hphp/doc/bytecode.specification

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


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

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

HHBC was designed with several competing goals in mind:

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

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

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


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

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

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

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

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

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

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


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

HipHop bytecode models the flow of execution using a stack of frames referred
to as the "call stack". A "frame" is a structure that logically consists of a
header, a program counter (PC), a local variable store, an iterator variable
store, 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. A value (or "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. Throughout this document, we use "cell" and "value"
interchangeably.

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 default value ("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 three varieties: the main entry point, default value ("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 class name specified by a litstr id and
the class metadata.


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 typed value 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
IterGetKey, IterGetValue, IterSetValue, IterNext and IterFree instructions 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
  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
  Class - denotes class pointer type
  LazyClass - denotes lazy class type
  EnumClassLabel - denotes enum class label type

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

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

$1 is used to refer to the value at the top of the evaluation stack, $2 is used
to refer to the value directly below $1 on the evaluation stack, $3 is used to
refer to the value directly below $2, and so forth. Also, %1 is used to refer
to the first immediate argument, and %2 is used to refer to the second
immediate argument. Thus, the indices used to refer to values on both
the evaluation stack and in the immediate arguments list are 1-indexed.

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 performs similar actions but takes different kinds of
inputs to specify the location to access.

There are numerous instructions which incorporate a readonly immediate.
These opcodes can be Mutable, Any, ReadOnly, CheckROCOW, or CheckMutROCOW and
specify the readonlyness constraint on the property read/written by the
instruction. The Any immediate is equivalent to no runtime check.
The ReadOnly immediate specifies this property must be readonly.
The Mutable immediate specifies this property must be mutable. The CheckROCOW
immediate specifies this property must be readonly and COW. The CheckMutROCOW
immediate specifies this property must be mutable unless it is readonly
and COW.

The member instructions provide functionality to operate on elements and
properties. Many of these instructions incorporate a readonly immediate
argument, as well as an immediate argument which specifies 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 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.

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.

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.


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

  Push immediate. Pushes %1 onto the stack.

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.

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) ]

NewVec <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.

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 arrays, 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.

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

  Get class constant (local). This instruction pushes the value of the class
  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.

EnumClassLabelName [C:EnumClassLabel] -> [C:Static Str]

  Push the name of the enum class label 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.

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

  Addition (+). Pushes ($2 + $1) onto the stack. This instruction throws a
  fatal error if either $1 or $2 is not numeric.

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 either $1 or $2 is not numeric.

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 either $1 or $2 is not numeric.

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 either $1 or $2 is not numeric, or if $1 is zero.

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

  Modulus (%). Pushes ((int)$2 % (int)$1) onto the stack. This instruction throws a
  fatal error if either $1 or $2 is not numeric, or if $1 is zero.

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

  Power. Pushes $2 raised to the power of $1 onto the stack. This instruction throws a
  fatal error if either $1 or $2 is not numeric.

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.

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.

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.

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 <as exception kind> [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. If the exception kind is Error, then throw
  a recoverable error. If it is Typehint raise a TypehintViolationException.

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.

StaticAnalysisError [] -> []

  Produces either a request level Fatal or a process level assertion indicating
  that code believed to be unreachable during static analysis has been executed.


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

Enter <rel offset>   []  ->  []

  Enter the function body. This instruction is used at the end of default value
  initializers to transfer control to the function body.

Jmp <rel offset>    []  ->  []

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

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

  If the local variable is not defined, this instruction raises a warning,
  pushes null onto the stack, and then pushes $1 back 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 <readonly op>    [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 <mode>                   [C]  ->  [C:Class]

  Fetch class. This instruction has different modes selected by the %1 enum.

  Normal mode:
  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.

  ExplicitConversion mode:
  This is an explicit str-to-class conversion through HH\classname_to_class().
  Checks if $1 is a string, class, or lazy class; objects are not admitted and
  classes pointers are pushed unchanged. When $1 is a string, we check whether
  the target class has the <<__DynamicallyReferenced>> attribute and raise a
  sampled notice otherwise. If the target class fails to load, we throw a
  catchable InvalidArgumentException.

