Remove deprecated slice() and kvzip() methods

[hiphop-php.git] / hphp / doc / bytecode.specification

**********************************
* HipHop Bytecode v1 revision 18 *
**********************************


Introduction
------------

HipHop bytecode (HHBC) v1 is intended to serve as the conceptual basis for
encoding the semantic meaning of HipHop source code into a format that is
appropriate for consumption by interpreters and just-in-time compilers. By
using simpler constructs to encode more complex expressions and statements,
HHBC makes it straightforward for an interpreter or a compiler to determine
the order of execution for a program.

HHBC was designed with several competing goals in mind:

1) Run-time efficiency. The design of HHBC should be congruous to implementing
an efficient execution engine, whether it be an interpreter or a just-in-time
compiler.

2) PHP 5.5 compatibility. It should be possible to compile valid PHP 5.5 source
code into HipHop bytecode in a way that preserves the semantic meaning of the
source.

3) Simplicity. The design of HHBC should avoid features that could be removed
or simplified without compromising PHP 5.5 compatibility, run-time efficiency,
or design cleanliness.


Compilation units
-----------------

Each HipHop source file is compiled into a separate "compilation unit", or
"unit" for short. Units are composed of bytecode and metadata.

A unit's bytecode is an array of bytes encoding a sequence of HHBC
instructions, where each instruction is encoded using one or more bytes. This
specification defines an instruction set and defines the behavior of each HHBC
instruction, but the exact byte values used to encode HHBC instructions is
currently unspecified.

A unit's metadata is a set of structures that provide essential information
that is needed at run time by the execution engine. This specification will
describe a unit's metadata as a set of named tables with ordered rows, but the
exact format of the metadata is currently unspecified.

Each instruction in a unit's bytecode can be referred to using a "bytecode
offset", which is the distance in bytes from the first byte of a unit's
bytecode to the first byte of the instruction.

A unit's bytecode is partitioned into sections called "functions". The unit's
metadata uses bytecode offsets to specify which instructions belong to which
functions.

When a unit is loaded at run time, the execution engine assigns the unit's
bytecode a logical range of addresses called "bytecode addresses". An
instruction is referred to at run time using its bytecode address.


Flow of execution
-----------------

HipHop bytecode models the flow of execution using a stack of frames referred
to as the "call stack". A "frame" is a structure that logically consists of a
header, a program counter (PC), a local variable store, an iterator variable
store, an evaluation stack, and a function parameter info (FPI) stack.

The frame at the top of the call stack is referred to as the "current frame".
The current frame represents the function that is currently executing. The
program counter (PC) of the current frame is referred to as the "current PC".
At any given time, the current PC holds the bytecode address of the current
instruction to execute. When the execution engine executes an instruction, the
current PC is updated to point to the next instruction. By default, the current
PC is updated to point to the byte that sequentially follows the last byte of
the current instruction in the bytecode. Some instructions override the default
behavior and explicitly update the current PC in a specific way.

HHBC provides special instructions to allow for calling a function and
returning from a function. When a function is called, a new frame is pushed
onto the call stack, and the PC of the new frame is initialized to the
appropriate entry point (typically the instruction of the function that is
sequentially first in the bytecode). The new frame becomes the current frame,
and the PC of the new frame becomes the current PC. When a function returns,
the current frame is popped off the call stack. The previous frame becomes the
current frame, and its PC becomes the current PC. The facility provided by the
execution engine that is responsible for handling function calls and returns is
called the "dispatcher".

Typically, a frame is removed from the call stack when its corresponding
function returns. However, a frame may be removed from the call stack before
its corresponding function returns in the course of processing an exception.
The facility provided by the execution engine that is responsible for
processing exceptions is called the "unwinder".


Values
------

HHBC instructions may push and pop values on the current frame's evaluation
stack and they may read and write values to the current frame's local
variables. Values come in three flavors: cells, refs, and classrefs.

A "cell" is a structure that contains a type identifier and either data (for
non-refcounted types) or a pointer to data (for refcounted types). When a cell
containing a pointer is duplicated, the new cell will point to the same data as
the original cell. When a cell containing a pointer is duplicated or discarded,
the execution engine is responsible for honoring the data's refcount logic.

A "ref" is a structure that contains a pointer to a cell container. When a ref
is duplicated, the new ref will point to the same container as the original
ref. When a ref is duplicated or destroyed, the execution engine is responsible
for honoring the container's refcount logic. When the container is destroyed,
the cell it contains is also destroyed.

A "classref" is a structure that contains a reference to a class. When a
classref is pushed onto the stack or popped of the stack, no refcounting is
required.

Values on the evaluation stack may be any of the three flavors listed above.
Values stored in local variables may only be cells or refs.


Functions
---------

A unit's bytecode is organized into functions. Each function has its own
metadata that provides essential information about the function, such as the
name of the function, how many local variables it has, how many iterator
variables it has, how many formal parameters it has, the names of the local
variables, the names of the formal parameters, how each parameter should be
passed (pass by value vs. pass by reference), whether each parameter has a
default value, and an upper bound for the maximum depth the evaluation stack
can reach at run time.

Each local variable and iterator variable has an id, and HHBC instructions can
reference these variables using these ids. The id space for local variables is
distinct from the id space for iterator variables. Thus local id 1 refers to a
different variable than iterator id 1. Local variable ids and iterator ids are
signed 32-bit integer values. No function may have more than 2^31 - 1 local
variables, and no function may have more than 2^31 - 1 iterator variables.

Some local variables have names associated with them (called "named local
variables"), while other local variables do not have names associated with them
(called "unnamed local variables"). All local variables that reference formally
declared parameters have names associated with them. Iterator variables do not
have names associated with them. Variables that have a name associated with
them will appear in the current variable environment (if they are defined),
while variables that do not have a name associated with them will never appear
in the current variable environment.

Formally declared parameters are considered to be local variables. Given a
function with n formally declared parameters, local ids 0 through n-1 will be
used to reference the formally declared parameters. Formal parameters without
default values are called "required parameters", while formal parameters with
default values are called "optional parameters".

The bytecode for each function is partitioned into a primary function body and
0 or more fault funclets. The metadata for each function specifies how many
fault funclets the function has. For the primary function body, the metadata
specifies a set of non-overlapping ranges of bytecode that compose the primary
function body, and it specifies the main entry point and 0 or more DV entry
points (entry points are discussed in more detail in the "Entry points"
section). For each fault funclet, the metadata specifies a set of non-
overlapping ranges of bytecode that compose the fault funclet body, and it
specifies an entry point for the fault funclet. The primary function body and
the fault funclets may not overlap with each other, and the union of the
primary function body and the fault funclets must cover all of the function's
bytecode. Fault funclets are discussed in more detail in the "Exception handler
(EH) table" and "Processing exceptions" sections.

The total size of the bytecode for the primary function body and all the fault
funclets must not exceed 2^31 - 1 bytes. The bytecode for a function must be
one contiguous range of bytecode.

Each function's metadata provides a "line number table" to allow mapping
bytecode offsets back to source line numbers. Each row in the line number table
consists of a source line number and a range of bytecode. The table is sorted
by starting bytecode offset, lowest offset first. The bytecode offset of the
beginning of each instruction in the function must belong to exactly one of the
ranges of bytecode in the line number table.


Classes
-------

Functions may be grouped into metadata for classes. Class metadata objects are
used to describe several PHP-level language features including traits,
interfaces, closures, and (of course) classes.

Class metadata includes information about the properties on the class, special
functions on the class such as constructors or internal property initialization
routines (86sinit, 86pinit), class constants, list of used traits, list of
extended classes, list of implemented interfaces, etc.

Classes also include a flag indicating their "hoistability". For now this isn't
documented much here. See class.h.


Closures
--------

Closures are implemented in hhbc as subclassses of Closure, in conjunction with
the CreateCl opcode. It is legal hhbc to create other subclasses of Closure (to
represent user code that attempts to do the same), but attempting to
instantiate one will result in a fatal error. The documentation of the CreateCl
opcode below lists the requirements for a closure subclass to be usable with
it.


Continuations
-------------

The basic compilation strategy for continuations is to create bytecode functions
consisting of two parts.

The first part, executed when the continuation function is called, must consist
of a CreateCont opcode followed by a RetC, which are responsible for suspending 
execution state into a new Continuation object (includes resume offset pointing 
to the start of the second part of the function) and returning it back to 
the caller.

The second part is where the real user-level code of the generator should be
placed. ContEnter and ContRaise opcodes used in Continuation's next(), send()
and raise() methods resume execution and transfer control to the resume offset
stored in the Continuation object. The user-level code yields values using
ContSuspend and ContSuspendK opcodes and returns using ContRetC opcode.


Async functions
---------------

Similarly to continuations, async functions are implemented as bytecode
functions consisting of two parts.

The first part represents "eager execution" of the async function, and the
second part is used when "suspended execution" is necessary (i.e. if it does
an await on something that is not yet available). Both parts are required to
'have the same effects', except the latter one works in a suspended mode.

The first part uses AsyncESuspend opcode to suspend execution state into a new
AsyncFunctionWaitHandle object. It includes the resume offset pointing to the
second part of the function. The second part suspends using ContSuspend opcode
and returns using ContRetC opcode.

The async function implementation is still changing and the implementation may
change significantly, so this spec is staying light on details for now.


Entry points
------------

Entry points come in four varieties: the main entry point, DV entry points,
fault entry points, and catch entry points.

Every function has exactly one main entry point. When a function is called, the
dispatcher will set the PC of the new frame to point to the main entry point if
either (1) the function does not have any optional parameters or (2) the caller
provides values for all of the optional parameters.

DV entry points are normally used to handle initializing optional parameters
that the caller did not provide. Generally the DV entries contain blocks that
initialize parameters, and then fall through directly into one another, with
the last block ending with a jump to the main entry point. This is not a
requirement, however. The dispatcher selects the appropriate DV entry point
based on the number of arguments passed into the function.

The main entry point and DV entry points are used by the dispatcher when
handling a function call. Each function's metadata provides an "entry point
table". Each row in the entry point table consists of a number of arguments and
the bytecode offset of the entry point that should be used by the dispatcher
(either the main entry point or a DV entry point).

Fault entry points are used by the unwinder to enter fault funclets as
appropriate to perform necessary cleanup when a region of code exits abnormally
through an exception. For a function with N fault funclets there are exactly N
fault entry points (one fault entry point per fault funclet). The bytecode
offset of a fault entry point must be inside its corresponding fault funclet.

Catch entry points are used by the unwinder to resume normal execution once a
matching "catch" block has been found and all the necessary cleanup has been
performed.

More details about the unwinder, fault funclets, fault entry points, catch
entry points can be found in the "Exception handler (EH) table" and "Processing
exceptions" sections.


Unit metadata
-------------

Every compilation unit has a litstr table, a scalar array table, a function
table, and a class table.

The litstr table maps litstr ids to literal strings. Bytecodes that refer to
literal strings do so by litstr id. Litstr ids are signed 32-bit integer
values, which must be between 0 and 2^31 - 2 inclusive. In addition to the
per-unit litstr tables, a global table is built when generating an
"authoritative" repo (one in which all the PHP code is known at bytecode
generation time, and is guaranteed not to change). Global litstr ids can be
used in any unit, and are encoded in the range [2^30..2^31-2].

The scalar array table maps scalar array ids to a description of the contents
of a scalar array. An array is a scalar array if and only if each element of
the array is a null, boolean, integer, double, string, or a scalar array.
Furthermore, each element of a scalar array must be a cell. Finally, scalar
arrays may not recurse infinitely. Each scalar array id must be between 0 and
2^31 - 2 inclusive.

Each row in the function table contains a unique function id, a function name
specified by a litstr id, the bytecode offset for the corresponding function, a
flag that indicates if the function is unconditionally declared in the
outermost scope, and the function metadata. Note that there may be multiple
rows in the function table with same function name. However, there may not be
multiple rows that are marked as being unconditionally declared in the
outermost scope with the same function name. Each function id must be between 0
and 2^31 - 2 inclusive.

Each row in the class table contains a unique class id, a class name specified
by a litstr id, a flag that indicates if the class declaration is hoisted to
the prelude of pseudo-main, and the class metadata. Note that there may be
multiple rows in the class table with same class name. However, there may not
be multiple rows that are marked as being hoisted with the same class name.
Each class id must be between 0 and 2^31 - 2 inclusive.


Function parameter info (FPI) structures and the FPI stack
----------------------------------------------------------

Every function has a function parameter info (FPI) structure associated with it
that can be retrieved at run time. The FPI structure contains the bytecode
address of the function, the number of parameters the function has, and a
parameter table that indicates whether each parameter is pass by value or pass
by reference.

In addition to the evaluation stack, each frame also contains another stack
called the FPI stack. Each entry on the FPI stack consists of a reference to an
FPI structure and a bytecode address of entry point into the corresponding
function. The entry on the top of the FPI stack is called the "current FPI".

The FPush* instructions push a new entry onto the FPI stack, initializing the
entry with a reference to the FPI structure for a given function and the
bytecode address of the appropriate entry point. The FPass* instructions
prepare the parameters that will be passed into the callee. The FCall*
instructions look at the current FPI to get the bytecode address of the
function (the callee), transfers the parameters from the evaluation stack to
the callee, pops the current FPI off of the FPI stack, and then invokes the
dispatcher to call the function.

Calls to builtin functions may be optimized to avoid pushing an entry on the
FPI stack if it is known that the builtin function does not need access to the
call stack. In this case, the arguments to the builtin are pushed on stack as
Cells and Vars, and the builtin can be invoked with the FCallBuiltin
instruction. Unless otherwise noted, subsequent references to FCall*
instructions should be meant to refer to non-optimized FCall instructions, i.e.
all FCall instructions other than FCallBuiltin.


Calling convention
------------------

The caller may pass any number of parameters to the callee by executing FPass*
instructions zero or more times prior to executing an FCall* instruction. The
caller must pass the parameters in forward order, i.e. the first use of FPass*
passes the first parameter, the second use of FPass* passes the second
parameter, and so forth.

The FPush*/FPass*/FCall* instructions can be used to call a global function, a
method on an object, or a method from a class. The caller is responsible for
evaluating all of the parameters in forward order. When the caller executes an
FCall* instruction, the dispatcher creates a new frame and moves the parameters
prepared by the caller into the callee's variable environment. The dispatcher
then transfers control to the appropriate entry point of the callee (either the
main entry point or a DV entry point) based on the number of parameters passed.

When the callee executes the Ret* instruction, the dispatcher pushes the return
value onto the caller's evaluation stack. Then the dispatcher destroys the
callee's frame and transfers control back to the caller.


Exception handler (EH) table
----------------------------

The metadata for each function provides an "exception handler (EH) table".
Each row in the EH table (called an "EH entry") consists of a kind ("fault" or
"catch"), a non-negative integer "region depth", a set of non-overlapping
ranges of bytecode that compose the "protected region", and an offset of a
fault funclet (if it's a "fault" EH entry) or a list of [class name, catch
entry point] pairs (if it's a "catch" EH entry).

Each range of bytecode is given by a starting offset and an ending offset,
where the starting offset is the bytecode offset of the first byte of the first
instruction in the range and the ending offset is the bytecode offset after the
last byte of the last instruction in the range.

Note that two or more EH entries may refer to the same fault funclet or the
same catch entry points. Regardless of whether multiple EH entries share the
same fault funclet or the same catch entry points, each EH entry in the EH
table will be considered to declare a distinct "protected region".

The EH entries in each EH table must honor the following rules:

1) If an EH entry covers one or more bytes of the primary function body, then
it may not cover any bytes in any fault funclets. If an EH entry covers one or
more bytes of a fault funclet, then it may not cover any bytes in the primary
function body or in other fault funclets.

2) If a catch EH entry covers one or more bytes of the primary function body,
then its catch entry points must point to instructions in the primary function
body. If a catch EH entry covers one or more bytes of a fault funclet F, then
its catch entry points must point to instructions in fault funclet F.

3) For each EH entry with a region depth of D and a protected region P, for all
other protected regions Q that overlap with P, one of the following must be
true: (i) Q has a region depth that is greater than D and P is a superset of
(or equal to) Q; or (ii) Q has a region depth that is less than D and P is a
subset of (or equal to) Q.

4) For each EH entry with a region depth of D and a protected region P, for
each integer I where 0 <= I < D there must be exactly one protected region Q in
the EH table where Q's region depth equals I and P overlaps with Q.


Processing exceptions
---------------------

The unwinder maintains a stack of exception infos called the "exception stack".
An "exception info" is a record that contains an exception object, a reference
to a frame on the call stack, a program counter (PC), and a non-negative
integer "region depth". When a thread of execution first begins, the exception
stack is initially empty.

HHBC allows programs to throw exceptions via the Throw instruction. When a
Throw instruction executes, it pushes a new exception info to the top of the
exception stack that contains the thrown exception object, a reference to the
frame that threw the exception, the PC at the time the exception was thrown,
and a region depth of D where D equals the number of protected regions in the
current frame's EH table that cover PC. Then it transfers control to the
unwinder which starts processing the exception info at the top of the exception
stack by following the steps given at the end of this section starting with
step 1 until control is transferred elsewhere.

HHBC also provides an Unwind instruction to allow a fault funclet to return
control back to the unwinder when it has finished its work. When the Unwind
instruction executes, it transfers control to the unwinder and the unwinder
resumes processing the exception info at the top of the exception stack by
following the steps below starting at step 1 until control is transferred
elsewhere.

Here are the steps that the unwinder follows to process the exception info at
the top of the stack (called the "current exception info"):

Step 1) Consult the EH table of the current exception info's function. Check if
        there are any EH entries that cover the current exception info's PC and
        have a region depth that is less than the current exception info's
        region depth. If one or more matching EH entries are found, choose the
        EH entry with the greatest region depth and continue on to step 2. If
        no matching EH entries are found go to step 5.

Step 2) Let E be the EH entry found in step 1, and let D be the region depth of
        E. Set the current exception info's region depth to D (overwriting the
        previous value). Continue on to step 3.

Step 3) If E is a fault EH entry, transfer control to E's fault funclet's entry
        point; eventually control will be transferred back to the unwinder,
        either via the Unwind instruction or via another exception being thrown
        with the Throw instruction. Otherwise continue on to step 4.

Step 4) E is a catch EH entry. Consult E's list of [class name, catch entry
        point] pairs. Check if there are any pairs in the list where the
        exception's type is compatible with the pair's class name. If one or
        more matching pairs are found, choose the one that occurs first in E's
        list and transfer control to that pair's catch entry point; the catch
        entry point begins with a Catch instruction which will take care of
        popping the current exception info off of the exception stack. If no
        matching pair is found, go to step 1.

Step 5) Check if the current exception info's PC is in a fault funclet. If it's
        not, continue to step 6. If it is, read the exception X from the
        current exception info and then pop the current exception info off of
        the exception stack. Then, read the new current exception info's
        exception Y, update exception X so that it's "previous" property chains
        to exception Y, and then update the new current exception info to point
        to exception X instead of exception Y. Then go to step 1.

Step 6) The current exception info's PC is in the primary function body. Pop
        the current frame off of the call stack and then check if the call
        stack is empty. If the call stack is empty, read the current exception
        info's exception X, clear the exception stack, and transfer control to
        the unhandled exception facility passing along exception X. If the call
        stack is not empty continue to step 7.

Step 7) Update the current exception info to refer to the new current frame at
        the top of the call stack, and set the current exception info's PC to
        point to the FCall* instruction which immediately precedes the PC of
        the current frame. Next, set the current exception info's region depth
        to D, where D equals the number of protected regions in the current
        frame's EH table that cover the current exception info's PC. Then go to
        step 1.


Property access
---------------

As object properties are accessed during execution, the execution engine is
responsible for following certain rules to honor each property's accessibility
and visibility.

The accessibility and visibility of a property in a given class is determined
by that class's definition and the definitions of all of that class's
ancestors. When a property is declared in a class definition (a "declared
property") it may be specified as being "public", "protected", or "private".
Accessibility and visibility are two related but distinct concepts. Depending
on the current context, a property may be visible and accessible, visible but
inaccessible, or invisible and inaccessible.

