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, and an evaluation 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 two flavors: cells and refs.

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.

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 and
iterator variables are all distinct from each other. 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 each
of local variables or 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 metadata of each function specifies a set of non-overlapping ranges of
bytecode that compose the 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).

The total size of the bytecode for the function body 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 subclasses 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.


Generators
----------

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

The first part, executed when the generator function is called, must consist
of a CreateCont, which is responsible for suspending execution state into a new
Generator 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 Generator's next(), send()
and raise() methods resume execution and transfer control to the resume offset
stored in the Generator object. The user-level code yields values using
Yield and YieldK opcodes and returns using RetC opcode.


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

Async functions are special type of functions representing asynchronous
execution. They can suspend while waiting for other asynchronous operations to
finish. This is achieved using Await opcode, which suspends execution into
an AsyncFunctionWaitHandle object. Once the given dependency is finished,
the scheduler resumes async function at the next opcode.

The async function body can be executed in 2 different modes. If the execution
was never suspended, we are in "eager execution" mode. The code executed after
the resume is executed in "resumed execution" mode.

The "eager execution" can end in 3 different ways. If a RetC opcode is reached,
the result is wrapped into a succeeded StaticWaitHandle and returned to the
caller. If an exception is thrown, it is wrapped into a failed StaticWaitHandle
and returned to the caller. Otherwise, if an Await opcode was reached and the
provided child WaitHandle has not finished, the current execution state is
suspended into an AsyncFunctionWaitHandle object and returned to the caller.
This mechanism allows fast execution if no blocking asynchronous operation was
reached.

The "resumed execution" mode is always entered from the scheduler. In this mode,
the async function either gets blocked on another dependency, or gets finished.
The scheduler is notified of these events using Await and RetC opcodes (or via
the unwinder if an exception was thrown) and the control is given back.

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, 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).

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. When catch entry points are entered, the stack contains a single
Throwable object.

More details about the unwinder and 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.


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

The caller may pass any number of parameters to the callee by pushing zero or
more cells or refs on the stack prior to executing a FCall* instruction. The
caller must pass the parameters in forward order, i.e. the first pushed value
corresponds to the first parameter, and so forth.

The 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 the 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 non-negative
integer "region depth", a set of non-overlapping ranges of bytecode that
compose the "protected region", and an offset of a catch entry point.

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 catch entry point.
Regardless of whether multiple EH entries share the same catch entry point,
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) 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.

2) 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
---------------------

HHBC allows programs to throw exceptions via the Throw instruction. When a
Throw instruction executes it transfers control to the unwinder, which follows
the steps below starting with step 1 until control is transferred elsewhere.

Step 1) Discard all temporary values on the evaluation stack.

Step 2) Consult the EH table of the current function. If there are any EH
        entries that cover the current PC, choose the EH entry with the
        greatest region depth and continue on to step 3. If no matching EH
        entries are found go to step 4.

Step 3) Push the exception object implementing the Throwable interface on the
        evaluation stack, then transfer control to the catch entry point. If
        this catch entry point corresponds to a PHP try/catch statement, it is
        responsible for finding the matching PHP catch clause (e.g. by using
        the InstanceOfD opcode) and rethrowing the exception if no matching
        clause was found.

Step 4) Check if we are handling user exception in an eagerly executed async
        function. If so, pop the current frame, wrap the exception into a
        failed StaticWaitHandle object, leave it on the stack as a return
        value from the async function and resume execution.

Step 5) Pop the current frame off of the call stack and then check if the call
        stack is empty. If the call stack is empty transfer control to the
        unhandled exception facility passing along the exception. If the call
        stack is not empty, then set the PC to point to the FCall* instruction
        which invoked the frame we just discarded and 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.


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

Any given value on the stack must either be a cell or ref 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
  U - uninit; specifies that the value must be an uninitialized null at run
      time; this is only used for FCallBuiltin, CreateCl, and CUGetL.

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, Fatal, Throw, or NativeImpl 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.
Finally, before executing the NativeImpl instruction, the evaluation stack must
be empty.

6) The code for the function body must be laid out in one contiguous block.

7) The last instruction of the function body must be either a control flow
without fallthrough or a terminal instruction.

8) 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 two
possible states: (1) uninitialized, and (2) "iter-initialized" (initialized
via IterInit*). Every range of bytecode for which an iterator variable i is
initialized must be protected by an EH entry with a catch handler that unsets
i by calling IterFree.

9) The iterator variable referenced by IterInit* 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. Note that
IterInit* conditionally initialize the iterator variable, and IterNext*
conditionally free the iterator variable.

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

11) Assertion (AssertRATL and AssertRATStk) instructions cannot be separated
from the following instruction by control flow. Practically speaking, this means
that the instruction immediately following an assertion cannot be a jump target.

12) Sequences of member instructions should be consistent and continuous. That
is, only member instructions and asserts may appear in the sequence, control
flow cannot interrupt the sequence, and the member op mode should be consistent
across all instructions in the sequence. This is because in the case of
exceptions the unwinder decides whether the member state is live by looking at
the instruction that threw.

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
  Vec - denotes the vec type
  Dict - denotes the dict type
  Keyset - denotes the keyset type
  Obj - denotes the object type
  Rec - denotes the record type
  ArrLike - denotes array, vec, dict, or keyset

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*, Switch,
or SSwitch instruction is relative to 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
      value of a cell, and whose name is given by value of a cell.
  C - cell; location is a temporary value given by a cell.
  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. Many of these instructions incorporate an immediate argument which
specifies a one of the following 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
  QT:<id>  - a nullsafe version of PT:<id>. The null-base doesn't issue
             a warning, and no stdClass promotion in write context for the
             base happens. Consume a litstr given by an immediate id
             as a property
  W        - synthesize a new element (no corresponding local variable or
             evaluation stack slot)

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. Generator creation and execution
  15. Async functions


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

Nop    []  ->  []

  No operation. This instruction does nothing.

EntryNop []  ->  []

  No operation. This instruction is occasionally inserted by the emitter to
  avoid having jumps to the entry of a function. Contrary to a Nop, EntryNop
  is guaranteed to not be optimized away.

PopC                  [C]  ->  []
PopU                  [U]  ->  []

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

PopU2                 [U C:<T>]  ->  [C:<T>]

  Pop two. Discards the uninit underneath the cell on top of the stack.

PopFrame<num-skip>    [UUUC..C] -> [C..C]

  Pops three Uninit values from the stack which sit below num-skip cell values.
  This instruction is used to pop the place holder uninit values pushed for a
  frame that was never called.

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

  Teleport value from the stack into a local. This instruction marks the local
  variable given by %1 as defined and pops and stores the value $1 into the
  local variable. This instruction behaves as if it was a SetL PopC pair, but
  might be implemented more efficiently.

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

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

CGetCUNop   [C|U:<T>]  ->  [C:<T>]

  Convert a cell or uninit value to a cell, no op. This is a flavor-safety only
  opcode and should only be used when $1 is statically known to be a cell.

