Home >Backend Development >PHP Tutorial >In-depth analysis of php execution principle (2): Compilation of functions, in-depth analysis of php_PHP tutorial
本文只探讨纯粹的函数,并不包含方法。对于方法,会放到类、对象中一起研究。
想讲清楚在zend vm中,函数如何被正确的编译成op指令、如何发生参数传递、如何模拟调用栈、如何切换作用域等等,的确是一个很大范畴的话题。但为了弄明白php的原理,必须要攻克它。
对函数的研究,大致可以分成两块。第一块是函数体的编译,主要涉及到如何将函数转化成zend_op指令。第二块是研究函数的调用,涉及到函数调用语句的编译,以及函数如何被执行等topic。这里先来看看函数如何被编译,我们下一篇再讲函数的调用。
对函数进行编译,最终目的是为了生成一份对应的op指令集,除了op指令集,编译函数还会产生其他一些相关的数据,比如说函数名称、参数列表信息、compiled variables,甚至函数所在文件,起始行数等等。这些信息作为编译的产出,都需要保存起来。保存这些产出的数据结构,正是上一节中所描述的zend_op_array。下文会以op_array作为简称。
下面列出了一个简单的例子:
<?<span>php </span><span>function</span> foo(<span>$arg1</span><span>) { </span><span>print</span>(<span>$arg1</span><span>); } </span><span>$bar</span> = 'hello php'<span>; foo(</span><span>$bar</span>);
这段代码包含了一个最简单的函数示例。
在这样一份php脚本中,最终其实会产生两个op_array。一个是由函数foo编译而来,另一个则是由除去foo之外代码编译生成的。同理可以推出,假如一份php脚本其中包含有2个函数和若干语句,则最终会产生3个op_array。也就是说,每个函数最终都会被编译成一个对应的op_array。
刚才提到,op_array中有一些字段是和函数息息相关的。比如function_name代表着函数的名称,比如num_args代表了函数的参数个数,比如required_num_args代表了必须的参数个数,比如arg_info代表着函数的参数信息...etc。
下面会继续结合这段代码,来研究foo函数详细的编译过程。
从zend_language_parser.y文件中可以看出,函数的语法分析大致涉及如下几个推导式:
top_statement:<br /> statement { zend_verify_namespace(TSRMLS_C); }<br /> | function_declaration_statement { zend_verify_namespace(TSRMLS_C); zend_do_early_binding(TSRMLS_C); }<br /> | class_declaration_statement { zend_verify_namespace(TSRMLS_C); zend_do_early_binding(TSRMLS_C); }<br /> ...<br /><br />function_declaration_statement: unticked_function_declaration_statement { DO_TICKS(); } ; unticked_function_declaration_statement: function is_reference T_STRING { zend_do_begin_function_declaration(&$1, &$3, 0, $2.op_type, NULL TSRMLS_CC); } '(' parameter_list ')' '{' inner_statement_list '}' { zend_do_end_function_declaration(&$1 TSRMLS_CC); } ;<br /><br />is_reference:<br /> /* empty */ { $$.op_type = ZEND_RETURN_VAL; }<br /> | '&' { $$.op_type = ZEND_RETURN_REF; }<br />;<br /><br />parameter_list:<br /> non_empty_parameter_list<br /> | /* empty */<br />; non_empty_parameter_list: optional_class_type T_VARIABLE { znode tmp; fetch_simple_variable(&tmp, &$2, 0 TSRMLS_CC); $$.op_type = IS_CONST; Z_LVAL($$.u.constant)=1; Z_TYPE($$.u.constant)=IS_LONG; INIT_PZVAL(&$$.u.constant); zend_do_receive_arg(ZEND_RECV, &tmp, &$$, NULL, &$1, &$2, 0 TSRMLS_CC); } | optional_class_type '&' T_VARIABLE { znode tmp; fetch_simple_variable(&tmp, &$3, 0 TSRMLS_CC); $$.op_type = IS_CONST; Z_LVAL($$.u.constant)=1; Z_TYPE($$.u.constant)=IS_LONG; INIT_PZVAL(&$$.u.constant); zend_do_receive_arg(ZEND_RECV, &tmp, &$$, NULL, &$1, &$3, 1 TSRMLS_CC); } | optional_class_type '&' T_VARIABLE '=' static_scalar { znode tmp; fetch_simple_variable(&tmp, &$3, 0 TSRMLS_CC); $$.op_type = IS_CONST; Z_LVAL($$.u.constant)=1; Z_TYPE($$.u.constant)=IS_LONG; INIT_PZVAL(&$$.u.constant); zend_do_receive_arg(ZEND_RECV_INIT, &tmp, &$$, &$5, &$1, &$3, 1 TSRMLS_CC); } | optional_class_type T_VARIABLE '=' static_scalar { znode tmp; fetch_simple_variable(&tmp, &$2, 0 TSRMLS_CC); $$.op_type = IS_CONST; Z_LVAL($$.u.constant)=1; Z_TYPE($$.u.constant)=IS_LONG; INIT_PZVAL(&$$.u.constant); zend_do_receive_arg(ZEND_RECV_INIT, &tmp, &$$, &$4, &$1, &$2, 0 TSRMLS_CC); } | non_empty_parameter_list ',' optional_class_type T_VARIABLE { znode tmp; fetch_simple_variable(&tmp, &$4, 0 TSRMLS_CC); $$=$1; Z_LVAL($$.u.constant)++; zend_do_receive_arg(ZEND_RECV, &tmp, &$$, NULL, &$3, &$4, 0 TSRMLS_CC); } | non_empty_parameter_list ',' optional_class_type '&' T_VARIABLE { znode tmp; fetch_simple_variable(&tmp, &$5, 0 TSRMLS_CC); $$=$1; Z_LVAL($$.u.constant)++; zend_do_receive_arg(ZEND_RECV, &tmp, &$$, NULL, &$3, &$5, 1 TSRMLS_CC); } | non_empty_parameter_list ',' optional_class_type '&' T_VARIABLE '=' static_scalar { znode tmp; fetch_simple_variable(&tmp, &$5, 0 TSRMLS_CC); $$=$1; Z_LVAL($$.u.constant)++; zend_do_receive_arg(ZEND_RECV_INIT, &tmp, &$$, &$7, &$3, &$5, 1 TSRMLS_CC); } | non_empty_parameter_list ',' optional_class_type T_VARIABLE '=' static_scalar { znode tmp; fetch_simple_variable(&tmp, &$4, 0 TSRMLS_CC); $$=$1; Z_LVAL($$.u.constant)++; zend_do_receive_arg(ZEND_RECV_INIT, &tmp, &$$, &$6, &$3, &$4, 0 TSRMLS_CC); } ;
这里并没有截取完整,主要是缺少函数体内语句的语法分析,但已经足够我们弄清楚编译过程中的一些细节。
函数体内的语句,其对应的语法为inner_statement_list。inner_statement_list和函数体之外一般的语句并无二致,可以简单当成普通的语句来编译。