  UnsafeBackdoor mode:
  Same as the last mode but skips the <<__DynamicallyReferenced>> check. Not
  valid in repo mode and asserts in HHBBC.

ClassGetTS                         [C:Dict]  ->  [C:Class]

  Fetch class from type-structure. This instruction checks if $1 is a valid
  type-structure (dict). If not, this instruction throws a fatal error.
  Otherwise, $1['classname'] is loaded as with ClassGetC in Normal mode
  and the resulting class is pushed.

ClassGetTSWithGenerics             [C:Dict]  ->  [C:Class, C:StaticVec]

  Fetch class from type-structure. This instruction checks if $1 is a valid
  type-structure (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, an empty vec is pushed.


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

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.

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

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

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.

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
      Vec      | Vec
      Dict     | Dict
      Keyset   | Keyset
      Obj      | Obj
      ArrLike  | Vec or Dict or Keyset
      Scalar   | Int or Dbl or Str or Bool
      Res      | Res
      Class    | Class or LazyClass

  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
      Vec      | Vec
      Dict     | Dict
      Keyset   | Keyset
      Obj      | Obj
      ArrLike  | Vec or Dict or Keyset
      Scalar   | Int or Dbl or Str or Bool
      Res      | Res
      Class    | Class or LazyClass

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

PopG    [C C]  ->  []

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

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.

SetS <readonly op>   [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, 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.

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]
NewObjD   <litstr id>                       []                     ->  [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
      NewObjD   | %1
      NewObjS   | %1

  When loading %1 into x, NewObjD 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/SelfCls/ParentCls 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.

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 1..%2
  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.

  LockWhileUnwinding: whether to lock newly constructed objects if unwinding the
  constructor call.

FCallFunc <fca>                 [U U C|V..C|V C]  ->  [C..C]
FCallFuncD <fca> <litstr id>    [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 C|V..C|V C]  ->  [C..C]
FCallObjMethodD <fca> <class hint> <nullsafe> <litstr id>
  [C 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
  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 C|V..C|V C C:Class]  ->  [C..C]
FCallClsMethodM <fca> <class hint> <op> <litstr id>
  [U U C|V..C|V C]    ->  [C..C]
FCallClsMethodD <fca> <litstr id> <litstr id>
  [U U C|V..C|V]    ->  [C..C]
FCallClsMethodS <fca> <class hint> <mode>
  [U U C|V..C|V C]  ->  [C..C]
FCallClsMethodSD <fca> <class hint> <mode> <litstr id>
  [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
      FCallClsMethodM   | $1 | %4
      FCallClsMethodD   | %2 | %3
      FCallClsMethodS   | %3 | $1
      FCallClsMethodSD  | %3 | %4

  When loading litstr id %2 into x, FCallClsMethodD will perform the work
  performed by the ClassGetC instruction to convert the name given by %2 into
  a class. Similarly, FCallClsMethodM will convert $1 into a class.

  When loading mode %3 into x, FCallClsMethodS and FCallClsMethodSD will
  perform the work performed by LateBoundCls/SelfCls/ParentCls 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
  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 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.


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.

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

  If y is an object that is not a hack collection, 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, hack array, or hack collection this operation outputs a
  base that points to the element at index x in 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 is not a hack collection, 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, hack array, or hack collection 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 non-empty string or an object that is not a hack collection, 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, hack array, or hack collection this operation outputs a
  base that points to the element at index x in y. If there is no element at
  index x, this operation outputs a base that points to null.

  If y is an object that is not a hack collection, 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, hack array, or hack collection this
  operation creates a new element with the next available numeric key in $1
  and outputs a base that points to the new element.

