Skip to content

Latest commit

 

History

History
1826 lines (1609 loc) · 70 KB

CanonicalABI.md

File metadata and controls

1826 lines (1609 loc) · 70 KB

Canonical ABI Explainer

This document defines the Canonical ABI used to convert between the values and functions of components in the Component Model and the values and functions of modules in Core WebAssembly.

Supporting definitions

The Canonical ABI specifies, for each component function signature, a corresponding core function signature and the process for reading component-level values into and out of linear memory. While a full formal specification would specify the Canonical ABI in terms of macro-expansion into Core WebAssembly instructions augmented with a new set of (spec-internal) administrative instructions, the informal presentation here instead specifies the process in terms of Python code that would be logically executed at validation- and run-time by a component model implementation. The Python code is presented by interleaving definitions with descriptions and eliding some boilerplate. For a complete listing of all Python definitions in a single executable file with a small unit test suite, see the canonical-abi directory.

The convention followed by the Python code below is that all traps are raised by explicit trap()/trap_if() calls; Python assert() statements should never fire and are only included as hints to the reader. Similarly, there should be no uncaught Python exceptions.

While the Python code appears to perform a copy as part of lifting the contents of linear memory into high-level Python values, a normal implementation should never need to make this extra intermediate copy. This claim is expanded upon below.

Lastly, independently of Python, the Canonical ABI defined below assumes that out-of-memory conditions (such as memory.grow returning -1 from within realloc) will trap (via unreachable). This significantly simplifies the Canonical ABI by avoiding the need to support the complicated protocols necessary to support recovery in the middle of nested allocations. In the MVP, for large allocations that can OOM, streams would usually be the appropriate type to use and streams will be able to explicitly express failure in their type. Post-MVP, adapter functions would allow fully custom OOM handling for all component-level types, allowing a toolchain to intentionally propagate OOM into the appropriate explicit return value of the function's declared return type.

Despecialization

In the explainer, component value types are classified as either fundamental or specialized, where the specialized value types are defined by expansion into fundamental value types. In most cases, the canonical ABI of a specialized value type is the same as its expansion so, to avoid repetition, the other definitions below use the following despecialize function to replace specialized value types with their expansion:

def despecialize(t):
  match t:
    case Tuple(ts)         : return Record([ Field(str(i), t) for i,t in enumerate(ts) ])
    case Enum(labels)      : return Variant([ Case(l, None) for l in labels ])
    case Option(t)         : return Variant([ Case("none", None), Case("some", t) ])
    case Result(ok, error) : return Variant([ Case("ok", ok), Case("error", error) ])
    case _                 : return t

The specialized value types string and flags are missing from this list because they are given specialized canonical ABI representations distinct from their respective expansions.

Alignment

Each value type is assigned an alignment which is used by subsequent Canonical ABI definitions. Presenting the definition of alignment piecewise, we start with the top-level case analysis:

def alignment(t):
  match despecialize(t):
    case Bool()             : return 1
    case S8() | U8()        : return 1
    case S16() | U16()      : return 2
    case S32() | U32()      : return 4
    case S64() | U64()      : return 8
    case F32()              : return 4
    case F64()              : return 8
    case Char()             : return 4
    case String() | List(_) : return 4
    case Record(fields)     : return alignment_record(fields)
    case Variant(cases)     : return alignment_variant(cases)
    case Flags(labels)      : return alignment_flags(labels)
    case Own(_) | Borrow(_) : return 4

Record alignment is tuple alignment, with the definitions split for reuse below:

def alignment_record(fields):
  a = 1
  for f in fields:
    a = max(a, alignment(f.t))
  return a

As an optimization, variant discriminants are represented by the smallest integer covering the number of cases in the variant (with cases numbered in order from 0 to len(cases)-1). Depending on the payload type, this can allow more compact representations of variants in memory. This smallest integer type is selected by the following function, used above and below:

def alignment_variant(cases):
  return max(alignment(discriminant_type(cases)), max_case_alignment(cases))

def discriminant_type(cases):
  n = len(cases)
  assert(0 < n < (1 << 32))
  match math.ceil(math.log2(n)/8):
    case 0: return U8()
    case 1: return U8()
    case 2: return U16()
    case 3: return U32()

def max_case_alignment(cases):
  a = 1
  for c in cases:
    if c.t is not None:
      a = max(a, alignment(c.t))
  return a

As an optimization, flags are represented as packed bit-vectors. Like variant discriminants, flags use the smallest integer that fits all the bits, falling back to sequences of i32s when there are more than 32 flags.

def alignment_flags(labels):
  n = len(labels)
  if n <= 8: return 1
  if n <= 16: return 2
  return 4

Handle types are passed as i32 indices into the Table[HandleElem] introduced below.

