---
title: 'V8 Torque user manual'
description: 'This document explains the V8 Torque language, as used in the V8 codebase.'
---
V8 Torque is a language that allows developers contributing to the V8 project to express changes in the VM by focusing on the _intent_ of their changes to the VM, rather than preoccupying themselves with unrelated implementation details. The language was designed to be simple enough to make it easy to directly translate the [ECMAScript specification](https://tc39.es/ecma262/) into an implementation in V8, but powerful enough to express the low-level V8 optimization tricks in a robust way, like creating fast-paths based on tests for specific object-shapes.
Torque will be familiar to V8 engineers and JavaScript developers, combining a TypeScript-like syntax that eases both writing and understanding V8 code with syntax and types that reflects concepts that are already common in the [`CodeStubAssembler`](/blog/csa). With a strong type system and structured control flow, Torque ensures correctness by construction. Torque’s expressiveness is sufficient to express almost all of the functionality that is [currently found in V8’s builtins](/docs/builtin-functions). It also is very interoperable with `CodeStubAssembler` builtins and `macro`s written in C++, allowing Torque code to use hand-written CSA functionality and vice-versa.
Torque provides language constructs to represent high-level, semantically-rich tidbits of V8 implementation, and the Torque compiler converts these morsels into efficient assembly code using the `CodeStubAssembler`. Both Torque’s language structure and the Torque compiler’s error checking ensure correctness in ways that were previously laborious and error-prone with direct usage of the `CodeStubAssembler`. Traditionally, writing optimal code with the `CodeStubAssembler` required V8 engineers to carry a lot of specialized knowledge in their heads — much of which was never formally captured in any written documentation — to avoid subtle pitfalls in their implementation. Without that knowledge, the learning curve for writing efficient builtins was steep. Even armed with the necessary knowledge, non-obvious and non-policed gotchas often led to correctness or [security](https://bugs.chromium.org/p/chromium/issues/detail?id=775888) [bugs](https://bugs.chromium.org/p/chromium/issues/detail?id=785804). With Torque, many of these pitfalls can be avoided and recognized automatically by the Torque compiler.
## Getting started
Most source written in Torque is checked into the V8 repository under [the `src/builtins` directory](https://github.com/v8/v8/tree/master/src/builtins), with the file extension `.tq`. Torque definitions of V8's heap-allocated classses are found alongside their C++ definitions, in `.tq` files with the same name as corresponding C++ files in `src/objects`. The actual Torque compiler can be found under [`src/torque`](https://github.com/v8/v8/tree/master/src/torque). Tests for Torque functionality are checked in under [`test/torque`](https://github.com/v8/v8/tree/master/test/torque), [`test/cctest/torque`](https://github.com/v8/v8/tree/master/test/cctest/torque), and [`test/unittests/torque`](https://github.com/v8/v8/tree/master/test/unittests/torque).
To give you a taste of the language, let’s write a V8 builtin that prints “Hello World!”. To do this, we’ll add a Torque `macro` in a test case and call it from the `cctest` test framework.
Begin by opening up the `test/torque/test-torque.tq` file and add the following code at the end (but before the last closing `}`):
```torque
@export
macro PrintHelloWorld(): void {
Print('Hello world!');
}
```
Next, open up `test/cctest/torque/test-torque.cc` and add the following test case that uses the new Torque code to build a code stub:
```cpp
TEST(HelloWorld) {
Isolate* isolate(CcTest::InitIsolateOnce());
CodeAssemblerTester asm_tester(isolate, JSParameterCount(0));
TestTorqueAssembler m(asm_tester.state());
{
m.PrintHelloWorld();
m.Return(m.UndefinedConstant());
}
FunctionTester ft(asm_tester.GenerateCode(), 0);
ft.Call();
}
```
Then [build the `cctest` executable](/docs/test), and finally execute the `cctest` test to print ‘Hello world’:
```bash
$ out/x64.debug/cctest test-torque/HelloWorld
Hello world!
```
## How Torque generates code
The Torque compiler doesn’t create machine code directly, but rather generates C++ code that calls V8’s existing `CodeStubAssembler` interface. The `CodeStubAssembler` uses the [TurboFan compiler’s](https://v8.dev/docs/turbofan) backend to generate efficient code. Torque compilation therefore requires multiple steps:
1. The `gn` build first runs the Torque compiler. It processes all `*.tq` files. Each Torque file `path/to/file.tq` causes the generation of the following files:
- `path/to/file-tq-csa.cc` and `path/to/file-tq-csa.h` containing generated CSA macros.