最重要的是看下unticked_function_declaration_statement,它定义了函数语法的骨架,同时还可以看出,函数编译中会执行zend_do_begin_function_declaration以及zend_do_end_function_declaration。这两步分别对应着下文提到的开始编译和结束编译。我们先来看zend_do_begin_function_declaration。
当解析器遇到一段函数声明时,会尝试开始编译函数,这是通过执行zend_do_begin_function_declaration来完成的。
有两点:
1,函数是否返回引用,通过is_reference判断。可以看到在对is_reference进行语法分析时,可能会将op_type赋予ZEND_RETURN_VAL或ZEND_RETURN_REF。根据我们文章开始给出的php代码示例,函数foo并不返回引用,因此这里$2.op_type为ZEND_RETURN_VAL。话说由 function & func_name(){ ... } 这种形式来决定是否返回引用,已经很古老了,还是在CI框架中见过,现在很少有类似需求。
2,zend_do_begin_function_declaration接受的第一个参数,是对function字面进行词法分析生成的znode。这个znode被使用得非常巧妙,因为在编译函数时,zend vm必须将CG(active_op_array)切换成函数自己的op_array,以便于存储函数的编译结果,当函数编译完成之后,zend vm又需要将将CG(active_op_array)恢复成函数体外层的op_array。利用该znode保存函数体外的op_array,可以很方便的在函数编译结束时进行CG(active_op_array)恢复,具体后面会讲到。
研究下zend_do_begin_function_declaration的实现,比较长,我们分段来看:
<span>//</span><span> 声明函数会变编译成的op_array</span> <span>zend_op_array op_array; </span><span>//</span><span> 函数名、长度、起始行数</span> <span>char</span> *name = function_name-><span>u.constant.value.str.val; </span><span>int</span> name_len = function_name-><span>u.constant.value.str.len; </span><span>int</span> function_begin_line = function_token-><span>u.opline_num; zend_uint fn_flags; </span><span>char</span> *<span>lcname; zend_bool orig_interactive; ALLOCA_FLAG(use_heap) </span><span>if</span><span> (is_method) { ... } </span><span>else</span><span> { fn_flags </span>= <span>0</span><span>; } </span><span>//</span><span> 对函数来说,fn_flags没用,对方法来说,fn_flags指定了方法的修饰符</span> <span>if</span> ((fn_flags & ZEND_ACC_STATIC) && (fn_flags & ZEND_ACC_ABSTRACT) && !(CG(active_class_entry)->ce_flags &<span> ZEND_ACC_INTERFACE)) { zend_error(E_STRICT, </span><span>"</span><span>Static function %s%s%s() should not be abstract</span><span>"</span>, is_method ? CG(active_class_entry)->name : <span>""</span>, is_method ? <span>"</span><span>::</span><span>"</span> : <span>""</span>, Z_STRVAL(function_name-><span>u.constant)); }</span>
这段代码一开始就印证了我们先前的说法,每个函数都有一份自己的op_array。所以会在开头先声明一个op_array变量。
<span>//</span><span> 第一个znode参数的妙处,它记录了当前的CG(active_op_array)</span> function_token->u.op_array =<span> CG(active_op_array); lcname </span>=<span> zend_str_tolower_dup(name, name_len); </span><span>//</span><span> 对op_array进行初始化,强制op_array.fn_flags会被初始化为0</span> orig_interactive =<span> CG(interactive); CG(interactive) </span>= <span>0</span><span>; init_op_array(</span>&<span>op_array, ZEND_USER_FUNCTION, INITIAL_OP_ARRAY_SIZE TSRMLS_CC); CG(interactive) </span>=<span> orig_interactive; </span><span>//</span><span> 对op_array的一些设置</span> op_array.function_name =<span> name; op_array.return_reference </span>=<span> return_reference; op_array.fn_flags </span>|=<span> fn_flags; op_array.pass_rest_by_reference </span>= <span>0</span><span>; op_array.scope </span>= is_method ?<span> CG(active_class_entry):NULL; op_array.prototype </span>=<span> NULL; op_array.line_start </span>= zend_get_compiled_lineno(TSRMLS_C);
function_token便是对function字面进行词法分析而生成的znode。这段代码一开始,就让它保存当前的CG(active_op_array),即函数体之外的op_array。保存好CG(active_op_array)之后,便会开始对函数自己的op_array进行初始化。
op_array.fn_flags是个多功能字段,还记得上一篇中提到的交互式么,如果php以交互式打开,则op_array.fn_flags会被初始化为ZEND_ACC_INTERACTIVE,否则会被初始化为0。这里在init_op_array之前设置CG(interactive) = 0,便是确保op_array.fn_flags初始化为0。随后会进一步执行op_array.fn_flags |= fn_flags,如果是在方法中,则op_array.fn_flags含义为static、abstract、final等修饰符,对函数来讲,op_array.fn_flags依然是0。
zend_op *opline =<span> get_next_op(CG(active_op_array) TSRMLS_CC); </span><span>//</span><span> 如果处于命名空间,则函数名还需要加上命名空间</span> <span>if</span><span> (CG(current_namespace)) { </span><span>/*</span><span> Prefix function name with current namespace name </span><span>*/</span><span> znode tmp; tmp.u.constant </span>= *<span>CG(current_namespace); zval_copy_ctor(</span>&<span>tmp.u.constant); zend_do_build_namespace_name(</span>&tmp, &<span>tmp, function_name TSRMLS_CC); op_array.function_name </span>=<span> Z_STRVAL(tmp.u.constant); efree(lcname); name_len </span>=<span> Z_STRLEN(tmp.u.constant); lcname </span>=<span> zend_str_tolower_dup(Z_STRVAL(tmp.u.constant), name_len); } </span><span>//</span><span> 设置opline</span> opline->opcode =<span> ZEND_DECLARE_FUNCTION; </span><span>//</span><span> 