Element Size

Each value type is also assigned an elem_size which is the number of bytes used when values of the type are stored as elements of a list. Having this byte size be a static property of the type instead of attempting to use a variable-length element-encoding scheme both simplifies the implementation and maps well to languages which represent lists as random-access arrays. Empty types, such as records with no fields, are not permitted, to avoid complications in source languages.

def elem_size(t):
  match despecialize(t):
    case Bool()             : return 1
    case S8() | U8()        : return 1
    case S16() | U16()      : return 2
    case S32() | U32()      : return 4
    case S64() | U64()      : return 8
    case F32()              : return 4
    case F64()              : return 8
    case Char()             : return 4
    case String() | List(_) : return 8
    case Record(fields)     : return elem_size_record(fields)
    case Variant(cases)     : return elem_size_variant(cases)
    case Flags(labels)      : return elem_size_flags(labels)
    case Own(_) | Borrow(_) : return 4

def elem_size_record(fields):
  s = 0
  for f in fields:
    s = align_to(s, alignment(f.t))
    s += elem_size(f.t)
  assert(s > 0)
  return align_to(s, alignment_record(fields))

def align_to(ptr, alignment):
  return math.ceil(ptr / alignment) * alignment

def elem_size_variant(cases):
  s = elem_size(discriminant_type(cases))
  s = align_to(s, max_case_alignment(cases))
  cs = 0
  for c in cases:
    if c.t is not None:
      cs = max(cs, elem_size(c.t))
  s += cs
  return align_to(s, alignment_variant(cases))

def elem_size_flags(labels):
  n = len(labels)
  assert(n > 0)
  if n <= 8: return 1
  if n <= 16: return 2
  return 4 * num_i32_flags(labels)

def num_i32_flags(labels):
  return math.ceil(len(labels) / 32)

Runtime State

The subsequent definitions of loading and storing a value from linear memory require additional runtime state, which is threaded through most subsequent definitions via the cx parameter of type CallContext:

@dataclass
class CallContext:
  opts: CanonicalOptions
  inst: ComponentInstance

The opts field of CallContext contains all the possible canonopt immediates that can be passed to the canon definition being implemented.

@dataclass
class CanonicalOptions:
  memory: Optional[bytearray] = None
  string_encoding: Optional[str] = None
  realloc: Optional[Callable] = None
  post_return: Optional[Callable] = None

The inst field of CallContext points to the component instance which the canon-generated function is closed over. Component instances contain all the core wasm instance as well as some extra state that is used exclusively by the Canonical ABI:

class ComponentInstance:
  # core module instance state
  may_leave: bool
  may_enter: bool
  handles: HandleTables

  def __init__(self):
    self.may_leave = True
    self.may_enter = True
    self.handles = HandleTables()

The may_enter and may_leave fields are used to enforce the component invariants: may_leave indicates whether the instance may call out to an import and the may_enter state indicates whether the instance may be called from the outside world through an export.

The handles field of ComponentInstance contains a mapping from ResourceType to Tables of HandleElems (defined next), establishing a separate i32-indexed array per resource type.

class HandleTables:
  rt_to_table: MutableMapping[ResourceType, Table[HandleElem]]

  def __init__(self):
    self.rt_to_table = dict()

  def table(self, rt):
    if rt not in self.rt_to_table:
      self.rt_to_table[rt] = Table[HandleElem]()
    return self.rt_to_table[rt]

  def get(self, rt, i):
    return self.table(rt).get(i)
  def add(self, rt, h):
    return self.table(rt).add(h)
  def remove(self, rt, i):
    return self.table(rt).remove(i)

While this Python code performs a dynamic hash-table lookup on each handle table access, as we'll see below, the rt parameter is always statically known such that a normal implementation can statically enumerate all Table objects at compile time and then route the calls to get, add and remove to the correct Table at the callsite. The net result is that each component instance will contain one handle table per resource type used by the component, with each compiled adapter function accessing the correct handle table as-if it were a global variable.

