This RFC proposes the introduction of an Abstract Syntax Tree (AST) as an intermediary structure in our compilation process. This replaces the existing practice of emitting opcodes directly from the parser.
Decoupling the parser and compiler allows us to remove a number of hacks and makes the implementation more maintainable and understandable in general. Furthermore it allows implementing syntax that was not feasible with a single-pass compilation process.
Apart from being standard practice in any compiler implementation, use of an AST has two primary advantages:
In the new AST-based implementation the compiler is fully decoupled from the parser, which leads to a code quality and maintainability improvement. In the following some examples of such improvements are discussed:
expr
production.close_bracket_token->u.op.opline_num
refer? Where was this opline emitted and what opcode does it have? What does it mean if the op2.opline_num
of that opline is changed?BP_VAR_W
fetches were inserted. Afterwards the oplines were modified or removed depending on the fetch mode that was eventually chosen. The AST-based implementation can directly compile using the correct fetch mode.list()
compilation.
With the current single-pass compiler some syntactical elements are very hard, or impossible, to implement. This has influenced a number of syntactical decisions in the past. One example is the restriction that yield
expressions must be wrapped in parentheses when used in expression context:
$result = yield fn(); // INVALID $result = (yield fn()); // VALID
This choice was made purely due to technical restrictions and is removed in the AST implementation.
Additionally the current compiler architecture prevents us from implementing some types of syntax altogether. Examples include:
[$a, $b, $c] = $array
instead of a dedicated list()
syntax. This is common in other languages, but not possible in PHP.[x * x for x in list]
in Python. In PHP only the reverse syntax is possible: [foreach ($list as $x) yield $x * $x]
Whether or not we actually want these things, I think it's important that syntax is evaluated based on its merit and not based on technical restrictions.
The abstract syntax tree has little impact on runtime performance or memory usage. While the AST does allow us to generate slightly better and smaller instruction sequences in some cases (e.g. for for
loops), the practical impact is not worth mentioning.
The introduction of an AST does however impact the performance and memory usage of the compilation process itself. It should be emphasized that this difference is only relevant when no opcode cache is in use. If an opcode cache is used then each file is only compiled once, as such any difference does not have a practical impact.
The script used to measure the following numbers is available as a Gist. Tests were performed on three files, with different sizes. The “small” one has about 100 lines of code, the “medium” one about 700 and the “large” one about 2800.
The following table shows the time that is needed to compile each of the files 1000 times:
php-ng | php-ast | diff | |
---|---|---|---|
small | 0.180s | 0.160s | -12.5% |
medium | 1.492s | 1.268s | -17.7% |
large | 6.703s | 5.736s | -16.9% |
The following table shows the peak memory usage during a single compilation1):
php-ng | php-ast | diff | |
---|---|---|---|
small | 378kB | 414kB | +9.5% |
medium | 507kB | 643kB | +26.8% |
large | 1084kB | 1857kB | +71.3% |
As compilation of individual files is not very representative of practical usage, I additionally measured the time and memory necessary to include (compile) all files from the PhpParser project. The results are shown in the following table:
php-ng | php-ast | diff | |
---|---|---|---|
time | 25.5ms | 22.8ms | -11.8% |
memory | 2360kB | 2482kB | +5.1% |
To summarize the results: The AST based implementation is 10-15% faster, but requires more memory. The amount of additional memory heavily depends on the file size. A small file needs an additional 10% of memory, whereas a very large file needs 70%. In the more realistic case where multiple files are compiled, the difference in memory usage drops to 5%, because at this point the memory necessary to hold all the compiled scripts is much larger than the memory necessary to compile one.
The introduction of the AST comes with minor changes to syntax and behavior, which are listed in the following:
This is the only syntax related change and was already mentioned previously. yield
in expression use no longer needs parentheses, so all of the following are valid:
$result = yield; $result = yield $v; $result = yield $k => $v;
An open issue of the Uniform Variable Syntax RFC was that ($foo)['bar'] = 'baz'
and $foo['bar'] = 'baz'
did not exhibit the same behavior, because they were compiled with different fetch modes.
Similarly byRef(func())
and byRef((func()))
will now both throw a strict-standards notice if byRef
takes its parameter by-reference, but func
does not return by reference.