第一个操作数</span> opline->op1.op_type =<span> IS_CONST; build_runtime_defined_function_key(</span>&opline-><span>op1.u.constant, lcname, name_len TSRMLS_CC); </span><span>//</span><span> 第二个操作数</span> opline->op2.op_type =<span> IS_CONST; opline</span>->op2.u.constant.type =<span> IS_STRING; opline</span>->op2.u.constant.value.str.val =<span> lcname; opline</span>->op2.u.constant.value.str.len =<span> name_len; Z_SET_REFCOUNT(opline</span>->op2.u.constant, <span>1</span><span>); opline</span>->extended_value =<span> ZEND_DECLARE_FUNCTION; </span><span>//</span><span> 切换CG(active_op_array)成函数自己的op_array</span> zend_hash_update(CG(function_table), opline->op1.u.constant.value.str.val, opline->op1.u.constant.value.str.len, &op_array, <span>sizeof</span>(zend_op_array), (<span>void</span> **) &CG(active_op_array));
上面这段代码很关键。有几点要说明的:
1,如果函数是处于命名空间中,则其名称会被扩展成命名空间\函数名。比如:
<?<span>php namespace MyProject; </span><span>function</span> foo(<span>$arg1</span>, <span>$arg2</span> = 100<span>) { </span><span>print</span>(<span>$arg1</span><span>); }</span>
则会将函数名改为MyProject\foo。扩展工作由zend_do_build_namespace_name来完成。
2,build_runtime_defined_function_key会生成一个“key”。除了用到函数名称之外,还用到了函数所在文件路径、代码在内存中的地址等等。具体的实现可以自行阅读。将函数放进CG(function_table)时,用的键便是这个“key”。
3,代码中的op_line获取时,尚未发生CG(active_op_array)的切换。也就是说,op_line依然是外层op_array的一条指令。该指令具体为ZEND_DECLARE_FUNCTION,有两个操作数,第一个操作数保存了第二点中提到的“key”,第二个操作数则保存了形如"myproject\foo"这样的函数名(小写)。
4,这段代码的最后,将函数自身对应的op_array存放进了CG(function_table),同时,完成了CG(active_op_array)的切换。从这条语句开始,CG(active_op_array)便开始指向函数自己的op_array,而不再是函数体外层的op_array了。
继续来看zend_do_begin_function_declaration的最后一段:
<span>//</span><span> 需要debuginfo,则函数体内的第一条zend_op,为ZEND_EXT_NOP</span> <span>if</span> (CG(compiler_options) &<span> ZEND_COMPILE_EXTENDED_INFO) { zend_op </span>*opline =<span> get_next_op(CG(active_op_array) TSRMLS_CC); opline</span>->opcode =<span> ZEND_EXT_NOP; opline</span>->lineno =<span> function_begin_line; SET_UNUSED(opline</span>-><span>op1); SET_UNUSED(opline</span>-><span>op2); } </span><span>//</span><span> 控制switch和foreach内声明的函数</span> <span>{ </span><span>/*</span><span> Push a seperator to the switch and foreach stacks </span><span>*/</span><span> zend_switch_entry switch_entry; switch_entry.cond.op_type </span>=<span> IS_UNUSED; switch_entry.default_case </span>= <span>0</span><span>; switch_entry.control_var </span>= <span>0</span><span>; zend_stack_push(</span>&CG(switch_cond_stack), (<span>void</span> *) &switch_entry, <span>sizeof</span><span>(switch_entry)); { </span><span>/*</span><span> Foreach stack separator </span><span>*/</span><span> zend_op dummy_opline; dummy_opline.result.op_type </span>=<span> IS_UNUSED; dummy_opline.op1.op_type </span>=<span> IS_UNUSED; zend_stack_push(</span>&, (<span>void</span> *) &dummy_opline, <span>sizeof</span><span>(zend_op)); } } </span><span>//</span><span> 保存函数的注释语句</span> <span>if</span><span> (CG(doc_comment)) { CG(active_op_array)</span>->doc_comment =<span> CG(doc_comment); CG(active_op_array)</span>->doc_comment_len =<span> CG(doc_comment_len); CG(doc_comment) </span>=<span> NULL; CG(doc_comment_len) </span>= <span>0</span><span>; } </span><span>//</span><span> 作用和上面switch,foreach是一样的,函数体内的语句并不属于函数体外的label</span> zend_stack_push(&CG(labels_stack), (<span>void</span> *) &CG(labels), <span>sizeof</span>(HashTable*<span>)); CG(labels) </span>= NULL;
可能初学者会对CG(switch_cond_stack),CG(foreach_copy_stack),CG(labels_stack)等字段有疑惑。其实也很好理解。以CG(labels_stack)为例,由于进入函数体内之后,op_array发生了切换,外层的CG(active_op_array)被保存到function znode的u.op_array中(如果记不清楚了回头看上文:-))。因此函数外层已经被parse出的一些label也需要被保存下来,用的正是CG(labels_stack)来保存。当函数体完成编译之后,zend vm可以从CG(labels_stack)中恢复出原先的label。举例来说,
<?<span>php label1</span>: <span>function</span> foo(<span>$arg1</span><span>) { </span><span>print</span>(<span>$arg1</span><span>); goto label2; label2</span>: <span>exit</span><span>; } </span><span>$bar</span> = 'hello php'<span>; foo(</span><span>$bar</span>);
解释器在进入zend_do_begin_function_declaration时,CG(labels)中保存的是“label1”。当解释器开始编译函数foo,则需要将“label1”保存到CG(labels_stack)中,同时清空CG(labels)。因为在编译foo的过程中,CG(labels)会保存“labe2”。当foo编译完成,会利用CG(labels_stack)来恢复CG(labels),则CG(labels)再次变成“label1”。
至此,整个zend_do_begin_function_declaration过程已经全部分析完成。最重要的是,一旦完成zend_do_begin_function_declaration,CG(active_op_array)就指向了函数自身对应的op_array。同时,也利用生成的“key”在CG(function_table)中替函数占了一个位。