UGetCUNop   [C|U:<T>]  ->  [U:<T>]

 Convert a cell or uninit value to an uninit, no op. This is a flavor-safety
 only opcode and should only be used when $1 is statically known to be an
 uninit.

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]
Vec <scalar vec id>                  []  ->  [C:Vec]
Dict <scalar dict id>                []  ->  [C:Dict]
Keyset <scalar keyset id>            []  ->  [C:Keyset]

  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.

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

  New array in mixed mode, 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.

NewDictArray <capacity hint>    []  ->  [C:Dict]

  New dict, with a capacity hint. Creates a new dict 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.

NewLikeArrayL <local variable id> <capacity hint>   []  ->  [C:Arr]

  New array with the same kind as the array stored in the local variable %1.
  If %1 is not an array, returns an array in mixed mode.  The implementation
  may make use of the hint in %2 to pre-size the array.  If %2 == 0, the
  implementation may instead use the size of %1 as the hint.

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) ]

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

  New dict-like array. Creates a new dict-like 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:

    NewStructDArray < "a" "b" >  [ 1 2 ]  ->  [ darray("a"=>1, "b"=>2) ]

NewStructDict <litstr id vector>    [C..C]  ->  [C:Dict]

  New dict array. Creates a new dict 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:

    NewStructDict < "a" "b" >  [ 1 2 ]  ->  [ dict("a"=>1, "b"=>2) ]

NewVecArray <num elems>   [C..C]  ->  [C:Vec]

  New vec. Creates a new vec from the top %1 cells on the stack, pops those
  cells, then pushes the new vec onto the stack. Elements are pushed on the
  stack in vec insertion order.

NewKeysetArray <num elems>    [C..C]  ->  [C:Keyset]

  New keyset. Creates a new keyset from the top %1 cells on the stack, pops
  those cells, then pushes the new keyset onto the stack.  Elements are pushed
  on the stack in keyset insertion order.

NewVArray <num elems>   [C..]  ->  [C:Arr]

 New vec-like array. Creates a new vec-like 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.

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

  New dict-like array with a capacity hint. Creates a new dict-like 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.

NewRecord <record-name><litstr id vector>    [C..C]  ->  [C:Rec]

  New record. Creates a new record of the record type corresponding to the name
  in %1. The vector of litstr ids in %2 gives the field names to initialize.
  One initial value for each name is popped from the stack. This instruction
  throws errors if the initial values do not match their corresponding
  type-hints or if a field in %2 is not declared in the record type or if any
  field declared in the record type is left uninitialized.

NewRecordArray <record-name><litstr id vector>    [C..C]  ->  [C:Arr]

  New record-array. Same as NewRecord, except it creates a record-array
  instead of a record.

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

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

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

AddNewElemC    [C C]  ->  [C:Arr|Vec|Keyset]

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

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

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

  New collection. Creates a empty new collection of type %1, and pushes it
  onto the stack. %1 must be one of the values of the CollectionType enum other
  than Pair.

NewPair    [C C] -> [C:Obj]

  New Pair collection. Creates a Pair from the top 2 cells on the stack, and
  pushes it onto the stack. Values were pushed onto the stack in the order
  they exist in the pair, so are added to it in reverse order (the top value
  on the stack will become the second element of the pair).

ColFromArray <coll type>    [C:Arr]  ->  [C:Obj]

  Create a collection of type %1 from array $1, and pushes the collection onto
  the stack. %1 must be one of the values of the CollectionType enum other
  than Pair. The array will be used to implement the collection without
  conversion or duplication, thus it should not contain references. $1 must be
  in packed mode if %1 is Vector or ImmVector, and must be in mixed mode
  otherwise.

  Note that integer-like string keys are converted to integers in array, but
  not in collections; thus not all collections can be created using this
  instruction.

CnsE <litstr id>    []  ->  [C:Null|Bool|Int|Dbl|Str|Arr|Vec|Dict|Keyset|Resource]

  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.

ClsCns <litstr id>  [C:Class]  ->  [C:Null|Bool|Int|Dbl|Str|Arr|Vec|Dict|Keyset|Resource]

  Get class constant. This instruction pushes the value of the class constant
  named %1 from the 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|Str|Arr|Vec|Dict|Keyset|Resource]

  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]
Method  []  ->  [C:Static Str]

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

FuncCred      []  -> [C:Obj]

  Push object holding information about current executing function
  onto the stack

ClassName    [C:Class] ->  [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.

ConcatN <n>   [C..C]  ->  [C:Str]

  Concatenation (.). Pushes ((string)$n . ... . (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)$2 - (double)$1.

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.

Pow    [C C]  ->  [C:Int|Dbl]

  Power. Pushes $2 raised to the power of $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.

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.

Cmp    [C C]  ->  [C:Int]

  Comparison. Pushes either -1, 0, or 1 onto the stack if ($1 < $2), ($1 ==
  $2), or ($1 > $2), respectively.

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.

CastVec       [C]  ->  [C:Vec]

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

CastDict      [C]  ->  [C:Dict]

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

CastKeyset    [C]  ->  [C:Keyset]

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

CastVArray    [C]  ->  [C:Arr]

  Cast to vec-like array. Performs an array cast, but drops any keys from the
  input and handles object inputs like CastVec.

CastDArray    [C]  ->  [C:Arr]

  Cast to dict-like array. Performs an array cast, but handles object inputs
  like CastDict.

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.

Select    [C C C] -> [C]

 Pushes (bool)$1 ? $2 : $3 onto the stack.

DblAsBits  [C]  ->   [C]

 If $1 is a double, reinterpret it as an integer (with the same bit-pattern)
 and push it onto the stack. Otherwise, push 0.

IsLateBoundCls [C] -> [C:Bool]

  If $1 is a subtype of the current late-bound class, this instruction pushes
  true onto the stack, otherwise it pushes false onto the stack.

IsTypeStructC <type struct resolve op>  [C C]  ->  [C:Bool]

  If $1 matches the type structure of a defined type and $2 is a subtype of $1,
  this instruction pushes true onto the stack, otherwise it pushes false onto
  the stack.
  If the type struct resolve op is Resolve, then resolves the type structure in
  $1 before performing the subtype check.
  If the type struct resolve op is DontResolve and the given type structure is
  unresolved, then this instruction raises an error.

ThrowAsTypeStructException [C C]  ->  []

  Throws a user catchable type assertion exception that indicates what the
  given type of $2 is, what the expected type is (given on $1) and which key
  it failed at, if applicable.

CombineAndResolveTypeStruct <num type structures> [C..C] -> [C]

  Consumes a type structure from the stack that potentially has holes in it,
  and (%1 - 1) amount of type structures from the stack and merges these type
  structures into the first type structure. Merging means that the hole on the
  first type structure denoted by the reified type kind will be replaced by the
  type structure whose id matches the id provided at this field. If the id at
  this field does not match that of any given type structures or the provided
  inputs are not valid type structures, this instruction throws a fatal error.
  After merging, this instruction resolves the final type structure and pushes
  it 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 (NS means "no surprise"). This behaves
  identically to the Jmp instruction, except that internal VM checks for things
  like OOM and timeouts do not need to be performed even if the offset is
  negative.  This instruction cannot have a zero offset (i.e. it cannot jump to
  itself).

