1.1.1. Design Goals

The design goals of WebAssembly are the following:

• Fast, safe, and portable semantics:

  • Fast: executes with near native code performance, taking advantage of capabilities common to all contemporary hardware.

  • Safe: code is validated and executes in a memory-safe 2, sandboxed environment preventing data corruption or security breaches.

  • Well-defined: fully and precisely defines valid programs and their behavior in a way that is easy to reason about informally and formally.

  • Hardware-independent: can be compiled on all modern architectures, desktop or mobile devices and embedded systems alike.

  • Language-independent: does not privilege any particular language, programming model, or object model.

  • Platform-independent: can be embedded in browsers, run as a stand-alone VM, or integrated in other environments.

  • Open: programs can interoperate with their environment in a simple and universal manner.

• Efficient and portable representation:

  • Compact: has a binary format that is fast to transmit by being smaller than typical text or native code formats.

  • Modular: programs can be split up in smaller parts that can be transmitted, cached, and consumed separately.

  • Efficient: can be decoded, validated, and compiled in a fast single pass, equally with either just-in-time (JIT) or ahead-of-time (AOT) compilation.

  • Streamable: allows decoding, validation, and compilation to begin as soon as possible, before all data has been seen.

  • Parallelizable: allows decoding, validation, and compilation to be split into many independent parallel tasks.

  • Portable: makes no architectural assumptions that are not broadly supported across modern hardware.

WebAssembly code is also intended to be easy to inspect and debug, especially in environments like web browsers, but such features are beyond the scope of this specification.

1

A contraction of “WebAssembly”, not an acronym, hence not using all-caps.

2

No program can break WebAssembly’s memory model. Of course, it cannot guarantee that an unsafe language compiling to WebAssembly does not corrupt its own memory layout, e.g. inside WebAssembly’s linear memory.

1.1.2. Scope

At its core, WebAssembly is a virtual instruction set architecture (virtual ISA). As such, it has many use cases and can be embedded in many different environments. To encompass their variety and enable maximum reuse, the WebAssembly specification is split and layered into several documents.

This document is concerned with the core ISA layer of WebAssembly. It defines the instruction set, binary encoding, validation, and execution semantics, as well as a textual representation. It does not, however, define how WebAssembly programs can interact with a specific environment they execute in, nor how they are invoked from such an environment.

Instead, this specification is complemented by additional documents defining interfaces to specific embedding environments such as the Web. These will each define a WebAssembly application programming interface (API) suitable for a given environment.

1.1.3. Security Considerations

WebAssembly provides no ambient access to the computing environment in which code is executed. Any interaction with the environment, such as I/O, access to resources, or operating system calls, can only be performed by invoking functions provided by the embedder and imported into a WebAssembly module. An embedder can establish security policies suitable for a respective environment by controlling or limiting which functional capabilities it makes available for import. Such considerations are an embedder’s responsibility and the subject of API definitions for a specific environment.

Because WebAssembly is designed to be translated into machine code running directly on the host’s hardware, it is potentially vulnerable to side channel attacks on the hardware level. In environments where this is a concern, an embedder may have to put suitable mitigations into place to isolate WebAssembly computations.

1.1.4. Dependencies

WebAssembly depends on two existing standards:

• [IEEE-754-2019], for the representation of floating-point data and the semantics of respective numeric operations.

• [UNICODE], for the representation of import/export names and the text format.

However, to make this specification self-contained, relevant aspects of the aforementioned standards are defined and formalized as part of this specification, such as the binary representation and rounding of floating-point values, and the value range and UTF-8 encoding of Unicode characters.

1.2.1. Concepts

WebAssembly encodes a low-level, assembly-like programming language. This language is structured around the following concepts.

Values

WebAssembly provides only four basic number types. These are integers and [IEEE-754-2019] numbers, each in 32 and 64 bit width. 32 bit integers also serve as Booleans and as memory addresses. The usual operations on these types are available, including the full matrix of conversions between them. There is no distinction between signed and unsigned integer types. Instead, integers are interpreted by respective operations as either unsigned or signed in two’s complement representation.

In addition to these basic number types, there is a single 128 bit wide vector type representing different types of packed data. The supported representations are 4 32-bit, or 2 64-bit [IEEE-754-2019] numbers, or different widths of packed integer values, specifically 2 64-bit integers, 4 32-bit integers, 8 16-bit integers, or 16 8-bit integers.