函数可以定义为不接受任何参数,对于参数列表为空的情况,其实不做任何处理。我们前文的例子foo函数,接受了一个参数$arg1,我们下面还是分析有参数的情况。
根据语法推导式non_empty_parameter_list的定义,参数列表一共有8种,前4种对应的是一个参数,后4种对应多个参数。我们只关心前4种,后4种编译的过程,仅仅是重复前4种的步骤而已。
optional_class_type T_VARIABLE { znode tmp; fetch_simple_variable(&tmp, &$2, 0 TSRMLS_CC); $$.op_type = IS_CONST; Z_LVAL($$.u.constant)=1; Z_TYPE($$.u.constant)=IS_LONG; INIT_PZVAL(&$$.u.constant); zend_do_receive_arg(ZEND_RECV, &tmp, &$$, NULL, &$1, &$2, 0 TSRMLS_CC); } optional_class_type '&' T_VARIABLE { znode tmp; fetch_simple_variable(&tmp, &$3, 0 TSRMLS_CC); $$.op_type = IS_CONST; Z_LVAL($$.u.constant)=1; Z_TYPE($$.u.constant)=IS_LONG; INIT_PZVAL(&$$.u.constant); zend_do_receive_arg(ZEND_RECV, &tmp, &$$, NULL, &$1, &$3, 1 TSRMLS_CC); } optional_class_type '&' T_VARIABLE '=' static_scalar { znode tmp; fetch_simple_variable(&tmp, &$3, 0 TSRMLS_CC); $$.op_type = IS_CONST; Z_LVAL($$.u.constant)=1; Z_TYPE($$.u.constant)=IS_LONG; INIT_PZVAL(&$$.u.constant); zend_do_receive_arg(ZEND_RECV_INIT, &tmp, &$$, &$5, &$1, &$3, 1 TSRMLS_CC); } optional_class_type T_VARIABLE '=' static_scalar { znode tmp; fetch_simple_variable(&tmp, &$2, 0 TSRMLS_CC); $$.op_type = IS_CONST; Z_LVAL($$.u.constant)=1; Z_TYPE($$.u.constant)=IS_LONG; INIT_PZVAL(&$$.u.constant); zend_do_receive_arg(ZEND_RECV_INIT, &tmp, &$$, &$4, &$1, &$2, 0 TSRMLS_CC); }
前4种情况,具体又可以分为2类,1类没有默认值,区别只在于参数的传递是否采用引用,而另1类,都有默认值“static_scalar”。
实际上区别并不大,它们的语法分析的处理过程也几乎一致。都是先调用fetch_simple_variable,再执行zend_do_receive_arg。有没有默认值,区别也仅仅在于zend_do_receive_arg的参数,会不会将默认值传递进去。先来看fetch_simple_variable。
fetch_simple_variable是用来获取compiled variables索引的。compiled variables被视作php的性能提升手段之一,因为它利用数组存储了变量,而并非内核中普遍使用的HashTable。这里可以看出,函数的任何一个参数,均会被编译为compiled variables,compiled variables被保存在函数体op_array->vars数组中。虽然根据变量名称去HashTable查询,效率并不低。但显然根据索引去op_array->vars数组中获取变量,会更加高效。
<span>void</span> fetch_simple_variable_ex(znode *result, znode *varname, <span>int</span> bp, zend_uchar op TSRMLS_DC) <span>/*</span><span> {{{ </span><span>*/</span><span> { zend_op opline; ... </span><span>if</span> (varname->op_type ==<span> IS_CONST) { </span><span>if</span> (Z_TYPE(varname->u.constant) !=<span> IS_STRING) { convert_to_string(</span>&varname-><span>u.constant); } </span><span>if</span> (!zend_is_auto_global(varname->u.constant.value.str.val, varname->u.constant.value.str.len TSRMLS_CC) && !(varname->u.constant.value.str.len == (<span>sizeof</span>(<span>"</span><span>this</span><span>"</span>)-<span>1</span>) && !memcmp(varname->u.constant.value.str.val, <span>"</span><span>this</span><span>"</span>, <span>sizeof</span>(<span>"</span><span>this</span><span>"</span>))) &&<span> (CG(active_op_array)</span>->last == <span>0</span> || CG(active_op_array)->opcodes[CG(active_op_array)->last-<span>1</span>].opcode !=<span> ZEND_BEGIN_SILENCE)) { </span><span>//</span><span> 节点的类型为IS_CV,表明是compiled variables</span> result->op_type =<span> IS_CV; </span><span>//</span><span> 用u.var来记录compiled variables在CG(active_op_array)->vars中的索引</span> result->u.<span>var</span> = lookup_cv(CG(active_op_array), varname->u.constant.value.str.val, varname-><span>u.constant.value.str.len); result</span>->u.EA.type = <span>0</span><span>; varname</span>->u.constant.value.str.val = CG(active_op_array)->vars[result->u.<span>var</span><span>].name; </span><span>return</span><span>; } } ... }</span>
这里不做详细的分析了。当fetch_simple_variable获取索引之后,znode中就不必再保存变量的名称,取而代之的是变量在vars数组中的索引,即znode->u.var,其类型为int。fetch_simple_variable完成,会进入zend_do_receive_arg。
zend_do_receive_arg目的是生成一条zend op指令,可以称作RECV。
一般而言,除非函数不存在参数,否则RECV是函数的第一条指令(这里表述不准,有extend info时也不是第一条)。该指令的opcode可能为ZEND_RECV或者ZEND_RECV_INIT,取决于是否有默认值。如果参数没有默认值,指令等于ZEND_RECV,有默认值,则为ZEND_RECV_INIT。zend_do_receive_arg的第二个参数,就是上面提到的compiled variables节点。
分析下zend_do_receive_arg的源码,也是分几段来看:
zend_op *<span>opline; zend_arg_info </span>*<span>cur_arg_info; </span><span>//</span><span> class_type主要用于限制函数参数的类型</span> <span>if</span> (class_type->op_type == IS_CONST && Z_TYPE(class_type->u.constant) == IS_STRING && Z_STRLEN(class_type->u.constant) == <span>0</span><span>) { </span><span>/*</span><span> Usage of namespace as class name not in namespace </span><span>*/</span><span> zval_dtor(</span>&class_type-><span>u.constant); zend_error(E_COMPILE_ERROR, </span><span>"</span><span>Cannot use 'namespace' as a class name</span><span>"</span><span>); </span><span>return</span><span>; } </span><span>//</span><span> 对静态方法来说,参数不能为this</span> <span>if</span> (<span>var</span>->op_type == IS_CV && <span>var</span>->u.<span>var</span> == CG(active_op_array)->this_var && (CG(active_op_array)->fn_flags & ZEND_ACC_STATIC) == <span>0</span><span>) { zend_error(E_COMPILE_ERROR, </span><span>"</span><span>Cannot re-assign $this</span><span>"</span><span>); } </span><span>else</span> <span>if</span> (<span>var</span>->op_type == IS_VAR && CG(active_op_array)->scope && ((CG(active_op_array)->fn_flags & ZEND_ACC_STATIC) == <span>0</span>) && (Z_TYPE(varname->u.constant) == IS_STRING) && (Z_STRLEN(varname->u.constant) == <span>sizeof</span>(<span>"</span><span>this</span><span>"</span>)-<span>1</span>) && (memcmp(Z_STRVAL(varname->u.constant), <span>"</span><span>this</span><span>"</span>, <span>sizeof</span>(<span>"</span><span>this</span><span>"</span>)) == <span>0</span><span>)) { zend_error(E_COMPILE_ERROR, </span><span>"</span><span>Cannot re-assign $this</span><span>"</span><span>); } </span><span>//</span><span> CG(active_op_array)此时已经是函数体的op_array了,这里拿一条指令</span> opline =<span> get_next_op(CG(active_op_array) TSRMLS_CC); CG(active_op_array)</span>->num_args++<span>; opline</span>->opcode =<span> op; opline</span>->result = *<span>var</span><span>; </span><span>//</span><span> op1节点表明是第几个参数</span> opline->op1 = *<span>offset; </span><span>//</span><span> op2节点可能为初始值,也可能为UNUSED</span> <span>if</span> (op ==<span> ZEND_RECV_INIT) { opline</span>->op2 = *<span>initialization; } </span><span>else</span><span> { CG(active_op_array)</span>->required_num_args = CG(active_op_array)-><span>num_args; SET_UNUSED(opline</span>-><span>op2); }</span>
上面这段代码,首先通过get_next_op(CG(active_op_array) TSRMLS_CC)一句获取了opline,opline是未被使用的一条zend_op指令。紧接着,会对opline的各个字段进行设置。opline->op1表明这是第几个参数,opline->op2可能为初始值,也可能被设置为UNUSED。
如果一个参数有默认值,那么在调用函数时,其实是可以不用传递该参数的。所以,required_num_args不会将这类非必须的参数算进去的。可以看到,在op == ZEND_RECV_INIT这段逻辑分支中,并没有处理required_num_args。
继续来看:
<span>//</span><span> 这里采用erealloc进行分配,因为期望最终会形成一个参数信息的数组</span> CG(active_op_array)->arg_info = erealloc(CG(active_op_array)->arg_info, <span>sizeof</span>(zend_arg_info)*(CG(active_op_array)-><span>num_args)); </span><span>//</span><span> 设置当前的zend_arg_info</span> cur_arg_info = &CG(active_op_array)->arg_info[CG(active_op_array)->num_args-<span>1</span><span>]; cur_arg_info</span>->name = estrndup(varname->u.constant.value.str.val, varname-><span>u.constant.value.str.len); cur_arg_info</span>->name_len = varname-><span>u.constant.value.str.len; cur_arg_info</span>->array_type_hint = <span>0</span><span>; cur_arg_info</span>->allow_null = <span>1</span><span>; cur_arg_info</span>->pass_by_reference =<span> pass_by_reference; cur_arg_info</span>->class_name =<span> NULL; cur_arg_info</span>->class_name_len = <span>0</span><span>; </span><span>//</span><span> 如果需要对参数做类型限定</span> <span>if</span> (class_type->op_type !=<span> IS_UNUSED) { cur_arg_info</span>->allow_null = <span>0</span><span>; </span><span>//</span><span> 限定为类</span> <span>if</span> (class_type->u.constant.type ==<span> IS_STRING) { </span><span>if</span> (ZEND_FETCH_CLASS_DEFAULT == zend_get_class_fetch_type(Z_STRVAL(class_type->u.constant), Z_STRLEN(class_type-><span>u.constant))) { zend_resolve_class_name(class_type, </span>&opline->extended_value, <span>1</span><span> TSRMLS_CC); } cur_arg_info</span>->class_name = class_type-><span>u.constant.value.str.val; cur_arg_info</span>->class_name_len = class_type-><span>u.constant.value.str.len; </span><span>//</span><span> 如果限定为类,则参数的默认值只能为NULL</span> <span>if</span> (op ==<span> ZEND_RECV_INIT) { </span><span>if</span> (Z_TYPE(initialization->u.constant) == IS_NULL || (Z_TYPE(initialization->u.constant) == IS_CONSTANT && !strcasecmp(Z_STRVAL(initialization->u.constant), <span>"</span><span>NULL</span><span>"</span><span>))) { cur_arg_info</span>->allow_null = <span>1</span><span>; } </span><span>else</span><span> { zend_error(E_COMPILE_ERROR, </span><span>"</span><span>Default value for parameters with a class type hint can only be NULL</span><span>"</span><span>); } } } </span><span>//</span><span> 