This is just an intial merge. It does not yet make generators and finally
work together.
Conflicts:
Zend/zend_language_scanner.c
Zend/zend_language_scanner_defs.h
Zend/zend_vm_def.h
Zend/zend_vm_execute.h
Zend/zend_vm_execute.skl
Zend/zend_vm_opcodes.h
Previously only an error was thrown when return occured after yield. Also
returns before the first yield would fail for by-ref generators.
Now the error message is handled in pass_two, so all returns are checked.
. zend_function.pass_rest_by_reference is replaced by
ZEND_ACC_PASS_REST_BY_REFERENCE in zend_function.fn_flags
. zend_function.return_reference is replaced by ZEND_ACC_RETURN_REFERENCE
in zend_function.fn_flags
. zend_arg_info.required_num_args removed. it was needed only for internal
functions. Now the first arg_info for internal function (which has special
meaning) is represented by zend_internal_function_info structure.
. zend_op_array.size, size_var, size_literal, current_brk_cont,
backpatch_count moved into CG(context), because they are used only during
compilation.
. zend_op_array.start_op is moved into EG(start_op), because it's used
only for 'interactive' execution of single top-level op-array.
. zend_op_array.done_pass_two is replaced by ZEND_ACC_DONE_PASS_TWO in
zend_op_array.fn_flags.
. op_array.vars array is trimmed (reallocated) during pass_two.
. zend_class_entry.constants_updated is replaced by
ZEND_ACC_CONSTANTS_UPDATED in zend_class_entry.ce_flags
. the size of zend_class_entry is reduced by sharing the same memory space
by different information for internal and user classes.
See zend_class_inttry.info union.
# - renamed php_runkit_function_copy_ctor to _duplicate_function
# REM: runkit does not compile with trunk at the moment, fixing it would introduce even more version #ifs, don't know what the best way is to fix it
# - extracted traits related stuff from destroy_zend_class into _destroy_zend_class_traits_info
# - need to investigate implementation/handling of internal classes further before enabling internal traits
# RFC http://wiki.php.net/rfc/horizontalreuse#traits_-_reuse_of_behavior
# Ok, here we go, I guess that will result in more discussion, which is fine
# by me. But now, the patch is here, and properly archived.
#
# See below a list of notes to the patch, it also includes a list of
# points which should be fixed
#
# Internals of the Traits Patch
# -----------------------------
#
# Open TODOs
# """"""""""
#
# - Reflection API
# - support for traits for internal classes
# - currently destroy_zend_class does not handle that case
#
# Introduced Structures
# """""""""""""""""""""
#
# Data structures to encode the composition information specified in the
# source:
# - zend_trait_method_reference
# - zend_trait_precedence
# - zend_trait_alias
#
# Changes
# """""""
#
# zend_class_entry
# - uses NULL terminated lists of pointers for
# - trait_aliases
# - trait_precedences
# - do you prefer an explicit counter?
# - the information is only necessary during class composition
# but might be interesting for reflection
# - did not want to blow up class further with not really necessary length counters
#
# added keywords
# - trait
# - insteadof
#
# Added opcodes
# ZEND_ADD_TRAIT
# - similar to ZEND_ADD_INTERFACE
# - adds the trait to the list of traits of a class, no actual composition done
# ZEND_BIND_TRAITS
# - emitted in zend_do_end_class_declaration
# - concludes the class definition and will initiate the trait composition
# when the class definition is encountered during runtime
#
# Added Flags
# ZEND_ACC_TRAIT = 0x120
# ZEND_ACC_IMPLEMENT_TRAITS = 0x400000
# ZEND_FETCH_CLASS_TRAIT = 14
#
# zend_vm_execute.h
# - not sure whether the handler initialization (ZEND_ADD_TRAIT_SPEC_HANDLER,
# ZEND_BIND_TRAITS_SPEC_HANDLER) is correct, maybe it should be more selective
#
# zend_compile.c
# - refactored do_inherit_method_check
# split into do_inherit_method_check and do_inheritance_check_on_method
# - added helper functions use a '_' as prefix and are not mentioned in the
# headers
# - _copy_functions
# prepare hash-maps of functions which should be merged into a class
# here the aliases are handled
# - _merge_functions
# builds a hash-table of the methods which need to be added to a class
# does the conflict detection
# - reused php_runkit_function_copy_ctor
# - it is not identical with the original code anymore, needed to update it
# think I fixed some bugs, not sure whether all have been reported back to runkit
# - has to be renamed, left the name for the moment, to make its origin obvious
# - here might be optimization potential
# - not sure whether everything needs to be copied
# - copying the literals might be broken
# - added it since the literals array is freed by efree and gave problems
# with doubled frees
# - all immutable parts of the zend_op array should not be copied
# - am not sure which parts are immutable
# - and not sure how to avoid doubled frees on the same arrays on shutdown
# - _merge_functions_to_class
# does the final merging with the target class to handle inherited
# and overridden methods
# - small helper for NULL terminated lists
# zend_init_list, zend_add_to_list
#
# zend_language_parser.y
# - reused class definition for traits
# - there should be something with regard to properties
# - if they get explicitly defined, it might be worthwhile to
# check that there are no collisions with other traits in a composition
# (however, I would not introduce elaborate language features to control that
# but a notice for such conflicts might be nice to the developers)
The following pseudo-code explains how it should be used in opcode cache.
function cache_compile_file($filename) {
if (!is_cached($filename)) {
...
orig_compiler_options = CG(compiler_optins);
CG(compiler_options) |= ZEND_COMPILE_IGNORE_INTERNAL_CLASSES |
ZEND_COMPILE_DELAYED_BINDING;
$op_array = orig_compile_file($filename);
CG(compiler_options) = orig_copiler_options;
...