The ResourceType class represents a concrete resource type that has been created by the component instance impl. ResourceType objects are used as keys by HandleTables above and thus we assume that Python object identity corresponds to resource type equality, as defined by [type checking] rules.

class ResourceType(Type):
  impl: ComponentInstance
  dtor: Optional[Callable]

  def __init__(self, impl, dtor = None):
    self.impl = impl
    self.dtor = dtor

The Table class, used by HandleTables above, encapsulates a single mutable, growable array of generic elements, indexed by Core WebAssembly i32s.

ElemT = TypeVar('ElemT')
class Table(Generic[ElemT]):
  array: list[Optional[ElemT]]
  free: list[int]

  def __init__(self):
    self.array = [None]
    self.free = []

  def get(self, i):
    trap_if(i >= len(self.array))
    trap_if(self.array[i] is None)
    return self.array[i]

  def add(self, e):
    if self.free:
      i = self.free.pop()
      assert(self.array[i] is None)
      self.array[i] = e
    else:
      i = len(self.array)
      trap_if(i >= 2**30)
      self.array.append(e)
    return i

  def remove(self, i):
    e = self.get(i)
    self.array[i] = None
    self.free.append(i)
    return e

Table maintains a dense array of elements that can contain holes created by the remove method (defined below). When table elements are accessed (e.g., by canon_lift and resource.rep, below), there are thus both a bounds check and hole check necessary. Upon initialization, table element 0 is allocated and set to None, effectively reserving index 0 which is both useful for catching null/uninitialized accesses and allowing 0 to serve as a sentinel value.

The add and remove methods work together to maintain a free list of holes that are used in preference to growing the table. The free list is represented as a Python list here, but an optimizing implementation could instead store the free list in the free elements of array.

The limit of 2**30 ensures that the high 2 bits of table indices are unset and available for other use in guest code (e.g., for tagging, packed words or sentinel values).

The HandleElem class defines the elements of the per-resource-type Tables stored in HandleTables:

class HandleElem:
  rep: int
  own: bool
  scope: Optional[ExportCall]
  lend_count: int

  def __init__(self, rep, own, scope = None):
    self.rep = rep
    self.own = own
    self.scope = scope
    self.lend_count = 0

The rep field of HandleElem stores the resource representation (currently fixed to be an i32) passed to resource.new.

The own field indicates whether this element was created from an own type (or, if false, a borrow type).

The scope field stores the ExportCall that created the borrowed handle. Until async is added to the Component Model, because of the non-reentrancy of components, there is at most one ExportCall alive for a given component at a time and thus this field does not actually need to be stored per HandleElem.

The lend_count field maintains a conservative approximation of the number of live handles that were lent from this own handle (by calls to borrow-taking functions). This count is maintained by the ImportCall bookkeeping functions (above) and is ensured to be zero when an own handle is dropped.

An optimizing implementation can enumerate the canonical definitions present in a component to statically determine that a given resource type's handle table only contains own or borrow handles and then, based on this, statically eliminate the own and the lend_count xor scope fields, and guards thereof.

One CallContext is created for each call for both the component caller and callee (in canon_lower and canon_lift, resp., as defined below). Thus, a cross-component call will create 2 CallContext objects for the call, while a component-to-host or host-to-component call will create a single CallContext for the component caller or callee, resp.

Additional per-call state is required to check that callers and callees uphold their respective parts of the call contract:

The ExportCall subclass of CallContext tracks the number of borrowed handles that were passed as parameters to the export that have not yet been dropped (which might dangle if the caller destroys the resource after the call):

class ExportCall(CallContext):
  borrow_count: int

  def __init__(self, opts, inst):
    super().__init__(opts, inst)
    self.borrow_count = 0

  def create_borrow(self):
    self.borrow_count += 1

  def drop_borrow(self):
    assert(self.borrow_count > 0)
    self.borrow_count -= 1

  def exit(self):
    trap_if(self.borrow_count != 0)