限定为数组</span> <span>else</span><span> { </span><span>//</span><span> 将array_type_hint设置为1</span> cur_arg_info->array_type_hint = <span>1</span><span>; cur_arg_info</span>->class_name =<span> NULL; cur_arg_info</span>->class_name_len = <span>0</span><span>; </span><span>//</span><span> 如果限定为数组,则参数的默认值只能为数组或NULL</span> <span>if</span> (op ==<span> ZEND_RECV_INIT) { </span><span>if</span> (Z_TYPE(initialization->u.constant) == IS_NULL || (Z_TYPE(initialization->u.constant) == IS_CONSTANT && !strcasecmp(Z_STRVAL(initialization->u.constant), <span>"</span><span>NULL</span><span>"</span><span>))) { cur_arg_info</span>->allow_null = <span>1</span><span>; } </span><span>else</span> <span>if</span> (Z_TYPE(initialization->u.constant) != IS_ARRAY && Z_TYPE(initialization->u.constant) !=<span> IS_CONSTANT_ARRAY) { zend_error(E_COMPILE_ERROR, </span><span>"</span><span>Default value for parameters with array type hint can only be an array or NULL</span><span>"</span><span>); } } } } opline</span>->result.u.EA.type |= EXT_TYPE_UNUSED;
这部分代码写的很清晰。注意,对于限定为数组的情况,class_type的op_type会被设置为IS_CONST,而u.constant.type会被设置为IS_NULL:
optional_class_type: /* empty */ { $$.op_type = IS_UNUSED; } | fully_qualified_class_name { $$ = $1; } | T_ARRAY { $$.op_type = IS_CONST; Z_TYPE($$.u.constant)=IS_NULL;}
因此,zend_do_receive_arg中区分限定为类还是数组,是利用class_type->u.constant.type == IS_STRING来判断的。如果类型限定为数组,则cur_arg_info->array_type_hint会被设置为1。
还有另一个地方需要了解,zend_resolve_class_name函数会修正类名。举例来说:
<?<span>php namespace A; </span><span>class</span><span> B { } </span><span>function</span> foo(B <span>$arg1</span>, <span>$arg2</span> = 100<span>) { </span><span>print</span>(<span>$arg1</span><span>); }</span>
我们期望参数arg1的类型为B,class_type中也保存了B。但是因为位于命名空间A下,所以,zend_resolve_class_name会将class_type中保存的类名B,修正为A\B。
OK,到这里,zend_do_receive_arg已经全部分析完。zend vm在分析函数参数时,每遇见一个参数,便会调用一次zend_do_receive_arg,生成一条RECV指令。因此,函数有几个参数,就会编译出几条RECV指令。
当编译完参数列表,zend vm便会进入函数内部了。函数体的编译其实和正常语句的编译一样。zend vm只需要将函数体内部的php语句,按照正常的statment,进行词法分析、语法分析来处理,最终形成一条条zend_op指令。
来看下语法文件:
unticked_function_declaration_statement: function is_reference T_STRING { zend_do_begin_function_declaration(&$1, &$3, 0, $2.op_type, NULL TSRMLS_CC); } '(' parameter_list ')' '{' inner_statement_list '}' { zend_do_end_function_declaration(&$1 TSRMLS_CC); }<br />;
函数体内部的语句,表示为inner_statement_list。
inner_statement_list: inner_statement_list { zend_do_extended_info(TSRMLS_C); } inner_statement { HANDLE_INTERACTIVE(); } | /* empty */ ;
而inner_statment正是由语句、函数声明、类声明组成的。
inner_statement: statement | function_declaration_statement | class_declaration_statement | T_HALT_COMPILER '(' ')' ';' { zend_error(E_COMPILE_ERROR, "__HALT_COMPILER() can only be used from the outermost scope"); } ;
inner_statement并非专门用于函数,其他譬如foreach,while循环等有block语句块中,都会被识别为inner_statement。从这里其实还能看到一些有意思的语法,比如说我们可以在函数里声明一个类。inner_statement就不展开叙述了,否则相当于将整个php的语法捋一遍,情况太多了。
我们最后来看下结束编译的过程。结束函数编译是通过zend_do_end_function_declaration来完成的。
zend_do_end_function_declaration接收的参数function_token,其实就是前面提到过的function字面对应的znode。根据我们在“开始编译”一节所述,function_token中保留了函数体之外的op_array。
<span>char</span> lcname[<span>16</span><span>]; </span><span>int</span><span> name_len; zend_do_extended_info(TSRMLS_C); </span><span>//</span><span> 返回NULL</span> zend_do_return(NULL, <span>0</span><span> TSRMLS_CC); </span><span>//</span><span> 通过op指令设置对应的handler函数</span> <span>pass_two(CG(active_op_array) TSRMLS_CC); </span><span>//</span><span> 释放当前函数的CG(labels),并从CG(labels_stack)中还原之前的CG(labels)</span> <span>zend_release_labels(TSRMLS_C); </span><span>if</span><span> (CG(active_class_entry)) { </span><span>//</span><span> 检查魔术方法的参数是否合法</span> zend_check_magic_method_implementation(CG(active_class_entry), (zend_function*<span>)CG(active_op_array), E_COMPILE_ERROR TSRMLS_CC); } </span><span>else</span><span> { </span><span>/*</span><span> we don't care if the function name is longer, in fact lowercasing only * the beginning of the name speeds up the check process </span><span>*/</span><span> name_len </span>= strlen(CG(active_op_array)-><span>function_name); zend_str_tolower_copy(lcname, CG(active_op_array)</span>->function_name, MIN(name_len, <span>sizeof</span>(lcname)-<span>1</span><span>)); lcname[</span><span>sizeof</span>(lcname)-<span>1</span>] = <span>'</span><span>\0</span><span>'</span>; <span>/*</span><span> zend_str_tolower_copy