bit 2 is set to false
  • It cannot be confused for null/undefined because bit 4 is set to true

This entire change is gated by USE(BIGINT32), to make it easier to disable if it turns out to have bugs.
It should also make it much easier to verify if a given bug comes from it or from something else.

Note that in this patch we create BigInt32s when parsing small BigInt constants, and most operations (e.g. Add or BitOr) produce a BigInt32 if both of their operands are BigInt32,
but we don't produce a BigInt32 from for example the substraction/division of two large heap-allocated JSBigInts, even if the result fits in 32-bits.
As a result, small BigInts can now either be heap-allocated or inlined in the JSValue.

This patch includes a significant refactor of various slow paths, which are now grouped together in Operations.h
Because this increased the size of Operations.h significantly, I split the parts of Operations.h which are only used by the GC into Scribble.h, to avoid bloating compile times.

In the DFG and FTL we now have 3 UseKinds for BigInts: HeapBigIntUse, BigInt32Use and AnyBigIntUse.
The latter is useful when we know that we are receiving BigInts, but speculation indicates a mix of heap-allocated and small (inlined) big-ints.

Unfortunately, a naive implementation of this patch significantly regresses the performance of StrictEq (and its variants), as it is no longer true that a cell and a non-cell cannot be equal.
Before this patch, the code was jumping to a slow path if either:

  • at least one operand is a double
  • or both operands are cells

Now, it also needs to jump to the slow path if at least one is a cell.
To recover this performance cost, I significantly rewrote this code, from

if (left is Cell && right is Cell) {

if (left == right)

return true;

goto slowPath;

}
if (! left is Int32) {

if (left is Number)

goto slowPath

}
if (! right is Int32) {

if (right is Number)

goto slowPath

}
return left == right

To the following:

if (left is Double
right is Double)

goto slowPath

if (left == right)

return true;

if (left is Cell
right is Cell)

goto slowPath

return false;

I believe this to be faster than just replacing (left is Cell && right is Cell) by an
, because I found a bit-trick to check (left is Double right is Double) which should help reduce the pressure on the branch predictor.

Early JetStream2 tests appear to confirm that this patch is roughly neutral while it was a 0.5% regression before I used this trick, but the numbers are still too noisy, I plan to do more measurements before landing this patch.

I don't yet have performance numbers for this patch on a BigInt benchmark, I will get such numbers before trying to land it, but I'd like some review in the meantime.

(JSC::X86Assembler::X86InstructionFormatter::SingleInstructionBufferWriter::memoryModRM):

  • bytecode/ArithProfile.cpp:

(JSC::ArithProfile<BitfieldType>::emitObserveResult):
(JSC::ArithProfile<BitfieldType>::shouldEmitSetBigInt32 const):
(JSC::ArithProfile<BitfieldType>::shouldEmitSetHeapBigInt const):
(JSC::ArithProfile<BitfieldType>::emitSetHeapBigInt const):
(JSC::ArithProfile<BitfieldType>::emitSetBigInt32 const):
(WTF::printInternal):

  • bytecode/ArithProfile.h:

(JSC::ObservedResults::didObserveNonInt32):
(JSC::ObservedResults::didObserveBigInt):
(JSC::ObservedResults::didObserveHeapBigInt):
(JSC::ObservedResults::didObserveBigInt32):
(JSC::ArithProfile::didObserveHeapBigInt const):
(JSC::ArithProfile::didObserveBigInt32 const):
(JSC::ArithProfile::setObservedHeapBigInt):
(JSC::ArithProfile::setObservedBigInt32):
(JSC::ArithProfile::observeResult):

  • bytecode/BytecodeList.rb:
  • bytecode/BytecodeLivenessAnalysisInlines.h:
  • bytecode/BytecodeUseDef.cpp:

(JSC::computeUsesForBytecodeIndexImpl):
(JSC::computeDefsForBytecodeIndexImpl):

  • bytecode/CodeBlock.cpp:
  • bytecode/DataFormat.h:
  • bytecode/MethodOfGettingAValueProfile.cpp:

(JSC::MethodOfGettingAValueProfile::emitReportValue const):

  • bytecode/MethodOfGettingAValueProfile.h:
  • bytecode/SpeculatedType.cpp:

(JSC::dumpSpeculation):
(JSC::speculationFromClassInfo):
(JSC::speculationFromStructure):
(JSC::speculationFromValue):
(JSC::speculationFromJSType):
(JSC::leastUpperBoundOfStrictlyEquivalentSpeculations):

  • bytecode/SpeculatedType.h:

(JSC::isBigInt32Speculation):
(JSC::isHeapBigIntSpeculation):
(JSC::isBigIntSpeculation):

  • bytecompiler/BytecodeGenerator.cpp:

(JSC::BytecodeGenerator::emitEqualityOpImpl):
(JSC::BytecodeGenerator::addBigIntConstant):

  • bytecompiler/BytecodeGenerator.h:
  • dfg/DFGAbstractInterpreterInlines.h:

(JSC::DFG::isToThisAnIdentity):
(JSC::DFG::AbstractInterpreter<AbstractStateType>::executeEffects):

  • dfg/DFGByteCodeParser.cpp:

(JSC::DFG::ByteCodeParser::parseBlock):

  • dfg/DFGCapabilities.cpp:

(JSC::DFG::capabilityLevel):

  • dfg/DFGClobberize.h:

(JSC::DFG::clobberize):

  • dfg/DFGConstantFoldingPhase.cpp:

(JSC::DFG::ConstantFoldingPhase::foldConstants):

  • dfg/DFGDoesGC.cpp:

(JSC::DFG::doesGC):

  • dfg/DFGFixupPhase.cpp:

(JSC::DFG::FixupPhase::fixupNode):
(JSC::DFG::FixupPhase::fixupToThis):
(JSC::DFG::FixupPhase::fixupToNumeric):
(JSC::DFG::FixupPhase::observeUseKindOnNode):
(JSC::DFG::FixupPhase::fixupCompareStrictEqAndSameValue):

  • dfg/DFGMayExit.cpp:
  • dfg/DFGNode.h:

(JSC::DFG::Node::shouldSpeculateBigInt32):
(JSC::DFG::Node::shouldSpeculateHeapBigInt):

  • dfg/DFGNodeType.h:
  • dfg/DFGOSRExit.cpp:

(JSC::DFG::OSRExit::compileExit):

  • dfg/DFGOSRExit.h:
  • dfg/DFGOperations.cpp:
  • dfg/DFGOperations.h:
  • dfg/DFGPredictionPropagationPhase.cpp:
  • dfg/DFGSafeToExecute.h:

(JSC::DFG::SafeToExecuteEdge::operator()):
(JSC::DFG::safeToExecute):

  • dfg/DFGSpeculativeJIT.cpp:

(JSC::DFG::SpeculativeJIT::compilePeepHoleBranch):
(JSC::DFG::SpeculativeJIT::compileValueBitNot):
(JSC::DFG::SpeculativeJIT::emitUntypedOrAnyBigIntBitOp):
(JSC::DFG::SpeculativeJIT::compileValueBitwiseOp):
(JSC::DFG::SpeculativeJIT::emitUntypedOrBigIntRightShiftBitOp):
(JSC::DFG::SpeculativeJIT::compileValueLShiftOp):
(JSC::DFG::SpeculativeJIT::compileValueBitRShift):
(JSC::DFG::SpeculativeJIT::compileShiftOp):
(JSC::DFG::SpeculativeJIT::compileValueAdd):
(JSC::DFG::SpeculativeJIT::compileValueSub):
(JSC::DFG::SpeculativeJIT::compileIncOrDec):
(JSC::DFG::SpeculativeJIT::compileValueNegate):
(JSC::DFG::SpeculativeJIT::compileValueMul):
(JSC::DFG::SpeculativeJIT::compileValueDiv):
(JSC::DFG::SpeculativeJIT::compileValueMod):
(JSC::DFG::SpeculativeJIT::compileValuePow):
(JSC::DFG::SpeculativeJIT::compare):
(JSC::DFG::SpeculativeJIT::compileStrictEq):
(JSC::DFG::SpeculativeJIT::speculateHeapBigInt):
(JSC::DFG::SpeculativeJIT::speculate):
(JSC::DFG::SpeculativeJIT::compileToNumeric):
(JSC::DFG::SpeculativeJIT::compileHeapBigIntEquality):

  • dfg/DFGSpeculativeJIT.h:

(JSC::DFG::SpeculateBigInt32Operand::SpeculateBigInt32Operand):
(JSC::DFG::SpeculateBigInt32Operand::~SpeculateBigInt32Operand):
(JSC::DFG::SpeculateBigInt32Operand::edge const):
(JSC::DFG::SpeculateBigInt32Operand::node const):
(JSC::DFG::SpeculateBigInt32Operand::gpr):
(JSC::DFG::SpeculateBigInt32Operand::use):

  • dfg/DFGSpeculativeJIT32_64.cpp:

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

  • dfg/DFGSpeculativeJIT64.cpp:

(JSC::DFG::SpeculativeJIT::fillJSValue):
(JSC::DFG::SpeculativeJIT::nonSpeculativePeepholeStrictEq):
(JSC::DFG::SpeculativeJIT::nonSpeculativeNonPeepholeStrictEq):
(JSC::DFG::SpeculativeJIT::fillSpeculateInt32Internal):
(JSC::DFG::SpeculativeJIT::fillSpeculateCell):
(JSC::DFG::SpeculativeJIT::fillSpeculateBoolean):
(JSC::DFG::SpeculativeJIT::speculateBigInt32):
(JSC::DFG::SpeculativeJIT::speculateAnyBigInt):
(JSC::DFG::SpeculativeJIT::fillSpeculateBigInt32):
(JSC::DFG::SpeculativeJIT::compileBigInt32Compare):
(JSC::DFG::SpeculativeJIT::compilePeepHoleBigInt32Branch):
(JSC::DFG::SpeculativeJIT::compile):

  • dfg/DFGStrengthReductionPhase.cpp:

(JSC::DFG::StrengthReductionPhase::handleNode):

  • dfg/DFGUseKind.cpp:

(WTF::printInternal):

  • dfg/DFGUseKind.h:

(JSC::DFG::typeFilterFor):
(JSC::DFG::isCell):

  • ftl/FTLCapabilities.cpp:

(JSC::FTL::canCompile):

  • ftl/FTLCommonValues.cpp:

(JSC::FTL::CommonValues::initializeConstants):

  • ftl/FTLCommonValues.h:
  • ftl/FTLLowerDFGToB3.cpp:

(JSC::FTL::DFG::LowerDFGToB3::compileNode):
(JSC::FTL::DFG::LowerDFGToB3::compileValueAdd):
(JSC::FTL::DFG::LowerDFGToB3::compileValueSub):
(JSC::FTL::DFG::LowerDFGToB3::compileValueMul):
(JSC::FTL::DFG::LowerDFGToB3::compileBinaryMathIC):
(JSC::FTL::DFG::LowerDFGToB3::compileValueDiv):
(JSC::FTL::DFG::LowerDFGToB3::compileValueMod):
(JSC::FTL::DFG::LowerDFGToB3::compileValuePow):
(JSC::FTL::DFG::LowerDFGToB3::compileValueBitNot):
(JSC::FTL::DFG::LowerDFGToB3::compileValueBitAnd):
(JSC::FTL::DFG::LowerDFGToB3::compileValueBitOr):
(JSC::FTL::DFG::LowerDFGToB3::compileValueBitXor):
(JSC::FTL::DFG::LowerDFGToB3::compileValueBitRShift):
(JSC::FTL::DFG::LowerDFGToB3::compileArithBitRShift):
(JSC::FTL::DFG::LowerDFGToB3::compileArithBitLShift):
(JSC::FTL::DFG::LowerDFGToB3::compileValueBitLShift):
(JSC::FTL::DFG::LowerDFGToB3::compileBitURShift):
(JSC::FTL::DFG::LowerDFGToB3::compileToNumeric):
(JSC::FTL::DFG::LowerDFGToB3::compileCompareEq):
(JSC::FTL::DFG::LowerDFGToB3::compileCompareStrictEq):
(JSC::FTL::DFG::LowerDFGToB3::compileIsBigInt):
(JSC::FTL::DFG::LowerDFGToB3::emitBinarySnippet):
(JSC::FTL::DFG::LowerDFGToB3::emitBinaryBitOpSnippet):
(JSC::FTL::DFG::LowerDFGToB3::boolify):
(JSC::FTL::DFG::LowerDFGToB3::buildTypeOf):
(JSC::FTL::DFG::LowerDFGToB3::lowHeapBigInt):
(JSC::FTL::DFG::LowerDFGToB3::lowBigInt32):
(JSC::FTL::DFG::LowerDFGToB3::isBigInt32):
(JSC::FTL::DFG::LowerDFGToB3::isNotBigInt32):
(JSC::FTL::DFG::LowerDFGToB3::unboxBigInt32):
(JSC::FTL::DFG::LowerDFGToB3::boxBigInt32):
(JSC::FTL::DFG::LowerDFGToB3::isNotAnyBigInt):
(JSC::FTL::DFG::LowerDFGToB3::speculate):
(JSC::FTL::DFG::LowerDFGToB3::isNotHeapBigIntUnknownWhetherCell):
(JSC::FTL::DFG::LowerDFGToB3::isNotHeapBigInt):
(JSC::FTL::DFG::LowerDFGToB3::isHeapBigInt):
(JSC::FTL::DFG::LowerDFGToB3::speculateHeapBigInt):
(JSC::FTL::DFG::LowerDFGToB3::speculateHeapBigIntUnknownWhetherCell):
(JSC::FTL::DFG::LowerDFGToB3::speculateBigInt32):
(JSC::FTL::DFG::LowerDFGToB3::speculateAnyBigInt):