The ImportCall subclass of CallContext tracks the owned handles that have been lent for the duration of an import call, ensuring that they aren't dropped during the call (which might create a dangling borrowed handle):

class ImportCall(CallContext):
  lenders: list[HandleElem]

  def __init__(self, opts, inst):
    super().__init__(opts, inst)
    self.lenders = []

  def track_owning_lend(self, lending_handle):
    assert(lending_handle.own)
    lending_handle.lend_count += 1
    self.lenders.append(lending_handle)

  def exit(self):
    for h in self.lenders:
      h.lend_count -= 1

Note, the lenders list usually has a fixed size (in all cases except when a function signature has borrows in lists) and thus can be stored inline in the native stack frame.

Loading

The load function defines how to read a value of a given value type t out of linear memory starting at offset ptr, returning the value represented as a Python value. Presenting the definition of load piecewise, we start with the top-level case analysis:

def load(cx, ptr, t):
  assert(ptr == align_to(ptr, alignment(t)))
  assert(ptr + elem_size(t) <= len(cx.opts.memory))
  match despecialize(t):
    case Bool()         : return convert_int_to_bool(load_int(cx, ptr, 1))
    case U8()           : return load_int(cx, ptr, 1)
    case U16()          : return load_int(cx, ptr, 2)
    case U32()          : return load_int(cx, ptr, 4)
    case U64()          : return load_int(cx, ptr, 8)
    case S8()           : return load_int(cx, ptr, 1, signed=True)
    case S16()          : return load_int(cx, ptr, 2, signed=True)
    case S32()          : return load_int(cx, ptr, 4, signed=True)
    case S64()          : return load_int(cx, ptr, 8, signed=True)
    case F32()          : return decode_i32_as_float(load_int(cx, ptr, 4))
    case F64()          : return decode_i64_as_float(load_int(cx, ptr, 8))
    case Char()         : return convert_i32_to_char(cx, load_int(cx, ptr, 4))
    case String()       : return load_string(cx, ptr)
    case List(t)        : return load_list(cx, ptr, t)
    case Record(fields) : return load_record(cx, ptr, fields)
    case Variant(cases) : return load_variant(cx, ptr, cases)
    case Flags(labels)  : return load_flags(cx, ptr, labels)
    case Own()          : return lift_own(cx, load_int(cx, ptr, 4), t)
    case Borrow()       : return lift_borrow(cx, load_int(cx, ptr, 4), t)

Integers are loaded directly from memory, with their high-order bit interpreted according to the signedness of the type.

def load_int(cx, ptr, nbytes, signed = False):
  return int.from_bytes(cx.opts.memory[ptr : ptr+nbytes], 'little', signed=signed)

Integer-to-boolean conversions treats 0 as false and all other bit-patterns as true:

def convert_int_to_bool(i):
  assert(i >= 0)
  return bool(i)

Floats are loaded directly from memory, with the sign and payload information of NaN values discarded. Consequently, there is only one unique NaN value per floating-point type. This reflects the practical reality that some languages and protocols do not preserve these bits. In the Python code below, this is expressed as canonicalizing NaNs to a particular bit pattern.

See the comments about lowering of float values for a discussion of possible optimizations.

DETERMINISTIC_PROFILE = False # or True
CANONICAL_FLOAT32_NAN = 0x7fc00000
CANONICAL_FLOAT64_NAN = 0x7ff8000000000000

def canonicalize_nan32(f):
  if math.isnan(f):
    f = core_f32_reinterpret_i32(CANONICAL_FLOAT32_NAN)
    assert(math.isnan(f))
  return f

def canonicalize_nan64(f):
  if math.isnan(f):
    f = core_f64_reinterpret_i64(CANONICAL_FLOAT64_NAN)
    assert(math.isnan(f))
  return f

def decode_i32_as_float(i):
  return canonicalize_nan32(core_f32_reinterpret_i32(i))

def decode_i64_as_float(i):
  return canonicalize_nan64(core_f64_reinterpret_i64(i))

def core_f32_reinterpret_i32(i):
  return struct.unpack('<f', struct.pack('<I', i))[0] # f32.reinterpret_i32

def core_f64_reinterpret_i64(i):
  return struct.unpack('<d', struct.pack('<Q', i))[0] # f64.reinterpret_i64

