Changeset 251886 in webkit for trunk/Source/JavaScriptCore/llint/LowLevelInterpreter64.asm

Timestamp:

Oct 31, 2019, 3:32:52 PM (6 years ago)

Author:

Tadeu Zagallo

Message:

[WebAssembly] Create a Wasm interpreter
​https://bugs.webkit.org/show_bug.cgi?id=194257
<rdar://problem/44186794>

Reviewed by Saam Barati.

Source/JavaScriptCore:

Add an interpreter tier to WebAssembly which reuses the LLInt infrastructure. The interpreter
currently tiers up straight to OMG and can OSR enter at the prologue and from loops. The initial
implementation of the interpreter is very naive, but despite the lack of optimizations it still
shows a 2x improvement on the WebAssembly subtests in JetStream2 and 2x improvement on the
PSPDFKit benchmark. It reduces "compilation" times by ~3x and it's neutral on throughput.

The interpreter follows the same calling conventions as the BBQ/OMG, this means that:

• We have to allocate locals for all argument registers and write all arguments registers to the stack in the prologue.
• Calls have to allocate space for at least as many arguments as the number of argument registers. Before each call, all argument registers must be loaded from the stack, and after we return from the call, all registers must be stored back to the stack, in case they contain return values. We carefully layout the stack so that the arguments that would already have to be passed in the stack end up in the right place. The stack layout for calls is:

[ gprs ][ fprs ][ optional stack arguments ][ callee frame ]

^sp

• The return opcode has to load all registers from the stack, since they might need to contain results of the function.
• The calling convention requires that the callee should store itself in the callee slot of the call frame, which is impossible in the interpreter, since the code we execute is the same for all callees. In order to work around that, we generate an entry thunk to the wasm interpreter for each function. All this thunk does is store the callee in the call frame and tail call the interpreter.

• CMakeLists.txt:
• DerivedSources-input.xcfilelist:
• DerivedSources-output.xcfilelist:
• DerivedSources.make:
• JavaScriptCore.xcodeproj/project.pbxproj:
• Sources.txt:
• bytecode/BytecodeDumper.cpp:

(JSC::BytecodeDumper<Block>::constantName const):
(JSC::BytecodeDumper<Block>::dumpValue):
(JSC::BytecodeDumper<Block>::dumpBytecode):
(JSC::CodeBlockBytecodeDumper<Block>::vm const):
(JSC::CodeBlockBytecodeDumper<Block>::identifier const):
(JSC::CodeBlockBytecodeDumper<Block>::dumpIdentifiers):
(JSC::CodeBlockBytecodeDumper<Block>::dumpConstants):
(JSC::CodeBlockBytecodeDumper<Block>::dumpExceptionHandlers):
(JSC::CodeBlockBytecodeDumper<Block>::dumpSwitchJumpTables):
(JSC::CodeBlockBytecodeDumper<Block>::dumpStringSwitchJumpTables):
(JSC::CodeBlockBytecodeDumper<Block>::dumpBlock):

• bytecode/BytecodeDumper.h:

(JSC::BytecodeDumper::dumpValue):
(JSC::BytecodeDumper::BytecodeDumper):

• bytecode/BytecodeGeneratorification.cpp:

(JSC::performGeneratorification):

• bytecode/BytecodeList.rb:
• bytecode/CodeBlock.cpp:

(JSC::CodeBlock::dumpBytecode):

• bytecode/Fits.h:
• bytecode/Instruction.h:

(JSC::BaseInstruction::BaseInstruction):
(JSC::BaseInstruction::Impl::opcodeID const):
(JSC::BaseInstruction::opcodeID const):
(JSC::BaseInstruction::name const):
(JSC::BaseInstruction::isWide16 const):
(JSC::BaseInstruction::isWide32 const):
(JSC::BaseInstruction::hasMetadata const):
(JSC::BaseInstruction::sizeShiftAmount const):
(JSC::BaseInstruction::size const):
(JSC::BaseInstruction::is const):
(JSC::BaseInstruction::as const):
(JSC::BaseInstruction::cast):
(JSC::BaseInstruction::cast const):
(JSC::BaseInstruction::wide16 const):
(JSC::BaseInstruction::wide32 const):