If a property P is declared with the "public" qualifier in the definition of
class C, for instances of class C and descendent classes the property P will be
visible and accessible in all contexts. If C has an ancestor that declares a
public property with the same name as P, C is said to "redeclare" property P,
and the declaration of P in class C is considered to refer to the same property
as the declaration in the ancestor class.

If a property P is declared as "protected" in the definition of class C, for
instances of class C the property P will be visible in all contexts, but only
accessible in the context of class C, an ancestor class, or descendent class.
When class C is loaded at run time, a semantic check must be performed to
ensure that all ancestor classes of C do not declare a property as "public"
with the same name as P. If C has an ancestor that declares a public property
with the same name as P, the execution engine must throw a fatal error when
class C is loaded. If C has an ancestor that declares a protected property with
the same name as P, C is said to "redeclare" property P, and the declaration of
P in class C is considered to refer to the same property as the declaration in
the ancestor class. Note that there may exist a class D that is a descendent of
C and declares a property as "public" with the same name as P. In such cases
the new "public" declaration in D is considered to refer to the same property
as the original "protected" declaration in C, and the "protected" qualifier
from the original declaration is effectively overridden by the "public"
qualifier from the new declaration. Class D is said to "redeclare" property P
with the "public" qualifier. Thus, for instances of class D and descendent
classes of D, property P will be visible and accessible in all contexts.
Finally, if a class E that is descendent of C does not redeclare P as public
and does not have an ancestor class that redeclares P as public, for instances
of class E the property P will be visible in all contexts, but only accessible
in the context of class E, an ancestor class of E, or a descendent class of E.

If a property P is declared with the "private" qualifier in the definition of
class C, for instances of class C the property P will be visible in all
contexts, but only accessible in the context of class C. For instances of
descendent classes of C, the property P will be visible and accessible in the
context of the class C, and in all other contexts property P will be invisible
and inaccessible. When class C is loaded at run time, a semantic check must be
performed to ensure that all ancestor classes of C do not declare a property as
"public" or "protected" with the same as P. If C has an ancestor that declares
a public or protected property with the same name as P, the execution engine
must throw a fatal error when class C is loaded. Note that descendent classes
of C may declare another property with the same name as P. The declaration of
property P as "private" in class C is considered to define a separate property
that is distinct from all other properties of the same name declared in
ancestor classes and descendent classes of C.

An instruction that accesses a property specifies the property by a name N via
a litstr id, a local variable id, or a cell consumed from the evaluation stack.
As noted above, it is possible for a class to have multiple distinct properties
named N. In cases where there are multiple distinct properties named N, the
visibility rules are used to determine which property is retrieved. If there is
a visible private property P named N, then property P is retrieved. Otherwise,
if there is a visible non-private property Q named N, then property Q is
retrieved. If there is no visible property named N, the behavior is determined
by the specific instruction. The semantic checks and the visibility rules
ensure that for any context there cannot be more than one visible private
property, and there cannot be more than one visible non-private property.

Some instructions can create a new property at run time with a name that is
different than the names of all declared properties that are visible in the
current context. Such properties are called "non-declared properties" or
"dynamic properties". Dynamic properties are considered to be visible and
accessible in all contexts.

If a declared property is unset, and then re-accessed/re-created, then it is
treated the same way as an invisible property with the same attributes as the
original declared property. Specifically, if the property gets created again,
it must have the same access attributes as the original declared property.


Magic property access methods
-----------------------------

Instructions that access properties may in some cases invoke a magic property
access method (__get, __set, __isset, or __unset) if an object implements the
method and the method is considered eligible for invocation. A magic property
access method is considered "eligible" for a given object if there is not a
frame on the call stack that corresponds to an invocation of the same method on
the same object.


Static property access
----------------------

As a class's static properties are accessed during execution, the execution
engine is responsible for following certain rules to honor each static
property's accessibility and visibility.

The accessibility and visibility of a static property in a given class is
determined by that class's definition and the definitions of all of that
class's ancestors. When a static property is declared in a class definition it
may be specified as being "public", "protected", or "private". Depending on the
current context, a static property may be visible and accessible, visible but
inaccessible, or invisible and inaccessible.

Conceptually, each class has a "static store" associated with it at run time
that provides storage for the static properties declared in the class's
definition. Static properties are accessed at run time by name through the
scope of a class. When an instruction accesses a static property through the
scope of class C, it will search the static store of C and then the static
stores of C's ancestors (starting with C's base class and moving up the
inheritance chain) for the first static property with the given name that is
visible in the current context.

If a static property S is declared with the "public" qualifier in the
definition of class C, the static property S when accessed through the scope of
class C or a descendent of C will be visible and accessible in all contexts.
Note that descendent classes of C may declare another static property with the
same name as S. The declaration in class C is considered to define a separate
static property that is distinct from all other static properties declared in
descendent classes of C.

If a static property S is declared with the "protected" qualifier in the
definition of class C, the static property S when accessed through the scope of
class C or a descendent of C will be visible in all contexts, but only
accessible in the context of class C, an ancestor class of C, or descendent
class of C. When class C is loaded at run time, a semantic check must be
performed to ensure that all ancestor classes of C do not declare a static
property as "public" with the same name as S. If C has an ancestor that
declares a public static property with the same name as S, the execution engine
must throw a fatal error when class C is loaded. Note that descendent classes
of C may declare another static property with the same name as S. The
declaration in class C is considered to define a separate static property that
is distinct from all other static properties declared in descendent classes of
C.

If a static property S is declared with the "private" qualifier in the
definition of class C, the static property S when accessed through the scope of
class C will be visible in all contexts, but only accessible in the context of
class C. The static property S when accessed through the scope of a descendent
of C will only be visible and accessible in the context of class C. When class
C is loaded at run time, a semantic check must be performed to ensure that all
ancestor classes of C do not declare a static property as "public" or
"protected" with the same name as S. If C has an ancestor that declares a
public or protected static property with the same name as S, the execution
engine must throw a fatal error when class C is loaded. Note that descendent
classes of C may declare another static property with the same name as S. The
declaration in class C is considered to define a separate static property that
is distinct from all other static properties declared in descendent classes of
C.

Note that instructions cannot create new static properties in a class that were
not declared in the class definition.


FPI regions
-----------

An FPI region is a contiguous range of bytecode that constitutes a call site.
Each FPI region begins immediately after an FPush* instruction that pushes an
FPI structure onto the FPI stack and must end with the corresponding FCall*
instruction that pops that FPI structure off of the FPI stack. If two FPI
regions overlap, one of the FPI regions must be completely enclosed by the
other FPI region. An FPI region may not contain backward jumps, nor may it
contain forward jumps that jump past the end of the FPI region.

Each function has an "FPI region table". Each row in the FPI region table
consists of the starting offset of the FPI region (the bytecode offset
immediately following the FPush* instruction), the ending offset of the FPI
region (the bytecode offset of the FCall* instruction), and the number of
parameters being passed.


Flavor descriptors
------------------

Any given value on the stack must either be a cell, ref, or classref at run
time. However, at bytecode generation time the specific flavor of a value on
the stack is not always known. HipHop bytecode uses symbols called "flavor
descriptors" to precisely describe what is known at bytecode generation about
the state of the evaluation stack at each instruction boundary.

Each instruction description specifies the flavor descriptor produced for each
of its outputs. Each description also specifies the flavor descriptor consumed
for each of the instruction's inputs.

Here is a description of each flavor descriptor:

  C - cell; specifies that the value must be a cell at run time
  V - ref; specifies that the value must be a ref at run time
  A - classref; specifies that the value must be a classref at run time
  R - return value; specifies that the value may be a cell or a ref at run
      time; this flavor descriptor is used for return values from function
      calls
  F - function argument; specifies that the value may be a cell or a ref at run
      time; this flavor descriptor is used for parameter values that are about
      to be passed into a function
  U - uninit; specifies that the value must be an uninitialized null at run
      time; this is only used for FCallBuiltin


Verifiability
-------------

Because instructions specify constraints on the flavor descriptor of each
input, it is important to be able to determine if a given HHBC program
satisfies these constraints. A program that satisfies the constraints on the
inputs to each instruction is said to be "flavor-safe".

HHBC provides a set of verification rules that can be mechanically applied to
verify that an HHBC program is flavor-safe. All valid HHBC programs must be
verifiably flavor-safe, and the execution engine may refuse to execute HHBC
programs that cannot be verified.

At bytecode generation time, what is known about the state of the evaluation
stack at a given instruction boundary can be precisely described using flavor
descriptors.

In addition to being flavor-safe, there are other invariants that valid HHBC
programs must uphold with respect to metadata and how certain instructions are
used.

Below is the complete list of verifiability rules. If the bytecode to be
executed does not come from a trusted source, it is the responsibility of the
bytecode execution engine to verify that these invariants hold.

1) The depth of the evaluation stack at any given point in the bytecode must be
the same for all possible control flow paths. The flavor descriptor of any
given slot on the evaluation stack at any given point in the bytecode must be
the same for all possible control flow paths.

2) No instruction may consume more values from the evaluation stack than are
available at that given point in the bytecode. Likewise, the flavor descriptor
of each slot on the evaluation stack must be compatible with the instruction's
inputs' flavor descriptors.

3) The evaluation stack must be empty at any offset listed as a catch entry
point.

4) If a given instruction is not the target of a forward branch and it follows
a Jmp, Switch, SSwitch, RetC, RetV, Unwind, Fatal, Throw, NativeImpl, or
ContHandle instruction, the evaluation stack before executing the given
instruction must be empty.

5) Before executing the RetC instruction, the evaluation stack must contain
exactly one value and the flavor descriptor of the value must be cell.
Likewise, before executing the RetV instruction, the evaluation stack must
contain exactly one value and the flavor descriptor of the value must be the
ref. Finally, before executing the Unwind instruction, the evaluation stack
must be empty.

6) The code for the primary function body and fault funclets must be laid out
in order in one contiguous block, starting with the primary function body and
optionally followed by one or more fault funclets. The code for primary
function body may not jump into the code for the funclets. Similarly, the code
for a funclet may not jump into the code for the primary function body or
another funclet.

7) Any bytecode instruction inside the primary function body or a fault funclet
that is immediately followed by an instruction not inside the same primary
function body or fault funclet must be one of the following instructions: Jmp,
Switch, SSwitch, RetC, RetV, Unwind, Fatal, Throw, NativeImpl, or ContHandle.

8) The primary function body may not contain the Unwind instruction, and fault
funclets may not contain the Ret* instructions. Also, each catch entry point
must point to a Catch instruction.

9) Each FPI region enumerated in the FPI region table must start immediately
after an FPush* instruction and it must end with an FCall* instruction. Each
use of the FPush* instruction must be the instruction immediately before
exactly one FPI region. Likewise, each use of the FCall* instruction must be
the last instruction in exactly one FPI region. Finally, FPass* instructions
may not be used outside an FPI region.

10) Each FPI region may not contain backward jumps, nor may it contain forward
jumps that jump outside the end of the FPI region. Also, there may not be jumps
anywhere in the function that transfer control from the outside of a given FPI
region to the inside of that region. An FPI region may not contain the Ret*,
Unwind, Throw, or Fatal instructions. Finally, an entry point may not point to
an instruction inside an FPI region.

11) The depth of the FPI stack at any given point in the bytecode must be the
same for all possible control flow paths. Also, for any given FPI region that
passes n parameters, all possible control flow paths from the beginning of the
region to the end must pass through exactly n FPass* instructions associated
with the region which pass the parameters in forward order.

12) Given an evaluation stack of depth n after an FPush* instruction, the
evaluation stack before the corresponding FCall* instruction must also have a
depth of n. Likewise, the evaluation stack after corresponding FPass*
instructions must have a depth of n as well. Finally, no instruction between an
FPush* and its corresponding FCall* may consume any of the values from the
evaluation stack that were pushed onto the stack before the FPush* instruction.

13) The initialization state of each iterator variable must be known at every
point in the code and must be the same for all control paths. There are four
possible states: (1) uninitialized, (2) "iter-initialized" (initialized via
IterInit*), (3) "miter-initialized" (initialized via MIterInit*), and (4)
"cuf-initialized" (initialized via DecodeCufIter). Every range of bytecode for
which an iterator variable i is initialized must be protected by a fault
funclet that unsets i by calling IterFree, MIterFree, or CIterFree.

14) The iterator variable referenced by IterInit* or MIterInit* or
DecodeCufIter must be in the uninitialized state when the instruction executes.
An iterator variable referenced by IterNext* and IterFree must be in the
"iter-initialized" state, an iterator variable referenced by MIterNext* or
MIterFree must be in the "miter-initialized" state, and an iterator variable
referenced by FPushCufIter or CIterFree must be in the citer-initialized state.
Note that IterInit* and MIterInit* conditionally initialize the iterator
variable, and IterNext* and MIterNext* conditionally free the iterator
variable.

15) Each EH table must follow all of the rules specified in the "Exception
handler (EH) table" section.


Instruction set
---------------

Each instruction description below consists of a mnemonic, followed by 0 or
more immediate operands, followed by a stack transition description of the form
"[xn,...,x2,x1] -> [ym,...,y2,y1]", where "[xn,...,x2,x1]" is a list of flavor
descriptors describing what the instruction consumes from the evaluation stack
and "[ym,...,y2,y1]" is the list of flavor descriptors describing what the
instruction pushes onto the stack. x1 and y1 represent the topmost stack
elements before and after execution, respectively.

Each element of a stack transition may also contain an optional type
annotation. Here is the list of the type annotations used in instruction
descriptions:

  Null - denotes the null type
  Bool - denotes the boolean type
  Int - denotes the integer type
  Dbl - denotes the double-precision floating-point type
  Str - denotes the string type
  Arr - denotes the array type
  Obj - denotes the object type

Multiple type annotations may be combined together using the "|" symbol. For
example, the type annotation "Int|Dbl" means that a value is either integer or
a double.

Some instructions may contain multiple stack transition descriptions to express
the relationship between the types of the values consumed from the stack and
types of the values pushed onto the stack. Also, in some stack transition
descriptions, "<T>" is used as shorthand to represent any one specific type.
For example, a transition such as "[C:<T>] -> [C:<T>]" indicates that the type
of value that the instruction pushes onto the stack will match the type of
value that it consumed from the stack. Likewise, "<F>" is used as shorthand to
represent any one specific flavor descriptor.

$1 is used to refer to the value at the top of the evaluation stack, $2 is used
to refer to the value directly below $1 on the evaluation stack, $3 is used to
refer to the value directly below $2, and so forth. Also, %1 is used to refer
to the first immediate argument, and %2 is used to refer to the second
immediate argument.

Note that the relative offset immediate used by a Jmp*, Iter*, MIter*, Switch,
SSwitch, CreateCont or AsyncESuspend instruction is relative the beginning 
of the instruction.

There are numerous instructions that operate on different kinds of locations.
Locations are specified using "location descriptors". The complete list of
location descriptors is given below:

  L - local id; location is the local variable whose id is given by an
      immediate.
  N - local name; location is the local variable whose name is given by the
      value of a cell.
  G - global name; location is the global variable whose name is given by the
      value of a cell.
  S - static property; location is the static property whose class is given by
      a classref and whose name is given by value of a cell.
  C - cell; location is a temporary value given by a cell.
  R - return value; location is a temporary value given by a cell or a ref
  H - $this; location is the $this pointer in the current frame. Must only be
      used in a frame that is known to have a non-null $this pointer; CheckThis
      is most commonly used to ensure this.

There are several groups of similarly named instructions where the name of each
instruction ends with a different location descriptor (for example, Set*). Each
instruction in the group perform similar actions but take different kinds of
inputs to specify the location to access.

The Member instructions provide functionality to operate on elements and
properties. These instructions incorporate an immediate argument vector which
specifies a location descriptor (defined above) followed by one or more member
descriptors:

  EC       - consume a cell from the evaluation stack as an element
  EL:<id>  - consume a local given by an immediate id as an element
  ET:<id>  - consume a litstr given by an immediate id as an element
  EI:<int> - consume a immediate integer as an element
  PC       - consume a cell from the evaluation stack as a property
  PL:<id>  - consume a local given by an immediate id as a property
  PT:<id>  - consume a litstr given by an immediate id as a property
  W        - synthesize a new element (no corresponding local variable or
             evaluation stack slot)

For example, the following correspondence exists (ignoring setup bytecode):

  Source code: $a[3][$b][]['hi'] = 42;
  Bytecode: SetM <L:0 EI:3 EL:1 W ET:hi>
  Stack:    []  ->  [42]

Instructions that have an immediate argument vector have different stack
transition descriptions depending on the kind of location descriptor and member
descriptors contained in the immediate argument vector. Member instructions
denote the immediate argument vector using the notation "<loc-desc/M-vector>"
and "C..C" is used to indicate that the member instructions consume a variable
number of cells from the stack. In most cases the immediate vector arguments
are ordered such that the loc-desc comes first (deepest in the stack), with the
last M-vector element last (shallowest in the stack). However, classrefs that
are part of the BaseSC and BaseSL loc-desc inputs always come last (the cell
input to BaseSC comes first though). Instructions accepting an immediate vector
containing a list of iterators and iterator types use the notation
"<iter-vec>".

In addition to describing each instruction, this instruction set documentation
also describes several operations that encapsulate fundamental, but non-trivial
processes that are shared by the Member instructions.

The instruction set is organized into the following sections:
   1. Basic instructions
   2. Literal and constant instructions
   3. Operator instructions
   4. Control flow instructions
   5. Get instructions
   6. Isset, Empty and type querying instructions
   7. Mutator instructions
   8. Call instructions
   9. Member operations
  10. Member instructions
  11. Iterator instructions
  12. Include, eval, and define instructions
  13. Miscellaneous instructions
  14. Continuation creation and execution


1. Basic instructions
---------------------

Nop    []  ->  []

  No operation. This instruction does nothing.

PopA    [A]  ->  []
PopC    [C]  ->  []
PopV    [V]  ->  []
PopR    [R]  ->  []

  Pop. Discards the value on the top of the stack.

Dup    [C:<T>]  ->  [C:<T> C:<T>]

  Duplicate. Duplicates the cell $1 and pushes it onto the stack.

Box    [C:<T>]  ->  [V:<T>]

  Box. Creates a new ref, sets the new ref to point at a copy of cell $1, and
  pushes the ref onto the stack.

Unbox    [V:<T>]  ->  [C:<T>]

  Unbox. Creates a copy of the cell that ref $1 points to, and pushes the cell
  onto the stack.

BoxR    [R:<T>]  ->  [V:<T>]

  Box. If $1 is a ref at run time, this instruction does nothing.

  If $1 is a cell at run time, this instruction creates a new ref, sets the new
  ref to point at a copy of cell $1, and pushes the ref onto the stack.

BoxRNop   [R:<T>]  ->  [C:<T>]

  Box, no op. $1 must be statically known to be boxed.

UnboxR    [R:<T>]  ->  [C:<T>]

  Unbox. If $1 is a cell at run time, this instruction does nothing.

  If $1 is a ref at run time, this instruction creates a copy of the cell that
  ref $1 points to, and pushes the cell onto the stack.

UnboxRNop   [R:<T>]  ->  [C:<T>]

  UnboxR, no op. $1 must be statically known to be unboxed. This instruction
  pushes $1 on the stack as a cell.

2. Literal and constant instructions
------------------------------------

Null     []  ->  [C:Null]
True     []  ->  [C:Bool]
False    []  ->  [C:Bool]

  Push constant. Null pushes null onto the stack, True pushes true onto the
  stack, and False pushes false onto the stack.

NullUninit                      []  ->  [U]

   Push an uninitialized null on the stack.

Int <signed 64-bit integer value>    []  ->  [C:Int]
Double <double value>                []  ->  [C:Dbl]
String <litstr id>                   []  ->  [C:Str]
Array <scalar array id>              []  ->  [C:Arr]

  Push immediate. Pushes %1 onto the stack.

NewArray <capacity hint>    []  ->  [C:Arr]

  New array, with a capacity hint. Creates a new array and pushes it onto the
  stack. The implementation may make use of the hint in %1 to pre-size the
  array. The hint %1 must be greater than or equal to 0.

