Validation Algorithm

The specification of WebAssembly validation is purely declarative. It describes the constraints that must be met by a module or instruction sequence to be valid.

This section sketches the skeleton of a sound and complete algorithm for effectively validating code, i.e., sequences of instructions. (Other aspects of validation are straightforward to implement.)

In fact, the algorithm is expressed over the flat sequence of opcodes as occurring in the binary format, and performs only a single pass over it. Consequently, it can be integrated directly into a decoder.

The algorithm is expressed in typed pseudo code whose semantics is intended to be self-explanatory.

Data Structures

The algorithm uses two separate stacks: the operand stack and the control stack. The former tracks the types of operand values on the stack, the latter surrounding structured control instructions and their associated blocks.

type val_type = I32 | I64 | F32 | F64

type opd_stack = stack(val_type | Unknown)

type ctrl_stack = stack(ctrl_frame)
type ctrl_frame = {
  opcode : opcode
  start_types : list(val_type)
  end_types : list(val_type)
  height : nat
  unreachable : bool
}

For each value, the operand stack records its value type, or Unknown when the type is not known.

For each entered block, the control stack records a control frame with the originating opcode, the types on the top of the operand stack at the start and end of the block (used to check its result as well as branches), the height of the operand stack at the start of the block (used to check that operands do not underflow the current block), and a flag recording whether the remainder of the block is unreachable (used to handle stack-polymorphic typing after branches).

For the purpose of presenting the algorithm, the operand and control stacks are simply maintained as global variables:

var opds : opd_stack
var ctrls : ctrl_stack

However, these variables are not manipulated directly by the main checking function, but through a set of auxiliary functions:

func push_opd(type : val_type | Unknown) =
  opds.push(type)

func pop_opd() : val_type | Unknown =
  if (opds.size() = ctrls[0].height && ctrls[0].unreachable) return Unknown
  error_if(opds.size() = ctrls[0].height)
  return opds.pop()

func pop_opd(expect : val_type | Unknown) : val_type | Unknown =
  let actual = pop_opd()
  if (actual = Unknown) return expect
  if (expect = Unknown) return actual
  error_if(actual =/= expect)
  return actual

func push_opds(types : list(val_type)) = foreach (t in types) push_opd(t)
func pop_opds(types : list(val_type)) = foreach (t in reverse(types)) pop_opd(t)

Pushing an operand simply pushes the respective type to the operand stack.

Popping an operand checks that the operand stack does not underflow the current block and then removes one type. But first, a special case is handled where the block contains no known operands, but has been marked as unreachable. That can occur after an unconditional branch, when the stack is typed polymorphically. In that case, an unknown type is returned.

A second function for popping an operand takes an expected type, which the actual operand type is checked against. The types may differ in case one of them is Unknown. The more specific type is returned.

Finally, there are accumulative functions for pushing or popping multiple operand types.

Note

The notation stack[i] is meant to index the stack from the top, so that ctrls[0] accesses the element pushed last.

The control stack is likewise manipulated through auxiliary functions:

func push_ctrl(opcode : opcode, in : list(val_type), out : list(val_type)) =
  let frame = ctrl_frame(opcode, in, out, opds.size(), false)
  ctrls.push(frame)
  push_opds(in)

func pop_ctrl() : ctrl_frame =
  error_if(ctrls.is_empty())
  let frame = ctrls[0]
  pop_opds(frame.end_types)
  error_if(opds.size() =/= frame.height)
  ctrls.pop()
  return frame

func label_types(frame : ctrl_frame) : list(val_types) =
  return (if frame.opcode == loop then frame.start_types else frame.end_types)

func unreachable() =
  opds.resize(ctrls[0].height)
  ctrls[0].unreachable := true

Pushing a control frame takes the types of the label and result values. It allocates a new frame record recording them along with the current height of the operand stack and marks the block as reachable.

Popping a frame first checks that the control stack is not empty. It then verifies that the operand stack contains the right types of values expected at the end of the exited block and pops them off the operand stack. Afterwards, it checks that the stack has shrunk back to its initial height.

The type of the label associated with a control frame is either that of the stack at the start or the end of the frame, determined by the opcode that it originates from.

Finally, the current frame can be marked as unreachable. In that case, all existing operand types are purged from the operand stack, in order to allow for the stack-polymorphism logic in pop_opd to take effect.

Note

Even with the unreachable flag set, consecutive operands are still pushed to and popped from the operand stack. That is necessary to detect invalid examples like \((\href{../syntax/instructions.html#syntax-instr-control}{\mathsf{unreachable}}~(\href{../syntax/types.html#syntax-valtype}{\mathsf{i32}}.\href{../syntax/instructions.html#syntax-instr-numeric}{\mathsf{const}})~\href{../syntax/types.html#syntax-valtype}{\mathsf{i64}}.\href{../syntax/instructions.html#syntax-instr-numeric}{\mathsf{add}})\). However, a polymorphic stack cannot underflow, but instead generates Unknown types as needed.

Validation of Opcode Sequences

The following function shows the validation of a number of representative instructions that manipulate the stack. Other instructions are checked in a similar manner.

Note

Various instructions not shown here will additionally require the presence of a validation context for checking uses of indices. That is an easy addition and therefore omitted from this presentation.

func validate(opcode) =
  switch (opcode)
    case (i32.add)
      pop_opd(I32)
      pop_opd(I32)
      push_opd(I32)

    case (drop)
      pop_opd()

    case (select)
      pop_opd(I32)
      let t1 = pop_opd()
      let t2 = pop_opd(t1)
      push_opd(t2)

    case (unreachable)
      unreachable()

    case (block t1*->t2*)
      pop_opds([t1*])
      push_ctrl(block, [t1*], [t2*])

    case (loop t1*->t2*)
      pop_opds([t1*])
      push_ctrl(loop, [t1*], [t2*])

    case (if t1*->t2*)
      pop_opd(I32)
      pop_opds([t1*])
      push_ctrl(if, [t1*], [t2*])

    case (end)
      let frame = pop_ctrl()
      push_opds(frame.end_types)

    case (else)
      let frame = pop_ctrl()
      error_if(frame.opcode =/= if)
      push_ctrl(else, frame.start_types, frame.end_types)

    case (br n)
      error_if(ctrls.size() < n)
      pop_opds(label_types(ctrls[n]))
      unreachable()

    case (br_if n)
      error_if(ctrls.size() < n)
      pop_opd(I32)
      pop_opds(label_types(ctrls[n]))
      push_opds(label_types(ctrls[n]))

    case (br_table n* m)
      error_if(ctrls.size() < m)
      foreach (n in n*)
        error_if(ctrls.size() < n || label_types(ctrls[n]) =/= label_types(ctrls[m]))
      pop_opd(I32)
      pop_opds(label_types(ctrls[m]))
      unreachable()

Note

It is an invariant under the current WebAssembly instruction set that an operand of Unknown type is never duplicated on the stack. This would change if the language were extended with stack instructions like dup. Under such an extension, the above algorithm would need to be refined by replacing the Unknown type with proper type variables to ensure that all uses are consistent.