  If $1 is a non-empty string or an object that is not a hack collection, 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, hack array, or hack collection this operation retrieves
  the element at index x from 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 is not a hack collection, 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, hack array, or hack collection this operation pushes
  !is_null(y[x]) onto the stack.

  If y is an object that is not a hack collection, 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.

SetElemC    [C C B]  ->  [C]

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

  If $1 is an object that is not a hack collection, 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, hack array, or hack collection this operation
  executes $1[%1] = $2 and then pushes $2 onto the stack.

  If $1 is an object that is not a hack collection, 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, hack array, or hack collection this
  operation first 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[$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 a non-empty string or an object that is not a hack collection, 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, hack array, or hack collection this
  operation first 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[%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 a non-empty string or an object that is not a hack collection, 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, hack array, or hack
  collection 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 is not a hack collection, this
  operation throws a fatal error.

  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, hack array or hack collection
  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 is not a hack collection, this
  operation throws a fatal error.

  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, hack array, or hack collection this operation removes the
  element at index x from y.

  If y is an object that is not a hack collection, 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, hack array, or hack collection 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 is not a hack collection, this
  operation throws a fatal error.

  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, hack array, or hack collection this
  operation first determines the next available integer offset k in $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 is not a hack collection, this
  operation throws a fatal error.

  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, hack array, or hack
  collection this operation first determines the next available integer offset
  k in $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 is not a hack collection, this
  operation throws a fatal error.

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

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

  Get property as cell.

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

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

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

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

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

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.

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

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

  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> <readonly op>    []  ->  []

  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}{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 and flags.

Flags must be set consistently across all instructions referring to the same
iterator. They carry information that could be used by the runtime to optimize
iteration:

  BaseConst: The base local is unmodified in the loop.
  WithKeys: Indicates whether IterGetKeys might be used during iteration.

IterBase    [C]  ->  [C:Arr|Obj]

  Extract iterator base from $1.

    - If $1 is array-like, pushes $1 onto the stack.

    - If $1 is a collection object, pushes the underlying vec or dict onto
      the stack.

    - If $1 is an object that implements Iterator, pushes $1 onto the stack.

    - 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, pushes x onto the stack.
      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 dict that contains all accessible properties of the base's class,
      in the order that they were defined, then push that dict onto the stack.
      Keys and values of the dict correspond to the property names and values.

    - If $1 is not an array-like or object, an exception is thrown.

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

  Initialize an iterator. This instruction takes a "base" in %2, which must be
  an array-like or an object implementing the Iterator interface. It creates
  an iterator with ID given in %1 pointing to the beginning of %2 and, if %2 is
  an Iterator object, rewinds %2. It then checks if the base is empty. If so,
  it frees the iterator (with an implicit IterFree) and transfers control to
  the target %3.

  The precise semantics of "rewind" and "is empty" depend on the type of
  the base:

    - If %2 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 %2.

    - If %2 is an object that implements Iterator, then we create a new object
      iterator. We call %2->rewind() to reset the base's internal state, then
      call %2->valid() to check if %2 is non-empty.

  All other operations on this iterator must provide the same local (containing
  the same iterator base) as immediates.

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

  Iterator next. This instruction first advances the iterator with the ID given
  in %1 using a base in %2. If the iterator has more elements, it transfers
  control to the location specified by %3. Otherwise, it frees the iterator
  (with an implicit IterFree).

  As with IterInit, the precise semantics of "advance" and "has more elements"
  depends on the type of the base:

    - If %2 is array-like, "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 %2 is an object that implements Iterator, we call %2->next() to update
      the base's internal state, then call %2->valid() to check whether it has
      more elements.

IterGetKey  <IterArgs> <local id>    []  ->  [C]

  Get the key at the current position of iterator with the ID given in %1 using
  a base in %2. Then push it to the evaluation stack.

  If %2 is an object that implements Iterator, the current key is obtained by
  calling %2->key().

  This instruction requires "WithKeys" flag to be set in %1 for this iterator.