  • ftl/FTLOSRExitCompiler.cpp:

(JSC::FTL::compileStub):

  • heap/HeapSnapshotBuilder.cpp:

(JSC::HeapSnapshotBuilder::json):

  • heap/MarkedBlockInlines.h:
  • heap/PreciseAllocation.cpp:
  • inspector/agents/InspectorHeapAgent.cpp:

(Inspector::InspectorHeapAgent::getPreview):

  • interpreter/Interpreter.cpp:

(JSC::sizeOfVarargs):

  • jit/AssemblyHelpers.cpp:

(JSC::AssemblyHelpers::emitConvertValueToBoolean):
(JSC::AssemblyHelpers::branchIfValue):

  • jit/AssemblyHelpers.h:

(JSC::AssemblyHelpers::branchIfBigInt32):
(JSC::AssemblyHelpers::branchIfBigInt32KnownNotNumber):
(JSC::AssemblyHelpers::branchIfNotBigInt32KnownNotNumber):
(JSC::AssemblyHelpers::branchIfHeapBigInt):
(JSC::AssemblyHelpers::branchIfNotHeapBigInt):
(JSC::AssemblyHelpers::unboxBigInt32):
(JSC::AssemblyHelpers::boxBigInt32):
(JSC::AssemblyHelpers::emitTypeOf):

  • jit/JIT.cpp:

(JSC::JIT::privateCompileMainPass):

  • jit/JIT.h:
  • jit/JITArithmetic.cpp:

(JSC::JIT::emit_op_negate):
(JSC::JIT::emitSlow_op_negate):

  • jit/JITOpcodes.cpp:

(JSC::JIT::emit_op_is_big_int):
(JSC::JIT::compileOpStrictEq):
(JSC::JIT::compileOpStrictEqJump):
(JSC::JIT::emit_op_to_numeric):

  • jit/JITOpcodes32_64.cpp:

(JSC::JIT::emit_op_is_big_int):
(JSC::JIT::emit_op_to_numeric):

  • jit/JITOperations.cpp:
  • jit/JITOperations.h:
  • llint/LLIntOfflineAsmConfig.h:
  • llint/LowLevelInterpreter.asm:
  • llint/LowLevelInterpreter64.asm:
  • parser/ParserArena.cpp:

(JSC::IdentifierArena::makeBigIntDecimalIdentifier):

  • runtime/ArrayPrototype.cpp:
  • runtime/BigIntConstructor.cpp:

(JSC::toBigInt):
(JSC::callBigIntConstructor):

  • runtime/BigIntObject.cpp:

(JSC::BigIntObject::create):
(JSC::BigIntObject::finishCreation):

  • runtime/BigIntObject.h:
  • runtime/BigIntPrototype.cpp:

(JSC::toThisBigIntValue):
(JSC::bigIntProtoFuncToStringImpl):

  • runtime/CommonSlowPaths.cpp:

(JSC::SLOW_PATH_DECL):
(JSC::updateArithProfileForUnaryArithOp):
(JSC::updateArithProfileForBinaryArithOp):

  • runtime/JSBigInt.cpp:

(JSC::JSBigInt::createStructure):
(JSC::JSBigInt::parseInt):
(JSC::JSBigInt::stringToBigInt):
(JSC::JSBigInt::inc):
(JSC::JSBigInt::dec):
(JSC::JSBigInt::bitwiseAnd):
(JSC::JSBigInt::toStringGeneric):
(JSC::JSBigInt::equalsToNumber):
(JSC::JSBigInt::equalsToInt32):