Finally, values can consist of opaque references that represent pointers towards different sorts of entities. Unlike with other types, their size or representation is not observable.

Instructions

The computational model of WebAssembly is based on a stack machine. Code consists of sequences of instructions that are executed in order. Instructions manipulate values on an implicit operand stack 1 and fall into two main categories. Simple instructions perform basic operations on data. They pop arguments from the operand stack and push results back to it. Control instructions alter control flow. Control flow is structured, meaning it is expressed with well-nested constructs such as blocks, loops, and conditionals. Branches can only target such constructs.

Traps

Under some conditions, certain instructions may produce a trap, which immediately aborts execution. Traps cannot be handled by WebAssembly code, but are reported to the outside environment, where they typically can be caught.

Functions

Code is organized into separate functions. Each function takes a sequence of values as parameters and returns a sequence of values as results. Functions can call each other, including recursively, resulting in an implicit call stack that cannot be accessed directly. Functions may also declare mutable local variables that are usable as virtual registers.

Tables

A table is an array of opaque values of a particular element type. It allows programs to select such values indirectly through a dynamic index operand. Currently, the only available element type is an untyped function reference or a reference to an external host value. Thereby, a program can call functions indirectly through a dynamic index into a table. For example, this allows emulating function pointers by way of table indices.

Linear Memory

A linear memory is a contiguous, mutable array of raw bytes. Such a memory is created with an initial size but can be grown dynamically. A program can load and store values from/to a linear memory at any byte address (including unaligned). Integer loads and stores can specify a storage size which is smaller than the size of the respective value type. A trap occurs if an access is not within the bounds of the current memory size.

Modules

A WebAssembly binary takes the form of a module that contains definitions for functions, tables, and linear memories, as well as mutable or immutable global variables. Definitions can also be imported, specifying a module/name pair and a suitable type. Each definition can optionally be exported under one or more names. In addition to definitions, modules can define initialization data for their memories or tables that takes the form of segments copied to given offsets. They can also define a start function that is automatically executed.

Embedder

A WebAssembly implementation will typically be embedded into a host environment. This environment defines how loading of modules is initiated, how imports are provided (including host-side definitions), and how exports can be accessed. However, the details of any particular embedding are beyond the scope of this specification, and will instead be provided by complementary, environment-specific API definitions.

1

In practice, implementations need not maintain an actual operand stack. Instead, the stack can be viewed as a set of anonymous registers that are implicitly referenced by instructions. The type system ensures that the stack height, and thus any referenced register, is always known statically.

1.2.2. Semantic Phases

Conceptually, the semantics of WebAssembly is divided into three phases. For each part of the language, the specification specifies each of them.

Decoding

WebAssembly modules are distributed in a binary format. Decoding processes that format and converts it into an internal representation of a module. In this specification, this representation is modelled by abstract syntax, but a real implementation could compile directly to machine code instead.

Validation

A decoded module has to be valid. Validation checks a number of well-formedness conditions to guarantee that the module is meaningful and safe. In particular, it performs type checking of functions and the instruction sequences in their bodies, ensuring for example that the operand stack is used consistently.

Execution

Finally, a valid module can be executed. Execution can be further divided into two phases:

Instantiation. A module instance is the dynamic representation of a module, complete with its own state and execution stack. Instantiation executes the module body itself, given definitions for all its imports. It initializes globals, memories and tables and invokes the module’s start function if defined. It returns the instances of the module’s exports.

Invocation. Once instantiated, further WebAssembly computations can be initiated by invoking an exported function on a module instance. Given the required arguments, that executes the respective function and returns its results.

Instantiation and invocation are operations within the embedding environment.

2.1.1. Grammar Notation

The following conventions are adopted in defining grammar rules for abstract syntax.

• Terminal symbols (atoms) are written in sans-serif font or in symbolic form: $$i32,\mathsf{end},\to ,[,]$$.

• Nonterminal symbols are written in italic font: $$valtype,instr$$.

• $${\mathrm{A}}^{\mathrm{n}}$$ is a sequence of $$\mathrm{n}\ge 0$$ iterations of $$\mathrm{A}$$.

• $${\mathrm{A}}^{\ast }$$ is a possibly empty sequence of iterations of $$\mathrm{A}$$. (This is a shorthand for $${\mathrm{A}}^{\mathrm{n}}$$ used where $$\mathrm{n}$$ is not relevant.)