• bytecode/InstructionStream.h:

(JSC::InstructionStream::iterator::operator+=):
(JSC::InstructionStream::iterator::operator++):
(JSC::InstructionStreamWriter::iterator::operator+=):
(JSC::InstructionStreamWriter::iterator::operator++):

• bytecode/Opcode.cpp:
• bytecode/Opcode.h:
• bytecode/PreciseJumpTargetsInlines.h:
• bytecode/UnlinkedCodeBlock.h:
• bytecode/VirtualRegister.cpp:

(JSC::VirtualRegister::VirtualRegister):

• bytecode/VirtualRegister.h:
• bytecompiler/BytecodeGenerator.cpp:

(JSC::GenericLabel<JSGeneratorTraits>::setLocation):
(JSC::BytecodeGenerator::BytecodeGenerator):

• bytecompiler/BytecodeGenerator.h:
• bytecompiler/BytecodeGeneratorBase.h: Added.
• bytecompiler/BytecodeGeneratorBaseInlines.h: Added.

(JSC::shrinkToFit):
(JSC::BytecodeGeneratorBase<Traits>::BytecodeGeneratorBase):
(JSC::BytecodeGeneratorBase<Traits>::newLabel):
(JSC::BytecodeGeneratorBase<Traits>::newEmittedLabel):
(JSC::BytecodeGeneratorBase<Traits>::reclaimFreeRegisters):
(JSC::BytecodeGeneratorBase<Traits>::emitLabel):
(JSC::BytecodeGeneratorBase<Traits>::recordOpcode):
(JSC::BytecodeGeneratorBase<Traits>::alignWideOpcode16):
(JSC::BytecodeGeneratorBase<Traits>::alignWideOpcode32):
(JSC::BytecodeGeneratorBase<Traits>::write):
(JSC::BytecodeGeneratorBase<Traits>::newRegister):
(JSC::BytecodeGeneratorBase<Traits>::newTemporary):
(JSC::BytecodeGeneratorBase<Traits>::addVar):
(JSC::BytecodeGeneratorBase<Traits>::allocateCalleeSaveSpace):

• bytecompiler/Label.h:

(JSC::GenericBoundLabel::GenericBoundLabel):
(JSC::GenericBoundLabel::target):
(JSC::GenericBoundLabel::saveTarget):
(JSC::GenericBoundLabel::commitTarget):

• dfg/DFGByteCodeParser.cpp:
• dfg/DFGCapabilities.cpp:

(JSC::DFG::capabilityLevel):

• dfg/DFGOperations.cpp:
• generator/Argument.rb:
• generator/DSL.rb:
• generator/GeneratedFile.rb:
• generator/Opcode.rb:
• generator/Options.rb:
• generator/Section.rb:
• generator/Wasm.rb: Added.
• interpreter/Register.h:
• interpreter/RegisterInlines.h:

(JSC::Register::operator=):

• jit/JITArithmetic.cpp:
• jit/JITOpcodes.cpp:
• llint/LLIntData.cpp:

(JSC::LLInt::initialize):

• llint/LLIntData.h:

(JSC::LLInt::wasmExceptionInstructions):

• llint/LLIntOfflineAsmConfig.h:
• llint/LLIntOffsetsExtractor.cpp:
• llint/LLIntSlowPaths.cpp:
• llint/LLIntThunks.cpp:

(JSC::LLInt::generateThunkWithJumpTo):
(JSC::LLInt::wasmFunctionEntryThunk):