won't necessarily set the zero byte </span><span>*/</span> <span>//</span><span> 检查__autoload函数的参数是否合法</span> <span>if</span> (name_len == <span>sizeof</span>(ZEND_AUTOLOAD_FUNC_NAME) - <span>1</span> && !memcmp(lcname, ZEND_AUTOLOAD_FUNC_NAME, <span>sizeof</span>(ZEND_AUTOLOAD_FUNC_NAME)) && CG(active_op_array)->num_args != <span>1</span><span>) { zend_error(E_COMPILE_ERROR, </span><span>"</span><span>%s() must take exactly 1 argument</span><span>"</span><span>, ZEND_AUTOLOAD_FUNC_NAME); } } CG(active_op_array)</span>->line_end =<span> zend_get_compiled_lineno(TSRMLS_C); </span><span>//</span><span> 很关键!将CG(active_op_array)还原成函数外层的op_array</span> CG(active_op_array) = function_token-><span>u.op_array; </span><span>/*</span><span> Pop the switch and foreach seperators </span><span>*/</span><span> zend_stack_del_top(</span>&<span>CG(switch_cond_stack)); zend_stack_del_top(</span>&CG(foreach_copy_stack));
有3处值得注意:
1,zend_do_end_function_declaration中会对CG(active_op_array)进行还原。用的正是function_token->u.op_array。一旦zend_do_end_function_declaration完成,函数的整个编译过程就已经结束了。zend vm会继续看接下来函数之外的代码,所以需要将CG(active_op_array)切换成原先的。
2,zend_do_return负责在函数最后添加上一条RETURN指令,因为我们传进去的是NULL,所以这条RETURN指令的操作数被强制设置为UNUSED。注意,不管函数本身是否有return语句,最后这条RETURN指令是必然存在的。假如函数有return语句,return语句也会产生一条RETURN指令,所以会导致可能出现多条RETURN指令。举例来说:
<span>function</span> foo(<span>)<br /><span>{</span> </span><span>return</span> <span>true</span><span>; }</span>
编译出来的OP指令最后两条如下:
RETURN true RETURN null
我们可以很明显在最后看到两条RETURN。一条是通过return true编译出来的。另一条,就是在zend_do_end_function_declaration阶段,强制插入的RETURN。
3,我们刚才讲解的所有步骤中,都只是设置了每条指令的opcode,而并没有设置这条指令具体的handle函数。pass_two会负责遍历每条zend_op指令,根据opcode,以及操作数op1和op2,去查找并且设置对应的handle函数。这项工作,是通过ZEND_VM_SET_OPCODE_HANDLER(opline)宏来完成的。
<span>#define</span> ZEND_VM_SET_OPCODE_HANDLER(opline) zend_vm_set_opcode_handler(opline)
zend_vm_set_opcode_handler的实现很简单:
<span>void</span> zend_init_opcodes_handlers(<span>void</span><span>) { </span><span>//</span><span> 超大的数组,里面存放了所有的handler</span> <span>static</span> <span>const</span> opcode_handler_t labels[] =<span> { ZEND_NOP_SPEC_HANDLER, ZEND_NOP_SPEC_HANDLER, ZEND_NOP_SPEC_HANDLER, ZEND_NOP_SPEC_HANDLER, ZEND_NOP_SPEC_HANDLER, ZEND_NOP_SPEC_HANDLER, ... }; zend_opcode_handlers </span>= (opcode_handler_t*<span>)labels; } </span><span>static</span> opcode_handler_t zend_vm_get_opcode_handler(zend_uchar opcode, zend_op*<span> op) { </span><span>static</span> <span>const</span> <span>int</span> zend_vm_decode[] =<span> { _UNUSED_CODE, </span><span>/*</span><span> 0 </span><span>*/</span><span> _CONST_CODE, </span><span>/*</span><span> 1 = IS_CONST </span><span>*/</span><span> _TMP_CODE, </span><span>/*</span><span> 2 = IS_TMP_VAR </span><span>*/</span><span> _UNUSED_CODE, </span><span>/*</span><span> 3 </span><span>*/</span><span> _VAR_CODE, </span><span>/*</span><span> 4 = IS_VAR </span><span>*/</span><span> _UNUSED_CODE, </span><span>/*</span><span> 5 </span><span>*/</span><span> _UNUSED_CODE, </span><span>/*</span><span> 6 </span><span>*/</span><span> _UNUSED_CODE, </span><span>/*</span><span> 7 </span><span>*/</span><span> _UNUSED_CODE, </span><span>/*</span><span> 8 = IS_UNUSED </span><span>*/</span><span> _UNUSED_CODE, </span><span>/*</span><span> 9 </span><span>*/</span><span> _UNUSED_CODE, </span><span>/*</span><span> 10 </span><span>*/</span><span> _UNUSED_CODE, </span><span>/*</span><span> 11 </span><span>*/</span><span> _UNUSED_CODE, </span><span>/*</span><span> 12 </span><span>*/</span><span> _UNUSED_CODE, </span><span>/*</span><span> 13 </span><span>*/</span><span> _UNUSED_CODE, </span><span>/*</span><span> 14 </span><span>*/</span><span> _UNUSED_CODE, </span><span>/*</span><span> 15 </span><span>*/</span><span> _CV_CODE </span><span>/*</span><span> 16 = IS_CV </span><span>*/</span><span> }; </span><span>//</span><span> 去handler数组里找到对应的处理函数</span> <span>return</span> zend_opcode_handlers[opcode * <span>25</span> + zend_vm_decode[op->op1.op_type] * <span>5</span> + zend_vm_decode[op-><span>op2.op_type]]; } ZEND_API </span><span>void</span> zend_vm_set_opcode_handler(zend_op*<span> op) { </span><span>//</span><span> 给zend op设置对应的handler函数</span> op->handler = zend_vm_get_opcode_handler(zend_user_opcodes[op-><span>opcode], op); }</span>