An i32 is converted to a char (a Unicode Scalar Value) by dynamically testing that its unsigned integral value is in the valid Unicode Code Point range and not a Surrogate:

def convert_i32_to_char(cx, i):
  assert(i >= 0)
  trap_if(i >= 0x110000)
  trap_if(0xD800 <= i <= 0xDFFF)
  return chr(i)

Strings are loaded from two i32 values: a pointer (offset in linear memory) and a number of bytes. There are three supported string encodings in canonopt: UTF-8, UTF-16 and latin1+utf16. This last options allows a dynamic choice between Latin-1 and UTF-16, indicated by the high bit of the second i32. String values include their original encoding and byte length as a "hint" that enables store_string (defined below) to make better up-front allocation size choices in many cases. Thus, the value produced by load_string isn't simply a Python str, but a tuple containing a str, the original encoding and the original byte length.

def load_string(cx, ptr):
  begin = load_int(cx, ptr, 4)
  tagged_code_units = load_int(cx, ptr + 4, 4)
  return load_string_from_range(cx, begin, tagged_code_units)

UTF16_TAG = 1 << 31

def load_string_from_range(cx, ptr, tagged_code_units):
  match cx.opts.string_encoding:
    case 'utf8':
      alignment = 1
      byte_length = tagged_code_units
      encoding = 'utf-8'
    case 'utf16':
      alignment = 2
      byte_length = 2 * tagged_code_units
      encoding = 'utf-16-le'
    case 'latin1+utf16':
      alignment = 2
      if bool(tagged_code_units & UTF16_TAG):
        byte_length = 2 * (tagged_code_units ^ UTF16_TAG)
        encoding = 'utf-16-le'
      else:
        byte_length = tagged_code_units
        encoding = 'latin-1'

  trap_if(ptr != align_to(ptr, alignment))
  trap_if(ptr + byte_length > len(cx.opts.memory))
  try:
    s = cx.opts.memory[ptr : ptr+byte_length].decode(encoding)
  except UnicodeError:
    trap()

  return (s, cx.opts.string_encoding, tagged_code_units)

Lists and records are loaded by recursively loading their elements/fields:

def load_list(cx, ptr, elem_type):
  begin = load_int(cx, ptr, 4)
  length = load_int(cx, ptr + 4, 4)
  return load_list_from_range(cx, begin, length, elem_type)

def load_list_from_range(cx, ptr, length, elem_type):
  trap_if(ptr != align_to(ptr, alignment(elem_type)))
  trap_if(ptr + length * elem_size(elem_type) > len(cx.opts.memory))
  a = []
  for i in range(length):
    a.append(load(cx, ptr + i * elem_size(elem_type), elem_type))
  return a

def load_record(cx, ptr, fields):
  record = {}
  for field in fields:
    ptr = align_to(ptr, alignment(field.t))
    record[field.label] = load(cx, ptr, field.t)
    ptr += elem_size(field.t)
  return record

As a technical detail: the align_to in the loop in load_record is guaranteed to be a no-op on the first iteration because the record as a whole starts out aligned (as asserted at the top of load).

Variants are loaded using the order of the cases in the type to determine the case index, assigning 0 to the first case, 1 to the next case, etc. To support the subtyping allowed by refines, a lifted variant value semantically includes a full ordered list of its refines case labels so that the lowering code (defined below) can search this list to find a case label it knows about. While the code below appears to perform case-label lookup at runtime, a normal implementation can build the appropriate index tables at compile-time so that variant-passing is always O(1) and not involving string operations.

def load_variant(cx, ptr, cases):
  disc_size = elem_size(discriminant_type(cases))
  case_index = load_int(cx, ptr, disc_size)
  ptr += disc_size
  trap_if(case_index >= len(cases))
  c = cases[case_index]
  ptr = align_to(ptr, max_case_alignment(cases))
  case_label = case_label_with_refinements(c, cases)
  if c.t is None:
    return { case_label: None }
  return { case_label: load(cx, ptr, c.t) }

def case_label_with_refinements(c, cases):
  label = c.label
  while c.refines is not None:
    c = cases[find_case(c.refines, cases)]
    label += '|' + c.label
  return label

def find_case(label, cases):
  matches = [i for i,c in enumerate(cases) if c.label == label]
  assert(len(matches) <= 1)
  if len(matches) == 1:
    return matches[0]
  return -1