• $${\mathrm{A}}^{+}$$ is a non-empty sequence of iterations of $$\mathrm{A}$$. (This is a shorthand for $${\mathrm{A}}^{\mathrm{n}}$$ where $$\mathrm{n}\ge 1$$.)

• $${\mathrm{A}}^{?}$$ is an optional occurrence of $$\mathrm{A}$$. (This is a shorthand for $${\mathrm{A}}^{\mathrm{n}}$$ where $$\mathrm{n}\le 1$$.)

• Productions are written $$sym::={\mathrm{A}}_1|\dots |{\mathrm{A}}_{\mathrm{n}}$$.

• Large productions may be split into multiple definitions, indicated by ending the first one with explicit ellipses, $$sym::={\mathrm{A}}_1|\dots$$, and starting continuations with ellipses, $$sym::=\dots |{\mathrm{A}}_2$$.

• Some productions are augmented with side conditions in parentheses, “$$(\text{if}condition)$$”, that provide a shorthand for a combinatorial expansion of the production into many separate cases.

• If the same meta variable or non-terminal symbol appears multiple times in a production, then all those occurrences must have the same instantiation. (This is a shorthand for a side condition requiring multiple different variables to be equal.)

2.1.2. Auxiliary Notation

When dealing with syntactic constructs the following notation is also used:

• $$ϵ$$ denotes the empty sequence.

• $$|s|$$ denotes the length of a sequence $$\mathrm{s}$$.

• $$\mathrm{s}[\mathrm{i}]$$ denotes the $$\mathrm{i}$$-th element of a sequence $$\mathrm{s}$$, starting from $$0$$.

• $$\mathrm{s}[\mathrm{i}:\mathrm{n}]$$ denotes the sub-sequence $$\mathrm{s}[\mathrm{i}]\dots \mathrm{s}[\mathrm{i}+\mathrm{n}-1]$$ of a sequence $$\mathrm{s}$$.

• $$\mathrm{s}\text{with}[\mathrm{i}]=\mathrm{A}$$ denotes the same sequence as $$\mathrm{s}$$, except that the $$\mathrm{i}$$-th element is replaced with $$\mathrm{A}$$.

• $$\mathrm{s}\text{with}[\mathrm{i}:\mathrm{n}]={\mathrm{A}}^{\mathrm{n}}$$ denotes the same sequence as $$\mathrm{s}$$, except that the sub-sequence $$\mathrm{s}[\mathrm{i}:\mathrm{n}]$$ is replaced with $${\mathrm{A}}^{\mathrm{n}}$$.

• $$\mathrm{concat}({\mathrm{s}}^{\ast })$$ denotes the flat sequence formed by concatenating all sequences $${\mathrm{s}}_{\mathrm{i}}$$ in $${\mathrm{s}}^{\ast }$$.

Moreover, the following conventions are employed:

• The notation $${\mathrm{x}}^{\mathrm{n}}$$, where $$\mathrm{x}$$ is a non-terminal symbol, is treated as a meta variable ranging over respective sequences of $$\mathrm{x}$$ (similarly for $${\mathrm{x}}^{\ast }$$, $${\mathrm{x}}^{+}$$, $${\mathrm{x}}^{?}$$).

• When given a sequence $${\mathrm{x}}^{\mathrm{n}}$$, then the occurrences of $$\mathrm{x}$$ in a sequence written $$({\mathrm{A}}_1\mathrm{x}{\mathrm{A}}_2{)}^{\mathrm{n}}$$ are assumed to be in point-wise correspondence with $${\mathrm{x}}^{\mathrm{n}}$$ (similarly for $${\mathrm{x}}^{\ast }$$, $${\mathrm{x}}^{+}$$, $${\mathrm{x}}^{?}$$). This implicitly expresses a form of mapping syntactic constructions over a sequence.

Productions of the following form are interpreted as records that map a fixed set of fields $${\mathsf{field}}_{\mathrm{i}}$$ to “values” $${\mathrm{A}}_{\mathrm{i}}$$, respectively:

The following notation is adopted for manipulating such records:

• $$r.\mathsf{field}$$ denotes the contents of the $$\mathsf{field}$$ component of $$\mathrm{r}$$.