- `path/to/file-tq.inc` to be included in in a corresponding header `path/to/file.h` containing class definitions.
- `path/to/file-tq-inl.inc` to be included in the corresponding inline header `path/to/file-inl.h`, containing C++ accessors of class definitions.
- `path/to/file-tq.cc` containing generated heap verifiers, printers, etc.
The Torque compiler also generates various other known `.h` files, meant to be consumed by the V8 build.
1. The `gn` build then compiles the generated `-csa.cc` files from step 1 into the `mksnapshot` executable.
1. When `mksnapshot` runs, all of V8’s builtins are generated and packaged in to the snapshot file, including those that are defined in Torque and any other builtins that use Torque-defined functionality.
1. The rest of V8 is built. All of Torque-authored builtins are made accessible via the snapshot file which is linked into V8. They can be called like any other builtin. In addition, the `d8` or `chrome` executable also includes the generated compilation units related to class definitions directly.
Graphically, the build process looks like this:
## Torque tooling { #tooling }
Basic tooling and development environment support is available for Torque.
- There is a [Visual Studio Code plugin](https://github.com/v8/vscode-torque) for Torque, which uses a custom language server to provide features like go-to-definition.
- There is also a formatting tool that should be used after changing `.tq` files: `tools/torque/format-torque.py -i `
## Troubleshooting builds involving Torque { #troubleshooting }
Why do you need to know this? Understanding how Torque files get converted into machine code is important because different problems (and bugs) can potentially arise in the different stages of translating Torque into the binary bits embedded in the snapshot:
- If you have a syntax or semantic error in Torque code (i.e. a `.tq` file), the Torque compiler fails. The V8 build aborts during this stage, and you will not see other errors that may be uncovered by later parts of the build.
- Once your Torque code is syntactically correct and passes the Torque compiler’s (more or less) rigorous semantic checks, the build of `mksnapshot` can still fail. This most frequently happens with inconsistencies in external definitions provided in `.tq` files. Definitions marked with the `extern` keyword in Torque code signal to the Torque compiler that the definition of required functionality is found in C++. Currently, the coupling between `extern` definitions from `.tq` files and the C++ code to which those `extern` definitions refer is loose, and there is no verification at Torque-compile time of that coupling. When `extern` definitions don’t match (or in the most subtle cases mask) the functionality that they access in the `code-stub-assembler.h` header file or other V8 headers, the C++ build of `mksnapshot` fails.
- Even once `mksnapshot` successfully builds, it can fail during execution. This might happen because Turbofan fails to compile the generated CSA code, for example because a Torque `static_assert` cannot be verified by Turbofan. Also, Torque-provided builtin that are run during snapshot creation might have a bug. For example, `Array.prototype.splice`, a Torque-authored builtin, is called as part of the JavaScript snapshot initialization process to setup the default JavaScript environment. If there is a bug in the implementation, `mksnapshot` crashes during execution. When `mksnapshot` crashes, it’s sometimes useful to call `mksnapshot` passing the `--gdb-jit-full` flag, which generates extra debug information that provides useful context, e.g. names for Torque-generated builtins in `gdb` stack crawls.
- Of course, even if Torque-authored code makes it through `mksnapshot`, it still may be buggy or crash. Adding test cases to `torque-test.tq` and `torque-test.cc` is a good way to ensure that your Torque code does what you actually expect. If your Torque code does end up crashing in `d8` or `chrome`, the `--gdb-jit-full` flag is again very useful.
## `constexpr`: compile-time vs. run-time { #constexpr }
Understanding the Torque build process is also important to understanding a core feature in the Torque language: `constexpr`.
Torque allows evaluation of expressions in Torque code at runtime (i.e. when V8 builtins are executed as part of executing JavaScript). However, it also allows expressions to be executed at compile time (i.e. as part of the Torque build process and before the V8 library and `d8` executable have even been created).
Torque uses the `constexpr` keyword to indicate that an expression must be evaluated at build-time. Its usage is somewhat analogous to [C++’s `constexpr`](https://en.cppreference.com/w/cpp/language/constexpr): in addition to borrowing the `constexpr` keyword and some of its syntax from C++, Torque similarly uses `constexpr` to indicate the distinction between evaluation at compile-time and runtime.
