[WebKit内核] JavaScript引擎深度解析--基础篇(一)字节码生成及语法树的构建详情分析

      看到HorkeyChen写的文章《[WebKit] JavaScriptCore解析--基础篇(三)从脚本代码到JIT编译的代码实现》,写的很好,深受启发。想补充一些Horkey没有写到的细节比如字节码是如何生成的等等,为此成文。


      JSC对JavaScript的处理,其实与Webkit对CSS的处理许多地方是类似的,它这么几个部分:

(1)词法分析->出来词语(Token);

(2)语法分析->出来抽象语法树(AST:Abstract Syntax Tree);

(3)遍历抽象语法树->生成字节码(Bytecode);

(4)用解释器(LLInt:Low Level Interpreter)执行字节码;

(5)如果性能不够好就用Baseline JIT编译字节码生成机器码、然后执行此机器码;

(6)如果性能还不够好,就用DFG JIT重新编译字节码生成更好的机器码、然后执行此机器码;

(7)最后,如果还不好,就祭出重器--虚拟器(LLVM:Low Level Virtual Machine)来编译DFG的中间表示代码、生成更高优化的机器码并执行。接下来,我将会用一下系列文章描述此过程。

      其中,步骤1、2是类似的,3、4、5步的思想,CSS JIT也是采用类似方法,请参考[1]。想写写JSC的文章,用菜鸟和愚公移山的方式,敲开JSC的冰山一角。

本篇主要描述词法和语法解析的细节。

 一、 JavaScriptCore的词法分析器工作流程分析

    W3C是这么解释词法和语法工作流程的:

[WebKit内核] JavaScript引擎深度解析--基础篇(一)字节码生成及语法树的构建详情分析_第1张图片

 词法器Tokenizer的工作过程如下,就是不断从字符串中寻找一个个的词(Token),比如找到连续的“true”字符串,就创建一个TokenTrue。词法器工作过程如下:

JavaScriptCore/interpreter/interpreter.cpp:
template 
template  TokenType LiteralParser::Lexer::lex(LiteralParserToken& token)
{
    while (m_ptr < m_end && isJSONWhiteSpace(*m_ptr))
        ++m_ptr;

    if (m_ptr >= m_end) {
        token.type = TokEnd;
        token.start = token.end = m_ptr;
        return TokEnd;
    }
    token.type = TokError;
    token.start = m_ptr;
    switch (*m_ptr) {
        case '[':
            token.type = TokLBracket;
            token.end = ++m_ptr;
            return TokLBracket;
        case ']':
            token.type = TokRBracket;
            token.end = ++m_ptr;
            return TokRBracket;
        case '(':
            token.type = TokLParen;
            token.end = ++m_ptr;
            return TokLParen;
        case ')':
            token.type = TokRParen;
            token.end = ++m_ptr;
            return TokRParen;
        case ',':
            token.type = TokComma;
            token.end = ++m_ptr;
            return TokComma;
        case ':':
            token.type = TokColon;
            token.end = ++m_ptr;
            return TokColon;
        case '"':
            return lexString(token);
        case 't':
            if (m_end - m_ptr >= 4 && m_ptr[1] == 'r' && m_ptr[2] == 'u' && m_ptr[3] == 'e') {
                m_ptr += 4;
                token.type = TokTrue;
                token.end = m_ptr;
                return TokTrue;
            }
            break;
        case '-':
        case '0':
        case '9':
            return lexNumber(token);
    }
    if (m_ptr < m_end) {
        if (*m_ptr == '.') {
            token.type = TokDot;
            token.end = ++m_ptr;
            return TokDot;
        }
        if (*m_ptr == '=') {
            token.type = TokAssign;
            token.end = ++m_ptr;
            return TokAssign;
        }
        if (*m_ptr == ';') {
            token.type = TokSemi;
            token.end = ++m_ptr;
            return TokAssign;
        }
        if (isASCIIAlpha(*m_ptr) || *m_ptr == '_' || *m_ptr == '$')
            return lexIdentifier(token);
        if (*m_ptr == '\'') {
            return lexString(token);
        }
    }
    m_lexErrorMessage = String::format("Unrecognized token '%c'", *m_ptr).impl();
    return TokError;
}