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

  Switch over integer case values. If bounded == SwitchKind::Unbounded, 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 == SwitchKind::Bounded, 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]  ->  []

  Return a cell. Returns $1 to the caller.

  If this instruction is used inside an async function executed in an "eager
  execution" mode, the $1 is wrapped into a StaticResultWaitHandle prior to
  return. In a "resumed execution" mode, the control is given back to the
  scheduler and it is informed that the async function has finished.

  If used in a generator, the Generator object is marked as finished and
  the control is given back to the next instruction after ContEnter or
  ContRaise instruction in a previous frame. The $1 must be Null.

RetCSuspended   [C]  ->  []

  Return a cell. Returns $1, which is an already suspended wait-handle, to the
  caller. This instruction can only be used within async functions. This is
  meant to be used within memoized async functions where the memoized value to
  be returned is already wrapped in a wait-handle.

RetM <num returns> [C..C] -> []

  RetM is a variant of RetC that allows multiple cells to be returned. The RetM
  bytecode must be the only form of return used in a single function, and all
  callers must use FCall* with the matching number of returned values to invoke
  the function.

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.

CGetQuietL <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 pushes null
  onto the stack.

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

CUGetL <local variable id>    []  ->  [C|U]

  Get local as cell or uninit. 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 pushes
  uninit 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.

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.

CGetS    [C C:Class]  ->  [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.

ClassGetC                          [C]  ->  [C:Class]

  Fetch class. This instruction checks if $1 is a string, object, or class. If
  a class, it pushes the input unchanged. If a string, it checks if $1 is the
  name of a defined class. If so, the class is pushed. If not, this instruction
  will invoke the autoload facility passing in $1, and then it will again check
  if $1 matches the name of a defined class. If still not defined, this
  instruction throws a fatal error. If $1 is an object, it pushes the runtime
  class of the object. If $1 is not any of the above cases, this instruction
  throws a fatal error.

ClassGetTS                         [C:DArr]  ->  [C:Class, C:StaticVArr|Null]

  Fetch class from type-structure. This instruction checks if $1 is a valid
  type-structure (darray or dict). If not, this instruction throws a fatal
  error. Otherwise, $1['classname'] is loaded, and mangled as dictated by the
  type-structure's generics information (if present). The possibly mangled
  classname is then processed like ClassGetC and the resulting class is
  pushed. If present, the type-structure's generics information is processed as
  in RecordReifiedGeneric and pushed. If not present, null is pushed.


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.

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 C:Class]  ->  [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.

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 C:Class]  ->  [C:Bool]

  Empty static property. This instruction first checks if $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 checks the type of a value on the stack, according
  to the following table:

    operand        t
    -----------+------
      Null     | Null
      Bool     | Bool
      Int      | Int
      Dbl      | Dbl
      Str      | Str
      Arr      | Arr
      Vec      | Vec
      Dict     | Dict
      Keyset   | Keyset
      Obj      | Obj
      ArrLike  | Arr or Vec or Dict or Keyset
      Scalar   | Int or Dbl or Str or Bool
      Res      | Res

  If t is Obj, this instruction checks if the operand in an object.
    Instances of a special class __PHP_Incomplete_Class are not considered
    objects.
  Otherwise, the result is true if $1 is of type t and false otherwise.
  The result is pushed on the stack.

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

  Is type. This instruction checks the type of a local, according to the
  following table:

    operand        t
    -----------+------
      Null     | Null
      Bool     | Bool
      Int      | Int
      Dbl      | Dbl
      Str      | Str
      Arr      | Arr
      Vec      | Vec
      Dict     | Dict
      Keyset   | Keyset
      Obj      | Obj
      ArrLike  | Arr or Vec or Dict or Keyset
      Scalar   | Int or Dbl or Str or Bool
      Res      | Res

  If the local variable given by %1 is defined, the logic is the same as for
    IsTypeC (see above).
  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 false onto the stack, unless if the operand is Null, in
  which case it pushes true.


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.

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 C:Class C]  ->  [C]

  Set static property. First this instruction checks if the 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.

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 C:Class 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.

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

  Increment/decrement. This instruction first computes x = (string)$1. Next, if
  the global variable 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 C:Class]  ->  [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.

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

  Unset local. Breaks any bindings the local variable given by %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
--------------------

NewObj                                      [C:Class]              ->  [C:Obj]
NewObjR                                     [C:Class, C:VArr|Null] ->  [C:Obj]
NewObjD   <litstr id>                       []                     ->  [C:Obj]
NewObjRD  <litstr id>                       [C:VArr|Null]          ->  [C:Obj]
NewObjS   <mode>                            []                     ->  [C:Obj]

  New object. First, these instructions load a class into x as given
  by the following table:

    instruction   x
    ------------+----
      NewObj    | $1
      NewObjR   | $2
      NewObjD   | %1
      NewObjRD  | %1
      NewObjS   | %1

  When loading %1 into x, NewObjD and NewObjRD will perform the work performed
  by the ClassGetC instruction to convert the name given by %1 into a class.
  NewObjS will perform the same work as LateBoundCls/Self/Parent depending on
  the specified mode.

  This instruction pushes a default-initialized object onto the stack. The
  initialization will complete by running a constructor with FCallCtor, and
  clearing the IsBeingConstructed flag using LockObj.

  NewObjR's and NewObjRD's $1 contains a list of reified generics, or null if
  none present.

LockObj                                    [C:Obj] -> [C:Obj]

  Clears the IsBeingConstructed flag on the object, leaving it on the stack.

FCall* opcodes
--------------

  FCall* opcodes are responsible for invoking the callee determined by the
  specific opcode and performing operations related to the function call as
  specified by the FCA (FCall arguments) immediate consisting of the following
  data:

  <flags> <num args> <num returns> <inout bool vector> <async eager offset>
  [C|V..C|V]  ->  [C..C]

  FCall* first looks up the callee function according to the specific opcode.

  The vector %4 must be either empty or it must contain exactly %2 booleans.
  If it is non-empty, FCall* checks whether inout-ness of parameters 0..(%2-1)
  of the callee matches the corresponding inout-ness values specified by the
  vector %4. Throws an exception if there is a mismatch.

  Finally, FCall* transfers the top %2 values from the stack to the callee as
  parameters and invokes the callee. When the callee returns, it will transfer
  %3 return values onto the caller's evaluation stack using the C flavor. The
  callee must return the matching number of values using either RetC opcode
  (if %3 was one) or RetM opcode (otherwise).

  If the optional offset %5 was specified, the callee supports async eager
  return and it would return a finished Awaitable, it may instead return the
  unpacked result of the Awaitable and continue execution of the caller at
  offset %5.