• $$\mathrm{r}\text{with}\mathsf{field}=\mathrm{A}$$ denotes the same record as $$\mathrm{r}$$, except that the contents of the $$\mathsf{field}$$ component is replaced with $$\mathrm{A}$$.

• $${\mathrm{r}}_1\oplus {\mathrm{r}}_2$$ denotes the composition of two records with the same fields of sequences by appending each sequence point-wise:

  $$\{{\mathsf{field}}_1{\mathrm{A}}_1^{\ast },{\mathsf{field}}_2{\mathrm{A}}_2^{\ast },\dots \}\oplus \{{\mathsf{field}}_1{\mathrm{B}}_1^{\ast },{\mathsf{field}}_2{\mathrm{B}}_2^{\ast },\dots \}=\{{\mathsf{field}}_1{\mathrm{A}}_1^{\ast }{\mathrm{B}}_1^{\ast },{\mathsf{field}}_2{\mathrm{A}}_2^{\ast }{\mathrm{B}}_2^{\ast },\dots \}$$

• $$⨁{\mathrm{r}}^{\ast }$$ denotes the composition of a sequence of records, respectively; if the sequence is empty, then all fields of the resulting record are empty.

The update notation for sequences and records generalizes recursively to nested components accessed by “paths” $$pth::=([\dots ]|.\mathsf{field}{)}^{+}$$:

• $$\mathrm{s}\text{with}[\mathrm{i}]pth=\mathrm{A}$$ is short for $$\mathrm{s}\text{with}[\mathrm{i}]=(\mathrm{s}[\mathrm{i}]\text{with}pth=\mathrm{A})$$,

• $$\mathrm{r}\text{with}\mathsf{field}pth=\mathrm{A}$$ is short for $$\mathrm{r}\text{with}\mathsf{field}=(r.\mathsf{field}\text{with}pth=\mathrm{A})$$,

where $$\mathrm{r}\text{with}.\mathsf{field}=\mathrm{A}$$ is shortened to $$\mathrm{r}\text{with}\mathsf{field}=\mathrm{A}$$.

2.1.3. Vectors

Vectors are bounded sequences of the form $${\mathrm{A}}^{\mathrm{n}}$$ (or $${\mathrm{A}}^{\ast }$$), where the $$\mathrm{A}$$ can either be values or complex constructions. A vector can have at most $$2^{32}-1$$ elements.

2.2.1. Bytes

The simplest form of value are raw uninterpreted bytes. In the abstract syntax they are represented as hexadecimal literals.

2.2.2. Integers

Different classes of integers with different value ranges are distinguished by their bit width $$\mathrm{N}$$ and by whether they are unsigned or signed.

The class $$i\mathrm{N}$$ defines uninterpreted integers, whose signedness interpretation can vary depending on context. In the abstract syntax, they are represented as unsigned values. However, some operations convert them to signed based on a two’s complement interpretation.

2.2.3. Floating-Point

Floating-point data represents 32 or 64 bit values that correspond to the respective binary formats of the [IEEE-754-2019] standard (Section 3.3).

Every value has a sign and a magnitude. Magnitudes can either be expressed as normal numbers of the form $${\mathrm{m}}_0.{\mathrm{m}}_1{\mathrm{m}}_2\dots {\mathrm{m}}_{\mathrm{M}}\cdot 2^{\mathrm{e}}$$, where $$\mathrm{e}$$ is the exponent and $$\mathrm{m}$$ is the significand whose most significant bit $${\mathrm{m}}_0$$ is $$1$$, or as a subnormal number where the exponent is fixed to the smallest possible value and $${\mathrm{m}}_0$$ is $$0$$; among the subnormals are positive and negative zero values. Since the significands are binary values, normals are represented in the form $$(1+\mathrm{m}\cdot 2^{-\mathrm{M}})\cdot 2^{\mathrm{e}}$$, where $$\mathrm{M}$$ is the bit width of $$\mathrm{m}$$; similarly for subnormals.

