Normally, we add classes without parents (and no interfaces or traits) directly
to the class map, early binding the class. However, if the same class has
already been registered, we would instead just add a ZEND_DECLARE_CLASS
instruction and let the handler throw a duplicate class declaration exception.
However, with opcache, if on the next request the files are included in the
opposite order, we won't perform early binding. To fix this, create a
ZEND_DECLARE_CLASS_DELAYED instruction instead and handle classes without
parents accordingly, skipping any linking for classes that are already linked in
delayed early binding.
Fixes GH-8846
If we bind the class to the runtime slot even if we're not the ones who have
performed early binding we'll miss the redeclaration error in the
ZEND_DECLARE_CLASS_DELAYED handler.
Closes GH-11226
When benchmarking the file cache of opcache on index.php from a dummy
WordPress install, I noticed that 36.42% of the time was spent in
zend_adler32 to verify the checksums of the files. Callgrind reported
that 332,731,216 instructions were executed during that run and average
time to execute the index file was around 91ms.
This patch implements an SSE2 accelerated version of zend_adler32, which
reduces the number of instructions executed on that bench to
248,600,983, which is a reduction of ~25%. There is also a decrease in
wallclock time measurable: around 10ms. Now only 16.05% of the time is
spent computing checksums.
The benchmark tests were performed using Callgrind, and time for the
wallclock time. These tests were executed multiple times and their
results were averaged. The WordPress install only contains two
almost-blank posts.
Previously, notify would only get called on classes with CE_CACHE.
However, during compilation class names are non-permanent which won't
get a CE_CACHE and thus wouldn't not get notified.
Closes GH-9550
It's possible for delayed early binding opcodes to get optimized
away if they are "unreachable". However, we still need to attempt
early binding for them. (In some cases we also corrupt the early
binding list outright during optimization, which is how I got here.)
Fix this by storing information about delayed early binding
independently of DECLARE_CLASS_DELAYED opcodes, so early binding is
performed even after the opcode has been dropped.
This can occur on opcache OOM conditions, where the function/class
names are not interned and the script does not get cached. In
that case the functions/classes get transferred from the persistent
script to the global tables, without incrementing the key refcount.
To mirror that, we should also not try to free the keys when freeing
the persistent script. For this by setting the number of elements
to zero, which will free only the hashtable structure itself.
Convert zend_hash_find_ex(..., 1) to zend_hash_find_known_hash(...)
Convert zend_hash_find_ex(..., 0) to zend_hash_find(...)
Also add serializable changes to UPGRADING.INTERNALS summary
1. Update: http://www.php.net/license/3_01.txt to https, as there is anyway server header "Location:" to https.
2. Update few license 3.0 to 3.01 as 3.0 states "php 5.1.1, 4.1.1, and earlier".
3. In some license comments is "at through the world-wide-web" while most is without "at", so deleted.
4. fixed indentation in some files before |
Currently, dynamically declared functions and closures are inserted
into the function table under a runtime definition key, and then later
possibly renamed. When opcache is not used and a file containing a
closure is repeatedly included, this leads to a very large memory leak,
as the no longer needed closure declarations will never be freed
(https://bugs.php.net/bug.php?id=76982).
With this patch, dynamic functions are instead stored in a
dynamic_func_defs member on the op_array, which opcodes reference
by index. When the parent op_array is destroyed, the dynamic_func_defs
it contains are also destroyed (unless they are stilled used elsewhere,
e.g. because they have been bound, or are used by a live closure). This
resolves the fundamental part of the leak, though doesn't completely
fix it yet due to some arena allocations.
The main non-obvious change here is to static variable handling:
We can't destroy static_variables_ptr in destroy_op_array, as e.g.
that would clear the static variables in a dynamic function when
the op_array containing it is destroyed. Static variable destruction
is separated out for this reason (we already do static variable
destruction separately for normal functions, so we only need to
handle main scripts).
Closes GH-5595.
When a method is inherited, the static variables will now always
use the initial values, rather than the values at the time of
inheritance. As such, behavior no longer depends on whether
inheritance happens before or after a method has been called.
This is implemented by always keeping static_variables as the
original values, and static_variables_ptr as the modified copy.