Note: The behavior oflist($a, $b) = $a
described below no longer applies. After this RFC was acceptedlist()
assignments that contain the same variable on the left- and right-hand side have been special cased to ensure the right-hand side always evaluates first. This means thatlist($a, $b) = $a
continues working as expected.
list()
currently assigns variables right-to-left, the AST implementation will assign them left-to-right instead:
list($array[], $array[], $array[]) = [1, 2, 3]; var_dump($array); // OLD: $array = [3, 2, 1] // NEW: $array = [1, 2, 3]
Another example where the assignment order is relevant is if both the left and right side of the list assignment use the same variable:
$a = [1, 2]; list($a, $b) = $a; // OLD: $a = 1, $b = 2 // NEW: $a = 1, $b = null + "Undefined index 1" $b = [1, 2]; list($a, $b) = $b; // OLD: $a = null + "Undefined index 0", $b = 2 // NEW: $a = 1, $b = 2
list()
will now access every offset only once:
list(list($a, $b)) = $array; // OLD: $b = $array[0][1]; $a = $array[0][0]; // NEW: $_tmp = $array[0]; $a = $_tmp[0]; $b = $_tmp[1];
The only visible change this has for most purposes is that an “Undefined index” notice is only thrown once per index, not many times.
Empty list()
s are now always disallowed. Previously they were only forbidden in some places:
list() = $a; // INVALID list($b, list()) = $a; // INVALID foreach ($a as list()) // INVALID (was also invalid previously)
Note: The auto-vivification order for reference assignments has been restored to the old behavior in PHP 7.1. The reason for this is that we found hard to avoid memory safety issues with the new order.
While by-reference assignments are (CVs notwithstanding) evaluated left-to-right, auto-vivification currently occurs right-to-left. In the AST implementation this will happen left-to-right instead:
$obj = new stdClass; $obj->a = &$obj->b; $obj->b = 1; var_dump($obj); // OLD: object(stdClass)#1 (2) { ["b"]=> &int(1) ["a"]=> &int(1) } // NEW: object(stdClass)#1 (2) { ["a"]=> &int(1) ["b"]=> &int(1) }
Note: The order can easily be changed, but the old behavior looks like a bug to me, so I decided to keep the new behavior.
Doing calls like $obj->__clone()
is now allowed. This was the only magic method that had a compile-time check preventing some calls to it, which doesn't make sense. If we allow all other magic methods to be called, there's no reason to forbid this one.
The process for converting a PHP file into opcodes now consists of three phases:
The lexer is defined in zend_language_scanner.l
and generated using re2c. The lexer returns token IDs one at a time and optionally provides a semantic value zval (e.g. containing the name of a variable token).
The parser is defined in zend_language_parser.y
and generated using bison. The parser consumes the tokens provided by the lexer and executes semantic actions based on the encountered token sequence. These semantic actions generate the abstract syntax tree, which is finally written into CG(ast)
.
The parser uses the LALR(1) parsing algorithm, which means that it only has one token of lookahead to distinguish different syntactic structures.
The compiler consumes the AST and generates op arrays from it (one per file and function). This happens by recursively walking the AST by invoking zend_compile_*
functions. These compilation functions emit opcodes, i.e. instructions for the Zend VM.
In the following the AST API is outlined, as well as its usage in the parser and the compiler.
A standard AST node is defined as follows:
typedef unsigned short zend_ast_kind; typedef unsigned short zend_ast_attr; typedef struct _zend_ast { zend_ast_kind kind; zend_ast_attr attr; zend_uint lineno; struct _zend_ast *child[1]; } zend_ast;
kind
is a ZEND_AST_*
enum constant indicating the type of the AST node, e.g. ZEND_AST_BINARY_OP
for a binary operation. attr
is a unsigned short that can be used to store kind-specific flags. lineno
is the start line number of the node.
Child nodes are stored in the child
array. The size of this array is determined during allocation based on the kind. Nodes are created using zend_ast_create
or zend_ast_create_ex
, depending on whether you want to make use of attr
:
zend_ast *zend_ast_create_ex(zend_ast_kind kind, zend_ast_attr attr, ...); zend_ast *zend_ast_create(zend_ast_kind kind, ...);
For example:
zend_ast *ast = zend_ast_create_ex(ZEND_AST_BINARY_OP, ZEND_ADD, left_ast, right_ast);
AST nodes created this way have a fixed number of children determined by the AST kind. For cases where the number of children is determined dynamically (e.g. arrays, argument lists, statement lists, etc) the type ``zend_ast_list`` is used instead. It is identical to ordinary AST nodes, but contains an additional children count:
typedef struct _zend_ast_list { zend_ast_kind kind; zend_ast_attr attr; zend_uint lineno; zend_uint children; zend_ast *child[1]; } zend_ast_list;
List nodes are created using zend_ast_create_list
and children are appended using zend_ast_list_add
.