• llint/LLIntThunks.h:
• llint/LowLevelInterpreter.asm:
• llint/LowLevelInterpreter32_64.asm:
• llint/LowLevelInterpreter64.asm:
• llint/WebAssembly.asm: Added.
• offlineasm/arm64.rb:
• offlineasm/instructions.rb:
• offlineasm/parser.rb:
• offlineasm/registers.rb:
• offlineasm/transform.rb:
• offlineasm/x86.rb:
• parser/Nodes.h:
• runtime/Error.cpp:

(JSC::FindFirstCallerFrameWithCodeblockFunctor::operator() const):

• runtime/ErrorInstance.cpp:

(JSC::ErrorInstance::finishCreation):

• runtime/Options.cpp:

(JSC::overrideDefaults):

• runtime/OptionsList.h:
• runtime/SamplingProfiler.cpp:

(JSC::FrameWalker::recordJITFrame):
(JSC::FrameWalker::resetAtMachineFrame):

• wasm/WasmAirIRGenerator.cpp:

(JSC::Wasm::AirIRGenerator::isControlTypeIf):
(JSC::Wasm::AirIRGenerator::emitLoopTierUpCheck):

• wasm/WasmB3IRGenerator.cpp:

(JSC::Wasm::B3IRGenerator::isControlTypeIf):

• wasm/WasmBBQPlan.cpp:

(JSC::Wasm::BBQPlan::prepareImpl):
(JSC::Wasm::BBQPlan::work):
(JSC::Wasm::BBQPlan::compileFunction):
(JSC::Wasm::BBQPlan::didCompleteCompilation):
(JSC::Wasm::BBQPlan::initializeCallees):

• wasm/WasmBBQPlan.h:
• wasm/WasmBBQPlanInlines.h: Removed.
• wasm/WasmCallee.cpp:

(JSC::Wasm::Callee::Callee):
(JSC::Wasm::Callee::dump const):
(JSC::Wasm::JITCallee::JITCallee):
(JSC::Wasm::LLIntCallee::setEntrypoint):
(JSC::Wasm::LLIntCallee::entrypoint const):
(JSC::Wasm::LLIntCallee::calleeSaveRegisters):
(JSC::Wasm:: const):

• wasm/WasmCallee.h:

(JSC::Wasm::Callee::setOSREntryCallee):
(JSC::Wasm::JITCallee::wasmToWasmCallsites):
(JSC::Wasm::JITCallee:: const):

• wasm/WasmCallingConvention.h:
• wasm/WasmCodeBlock.cpp:

(JSC::Wasm::CodeBlock::CodeBlock):

• wasm/WasmCodeBlock.h:

(JSC::Wasm::CodeBlock::wasmEntrypointCalleeFromFunctionIndexSpace):
(JSC::Wasm::CodeBlock::wasmBBQCalleeFromFunctionIndexSpace):
(JSC::Wasm::CodeBlock::wasmToWasmExitStub):

• wasm/WasmCompilationMode.cpp:

(JSC::Wasm::makeString):

• wasm/WasmCompilationMode.h:
• wasm/WasmEmbedder.h:
• wasm/WasmEntryPlan.cpp: Added.

(JSC::Wasm::EntryPlan::EntryPlan):
(JSC::Wasm::EntryPlan::stateString):
(JSC::Wasm::EntryPlan::moveToState):
(JSC::Wasm::EntryPlan::didReceiveFunctionData):
(JSC::Wasm::EntryPlan::parseAndValidateModule):
(JSC::Wasm::EntryPlan::prepare):
(JSC::Wasm::EntryPlan::ThreadCountHolder::ThreadCountHolder):
(JSC::Wasm::EntryPlan::ThreadCountHolder::~ThreadCountHolder):
(JSC::Wasm::EntryPlan::complete):
(JSC::Wasm::EntryPlan::compileFunctions):
(JSC::Wasm::EntryPlan::work):

• wasm/WasmEntryPlan.h: Copied from Source/JavaScriptCore/wasm/WasmBBQPlan.h.