  If %5 was specified and the callee raised an exception, the exception will
  continue to propagate thru the caller instead of being wrapped into Awaitable.
  Note that for the purposes of exception handling inside the caller, the PC
  will point after the FCall* rather than %5 so it is not advised to have
  different EH entries for these two locations.

  Async eager offset feature is used to avoid the cost of construction of short
  lived Awaitables that are produced by eagerly finishing asynchronous code and
  then immediately awaited by the caller.

  The %1 contains a list of boolean flags:

  Unpack: if enabled, %2 arguments on the stack are followed by an additional
  value, which must be an array. Its elements are transferred to the callee as
  parameters, following the regular %2 parameters.

  Generics: if enabled, %2 arguments and an optional unpack array on the stack
  are followed by an additional value containing the list of reified generics.
  Only FCall*D opcodes are allowed to pass generics.

  SupportsAER: hint from static analysis indicating whether the callee is known
  to support async eager return. Requires specified async eager offset in %5.

FCallFunc <fca>                 [U U U C|V..C|V C]  ->  [C..C]
FCallFuncD <fca> <litstr id>    [U U U C|V..C|V]  ->  [C..C]

  Call a callable. First, these instructions load a value into x as given by
  the following table:

    instruction      x
    --------------+----
      FCallFunc   | $1
      FCallFuncD  | %2

  If x is a string, this instruction attempts to lookup a function named x. If
  a function named x is defined, this instruction calls it. Otherwise it throws
  a fatal error. With FCallFunc*D the litstr in %2 must not start with a '\'
  character, or be of the form "Class::Method". 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 have an __invoke method, this instruction calls
  it. Otherwise it throws a fatal error.

  if x is an array, this instruction will check that the first array element is
  either the name of a class, or an instance of one, and that the second array
  element is the name of a method implemented by the class. If a method exists,
  this instruction calls it. Otherwise it throws a fatal error.

  If x is a func or a clsmeth, this instruction calls it.

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

FCallObjMethod <fca> <class hint> <nullsafe>
  [C U U C|V..C|V C]  ->  [C..C]
FCallObjMethodD <fca> <class hint> <nullsafe> <litstr id>
  [C U U C|V..C|V]  ->  [C..C]

  Call an instance method. First, these instructions load values into x
  and y as given by the following table:

    instruction        x                                             y
    -----------------+---------------------------------------------+-----
    FCallObjMethod   | $(num args + has unpack + 4)                | $1
    FCallObjMethodD  | $(num args + has unpack + has generics + 3) | %3

  If x is not an object and nullsafe != ObjMethodOp::NullThrows, 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
  calls that method.

  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 packs arguments into a varray z and calls the __call method with
  two arguments -- string y and varray z.

  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.

  The string in %2 provides a static analysis hint. If it is non-empty, it is
  the class name with the implementation of method y that will be called.

FCallClsMethod <fca> <class hint> <op>
  [U U U C|V..C|V C C:Class]  ->  [C..C]
FCallClsMethodD <fca> <class hint> <litstr id> <litstr id>
  [U U U C|V..C|V]    ->  [C..C]
FCallClsMethodS <fca> <class hint> <mode>
  [U U U C|V..C|V C]  ->  [C..C]
FCallClsMethodSD <fca> <class hint> <mode> <litstr id>
  [U U U C|V..C|V]    ->  [C..C]

  Call a static method. First, these instructions load values into x and
  y as given by the following table:

    instruction            x    y
    --------------------+----+-----
      FCallClsMethod    | $1 | $2
      FCallClsMethodD   | %3 | %4
      FCallClsMethodS   | %3 | $1
      FCallClsMethodSD  | %3 | %4

  When loading litstr id %3 into x, FCallClsMethodD will perform the work
  performed by the ClassGetC instruction to convert the name given by %3 into
  a class.

  When loading mode %3 into x, FCallClsMethodS and FCallClsMethodSD will
  perform the work performed by LateBoundCls/Self/Parent depending on the
  specified mode.

  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 it does,
  this instruction calls that method.

  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 a
  suitable __call method is found, this instruction packs arguments into
  a varray z and calls the __call method with two arguments -- string y and
  varray z.

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

  The string in %2 provides a static analysis hint. If it is non-empty, it is
  the class name with the implementation of method y that will be called.

  If op is DontLogAsDynamicCall all logging set up for dynamic calls will be
  skipped.

FCallCtor <fca> <class hint>  [C:Obj U U C|V..C|V]  ->  [C]

  This instruction calls a constructor for class of the object given by
  $(num args + has unpack + 3).

  The string in %2 provides a static analysis hint. If it is non-empty, it is
  the class name with the implementation of __construct() that will be called.

  Constructors do not support inout, so the "num returns" in %1 must be 1.

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

  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 C flavor.

  FCallBuiltin supports standalone methods and static methods for which the
  class name is statically known. For static methods, the litstr id will be a
  fully-qualified cls::fn name (e.g. "HH\Map::fromItems").


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

Member operations represent one part of a member expression such as
"$a[0]['name'] = $foo". Each operation corresponds to one bytecode instruction,
but the operations are described separately from their instruction mapping to
separate them from any concerns about instruction encoding.

Operations can produce and consume intermediate values called "bases". A "base"
is a pointer to a memory location that is occupied by a cell or a ref, typically
a local variable, array element, or object property. The current base is stored
in a VM register called the member base register, or MBR for short. Bases are
never stored on the evaluation stack or in any VM location other than the MBR.

A base never owns a reference to the value it points to. It may point to a
temporary value in a scratch register, but the lifetime of the value is always
managed elsewhere.

There are three categories of member operations: base, intermediate, and
final. Base operations produce a base, intermediate operations consume the
current base and produce a new base, and final operations consume the current
base without producing a new one.

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.

9.1 Member base operations
--------------------------

BaseC     [C]  ->  [B]

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

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

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

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

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

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

  Get base from local. This operation outputs a base that points to the local
  given by %1, whether or not it is defined.

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

  Get base from global name. This operation outputs a base that points to 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 points to null.

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

  Get base from global name. This operation outputs a base that points to 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 points to null.

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

  Get base from global name. This operation outputs a base that points to the
  global variable whose name is given by (string)%1 or (string)$1, defining it
  first if necessary.

BaseSC                         [C C:Class]  ->  [B]

  Get base from static property. First, this operation computes x = (string)$2.
  Then this instruction checks if class $1 has an accessible property named
  x. If it does, this operation outputs a base that points to 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 pointing to the object in $this.

9.2 Intermediate member operations
----------------------------------

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

  If y is an object that implements the ArrayAccess interface, this operation
  outputs a base that points to 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 points to the empty string.

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

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

  If y is an object that implements the ArrayAccess interface, this operation
  outputs a base that points to 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 points to the new string.
  Otherwise, this operation raises a warning and outputs a base that points to
  the empty string.

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

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 points to
  the element.

  If y is true, integer, double, this operation raises a warning and outputs a
  base that points to 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 points to the element
  at index x from array y. If there is no element at index x, this operation
  outputs a base that points to null.