zend_ast_list *zend_ast_create_list(zend_uint init_children, zend_ast_kind kind, ...); zend_ast_list *zend_ast_list_add(zend_ast_list *list, zend_ast *op);
For example, creating and appending to an array AST:
/* Initialize array with two elems */ zend_ast_list *list = zend_ast_create_list(2, ZEND_AST_ARRAY, zend_ast_create(ZEND_AST_ARRAY_ELEM, value1_ast, key1_ast), zend_ast_create(ZEND_AST_ARRAY_ELEM, value2_ast, key2_ast)); /* Add another element afterwards */ list = zend_ast_list_add(list, zend_ast_create(ZEND_AST_ARRAY_ELEM(value3_ast, key3_ast));
Lastly an AST node can store a zval
. For this purpose the ZEND_AST_ZVAL
kind is used in conjunction with the zend_ast_zval
structure:
/* Lineno is stored in val.u2.lineno */ typedef struct _zend_ast_zval { zend_ast_kind kind; zend_ast_attr attr; zval val; } zend_ast_zval;
Zval AST nodes are created using zend_ast_create_zval
or zend_ast_create_zval_ex
(in case attr
is used). There are two additional convenience functions which create a zval AST node from a string or a long:
zend_ast *zend_ast_create_zval_ex(zval *zv, zend_ast_attr attr); zend_ast *zend_ast_create_zval(zval *zv); zend_ast *zend_ast_create_zval_from_str(zend_string *str); zend_ast *zend_ast_create_zval_from_long(long lval);
These functions return the node cast to zend_ast*
for practical purposes. The zend_ast_zval
structure is only used internally and all external code works through zend_ast*
.
There is another special node type for class and function declarations, which is not documented here.
The parser stack now uses zend_parser_stack_elem
unions rather than znode
s. The union is defined as follows:
typedef union _zend_parser_stack_elem { zend_ast *ast; zend_ast_list *list; zend_string *str; zend_ulong num; } zend_parser_stack_elem;
The ast
member is used when creating ordinary AST nodes:
callable_variable: simple_variable { $$.ast = zend_ast_create(ZEND_AST_VAR, $1.ast); } | dereferencable '[' dim_offset ']' { $$.ast = zend_ast_create(ZEND_AST_DIM, $1.ast, $3.ast); } | dereferencable '{' expr '}' { $$.ast = zend_ast_create(ZEND_AST_DIM, $1.ast, $3.ast); } | dereferencable T_OBJECT_OPERATOR member_name argument_list { $$.ast = zend_ast_create(ZEND_AST_METHOD_CALL, $1.ast, $3.ast, $4.ast); } | function_call { $$.ast = $1.ast; } ;
When lists are created or modified, the list
member is used instead:
inner_statement_list: inner_statement_list inner_statement { $$.list = zend_ast_list_add($1.list, $2.ast); } | /* empty */ { $$.list = zend_ast_create_list(0, ZEND_AST_STMT_LIST); } ;
The str
member is used to back up doc comments during parsing and num
is utilized to back up line numbers or store flags.
For ordinary AST nodes you can directly access the children using ast->child[0]
and so on.
When working with a list node you must first retrieve the list using zend_ast_get_list
(this is effectively just a cast to the zend_ast_list*
type). Afterwards you can iterate through the list as follows:
zend_ast_list *list = zend_ast_get_list(ast); zend_uint i; for (i = 0; i < list->children; ++i) { zend_ast *elem = list->child[i]; /* ... */ }
The zval from a zval AST node is fetched using zend_ast_get_zval
. As the zval is commonly known to be a string an additional zend_ast_get_str
function is provided, which returns a zend_string*
.
Apart from these, there are a number of introspection function, which are useful work generic code working on AST nodes:
zend_ast_get_lineno
will return the starting line number for all AST node types. zend_ast_is_list
returns whether an AST node is a listzend_ast_get_num_children
returns the number of children a non-list node has.
As the abstract syntax tree is only necessary during compilation and discarded afterwards, AST nodes make use of an arena allocator. The arena is stored in CG(ast_arena)
. Before invoking zendparse
this arena must be allocated using zend_arena_create
.