Flags are converted from a bit-vector to a dictionary whose keys are derived from the ordered labels of the flags type. The code here takes advantage of Python's support for integers of arbitrary width.

def load_flags(cx, ptr, labels):
  i = load_int(cx, ptr, elem_size_flags(labels))
  return unpack_flags_from_int(i, labels)

def unpack_flags_from_int(i, labels):
  record = {}
  for l in labels:
    record[l] = bool(i & 1)
    i >>= 1
  return record

own handles are lifted by removing the handle from the current component instance's handle table, so that ownership is transferred to the lowering component. The lifting operation fails if unique ownership of the handle isn't possible, for example if the index was actually a borrow or if the own handle is currently being lent out as borrows.

def lift_own(cx, i, t):
  h = cx.inst.handles.remove(t.rt, i)
  trap_if(h.lend_count != 0)
  trap_if(not h.own)
  return h.rep

The abstract lifted value for handle types is currently just the internal resource representation i32, which is kept opaque from the receiving component (it's stored in the handle table and only accessed indirectly via index). (This assumes that resource representations are immutable. If representations were to become mutable, the address of the mutable cell would be passed as the lifted value instead.)

In contrast to own, borrow handles are lifted by reading the representation from the source handle, leaving the source handle intact in the current component instance's handle table:

def lift_borrow(cx, i, t):
  assert(isinstance(cx, ImportCall))
  h = cx.inst.handles.get(t.rt, i)
  if h.own:
    cx.track_owning_lend(h)
  return h.rep

The track_owning_lend call to CallContext participates in the enforcement of the dynamic borrow rules, which keep the source own handle alive until the end of the call (as an intentionally-conservative upper bound on how long the borrow handle can be held). This tracking is only required when h is an own handle because, when h is a borrow handle, this tracking has already happened (when the originating own handle was lifted) for a strictly longer call scope than the current call.

Storing

The store function defines how to write a value v of a given value type t into linear memory starting at offset ptr. Presenting the definition of store piecewise, we start with the top-level case analysis:

def store(cx, v, t, ptr):
  assert(ptr == align_to(ptr, alignment(t)))
  assert(ptr + elem_size(t) <= len(cx.opts.memory))
  match despecialize(t):
    case Bool()         : store_int(cx, int(bool(v)), ptr, 1)
    case U8()           : store_int(cx, v, ptr, 1)
    case U16()          : store_int(cx, v, ptr, 2)
    case U32()          : store_int(cx, v, ptr, 4)
    case U64()          : store_int(cx, v, ptr, 8)
    case S8()           : store_int(cx, v, ptr, 1, signed=True)
    case S16()          : store_int(cx, v, ptr, 2, signed=True)
    case S32()          : store_int(cx, v, ptr, 4, signed=True)
    case S64()          : store_int(cx, v, ptr, 8, signed=True)
    case F32()          : store_int(cx, encode_float_as_i32(v), ptr, 4)
    case F64()          : store_int(cx, encode_float_as_i64(v), ptr, 8)
    case Char()         : store_int(cx, char_to_i32(v), ptr, 4)
    case String()       : store_string(cx, v, ptr)
    case List(t)        : store_list(cx, v, ptr, t)
    case Record(fields) : store_record(cx, v, ptr, fields)
    case Variant(cases) : store_variant(cx, v, ptr, cases)
    case Flags(labels)  : store_flags(cx, v, ptr, labels)
    case Own()          : store_int(cx, lower_own(cx.opts, v, t), ptr, 4)
    case Borrow()       : store_int(cx, lower_borrow(cx.opts, v, t), ptr, 4)

Integers are stored directly into memory. Because the input domain is exactly the integers in range for the given type, no extra range checks are necessary; the signed parameter is only present to ensure that the internal range checks of int.to_bytes are satisfied.

def store_int(cx, v, ptr, nbytes, signed = False):
  cx.opts.memory[ptr : ptr+nbytes] = int.to_bytes(v, nbytes, 'little', signed=signed)

Floats are stored directly into memory, with the sign and payload bits of NaN values modified non-deterministically. This reflects the practical reality that different languages, protocols and CPUs have different effects on NaNs.