NewPackedArray <num elems>    [C..C]  ->  [C]

  New array. Creates a new array from the top %1 cells on the stack, pops those
  cells, then pushes the new array onto the stack. Elements are pushed on the
  stack in array insertion order and are implicitly numbered from 0 to %1 - 1.
  $1 is at index %i - 1, $2 at %1-2, and so on; $(%1) is at index 0.

NewStructArray <litstr id vector>    [C..C]  ->  [C]

  New array. Creates a new array from the names given in %1 and values from the
  stack. The vector of litstr ids gives the element names, one value for each
  name is popped from the stack. Names are in array insertion order, and values
  were pushed onto the stack in insertion order, so are added to the array in
  reverse order (the topmost value will become the last element in the array).
  For example:

    NewStructArray < "a" "b" >  [ 1 2 ]  ->  [ array("a"=>1, "b"=>2) ]

AddElemC    [C C C]  ->  [C:Arr]

  Add element. If $3 is an array, this instruction executes $3[$2] = $1 and
  then pushes $3 onto the stack.

  If $3 is not an array, this instruction throws a fatal error.

AddElemV    [C C V]  ->  [C:Arr]

  Add element. If $3 is an array, this instruction executes $3[$2] = &$1 and
  then pushes $3 onto the stack.

  If $3 is not an array, this instruction throws a fatal error.

AddNewElemC    [C C]  ->  [C:Arr]

  Add new element. If $2 is an array, this instruction executes $2[] = $1 and
  then pushes $2 onto the stack.

  If $2 is not an array, this instruction throws a fatal error.

AddNewElemV    [C V]  ->  [C:Arr]

  Add new element. If $2 is an array, this instruction executes $2[] = &$1 and
  then pushes $2 onto the stack.

  If $2 is not an array, this instruction throws a fatal error.

NewCol <coll type> <num elems>    []  ->  [C:Obj]

  New collection. Creates a new collection of type %1 with an initial capacity
  sufficient to hold the number of elements specified by %2, and pushes the
  collection onto the stack.

ColAddElemC    [C C C]  ->  [C:Obj]

  Collection add key/value pair. If $3 is a collection object, this instruction
  executes $3[$2] = $1 and then pushes $3 onto the stack.

  If $3 is not a collection object, this instruction throws a fatal error.

ColAddNewElemC    [C C]  ->  [C:Obj]

  Collection add value. If $2 is a collection object, this instruction executes
  $2[] = $1 and then pushes $2 onto the stack.

  If $2 is not a collection object, this instruction throws a fatal error.

Cns <litstr id>    []  ->  [C:Null|Bool|Int|Dbl|Str]

  Get constant. Pushes the value of the global constant named %1 onto the stack
  as a cell. If there is no constant named %1, this instruction raises a notice
  and pushes the string %1 onto the stack as a cell.

CnsE <litstr id>    []  ->  [C:Null|Bool|Int|Dbl|Str]

  Get constant. Pushes the value of the global constant named %1 onto the stack
  as a cell. If there is no constant named %1, throws a fatal error.

CnsU <litstr id> <litstr fallback>  []  ->  [C:Null|Bool|Int|Dbl|Str]

  Get constant. Identical to Cns except returns constant named %2 if the
  constant named %1 is undefined.

ClsCns <litstr id>    [A]  ->  [C:Null|Bool|Int|Dbl|Static Str]

  Get class constant. This instruction pushes the value of the class constant
  named %1 from class $1 onto the stack. If there is no class constant named %1
  in class $1, this instruction throws a fatal error.

ClsCnsD <litstr id> <litstr id>    []  ->  [C:Null|Bool|Int|Dbl|Static Str]

  Get class constant (direct). This instruction first checks if %2 matches the
  name of a defined class. If %2 does not match the name of a defined class,
  this instruction will invoke the autoload facility passing in the class name
  %2, and then it will again check if %2 matches the name of a defined class.
  If %2 still does not match the name of a defined class this instruction
  throws a fatal error.

  Next, this instruction pushes the value of the class constant named %1 from
  class %2 onto the stack. If there is no class constant named %1 in class %2,
  this instruction throws a fatal error.

File    []  ->  [C:Static Str]
Dir     []  ->  [C:Static Str]

  Push string. File pushes __FILE__ onto the stack, and Dir pushes __DIR__ onto
  the stack.

NameA   [A] ->  [C:Static Str]

  Push the name of the class in $1 as a string.


3. Operator instructions
------------------------

Concat    [C C]  ->  [C:Str]

  Concatenation (.). Pushes ((string)$2 . (string)$1) on the stack.

Abs    [C] -> [C:Int|Dbl|Bool]

  Absolute value. Computes the absolute value of $1 and pushes the result onto
  the stack.

Add    [C:Arr C:Arr]  ->  [C:Arr]
       [C:<T2> C:<T1>]  ->  [C:Dbl]    (where T1 == Dbl || T2 == Dbl)
       [C:<T2> C:<T1>]  ->  [C:Int]    (where T1 != Dbl && T2 != Dbl &&
                                              (T1 != Arr || T2 != Arr))

  Addition (+). Performs addition (or plus-merge if $1 and $2 are both arrays).
  Pushes ($2 + $1) onto the stack. This instruction throws a fatal error if
  is_array($1) xor is_array($2) is true.

Sub    [C:<T2> C:<T1>]  ->  [C:Dbl]    (where T1 == Dbl || T2 == Dbl)
       [C:<T2> C:<T1>]  ->  [C:Int]    (where T1 != Dbl && T2 != Dbl)

  Subtraction (-). Pushes ($2 - $1) onto the stack. This instruction throws a
  fatal error if is_array($1) || is_array($2) is true.

Mul    [C:<T2> C:<T1>]  ->  [C:Dbl]    (where T1 == Dbl || T2 == Dbl)
       [C:<T2> C:<T1>]  ->  [C:Int]    (where T1 != Dbl && T2 != Dbl)

  Multiplication (*). Pushes ($2 * $1) onto the stack. This instruction throws
  a fatal error if is_array($1) || is_array($2) is true.

AddO   [C:Arr C:Arr]  ->  [C:Arr]
       [C:<T2> C:<T1>]  ->  [C:Dbl]     (where T1 == Dbl || T2 == Dbl)
       [C:<T2> C:<T1>]  ->  [C:Int|dbl] (where T1 != Dbl && T2 != Dbl &&
                                               (T1 != Arr || T2 != Arr))

  Same behavior as Add, except for when both inputs have type Int and the
  result would not fit in a 64-bit integer. Then this instruction will push
  (double)$1 + (double)$2.

SubO   [C:<T2> C:<T1>]  ->  [C:Dbl]     (where T1 == Dbl || T2 == Dbl)
       [C:<T2> C:<T1>]  ->  [C:Int|Dbl] (where T1 != Dbl && T2 != Dbl)

  Same behavior as Sub, except for when both inputs have type Int and the
  result would not fit in a 64-bit integer. Then this instruction will push
  (double)$1 - (double)$2.

MulO   [C:<T2> C:<T1>]  ->  [C:Dbl]     (where T1 == Dbl || T2 == Dbl)
       [C:<T2> C:<T1>]  ->  [C:Int|Dbl] (where T1 != Dbl && T2 != Dbl)

  Same behavior as Mul, except for when both inputs have type Int and the
  result would not fit in a 64-bit integer. Then this instruction will push
  (double)$1 * (double)$2.

Div    [C C]  ->  [C:Bool|Int|Dbl]
       [C:Dbl C:Int]  ->  [C:Bool|Dbl]
       [C:Int C:Dbl]  ->  [C:Bool|Dbl]
       [C:Dbl C:Dbl]  ->  [C:Bool|Dbl]

  Division (/). Pushes ($2 / $1) onto the stack. This instruction throws a
  fatal error if is_array($1) || is_array($2) is true.

Mod    [C C]  ->  [C:Bool|Int]

  Modulus (%). Pushes ((int)$2 % (int)$1) onto the stack. This instruction
  never throws a fatal error.

Sqrt   [C]    ->  [C:Null|Dbl]

  Square root. Computes the square root of $1 and pushes the result onto the
  stack. If $1 is not null, a bool, an int, a double, or a numeric string, it
  raises a warning and pushes null onto the stack.

  If $1 is a negative number, this instruction pushes a floating-point value
  representing NAN onto the stack.

Strlen  [C] -> [C:Null|Int]

  String length. If $1 is a string push the length of the string on the stack.
  If $1 is an object with a __toString method, call this method and push the
  length of the resulting string on the stack. If $1 is an array or resource,
  raise a warning and push null on the stack. Otherwise convert $1 to a string
  and push the length of that string on the stack.

Xor    [C C]  ->  [C:Bool]

  Logical xor (xor). Pushes ((bool)$2 xor (bool)$1) onto the stack.

Not    [C]  ->  [C:Bool]

  Logical not (!). Pushes (!(bool)$1) onto the stack.

Same    [C C]  ->  [C:Bool]

  Same (===). Pushes ($2 === $1) onto the stack.

NSame    [C C]  ->  [C:Bool]

  Not same (!==). Pushes ($2 !== $1) onto the stack.

Eq    [C C]  ->  [C:Bool]

  Equals (==). Pushes ($2 == $1) onto the stack.

Neq    [C C]  ->  [C:Bool]

  Not equal (!=). Pushes ($2 != $1) onto the stack.

Lt    [C C]  ->  [C:Bool]

  Less than (<). Pushes ($2 < $1) onto the stack.

Lte    [C C]  ->  [C:Bool]

  Less than or equal to (<=). Pushes ($2 <= $1) onto the stack.

Gt    [C C]  ->  [C:Bool]

  Greater than (>). Pushes ($2 > $1) onto the stack.

Gte    [C C]  ->  [C:Bool]

  Greater than or equal to (>=). Pushes ($2 >= $1) onto the stack.

BitAnd    [C:<T2> C:<T1>]  ->  [C:Int]    (where T1 != Str || T2 != Str)
          [C:Str C:Str]  ->  [C:Str]

  Bitwise and (&). Pushes ($2 & $1) onto the stack. If either $1 or $2 is an
  object, this instruction throws a fatal error.

BitOr    [C:<T2> C:<T1>]  ->  [C:Int]    (where T1 != Str || T2 != Str)
         [C:Str C:Str]  ->  [C:Str]

  Bitwise or (|). Pushes ($2 | $1) onto the stack. If either $1 or $2 is an
  object, this instruction throws a fatal error.

BitXor    [C:<T2> C:<T1>]  ->  [C:Int]    (where T1 != Str || T2 != Str)
          [C:Str C:Str]  ->  [C:Str]

  Bitwise xor (^). Pushes ($2 ^ $1) onto the stack. If either $1 or $2 is an
  object, this instruction throws a fatal error.

BitNot    [C:<T>]  ->  [C:Int]    (where T != Str)
          [C:Str]  ->  [C:Str]

  Bitwise not (~). Pushes (~$1) onto the stack. If $1 is null, a boolean, an
  array, or an object, this instruction throws a fatal error.

Shl    [C C]  ->  [C:Int]

  Shift left (<<). Pushes ((int)$2 << (int)$1) onto the stack. This instruction
  never throws a fatal error.

Shr    [C C]  ->  [C:Int]

  Shift right (>>). Pushes ((int)$2 >> (int)$1) onto the stack. This
  instruction never throws a fatal error.

Floor    [C]  ->  [C:Dbl]

  Round $1 to nearest integer value not greater than $1. Converts $1 to numeric
  as appropriate and then takes floor of resulting numeric value.

Ceil    [C]  ->  [C:Dbl]

  Round $1 to nearest integer value not less than $1. Converts $1 to numeric as
  appropriate and then takes ceil of resulting numeric value.

CastBool    [C]  ->  [C:Bool]

  Cast to boolean ((bool),(boolean)). Pushes (bool)$1 onto the stack.

CastInt    [C]  ->  [C:Int]

  Cast to integer ((int),(integer)). Pushes (int)$1 onto the stack.

CastDouble    [C]  ->  [C:Dbl]

  Cast to double ((float),(double),(real)). Pushes (double)$1 onto the stack.

CastString    [C]  ->  [C:Str]

  Cast to string ((string),(binary)). Pushes (string)$1 onto the stack. If $1
  is an object that implements the __toString method, the string cast returns
  $1->__toString(). If $1 is an object that does not implement __toString
  method, the string cast throws a fatal error.

CastArray    [C]  ->  [C:Arr]

  Cast to array ((array)). Pushes (array)$1 onto the stack.

CastObject    [C]  ->  [C:Obj]

  Cast to object ((object)). Pushes (object)$1 onto the stack.

InstanceOf    [C C]  ->  [C:Bool]

  Instance of (instanceof). If $1 is a string and it matches the name of a
  defined class and $2 is an object that is an instance of $1, this instruction
  pushes true onto the stack. If $1 is an object and get_class($1) matches the
  name of a defined class and $2 is an object that is an instance of
  get_class($1), this instruction pushes true onto the stack. If $1 is not a
  string or an object, this instruction throws a fatal error.

InstanceOfD <litstr id>    [C]  ->  [C:Bool]

  Instance of direct (instanceof). If %1 matches the name of a defined class
  and $1 is an instance of the %1, this instruction pushes true onto the stack,
  otherwise it pushes false onto the stack.

Print    [C]  ->  [C:Int]

  Print (print). Outputs (string)$1 to STDOUT and pushes the integer value 1
  onto the stack.

Clone    [C]  ->  [C:Obj]

  Clone (clone). Clones $1 and pushes it onto the stack. If $1 is not an
  object, this instruction throws a fatal error.

Exit    [C]  ->  [C:Null]

  Exit (exit). Terminates execution of the program.

  If $1 is an integer, this instruction will set the exit status to $1, push
  null onto the stack, and then it will terminate execution.

  If $1 is not an integer, this instruction will output (string)$1 to STDOUT,
  set the exit status to 0, push null onto the stack, and then it will
  terminate execution.

Fatal <fatal subop>  [C]  ->  []

  Fatal. This instruction throws a fatal error using $1 as the error message.
  If $1 is not a string, this instruction throws a fatal error with an error
  message that indicates that the error message was not a string. Setting %1 to
  0 will throw a runtime fatal error with a full backtrace. Setting %1 to 1
  will throw a parse fatal error with a full backtrace. Setting %1 to 2 will
  throw a runtime fatal error with the backtrace omitting the top frame.


4. Control flow instructions
----------------------------

Jmp <rel offset>    []  ->  []

  Jump. Transfers control to the location specified by %1.

JmpNS <rel offset>   []  ->  []

  Jump, with no surprise flag checks. This behaves identically to the Jmp
  instruction, except that internal VM checks for things like OOM do not need
  to be performed even if the offset is negative.

JmpZ <rel offset>    [C]  ->  []

  Jump if zero. Conditionally transfers control to the location specified by %1
  if (bool)$1 == (bool)0.

JmpNZ <rel offset>    [C]  ->  []

  Jump if not zero. Conditionally transfers control to the location specified
  by %1 if (bool)$1 != (bool)0.

Switch <offset vector> <base> <bounded>    [C]  ->  []

  Switch over integer case values. If bounded == 0, the implementation will
  assume that $1 is an integer in the range [0, length(vector)) and
  unconditionally transfer control to the location specified by vector[$1].
  Undefined behavior will result if $1 is not an integer inside this range. If
  bounded != 0, the following rules take over:

  For a bounded Switch, the last two elements of the offset vector are special:
  they represent the first non-zero case and the default case, respectively.
  base + length(vector) - 2 must not be greater than 2^63-1. If $1 === true,
  control will be transferred to the location specified by
  vector[length(vector) - 2]. If $1 is equal (as defined by Eq) to any integer
  $n in the range [base, base + length(vector) - 2), control will be
  transferred to the location specified by vector[$n - base]. Otherwise,
  control will be transferred to the location specified by
  vector[length(vector) - 1].

SSwitch <litstr id/offset vector>    [C]  ->  []

  Switch over string case values. This instruction will search the
  string/offset vector from the beginning until it finds a string that is equal
  to $1. If one is found, control will be transferred to the location specified
  by the offset corresponding to that string. If a matching string is not
  found, control is transferred to the location specified by the final element
  in the vector, which must have a litstr id of -1.

RetC    [C]  ->  []
RetV    [V]  ->  []

  Return. Returns $1 to the caller. This instruction may not be used inside
  default value funclets or fault funclets.

Unwind    []  ->  []

  Unwind. Transfers control back to the unwinder. This instruction may only be
  used inside a fault funclet.

Throw    [C]  ->  []

  Throw. Throws the object $1. If $1 is not an object that extends the
  Exception class, this instruction throws a fatal error.


5. Get instructions
-------------------

CGetL <local variable id>    []  ->  [C]

  Get local as cell. If the local variable given by %1 is defined, this
  instruction gets the value of the local variable and pushes it onto the stack
  as a cell. If the local variable is not defined, this instruction raises a
  warning and pushes null onto the stack.

CGetL2 <local variable id>    [<F>:<T>]  ->  [C <F>:<T>]

  Get local as cell. If the local variable given by %1 is defined, this
  instruction gets the value of the local variable, pushes it onto the stack as
  a cell, and then pushes $1 onto the stack.

  If the local variable is not defined, this instruction raises a warning,
  pushes null onto the stack, and then pushes $1 onto the stack.

CGetL3 <local variable id>    [<F2>:<T2> <F1>:<T1>]  ->  [C <F2>:<T1> <F1>:<T1>]

  Get local as cell. If the local variable given by %1 is defined, this
  instruction gets the value of the local variable, pushes it onto the stack as
  a cell, then pushes $2 onto the stack, and then pushes $1 onto the stack.

  If the local variable given by %1 is not defined, this instruction raises a
  warning, pushes null onto the stack, then pushes $2 onto the stack, and then
  pushes $1 onto the stack.

PushL <local variable id>    []  ->  [C]

  Teleport local value to eval stack. The local variable given by %1 must be
  defined and must not contain a reference. This instruction pushes the local's
  value on the stack, then unsets it, equivalent to the behavior of UnsetL.

CGetN    [C]  ->  [C]

  Get local as cell. This instruction first computes x = (string)$1. Next, this
  instruction reads the local variable named x pushes its value onto the stack
  as a cell.

  If there is no local variable defined named x, this instruction pushes null
  onto the stack and raises a warning.

CGetG    [C]  ->  [C]

  Get global as cell. This instruction first computes x = (string)$1. Next,
  this instruction reads the global variable named x pushes its value onto the
  stack as a cell.

  If there is not a global variable defined named x, this instruction pushes
  null onto the stack and raises a warning.

CGetS    [C A]  ->  [C]

  Get static property as cell. This instruction first checks if class $1 has a
  visible and accessible static property named (string)$2. If it doesn't, this
  instruction throws a fatal error. Otherwise, this instruction pushes the
  static property onto the stack as a cell.

VGetL <local variable id>    []  ->  [V]

  Get local as ref. This instruction boxes the local variable given by %1 if
  necessary and pushes it onto the stack as a ref. If the given local variable
  is not defined, this instruction defines it, sets it to null, boxes it, and
  pushes a the value of the local variable onto the stack as a ref.

VGetN    [C]  ->  [V]

  Get local as ref. This instruction first computes x = (string)$1. Next, this
  instruction boxes the local variable named x (if the local is a cell) and
  pushes its value onto the stack as a ref. If there is no local variable
  defined named x, this instruction defines a local variable named x, sets it
  to null, boxes it, and pushes the value of the local variable onto the stack
  as a ref.

VGetG    [C]  ->  [V]

  Get global as ref. This instruction first computes x = (string)$1. Next, this
  instruction boxes the global variable named x (if the local is a cell) and
  pushes its value onto the stack as a ref. If there is no global variable
  defined named x, this instruction defines a global variable named x, sets it
  to null, boxes it, and pushes the value of the global variable onto the stack
  as a ref.

VGetS    [C A]  ->  [V]

  Get static property as ref. This instruction first checks if class $1 has a
  visible and accessible static property named (string)$2. If it doesn't, this
  instruction throws a fatal error. Otherwise, this instruction boxes the
  static property and pushes it onto the stack as a ref.