IterGetValue  <IterArgs> <local id>    []  ->  [C]

  Get the value at the current position of iterator with the ID given in %1
  using a base in %2. Then push it to the evaluation stack.

  If %2 is an object that implements Iterator, the current value is obtained by
  calling %2->current().

IterSetValue  <IterArgs> <local id>    [C]  ->  []

  Move the value on the evaluation stack to the current position of iterator
  with the ID given in %1 using a base in %2.

  The base %2 must be a vec or dict.

  This instruction requires "BaseConst" flag to not be set in %1.

IterFree <iterator id>    []  ->  []

  Iterator free. This instruction frees the iterator with ID %1, which must have
  been initialized by IterInit. Since the iterator does not contain the base,
  this does not actually free anything. This instruction is just used to inform
  the VM that the iterator slot is no longer used.


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.

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.

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.

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

  Verify out param type. Check that $1 is a value compatible with the declared
  parameter type specified by the given parameter. In case of a mismatch
  a warning or recoverable error is raised.

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

  Verify parameter type. Functions and methods can optionally specify the types
  of arguments they will accept.

  VerifyParamType checks the type of the parameter in $1 against the enclosing
  function's corresponding parameter constraints for `parameter id`. 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.

VerifyTypeTS [C C]  ->  [C]

  VerifyTypeTS pops a type structure from the stack and checks whether
  $2 is compatible with this type structure, checking tuples and shape elements,
  and enum values, in contrast to other Verify instructions that only do
  shallow checking. In case of a mismatch, a warning is issued. If the popped cell
  is not a type structure, an error is raised. This instruction also verifies
  the reified generic type parameters of $2.

SelfCls                  []  ->  [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.

ParentCls                []  ->  [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:Vec] -> [C:Vec]

  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.

CheckClsReifiedGenericMismatch [C:Vec] -> []

  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.

ClassHasReifiedGenerics [C:Class|LazyClass] -> [C:Bool]

  Checks if the class in $1 has reified generics and pushes the
  resulting boolean onto the stack. Throws a fatal error if $1 is not a class
  or lazy class.

GetClsRGProp [C:Class|LazyClass] -> [C:Vec|Null]

  Gets the reified generics property for current instance, using the index of this
  property as stored in class $1. Raises a fatal error if the current instance
  is null or $1 is not a class or lazy class. Returns null if the current
  instance or $1 doesn't have reified generics.

HasReifiedParent [C:Class|LazyClass] -> [C:Bool]

  Checks if the parent of the class in $1 has reified generics and pushes the
  resulting boolean onto the stack. Throws a fatal error if $1 is not a class.
  or lazy class.

CheckClsRGSoft [C:Class|LazyClass] -> []

  Raises a warning if every reified generic in class $1 is soft.
  Otherwise, raises a fatal error. This bytecode should only be emitted
  when a class which expects reified generics is being instantiated without
  any reified generics given (thus guarded by a ClassHasReifiedGenerics bytecode).
  Throws a fatal error if $1 is not a class or lazy class.

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.

ReifiedInit <local variable id> [C:Class|LazyClass C:Obj] -> []

  Checks if the class in $1 or its parent has reified generics. If yes, assigns vec %1
  to the reified generics of object $2. Throws a fatal error if $1 is not a class
  or lazy class.

AKExists    [C C] -> [C:Bool]

  Checks if array (object) in $1 contains key (property) in $2 and pushes the
  resulting boolean onto the stack. If $2 is null, uses the empty string as
  key. Throws a fatal error if $1 is not an array or object, and raises a
  warning if $2 is not a string, integer, or null.

CreateCl <num args> <class name>  [C|U..C|U]  ->  [C]

  Creates an instance of the class specified by <class name> 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 array,
  hack array, or hack collection.

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.