Possible magnitudes also include the special values $$\infty$$ (infinity) and $$\mathsf{nan}$$ (NaN, not a number). NaN values have a payload that describes the mantissa bits in the underlying binary representation. No distinction is made between signalling and quiet NaNs.

where $$\mathrm{M}=\mathrm{signif}(\mathrm{N})$$ and $$\mathrm{E}=\mathrm{expon}(\mathrm{N})$$ with

A canonical NaN is a floating-point value $$\pm \mathsf{nan}({\mathrm{canon}}_{\mathrm{N}})$$ where $${\mathrm{canon}}_{\mathrm{N}}$$ is a payload whose most significant bit is $$1$$ while all others are $$0$$:

An arithmetic NaN is a floating-point value $$\pm \mathsf{nan}(\mathrm{n})$$ with $$\mathrm{n}\ge {\mathrm{canon}}_{\mathrm{N}}$$, such that the most significant bit is $$1$$ while all others are arbitrary.

2.2.4. Vectors

Numeric vectors are 128-bit values that are processed by vector instructions (also known as SIMD instructions, single instruction multiple data). They are represented in the abstract syntax using $$i128$$. The interpretation of lane types (integer or floating-point numbers) and lane sizes are determined by the specific instruction operating on them.

2.2.5. Names

Names are sequences of characters, which are scalar values as defined by [UNICODE] (Section 2.4).

Due to the limitations of the binary format, the length of a name is bounded by the length of its UTF-8 encoding.

2.3.1. Number Types

Number types classify numeric values.

The types $$i32$$ and $$i64$$ classify 32 and 64 bit integers, respectively. Integers are not inherently signed or unsigned, their interpretation is determined by individual operations.

The types $$f32$$ and $$f64$$ classify 32 and 64 bit floating-point data, respectively. They correspond to the respective binary floating-point representations, also known as single and double precision, as defined by the [IEEE-754-2019] standard (Section 3.3).

Number types are transparent, meaning that their bit patterns can be observed. Values of number type can be stored in memories.

2.3.2. Vector Types

Vector types classify vectors of numeric values processed by vector instructions (also known as SIMD instructions, single instruction multiple data).

The type $$v128$$ corresponds to a 128 bit vector of packed integer or floating-point data. The packed data can be interpreted as signed or unsigned integers, single or double precision floating-point values, or a single 128 bit type. The interpretation is determined by individual operations.

Vector types, like number types are transparent, meaning that their bit patterns can be observed. Values of vector type can be stored in memories.

2.3.3. Reference Types

Reference types classify first-class references to objects in the runtime store.

The type $$\mathsf{funcref}$$ denotes the infinite union of all references to functions, regardless of their function types.

The type $$\mathsf{externref}$$ denotes the infinite union of all references to objects owned by the embedder and that can be passed into WebAssembly under this type.

Reference types are opaque, meaning that neither their size nor their bit pattern can be observed. Values of reference type can be stored in tables.

2.3.4. Value Types

Value types classify the individual values that WebAssembly code can compute with and the values that a variable accepts. They are either number types, vector types, or reference types.

2.3.5. Result Types

Result types classify the result of executing instructions or functions, which is a sequence of values, written with brackets.

2.3.6. Function Types

Function types classify the signature of functions, mapping a vector of parameters to a vector of results. They are also used to classify the inputs and outputs of instructions.

2.3.7. Limits

Limits classify the size range of resizeable storage associated with memory types and table types.

If no maximum is given, the respective storage can grow to any size.

2.3.8. Memory Types

Memory types classify linear memories and their size range.

The limits constrain the minimum and optionally the maximum size of a memory. The limits are given in units of page size. The memory type also determines whether this memory is shared.

2.3.9. Table Types

Table types classify tables over elements of reference type within a size range.

Like memories, tables are constrained by limits for their minimum and optionally maximum size. The limits are given in numbers of entries.

2.3.10. Global Types

Global types classify global variables, which hold a value and can either be mutable or immutable.

2.3.11. External Types

External types classify imports and external values with their respective types.

2.4.1. Numeric Instructions

Numeric instructions provide basic operations over numeric values of specific type. These operations closely match respective operations available in hardware.

Numeric instructions are divided by number type. For each type, several subcategories can be distinguished:

• Constants: return a static constant.

• Unary Operations: consume one operand and produce one result of the respective type.

• Binary Operations: consume two operands and produce one result of the respective type.

• Tests: consume one operand of the respective type and produce a Boolean integer result.

• Comparisons: consume two operands of the respective type and produce a Boolean integer result.

• Conversions: consume a value of one type and produce a result of another (the source type of the conversion is the one after the “$$\_$$”).

Some integer instructions come in two flavors, where a signedness annotation $$sx$$ distinguishes whether the operands are to be interpreted as unsigned or signed integers. For the other integer instructions, the use of two’s complement for the signed interpretation means that they behave the same regardless of signedness.