  If y is an object that implements the ArrayAccess interface, this operation
  outputs a base that points to 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 base
  pointing to null.

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
  points to the new element.

  If $1 is an object that implements the ArrayAccess interface, this operation
  outputs a base that points to 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 points to the new element.

  If $1 is true, integer, or double, this operation raises a warning and
  outputs a base that points to 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, produce a base pointing to:

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

    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; null
    10X0X | raise warning; null
    10X1X | y->__get(x)
    1100X | throw fatal error
    1101X | y->__get(x)
    111X0 | y->x
    11101 | raise warning; null
    11111 | 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, produce a base pointing to:

    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 | null
    01XXXX | y = new stdclass; create property y->x; y->x
    1X0X0X | create property y->x; y->x
    1X0X1X | y->__get(x)
    1X100X | throw fatal error
    1X101X | y->__get(x)
    1X11X0 | y->x
    1X1101 | re-create property y->x, y->x
    1X1111 | 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
       ----------+----+-----
         PropCU  | $2 | $1
         PropLU  | %1 | $1

  Next, produce a base pointing to:

    y is an object
     y->x is visible
      y->x is accessible
       y->x has been unset previously
    -----+----------------------------------------------------------------------
    0XXX | null
    10XX | create property y->x; y->x
    110X | throw fatal error
    1110 | y->x
    1111 | re-create property y->x; y->x

9.3 Final member operations
---------------------------

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.

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 z = (int)x and then it pushes (z >=
  0 && z < 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.

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.

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.

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.

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

  Set op property.

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

         operation    k    x    y
       ------------+----+----+----
        SetOpPropC | $3 | $2 | $1
        SetOpPropL | %1 | $2 | $1

  Next, perform one of the following actions:

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

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

  Increment/decrement property.

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

         operation     x    y
       -------------+----+----
        IncDecPropC | $2 | $1
        IncDecPropL | %1 | $1

  Next, perform 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 has eligible __set method
          y->x has been unset previously
    --------+-------------------------------------------------------------------
    00XXXXX | raise warning; push null
    01XXXXX | y = new stdclass; b = null; a = b; <op>a; y->x = a;
            |   push a (Pre*) or b (Post*)
    100X0XX | b = null; a = b; <op>a; y->x = a; push a (Pre*) or b (Post*)
    100X10X | b = y->__get(x); a = b; <op>a; y->x = a;
            |   push a (Pre*) or b (Post*)
    100X11X | b = y->__get(x); a = b, <op>a; y->__set(x, a);
            |   push a (Pre*) or b (Post*)
    10100XX | throw fatal error
    101010X | throw fatal error
    101011X | b = y->__get(x); a = b, <op>a; y->__set(x, a);
            |   push a (Pre*) or b (Post*)
    1011XX0 | b = y->x; a = b; <op>a; y->x = a; push a (Pre*) or b (Post*)
    10110X1 | b = null; a = b; <op>a; re-create y->x; y->x = a;
            |   push a (Pre*) or b (Post*)
    1011101 | b = y->__get(x); a = b; <op>a; re-create y->x; y->x = a;
            |   push a (Pre*) or b (Post*)
    1011111 | b = y->__get(x); a = b; <op>a; y->__set(x, a);
            |   push a (Pre*) or b (Post*)

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
-----------------------

Each instruction in this section corresponds to one member operation from the
previous section. The same bytecode may represent multiple different member
operations, differentiating between the options using MOpMode immediates.

Since they represent member operations, these instructions produced and/or
consume a base in the member base register. The MBR is live starting after a
Base* bytecode, modified by zero or more Dim* bytecodes, then finally consumed
by a final operation:

  bytecode  | MBR in-state | MBR out-state
  ----------+--------------+--------------
  Base*     | dead         | live
  Dim*      | live         | live
  Final Ops | live         | dead

Finally, many of these instructions have a <member key> immediate. This is
described in the "Instruction set" introduction section.

10.1 Base Operations
---------------------

BaseGC <stack index> <member op mode>    []  ->  []
BaseGL <local id>    <member op mode>    []  ->  []

  BaseG{C,L}{,W,D} member operation.

BaseSC <stack index> <stack index> <member op mode>    []  ->  []

  BaseSC member operation. %1 gives the location of the static property name,
  and %2 gives the location of the class.

BaseL <local id> <member op mode>    []  ->  []

  BaseL{,W,D} member operation.

BaseC <stack index> <member op mode>   []  ->  []

  BaseC member operation.

BaseH    []  ->  []

  BaseH member operation.

10.2 Intermediate operations
-----------------------------

Dim  <member op mode> <member key>     []  ->  []

  {Prop,Elem}{L,C,I,T}{W,D,U} member operation.

  NewElem operation.

10.3 Final operations
----------------------

All final operations take a <stack count> immediate, which indicates the number
of elements on the eval stack that must be consumed before pushing the final
result. These are elements read by Base*C instructions, and member keys.

QueryM <stack count> <query op> <member key>    [...]  ->  [C]

  {CGet,Isset,Empty}{Prop,Elem} member operation.

SetM <stack count> <member key>    [... C]  ->  [C]

  Set{Prop,Elem} or SetNewElem member operation.

SetRangeM <stack count> <op> <elem size> [... C C C]  ->  []

  Store raw data into a string, optionally reversing the order of elements
  based on op, which may be Forward or Reverse.

  The current member base must be a string (if this or any other required
  conditions are violated, an exception will be thrown). $3, and $1 are cast to
  Int before inspecting their values. $3 gives the offset within the base
  string to begin copying data into. The data comes from a source value in $2;
  supported types are described below. $1 is the count of items to copy from
  $2, and it maybe be -1 to request that an appropriate value is inferred from
  $2. The range [$3, count * size) must fit within [0, length of base).

  The following types are supported as data sources (the value in $2):

  - Bool: op must be Forward, count is ignored, and size must be 1. Stored as a
    1-byte value, either 0 or 1.

  - Int: op must be Forward, count is ignored, and size must be 1, 2, 4, or
    8. The value is truncated to the requested size and stored using the
    current machine's byte ordering.

  - Dbl: op must be Forward, count is ignored, and size must be 4 or 8. The
    value is converted to the requested size and stored using the current
    machine's byte ordering.

  - Str: count indicates the number of characters to copy, starting at the
    beginning of $2, and size must be 1. If op is Reverse, the characters are
    copied in reverse order. Note that characters are still copied starting at
    the beginning of $2, so Forward vs. Reverse never affects which characters
    are copied, just their order as they're written to the base string.

  - Vec: count indicates the number of elements to copy, and size indicates the
    size of each element. All elements of the vec must have the same type,
    which must be Bool, Int, or Dbl. The operation may modify the base string
    before failing if there are elements with mismatched types. Size must be
    one of the allowed values for the contained type, described above. Count
    must not be greater than the size of the vec. If op is Reverse, the
    elements will be copied in reverse order (always starting from offset 0 of
    the vec, as with string sources).

IncDecM <stack count> <op> <member key>    [...]  ->  [C]

  IncDec{Prop,Elem} or IncDecNewElem member operation.