      经过此过程,一个完整的JSC世界的Token就生成了。然后,再进行语法分析,生成抽象语法树.下图就是JavaScriptCore世界语法节点的静态类关系:

[WebKit内核] JavaScript引擎深度解析--基础篇(一)字节码生成及语法树的构建详情分析_第2张图片

      下面我们看看,语法解析具体过程:

JavaScriptCore/parser/parser.cpp:

PassRefPtr Parser::parse(JSGlobalObject* lexicalGlobalObject, Debugger* debugger, ExecState* debuggerExecState, JSObject** exception)
{
    ASSERT(lexicalGlobalObject);
    ASSERT(exception && !*exception);
    int errLine;
    UString errMsg;

    if (ParsedNode::scopeIsFunction)
        m_lexer->setIsReparsing();

    m_sourceElements = 0;

    errLine = -1;
    errMsg = UString();

    UString parseError = parseInner();
    。。。
}

创建抽象语法树Builder,并用来解析、生成语法节点:

UString Parser::parseInner(){
    UString parseError = UString();
   unsigned oldFunctionCacheSize = m_functionCache ? m_functionCache->byteSize() : 0;
    //抽象语法树Builder:
    ASTBuilder context(const_cast(m_globalData), const_cast(m_source));
    if (m_lexer->isReparsing())
        m_statementDepth--;
    ScopeRef scope = currentScope();
    //开始解析生成语法树的一个节点:
    SourceElements* sourceElements = parseSourceElements(context);
    if (!sourceElements || !consume(EOFTOK))
}

         举例说来,根据Token的类型,JSC认为输入的Token是一个常量声明,就会使用如下的模板函数生成语法节点(Node),然后放入ASTBuilder里面,我们先看ASTBuilder的结构:

class ASTBuilder {
    ......
    Scope m_scope;
    Vector m_binaryOperandStack;
    Vector m_assignmentInfoStack;
    Vector, 10> m_binaryOperatorStack;
    Vector, 10> m_unaryTokenStack;
    int m_evalCount;
};

再看主要语法解析过程(Parser/parser.cpp):

template 
template  TreeSourceElements Parser::parseSourceElements(TreeBuilder& context)
{
    const unsigned lengthOfUseStrictLiteral = 12; // "use strict".length
    TreeSourceElements sourceElements = context.createSourceElements();
    bool seenNonDirective = false;
    const Identifier* directive = 0;
    unsigned directiveLiteralLength = 0;
    unsigned startOffset = m_token.m_info.startOffset;
    unsigned oldLastLineNumber = m_lexer->lastLineNumber();
    unsigned oldLineNumber = m_lexer->lineNumber();
    bool hasSetStrict = false;
    //解析语法节点--语句
    while (TreeStatement statement = parseStatement(context, directive, &directiveLiteralLength)) {
        if (mode == CheckForStrictMode && !seenNonDirective) {
            if (directive) {
                // "use strict" must be the exact literal without escape sequences or line continuation.
                if (!hasSetStrict && directiveLiteralLength == lengthOfUseStrictLiteral && m_globalData->propertyNames->useStrictIdentifier == *directive) {
                    setStrictMode();
                    hasSetStrict = true;
                    failIfFalse(isValidStrictMode());
                    m_lexer->setOffset(startOffset);
                    next();
                    m_lexer->setLastLineNumber(oldLastLineNumber);
                    m_lexer->setLineNumber(oldLineNumber);
                    failIfTrue(m_error);
                    continue;
                }
            } else
                seenNonDirective = true;
        }
        context.appendStatement(sourceElements, statement); //添加语法节点到ASTBuilder
    }
    
    if (m_error)
        fail();
    return sourceElements;
}

解析语句就是各种switch case,效率不高啊!