所有的opcode都定义在zend_vm_opcodes.h里,从php5.3-php5.6,大概从150增长到170个opcode。上面可以看到通过opcode查找handler的准确算法:
zend_opcode_handlers[opcode * <span>25</span> + zend_vm_decode[op->op1.op_type] * <span>5</span> + zend_vm_decode[op->op2.op_type]
不过zend_opcode_handlers数组太大了...找起来很麻烦。
下面回到文章开始的那段php代码,我们将函数foo进行编译,最终得到的指令如下:
可以看出,因为foo指接受一个参数,所以这里只有一条RECV指令。
print语句的参数为!0,!0是一个compiled variables,其实就是参数中的arg1。0代表着索引,回忆一下,函数的op_array有一个数组专门用于保存compiled variables,0表明arg1位于该数组的开端。
print语句有返回值,所以会存在一个临时变量保存其返回值,即~0。由于我们在函数中并未使用~0,所以随即便会有一条FREE指令对其进行释放。
在函数的最后,是一条RETURN指令。
函数编译完成之后,还需要进行的一步是绑定。zend vm通过zend_do_early_binding来实现绑定。这个名字容易让人产生疑惑,其实只有在涉及到类和方法的时候,才会有早期绑定,与之相对的是延迟绑定,或者叫后期绑定。纯粹函数谈不上这种概念,不过zend_do_early_binding是多功能的,并非仅仅为绑定方法而实现。
来看下zend_do_early_binding:
<span>//</span><span> 拿到的是最近一条zend op,对于函数来说,就是ZEND_DECLARE_FUNCTION</span> zend_op *opline = &CG(active_op_array)->opcodes[CG(active_op_array)->last-<span>1</span><span>]; HashTable </span>*<span>table; </span><span>while</span> (opline->opcode == ZEND_TICKS && opline > CG(active_op_array)-><span>opcodes) { opline</span>--<span>; } </span><span>switch</span> (opline-><span>opcode) { </span><span>case</span><span> ZEND_DECLARE_FUNCTION: </span><span>//</span><span> 真正绑定函数</span> <span>if</span> (do_bind_function(opline, CG(function_table), <span>1</span>) ==<span> FAILURE) { </span><span>return</span><span>; } table </span>=<span> CG(function_table); </span><span>break</span><span>; </span><span>case</span><span> ZEND_DECLARE_CLASS: ... </span><span>case</span><span> ZEND_DECLARE_INHERITED_CLASS: ... } </span><span>//</span><span> op1中保存的是函数的key,这里其从将CG(function_table)中删除</span> zend_hash_del(table, opline->op1.u.constant.value.str.val, opline-><span>op1.u.constant.value.str.len); zval_dtor(</span>&opline-><span>op1.u.constant); zval_dtor(</span>&opline-><span>op2.u.constant); </span><span>//</span><span> opline置为NOP</span> MAKE_NOP(opline);
这个函数实现也很简单,主要就是调用了do_bind_function。
ZEND_API <span>int</span> do_bind_function(zend_op *opline, HashTable *function_table, zend_bool compile_time) <span>/*</span><span> {{{ </span><span>*/</span><span> { zend_function </span>*<span>function; </span><span>//</span><span> 找出函数</span> zend_hash_find(function_table, opline->op1.u.constant.value.str.val, opline->op1.u.constant.value.str.len, (<span>void</span> *) &<span>function); </span><span>//</span><span> 以函数名称作为key,重新加入function_table</span> <span>if</span> (zend_hash_add(function_table, opline->op2.u.constant.value.str.val, opline->op2.u.constant.value.str.len+<span>1</span>, function, <span>sizeof</span>(zend_function), NULL)==<span>FAILURE) { </span><span>int</span> error_level = compile_time ?<span> E_COMPILE_ERROR : E_ERROR; zend_function </span>*<span>old_function; </span><span>//</span><span> 加入失败,可能发生重复定义了</span> <span>if</span> (zend_hash_find(function_table, opline->op2.u.constant.value.str.val, opline->op2.u.constant.value.str.len+<span>1</span>, (<span>void</span> *) &old_function)==<span>SUCCESS </span>&& old_function->type ==<span> ZEND_USER_FUNCTION </span>&& old_function->op_array.last > <span>0</span><span>) { zend_error(error_level, </span><span>"</span><span>Cannot redeclare %s() (previously declared in %s:%d)</span><span>"</span><span>, function</span>->common.function_name, old_function->op_array.filename, old_function->op_array.opcodes[<span>0</span><span>].lineno); } </span><span>else</span><span> { zend_error(error_level, </span><span>"</span><span>Cannot redeclare %s()</span><span>"</span>, function-><span>common.function_name); } </span><span>return</span><span> FAILURE; } </span><span>else</span><span> { (</span>*function->op_array.refcount)++<span>; function</span>->op_array.static_variables = NULL; <span>/*</span><span> NULL out the unbound function </span><span>*/</span> <span>return</span><span> SUCCESS; } }</span>
在进入do_bind_function之前,其实CG(function_table)中已经有了函数的op_array。不过用的键并非函数名,而是build_runtime_defined_function_key生成的“key”,这点在前面“开始编译”一节中有过介绍。do_bind_function所做的事情,正是利用这个“key”,将函数查找出来,并且以真正的函数名为键,重新插入到CG(function_table)中。
因此当do_bind_function完成时,function_table中有2个键可以查询到该函数。一个是“key”为索引的,另一个是以函数名为索引的。在zend_do_early_binding的最后,会通过zend_hash_del来删除“key”,从而保证function_table中,该函数只能够以函数名为键值查询到。
这篇其实主要是为了弄清楚,函数如何被编译成op_array。一些关键的步骤如下图:
至于函数的调用,又是另外一个话题了。