Due to the usage of an arena allocator AST nodes do not need to be individually freed, however zvals held by them still need to be destroyed. This is accomplished using zend_ast_destroy
, which will recursively walk the AST and dtor all held values. The following code features a sample invocation of the parser, including arena handling:
CG(ast_arena) = zend_arena_create(1024 * 32); compiler_result = zendparse(TSRMLS_C); if (compiler_result != 0) { zend_bailout(); } zend_compile_top_stmt(CG(ast) TSRMLS_CC); zend_ast_destroy(CG(ast)); zend_arena_destroy(CG(ast_arena));
For constant scalar expressions AST nodes need to be preserved after compilation. For this purpose they need to be copied from the arena into ZMM allocated memory. This is accomplished using the zend_ast_copy
function. The heap-allocated AST can then be destroyed using zend_ast_destroy_and_free
.
The compiler continues to use znode
s to store operands during compilation. Oplines are usually created using the zend_emit_op
or zend_emit_op_tmp
functions:
zend_op *zend_emit_op(znode *result, zend_uchar opcode, znode *op1, znode *op2 TSRMLS_DC); zend_op *zend_emit_op_tmp(znode *result, zend_uchar opcode, znode *op1, znode *op2 TSRMLS_DC);
Both will emit (and return) an opline with the given opcode and operands. zend_emit_op_tmp
will use an IS_TMP_VAR
variable for the result. zend_emit_op
uses an IS_VAR
variable instead. zend_emit_op
also accepts NULL
for the result, in which case the opline will have no result (IS_UNUSED
).
Both op1
and op2
can also be NULL
, in which case they are unused.
There are additional functions for handling jump instructions. zend_emit_jump
and zend_emit_cond_jump
will emit jumps / conditional jumps. In many cases the jump target is not known at the time the opline is emitted. In this case 0
should be passed for opnum_target
and the opnum returned by the function backed up. Afterwards the returned opnum can be used to update the jump target using zend_update_jump_target
:
zend_uint zend_emit_jump(zend_uint opnum_target TSRMLS_DC); zend_uint zend_emit_cond_jump(zend_uchar opcode, znode *cond, zend_uint opnum_target TSRMLS_DC); void zend_update_jump_target(zend_uint opnum_jump, zend_uint opnum_target TSRMLS_DC);
Expressions are compiled using the zend_compile_expr
function, which accepts a result
znode and an ast
. This function merely dispatches to a more concrete compilation function based on the AST kind:
switch (ast->kind) { /* ... */ case ZEND_AST_ASSIGN_OP: zend_compile_compound_assign(result, ast TSRMLS_CC); return; case ZEND_AST_BINARY_OP: zend_compile_binary_op(result, ast TSRMLS_CC); return; /* ... */ }
The zend_compile_binary_op
function referenced above is defined as follows (compile-time pre-evaluation code has been removed for conciseness):
void zend_compile_binary_op(znode *result, zend_ast *ast TSRMLS_DC) { zend_ast *left_ast = ast->child[0]; zend_ast *right_ast = ast->child[1]; zend_uint opcode = ast->attr; znode left_node, right_node; zend_compile_expr(&left_node, left_ast TSRMLS_CC); zend_compile_expr(&right_node, right_ast TSRMLS_CC); zend_emit_op_tmp(result, opcode, &left_node, &right_node TSRMLS_CC); }
By convention, all zend_compile_*
functions start by extracting child nodes into named local variables with an _ast
suffix. Afterwards the function compiles the child nodes into znode
s (which have the same name as the corresponding AST nodes, but using a _node
suffix). Finally an opline is emitted, which has a TMP_VAR result and uses the two znodes as operands.
When zend_compile_expr
is passed the AST of a variable expression, it will compile it using the BP_VAR_R
fetch mode. To explicitly specify the fetch mode the function zend_compile_var
can be used instead, which accepts a BP_VAR_*
mode as the last argument:
void zend_compile_pre_incdec(znode *result, zend_ast *ast TSRMLS_DC) { zend_ast *var_ast = ast->child[0]; ZEND_ASSERT(ast->kind == ZEND_AST_PRE_INC || ast->kind == ZEND_AST_PRE_DEC); if (var_ast->kind == ZEND_AST_PROP) { zend_op *opline = zend_compile_prop_common(result, var_ast, BP_VAR_RW TSRMLS_CC); opline->opcode = ast->kind == ZEND_AST_PRE_INC ? ZEND_PRE_INC_OBJ : ZEND_PRE_DEC_OBJ; } else { znode var_node; zend_compile_var(&var_node, var_ast, BP_VAR_RW TSRMLS_CC); zend_emit_op(result, ast->kind == ZEND_AST_PRE_INC ? ZEND_PRE_INC : ZEND_PRE_DEC, &var_node, NULL TSRMLS_CC); } }
The preceding code sample also illustrates another pattern that sometimes occurs when dealing with operations on variables: A property increment has the same structure as a property fetch, just using a different opcode. Such situations are handled by providing a _common
variant of variable-compilation functions, which return the generated opline for further adjustments.