  • runtime/JSBigInt.h:

(JSC::asHeapBigInt):

  • runtime/JSCJSValue.cpp:

(JSC::JSValue::toNumberSlowCase const):
(JSC::JSValue::toObjectSlowCase const):
(JSC::JSValue::toThisSlowCase const):
(JSC::JSValue::synthesizePrototype const):
(JSC::JSValue::dumpInContextAssumingStructure const):
(JSC::JSValue::dumpForBacktrace const):
(JSC::JSValue::toStringSlowCase const):

  • runtime/JSCJSValue.h:
  • runtime/JSCJSValueInlines.h:

(JSC::JSValue::JSValue):
(JSC::JSValue::asHeapBigInt const):
(JSC::JSValue::isBigInt const):
(JSC::JSValue::isHeapBigInt const):
(JSC::JSValue::isBigInt32 const):
(JSC::JSValue::bigInt32AsInt32 const):
(JSC::JSValue::isPrimitive const):
(JSC::JSValue::getPrimitiveNumber):
(JSC::JSValue::toNumeric const):
(JSC::JSValue::toBigIntOrInt32 const):
(JSC::JSValue::equalSlowCaseInline):
(JSC::JSValue::strictEqualForCells):
(JSC::JSValue::strictEqual):
(JSC::JSValue::pureStrictEqual):
(JSC::JSValue::pureToBoolean const):

  • runtime/JSCell.cpp:

(JSC::JSCell::put):
(JSC::JSCell::putByIndex):
(JSC::JSCell::toPrimitive const):
(JSC::JSCell::getPrimitiveNumber const):
(JSC::JSCell::toNumber const):
(JSC::JSCell::toObjectSlow const):

  • runtime/JSCell.h:
  • runtime/JSCellInlines.h:

(JSC::JSCell::isHeapBigInt const):
(JSC::JSCell::toBoolean const):
(JSC::JSCell::pureToBoolean const):

  • runtime/JSString.h:

(JSC::JSValue::toBoolean const):

  • runtime/JSType.cpp:

(WTF::printInternal):

  • runtime/JSType.h:
  • runtime/JSTypeInfo.h:
  • runtime/ObjectInitializationScope.cpp:
  • runtime/Operations.cpp:

(JSC::jsAddSlowCase):
(JSC::jsIsObjectTypeOrNull):

  • runtime/Operations.h:

(JSC::compareBigIntToOtherPrimitive):
(JSC::bigIntCompare):
(JSC::jsLess):
(JSC::jsLessEq):
(JSC::arithmeticBinaryOp):
(JSC::jsSub):
(JSC::jsMul):
(JSC::jsDiv):
(JSC::jsRemainder):
(JSC::jsPow):
(JSC::jsInc):
(JSC::jsDec):
(JSC::jsBitwiseNot):
(JSC::shift):
(JSC::jsLShift):
(JSC::jsRShift):
(JSC::bitwiseBinaryOp):
(JSC::jsBitwiseAnd):
(JSC::jsBitwiseOr):
(JSC::jsBitwiseXor):

  • runtime/Scribble.h: Copied from Source/JavaScriptCore/runtime/BigIntObject.h.

(JSC::scribbleFreeCells):
(JSC::isScribbledValue):
(JSC::scribble):

  • runtime/StructureInlines.h:

(JSC::prototypeForLookupPrimitiveImpl):

Source/WTF:

Add a USE(BIGINT32) flag.

  • wtf/PlatformUse.h:
File:
1 edited

Legend:

Unmodified
Added
Removed
  • trunk/Source/JavaScriptCore/ftl/FTLCapabilities.cpp

    r260323 r260331  
    261261    case IsBoolean:
    262262    case IsNumber:
     263    case IsBigInt:
    263264    case NumberIsInteger:
    264265    case IsObject:
     
    484485                case StringOrStringObjectUse:
    485486                case SymbolUse:
    486                 case BigIntUse:
     487                case AnyBigIntUse:
     488                case BigInt32Use:
     489                case HeapBigIntUse:
    487490                case DateObjectUse:
    488491                case MapObjectUse:
Note: See TracChangeset for help on using the changeset viewer.