ArrayMarkLegacy   [C C] -> [C]

  Marks the array in $2 as a legacy array and pushes it onto the stack. If $1
  is true then the it is done recursively. If $1 isn't a bool then an exception
  is thrown.

ArrayUnmarkLegacy   [C C] -> [C]

  Marks the array in $2 as a non-legacy array and pushes it onto the stack. If
  $1 is true then the it is done recursively. If $1 isn't a bool then an
  exception is thrown.

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.

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.

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

  Resolve %1 as a function name to a function pointer value corresponding to a
  MethCaller. If the method called is not available in the current context then
  an exception is thrown. Otherwise, the function pointer is pushed to the top
  of the stack. The meth caller must exist in the same unit as the resolving
  function.

ResolveRFunc <litstr id> [C:Vec]  ->  [C]

  Similar to ResolveFunc, resolve %1 as a function name to a function pointer
  and raises an error if the resolution fails. $1 contains a list of reified
  generics. If the function pointer takes reified generics, pushes a value
  capturing the function pointer and reified generics onto the top of stack.
  If the function pointer does not take reified generics, pushes just the
  function pointer into the top of stack.

ResolveClsMethod  <litstr id>               [C:Class] -> [C]
ResolveClsMethodD <litstr id> <litstr id>   []        -> [C]
ResolveClsMethodS <mode>      <litstr id>   []        -> [C]

  Push a class method pointer value. First, these instructions load values
  into x and y as given by the following table:

    instruction             x    y
    ---------------------+----+-----
      ResolveClsMethod   | $1 | %1
      ResolveClsMethodD  | %1 | %2
      ResolveClsMethodS  | %1 | %2

  When loading litstr id %1 into x, ResolveClsMethodD will perform the work
  done by the ClassGetC instruction to convert the name given by %1 into a
  class.

  When loading mode %1 into x, ResolveClsMethodS will perform the same work
  as LateBoundCls/SelfCls/ParentCls depending on the specified mode.

  This instruction checks if class x has an accessible static method named y.
  If not, it raises a fatal error. Otherwise, it creates a value that can be
  used to call that method and pushes the resulting class method pointer onto
  the stack.

ResolveRClsMethod  <litstr id>               [C:Vec C:Class] -> [C]
ResolveRClsMethodD <litstr id> <litstr id>   [C:Vec]         -> [C]
ResolveRClsMethodS <mode>      <litstr id>   [C:Vec]         -> [C]

  Similar to their non-reified counterparts (ResolveClsMethod*), these
  instructions load values into x and y based on the same table and performing
  the same work and the same checks to resolve the class method pointer.

  $1 contains a list of reified generics. If the class x has accessible
  static method y and takes the given reified generics, pushes a value
  capturing the class method pointer and the reified generics. If the
  class method does not accept the reified generics, pushes the class
  method pointer onto the stack.

ThrowNonExhaustiveSwitch [] -> []

  Throws an exception indicating that the switch statement is non exhaustive.
  This exception can be downgraded to a warning or a noop through a runtime
  option.
  This bytecode instruction does not do any checks, it assumes that it was
  emitted correctly.

ResolveClass <litstr id>   []  ->  [C:Class] (where %1 is a class name)

  If %1 is a valid class name, resolve it to a class pointer value, then push
  the pointer onto the top of stack. When resolution fails, raise an error.

RaiseClassStringConversionNotice [] -> []

  Raises a notice indicating an implicit class to string conversion.
  This notice is emitted at a sampled rate through a runtime option.

SetImplicitContextByValue [C:Obj] -> [C:Obj]

  Sets the implicit context to %1 and returns the previous implicit context.

GetMemoAgnosticImplicitContext [] -> [C:Obj]

  Returns the memo agnostic branch of the current IC.

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.

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.

AwaitLowPri []  ->  [C:Null]

  This instruction may only appear in bodies of async functions, and there
  cannot be another await instruction before this one. This instruction suspends
  execution into a new low priority AsyncFunctionWaitHandle. 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.

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) 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