(JSC::Wasm::EntryPlan::parseAndValidateModule):
(JSC::Wasm::EntryPlan::exports const):
(JSC::Wasm::EntryPlan::internalFunctionCount const):
(JSC::Wasm::EntryPlan::takeModuleInformation):
(JSC::Wasm::EntryPlan::takeWasmToWasmExitStubs):
(JSC::Wasm::EntryPlan::takeWasmToWasmCallsites):
(JSC::Wasm::EntryPlan::hasBeenPrepared const):
(JSC::Wasm::EntryPlan::tryReserveCapacity):

• wasm/WasmFunctionCodeBlock.cpp: Added.

(JSC::Wasm::FunctionCodeBlock::setInstructions):
(JSC::Wasm::FunctionCodeBlock::dumpBytecode):
(JSC::Wasm::FunctionCodeBlock::addOutOfLineJumpTarget):
(JSC::Wasm::FunctionCodeBlock::outOfLineJumpOffset):
(JSC::Wasm::FunctionCodeBlock::outOfLineJumpTarget):
(JSC::Wasm::FunctionCodeBlock::addSignature):
(JSC::Wasm::FunctionCodeBlock::signature const):
(JSC::Wasm::FunctionCodeBlock::addJumpTable):
(JSC::Wasm::FunctionCodeBlock::jumpTable const const):
(JSC::Wasm::FunctionCodeBlock::numberOfJumpTables const):

• wasm/WasmFunctionCodeBlock.h: Added.

(JSC::Wasm::FunctionCodeBlock::FunctionCodeBlock):
(JSC::Wasm::FunctionCodeBlock::getConstant const):
(JSC::Wasm::FunctionCodeBlock::functionIndex const):
(JSC::Wasm::FunctionCodeBlock::addJumpTarget):
(JSC::Wasm::FunctionCodeBlock::numberOfJumpTargets):
(JSC::Wasm::FunctionCodeBlock::lastJumpTarget):
(JSC::Wasm::FunctionCodeBlock::outOfLineJumpOffset):
(JSC::Wasm::FunctionCodeBlock::bytecodeOffset):
(JSC::Wasm::FunctionCodeBlock::tierUpCounter):

• wasm/WasmFunctionParser.h:

(JSC::Wasm::FunctionParser<Context>::parseExpression):
(JSC::Wasm::FunctionParser<Context>::parseUnreachableExpression):

• wasm/WasmInstance.h:
• wasm/WasmLLIntGenerator.cpp: Added.