Closes GH-6705.
This is a new transparent technology that eliminates overhead of PHP class inheritance.
PHP classes are compiled and cached (by opcahce) separately, however their "linking" was done at run-time - on each request. The process of "linking" may involve a number of compatibility checks and borrowing methods/properties/constants form parent and traits. This takes significant time, but the result is the same on each request.
Inheritance Cache performs "linking" for unique set of all the depending classes (parent, interfaces, traits, property types, method types involved into compatibility checks) once and stores result in opcache shared memory. As a part of the this patch, I removed limitations for immutable classes (unresolved constants, typed properties and covariant type checks). So now all classes stored in opcache are "immutable". They may be lazily loaded into process memory, if necessary, but this usually occurs just once (on first linking).
The patch shows 8% improvement on Symphony "Hello World" app.
- ZCG(bind_hash) is not used anymore
- zend_accel_function_hash_copy() and zend_accel_function_hash_copy_from_shm() are the same
- zend_accel_class_hash_copy() and zend_accel_class_hash_copy_from_shm() are almost the same
Instead of having a completely independent encoding for type list
entries. This is going to use more memory, but I'm not particularly
concerned about that, as type unions that contain multiple classes
should be uncommon. On the other hand, this allows us to treat
top-level types and types inside lists mostly the same.
A new ZEND_TYPE_FOREACH macros allows to transparently treat list
and non-list types the same way. I'm not using it everywhere it could be
used for now, just the places that seemed most obvious.
Of course, this will make any future type system changes much simpler,
as it will not be necessary to duplicate all logic two times.
Also generate a fatal error if a collision occurs in zend_compile.
This is not perfect, because collisions might still be introduced
via opcache, if one file is included multiple times during a request,
invalidate in the meantime and recompiled by different processes.
This still needs to be addressed, but this patch fixes the much
more common case of collisions occuring when opcache is not used.
Fixes bug #78903.
According to RFC: https://wiki.php.net/rfc/union_types_v2
The type representation now makes use of both the pointer payload
and the type mask at the same time. Additionall, zend_type_list is
introduced as a new kind of pointer payload, which is used to store
multiple class types. Each of the class types is a tagged pointer,
which may be either a class name or class entry. The latter is only
used for typed properties, while arguments/returns will instead use
cache slots. A type list can contain a mix of both names and CEs at
the same time, as not all classes may be resolvable.
One thing this is missing is support for union types in arginfo
and stubs, which I want to handle separately.
I've also dropped the special object code from the JIT implementation
for now -- I plan to add this back in a different form at a later time.
For now I did not want to include non-trivial JIT changes together
with large functional changes.
Another possible piece of follow-up work is to implement "iterable"
as an internal alias for "array|Traversable". I believe this will
eliminate quite a few special-cases that had to be implemented.
Closes GH-4838.
We now store the pointer payload and the type mask separately. This
is in preparation for union types, where we will be using both at
the same time.
To avoid increasing the size of arginfo structures, the
pass_by_reference and is_variadic fields are now stored as part of
the type_mask (8-bit are reserved for custom use).
Different types of pointer payloads are distinguished based on bits
in the type_mask.
Assigning to an uninitialized typed property will no longer trigger
a call to __set(). However, calls to __set() are still triggered if
the property is explicitly unset().
This gives us both the behavior people generally expect, and still
allows ORMs to do lazy initialization by unsetting properties.
For PHP 8, we should fine a way to forbid unsetting of declared
properties entirely, and provide a different way to achieve lazy
initialization.
The AC_CHECK_TYPE was refactored in more recent versions of Autoconf
and the call with two arguments is obsolete and not recommended anymore.
This patch also refactors some leftovers of using ulong and uint which
are not standard nor common usages of types in C.
The ulong can be used as zend_ulong and uint usage is actually
`unsigned int`.
The usage of HAVE_ULONG removed since it is not used in current code
base.
Legacy edgecase for some legacy HPUX systems removed:
- sys/stream.h header is not checked and the HAVE_SYS_STREAM_H is
not defined with current build system.
- flags are unsigned int
- max_allowed_packet changed to unsigned int