2.4.2. Vector Instructions

Vector instructions (also known as SIMD instructions, single instruction multiple data) provide basic operations over values of vector type.

Vector instructions have a naming convention involving a prefix that determines how their operands will be interpreted. This prefix describes the shape of the operand, written $$\mathrm{t}\mathsf{x}\mathrm{N}$$, and consisting of a packed numeric type $$\mathrm{t}$$ and the number of lanes $$\mathrm{N}$$ of that type. Operations are performed point-wise on the values of each lane.

Instructions prefixed with $$v128$$ do not involve a specific interpretation, and treat the $$v128$$ as an $$i128$$ value or a vector of 128 individual bits.

Vector instructions can be grouped into several subcategories:

• Constants: return a static constant.

• Unary Operations: consume one $$v128$$ operand and produce one $$v128$$ result.

• Binary Operations: consume two $$v128$$ operands and produce one $$v128$$ result.

• Ternary Operations: consume three $$v128$$ operands and produce one $$v128$$ result.

• Tests: consume one $$v128$$ operand and produce a Boolean integer result.

• Shifts: consume a $$v128$$ operand and a $$i32$$ operand, producing one $$v128$$ result.

• Splats: consume a value of numeric type and produce a $$v128$$ result of a specified shape.

• Extract lanes: consume a $$v128$$ operand and return the numeric value in a given lane.

• Replace lanes: consume a $$v128$$ operand and a numeric value for a given lane, and produce a $$v128$$ result.

Some vector instructions have a signedness annotation $$sx$$ which distinguishes whether the elements in the operands are to be interpreted as unsigned or signed integers. For the other vector instructions, the use of two’s complement for the signed interpretation means that they behave the same regardless of signedness.

2.4.3. Reference Instructions

Instructions in this group are concerned with accessing references.

These instructions produce a null value, check for a null value, or produce a reference to a given function, respectively.

2.4.4. Parametric Instructions

Instructions in this group can operate on operands of any value type.

The $$\mathsf{drop}$$ instruction simply throws away a single operand.

The $$\mathsf{select}$$ instruction selects one of its first two operands based on whether its third operand is zero or not. It may include a value type determining the type of these operands. If missing, the operands must be of numeric type.

2.4.5. Variable Instructions

Variable instructions are concerned with access to local or global variables.

These instructions get or set the values of variables, respectively. The $$local.tee$$ instruction is like $$local.set$$ but also returns its argument.

2.4.6. Table Instructions

Instructions in this group are concerned with tables table.

The $$table.get$$ and $$table.set$$ instructions load or store an element in a table, respectively.

The $$table.size$$ instruction returns the current size of a table. The $$table.grow$$ instruction grows table by a given delta and returns the previous size, or $$-1$$ if enough space cannot be allocated. It also takes an initialization value for the newly allocated entries.

The $$table.fill$$ instruction sets all entries in a range to a given value.

The $$table.copy$$ instruction copies elements from a source table region to a possibly overlapping destination region; the first index denotes the destination. The $$table.init$$ instruction copies elements from a passive element segment into a table. The $$elem.drop$$ instruction prevents further use of a passive element segment. This instruction is intended to be used as an optimization hint. After an element segment is dropped its elements can no longer be retrieved, so the memory used by this segment may be freed.

An additional instruction that accesses a table is the control instruction $$call\_indirect$$.

2.4.7. Memory Instructions

Instructions in this group are concerned with linear memory.

Memory is accessed with $$\mathsf{load}$$ and $$\mathsf{store}$$ instructions for the different number types. They all take a memory immediate $$memarg$$ that contains an address offset and the expected alignment (expressed as the exponent of a power of 2). Integer loads and stores can optionally specify a storage size that is smaller than the bit width of the respective value type. In the case of loads, a sign extension mode $$sx$$ is then required to select appropriate behavior.

Vector loads can specify a shape that is half the bit width of $$v128$$. Each lane is half its usual size, and the sign extension mode $$sx$$ then specifies how the smaller lane is extended to the larger lane. Alternatively, vector loads can perform a splat, such that only a single lane of the specified storage size is loaded, and the result is duplicated to all lanes.