(JSC::Wasm::LLIntGenerator::ControlType::ControlType):
(JSC::Wasm::LLIntGenerator::ControlType::loop):
(JSC::Wasm::LLIntGenerator::ControlType::topLevel):
(JSC::Wasm::LLIntGenerator::ControlType::block):
(JSC::Wasm::LLIntGenerator::ControlType::if_):
(JSC::Wasm::LLIntGenerator::ControlType::targetLabelForBranch const):
(JSC::Wasm::LLIntGenerator::fail const):
(JSC::Wasm::LLIntGenerator::unifyValuesWithBlock):
(JSC::Wasm::LLIntGenerator::emptyExpression):
(JSC::Wasm::LLIntGenerator::createStack):
(JSC::Wasm::LLIntGenerator::isControlTypeIf):
(JSC::Wasm::LLIntGenerator::addEndToUnreachable):
(JSC::Wasm::LLIntGenerator::setParser):
(JSC::Wasm::LLIntGenerator::dump):
(JSC::Wasm::LLIntGenerator::virtualRegisterForLocal):
(JSC::Wasm::LLIntGenerator::tmpsForSignature):
(JSC::Wasm::LLIntGenerator::jsNullConstant):
(JSC::Wasm::LLIntGenerator::isConstant):
(JSC::Wasm::parseAndCompileBytecode):
(JSC::Wasm::LLIntGenerator::LLIntGenerator):
(JSC::Wasm::LLIntGenerator::finalize):
(JSC::Wasm::LLIntGenerator::callInformationFor):
(JSC::Wasm::LLIntGenerator::addArguments):
(JSC::Wasm::LLIntGenerator::addLocal):
(JSC::Wasm::LLIntGenerator::addConstant):
(JSC::Wasm::LLIntGenerator::getLocal):
(JSC::Wasm::LLIntGenerator::setLocal):
(JSC::Wasm::LLIntGenerator::getGlobal):
(JSC::Wasm::LLIntGenerator::setGlobal):
(JSC::Wasm::LLIntGenerator::addLoop):
(JSC::Wasm::LLIntGenerator::addTopLevel):
(JSC::Wasm::LLIntGenerator::addBlock):
(JSC::Wasm::LLIntGenerator::addIf):
(JSC::Wasm::LLIntGenerator::addElse):
(JSC::Wasm::LLIntGenerator::addElseToUnreachable):
(JSC::Wasm::LLIntGenerator::addReturn):
(JSC::Wasm::LLIntGenerator::addBranch):
(JSC::Wasm::LLIntGenerator::addSwitch):
(JSC::Wasm::LLIntGenerator::endBlock):
(JSC::Wasm::LLIntGenerator::addCall):
(JSC::Wasm::LLIntGenerator::addCallIndirect):
(JSC::Wasm::LLIntGenerator::addRefIsNull):
(JSC::Wasm::LLIntGenerator::addRefFunc):
(JSC::Wasm::LLIntGenerator::addTableGet):
(JSC::Wasm::LLIntGenerator::addTableSet):
(JSC::Wasm::LLIntGenerator::addTableSize):
(JSC::Wasm::LLIntGenerator::addTableGrow):
(JSC::Wasm::LLIntGenerator::addTableFill):
(JSC::Wasm::LLIntGenerator::addUnreachable):
(JSC::Wasm::LLIntGenerator::addCurrentMemory):
(JSC::Wasm::LLIntGenerator::addGrowMemory):
(JSC::Wasm::LLIntGenerator::addSelect):
(JSC::Wasm::LLIntGenerator::load):
(JSC::Wasm::LLIntGenerator::store):
(JSC::GenericLabel<Wasm::GeneratorTraits>::setLocation):

• wasm/WasmLLIntGenerator.h: Copied from Source/JavaScriptCore/wasm/WasmCompilationMode.h.
• wasm/WasmLLIntPlan.cpp: Added.

(JSC::Wasm::LLIntPlan::prepareImpl):
(JSC::Wasm::LLIntPlan::compileFunction):
(JSC::Wasm::LLIntPlan::didCompleteCompilation):
(JSC::Wasm::LLIntPlan::initializeCallees):

• wasm/WasmLLIntPlan.h: Copied from Source/JavaScriptCore/wasm/WasmOMGForOSREntryPlan.h.
• wasm/WasmLLIntTierUpCounter.cpp: Copied from Source/JavaScriptCore/wasm/WasmCompilationMode.cpp.

(JSC::Wasm::LLIntTierUpCounter::addOSREntryDataForLoop):
(JSC::Wasm::LLIntTierUpCounter::osrEntryDataForLoop const const):

• wasm/WasmLLIntTierUpCounter.h: Copied from Source/JavaScriptCore/wasm/WasmOMGForOSREntryPlan.h.