Although this non-determinism is expressed in the Python code below as generating a "random" NaN bit-pattern, native implementations do not need to use the same "random" algorithm, or even any random algorithm at all. Hosts may instead chose to canonicalize to an arbitrary fixed NaN value, or even to the original value of the NaN before lifting, allowing them to optimize away both the canonicalization of lifting and the randomization of lowering.

When a host implements the deterministic profile, NaNs are canonicalized to a particular NaN bit-pattern.

def maybe_scramble_nan32(f):
  if math.isnan(f):
    if DETERMINISTIC_PROFILE:
      f = core_f32_reinterpret_i32(CANONICAL_FLOAT32_NAN)
    else:
      f = core_f32_reinterpret_i32(random_nan_bits(32, 8))
    assert(math.isnan(f))
  return f

def maybe_scramble_nan64(f):
  if math.isnan(f):
    if DETERMINISTIC_PROFILE:
      f = core_f64_reinterpret_i64(CANONICAL_FLOAT64_NAN)
    else:
      f = core_f64_reinterpret_i64(random_nan_bits(64, 11))
    assert(math.isnan(f))
  return f

def random_nan_bits(total_bits, exponent_bits):
  fraction_bits = total_bits - exponent_bits - 1
  bits = random.getrandbits(total_bits)
  bits |= ((1 << exponent_bits) - 1) << fraction_bits
  bits |= 1 << random.randrange(fraction_bits - 1)
  return bits

def encode_float_as_i32(f):
  return core_i32_reinterpret_f32(maybe_scramble_nan32(f))

def encode_float_as_i64(f):
  return core_i64_reinterpret_f64(maybe_scramble_nan64(f))

def core_i32_reinterpret_f32(f):
  return struct.unpack('<I', struct.pack('<f', f))[0] # i32.reinterpret_f32

def core_i64_reinterpret_f64(f):
  return struct.unpack('<Q', struct.pack('<d', f))[0] # i64.reinterpret_f64

The integral value of a char (a Unicode Scalar Value) is a valid unsigned i32 and thus no runtime conversion or checking is necessary:

def char_to_i32(c):
  i = ord(c)
  assert(0 <= i <= 0xD7FF or 0xD800 <= i <= 0x10FFFF)
  return i

Storing strings is complicated by the goal of attempting to optimize the different transcoding cases. In particular, one challenge is choosing the linear memory allocation size before examining the contents of the string. The reason for this constraint is that, in some settings where single-pass iterators are involved (host calls and post-MVP adapter functions), examining the contents of a string more than once would require making an engine-internal temporary copy of the whole string, which the component model specifically aims not to do. To avoid multiple passes, the canonical ABI instead uses a realloc approach to update the allocation size during the single copy. A blind realloc approach would normally suffer from multiple reallocations per string (e.g., using the standard doubling-growth strategy). However, as already shown in load_string above, string values come with two useful hints: their original encoding and byte length. From this hint data, store_string can do a much better job minimizing the number of reallocations.

We start with a case analysis to enumerate all the meaningful encoding combinations, subdividing the latin1+utf16 encoding into either latin1 or utf16 based on the UTF16_BIT flag set by load_string:

def store_string(cx, v, ptr):
  begin, tagged_code_units = store_string_into_range(cx, v)
  store_int(cx, begin, ptr, 4)
  store_int(cx, tagged_code_units, ptr + 4, 4)

def store_string_into_range(cx, v):
  src, src_encoding, src_tagged_code_units = v

  if src_encoding == 'latin1+utf16':
    if bool(src_tagged_code_units & UTF16_TAG):
      src_simple_encoding = 'utf16'
      src_code_units = src_tagged_code_units ^ UTF16_TAG
    else:
      src_simple_encoding = 'latin1'
      src_code_units = src_tagged_code_units
  else:
    src_simple_encoding = src_encoding
    src_code_units = src_tagged_code_units