However, there are some subtle differences in Torque’s `constexpr` semantics. In C++, `constexpr` expressions can be evaluated completely by the C++ compiler. In Torque `constexpr` expressions cannot fully be evaluated by the Torque compiler, but instead map to C++ types, variables and expressions that can be (and must be) fully evaluated when running `mksnapshot`. From the Torque-writer’s perspective, `constexpr` expressions do not generate code executed at runtime, so in that sense they are compile-time, even though they are technically evaluated by C++ code external to Torque that `mksnapshot` runs. So, in Torque, `constexpr` essentially means “`mksnapshot`-time”, not “compile time”.
In combination with generics, `constexpr` is a powerful Torque tool that can be used to automate the generation of multiple very efficient specialized builtins that differ from each other in a small number of specific details that can be anticipated by V8 developers in advance.
## Files
Torque code is packaged in individual source files. Each source file consists of a series of declarations, which themselves can optionally wrapped in a namespace declaration to separate the namespaces of declarations. The following description of the grammar is likely out-of-date. The source-of-truth is [the grammar definition in the Torque compiler](https://source.chromium.org/chromium/chromium/src/+/master:v8/src/torque/torque-parser.cc?q=TorqueGrammar::TorqueGrammar), which is written using contex-free grammar rules.
A Torque file is a sequence of declarations. The possible declarations are listed [in `torque-parser.cc`](https://source.chromium.org/chromium/chromium/src/+/master:v8/src/torque/torque-parser.cc?q=TorqueGrammar::declaration).
## Namespaces
Torque namespaces allow declarations to be in independent namespaces. They are similar to C++ namespaces. They allow you to create declarations that are not automatically visible in other namespaces. They can be nested, and declarations inside a nested namespace can access the declarations in the namespace that contains them without qualification. Declarations that are not explicitly in a namespace declaration are put in a shared global default namespace that is visible to all namespaces. Namespaces can be reopened, allowing them to be defined over multiple files.
For example:
```torque
macro IsJSObject(o: Object): bool { … } // In default namespace
namespace array {
macro IsJSArray(o: Object): bool { … } // In array namespace
};
namespace string {
// …
macro TestVisibility() {
IsJsObject(o); // OK, global namespace visible here
IsJSArray(o); // ERROR, not visible in this namespace
array::IsJSArray(o); // OK, explicit namespace qualification
}
// …
};
namespace array {
// OK, namespace has been re-opened.
macro EnsureWriteableFastElements(array: JSArray){ … }
};
```
## Declarations
### Types
Torque is strongly typed. Its type system is the basis for many of the security and correctness guarantees it provides.
For many basic types, Torque doesn’t actually inherently know very much about them. Instead, many types are just loosely coupled with `CodeStubAssembler` and C++ types through explicit type mappings and rely on the C++ compiler to enforce the rigor of that mapping. Such types are realized as abstract types.
#### Abstract types
Torque’s abstract types map directly to C++ compile-time and CodeStubAssembler runtime values. Their declarations specify a name and a relationship to C++ types:
```grammar
AbstractTypeDeclaration :
type IdentifierName ExtendsDeclaration opt GeneratesDeclaration opt ConstexprDeclaration opt
ExtendsDeclaration :
extends IdentifierName ;
GeneratesDeclaration :
generates StringLiteral ;
ConstexprDeclaration :
constexpr StringLiteral ;
```
`IdentifierName` specifies the name of the abstract type, and `ExtendsDeclaration` optionally specifies the type from which the declared type derives. `GeneratesDeclaration` optionally specifies a string literal which corresponds to the C++ `TNode` type used in `CodeStubAssembler` code to contain a runtime value of its type. `ConstexprDeclaration` is a string literal specifying the C++ type corresponding to the `constexpr` version of the Torque type for build-time (`mksnapshot`-time) evaluation.
Here’s an example from `base.tq` for Torque’s 31- and 32-bit signed integer types:
```torque
type int32 generates 'TNode' constexpr 'int32_t';
type int31 extends int32 generates 'TNode' constexpr 'int31_t';
```
#### Union types
Union types express that a value belongs to one of several possible types. We only allow union types for tagged values, because they can be distinguished at runtime using the map pointer. For example, JavaScript numbers are either Smi values or allocated `HeapNumber` objects.