AGetC                        [C]  ->  [A]
AGetL <local variable id>    []  ->  [A]

  Fetch class. This instruction first loads a value into x as shown by the
  following table:

    instruction    x
    ------------+----
      AGetC     | $1
      AGetL     | %1

  Next this instruction checks if x is a string or an object. If x is not a
  string or object, this instruction throws a fatal error. Otherwise, this
  instruction executes y = (is_object(x) ? get_class(x) : (string)x) and checks
  if y matches the name of a defined class. If y does not match the name of a
  defined class, this instruction will invoke the autoload facility passing in
  the class name y, and then it will again check if y matches the name of a
  defined class. If y still does not match the name of a defined class this
  instruction throws a fatal error.

  Next, this instruction pushes a classref that refers to the class named y.


6. Isset, Empty, and type querying instructions
-----------------------------------------------

IssetC    [C]  ->  [C:Bool]

  Isset. If $1 is null this instruction pushes false onto the stack, otherwise
  it pushes true.

IssetL <local variable id>    []  ->  [C:Bool]

  Isset local. This instruction reads the local variable given by %1. If the
  local variable is undefined or null, this instruction pushes false onto the
  stack, otherwise it pushes true.

IssetN    [C]  ->  [C:Bool]

  Isset local. This instruction reads the local variable named (string)$1. If
  the local variable is undefined or null, this instruction pushes false onto
  the stack, otherwise it pushes true.

IssetG    [C]  ->  [C:Bool]

  Isset global. This instruction reads the global variable named (string)$1. If
  the global variable is undefined or null, this instruction pushes false onto
  the stack, otherwise it pushes true.

IssetS    [C A]  ->  [C:Bool]

  Isset static property. This instruction first computes x = (string)$2. Next
  it checks if class $1 has an accessible static property named x. If it
  doesn't, this instruction pushes false.

  If class $1 does have an accessible property named x, this instruction reads
  the static property named x. If the static property is null, this instruction
  pushes false onto the stack, otherwise it pushes true.

EmptyL <local variable id>     []  ->  [C:Bool]

  Empty local. This instruction reads the local variable named %1 into x. If
  the local variable is defined this instruction pushes !(x) onto the stack,
  otherwise it pushes true.

EmptyN    [C]  ->  [C:Bool]

  Empty local. This instruction reads the local variable named (string)$1 into
  x. If the local variable is defined this instruction pushes !(x) onto the
  stack, otherwise it pushes true.

EmptyG    [C]  ->  [C:Bool]

  Empty global. This instruction reads the global variable named (string)$1
  into x. If the global variable is defined this instruction pushes !(x) onto
  the stack, otherwise it pushes true.

EmptyS    [C A]  ->  [C:Bool]

  Empty static property. This instruction first checks if class $1 has an
  accessible static property named (string)$2. If it doesn't, this instruction
  pushes true, otherwise this instruction reads the static property into x and
  pushes !(x) onto the stack.

IsTypeC  <op>                     [C]  ->  [C:Bool]

  Is type. This instruction first loads a type into t based on the operand op,
  according to the following table:

    operand        t
    -----------+------
      Null     | Null
      Bool     | Bool
      Int      |  Int
      Dbl      |  Dbl
      Str      |  Str
      Arr      |  Arr
      Obj      |  Obj
      Scalar   | Int or Dbl or Str or Bool

  If $1 is of type t, this instruction pushes true onto the stack, otherwise it
  pushes false.

IsTypeL  <local variable id> <op>     []  ->  [C:Bool]

  Is type. This instruction first loads a type into t and a value into x as
  given by the following table:

    operand        t                           x
    -----------+----------------------------+-------
      Null     | Null                       |  true
      Bool     | Bool                       | false
      Int      |  Int                       | false
      Dbl      |  Dbl                       | false
      Str      |  Str                       | false
      Arr      |  Arr                       | false
      Obj      |  Obj                       | false
      Scalar   |  Int or Dbl or Str or Bool | false

  If the local variable given by %1 is defined, this pushes true onto the stack
  if the local variable is of type t, otherwise it pushes false. If the local
  is of kind reference, then the inner value is used to determine the type.

  If the local variable given by %1 is not defined, this instruction raises a
  warning and pushes x onto the stack.


7. Mutator instructions
-----------------------

SetL <local variable id>    [C]  ->  [C]

  Set local. This instruction marks the local variable given by %1 as defined,
  stores the value $1 into the local variable, and then pushes $1 onto the
  stack.

SetN    [C C]  ->  [C]

  Set local. This instruction marks the local variable named (string)$2 as
  defined, assigns the value $1 to the local variable, and then pushes $1 onto
  the stack.

SetG    [C C]  ->  [C]

  Set global. This instruction marks the global variable named (string)$2 as
  defined, assigns the value $1 to the global variable, and then pushes $1 onto
  the stack.

SetS    [C A C]  ->  [C]

  Set static property. First this instruction checks if class $2 has an
  accessible static property named (string)$3. If it doesn't, this instruction
  throws a fatal error. Otherwise, this instruction assigns the value $1 to the
  static property, and then it pushes $1 onto the stack.

SetOpL <local variable id> <op>    [C]  ->  [C]

  Set op local. If the local variable given %1 is not defined, this instruction
  marks it as defined, sets it to null, and raises a warning.

  Next, this instruction reads the local variable into x, then executes y = x
  <op> $1, assigns y into local variable %1, and then pushes y onto the stack.
  The immediate value must be one of the following opcodes:
    Add, AddO, Sub, SubO, Mul, MulO, Div, Mod, Shl, Shr, Concat, BitAnd,
    BitOr, BitXor.

SetOpN <op>    [C C]  ->  [C]

  Set op local. This instruction first computes x = (string)$2. If the local
  variable named n is not defined, this instruction marks it as defined, sets
  it to null, and raises a warning.

  Next, this instruction reads the local variable named x into y, executes z =
  y <op> $1, assigns z into the local variable named x, and then pushes z onto
  the stack as a cell. The immediate value must be one of the following
  opcodes:
    Add, Sub, Mul, Div, Mod, Shl, Shr, Concat, BitAnd, BitOr, BitXor.

SetOpG <op>    [C C]  ->  [C]

  Set op global. This instruction first computes x = (string)$2. If the global
  variable named n is not defined, this instruction marks it as defined, sets
  it to null, and raises a warning.

  Next, this instruction reads the global variable named x into y, executes z =
  y <op> $1, assigns z into the global variable named x, and then pushes z onto
  the stack as a cell. The immediate value must be one of the following
  opcodes:
    Add, Sub, Mul, Div, Mod, Shl, Shr, Concat, BitAnd, BitOr, BitXor.

SetOpS <op>    [C A C]  ->  [C]

  Set op static property. This instruction first computes x = (string)$3. Next
  it checks if class $2 has an accessible static property named x. If it
  doesn't, this instruction throws a fatal error. Otherwise, this instruction
  reads the static property named x into y, executes z = y <op> $1, assigns z
  into the static property, and then pushes z onto the stack. The immediate
  value must be one of the following opcodes:
    Add, Sub, Mul, Div, Mod, Shl, Shr, Concat, BitAnd, BitOr, BitXor.

IncDecL <local variable id> <op>    []  ->  [C]

  Increment/decrement local. If the local variable given by %1 is not defined,
  this instruction marks it as defined, sets it to null, and raises a warning.

  Where x is the local given by %1, this instruction then does the following:

  If op is PreInc, this instruction executes ++x and then pushes x onto the
  stack as a cell.

  If op is PostInc, this instruction pushes x onto the stack and then it
  executes ++x.

  If op is PreDec, this instruction executes --x and then pushes x onto the
  stack.

  If op is PostDec, this instruction pushes x onto the stack and then it
  executes --x.

IncDecN <op>    [C]  ->  [C]
IncDecG <op>    [C]  ->  [C]

  Increment/decrement. This instruction first computes x = (string)$1. Next, if
  the local variable (IncDecN) or global variable (IncDecG) named x is not
  defined, this instruction first defines it, sets it to null, and raises a
  warning.

  Where v is the local variable or global variable named x, this instruction
  performs the following:

  If op is PreInc, this instruction executes ++v and then pushes v onto the
  stack as a cell.

  If op is PostInc, this instruction pushes v onto the stack and then it
  executes ++v.

  If op is PreDec, this instruction executes --v and then pushes v onto the
  stack.

  If op is PostDec, this instruction pushes v onto the stack and then it
  executes --v.

IncDecS <op>    [C A]  ->  [C]

  Increment/decrement static property. This instruction first computes x =
  (string)$2. Next it checks if class $1 has an accessible static property
  named x. If it doesn't, this instruction throws a fatal error.

  Where s is the static property named x, this instruction performs the
  following:

  If op is PreInc, this instruction increments the ++s and then pushes s onto
  the stack.

  If op is PostInc, this instruction pushes s onto the stack and then it
  executes ++s.

  If op is PreDec, this instruction executes --s and then pushes s onto the
  stack.

  If op is PostDec, this instruction pushes s onto the stack and then it
  executes --s.

BindL <local variable id>    [V]  ->  [V]

  Bind local. This instruction marks the local variable given by %1 as defined,
  binds the local variable to $1, and pushes $1 onto the stack.

BindN    [C V]  ->  [V]

  Bind local. This instruction marks the local variable named (string)$2 as
  defined, binds the local variable to $1, and pushes $1 onto the stack.

BindG    [C V]  ->  [V]

  Bind global. This instruction marks the global variable named (string)$2 as
  defined, binds the global variable to $1, and pushes $1 onto the stack.

BindS    [C A V]  ->  [V]

  Bind static property. This instruction first checks if class $2 has an
  accessible static property named (string)$3. If it doesn't, this instruction
  throws a fatal error. Otherwise, this instruction binds the static property
  to $1, and pushes $1 onto the stack.

UnsetL <local variable id>    []  ->  []

  Unset local. Breaks any bindings the local variable given by %1 may have and
  marks the local variable as undefined.

UnsetN    [C]  ->  []

  Unset local. This instruction breaks any bindings the local variable named
  (string)$1 may have and marks the local variable as undefined.

UnsetG    [C]  ->  []

  Unset global. This instruction breaks any bindings the global variable named
  (string)$1 may have and marks the global variable as undefined.

CheckProp <propName> [] -> [C:Bool]

  Check non-scalar property initializer. This instruction checks the
  initializer for property named %1 in the context class, and pushes
  true on the stack if it is initialized, and false otherwise.

InitProp <propName> <op> [C] -> []

  Initialize non-scalar property. If %2 is 'NonStatic', this instruction sets
  the initializer for the property named %1 in the context class to $1. If %2
  is 'Static', this instruction sets the initializer for the static property
  named %1 in the context class to $1.

  The CheckProp and InitProp opcodes should only be used in 86pinit methods.
  86pinit methods are HHVM-internal property initialization methods that
  cannot be called from user-land. After 86pinit runs, no declared properties
  of the class can be of type NullUninit.

8. Call instructions
--------------------

FPushFunc <num params>                 [C]  ->  []
FPushFuncD <num params> <litstr id>    []  ->  []

  FPI push function. First, these instructions load a value into x as given by
  the following table:

    instruction      x
    --------------+----
      FPushFunc   | $1
      FPushFuncD  | %2

  If x is a string, this instruction attempts to lookup a function named x. If
  no function named x is defined, this instruction throws a fatal error.
  Otherwise this instruction pushes a new entry on the FPI stack, initializing
  it with the number of parameters being passed (given by %1) and a reference
  to the FPI structure for the function named x. With FPushFuncD the litstr in
  %2 must not start with a '\' character. Function names should be normalized
  with respect to namespace and never start with a '\'.

  If x is an object, this instruction checks if the object has an __invoke
  method. If the object does not have an __invoke method, this instruction
  throws a fatal error. Otherwise this instruction pushes a new entry on the
  FPI stack, initializing it with the number of parameters being passed (given
  by %1) and a reference to the FPI structure for the __invoke method from
  object x.

  If x is not a string or object, this instruction throws a fatal error.

FPushFuncU <num params> <litstr id> <litstr fallback> []  ->  []

  FPI push function unqualified. Identical to FPushFuncD except first tries to
  lookup the function named %2 and if it isn't defined calls the function named
  %3. As for FPushFuncD the litstr in %2 and %3 must not start with a '\'
  %character.

FPushObjMethod <num params>                 [C C]  ->  []
FPushObjMethodD <num params> <litstr id>    [C]  ->  []

  FPI push object-based method. First, these instructions load values into x
  and y as given by the following table:

    instruction           x    y
    -------------------+----+-----
      FPushObjMethod   | $2 | $1
      FPushObjMethodD  | $1 | %2

  If x is not an object or if y is not a string, this instruction throws a
  fatal error. Next, this instruction checks if object x has an accessible
  method named y. If it does, this instruction pushes a new entry on the FPI
  stack, initializing it with the number of parameters being passed (given by
  %1) and a reference to the FPI structure for the method named y from object
  %x.

  If object x does not have an accessible method named y, this instruction
  checks if object x has a __call method. If a __call method is found, this
  instruction pushes a new entry on the FPI stack, initializing it with the
  number of parameters being passed (given by %1) and a reference to the FPI
  structure for the __call from object x, and stores the original name y in the
  FPI stack entry.

  If object x does not have an accessible method named y and it does not have a
  __call method, this instruction throws a fatal error.

FPushClsMethod <num params>                             [C A]  ->  []
FPushClsMethodF <num params>                            [C A]  ->  []
FPushClsMethodD <num params> <litstr id> <litstr id>    []  ->  []

  FPI push class-based method. First, these instructions load values into x and
  y as given by the following table:

    instruction           x    y
    -------------------+----+-----
      FPushClsMethod   | $1 | $2
      FPushClsMethodF  | $1 | $2
      FPushClsMethodD  | %3 | %2

  When loading %3 into x, FPushClsMethodD will perform the work performed by
  the AGetC instruction to convert the name given by %3 into a classref.

  If y is not a string, this instruction throws a fatal error. Next, this
  instruction checks if class x has an accessible method named y. If class x
  has a method named y. If it does, this instruction pushes a new entry on the
  FPI stack, initializing it with the number of parameters being passed (given
  by %1) and a reference to the FPI structure for the method named y from class
  x.

  If class x does not have an accessible method named y, this instruction
  checks if the current function's $this is non-null, if the class of $this is
  the same or derived from class x, and if $this has a __call method. If no
  suitable __call method is found, this instruction will check if class x has a
  __callStatic method. If a suitable __call method or a __callStatic method is
  found, this instruction pushes a new entry on the FPI stack, initializing it
  with the number of parameters being passed (given by %1) and a reference to
  the FPI structure for the __call or __callStatic method that was found, and
  stores the original name y in the FPI stack entry.

  If class x does not have an accessible method named y, and if a suitable
  __call method or a __callStatic method could not be found, this instruction
  throws a fatal error.

FPushCtor <num params>                 [A]  ->  [C]
FPushCtorD <num params> <litstr id>    []  ->  [C]

  FPI push constructor. First, these instructions load a value into x as given
  by the following table:

    instruction     x
    --------------+----
      FPushCtor   | $1
      FPushCtorD  | %2

  When loading %2 into x, FPushCtorD will perform the work performed by the
  AGetC instruction to convert the name given by %2 into a classref.

  This instruction pushes an uninitialized object onto the stack (to be
  initialized during FCall*) prior to entering the FPI region, then pushes a
  new entry on the FPI stack, initializing it with the number of parameters
  being passed (given by %1) and a reference to the FPI structure for the
  constructor for class x.

DecodeCufIter <iterator id>    [C]  ->  []

  This instruction looks up $1 as a callable, and writes enough information to
  iterator %1 for FPushCufIter to be able to push an actrec, as if it had been
  given the callable. If the function is not successfully decoded, sets up iter
  to call a function that does nothing and returns Null. No warning is raised.

FPushCufIter <num params> <iterator id>   []  ->  []

  FPI push the result of a previous DecodeCufIter. No warning is raised.

FPushCuf <num params>     [C]   ->  []
FPushCufF <num params>    [C]   ->  []

  FPI push call user function. These instructions lookup $1 as a callable, and
  push a new entry onto the FPI stack. If $1 is not callable, they issue a
  warning, and push an entry representing a function which does nothing, takes
  no argument, and returns null.

FPushCufSafe <num params> [C C] ->  [C C:Bool]

  FPI push call user function. This instruction pops $1 and $2, then pushes $1
  back onto the stack. It then looks up $2 as a callable, and pushes a new
  entry onto the FPI stack. If $2 is not callable, it pushes an entry
  representing a function which does nothing, takes no argument, and returns
  null, and in addition pushes boolean false onto the evaluation stack;
  otherwise it pushes true onto the evaluation stack.

CufSafeArray   [C C R] -> [C]

  Pops 3 elements from the stack, and pushes array($2, $1), preserving refs.

CufSafeReturn  [C C R] -> [R]

  Pops 3 elements from the stack, and pushes $2 ? $1 : $3, preserving refs.

FPassC <param id>     [C]  ->  [F]
FPassCW <param id>    [C]  ->  [F]
FPassCE <param id>    [C]  ->  [F]

  FPI pass parameter. This instruction pushes $1 onto the stack as a cell
  regardless of whether parameter %1 is pass by value or pass by reference.

  If parameter %1 is pass by reference, FPassCW and FPassCE check if the
  function associated with the current FPI (the callee) is an extension
  function that can accept a cell for parameter %1. If this condition is not
  met, FPassCW will raise a warning while FPassCE will throw a fatal error.

FPassV <param id>    [V]  ->  [F]

  FPI pass parameter. If parameter %1 is pass by value, this instruction will
  unbox $1 and push it onto the stack as a cell. If parameter %1 is pass by
  reference, this instruction will push $1 onto the stack as a ref.

FPassVNop <param id> [V]  ->  [F]

  FPI pass parameter, no op. Parameter %1 must be statically known to be pass
  by reference. This instruction pushes $1 onto the stack as a ref.

FPassR <param id>    [R]  ->  [F]

  FPI pass parameter. If $1 is a cell at run time, this instruction will behave
  like FPassC. Otherwise, this instruction will behave like FPassV.

FPassL <param id> <local variable id>    []  ->  [F]

  FPI pass local as parameter. This instruction behaves as CGetL if parameter
  %1 is pass by value, or it behaves like VGetL if parameter %1 is pass by
  reference.

FPassN <param id>    [C]  ->  [F]

  FPI pass local as parameter. This instruction behaves as CGetN if parameter
  %1 is pass by value, or it behaves like VGetN if parameter %1 is pass by
  reference.

FPassG <param id>    [C]  ->  [F]

  FPI pass global as parameter. This instruction behaves as CGetG if parameter
  %1 is pass by value, or it behaves like VGetG if parameter %1 is pass by
  reference.

FPassS <param id>    [C A]  ->  [F]

  FPI pass parameter. This instruction behaves as CGetS if parameter %1 is pass
  by value, or it behaves like VGetS if parameter %1 is pass by reference.

FCall <num params>    [F..F]  ->  [R]

  FPI call. This instruction gets the bytecode address of the function
  associated with the current FPI (the callee), transfers the top %1 values
  from the stack to the callee as parameters, pops the current FPI off of the
  FPI stack, and then invokes the dispatcher to call the callee. When the
  callee returns, it will transfer the return value onto the caller's
  evaluation stack using the R flavor.

FCallD <num params> <class name> <func name>   [F..F]  ->  [R]

  FPI call direct. This instruction has exactly the effects of FCall %1, but
  provides hints from static analysis to assist the region selector in
  determining the callee. The strings in %2 and %3 are statically known names
  of the class (if any) and method being called. If the call is targeting a
  non-method, %2 must be the empty string.

FCallArray    [F]  ->  [R]

  FPI call with array. This instruction gets the bytecode address of the
  function associated with the current FPI (the callee), transfers the elements
  of $1 (which must be an array) to the callee as parameters, pops the current
  FPI off of the FPI stack, and then invokes the dispatcher to call the callee.
  When the callee returns, it will transfer the return value onto the caller's
  evaluation stack using the R flavor.