SetOpM <stack count> <op> <member key>    [... C]  ->  [C]

  SetOp{Prop,Elem} or SetOpNewElem member operation.

UnsetM <stack count> <member key>    [...]  ->  []

  Unset{Prop,Elem} member operation.

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

Several iterator instructions take an IterArgs struct. This struct contains an
iterator ID, an value output local ID, and optionally a key output local ID.
Below, when we refer to "the iterator ID in %1", "the value local given in %1",
etc., we're referring to these IDs.

IterInit  <IterArgs> <rel offset>    [C]  ->  []

  Initialize an iterator. This instruction takes a "base" in $1. It creates an
  iterator with ID given in %1 pointing to the beginning of $1 and, if $1 has
  base-internal iteration state (see below), rewinds $1. It then checks if the
  base is empty. If so, it frees the iterator (with an implicit IterFree) and
  transfers control to the target %2.

  If the base is non-empty, this instruction writes the value of the base's
  first element to the value local given in %1. If the iterator is a key-value
  iterator, then this instruction also writes the key of the base's first
  element to the key local given in %1.

  This instruction stores to its key and value output locals with the same
  semantics as SetL (non-binding assignment).

  The precise semantics of "rewind", "is empty", "get key", and "get value"
  depend on the type of the base:

    - If $1 is array-like, we will create a new array iterator. "rewind" does
      nothing in this case - array iterators don't have base-internal state.
      The "is empty" check is a check on the length of $1, and "get key" and
      "get value" load $1's first element.

    - If $1 is a collection object, then we'll unpack or create the underlying
      vec or dict and then create a new array iterator, as above.

    - If $1 is an object that implements Iterator, or if it is an instance of
      an extension class that implements Traversable, then we create a new
      object iterator. We call $1->rewind() to reset the base's internal state,
      then call $1->valid() to check if $1 is non-empty. If $1 is valid, then we
      call $1->current() (and $1->key()) to get the $1's first value (and key).

    - If $1 is an object that implements the IteratorAggregate interface, then
      we repeatedly execute "x = x->getIterator()" until x is no longer an
      object that implements the IteratorAggregate interface. If x is now an
      object that implements the Iterator interface, we create a new object
      iterator as above. Otherwise, we throw an object of type Exception.

    - If $1 is an object that does not match any of the cases above, then we
      create a new default class iterator (with no base-internal state) that
      iterates over accessible properties of the base's class, in the order
      that they were defined. "get key" returns the name of the first property,
      and "get value" returns the $1's value for that property.

    - If $1 is not an array-like or object, this method raises a warning
      and transfers control to the target %2.

LIterInit  <IterArgs> <local id> <rel offset>    []   ->  []

  Initialize iterator with local. If the local specified by %2 is array-like,
  this instruction creates an array iterator pointing to the beginning of %2,
  but leaves %2 in its local rather than copying it into the iterator and
  inc-ref-ing it. Otherwise, it behaves as IterInit, except that the base comes
  from the local %2 rather than from the stack.

  If the base is non-empty, this instruction sets the output value (and, for
  key-value iterators, the output key) to the first element of the base.
  Otherwise, it frees the iterator and transfers control to the target %3.

  Since we don't store array-like bases in the iterator, all other operations
  on this iterator must use the LIter variants (LIterNext, LIterFree) and must
  provide the same local (containing the same array-like) as immediates.

IterNext  <IterArgs> <rel offset>    []   ->  []

  Iterator next. This instruction first advances the iterator with the ID given
  in %1. If the iterator has more elements, then it writes the next element's
  value (and, for key-value iters, its key) to the output locals given in %1,
  Otherwise, the base frees the iterator (with an implicit IterFree) and
  transfers control to the location specified by %2.

  As with IterInit, the precise semantics of "advance" and "has more elements"
  depend on the iterator type. (We could also say it depends on the type of
  the base, but we use some iterator types (e.g. array iterator) for multiple
  base types (e.g. array-likes and collection objects).)

    - For array iterators: "advance" increments the iterator's position, and
      the base "has more elements" if this position is not the final position
      for the stored array. If the array has more elements, we take its key
      and value at the new position.

    - For object iterators: we call $base->next() to update the base's internal
      state, then call $base->valid() to check whether it has more elements. If
      so, we use $base->key() and $base->current() to get the new key and value.

    - For default class iterators, we advance to the base's class's next
      accessible property (if there are more) and we use its name and value as
      the new key and value.

LIterNext  <IterArgs> <local id> <rel offset>    []   ->  []

  Iterator next with local. The iterator with ID given in %1 must have been
  initialized with LIterInit. This instructions behaves similarily to IterNext,
  except that if %2 is an array-like, it will use this local as the iterator's
  base rather than the one stored in the iterator.

IterFree <iterator id>    []  ->  []

  Iterator free. This instruction frees the iterator with ID %1. An iterator is
  typically freed by IterInit or IterNext when its base has no more elements,
  so IterFree is only needed for guarding against exceptions.

LIterFree <iterator id> <local id>   []  ->  []

  Iterator free with local. This instruction frees the iterator with ID %1,
  which must have been initialized by LIterInit with base %2.

IterBreak <rel offset> <iter-vec>    []  ->  []

  Iterator break. Frees iterators in %2 in left to right order, then transfers
  control to the target %1.


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.

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.

DefClsNop <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.)

AliasCls <class-name, alias-name>    [C]  ->  [C:bool]

  This instruction will attempt to bind the name alias-name as an alias for the
  class class-name. If class-name is not an existing class, and $1 is true,
  this instruction will attempt to autoload the class. If class-name is still
  not an existing class, this instruction will raise a warning and return
  false. If alias-name is already bound, this instruction will raise a warning
  and return false. Otherwise it binds the name and returns true.

DefCns <litstr id>    [C]  ->  [C]

  Define constant. If there is already a global constant named %1, raises a
  notice and pushes false. If $1 isn't a scalar, array, vec, dict, or keyset,
  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 throws a fatal error.

DefRecord <record id>    []  ->  []

  Define record. Bind the record specified by %1. If the record specified by %1
  is already bound, this instruction does nothing. If another record or class
  or type alias is already bound to the associated name, this instruction throws
  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.

ChainFaults    [C C]  ->  [C]

  Chain exception objects. If either $1 or $2 is not an object that implements
  Throwable, raise a fatal error. Otherwise, start at $1 and walk the chain of
  "previous" properties until an unset one is found. Set that property to $2,
  unless the previous chain of $1 or $2 forms a cycle. In either case, $1 is
  left on the top of the stack.

OODeclExists <Class|Interface|Trait> [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.

  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]

  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.

VerifyRetNonNullC [C] -> [C]

  This is intended to provide the same behavior as VerifyRetTypeC, except in
  only checks that $1 is non null. This should only be emitted by HHBBC if it
  can statically verify that return value will pass the function's type
  annotation if it is non null.