The static address offset is added to the dynamic address operand, yielding a 33 bit effective address that is the zero-based index at which the memory is accessed. All values are read and written in little endian byte order. A trap results if any of the accessed memory bytes lies outside the address range implied by the memory’s current size.

The $$memory.size$$ instruction returns the current size of a memory. The $$memory.grow$$ instruction grows memory by a given delta and returns the previous size, or $$-1$$ if enough memory cannot be allocated. Both instructions operate in units of page size.

The $$memory.fill$$ instruction sets all values in a region to a given byte. The $$memory.copy$$ instruction copies data from a source memory region to a possibly overlapping destination region. The $$memory.init$$ instruction copies data from a passive data segment into a memory. The $$data.drop$$ instruction prevents further use of a passive data segment. This instruction is intended to be used as an optimization hint. After a data segment is dropped its data can no longer be retrieved, so the memory used by this segment may be freed.

2.4.8. Atomic Memory Instructions

Instructions in this group are concerned with accessing linear memory atomically.

Memory is accessed atomically using the $$atomic.load$$, $$atomic.store$$, and $$atomic.rmw$$ instructions. All instructions take a memory immediate $$memarg$$, just like their non-atomic equivalents. Unlike non-atomic memory access instructions, only integer value types can be used. Also unlike non-atomic memory access instructions, there are no sign extension modes; atomic memory accesses are always zero-extending.

The $$atomic.rmw$$ instructions are read-modify-write instructions. They each have an atomic operator, which specifies how memory will be modified. Each instruction returns the value read from memory before modification. The $$\mathsf{xchg}$$ operator doesn’t use the read value, but instead stores its argument unmodified. The $$\mathsf{cmpxchg}$$ operator is similar, but only performs this action conditionally, if the read value is equal to a provided comparison argument. All other atomic operators have the same behavior as the binary operator of the same name.

The $$memory.atomic.wait$$, $$memory.atomic.notify$$, and $$atomic.fence$$ instructions provide primitive synchronization between threads. The $$memory.atomic.wait$$ instructions atomically load a value from the calculated effective address and compare it to an expected value. If they are equal, the thread is then suspended until a given timeout expires or another thread wakes it. The $$memory.atomic.notify$$ instruction wakes threads that are waiting on a given address, up to a given maximum. The $$atomic.fence$$ instruction takes no operands, and returns nothing. It is intended to preserve the synchronization guarantees of the fence operators of higher-level languages. Unlike other atomic operators, it does not target a particular linear memory.

2.4.9. Control Instructions

Instructions in this group affect the flow of control.

The $$\mathsf{nop}$$ instruction does nothing.

The $$\mathsf{unreachable}$$ instruction causes an unconditional trap.

The $$\mathsf{block}$$, $$\mathsf{loop}$$ and $$\mathsf{if}$$ instructions are structured instructions. They bracket nested sequences of instructions, called blocks, terminated with, or separated by, $$\mathsf{end}$$ or $$\mathsf{else}$$ pseudo-instructions. As the grammar prescribes, they must be well-nested.

A structured instruction can consume input and produce output on the operand stack according to its annotated block type. It is given either as a type index that refers to a suitable function type, or as an optional value type inline, which is a shorthand for the function type $$[]\to [{valtype}^{?}]$$.

Each structured control instruction introduces an implicit label. Labels are targets for branch instructions that reference them with label indices. Unlike with other index spaces, indexing of labels is relative by nesting depth, that is, label $$0$$ refers to the innermost structured control instruction enclosing the referring branch instruction, while increasing indices refer to those farther out. Consequently, labels can only be referenced from within the associated structured control instruction. This also implies that branches can only be directed outwards, “breaking” from the block of the control construct they target. The exact effect depends on that control construct. In case of $$\mathsf{block}$$ or $$\mathsf{if}$$ it is a forward jump, resuming execution after the matching $$\mathsf{end}$$. In case of $$\mathsf{loop}$$ it is a backward jump to the beginning of the loop.

Branch instructions come in several flavors: $$\mathsf{br}$$ performs an unconditional branch, $$br\_if$$ performs a conditional branch, and $$br\_table$$ performs an indirect branch through an operand indexing into the label vector that is an immediate to the instruction, or to a default target if the operand is out of bounds. The $$\mathsf{return}$$ instruction is a shortcut for an unconditional branch to the outermost block, which implicitly is the body of the current function. Taking a branch unwinds the operand stack up to the height where the targeted structured control instruction was entered. However, branches may additionally consume operands themselves, which they push back on the operand stack after unwinding. Forward branches require operands according to the output of the targeted block’s type, i.e., represent the values produced by the terminated block. Backward branches require operands according to the input of the targeted block’s type, i.e., represent the values consumed by the restarted block.