```torque
type Number = Smi | HeapNumber;
```
Union types satisfy the following equalities:
- `A | B = B | A`
- `A | (B | C) = (A | B) | C`
- `A | B = A` if `B` is a subtype of `A`
It is only allowed to form union types from tagged types because untagged types cannot be distinguished at runtime.
When mapping union types to CSA, the most specific common supertype of all the types of the union type is selected, with the exception of `Number` and `Numeric`, which are mapped to the corresponding CSA union types.
#### Class types
Class types make it possible to define, allocate and manipulate structured objects on the V8 GC heap from Torque code. Each Torque class type must correspond to a subclass of HeapObject in C++ code. In order to minimize the expense of maintaining boilerplate object-accessing code between V8’s C++ and Torque implementation, the Torque class definitions are used to generate the required C++ object-accessing code whenever possible (and appropriate) to reduce the hassle of keeping C++ and Torque synchronized by hand.
```grammar
ClassDeclaration :
ClassAnnotation* extern opt transient opt class IdentifierName ExtendsDeclaration opt GeneratesDeclaration opt {
ClassMethodDeclaration*
ClassFieldDeclaration*
}
ClassAnnotation :
@doNotGenerateCppClass
@generateBodyDescriptor
@generatePrint
@abstract
@export
@noVerifier
@hasSameInstanceTypeAsParent
@highestInstanceTypeWithinParentClassRange
@lowestInstanceTypeWithinParentClassRange
@reserveBitsInInstanceType ( NumericLiteral )
@apiExposedInstanceTypeValue ( NumericLiteral )
ClassMethodDeclaration :
transitioning opt IdentifierName ImplicitParameters opt ExplicitParameters ReturnType opt LabelsDeclaration opt StatementBlock
ClassFieldDeclaration :
ClassFieldAnnotation* weak opt const opt FieldDeclaration;
ClassFieldAnnotation :
@noVerifier
@if ( Identifier )
@ifnot ( Identifier )
FieldDeclaration :
Identifier ArraySpecifier opt : Type ;
ArraySpecifier :
[ Expression ]
```
An example class:
```torque
extern class JSProxy extends JSReceiver {
target: JSReceiver|Null;
handler: JSReceiver|Null;
}
```
`extern` signifies that this class is defined in C++, rather than defined only in Torque.
The field declarations in classes implicitly generate field getters and setters that can be used from CodeStubAssembler, e.g.:
```cpp
// In TorqueGeneratedExportedMacrosAssembler:
TNode LoadJSProxyTarget(TNode p_o);
void StoreJSProxyTarget(TNode p_o, TNode p_v);
```
As described above, the fields defined in Torque classes generate C++ code that removes the need for duplicate boilerplate accessor and heap visitor code. The hand-written definition of JSProxy must inherit from a generated class template, like this:
```cpp
// In js-proxy.h:
class JSProxy : public TorqueGeneratedJSProxy {
// Whatever the class needs beyond Torque-generated stuff goes here...
// At the end, because it messes with public/private:
TQ_OBJECT_CONSTRUCTORS(JSProxy)
}
// In js-proxy-inl.h:
TQ_OBJECT_CONSTRUCTORS_IMPL(JSProxy)
```
The generated class provides cast functions, field accessor functions, and field offset constants (e.g. `kTargetOffset` and `kHandlerOffset` in this case) representing the byte offset of each field from the beginning of the class.
##### Class type annotations
Some classes can't use the inheritance pattern shown in the example above. In those cases, the class can specify `@doNotGenerateCppClass`, inherit directly from its superclass type, and include a Torque-generated macro for its field offset constants. Such classes must implement their own accessors and cast functions. Using that macro looks like this:
```cpp
class JSProxy : public JSReceiver {
public:
DEFINE_FIELD_OFFSET_CONSTANTS(
JSReceiver::kHeaderSize, TORQUE_GENERATED_JS_PROXY_FIELDS)
// Rest of class omitted...
}
```
`@generateBodyDescriptor` causes Torque to emit a class `BodyDescriptor` within the generated class, which represents how the garbage collector should visit the object. Otherwise the C++ code must either define its own object visitation, or use one of the existing patterns (for example, inheriting from `Struct` and including the class in `STRUCT_LIST` means that the class is expected to contain only tagged values).