VerifyParamTypeTS <parameter id>    [C]  ->  []

  VerifyParamTypeTS pops a type structure from the stack and checks the
  specified parameter against this type structure. In case of a mismatch, a
  recoverable error is raised. If the popped cell is not a type structure,
  an error is raised. This instruction also verifies the reified generic type
  parameters of the specified parameter.

VerifyRetTypeTS [C C]  ->  [C]

  VerifyRetTypeTS pops a type structure from the stack and checks whether
  $2 is compatible with this type structure. In case of a mismatch, a
  recoverable error is raised. If the popped cell is not a type structure,
  an error is raised. This instruction also verifies the reified generic type
  parameters of $2.

Self                     []  ->  [C:Class]

  Push a class 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.

Parent                   []  ->  [C:Class]

  Push a class 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.

LateBoundCls             []  ->  [C:Class]

  Late-bound class. Push a class that refers to the current late-bound
  class.

RecordReifiedGeneric [C:VArr] -> [C:VArr]

  Takes a varray or vec based on runtime flag of type structures from $1
  and unless the entry already exists adds a mapping from the grouped name of
  these type structures to a static array that contains the runtime
  representation of these type structures to the global reified generics table.
  Pushes the resulting static list of type structures.

CheckReifiedGenericMismatch [C:VArr] -> []

  Throws a fatal error unless whether each generic in $1 is reified or erased
  matches exactly to the expectations of the current class. If there is
  no class in the current context, throws a fatal error as well.

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.

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 id>  [C|U..C|U]  ->  [C]

  Creates an instance of the class specified by <class id> and pushes it on the
  stack.

  The specified class must be a subclass of "Closure", must have a single
  public method named __invoke, and must be defined in the same unit as the
  CreateCl opcode.

  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.

AssertRATL    <local id>     <repo auth type>    []  ->  []
AssertRATStk  <stack offset> <repo auth type>    []  ->  []

  Assert known "repo authoritative type", for locals or stack offsets.

  These opcodes may be used to communicate the results of ahead of time static
  analysis (hhbbc) to the runtime. They indicate that the value in the
  specified local or stack offset is statically known to have a particular
  type. The "repo auth type" immediate is an encoded RepoAuthType struct (for
  details see runtime/base/repo-auth-type.h).

  As suggested by the name, these opcodes are generally for use with
  RepoAuthoritative mode. They may appear in non-RepoAuthoritative mode with
  one restriction: "specialized" array type information may not be asserted,
  because the global array type table may only be present in RepoAuthoritative
  mode.

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.

Silence  <local id> <Start|End>   []  ->  []

  With %2 = Start, sets the error reporting level to 0 and stores the previous
  one in the local variable %1. The local variable will be overwritten without
  reference counting.

  With %2 = End, if the error reporting level is 0, restores the error
  reporting level to the previous value (stored in local variable %1); if the
  error reporting level is not 0, does nothing.

  The verifier requires that all code paths to an End on local variable %1
  contain a Start on %1, and that all code paths with a Start lead to an End on
  the same variable.  It additionally requires that none of these paths store
  any value in %1 between the Start and End operations. Lastly, the set of
  variables storing the error reporting state must be consistent across block
  boundaries.

  In either case, the local variable %1 must be an unnamed local.

GetMemoKeyL  <local id>  [] ->  [C:<Int/String>]

  Push an int or string which is an appropriate memoize cache key for the
  specified local. The local should be one of the function's parameters. The
  exact scheme for the cache key generation depends on whether the parameter is
  constrained by an appropriate type constraint. This op may throw if the input
  value is one that cannot be converted to a cache key (IE, an object that does
  not implement IMemoizeParam). This op can only be used within a function
  marked as being a memoize wrapper.

MemoGet  <rel offset> <local range>   []   ->  [C]

  Retrieve a memoization value associated with the current function and push it
  onto the stack. The values of the specified range of locals are used as the
  keys to perform the lookup (if any). If any of the locals are not ints or
  strings, fatal. The number of locals must match the number of formal
  parameters to the function. If no value is present, branch to the specified
  offset (without pushing anything). This op can only be used within a function
  marked as being a memoize wrapper.

MemoGetEager  <rel offset> <rel offset> <local range>  []  -> [C]

  Retrieve a memoization value associated with the current function and push it
  onto the stack. This instruction behaves similarily to MemoGet, but is meant
  to be used within an async memoize wrapper. If no value is present, branch to
  the first specified offset (without pushing anything). If a value is present,
  but it is a suspended wait-handle, push it onto the stack and branch to the
  second specified offset. If a value is present, and it represents an eagerly
  returned value (not a suspended wait-handle), push it without branching.

MemoSet  <local range>   [C]  ->  [C]

  Store $1 as a memoization value associated with the current function and
  leave it on the stack. The values of the specified range of locals are used
  as keys to perform the lookup (if any). If any of the locals are not ints or
  strings, fatal. The number of locals must match the number of formal
  parameters to the function. If there is already a value stored with that
  particular set of keys, it is overwritten. This op can only be used within a
  function marked as being a memoize wrapper. If the function is an async
  memoize wrapper, this marks the value as representing a suspended return
  value from the wrapped async function (and therefore must be a wait-handle).

MemoSetEager  <local range>  [C]  -> [C]

  Store $1 as a memoization value associated with the current function and
  leave it on the stack. This instruction behaves similarily as MemoSet, but is
  meant to be used within async memoize wrappers. It indicates that the value
  being stored represents an eager return from the wrapped async function (and
  is not a suspended wait-handle).

ResolveFunc <litstr id>   []  ->  [C]

  Resolve %1 as a function name to a function pointer value, then push the
  pointer onto the top of stack. When resolution fails, raise an error.

ResolveObjMethod [C C] -> [C]

  Resolve $1 as a method of $2 to a method array, then push the array onto the
  top of stack. When $1 represents a class method of $2, the first element of
  the resulting array is the class pointer of $2, otherwise the first element is
  always $2. Fatal if the function resolves to a magic __call method. When
  resolution fails, raise an error.

ResolveClsMethod <cls meth resolve op> [C C] -> [C]

  Resolve $1 as a class method of $2 to a method array, then push the array onto
  the top of stack. Fatal if there is an invoke name for PHP magic call. When
  resolution fails, raise an error.

  If the cls meth resolve op is Warn, then raises warning about parameters used
  in method not being string literals.

14. Generator creation and execution
---------------------------------------

CreateCont   []  ->  [C:Null]

  This instruction may only appear in bodies of generators. Creates a new
  Generator object, moves all local variables from the current frame into
  the object, sets resume offset at the next opcode and suspends execution by
  transferring control flow back to the caller, returning the Generator
  object. Once the execution is resumed, the Null value sent by ContEnter
  becomes available on the stack. It is illegal to resume newly constructed
  Generator using ContEnter with a non-null value or ContRaise opcodes.

ContEnter   [C]  ->  [C]

  This instruction may only appear in non-static methods of the Generator
  class. It transfers control flow to the saved resume offset of a function
  associated with $this Generator object. The $1 will remain available
  on the stack after the control is transferred. Once the control is
  transferred back, a value determined by suspending opcode (Await, Yield,
  YieldK or RetC) will be pushed on the stack. This value corresponds to
  the next()/send() return value -- null for non-async generators, and
  WaitHandle or null for async generators.

