Copy as Markdown
the *Mozilla Firefox* web browser. The implementation behaviour is defined by
the `ECMAScript <https://tc39.es/ecma262/>`_ and `WebAssembly
Much of the internal technical documentation of the engine can be found
throughout the source files themselves by looking for comments labelled with
`[SMDOC]`_. Information about the team, our processes, and about embedding
*SpiderMonkey* in your own projects can be found at https://spidermonkey.dev.
Specific documentation on a few topics is available at:
Components of SpiderMonkey
🧹 Garbage Collector
Rooting Hazard Analysis <HazardAnalysis/index>
Running the Analysis <HazardAnalysis/running>
we manage a garbage-collected memory heap. Elements of this heap have a base
C++ type of `gc::Cell`_. Each round of garbage collection will free up any
*Cell* that is not referenced by a *root* or another live *Cell* in turn.
See :doc:`GC overview<gc>` for more details.
📦 JS::Value and JSObject
(*Undefined*, *Null*, *Boolean*, *Number*, *BigInt*, *String*, or *Symbol*).
Values are represented with the `JS::Value`_ type which may in turn point to
an object that extends from the `JSObject`_ type. Objects include both plain
functions to *ArrayBuffers* to *HTML Elements* and more.
Most objects extend ``NativeObject`` (which is a subtype of ``JSObject``)
which provides a way to store properties as key-value pairs similar to a hash
table. These objects hold their *values* and point to a *Shape* that
represents the set of *keys*. Similar objects point to the same *Shape* which
saves memory and allows the JITs to quickly work with objects similar to ones
it has seen before. See the `[SMDOC] Shapes`_ comment for more details.
C++ (and Rust) code may create and manipulate these objects using the
collection of interfaces we traditionally call the **JSAPI**.
In order to evaluate script text, we parse it using the *Parser* into an
`Abstract Syntax Tree`_ (AST) temporarily and then run the *BytecodeEmitter*
(BCE) to generate `Bytecode`_ and associated metadata. We refer to this
resulting format as `Stencil`_ and it has the helpful characteristic that it
does not utilize the Garbage Collector. The *Stencil* can then be
instantiated into a series of GC *Cells* that can be mutated and understood
by the execution engines described below.
Each function as well as the top-level itself generates a distinct script.
This is the unit of execution granularity since functions may be set as
callbacks that the host runs at a later time. There are both
``ScriptStencil`` and ``js::BaseScript`` forms of scripts.
By default, the parser runs in a mode called *syntax* or *lazy* parsing where
we avoid generating full bytecode for functions within the source that we are
parsing. This lazy parsing is still required to check for all *early errors*
that the specification describes. When such a lazily compiled inner function
is first executed, we recompile just that function in a process called
*delazification*. Lazy parsing avoids allocating the AST and bytecode which
saves both CPU time and memory. In practice, many functions are never
executed during a given load of a webpage so this delayed parsing can be
The *bytecode* generated by the parser may be executed by an interpreter
written in C++ that manipulates objects in the GC heap and invokes native
code of the host (eg. web browser). See `[SMDOC] Bytecode Definitions`_ for
descriptions of each bytecode opcode and ``js/src/vm/Interpreter.cpp`` for
In order to speed up execution of *bytecode*, we use a series of Just-In-Time
(JIT) compilers to generate specialized machine code (eg. x86, ARM, etc)
As an individual script runs more times (or has a loop that runs many times)
we describe it as getting *hotter* and at certain thresholds we *tier-up* by
JIT-compiling it. Each subsequent JIT tier spends more time compiling but
aims for better execution performance.
The *Baseline Interpreter* is a hybrid interpreter/JIT that interprets the
*bytecode* one opcode at a time, but attaches small fragments of code called
*Inline Caches* (ICs) that rapidly speed-up executing the same opcode the next
time (if the data is similar enough). See the `[SMDOC] JIT Inline Caches`_
comment for more details.
The *Baseline Compiler* use the same *Inline Caches* mechanism from the
*Baseline Interpreter* but additionally translates the entire bytecode to
native machine code. This removes dispatch overhead and does minor local
optimizations. This machine code still calls back into C++ for complex
operations. The translation is very fast but the ``BaselineScript`` uses
memory and requires ``mprotect`` and flushing CPU caches.
The *WarpMonkey* JIT replaces the former *IonMonkey* engine and is the
highest level of optimization for the most frequently run scripts. It is able
to inline other scripts and specialize code based on the data and arguments
We translate the *bytecode* and *Inline Cache* data into a Mid-level
`Intermediate Representation`_ (Ion MIR) representation. This graph is
transformed and optimized before being *lowered* to a Low-level Intermediate
Representation (Ion LIR). This *LIR* performs register allocation and then
generates native machine code in a process called *Code Generation*.
See `MIR Optimizations`_ for an overview of MIR optimizations.
The optimizations here assume that a script continues to see data similar
what has been seen before. The *Baseline* JITs are essential to success here
because they generate *ICs* that match observed data. If after a script is
compiled with *Warp*, it encounters data that it is not prepared to handle it
performs a *bailout*. The *bailout* mechanism reconstructs the native machine
stack frame to match the layout used by the *Baseline Interpreter* and then
branches to that interpreter as though we were running it all along. Building
this stack frame may use special side-table saved by *Warp* to reconstruct
values that are not otherwise available.
This engine performs fast translation to machine code in order to minimize
latency to first execution.
This engine translates the WASM input into same *MIR* form that *WarpMonkey*
uses and uses the *IonBackend* to optimize. These optimizations (and in
particular, the register allocation) generate very fast native machine code.
.. _[SMDOC] Shapes: https://searchfox.org/mozilla-central/search?q=[SMDOC]+Shapes
.. _[SMDOC] Bytecode Definitions: https://searchfox.org/mozilla-central/search?q=[SMDOC]+Bytecode+Definitions&path=js%2F
.. _[SMDOC] JIT Inline Caches: https://searchfox.org/mozilla-central/search?q=[SMDOC]+JIT+Inline+Caches
.. _Bytecode: https://en.wikipedia.org/wiki/Bytecode
.. _Abstract Syntax Tree: https://en.wikipedia.org/wiki/Abstract_syntax_tree
.. _Intermediate Representation: https://en.wikipedia.org/wiki/Intermediate_representation
.. _MIR Optimizations: ./MIR-optimizations/index.html