} else {
$op_array = restore_from_cache($filename);
}
zend_do_delayed_early_binding($op_array);
}
- Extensions which delete global variables need to use new special function
- delete_global_variable() (I'm about to rename it) to remove them.
- Will post to internals@ or via commit messages if there's anything else.
a) We specialize opcodes according to op_type fields. Each opcode has to
be marked with which op_type's it uses.
b) We support different execution methods. Function handlers, switch()
and goto dispatching. goto seems to be the fastest but it really
depends on the compiler and how well it optimizes. I suggest playing
around with optimization flags.
- Warning: Things might break so keep us posted on how things are going.
(Dmitry, Andi)
implementation, and allows exceptions to 'fire' much earlier than before.
Instructions on how to use the new mechanism will follow on internals@
shortly...
Note - this (most probably) breaks the current implementation of
set_exception_handler()
including:
- Whether or not to pass by ref (replaces the old arg_types, with arg_info)
- Argument name (for future use, maybe introspection)
- Class/Interface name (for type hints)
- If a class/interface name is available, whether to allow a null instance
Both user and builtin functions share the same data structures.
To declare a builtin function that expects its first arg to be an instance
of class 'Person', its second argument as a regular arg, and its third by
reference, use:
ZEND_BEGIN_ARG_INFO(my_func_arg_info, 0)
ZEND_ARG_OBJ_INFO(0, someone, Person, 1)
ZEND_ARG_PASS_INFO(0)
ZEND_ARG_PASS_INFO(1)
ZEND_END_ARG_INFO();
and use my_func_arg_info as the arg_info parameter to the ZEND_FE() family
of macros.
The first arg to each ZEND_ARG_*() macro is whether or not to pass by ref.
The boolean arg to ZEND_BEGIN_ARG_INFO() tells the engine whether to treat
the arguments for which there's no explicit information as pass by reference
or not.
The boolean argument to ZEND_ARG_OBJ_INFO() (4th arg) is whether or not to allownull values.
- The fields of zend_namespace were not completely initialized which
led to a variety of problems.
- The occurrence of class/interface/namespace definition is now
captured.
- Functions/classes/interfaces/namespaces can be preceded by doc
comments which are stored for use by extensions.
1. Nested classes are gone.
2. New syntax for namespaces:
namespace foo {
class X { ... }
function bar { ... }
var x = 1;
const ZZ = 2;
}
3. Namespaced symbol access: $x = new foo::X; - etc.
For now, namespaces are case insensitive, just like classes.
Also, there can be no global class and namespace with the same name
(to avoid ambiguities in :: resolution).
We need separate cleanup stage because of the following problem:
Suppose we destroy class X, which destroys function table,
and in function table we have function foo() that has static $bar. Now if
object of class X was assigned to $bar, its destructor will be called and will
fail since X's function table is in mid-destruction.
So we want first of all to clean up all data and then move to tables
destruction.
Note that only run-time accessed data need to be cleaned up, pre-defined
data can not contain objects and thus are not probelmatic.
# Looks like we are having a lots of pain in the various parts of the body
# because of the destructors...
- some fixes by me).
- You can't access protected variables from outside the object. If you want
- to see a protected member from your ancestors you need to declare the
- member as protected in the class you want to use it in. You can't
- redeclare a protected variable as private nor the other way around.
- destructor could be run after its class was already dead. Right now
- object destructors is the first thing whic happens during shutdown in
- order to prevent this problem. It's very likely that destructors will
- cause more grief and we'll have to outline exactly when you should use
- them and what kind of logic you're allowed to do inside of them.
- This bug was reported by sebastian.
Advantages:
- Smaller memory footprint for the op arrays
- Slightly faster compilation times (due to saved erealloc() calls and faster zend_op
initialization)
- include_once() & require_once() share the same file list
- Consistency between include() and require() - this mostly means that return()
works inside require()'d files just as it does in include() files (it used to
be meaningless in require()'d files, most of the time (see below))
- Made require() consistent with itself. Before, if the argument was not a constant
string, require() took the include() behavior (with return()).
- Removed lots of duplicate code.
Bottom line - require() and include() are very similar now; require() is simply an include()
which isn't allowed to fail. Due to the erealloc() calls for large op arrays, require()
didn't end up being any faster than include() in the Zend engine.
- In function declaration instead of the return statement
- In the assignment phase
- Implement ability to turn off support for call-time pass by reference
Later on we will want to delay the write fetches even longer until after their
resulting expression is parsed. The way it is now, will make it very easy
to delay as long as we need.
* updated alloc_persist to use critical sections
* changed extension shutdown to two-phase
* updated dependencies
* PR support (don't remember if there was any really)