FCallBuiltin  <total params> <passed params> <litstr id> [C|V|U..C|V|U]  -> [R]

  Optimized builtin call without an ActRec. This instruction attempts to lookup
  a builtin function named %3. If no function named %3 is defined, this
  instruction throws a fatal error. Otherwise, this function gets address of
  the builtin function named %3, transfers the top %1 values from the stack to
  the callee as parameters, and then invokes the dispatcher to call the callee.
  %2 denotes the number of non-default parameters pushed onto stack by user
  level code. When the callee returns, it will transfer the return value onto
  the caller's evaluation stack using the R flavor.


9. Member operations
--------------------

The following operations describe processes that are shared across the Member
instructions. Operations are not considered instructions; they do not have
opcodes associated with them.

Operations can produce and consume intermediate values called "bases". A "base"
is a structure that contains either a cell or a ref or a reference to a memory
location that is occupied by a cell or a ref. Bases are never pushed onto the
evaluation stack.

For operations that create a base, the operation descriptions specify whether
the base created "contains" a value or "references" a location. In the former
case, the base created contains a cell or a ref. In the latter case, the base
created contains a reference to a memory location occupied by a cell or a ref.

When a base that contains a cell is destroyed, if the cell points to data then
the execution engine is responsible for honoring the data's refcount logic.
Likewise when a base that contains a ref is destroyed, the execution engine is
responsible for honoring the refcount logic of the cell container pointed to by
the ref. When a base that contains a reference to a memory location occupied by
a cell or a ref is destroyed, no refcounting is required.

Some operations that take a base as input can modify that base as part of the
work performed by the operation. Such operations are said to "set" the base to
a new value. When a base that contains a cell or a reference to a memory
location occupied by a cell is set to a new value, the new value overwrites the
previous value contained in the cell (honoring the data refcount logic if the
previous value was a refcounted type). When a base that contains a ref or a
reference to a memory location occupied by a ref is set to the new value, the
new value is written into the cell container referenced by the ref, overwriting
the previous cell in that container (honoring the data refcount logic if the
previous cell was a refcounted type). Note that for bases that contain a
reference to a memory location, "setting" the base does not change which memory
location the base references.

Operations are specified as if they directly operate on the top of the
evaluation stack in the name of consistency and clarity, but in fact their
inputs and outputs may reside elsewhere. The symbol 'B' is used in the input
descriptions and output descriptions of operations to indicate that a given
operation consumes a base as input or produces a base as output.

BaseC     [C]  ->  [B]

  Get base from value. This operation outputs a base that contains the value
  given by $1.

BaseR     [R]  ->  [B]

  Get base from return value. This operation outputs a base that contains the
  return value given by $1.

BaseL <local variable id>    []  ->  [B]

  Get base from local. This operation outputs a base that references the local
  given by %1. If the local is not defined, this operation outputs a base that
  contains null.

BaseLW <local variable id>    []  ->  [B]

  Get base from local. This operation outputs a base that references the local
  given by %1. If the local is not defined, this operation raises a warning and
  outputs a base that contains null.

BaseLD <local variable id>    []  ->  [B]

  Get base from local. This operation outputs a base that references the local
  given by %1. If the local is not defined, this operation defines it and
  returns a base that references the local.

BaseLWD <local variable id>    []  ->  [B]

  Get base from local. This operation outputs a base that references the local
  variable given by %1. If the local is not defined, this operation defines it,
  raises a warning, and returns a base that references the local.

BaseNC                       [C]  ->  [B]
BaseNL <local variable id>   []  ->  [B]

  Get base from name. This operation outputs a base that references the local
  variable whose name is given by (string)%1 or (string)$1. If the local is not
  defined, this operation outputs a base that contains null.

BaseNCW                      [C]  ->  [B]
BaseNLW <local variable id>  []  ->  [B]

  Get base from name. This operation outputs a base that references the local
  variable whose name is given by (string)%1 or (string)$1. If the local is not
  defined, this operation raises a warning and outputs a base that contains
  null.

BaseNCD                      [C]  ->  [B]
BaseNLD <local variable id>  []  ->  [B]

  Get base from name. This operation outputs a base that references the local
  variable whose name is given by (string)%1 or (string)$1. If the local is not
  defined, this operation defines it and returns a base that references the
  local.

BaseNCWD                        [C]  ->  [B]
BaseNLWD <local variable id>    []  ->  [B]

  Get base from name. This operation outputs a base that references the local
  variable whose name is given by (string)%1 or (string)$1. If the local is not
  defined, this operation defines it, raises a warning, and returns a base that
  references the local.

BaseGC                        [C]  ->  [B]
BaseGL <local variable id>    []  ->  [B]

  Get base from global name. This operation outputs a base that references the
  global variable whose name is given by (string)%1 or (string)$1. If the
  global is not defined, this operation produces a base that contains null.

BaseGCW                        [C]  ->  [B]
BaseGLW <local variable id>    []  ->  [B]

  Get base from global name. This operation outputs a base that references the
  global variable whose name is given by (string)%1 or (string)$1. If the
  global is not defined, this operation raises a warning and outputs a base
  that contains null.

BaseGCD                        [C]  ->  [B]
BaseGLD <local variable id>    []  ->  [B]

  Get base from global name. This operation outputs a base that references the
  global variable whose name is given by (string)%1 or (string)$1. If the
  global is not defined, this operation defines it and returns a base that
  references the global.

BaseGCWD                        [C]  ->  [B]
BaseGLWD <local variable id>    []  ->  [B]

  Get base from global name. This operation outputs a base that references the
  global variable whose name is given by (string)%1 or (string)$1. If the
  global is not defined, this operation defines it, raises a warning, and
  returns a base that references the global.

BaseSC                        [C A]  ->  [B]
BaseSL <local variable id>    [A]  ->  [B]

  Get base from static property. First, this operation loads a value into x as
  given by the following table:

       operation     x
       -----------+----
         BaseSC   | $2
         BaseSL   | %1

  Next this operation computes y = (string)x. Then this instruction checks if
  class $1 has an accessible property named y. If it does, this operation
  outputs a base that references the static property. Otherwise, this operation
  throws a fatal error.

BaseH     []  ->  [B]

  Get base from $this. This operation assumes that the current frame contains a
  valid $this pointer and outputs a base containing the object in $this.

ElemC                        [C B]  ->  [B]
ElemL <local variable id>    [B]  ->  [B]

  Fetch element if it exists. First, these operations load a value into x and a
  base into y, as given by the following table:

       operation    x    y
       ----------+----+-----
         ElemC   | $2 | $1
         ElemL   | %1 | $1

  Then, if y is an array, this operation outputs a base that references the
  element at index x from array y. If there is no element at index x, this
  operation outputs a base that contains null.

  If y is an object that implements the ArrayAccess interface, this operation
  outputs a base that contains the result of y->offsetGet(x).

  If y is an object that does not implement the ArrayAccess interface, this
  operation throws a fatal error.

  If y is a string, this operation computes z = (int)x. If z >= 0 and z <
  strlen(z), this operation builds a new string consisting of the character at
  offset z from y and outputs a base that contains the new string. Otherwise,
  this operation outputs a base that contains the empty string.

  If y is not a string, array, or object, this operation will output a null
  base.

ElemCW                        [C B]  ->  [B]
ElemLW <local variable id>    [B]  ->  [B]

  Fetch element; warn if it doesn't exist.

  First, these operations load a value into x and a base into y, as given by
  the following table:

       operation    x    y
       ----------+----+-----
         ElemCW  | $2 | $1
         ElemLW  | %1 | $1

  If y is an array, this operation outputs a base that references the element
  at index x from array y. If there is no element at index x, this operation
  outputs a base that contains null and raises a warning.

  If y is an object that implements the ArrayAccess interface, this operation
  outputs a base that contains the result of y->offsetGet(x).

  If y is an object that does not implement the ArrayAccess interface, this
  operation throws a fatal error.

  If y is a string, this operation continues to compute z = (int)x. If z >= 0
  and z < strlen(z), this operation builds a new string consisting of the
  character at offset z from y and outputs a base that contains the new string.
  Otherwise, this operation raises a warning and outputs a base that contains
  the empty string.

  If y is not a string, array, or object, this operation will output a null
  base.

ElemCD                        [C B]  ->  [B]
ElemLD <local variable id>    [B]  ->  [B]

  Fetch element; define it if it doesn't exist.

  First, these operations load a value into x and a base into y, as given by
  the following table:

       operation    x    y
       ----------+----+-----
         ElemCD  | $2 | $1
         ElemLD  | %1 | $1

  If y is an array, this operation outputs a base that references the element
  at index x. If there is no element at index x, this operation creates an
  element at index x, and outputs a base that references the element.

  If y is an object that implements the ArrayAccess interface, this operation
  outputs a base that contains the result of y->offsetGet(x).

  If y is non-empty string or an object that does not implement the ArrayAccess
  interface, this operation throws a fatal error.

  If y is null, the empty string, or false, this operation will set y to a new
  empty array, create an element at index x, and output a base that references
  the element.

  If y is true, integer, double, this operation raises a warning and outputs a
  base that contains null.

ElemCWD                        [C B]  ->  [B]
ElemLWD <local variable id>    [B]  ->  [B]

  Fetch element; warn and define it if it doesn't exist.

  First, these operations load a value into x and a base into y, as given by
  the following table:

       operation    x    y
       ----------+----+-----
         ElemCWD | $2 | $1
         ElemLWD | %1 | $1

  If y is an array, this operation outputs a base that references the element
  at index x. If there is no element at index x, this operation creates an
  element at index x, raises a warning, and outputs a base that references the
  element.

  If y is an object that implements the ArrayAccess interface, this operation
  outputs a base that contains the result of y->offsetGet(x).

  If y is non-empty string or an object that does not implement the ArrayAccess
  interface, this operation throws a fatal error.

  If y is null, the empty string, or false, this operation will set y to a new
  empty array, create an element at index x, and output a base that references
  the element.

  If y is true, integer, or double, this operation raises a warning and outputs
  a base that contains null.

ElemCU                        [C B]  ->  [B]
ElemLU <local variable id>    [B]  ->  [B]

  Fetch element for unset.

  First, these operations load a value into x and a base into y, as given by
  the following table:

       operation    x    y
       ----------+----+-----
         ElemCU  | $2 | $1
         ElemLU  | %1 | $1

  If y is an array, this operation outputs a base that references the element
  at index x from array y. If there is no element at index x, this operation
  outputs a base that contains null.

  If y is an object that implements the ArrayAccess interface, this operation
  outputs a base that contains the result of y->offsetGet(x).

  If y is an object that does not implement the ArrayAccess interface, this
  operation throws a fatal error.

  If y is a string, this operation throws a fatal error.

  If y is not a string, array, or object, this operation will output a null
  base.

NewElem    [B]  ->  [B]

  Fetch new element. If $1 is an array, this operation creates a new element
  with the next available numeric key in array $1 and outputs a base that
  references the new element.

  If $1 is an object that implements the ArrayAccess interface, this operation
  outputs a base that contains the result of $1->offsetGet(null).

  If $1 is a non-empty string or an object that does not implement the
  ArrayAccess interface, this operation throws a fatal error.

  If $1 is null, false, or the empty string, this operation sets $1 to a new
  empty array, creates a new element with the next available numeric key in
  array $1, and then outputs a base that references the new element.

  If $1 is true, integer, or double, this operation raises a warning and
  outputs a base that contains null.

PropC                        [C B]  ->  [B]
PropL <local variable id>    [B]  ->  [B]

  Fetch property if it exists.

  First, these operations load a value into x and a base into y, as given by
  the following table:

       operation    x    y
       ----------+----+-----
         PropC   | $2 | $1
         PropL   | %1 | $1

  Next, performs one of the following actions:

    y is an object
     y->x is visible
      y->x is accessible
       y has eligible __get method
        y->x has been unset previously
    ------+---------------------------------------------------------------------
    0XXXX | push null
    10X0X | push null
    10X1X | push ref(y->__get(x))
    1100X | throw fatal error
    1101X | push ref(y->__get(x))
    111X0 | push ref(y->x)
    11101 | push null
    11111 | push ref(y->__get(x))

PropCW                        [C B]  ->  [B]
PropLW <local variable id>    [B]  ->  [B]

  Fetch property; warn if it doesn't exist.

  First, these operations load a value into x and a base into y, as given by
  the following table:

       operation    x    y
       ----------+----+-----
         PropCW  | $2 | $1
         PropLW  | %1 | $1

  Next, performs one of the following actions:

    y is an object
     y->x is visible
      y->x is accessible
       y has eligible __get method
        y->x has been unset previously
    -----+----------------------------------------------------------------------
    0XXXX | raise warning; push null
    10X0X | raise warning; push null
    10X1X | push ref(y->__get(x))
    1100X | throw fatal error
    1101X | push ref(y->__get(x))
    111X0 | push ref(y->x)
    11101 | raise warning; push null
    11111 | push ref(y->__get(x))

PropCD                        [C B]  ->  [B]
PropLD <local variable id>    [B]  ->  [B]

  Fetch property; define it if it doesn't exist.

  First, these operations load a value into x and a base into y, as given by
  the following table:

       operation    x    y
       ----------+----+-----
         PropCD  | $2 | $1
         PropLD  | %1 | $1

  Next, performs one of the following actions:

    y is an object
     y is null/false/""
      y->x is visible
       y->x is accessible
        y has eligible __get method
         y->x has been unset previously
    ------+---------------------------------------------------------------------
    00XXXX | push null
    01XXXX | y = new stdclass; create property y->x; push ref(y->x)
    1X0X0X | create property y->x; push ref(y->x)
    1X0X1X | push ref(y->__get(x))
    1X100X | throw fatal error
    1X101X | push ref(y->__get(x))
    1X11X0 | push ref(y->x)
    1X1101 | re-create property y->x, push ref(y->x)
    1X1111 | push ref(y->__get(x))

PropCWD                        [C B]  ->  [B]
PropLWD <local variable id>    [B]  ->  [B]

  Fetch property; warn and define it if it doesn't exist.

  First, these operations load a value into x and a base into y, as given by
  the following table:

       operation    x    y
       ----------+----+-----
         PropCWD | $2 | $1
         PropLWD | %1 | $1

  Next, performs one of the following actions:

    y is an object
     y is null/false/""
      y->x is visible
       y->x is accessible
        y has eligible __get method
         y->x has been unset previously
    ------+---------------------------------------------------------------------
    00XXXX | raise warning; push null
    01XXXX | raise warning; y = new stdclass; create property y->x;
           | push ref(y->x)
    1X0X0X | raise warning; create property y->x; push ref(y->x)
    1X0X1X | push ref(y->__get(x))
    1X100X | throw fatal error
    1X101X | push ref(y->__get(x))
    1X11X0 | push ref(y->x)
    1X1101 | re-create property y->x, push ref(y->x)
    1X1111 | push ref(y->__get(x))

PropCU                        [C B]  ->  [B]
PropLU <local variabld id>    [B]  ->  [B]

  Fetch property for unset.

  First, these operations load a value into x and a base into y, as given by
  the following table:

       operation    x    y
       ----------+----+-----
         PropCW  | $2 | $1
         PropLW  | %1 | $1

  Next, performs one of the following actions:

    y is an object
     y->x is visible
      y->x is accessible
       y has eligible __get method
        y->x has been unset previously
    -----+----------------------------------------------------------------------
    0XXXX | push null
    10XXX | create property y->x; push ref(y->x)
    110XX | throw fatal error
    111X0 | push ref(y->x)
    111X1 | re-create property y->x, push ref(y->x)

CGetElemC                        [C B]  ->  [C]
CGetElemL <local variable id>    [B]  ->  [C]

  Get element as cell.

  These instructions first load a value into x and a base into y, as given by
  the following table:

         operation    x    y
       ------------+----+-----
        CGetElemC  | $2 | $1
        CGetElemL  | %1 | $1

  If y is an array, this operation retrieves the element at index x from array
  y and pushes it onto the stack as a cell. If there is no element at index x,
  this operation raises a warning and pushes null onto the stack.

  If y is an object that implements the ArrayAccess interface, this operation
  pushes x->offsetGet($2) onto the stack.

  If y is an object that does not implement the ArrayAccess interface, this
  operation throws a fatal error.

  If y is a string, this operation continues to compute z = (int)x. If z >= 0
  and z < strlen(z), this operation builds a new string consisting of the
  character at offset z from y and pushes it onto the stack. Otherwise, this
  operation raises a warning and pushes the empty string onto the stack.

  If y is not a string, array, or object, this operation will push null onto
  the stack.

VGetElemC                        [C B]  ->  [V]
VGetElemL <local variable id>    [B]  ->  [V]

  Get element as ref.

  These instructions first load a value into x and a base into y, as given by
  the following table:

         operation    x    y
       ------------+----+-----
        VGetElemC  | $2 | $1
        VGetElemL  | %1 | $1

  If y is an array, this operation retrieves the element at index x from array
  y and pushes it onto the stack as a ref. If there is no element at index x,
  this operation creates a new element at index x, and pushes it onto the stack
  as a ref.

  If y is an object that implements the ArrayAccess interface, this operation
  pushes y->offsetGet(x) onto the stack as a ref.

  If y is a non-empty string or an object that does not implement the
  ArrayAccess interface, this operation throws a fatal error.

  If y is null, false, or the empty string, this operation sets y to a new
  empty array. Then this operation retrieves the element at index x from array
  y and pushes it onto the stack as a ref. If there is no element at index x,
  this operation creates a new element at index x, and pushes it onto the stack
  as a ref.

  If y is true, integer, or double, this operation raises a warning and pushes
  null onto the stack.

IssetElemC                        [C B]  ->  [C:Bool]
IssetElemL <local variable id>    [B]  ->  [C:Bool]

  Isset element.

  These instructions first load a value into x and a base into y, as given by
  the following table:

         operation    x    y
       ------------+----+-----
        IssetElemC | $2 | $1
        IssetElemL | %1 | $1

  If y is an array, this operation pushes !is_null(y[x]) onto the stack.

  If y is an object that implements the ArrayAccess interface, this operation
  pushes y->offsetExists(x) onto the stack.

  If y is an object that does not implement the ArrayAccess interface, this
  operation throws a fatal error.

  If y is a string, this operation computes x = (int)x and then it pushes (x >=
  0 && x < strlen(y)) onto the stack.

  If y is a not a string, array, or object, this operation pushes false onto
  the stack.

EmptyElemC                        [C B]  ->  [C]
EmptyElemL <local variable id>    [B]  ->  [C]

  Empty element.

  These instructions first load a value into x and a base into y, as given by
  the following table:

         operation    x    y
       ------------+----+-----
        EmptyElemC | $2 | $1
        EmptyElemL | %1 | $1

  If y is an array, this operation pushes !(y[x]) onto the stack.

  If y is an object that implements the ArrayAccess interface, this operation
  first calls y->offsetExists(x); if that returns false this operation pushes
  true onto the stack, otherwise it pushes !(y->offsetGet(x)) onto the stack.

  If y is an object that does not implement the ArrayAccess interface, this
  operation throws a fatal error.

  If y is a string, this operation computes z = (int)x, then pushes true if (z
  < 0 || z >= strlen(y)), !(y[z]) otherwise.

  If y is, not an array, object, or string, this operation pushes true onto the
  stack.

SetElemC    [C C B]  ->  [C]

  Set element. If $1 is an array, this operation executes $1[$3] = $2 and then
  pushes $2 onto the stack.

  If $1 is an object that implements the ArrayAccess interface, this operation
  executes $1->offsetSet($3, $2) and then pushes $2 onto the stack.

  If $1 is an object that does not implement the ArrayAccess interface, this
  operation throws a fatal error.

  If $1 is null, the empty string, or false, this operation sets $1 to a new
  empty array, executes $1[$3] = $2, and then pushes $2 onto the stack.

  If $1 is a non-empty string, this operation first computes x = (int)$3. If x
  is negative, this operation raises a warning and does nothing else. If x is
  non-negative, this operation appends spaces to the end of $1 as needed to
  ensure that x is in bounds, then it computes y = substr((string)$2,0,1), and
  then it sets the character at index x in $1 equal to y (if y is not empty) or
  it sets the character at index x in $1 to "\0" (if y is empty). Then this
  operation pushes y on to the stack.