ContRaise   [C:Obj]  ->  [C]

  This instruction may only appear in non-static methods of the Generator
  class. It transfers control flow to the saved resume offset of a function
  associated with $this Generator object. The Exception stored at $1 is
  thrown instead of invoking code at the resume offset. Once the control is
  transferred back, a value determined by suspending opcode (Await, Yield,
  YieldK or RetC) will be pushed on the stack. This value corresponds to
  the raise() return value -- null for non-async generators, and WaitHandle
  or null for async generators.

Yield   [C]  ->  [C]

  This instruction may only appear in bodies of generators. Stores $1
  in the generator 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.

YieldK   [C C]  ->  [C]

  This instruction may only appear in bodies of generators. Stores $1
  in the generator 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.

ContCheck <check started>    []  ->  []

  Check whether generator can be iterated. $this must be a Generator
  object. If the generator is finished, already running, or not yet started
  and <check started> is enabled, an exception will be thrown.

ContValid    []  ->  [C:Bool]

  Check generator validity. $this must be a Generator object. Pushes true
  onto the stack if the generator can be iterated further, false otherwise.

ContKey    []  ->  [C]

  Get generator key. $this must be a Generator object. Pushes the most
  recently yielded key from the generator onto the stack.

ContCurrent    []  ->  [C]

  Get generator value. $this must be a Generator object. Pushes the most
  recently yielded value from the generator onto the stack.

ContGetReturn    [] -> [C]

  Get generator's return value. $this must be a Generator object. Pushes the
  return value of the generator onto the stack.

14.1 Generator Delegation
-------------------------

The following bytecodes are used to implement the Generator Delegation feature
of PHP7. Generator Delegation (which is exposed to programmers as `yield from`)
allows generators to yield results from other generators (referred to as their
"delegates") as if they were the ones yielding the value. For a full
explanation of the feature, see https://wiki.php.net/rfc/generator-delegation.

The important part of this feature is how it interacts with different inputs.
The `yield from` pseudo-expression takes two different types of inputs, and
interacts with them very differently. The first is other generators. When
When the delegate is a generator we resume execution of the subgenerator and
return to the caller of the delegating generator (the one containing
`yield from`) when we encounter a yield. All of the features of generators
(`send`, `throw`, etc) are piped through to the delegate, and the result of the
`yield from` expression is the return value of the subgenerator.

The second valid type of input to `yield from` is an object conforming to the
Traversable interface. When the delegate of a generator is a Traversable the
`send()` method does nothing more than continue execution of the delegating
generator. The `throw()` method will throw from the context of the delegating
generator, and the result of the `yield from` expression will always be NULL.
At a high level, you can think of `yield from <traversable>` as being
equivalent to `foreach(<traversable> as $val) { yield $val; }`.

While the bytecodes implementing Generator Delegation have their own API and
well defined contracts, they generally expect to be used in the following
pattern:

```
ContAssignDelegate <expr>
ContEnterDelegate
YieldFromDelegate
ContUnsetDelegate
```

ContAssignDelegate [C] -> []

  This instruction may only appear in bodies of non-async generators. Stores $1
  in the generator as it's delegate, for use later by the ContEnterDelegate
  and YieldFromDelegate bytecodes. Throws an exception if $1 is not either a
  Traversable or another Generator.

ContEnterDelegate [C] -> []

  This instruction may only appear in bodies of non-async generators, and the
  delegate property of the current generator must be non-null. If the delegate
  is a Generator, control flow is transfered to the saved resume offset of the
  delegate. $1 describes the value given to either ContEnter or ContRaise when
  execution of the current generator was resumed, and said value is then passed
  to the delegate generator for use there.

YieldFromDelegate [] -> [C]

  This instruction may only appear in bodies of non-async generators, and the
  delegate property of the current generator must be non-null. If the delegate
  is a Traversable the current key/value pair is stored in the generator as
  the result of the current iteration and the iterator is advanced. When the
  delegate is a Generator, the ContCurrent and ContKey opcodes handle
  returning the result from the deepest active generator. When iteration has
  finished, this opcode continues executing the next instruction, otherwise it
  sets the resume offset to the offset specified by its immediate and returns
  control to the caller.

ContUnsetDelegate <free iterator> [] -> []

  This instruction may only appear in bodies of non-async generators. Removes
  the currently assigned delegate of the generator. If <free iterator> is
  "FreeIter", and the delegate is not another generator, then this opcode is
  responsible for freeing the internal iterator used by this group of opcodes.
  If "IgnoreIter" then it is expected that the iterator is freed elsewhere.
  When the delegate is another generator then it assumes that the iterator was
  never initialized in the first place.

15. Async functions
-------------------

WHResult [C:Obj] -> [C]

  If $1 is not a subclass of WaitHandle, throws a fatal error. If $1 succeeded,
  this instruction pushes the result value from the WaitHandle. If $1 failed,
  this instruction throws the exception stored in the WaitHandle. If $1 is not
  finished, throws an Exception.

Await    [C]  ->  [C]

  This instruction may only appear in bodies of async functions. Awaits
  a WaitHandle provided by $1, suspending the execution if the WaitHandle
  was not yet ready.

  If $1 is not a subclass of WaitHandle, throws a fatal error. If $1 succeeded,
  this instruction pushes the result value from the WaitHandle. If $1 failed,
  this instruction throws the exception from the WaitHandle. Otherwise the
  execution needs to be suspended:

  If the async function is executed eagerly, creates an AsyncFunctionWaitHandle
  object, moves all local variables and iterators from the current frame into
  the object, sets resume offset at the next opcode, marks the
  AsyncFunctionWaitHandle as blocked on the WaitHandle provided by $1 and
  suspends execution by transferring control flow back to the caller, returning
  the AsyncFunctionWaitHandle object.

  If the async function is executed in resumed mode, sets resume offset at
  the next opcode, marks the AsyncFunctionWaitHandle as blocked on the
  WaitHandle provided by $1 and suspends execution by transferring control
  flow back to the scheduler.

  Once the execution is resumed, the result of the WaitHandle provided by $1
  becomes available on the stack.

AwaitAll<local-range>    []  ->  [C:Null]

  Fetches instances of Awaitables from the locals in range %1, and suspends
  until all of them have completed, at which point execution is resumed with a
  single null on the stack.

  Nulls in %1 are ignored, a fatal error is thrown if other non-Awaitables are
  encountered. The stack must be empty. Should all of the Awaitables in %1
  already be complete a null will be pushed to the stack without suspending
  the current function.

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*
instructions. Each foreach loop must be protected by an EH catch entry 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*
and IterFree 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 QueryM
and SetL instructions. In case of exception, the unnamed local variable is
freed using EH entry.

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 NewObj*, FCallCtor, and LockObj
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 an EH catch entry. 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 Throw 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