Statements are compiled using the zend_compile_stmt
function, which accepts an ast
(but no longer a result znode, as statements have no result). Once again this function will delegate to more specific compilation functions based on the AST kind. For example, this is the compilation function for a while
loop:
void zend_compile_while(zend_ast *ast TSRMLS_DC) { zend_ast *cond_ast = ast->child[0]; zend_ast *stmt_ast = ast->child[1]; znode cond_node; zend_uint opnum_start, opnum_jmpz; opnum_start = get_next_op_number(CG(active_op_array)); zend_compile_expr(&cond_node, cond_ast TSRMLS_CC); opnum_jmpz = zend_emit_cond_jump(ZEND_JMPZ, &cond_node, 0 TSRMLS_CC); zend_begin_loop(TSRMLS_C); zend_compile_stmt(stmt_ast TSRMLS_CC); zend_emit_jump(opnum_start TSRMLS_CC); zend_update_jump_target_to_next(opnum_jmpz TSRMLS_CC); zend_end_loop(opnum_start, 0 TSRMLS_CC); }
After extracting the child nodes into local variables, this function compiles the loop condition and generates a conditional jump of type JMPZ
based on it. Both the opnum of the condition and of the JMPZ instructions are backed up.
Afterwards the body of the loop is compiled (stmt_ast
) and a jump back to the condition generated. Lastly the jump target of the JMPZ instruction is updated to point after the loop (i.e. if the condition is false, execution should continue after the loop.)
The zend_begin_loop
and zend_end_loop
functions store information for break/continue and try/catch.
The generated AST can be exposed to userland via an extension, for use by static analysers. This should be relatively easy to implement and we might even want to provide this as a bundled extension (like ext/tokenizer).
More interestingly, we could allow extensions to hook into the compilation process (the current AST implementation does not provide hooks, but they can be added if we want them). This would allow extensions to implement some types of “language features”.
As an example, this is roughly how an implementation of the ifsetor RFC could look like using an AST hook:
/* Works by rewriting ifsetor($foo, 'bar') to isset($foo) ? $foo : 'bar' */ void ext_ifsetor_hook(zend_ast **ast_ptr TSRMLS_DC) { zend_ast *ast = *ast_ptr; if (ast->kind == ZEND_AST_CALL && ast->child[0]->kind == ZEND_AST_ZVAL) { zend_string *name = zval_get_string(zend_ast_get_zval(ast->child[0])); zend_ast_list *args = zend_ast_get_list(ast->child[1]); if (zend_str_equals_literal_ci(name, "ifsetor") && args->children == 2 && !zend_args_contain_unpack(args) ) { if (!zend_is_variable(args->child[0])) { zend_error_noreturn(E_COMPILE_ERROR, "First argument of ifsetor " "must be a variable"); } /* Note: One would need a function for adding refs to args->child[0] here, * as it is used two times - as written here it won't work correctly. */ *ast_ptr = zend_ast_create(ZEND_AST_CONDITIONAL, zend_ast_create(ZEND_AST_ISSET, args->child[0]), args->child[0], args->child[1] ); } STR_RELEASE(name); } }
I don't know how useful this is and how many things can be implemented in this way, but I think it's worth considering.
An additional possibility is to drop the keywords for isset
and empty
and just compile them as special function calls (using similar checks as the code above). Maybe other keywords can be dropped as well.
The AST implementation can be found at https://github.com/nikic/php-src/tree/ast. Some quick links to the most important files:
zend_compile.c
(The code relevant to the AST begins somewhere around line 3200, everything before that is largely untouched.)The branch already includes the Uniform Variable Syntax RFC, as it was a necessary prerequisite for the implementation.
The implementation has everything ported, but probably still has some bugs and needs some cleanup :)
The vote started on 2014-08-18 and ended on 2014-08-25. The necessary 2/3 majority was reached, as such the RFC is accepted.