  If $1 is true, integer, or double, this operation raises a warning and pushes
  null onto the stack as a cell.

SetElemL <local variable id>    [C B]  ->  [C]

  Set element. If $1 is an array, this operation executes $1[%1] = $2 and then
  pushes $2 onto the stack.

  If $1 is an object that implements the ArrayAccess interface, this operation
  executes $1->offsetSet(%1, $2) and then pushes $2 onto the stack.

  If $1 is an object that does not implement the ArrayAccess interface, this
  operation throws a fatal error.

  If $1 is null, the empty string, or false, this operation sets $1 to a new
  empty array, executes $1[%1] = $2, and then pushes $2 onto the stack.

  If $1 is a non-empty string, this operation first computes x = (int)%1. If x
  is negative, this operation raises a warning and does nothing else. If x is
  non-negative, this operation appends spaces to the end of $1 as needed to
  ensure that x is in bounds, then it computes y = substr((string)$2,0,1), and
  then it sets the character at index x in $1 equal to y (if y is not empty) or
  it sets the character at index x in $1 to "\0" (if y is empty). Then this
  operation pushes y on to the stack.

  If $1 is true, integer, or double, this operation raises a warning and pushes
  null onto the stack as a cell.

SetOpElemC <op>    [C C B]  ->  [C]

  Set element op. If $1 is an array, this operation first checks $1 contains an
  element at offset $2. If it does not, this operation creates an element at
  offset $2, sets it to null, and raises a warning. Next, this operation
  executes x = $1[$3], y = x <op> $2, and $1[$3] = y, and then it pushes y onto
  the stack as a cell.

  If $1 is null, false, or the empty string, this operation first sets $1 to a
  new empty array. Then it follows the rules described in the case above.

  If $1 is an object that implements the ArrayAccess interface, this operation
  executes x = $1->offsetGet($3), y = x <op> $2, and $1->offsetSet($3, y), and
  then it pushes y onto the stack as a cell.

  If $1 is a non-empty string or an object that does not implement the
  ArrayAccess interface, this operation throws a fatal error.

  If $1 is true, integer, or double, this operation raises a warning and pushes
  null onto the stack.

SetOpElemL <op> <local variable id>    [C B]  ->  [C]

  Set element op. If $1 is an array, this operation first checks $1 contains an
  element at offset $2. If it does not, this operation creates an element at
  offset $2, sets it to null, and raises a warning. Next, this operation
  executes x = $1[%1], y = x <op> $2, and $1[%1] = y, and then it pushes y onto
  the stack as a cell.

  If $1 is null, false, or the empty string, this operation first sets $1 to a
  new empty array. Then it follows the rules described in the case above.

  If $1 is an object that implements the ArrayAccess interface, this operation
  executes x = $1->offsetGet(%1), y = x <op> $2, and $1->offsetSet(%1, y), and
  then it pushes y onto the stack as a cell.

  If $1 is a non-empty string or an object that does not implement the
  ArrayAccess interface, this operation throws a fatal error.

  If $1 is true, integer, or double, this operation raises a warning and pushes
  null onto the stack.

IncDecElemC <op>    [C B]  ->  [C]

  Increment/decrement element. If $1 is an array, this operation checks if $1
  contains an element at offset $2. If it does not, this operation creates an
  element at offset $2, sets it to null, and raises a warning. Next, this
  operation executes x = $1[$2], y = x, and either ++y (if op is PreInc or
  PostInc) or --y (if op is PreDec or PostDec). Then it assigns y to $1[$2] and
  pushes either y (if op is PreInc or PreDec) or x (if op is PostInc or
  PostDec) onto the stack.

  If $1 is null, false, or the empty string, this operation first sets $1 to an
  empty array. Then it follows the rules described in the case above.

  If $1 is a non-empty string or an object that does not implement the
  ArrayAccess interface, this operation throws a fatal error.

  If $1 is an object that implements ArrayAccess, this operation executes x =
  $1->offsetGet($2), y = x, and either ++y (if op is PreInc or PostInc) or --y
  (if op is PreDec or PostDec). Then it pushes either y (if op is PreInc or
  PreDec) or x (if op is PostInc or PostDec) onto the stack.

  If $1 is true, integer, or double, this operation raises a warning and pushes
  null onto the stack.

IncDecElemL <op> <local variable id>    [B]  ->  [C]

  Increment/decrement element. If $1 is an array, this operation checks if $1
  contains an element at offset %1. If it does not, this operation creates an
  element at offset %1, sets it to null, and raises a warning. Next, this
  operation executes x = $1[%1], y = x, and either ++y (if op is PreInc or
  PostInc) or --y (if op is PreDec or PostDec). Then it assigns y to $1[%1] and
  pushes either y (if op is PreInc or PreDec) or x (if op is PostInc or
  PostDec) onto the stack.

  If $1 is null, false, or the empty string, this operation first sets $1 to an
  empty array. Then it follows the rules described in the case above.

  If $1 is a non-empty string or an object that does not implement the
  ArrayAccess interface, this operation throws a fatal error.

  If $1 is an object that implements ArrayAccess, this operation executes x =
  $1->offsetGet(%1), y = x, and either ++y (if op is PreInc or PostInc) or --y
  (if op is PreDec or PostDec). Then it pushes either y (if op is PreInc or
  PreDec) or x (if op is PostInc or PostDec) onto the stack.

  If $1 is true, integer, or double, this operation raises a warning and pushes
  null onto the stack.

BindElemC                        [C V B]  ->  [V]
BindElemL <local variable id>    [V B]  ->  [V]

  Bind element.

  This instruction first loads a value into x, from $3 or the local referred to
  by %1.

  If $1 is an array, this operation executes $1[x] =& $2 and pushes $2 onto the
  stack as a ref.

  If $1 is an object, this operation throws a fatal error.

  If $1 is null, false, or the empty string, this operation sets $1 to a new
  empty array, executes $1[x] =& $2, and pushes $2 onto the stack as a ref.

  If $1 is a non-empty string, this operation throws a fatal error.

  If $1 is true, integer, or double, this operation raises a warning.

UnsetElemC                        [C B]  ->  []
UnsetElemL <local variable id>    [B]  ->  []

  Unset element.

  These instructions first load a value into x and a base into y, as given by
  the following table:

         operation    x    y
       ------------+----+-----
        UnsetElemL | %1 | $1
        UnsetElemC | $2 | $1

  If y is an array, this operation removes the element at index x from array y.

  If y is an object that implements ArrayAccess interface, this operation
  executes y->offsetUnset(x).

  If y is an object that does not implement the ArrayAccess interface, this
  operation throws a fatal error.

  If y is a string, this operation throws a fatal error.

  If y is not a string, array, or object, this operation does nothing.

VGetNewElem    [B]  ->  [V]

  Get new element as ref.

  If $1 is an array, this operation creates a new element with the next
  available numeric key in array $1 and pushes it onto the stack as a ref.

  If $1 is an object that implements the ArrayAccess interface, this operation
  pushes $1->offsetGet($2) onto the stack as a ref.

  If $1 is a non-empty string or an object that does not implement the
  ArrayAccess interface, this operation throws a fatal error.

  If $1 is null, false, or the empty string, this operation first sets $1 to a
  new empty array. Then it creates a new element with the next available
  numeric key in array $1 and pushes it onto the stack as a ref.

  If $1 is true, integer, or double, this operation raises a warning and pushes
  null onto the stack.

SetNewElem    [C B]  ->  [C]

  Set new element. If $1 is an array, this operation executes $1[] = $2 and
  then pushes $2 onto the stack.

  If $1 is null, false, or the empty string, this operation sets $1 to a new
  empty array, and then it executes $1[] = $2 and pushes $2 onto the stack.

  If $1 is a non-empty string or an object that does not implement the
  ArrayAccess interface, this operation throws a fatal error.

  If $1 is an object that implements the ArrayAccess interface, this operation
  executes $1->offsetSet(null, $2) and then pushes $2 onto the stack.

  If $1 is true, integer, or double, this operation raises a warning and pushes
  null onto the stack.

SetOpNewElem <op>    [C B]  ->  [C]

  Set op new element. If $1 is an array, this operation first determines the
  next available integer offset k in array $1. Next, this operation executes
  $1[k] = null, x = $1[k], and y = x <op> $2. Then it assigns y to $1[k] and
  pushes y onto the stack.

  If $1 is null, false, or the empty string, this operation first sets $1 to an
  empty array. Then it follows the rules described in the case above.

  If $1 is a non-empty string or an object that does not implement the
  ArrayAccess interface, this operation throws a fatal error.

  If $1 is an object that implements ArrayAccess, this operation executes x =
  $1->offsetGet(null), y = x <op> $2, and $1->offsetSet(null, y). Then it
  pushes y onto the stack.

  If $1 is true, integer, or double, this operation raises a warning and pushes
  null onto the stack.

IncDecNewElem <op>    [B]  ->  [C]

  Increment/decrement new element. If $1 is an array, this operation first
  determines the next available integer offset k in array $1. Next, this
  operation executes $1[k] = null, x = $1[k], y = x, and either ++y (if op is
  PreInc or PostInc) or --y (if op is PreDec or PostDec). Then it assigns y to
  $1[k] and pushes either y (if op is PreInc or PreDec) or x (if op is PostInc
  or PostDec) onto the stack.

  If $1 is null, false, or the empty string, this operation first sets $1 to an
  empty array. Then it follows the rules described in the case above.

  If $1 is a non-empty string or an object that does not implement the
  ArrayAccess interface, this operation throws a fatal error.

  If $1 is an object that implements ArrayAccess, this operation executes x =
  $1->offsetGet(null), y = x, and either ++y (if op is PreInc or PostInc) or
  --y (if op is PreDec or PostDec). Then it pushes either y (if op is PreInc or
  PreDec) or x (if op is PostInc or PostDec) onto the stack.

  If $1 is true, integer, or double, this operation raises a warning and pushes
  null onto the stack.

BindNewElem    [V B]  ->  [V]

  Bind new element. If $1 is an array, this operation executes $1[] =& $2 and
  then it pushes $2 onto the stack.

  If $1 is null, false, or empty string, this operation sets $1 to a new empty
  array, executes $1[] =& $2, and pushes $2 onto the stack.

  If $1 is a non-empty string or an object, this operation throws a fatal
  error.

  If $1 is true, integer, or double, this operation raises a warning and pushes
  null onto the stack.

CGetPropC                        [C B]  ->  [C]
CGetPropL <local variable id>    [B]  ->  [C]

  Get property as cell.

  These instructions first load a value into x and a base into y, as given by
  the following table:

         operation    x    y
       ------------+----+-----
        CGetPropC  | $2 | $1
        CGetPropL  | %1 | $1

  If y is an object that does not have an eligible __get method, this operation
  first checks if y has a visible property named x. If it does not, this
  operation raises a warning and pushes null. Otherwise, this operation
  continues to check if the property named x is accessible. If the property
  named x is accessible this operation pushes it onto the stack as a cell,
  otherwise this operation throws a fatal error.

  If y is an object that has an eligible __get method, this operation checks if
  y has a visible and accessible property named x. If it does, this operation
  pushes the property onto the stack. Otherwise, this operation pushes
  y->__get(x) onto the stack.

  If y is not an object, this operation will raise a warning and push null onto
  the stack.

VGetPropC                        [C B]  ->  [V]
VGetPropL <local variable id>    [B]  ->  [V]

  Get property as ref.

  These instructions first load a value into x and a base into y, as given by
  the following table:

         operation    x    y
       ------------+----+-----
        VGetPropC  | $2 | $1
        VGetPropL  | %1 | $1

  If y is an object that does not have an eligible __get method, this operation
  first checks if y has a visible property named x. If it does not, this
  operation will create a new property named x and push it onto the stack as a
  ref. Otherwise this operation continues to check if the property named x is
  accessible. If it the property named x is accessible this operation pushes it
  onto the stack as a ref, otherwise this operation throws a fatal error.

  If y is an object has an eligible __get method, this operation checks if y
  has a visible and accessible property named x. If it does, this operation
  pushes the property onto the stack. Otherwise, this operation pushes
  y->__get(x) onto the stack.

  If y is null, false, or the empty string, this operation will set y to a new
  object of type stdclass, create a new property named x, and pushes it onto
  the stack.

  If y is true, integer, double, a non-empty string, or an array, this
  operation raises a warning and pushes null.

IssetPropC                        [C B]  ->  [C:Bool]
IssetPropL <local variable id>    [B]  ->  [C:Bool]

  Isset property.

  These instructions first load a value into x and a base into y, as given by
  the following table:

         operation     x    y
       -------------+----+-----
        IssetPropC  | $2 | $1
        IssetPropL  | %1 | $1

  If y is an object that does not have an eligible __isset method, this
  operation checks if y has a visible accessible property named x. If it does,
  this operation pushes !is_null(y->x) onto the stack. Otherwise this operation
  pushes false onto the stack.

  If y is an object that has an eligible __isset method, this operation checks
  if y has a visible and accessible property named x. If it does, this
  operation pushes !is_null(y->x) onto the stack. Otherwise this operation
  pushes y->__isset(x) onto the stack.

  If y is an array, this operation pushes !is_null(y[x]) onto the stack.

  If y is not an object or array, this operation pushes false.

EmptyPropC                        [C B]  ->  [C:Bool]
EmptyPropL <local variable id>    [B]  ->  [C:Bool]

  Empty property.

  These instructions first load a value into x and a base into y, as given by
  the following table:

         operation     x    y
       -------------+----+-----
        EmptyPropC  | $2 | $1
        EmptyPropL  | %1 | $1

  If y is an object that does not have an eligible __isset method, this
  operation first checks if y has a visible and accessible property named x.
  If it does, this operation pushes !(y->x) onto the stack. Otherwise this
  operation pushes true onto the stack.

  If y is an object that has an eligible __isset method but it does not have an
  eligible __get method, this operation checks if y has a visible and
  accessible property named x. If it does, this operation pushes !(y->x) onto
  the stack. Otherwise this operation pushes !(y->__isset(x)) onto the stack.

  If y is an object that has an eligible __isset method and an eligible __get
  method, this operation checks if y has a visible and accessible property
  named x. If it does, this operation pushes !(y->x) onto the stack. Otherwise
  this operation continues to execute x = y->__isset(x). If x is false this
  operation pushes true onto the stack, otherwise this operation pushes
  !(y->__get(x)) onto the stack.

  If y is an array, this operation pushes !(y[x]) onto the stack.

  If y is not an object or array, this operation pushes true.

SetPropC                        [C C B]  ->  [C]
SetPropL <local variable id>    [C B]  ->  [C]

  Set property. Perform one of the following actions:

  First, these operations load values into k and x, and a base into y, as given
  by the following table:

       operation    k    x    y
       ----------+----+----+----
        SetPropC | $3 | $2 | $1
        SetPropL | %1 | $2 | $1

  Next, performs one of the following actions:

    y is an object
     y is null/false/""
      y->k is visible
       y->k is accessible
        y has eligible __set method
         y->k has been unset previously
    ------+---------------------------------------------------------------------
    00XXXX | raise warning; push null
    01XXXX | y = new stdclass; y->k = x; push x
    1X0X0X | create property y->k; y->k = x; push x
    1X0X1X | y->__set(k, x); push x
    1X100X | throw fatal error
    1X101X | y->__set(k, x); push x
    1X11X0 | y->k = x; push x
    1X1101 | re-create property y->k; y->k = x; push x
    1X1111 | y->__set(k, x); push x

SetOpPropC <op>    [C C B]  ->  [C]

  Set op property. Perform one of the following actions:

    $1 is an object
     $1 is null/false/""
      $1->$3 is visible
       $1->$3 is accessible
        $1 has eligible __get method
         $1 has eligible __set method
          $1->$3 has been unset previously
    -------+--------------------------------------------------------------------
    00XXXXX | raise warning; push null
    01XXXXX | $1 = new stdclass; y = null <op> $2; $1->$3 = y; push y
    100X0XX | y = null <op> $2; $1->$3 = y; push y
    100X10X | x = $1->__get($3); y = x <op> $2; $1->$3 = y; push y
    100X11X | x = $1->__get($3); y = x <op> $2; $1->__set($3, y), push y
    10100XX | throw fatal error
    101010X | throw fatal error
    101011X | x = $1->__get($3); y = x <op> $2; $1->__set($3, y), push y
    1011XX0 | x = $1->$3; y = x <op> $2; $1->$3 = y; push y
    10110X1 | y = null <op> $2; re-create $1->$3; $1->$3 = y; push y
    1011101 | x = $1->__get($3); y = x <op> $2; re-create $1->$3; $1->$3 = y;
            |   push y
    1011111 | x = $1->__get($3); y = x <op> $2; $1->__set($3, y); push y

SetOpPropL <op> <local variable id>    [C B]  ->  [C]

  Set op property. Perform one of the following actions, where k is the value
  of the local given by %2.

    $1 is an object
     $1 is null/false/""
      $1->k is visible
       $1->k is accessible
        $1 has eligible __get method
         $1 has eligible __set method
          $1->k has been unset previously
    -------+--------------------------------------------------------------------
    00XXXXX | raise warning; push null
    01XXXXX | $1 = new stdclass; y = null <op> $2; $1->k = y; push y
    100X0XX | y = null <op> $2; $1->k = y; push y
    100X10X | x = $1->__get(k); y = x <op> $2; $1->k = y; push y
    100X11X | x = $1->__get(k); y = x <op> $2; $1->__set(k, y), push y
    10100XX | throw fatal error
    101010X | throw fatal error
    101011X | x = $1->__get(k); y = x <op> $2; $1->__set(k, y), push y
    1011XX0 | x = $1->k; y = x <op> $2; $1->k = y; push y
    10110X1 | y = null <op> $2; re-create $1->k; $1->k = y; push y
    1011101 | x = $1->__get(k); y = x <op> $2; re-create $1->k; $1->k = y;
            |   push y
    1011111 | x = $1->__get(k); y = x <op> $2; $1->__set(k, y); push y

IncDecPropC <op>    [C B]  ->  [C]

  Increment/decrement property. Perform one of the following actions:

    $1 is an object
     $1 is null/false/""
      $1->$2 is visible
       $1->$2 is accessible
        $1 has eligible __get method
         $1 has eligible __set method
          $1->$2 has been unset previously
    -------+--------------------------------------------------------------------
    00XXXXX | raise warning; push null
    01XXXXX | $1 = new stdclass; x = null; y = x; <op>y; $1->$2 = y;
            |   push y (Pre*) or x (Post*)
    100X0XX | x = null; y = x; <op>y; $1->$2 = y; push y (Pre*) or x (Post*)
    100X10X | x = $1->__get($2); y = x; <op>y; $1->$2 = y;
            |   push y (Pre*) or x (Post*)
    100X11X | x = $1->__get($2); y = x, <op>y; $1->__set($2, y);
            |   push y (Pre*) or x (Post*)
    10100XX | throw fatal error
    101010X | throw fatal error
    101011X | x = $1->__get($2); y = x, <op>y; $1->__set($2, y);
            |   push y (Pre*) or x (Post*)
    1011XX0 | x = $1->$2; y = x; <op>y; $1->$2 = y; push y (Pre*) or x (Post*)
    10110X1 | x = null; y = x; <op>y; re-create $1->$2; $1->$2 = y;
            |   push y (Pre*) or x (Post*)
    1011101 | x = $1->__get($2); y = x; <op>y; re-create $1->$2; $1->$2 = y;
            |   push y (Pre*) or x (Post*)
    1011111 | x = $1->__get($2); y = x; <op>y; $1->__set($2, y);
            |   push y (Pre*) or x (Post*)

IncDecPropL <op> <local variable id>    [B]  ->  [C]

  Increment/decrement property. Perform one of the following actions, where k
  is the value of the local variable given by %2.