If the `@generatePrint` annotation is added, then the generator will implement a C++ function that prints the field values as defined by the Torque layout. Using the JSProxy example, the signature would be `void TorqueGeneratedJSProxy::JSProxyPrint(std::ostream& os)`, which can be inherited by `JSProxy`.
The Torque compiler also generates verification code for all `extern` classes, unless the class opts out with the `@noVerifier` annotation. For example, the JSProxy class definition above will generate a C++ method `void TorqueGeneratedClassVerifiers::JSProxyVerify(JSProxy o, Isolate* isolate)` which verifies that its fields are valid according to the Torque type definition. It will also generate a corresponding function on the generated class, `TorqueGeneratedJSProxy::JSProxyVerify`, which calls the static function from `TorqueGeneratedClassVerifiers`. If you want to add extra verification for a class (such as a range of acceptable values on a number, or a requirement that field `foo` is true if field `bar` is non-null, etc.), then add a `DECL_VERIFIER(JSProxy)` to the C++ class (which hides the inherited `JSProxyVerify`) and implement it in `src/objects-debug.cc`. The first step of any such custom verifier should be to call the generated verifier, such as `TorqueGeneratedClassVerifiers::JSProxyVerify(*this, isolate);`. (To run those verifiers before and after every GC, build with `v8_enable_verify_heap = true` and run with `--verify-heap`.)
`@abstract` indicates that the class itself is not instantiated, and does not have its own instance type: the instance types that logically belong to the class are the instance types of the derived classes.
The `@export` annotation causes the Torque compiler to generate a concrete C++ class (such as `JSProxy` in the example above). This is obviously only useful if you don't want to add any C++ functionality beyond that provided by the Torque-generated code. Cannot be used in conjunction with `extern`. For a class that is defined and used only within Torque, it is most appropriate to use neither `extern` nor `@export`.
`@hasSameInstanceTypeAsParent` indicates classes that have the same instance types as their parent class, but rename some fields, or possibly have a different map. In such cases, the parent class is not abstract.
The annotations `@highestInstanceTypeWithinParentClassRange`, `@lowestInstanceTypeWithinParentClassRange`, `@reserveBitsInInstanceType`, and `@apiExposedInstanceTypeValue` all affect generation of instance types. Generally you can ignore these and be okay. Torque is responsible for assigning a unique value in the enum `v8::internal::InstanceType` for every class so that V8 can determine at runtime the type any object in the JS heap. Torque's assignment of instance types should be adequate in the vast majority of cases, but there are a few cases where we want an instance type for a particular class to be stable across builds, or to be at the beginning or end of the range of instance types assigned to its superclass, or to be a range of reserved values that can be defined outside of Torque.
##### Class fields
As well as plain values, as in the example above, class fields may contain indexed data. Here's an example:
```torque
extern class CoverageInfo extends HeapObject {
const slot_count: int32;
slots[slot_count]: CoverageInfoSlot;
}
```
This means that instances of `CoverageInfo` are of varying sizes based on the data in `slot_count`.
Unlike C++, Torque will not implicitly add padding between fields; instead, it will fail and emit an error if fields are not properly aligned. Torque also requires that strong fields, weak fields, and scalar fields be together with other fields of the same category in the field order.
`const` means that a field cannot be altered at runtime (or at least not easily; Torque will fail compilation if you attempt to set it). This is a good idea for length fields, which should only be reset with great care because they would require freeing any released space and might cause data races with a marking thread.
In fact, Torque requires length fields used for indexed data to be `const`.
`weak` at the beginning of a field declaration means that the field is a custom weak reference, as opposed to the `MaybeObject` tagging mechanism for weak fields.
In addition `weak` affects generation of constants such as `kEndOfStrongFieldsOffset` and `kStartOfWeakFieldsOffset`, which is a legacy feature used in some custom `BodyDescriptor`s and currently also still requires grouping fields marked as `weak` together. We hope to remove this keyword once Torque is fully capable of generating all `BodyDescriptor`s.
If the object stored in a field may be a `MaybeObject`-style weak reference (with the second bit set), then `Weak` should be used in the type and the `weak` keyword should **not** be used. There are still some exceptions to this rule, like this field from `Map`, which can contain some strong and some weak types, and is also marked as `weak` for inclusion in the weak section:
```torque
weak transitions_or_prototype_info: Map|Weak