The $$\mathsf{call}$$ instruction invokes another function, consuming the necessary arguments from the stack and returning the result values of the call. The $$call\_indirect$$ instruction calls a function indirectly through an operand indexing into a table that is denoted by a table index and must have type $$\mathsf{funcref}$$. Since it may contain functions of heterogeneous type, the callee is dynamically checked against the function type indexed by the instruction’s second immediate, and the call is aborted with a trap if it does not match.

2.4.10. Expressions

Function bodies, initialization values for globals, and offsets of element or data segments are given as expressions, which are sequences of instructions terminated by an $$\mathsf{end}$$ marker.

In some places, validation restricts expressions to be constant, which limits the set of allowable instructions.

2.5.1. Indices

Definitions are referenced with zero-based indices. Each class of definition has its own index space, as distinguished by the following classes.

The index space for functions, tables, memories and globals includes respective imports declared in the same module. The indices of these imports precede the indices of other definitions in the same index space.

Element indices reference element segments and data indices reference data segments.

The index space for locals is only accessible inside a function and includes the parameters of that function, which precede the local variables.

Label indices reference structured control instructions inside an instruction sequence.

2.5.2. Types

The $$\mathsf{types}$$ component of a module defines a vector of function types.

All function types used in a module must be defined in this component. They are referenced by type indices.

2.5.3. Functions

The $$\mathsf{funcs}$$ component of a module defines a vector of functions with the following structure:

The $$\mathsf{type}$$ of a function declares its signature by reference to a type defined in the module. The parameters of the function are referenced through 0-based local indices in the function’s body; they are mutable.

The $$\mathsf{locals}$$ declare a vector of mutable local variables and their types. These variables are referenced through local indices in the function’s body. The index of the first local is the smallest index not referencing a parameter.

The $$\mathsf{body}$$ is an instruction sequence that upon termination must produce a stack matching the function type’s result type.

Functions are referenced through function indices, starting with the smallest index not referencing a function import.

2.5.4. Tables

The $$\mathsf{tables}$$ component of a module defines a vector of tables described by their table type:

A table is a vector of opaque values of a particular reference type. The $$\mathsf{min}$$ size in the limits of the table type specifies the initial size of that table, while its $$\mathsf{max}$$, if present, restricts the size to which it can grow later.

Tables can be initialized through element segments.

Tables are referenced through table indices, starting with the smallest index not referencing a table import. Most constructs implicitly reference table index $$0$$.

2.5.5. Memories

The $$\mathsf{mems}$$ component of a module defines a vector of linear memories (or memories for short) as described by their memory type:

A memory is a vector of raw uninterpreted bytes. The $$\mathsf{min}$$ size in the limits of the memory type specifies the initial size of that memory, while its $$\mathsf{max}$$, if present, restricts the size to which it can grow later. Both are in units of page size.

Memories can be initialized through data segments.

Memories are referenced through memory indices, starting with the smallest index not referencing a memory import. Most constructs implicitly reference memory index $$0$$.

2.5.6. Globals

The $$\mathsf{globals}$$ component of a module defines a vector of global variables (or globals for short):

Each global stores a single value of the given global type. Its $$\mathsf{type}$$ also specifies whether a global is immutable or mutable. Moreover, each global is initialized with an $$\mathsf{init}$$ value given by a constant initializer expression.

Globals are referenced through global indices, starting with the smallest index not referencing a global import.

2.5.7. Element Segments

The initial contents of a table is uninitialized. Element segments can be used to initialize a subrange of a table from a static vector of elements.

The $$\mathsf{elems}$$ component of a module defines a vector of element segments. Each element segment defines a reference type and a corresponding list of constant element expressions.

Element segments have a mode that identifies them as either passive, active, or declarative. A passive element segment’s elements can be copied to a table using the $$table.init$$ instruction. An active element segment copies its elements into a table during instantiation, as specified by a table index and a constant expression defining an offset into that table. A declarative element segment is not available at runtime but merely serves to forward-declare references that are formed in code with instructions like $$\mathsf{ref}.\mathsf{func}$$.