    $1 is an object
     $1 is null/false/""
      $1->k is visible
       $1->k is accessible
        $1 has eligible __get method
         $1 has eligible __set method
          $1->k has been unset previously
    -------+--------------------------------------------------------------------
    00XXXXX | raise warning; push null
    01XXXXX | $1 = new stdclass; x = null; y = x; <op>y; $1->k = y;
            |   push y (Pre*) or x (Post*)
    100X0XX | x = null; y = x; <op>y; $1->k = y; push y (Pre*) or x (Post*)
    100X10X | x = $1->__get(k); y = x; <op>y; $1->k = y;
            |   push y (Pre*) or x (Post*)
    100X11X | x = $1->__get(k); y = x, <op>y; $1->__set(k, y);
            |   push y (Pre*) or x (Post*)
    10100XX | throw fatal error
    101010X | throw fatal error
    101011X | x = $1->__get(k); y = x, <op>y; $1->__set(k, y);
            |   push y (Pre*) or x (Post*)
    1011XX0 | x = $1->k; y = x; <op>y; $1->k = y; push y (Pre*) or x (Post*)
    10110X1 | x = null; y = x; <op>y; re-create $1->k; $1->k = y;
            |   push y (Pre*) or x (Post*)
    1011101 | x = $1->__get(k); y = x; <op>y; re-create $1->k; $1->k = y;
            |   push y (Pre*) or x (Post*)
    1011111 | x = $1->__get(k); y = x; <op>y; $1->__set(k, y);
            |   push y (Pre*) or x (Post*)

BindPropC    [C V B]  ->  [V]

  Bind property. If $1 is an object that does not have an eligible __set
  method, this operation first checks if $1 has a visible property named $3. If
  it does not, this operation creates a new property named $3, executes $1->$3
  =& $2, and pushes $2 onto the stack. Otherwise, this operation continues to
  check if the property named $3 is accessible. If the property named $3 is not
  accessible, this operation throws a fatal error. Otherwise, this operation
  executes $1->$3 =& $2 and pushes $2 onto the stack.

  If $1 is an object that has an eligible __set method, this operation checks
  if $1 has a visible and accessible property named $3. If it does, this
  operation follows the rules described in the first case given above.
  Otherwise this operation throws a fatal error.

  If $1 is null, false, or empty string, this operation sets $1 to a new object
  of type stdclass, executes $1->$3 =& $2, and pushes $2 onto the stack.

  If $1 is true, integer, double, a non-empty string, or an array, this
  operation raises a warning and pushes null onto the stack.

BindPropL <local variable id>    [V B]  ->  [V]

  Bind property. Where k is the value of the local variable given by %1:

  If $1 is an object that does not have an eligible __set method, this
  operation first checks if $1 has a visible property named k. If it does not,
  this operation creates a new property named k, executes $1->k =& $2, and
  pushes $2 onto the stack. Otherwise, this operation continues to check if the
  property named k is accessible. If the property named k is not accessible,
  this operation throws a fatal error. Otherwise, this operation executes $1->k
  =& $2 and pushes $2 onto the stack.

  If $1 is an object that has an eligible __set method, this operation checks
  if $1 has a visible and accessible property named k. If it does, this
  operation follows the rules described in the first case given above.
  Otherwise this operation throws a fatal error.

  If $1 is null, false, or empty string, this operation sets $1 to a new object
  of type stdclass, executes $1->k =& $2, and pushes $2 onto the stack.

  If $1 is true, integer, double, a non-empty string, or an array, this
  operation raises a warning and pushes null onto the stack.

UnsetPropC                        [C B]  ->  []
UnsetPropL <local variable id>    [B]  ->  []

  Unset property.

  These instructions first load a value into x and a base into y, as given by
  the following table:

         operation     x    y
       -------------+----+-----
        UnsetPropC  | $2 | $1
        UnsetPropL  | %1 | $1

  Next, performs one of the following actions:

    y is an object
     y->x is visible
      y->x is accessible
       y has eligible __unset method
    -----+----------------------------------------------------------------------
    0XXX | do nothing
    10X0 | do nothing
    10X1 | y->__unset(x)
    1100 | throw fatal error
    1101 | y->__unset(x)
    111X | unset(y->x)


10. Member instructions
-----------------------

Member instructions perform series of operations that are structurally
identical, but each instruction utilizes a distinct set of operations. For each
member instruction, first use a Base* operation depending on the kind of
location code. Next perform a series of intermediate operations depending on
member code to process all but the last member. Finally, perform a final
operation depending on member code to process the last member. See the
instruction-specific tables for details.

The member codes that represent immediate literal data (ET, EI, PT) are
implemented using the corresponding EC or PC intermediate operation: they
behave exactly as though that literal data had been pushed on the stack as a
cell, then consumed by an ElemC* or PropC* operation.

CGetM <loc-desc/M-vector>    [C..C]  ->  [C]

  Get member as cell.

    location     Base*      member   intermediate   final
    descriptor   operation  code     operation      operation
    -----------+----------  -------+--------------+----------
    C          | BaseC      EC     | ElemCW       | CGetElemC
    R          | BaseR      PC     | PropCW       | CGetPropC
    L          | BaseLW     EL     | ElemLW       | CGetElemL
    NC         | BaseNCW    PL     | PropLW       | CGetPropL
    NL         | BaseNLW    W      | N/A          | N/A
    GC         | BaseGCW
    GL         | BaseGLW
    SC         | BaseSC
    SL         | BaseSL
    H          | BaseH

VGetM <loc-desc/M-vector>    [C..C]  ->  [V]

  Get member as ref.

    location     Base*      member   intermediate   final
    descriptor   operation  code     operation      operation
    -----------+----------  -------+--------------+------------
    C          | BaseC      EC     |  ElemCD      | VGetElemC
    R          | BaseR      PC     |  PropCD      | VGetPropC
    L          | BaseLD     EL     |  ElemLD      | VGetElemL
    NC         | BaseNCD    PL     |  PropLD      | VGetPropL
    NL         | BaseNLD    W      |  NewElem     | VGetNewElem
    GC         | BaseGCD
    GL         | BaseGLD
    SC         | BaseSC
    SL         | BaseSL
    H          | BaseH

FPassM <param id> <loc-desc/M-vector>    [C..C]  ->  [F]

  FPI pass parameter. This instruction behaves as CGetM if parameter %1 is pass
  by value, or it behaves like VGetM if parameter %1 is pass by reference. Then
  it passes the value produced to the callee.

IssetM <loc-desc/M-vector>    [C..C]  ->  [C:Bool]

  Isset member.

    location     Base*      member   intermediate   final
    descriptor   operation  code     operation      operation
    -----------+----------  -------+--------------+-----------
    C          | BaseC      EC     | ElemC        | IssetElemC
    R          | BaseR      PC     | PropC        | IssetPropC
    L          | BaseL      EL     | ElemL        | IssetElemL
    NC         | BaseNC     PL     | PropL        | IssetPropL
    NL         | BaseNL     W      | N/A          | N/A
    GC         | BaseGC
    GL         | BaseGL
    SC         | BaseSC
    SL         | BaseSL
    H          | BaseH

EmptyM <loc-desc/M-vector>    [C..C]  ->  [C:Bool]

  Empty member.

    location     Base*      member   intermediate   final
    descriptor   operation  code     operation      operation
    -----------+----------  -------+--------------+-----------
    C          | BaseC      EC     | ElemC        | EmptyElemC
    R          | BaseR      PC     | PropC        | EmptyPropC
    L          | BaseL      EL     | ElemL        | EmptyElemL
    NC         | BaseNC     PL     | PropL        | EmptyPropL
    NL         | BaseNL     W      | N/A          | N/A
    GC         | BaseGC
    GL         | BaseGL
    SC         | BaseSC
    SL         | BaseSL
    H          | BaseH

SetM <loc-desc/M-vector>    [C..C C]  ->  [C]

  Set member.

    location     Base*      member   intermediate   final
    descriptor   operation  code     operation      operation
    -----------+----------  -------+--------------+-----------
    C          | BaseC      EC     | ElemCD       | SetElemC
    R          | BaseR      PC     | PropCD       | SetPropC
    L          | BaseLD     EL     | ElemLD       | SetElemL
    NC         | BaseNCD    PL     | PropLD       | SetPropL
    NL         | BaseNLD    W      | NewElem      | SetNewElem
    GC         | BaseGCD
    GL         | BaseGLD
    SC         | BaseSC
    SL         | BaseSL
    H          | BaseH

SetWithRefLM <loc-desc/M-vector> <local variable id> [C..C] -> []

  Set member preserving the reffiness of the local.

    location     Base*      member   intermediate   final
    descriptor   operation  code     operation      operation
    -----------+----------  -------+--------------+-----------
    C          | BaseC      EC     | ElemCD       | SetWithRefElemC
    R          | BaseR      PC     | PropCD       | SetPropC
    L          | BaseLD     EL     | ElemLD       | SetWithRefElemL
    NC         | BaseNCD    PL     | PropLD       | SetPropL
    NL         | BaseNLD    W      | NewElem      | SetWithRefNewElem
    GC         | BaseGCD
    GL         | BaseGLD
    SC         | BaseSC
    SL         | BaseSL
    H          | BaseH

SetWithRefRM <loc-desc/M-vector> [C..C R] -> []

  Set member preserving the reffiness of the stack element.

    location     Base*      member   intermediate   final
    descriptor   operation  code     operation      operation
    -----------+----------  -------+--------------+-----------
    C          | BaseC      EC     | ElemCD       | SetWithRefElemC
    R          | BaseR      PC     | PropCD       | SetPropC
    L          | BaseLD     EL     | ElemLD       | SetWithRefElemL
    NC         | BaseNCD    PL     | PropLD       | SetPropL
    NL         | BaseNLD    W      | NewElem      | SetWithRefNewElem
    GC         | BaseGCD
    GL         | BaseGLD
    SC         | BaseSC
    SL         | BaseSL
    H          | BaseH

SetOpM <op> <loc-desc/M-vector>    [C..C C]  ->  [C]

  Set op member.

    location     Base*      member   intermediate   final
    descriptor   operation  code     operation      operation
    -----------+----------  -------+--------------+-------------
    C          | BaseC      EC     | ElemCWD      | SetOpElemC
    R          | BaseR      PC     | PropCWD      | SetOpPropC
    L          | BaseLWD    EL     | ElemLWD      | SetOpElemL
    NC         | BaseNCWD   PL     | PropLWD      | SetOpPropL
    NL         | BaseNLWD   W      | NewElem      | SetOpNewElem
    GC         | BaseGCWD
    GL         | BaseGLWD
    SC         | BaseSC
    SL         | BaseSL
    H          | BaseH

IncDecM <op> <loc-desc/M-vector>    [C..C]  ->  [C]

  Increment/decrement member.

    location     Base*      member   intermediate   final
    descriptor   operation  code     operation      operation
    -----------+----------  -------+--------------+--------------
    C          | BaseC      EC     | ElemCWD      | IncDecElemC
    R          | BaseR      PC     | PropCWD      | IncDecPropC
    L          | BaseLWD    EL     | ElemLWD      | IncDecElemL
    NC         | BaseNCWD   PL     | PropLWD      | IncDecPropL
    NL         | BaseNLWD   W      | NewElem      | IncDecNewElem
    GC         | BaseGCWD
    GL         | BaseGLWD
    SC         | BaseSC
    SL         | BaseSL
    H          | BaseH

BindM <loc-desc/M-vector>    [C..C V]  ->  [V]

  Bind member.

    location     Base*      member   intermediate   final
    descriptor   operation  code     operation      operation
    -----------+----------  -------+--------------+------------
    C          | BaseC      EC     | ElemCD       | BindElemC
    R          | BaseR      PC     | PropCD       | BindPropC
    L          | BaseLD     EL     | ElemLD       | BindElemL
    NC         | BaseNCD    PL     | PropLD       | BindPropL
    NL         | BaseNLD    W      | NewElem      | BindNewElem
    GC         | BaseGCD
    GL         | BaseGLD
    SC         | BaseSC
    SL         | BaseSL
    H          | BaseH

UnsetM <loc-desc/M-vector>    [C..C]  ->  []

  Unset member.

    location     Base*      member   intermediate   final
    descriptor   operation  code     operation      operation
    -----------+----------  -------+--------------+-----------
    C          | BaseC      EC     | ElemCU       | UnsetElemC
    R          | BaseR      PC     | PropCU       | UnsetPropC
    L          | BaseL      EL     | ElemLU       | UnsetElemL
    NC         | BaseNC     PL     | PropLU       | UnsetPropL
    NL         | BaseNL     W      | N/A          | N/A
    GC         | BaseGC
    GL         | BaseGL
    SC         | BaseSC
    SL         | BaseSL
    H          | BaseH


11. Iterator instructions
-------------------------

IterInit  <iterator id> <rel offset> <local id>                [C]  ->  []
IterInitK <iterator id> <rel offset> <local id> <local id>     [C]  ->  []
WIterInit  <iterator id> <rel offset> <local id>               [C]  ->  []
WIterInitK <iterator id> <rel offset> <local id> <local id>    [C]  ->  []

  Initialize iterator. If $1 is an array, these instructions create an array
  iterator, rewind the array iterator to point to the beginning of the array,
  and store the array iterator in the iterator variable %1. Then these
  instructions check if the iterator is at the end, and if it is, these
  instructions free the iterator and transfer control to the location specified
  by %2.

  If $1 is an object that is an instance of an extension class that implements
  the Traversable interface, these instructions create an extension class
  iterator and store it in the iterator variable %1. Then these instructions
  check if the iterator is at the end, and if it is these instructions free the
  iterator and transfer control the location specified by %2.

  If $1 is an object that implements the Iterator interface, these instructions
  create an user class iterator, call $1->rewind(), and store the user class
  iterator in the iterator variable %1. Then these instructions check if
  $1->valid() returns false, and if it does these instructions free the
  iterator and transfer control to the location specified by %2.

  If $1 is an object that implements the IteratorAggregate interface, these
  instructions call $1->getIterator() and inspect the object x that is
  returned. If x is an instance of IteratorAggregate, these instructions will
  repeatedly execute "x = x->getIterator()" until x is not an object that is an
  instance of IteratorAggregate. If x is an object that implements the
  Traversable interface, then this instruction will behave according to the
  appropriate case described above. Otherwise, these instructions will throw an
  exception of type Exception.

  If $1 is an object that does not match any of the three cases above, these
  instructions create a default class iterator, rewind the default class
  iterator to point to the first accessible property, and store the default
  class iterator in the iterator variable %1. Then these instructions check if
  the iterator is at the end, and if it is these instructions free the iterator
  and transfer control the location specified by %2.

  If $1 is not an array or an object, these instructions raise a warning and
  transfer control to the location specified by %2.

  The local ids in %3 (and %4, for the *K variants) represent the locals that
  should receive the value (in %3) and key (in %4) for the iteration, in
  accordance with the type of iterator initialized in %1.

  For the non-W* flavors of these instructions, the locals are stored to with
  the same semantics as SetL (non-binding assignment). The W* flavors of these
  instructions do a binding assignment to %3 if the rhs was a reference, or if
  not they unset the old value of the local and then do a non-binding
  assignment. The W* flavors still do a non-binding assignment to %4.

  The logical value is computed differently depending on the iterator type that
  is initialized in %1:

    If the iterator specified by %1 is a non-mutable array iterator or an
    extension class iterator, these instructions of the current value in %3.

    If the iterator specified by %1 is a user class iterator for object $x,
    these instructions store the return value of $x->current() in %3.

    If the iterator specified by %1 is a non-mutable default class iterator,
    these instructions store a copy of the current property in %3.

  For the *K variants, the logical key to be stored in %4 is computed
  differently depending on the iterator type initialized in %1:

    If the iterator specified by %1 is an array iterator or an extension class
    iterator, this instruction stores a copy of the current key in %4.

    If the iterator specified by %1 is a user class iterator for object $x,
    this instruction stores the return value of $x->key() in %4.

    If the iterator specified by %1 is a non-mutable default class iterator,
    this instruction stores a copy of the name of the current property in %4.

MIterInit  <iterator id> <rel offset> <local id>               [V]  ->  []
MIterInitK <iterator id> <rel offset> <local id> <local id>    [V]  ->  []

  Initialize mutable iterator. If $1 is an array, these instructions create a
  mutable array iterator, rewind the mutable array iterator to point to the
  beginning of the array, and store the mutable array iterator in the iterator
  variable %1. Then these instructions check if the iterator is at the end, and
  if it is these instructions free the iterator and transfers control the
  location specified by %2.

  If $1 is an object that is an instance of an extension class that implements
  the Traversable interface, these instructions raise a fatal error.

  If $1 is an object that implements the Iterator interface, these instructions
  throw a fatal error.

  If $1 is an object that implements the IteratorAggregate interface, these
  instructions throw a fatal error.

  If $1 is an object that does not match any of the three cases above, these
  instructions create a mutable default class iterator, rewind it to point to
  the first accessible property, and store the it in the iterator variable %1.
  Then these instructions check if the iterator is at the end, and if it is
  these instructions free the iterator and transfer control to the location
  specified by %2.

  If $1 is not an array or an object, these instructions raise a warning and
  transfer control to the location specified by %2.

  If the iterator specified by %1 is a mutable array iterator, these
  instructions store the current value in %3 as a ref.

  If the iterator specified by %1 is a mutable default class iterator, these
  instructions store the current property in %3 as a ref.

  For the MIterInitK version, the following also happens:

  If the iterator specified by %1 is an array iterator, this instruction stores
  a copy of the current key in %4 as a cell.

  If the iterator specified by %1 is a mutable default class iterator, this
  instruction stores a copy of the name of the current property in %4 as a
  cell.

IterNext  <iterator id> <rel offset> <local id>               []  ->  []
IterNextK <iterator id> <rel offset> <local id> <local id>    []  ->  []
WIterNext  <iterator id> <rel offset> <local id>              []  ->  []
WIterNextK <iterator id> <rel offset> <local id> <local id>   []  ->  []

  Iterator next. If the specified iterator is a non-mutable array iterator or
  an extension class iterator, advance the iterator to point to the next
  element. If the iterator is not at the end, these instructions transfer
  control to the location specified by %2.

  If the specified iterator is a user class iterator for object $x, this
  instruction executes $x->next(). Then these instructions check if $x->valid()
  returns true, and if it does these instructions transfer control to the
  location specified by %2.

  If the specified iterator is a non-mutable default class iterator, advance
  the iterator to point to the next accessible property in the object. If the
  iterator is not at the end, these instructions transfer control to the
  location specified by %2.

  If the specified iterator is at the end, free the iterator variable with an
  implicit IterFree, then fall through to the next instruction.

  If the specified iterator is not at the end, the local ids in %3 (and %4, for
  the *K variants) represent the locals that should receive the value (in
  %3) and key (in %4) for the iteration, in accordance with the type of
  iterator initialized in %1.

  These locals are stored to with the same semantics as SetL (non-binding
  assignment), except the W* flavors of these instructions do a "with
  reference" assignment (i.e. the local gets a binding assignment if and only
  if the value to be assigned is already a reference).

  The semantics of how to determine what the key and value are depend on %1 in
  an analogous way to {W,}IterInit{K,}.

MIterNext  <iterator id> <rel offset> <local id>               []  ->  []
MIterNextK <iterator id> <rel offset> <local id> <local id>    []  ->  []

  Iterator next. If the specified iterator is a mutable array iterator, advance
  the iterator to point to the next element. If the iterator is not at the end,
  these instructions transfer control to the location specified by %2.

  If the specified iterator is a mutable default class iterator, advance the
  iterator to point to the next accessible property in the object. If the
  iterator is not at the end, these instructions transfer control to the
  location specified by %2.

  If the specified iterator is at the end, free the iterator variable with an
  implicit MIterFree, then fall through to the next instruction.

  If the specified iterator is not at the end, retrieve the key and value:

  If the iterator specified by %1 is a mutable array iterator, these
  instructions store the new current value in %3 as a ref.

  If the iterator specified by %1 is a mutable default class iterator, these
  instructions store the new current property in %3 as a ref.

  For the MIterNextK version, the following also happens:

  If the iterator specified by %1 is an array iterator, this instruction stores
  a copy of the new current key in %4 as a cell.

  If the iterator specified by %1 is a mutable default class iterator, this
  instruction stores a copy of the name of the new current property in %4 as a
  cell.