template 
template  TreeStatement Parser::parseStatement(TreeBuilder& context, const Identifier*& directive, unsigned* directiveLiteralLength)
{
    DepthManager statementDepth(&m_statementDepth);
    m_statementDepth++;
    directive = 0;
    int nonTrivialExpressionCount = 0;
    failIfStackOverflow();
    switch (m_token.m_type) {
    case OPENBRACE:
        return parseBlockStatement(context);
    case VAR:
        return parseVarDeclaration(context);
    case CONSTTOKEN:
        return parseConstDeclaration(context);
    case FUNCTION:
        failIfFalseIfStrictWithMessage(m_statementDepth == 1, "Functions cannot be declared in a nested block in strict mode");
        return parseFunctionDeclaration(context);
    case SEMICOLON:
        next();
        return context.createEmptyStatement(m_lexer->lastLineNumber());
    case IF:
        return parseIfStatement(context);
    case DO:
        return parseDoWhileStatement(context);
    case WHILE:
        return parseWhileStatement(context);
    case FOR:
        return parseForStatement(context);
    case CONTINUE:
        return parseContinueStatement(context);
    case BREAK:
        return parseBreakStatement(context);
    case RETURN:
        return parseReturnStatement(context);
    case WITH:
        return parseWithStatement(context);
    case SWITCH:
        return parseSwitchStatement(context);
    case THROW:
        return parseThrowStatement(context);
    case TRY:
        return parseTryStatement(context);
    case DEBUGGER:
        return parseDebuggerStatement(context);
    case EOFTOK:
    case CASE:
    case CLOSEBRACE:
    case DEFAULT:
        // These tokens imply the end of a set of source elements
        return 0;
    case IDENT:
        return parseExpressionOrLabelStatement(context);
    case STRING:
        directive = m_token.m_data.ident;
        if (directiveLiteralLength)
            *directiveLiteralLength = m_token.m_info.endOffset - m_token.m_info.startOffset;
        nonTrivialExpressionCount = m_nonTrivialExpressionCount;
    default:
        TreeStatement exprStatement = parseExpressionStatement(context);
        if (directive && nonTrivialExpressionCount != m_nonTrivialExpressionCount)
            directive = 0;
        return exprStatement;
    }
}

举其中一个例子:

JavaScriptCore/parser/parser.cpp:

template 
template  TreeConstDeclList Parser::parseConstDeclarationList(TreeBuilder& context)
{
    failIfTrue(strictMode());
    TreeConstDeclList constDecls = 0;
    TreeConstDeclList tail = 0;
    do {
        next();
        matchOrFail(IDENT);
        //取出词(Token):
        const Identifier* name = m_token.m_data.ident;
        next();
        //是一个=吗?
        bool hasInitializer = match(EQUAL);
        //
        declareVariable(name);
        context.addVar(name, DeclarationStacks::IsConstant | (hasInitializer ? DeclarationStacks::HasInitializer : 0));
        TreeExpression initializer = 0;
        if (hasInitializer) {
            next(TreeBuilder::DontBuildStrings); // consume '='
            initializer = parseAssignmentExpression(context);
        }
	新建一个“常量申明节点”放入ASTBuilder里面:
        tail = context.appendConstDecl(m_lexer->lastLineNumber(), tail, name, initializer);
        if (!constDecls)
            constDecls = tail;
    } while (match(COMMA));
    return constDecls;
}

ASTBuilder.h:

ConstDeclNode* appendConstDecl(int lineNumber, ConstDeclNode* tail, const Identifier* name, ExpressionNode* initializer)
{
        ConstDeclNode* result = new (m_globalData) ConstDeclNode(lineNumber, *name, initializer);
        if (tail)
            tail->m_next = result;
        return result;
}

调用堆栈 如下:

#0  JSC::ASTBuilder::BinaryExprContext::BinaryExprContext (this=0x7fffffffbb6f)    at JavaScriptCore/parser/ASTBuilder.h:85
#1  JSC::Parser >::parseBinaryExpression (this=0x7fffffffc330, context=...)JavaScriptCore/parser/Parser.cpp:1143
#2  JSC::Parser >::parseConditionalExpression (this=0x7fffffffc330, context=...)   at JavaScriptCore/parser/Parser.cpp:1109
#3  JSC::Parser >::parseAssignmentExpression (this=0x7fffffffc330, context=...)
    at /opt/src/opt/src/mp50/framework/webkit/WebKit_123412/Source/JavaScriptCore/parser/Parser.cpp:1051
#4  JSC::Parser >::parseVarDeclarationList (this=, context=...,   declarations=@0x7fffffffbd3c: 1, lastIdent=@0x7fffffffbd30: 0xdb3060, lastInitializer=@0x7fffffffbd28: 0x0, identStart=@0x7fffffffbd24: 5,   initStart=@0x7fffffffbd24: 5, initEnd=@: 5)    at parser/Parser.cpp:263
#5  JSC::Parser >::parseVarDeclaration (this=0x7fffffffc330, context=...)   at JavaScriptCore/parser/Parser.cpp:181
#6  JSC::Parser >::parseStatement (this=0x7fffffffc330, context=..., directive=: 0x0,directiveLiteralLength=)   Parser.cpp:682
#7  JSC::Parser >::parseSourceElements<(JSC::SourceElementsMode)0, JSC::ASTBuilder> (this, context=...) at parser/Parser.cpp:145
#8  JSC::Parser >::parseInner (this=0x7fffffffc330)    at Parser.cpp:93
#9  JSC::Parser >::parse (this=, lexicalGlobalObject=,   debugger=0x0, debuggerExecState=, exception=) Parser.h:990
#10 JSC::parse (globalData=, lexicalGlobalObject=source,parameters,  strictness=JSParseNormal,parserMode=JSParseProgramCode, debugger, execState=, exception=) Parser.h:1048
#11 JSC::ProgramExecutable::compileInternal (this=, exec=, scopeChainNode=, jitType=JSC::JITCode::BaselineJIT) at JavaScriptCore/runtime/Executable.cpp:338
#12 JSC::ProgramExecutable::compile (this=0x7ffff7fbb580, exec=0x7ffff7f9fb90, scopeChainNode=0x7ffff7f7ffc0)JavaScriptCore/runtime/Executable.h:446
#13 JSC::Interpreter::execute (this=, program=, callFrame=, scopeChain=,  thisObj=0x7ffff7f9f980) at JavaScriptCore/interpreter/Interpreter.cpp:1224
#14 JSC::evaluate (exec=, scopeChain=, source=..., thisValue=..., returnedException=) JavaScriptCore/runtime/Completion.cpp:75
#15 runWithScripts (globalObject=0x7ffff7f9f980, scripts=, dump=false)   at JavaScriptCore/jsc.cpp:545
#16 jscmain (argc=2, argv=0x7fffffffdc88) at JavaScriptCore/jsc.cpp:733
#17 main (argc=2, argv=0x7fffffffdc88) atavaScriptCore/jsc.cpp:510

     

      接下来,就会调用BytecodeGenerator::generate生成字节码,具体分下节分析。我们先看看下面来自JavaScript的一个个语法树节点生成字节码的过程:

JavaScriptCore/bytecompiler/BytecodeGenerator.cpp:
RegisterID* BooleanNode::emitBytecode(BytecodeGenerator& generator, RegisterID* dst)

{
    if (dst == generator.ignoredResult())
        return 0;
    return generator.emitLoad(dst, m_value);
}

     以下是我准备写的文章题目:

一、 JavaScriptCore的词法分析器工作流程分析;

二、 JavaScriptCore的语法分析器工作流程分析;

三、 JavaScriptCore的字节码生成流程分析;

四、 LLInt解释器工作流程分析;

五、 Baseline JIT编译器的工作流程分析;

六、 DFG JIT编译器的工作流程分析;

七、LLVM虚拟机的工作流程分析;

八、JavaScriptCore的未来展望;

     文笔粗糙,不善表达,希望能越写越好。

原创,转载请注明:http://blog.csdn.net/lichwei1983/article/details/44658533

第一时间获得博客更新提醒,以及更多技术信息分享,欢迎关注个人微信公众平台:程序员互动联盟(coder_online),扫一扫下方二维码或搜索微信号coder_online即可关注,我们可以在线交流。

引用:

1 https://www.webkit.org/blog/3271/webkit-css-selector-jit-compiler/

2 http://blog.csdn.net/horkychen/article/details/8928578


你可能感兴趣的:(Webkit)