(JSC::Wasm::LLIntTierUpCounter::LLIntTierUpCounter):
(JSC::Wasm::LLIntTierUpCounter::optimizeAfterWarmUp):
(JSC::Wasm::LLIntTierUpCounter::checkIfOptimizationThresholdReached):
(JSC::Wasm::LLIntTierUpCounter::optimizeSoon):

• wasm/WasmMemoryInformation.cpp:

(JSC::Wasm::PinnedRegisterInfo::get):

• wasm/WasmModule.cpp:

(JSC::Wasm::makeValidationResult):
(JSC::Wasm::makeValidationCallback):
(JSC::Wasm::Module::validateSync):
(JSC::Wasm::Module::validateAsync):

• wasm/WasmOMGForOSREntryPlan.cpp:

(JSC::Wasm::OMGForOSREntryPlan::OMGForOSREntryPlan):
(JSC::Wasm::OMGForOSREntryPlan::work):

• wasm/WasmOMGForOSREntryPlan.h:
• wasm/WasmOMGPlan.cpp:

(JSC::Wasm::OMGPlan::work):

• wasm/WasmSlowPaths.cpp: Added.

(JSC::LLInt::jitCompileAndSetHeuristics):
(JSC::LLInt::WASM_SLOW_PATH_DECL):
(JSC::LLInt::doWasmCall):
(JSC::LLInt::doWasmCallIndirect):
(JSC::LLInt::slow_path_wasm_throw_exception):
(JSC::LLInt::slow_path_wasm_popcount):
(JSC::LLInt::slow_path_wasm_popcountll):

• wasm/WasmSlowPaths.h: Added.
• wasm/WasmTable.cpp:

(JSC::Wasm::FuncRefTable::function const):
(JSC::Wasm::FuncRefTable::instance const):

• wasm/WasmTable.h:
• wasm/WasmTierUpCount.h:
• wasm/WasmValidate.cpp:

(JSC::Wasm::Validate::isControlTypeIf):

• wasm/js/JSToWasm.cpp:

(JSC::Wasm::createJSToWasmWrapper):

• wasm/js/JSToWasm.h:
• wasm/js/WebAssemblyFunction.cpp:

(JSC::WebAssemblyFunction::calleeSaves const):

Tools:

Add a mode that runs WebAssembly tests without the LLInt (i.e. only Air)
and update the no-air mode to also disable the LLInt tier.

• Scripts/run-jsc-stress-tests:

File:

• trunk/Source/JavaScriptCore/llint/LowLevelInterpreter64.asm (modified) (7 diffs)

trunk/Source/JavaScriptCore/llint/LowLevelInterpreter64.asm

@@ -388 +388 @@
 end
 
-# Call a slow path for call call opcodes.
+# Call a slow path for call opcodes.
 macro callCallSlowPath(slowPath, action)
     storei PC, ArgumentCount + TagOffset[cfr]
@@ -1020 +1020 @@
     .op1NotIntOp2Int:
         profile(ArithProfileNumberInt)
-        ci2d t1, ft1
+        ci2ds t1, ft1
     .op1NotIntReady:
         get(m_dst, t2)
@@ -1036 +1036 @@
         btqz t1, numberTag, .slow
         profile(ArithProfileIntNumber)
-        ci2d t0, ft0
+        ci2ds t0, ft0
         addq numberTag, t1
         fq2d t1, ft1
@@ -1679 +1679 @@
                 loadConstantOrVariable(size, operand, scratch)
                 bqb scratch, numberTag, .notInt
-                ci2d scratch, ft0
+                ci2ds scratch, ft0
                 jmp .ready
             .notInt:
@@ -1828 +1828 @@
     storeb 1, OpJneqPtr::Metadata::m_hasJumped[t5]
     get(m_targetLabel, t0)
-    jumpImpl(t0)
+    jumpImpl(dispatchIndirect, t0)
 end)
 
@@ -1846 +1846 @@
         btqz t0, numberTag, .slow
         bqb t1, numberTag, .op1NotIntOp2NotInt
-        ci2d t1, ft1
+        ci2ds t1, ft1
         jmp .op1NotIntReady
     .op1NotIntOp2NotInt:
@@ -1859 +1859 @@
 
     .op2NotInt:
-        ci2d t0, ft0
+        ci2ds t0, ft0
         btqz t1, numberTag, .slow
         addq numberTag, t1