IterFree <iterator id>    []  ->  []

  Iterator free. This instruction frees the specified iterator variable.
  Typically an iterator gets freed by IterNext, so IterFree is only needed for
  guarding against exceptions and implementing break and return control flow
  statements inside iterator loops.

MIterFree <iterator id>    []  ->  []

  Mutable iterator free. This instruction frees the specified iterator
  variable. Typically an iterator gets freed by MIterNext*, so MIterFree is
  only needed for guarding against exceptions and implementing break and return
  control flow statements inside iterator loops.

CIterFree <iterator id>    []  ->  []

  Cuf iterator free. This instruction frees the specified iterator variable.

IterBreak <iter-vec> <rel offset>   []  ->  []

  Iterator break. Frees vectors in %1 in left to right order then transfers
  control to the location specified by %2. Surprise checks are performed before
  iterators are freed so that in the event of an exception iterators are not
  double freed. Note that as with normal jumps surprise checks will only be
  performed if %2 is negative.


12. Include, eval, and define instructions
------------------------------------------

Incl    [C]  ->  [C]

  Include. Includes the compilation unit containing the file (string)$1. The
  instruction eagerly marks all functions and classes that are unconditionally
  declared in the outermost scope as defined. Next this instruction calls the
  pseudo-main function from the file (string)$1. The pseudo-main function
  inherits the caller's variable environment. If the execution engine cannot
  find a compilation unit containing the file (string)$1, this instruction
  raises a warning.

InclOnce    [C]  ->  [C]

  Include once. Include the compilation unit containing the file (string)$1 if
  it hasn't been included already. This instruction eagerly marks all functions
  and classes that are unconditionally declared in the outermost scope as
  defined, and then calls the pseudo-main function from (string)$1 if it hasn't
  run already. The pseudo-main function inherits the caller's variable
  environment. If the execution engine cannot find a compilation unit
  containing the file (string)$1, this instruction raises a warning.

Req    [C]  ->  [C]

  Require. Includes the compilation unit containing the file (string)$1. The
  instruction eagerly marks all functions and classes that are unconditionally
  declared in the outermost scope as defined. Next this instruction calls the
  pseudo-main function from the file (string)$1. The pseudo-main function
  inherits the caller's variable environment. If the execution engine cannot
  find a compilation unit containing the file (string)$1, this instruction
  throws a fatal error.

ReqOnce    [C]  ->  [C]

  Require once. Include the compilation unit containing the file (string)$1 if
  it hasn't been included already. This instruction eagerly marks all functions
  and classes that are unconditionally declared in the outermost scope as
  defined, and then calls the pseudo-main function from (string)$1 if it hasn't
  run already. The pseudo-main function inherits the caller's variable
  environment. If the execution engine cannot find a compilation unit
  containing the file (string)$1, this instruction throws a fatal error.

ReqDoc    [C]  ->  [C]

  As ReqOnce except the string is always taken to be relative to the document
  root (ie SourceRoot).

Eval    [C]  ->  [C]

  Eval. Executes the source code in (string)$1. This instruction eagerly marks
  all functions and classes that are unconditionally declared in the outermost
  scope as defined, and then calls the pseudo-main function from (string)$1.
  The pseudo-main function from (string)$1 inherits the caller's variable
  environment.

DefFunc <function id>    []  ->  []

  Define function. Bind the function specified by %1. If the function specified
  by %1 is already bound, this instruction does nothing. If another function is
  already bound to the name associated with %1, this instruction throws a fatal
  error.

DefCls <class id>    []  ->  []

  Define class. Bind the class specified by %1. If the class specified by %1 is
  already bound, this instruction does nothing. If another class is already
  bound to the associated name, this instruction throws a fatal error.

NopDefCls <class id>  []  ->  []

  For always-hoistable classes (which are automatically defined when the unit
  is loaded). This instruction is used as a marker for the location in a
  pseudo-main where a DefCls would've existed, but doesn't need to be. (It is
  used as a place to raise errors from if the class fails to define.)

DefCns <litstr id>    [C]  ->  [C]

  Define constant. If there is already a global constant named %1, raises a
  notice and pushes false. If $1 is an array or an object, raises a notice, and
  pushes false. Otherwise defines the constant named %1 to have the value $1,
  and pushes true.

DefTypeAlias <litstr id>   []  ->  []

  Define type alias. Type aliases are a hhvm extension to PHP that allow
  declaring new names for existing types. The unit contains a table of the type
  aliases it was compiled with. This instruction looks up the type alias given
  by %1 in the table. If there is an existing class or type alias defined with
  the same name as %1, this function checks whether it is compatible with the
  type alias given by %1, and if it isn't it raises a fatal error.


13. Miscellaneous instructions
------------------------------

This    []  ->  [C:Obj]

  This. This instruction checks the current instance, and if it is null, this
  instruction throws a fatal error. Next, this instruction pushes the current
  instance onto the stack.

BareThis <notice>   []  ->  [C:Obj|Null]

  This. This instruction pushes the current instance onto the stack. If %1 is
  BareThisOp::Notice, and the current instance is null, emits a notice. If %1
  is BareThisOp::NeverNull the current value of $this is guaranteed to be
  available and can be loaded with no null check.

CheckThis    []  ->  []

  Check existence of this. This instruction checks the current instance, and if
  it is null, throws a fatal error.

InitThisLoc <local variable id>    []  ->  []

  Initialize this local variable. This instruction checks the current instance,
  and if it is not null this instruction stores it to the specified local
  variable. If the current instance is null, or if this bytecode appears in a
  function body that is not a class method, this instruction does nothing.

StaticLoc <local variable id> <litstr id>    []  ->  [C:Bool]

  Static variable. This instruction first checks if the static variable named
  %2 has been marked as initialized. If the static variable has been marked as
  initialized, this instruction binds the static variable to the local variable
  %1 and pushes true. Otherwise, this instruction binds the static variable to
  the local variable %1, marks the static variable as initialized, and pushes
  false.

StaticLocInit <local variable id> <litstr id>    [C]  ->  []

  Static variable with initializer. This instruction first checks if the static
  variable named %2 has been marked as initialized. If the static variable has
  been marked as initialized, this instruction binds the static variable to the
  local variable %1. Otherwise, this instruction binds the static variable to
  the local variable, assigns $1 to the local variable, and marks the static
  variable as initialized.

  The cell in $1 must not be a reference counted type.

Catch    []  ->  [C:Obj]

  Catch. Retrieves the current exception object and pushes it onto the stack.
  The exception object must be a subclass of the builtin Exception class. This
  instruction may only be used at the beginning of a catch entry point.

ClassExists        [C C]  ->  [C:Bool]
InterfaceExists    [C C]  ->  [C:Bool]
TraitExists        [C C]  ->  [C:Bool]

  Check for class/interface/trait existence. If $1 cannot be cast to a bool or
  $2 cannot be cast to a string, this instruction will throw a fatal error.
  Otherwise, it will check for existence of the entity named by $2, invoking
  the autoloader if needed and if $1 is true. The result of the existence check
  will be pushed on the stack.

VerifyParamType <parameter id>    []  ->  []

  Verify parameter type. Functions and methods can optionally specify the types
  of arguments they will accept. These type constraints are memoized into each
  function's FPI structure.

  VerifyParamType checks the specified parameter against the enclosing
  function's corresponding parameter constraints. In case of a mismatch, a
  recoverable error is raised.

VerifyRetTypeC    [C] -> [C]
VerifyRetTypeV    [V] -> [V]

  Verify return type. This instruction pops $1 off of the stack, checks if $1
  is compatible with the current function's return type annotation and raises
  a warning if there is a mismatch, and then it pushes $1 back onto the stack.

Self    []  ->  [A]

  Creates a classref that refers to the class in which the current function is
  defined. This instruction throws a fatal error if the current method is
  defined outside of a class, otherwise it pushes a classref on the stack.

Parent    []  ->  [A]

  Creates a classref that refers to the parent of the class in which the
  current method is defined. This instruction throws a fatal error if the
  current method is defined outside of a class or if the class in which the
  current method is defined has no parent, otherwise it pushes a classref on
  the stack.

LateBoundCls    []  ->  [A]

  Late-bound class. Creates a classref that refers to the current late-bound
  class and pushes it onto the stack.

NativeImpl    []  ->  []

  Native implementation. This instruction invokes the native implementation
  associated with current function and returns the return value to the caller
  of the current function.

IncStat <counter id> <value>    []  ->  []

  Increment stat counter. If stats are enabled, this instruction adds <value>
  to the counter specified by <counter id>. The meaning of the <counter id>
  immediate is implementation defined

AKExists    [C C] -> [C:Bool]

  Checks if array (object) in $1 contains key (property) in $2 and pushes the
  resulting boolean onto the stack. If $2 is null, uses the empty string as
  key. Throws a fatal error if $1 is not an array or object, and raises a
  warning if $2 is not a string, integer, or null.

CreateCl <num args> <class name>  [C|V..C|V]  ->  [C]

  Creates an instance of <class name> and pushes it on the stack.

  The class named by %2 must be a subclass of "Closure", must have a single
  public method named __invoke, must be defined at the point of the CreateCl
  opcode, must be defined in the same unit as the CreateCl opcode, and must
  always be tagged as "ClosureHoistable" (see comments near
  PreClass---hoistability metadata is currently not covered by this spec).

  If there is more than one CreateCl opcode in the unit for the Closure
  subclass named by %2, all of the opcodes must be possible to associate with
  the same class (or trait), or none if the closure will not inherit a class
  context at runtime. This is intended to mean that CreateCl opcodes for a
  given closure may only occur in bytecode bodies of functions that are
  generated to represent a single user-visible PHP function, async function,
  async closure, generator, or generator closure.

  Moreover, for normal (non-async, non-generator) functions and methods, there
  must be at most a single CreateCl opcode in the unit for a given Closure
  subclass contained in the unit.

Idx         [C C C] -> [C]

  Checks if object in $3 contains key in $2 and pushes the result onto the
  stack if found. Otherwise, $1 is pushed onto the stack. $3 must be an object
  that supports array or indexed access (e.g. arrays, collections,
  implementations of ArrayAccess).

ArrayIdx    [C C C] -> [C]

  Checks if array in $3 contains key in $2 and pushes the result onto the stack
  if found. Otherwise, $1 is pushed onto the stack. A fatal error will be
  thrown if $3 is not an array.

AssertTL   <local id>     <op>   []  ->  []
AssertTStk <stack offset> <op>   []  ->  []

  These opcodes may be used to communicate the results of ahead of time static
  analysis to the runtime. They indicate that the value in the specified local
  or stack offset must be statically known to have a particular type. The sub
  opcodes should be used to indicate the type, in an unspecified way.

AssertObjL    <local id>     <class name> <op>   []  ->  []
AssertObjStk  <stack offset> <class name> <op>   []  ->  []

  These opcodes may be used to communicate the results of ahead of time static
  analysis to the runtime.

  Depending on the value of %3, they indicate that the value in the specified
  local or stack offset must be statically known to be an object of a
  particular class or interface type, or an object that is a non-strict subtype
  of a particular class or interface type, or either of those cases with
  nullability.

  %2 is a litstr id of a class or interface name. %3 has the following
  semantics regarding what it asserts about the type of the local or stack
  slot:

        Subop                    Semantics
      -----------------------+------------------------------------
      AssertObjOp::Exact     |  exactly the type named in %2
      AssertObjOp::Sub       |  non-strict subtype of %2
      AssertObjOp::OptExact  |  null, or exactly the type in %2
      AssertObjOp::OptSub    |  null, or non-strict subtype of %2


PredictTL   <local id>     <op>   []  ->  []
PredictTStk <stack offset> <op>   []  ->  []

  These opcodes may be used to provide hints about the types of values at
  runtime. Runtime behavior is implementation defined, although incorrect
  predictions should not affect program correctness.

BreakTraceHint   []  ->  []

  This opcode has no effects, but is a hint that code immediately following it
  is probably not worth including in the same compilation unit as the code in
  front of it. In HHVM, this is used to tell the JIT to break a Tracelet when
  it sees this opcode.

14. Continuation creation and execution
---------------------------------------

CreateCont  <rel offset>   []  ->  [C]

  Creates a Continuation object and pushes it on the stack. The Continuation
  will capture all local variables in the current function, and marks them
  uninitialized on the current frame. The continuation will also store
  resume offset specified by %1.
  
ContEnter   [C]  ->  []

  This instruction may only appear in non-static methods of the Continuation
  class. It transfers control flow to the saved resume offset of a function
  associated with $this Continuation object. The $1 will remain available
  on the stack after the control is transferred.
 
ContRaise   [C:Obj]  ->  []

  This instruction may only appear in non-static methods of the Continuation
  class. It transfers control flow to the saved resume offset of a function
  associated with $this Continuation object. The Exception stored at $1 is
  raised instead of invoking code at the resume offset.

ContSuspend   [C]  ->  [C]

  This instruction may only appear in bodies of continuations and async
  functions. Stores $1 in the continuation as the result of the current
  iteration, sets resume offset at the next opcode and suspends execution
  by transferring control flow back to the ContEnter or ContRaise. Once
  the execution is resumed, the value sent by ContEnter becomes available
  on the stack, or an exception sent by ContRaise is thrown.

ContSuspendK   [C C]  ->  []

  This instruction may only appear in bodies of continuations. Stores $1
  in the continuation as the result and $2 as the key of the current
  iteration, sets resume offset at the next opcode and suspends execution 
  by transferring control flow back to the ContEnter or ContRaise. Once 
  the execution is resumed, the value sent by ContEnter becomes available
  on the stack, or an exception sent by ContRaise is thrown.

ContRetC   [C]  ->  []

  This instruction may only appear in bodies of continuations and async
  functions. Stores $1 in the continuation as the result of the execution,
  marks the continuation as finished and returns by transferring control flow
  back to the ContEnter or ContRaise. Further attempts to iterate the
  Continuation will fail.
 
ContCheck <check started>    []  ->  []

  Check whether continuation can be iterated. $this must be a Continuation
  object. If the continuation is finished, already running, or not yet started
  and <check started> is enabled, an exception will be thrown.

ContValid    []  ->  [C:Bool]

  Check continuation validity. $this must be a Continuation object. Pushes true
  onto the stack if the continuation can be iterated further, false otherwise.

ContKey    []  ->  [C]

  Get continuation key. $this must be a Continuation object. Pushes the most
  recently yielded key from the continuation onto the stack.

ContCurrent    []  ->  [C]

  Get continuation value. $this must be a Continuation object. Pushes the most
  recently yielded value from the continuation onto the stack.

ContStopped    []  ->  []

  Mark continuation as stopped. $this must be a Continuation object.

ContHandle    [C]  ->  []

  Handle exception from continuation body. $this must be a Continuation object.
  Marks the continuation as no longer running and finished and rethrows the
  exception on the top of the stack.

15. Async functions
-------------------

AsyncAwait   [C]  ->  [C C:Bool]

  Perform an await operation on a wait handle. If $1 is not a subclass of
  WaitHandle, raises a fatal error. If $1 is not finished, this instruction
  pushes $1, and pushes False. Otherwise, if $1 succeeded, this instruction
  pushes the result value from the WaitHandle, and then pushes true. If $1
  failed, this instruction throws the exception from the WaitHandle.

AsyncESuspend  <rel offset> <number of iterators>   [C:Obj]  ->  [C:Obj]

  Suspends an eagerly executed async function into an AsyncFunctionWaitHandle 
  object. The AsyncFunctionWaitHandle object will capture all local variables
  and the first %2 iterators in the current eagerly executed async function,
  and mark them uninitialized on the current frame. The object will store
  a reference to the associated resumable async function and a resume offset 
  %1 to be used to resume the execution once the WaitHandle object given 
  by $1 is finished.

  If $1 is not a WaitHandle, or is a WaitHandle that is already finished, the
  behavior is undefined. This opcode must only occur in the eagerly executed
  part of the async function. The value this opcode pushes on the stack must 
  be passed to a RetC opcode.

AsyncResume   []  ->  []

  No operation. Used to teach hhbbc that an exception may be thrown right after
  unreachable ContSuspend. Must appear as the target of AsyncESuspend opcode.

AsyncWrapResult   [C]  ->  [C:Obj]

  Wraps $1 into a StaticResultWaitHandle object.


Basic statement transformations
-------------------------------

To achieve HHBC's goal of making it straightforward for an interpreter or a
compiler to determine order of execution, control flow statements are
transformed to use the simpler constructs. Most control flow statements such as
"if", "while", and "for" are implemented in a straightforward manner using the
Jmp* instructions.

HHBC provides the Switch instruction for implementing very simple switch
statements; most real switch statements are implemented naively using the Eq
and JmpNZ instructions. Also, the functionality of both the echo statement and
the print statement is implemented with the Print instruction.

Foreach statements are implemented using iterator variables and the Iter* and
MIter* instructions. Each foreach loop must be protected by a fault funclet to
ensure that the iterator variable is freed when a foreach loop exits abnormally
through an exception.

Simple break statements and continue statements are implemented using the Jmp*,
IterFree, MIterFree and CIterFree instructions. Dynamic break is implemented
using an unnamed local (to store the 'break count') and a chain of basic
blocks, where each block decrements the unnamed local variable and compares it
with 0, and then decides where to jump next.


Basic expression transformations
--------------------------------

To reduce the size of the instruction set, certain types of expressions are
transformed:

1) Unary plus and negation
Unary plus and negation "+(<expression>)" gets converted to "(0 +
(<expression>))", and "-(<expression>)" gets converted to "(0 -
(<expression>))".

2) Assignment-by operators (+=, -=, etc)
Assignment-by operators are converted to use the SetOp* instructions.

3) List assignment (list)
List assignments are converted to use an unnamed local variable and the SetM
and VGet* instructions. If the function contains any catch funclets, then list
assignment requires a fault funclet as well.

4) Logical and and logical or operators (and/&&, or/||)
If any of the operands side-effect, these operators are implemented using Jmp*
instructions instead of using the "and" and "or" instructions to implement
short-circuit semantics correctly. All Jmp* instructions used to implement
"and" and "or" operators will be forward jumps.

5) The new expression
The new expression is implemented by using the FPushCtor*, FPass*, and FCall
instructions.

6) The ternary operator (?:)
The functionality of the ternary operator is implemented using Jmp*
instructions. All Jmp* instructions used to implement the ternary operator will
be forward jumps.

7) Silence operator (@)
The silence operator is implemented by using various instructions (including
the Jmp* instructions), unnamed local variables, and a fault funclet. All Jmp*
instructions used to implement the silence operator will be forward jumps.

8) The $this expression
The $this expression has different effects depending on whether or not $this is
the direct base of a property expression (such as "$this->x") or a method call
expression (such as "$this->foo()"). When the $this expression is the direct
base of a property expression or a method call expression, the This instruction
is used.

A bare $this expression within an instance method is handled one of two ways:
general or BareThis-optimized (optional). The general solution accesses a local
variable named "this", which is initialized at the beginning of the method
using the InitThisLoc instruction. The BareThis optimization applies to bare
$this access as long as $this is not passed by reference and there are no
dynamic method variables. In such cases, the BareThis instruction can be used
to directly access $this, and the InitThisLoc instruction is not needed.


Warning and errors at parse time
--------------------------------

Certain syntactically correct source code may cause warnings or errors to be
raised when the source file is parsed. Examples of this include using "$this"
on the left hand side of the assignment, using "$this" with binding assignment,
using "$a[]" in an r-value context, and doing "unset($a[])". HHBC handles these
cases by generating Raise or Fatal instructions at the beginning of the body
for the pseudo-main function.


Not yet implemented
-------------------

At the time of this writing, the HipHop bytecode specification is missing the
following details:

1) Description of traits
2) Description of metadata for class statements, trait statements, and method
   statements
3) Description and examples for the yield generator feature
4) Description of the late static binding feature
5) Description of the resource type
6) Definitions of operators (ex. +, -, !) and other helper functions (ex.
   is_null, get_class, strlen)
7) High level description of how namespaces are dealt with and any relevant
   details
8) Description of async function implementation


/* Local Variables: */
/* fill-column: 79 */
/* End: */
vim:textwidth=80