  match cx.opts.string_encoding:
    case 'utf8':
      match src_simple_encoding:
        case 'utf8'         : return store_string_copy(cx, src, src_code_units, 1, 1, 'utf-8')
        case 'utf16'        : return store_utf16_to_utf8(cx, src, src_code_units)
        case 'latin1'       : return store_latin1_to_utf8(cx, src, src_code_units)
    case 'utf16':
      match src_simple_encoding:
        case 'utf8'         : return store_utf8_to_utf16(cx, src, src_code_units)
        case 'utf16'        : return store_string_copy(cx, src, src_code_units, 2, 2, 'utf-16-le')
        case 'latin1'       : return store_string_copy(cx, src, src_code_units, 2, 2, 'utf-16-le')
    case 'latin1+utf16':
      match src_encoding:
        case 'utf8'         : return store_string_to_latin1_or_utf16(cx, src, src_code_units)
        case 'utf16'        : return store_string_to_latin1_or_utf16(cx, src, src_code_units)
        case 'latin1+utf16' :
          match src_simple_encoding:
            case 'latin1'   : return store_string_copy(cx, src, src_code_units, 1, 2, 'latin-1')
            case 'utf16'    : return store_probably_utf16_to_latin1_or_utf16(cx, src, src_code_units)

The simplest 4 cases above can compute the exact destination size and then copy with a simply loop (that possibly inflates Latin-1 to UTF-16 by injecting a 0 byte after every Latin-1 byte).

MAX_STRING_BYTE_LENGTH = (1 << 31) - 1

def store_string_copy(cx, src, src_code_units, dst_code_unit_size, dst_alignment, dst_encoding):
  dst_byte_length = dst_code_unit_size * src_code_units
  trap_if(dst_byte_length > MAX_STRING_BYTE_LENGTH)
  ptr = cx.opts.realloc(0, 0, dst_alignment, dst_byte_length)
  trap_if(ptr != align_to(ptr, dst_alignment))
  trap_if(ptr + dst_byte_length > len(cx.opts.memory))
  encoded = src.encode(dst_encoding)
  assert(dst_byte_length == len(encoded))
  cx.opts.memory[ptr : ptr+len(encoded)] = encoded
  return (ptr, src_code_units)

The choice of MAX_STRING_BYTE_LENGTH constant ensures that the high bit of a string's byte length is never set, keeping it clear for UTF16_BIT.

The 2 cases of transcoding into UTF-8 share an algorithm that starts by optimistically assuming that each code unit of the source string fits in a single UTF-8 byte and then, failing that, reallocates to a worst-case size, finishes the copy, and then finishes with a shrinking reallocation.

def store_utf16_to_utf8(cx, src, src_code_units):
  worst_case_size = src_code_units * 3
  return store_string_to_utf8(cx, src, src_code_units, worst_case_size)

def store_latin1_to_utf8(cx, src, src_code_units):
  worst_case_size = src_code_units * 2
  return store_string_to_utf8(cx, src, src_code_units, worst_case_size)

def store_string_to_utf8(cx, src, src_code_units, worst_case_size):
  assert(src_code_units <= MAX_STRING_BYTE_LENGTH)
  ptr = cx.opts.realloc(0, 0, 1, src_code_units)
  trap_if(ptr + src_code_units > len(cx.opts.memory))
  for i,code_point in enumerate(src):
    if ord(code_point) < 2**7:
      cx.opts.memory[ptr + i] = ord(code_point)
    else:
      trap_if(worst_case_size > MAX_STRING_BYTE_LENGTH)
      ptr = cx.opts.realloc(ptr, src_code_units, 1, worst_case_size)
      trap_if(ptr + worst_case_size > len(cx.opts.memory))
      encoded = src.encode('utf-8')
      cx.opts.memory[ptr+i : ptr+len(encoded)] = encoded[i : ]
      if worst_case_size > len(encoded):
        ptr = cx.opts.realloc(ptr, worst_case_size, 1, len(encoded))
        trap_if(ptr + len(encoded) > len(cx.opts.memory))
      return (ptr, len(encoded))
  return (ptr, src_code_units)

Converting from UTF-8 to UTF-16 performs an initial worst-case size allocation (assuming each UTF-8 byte encodes a whole code point that inflates into a two-byte UTF-16 code unit) and then does a shrinking reallocation at the end if multiple UTF-8 bytes were collapsed into a single 2-byte UTF-16 code unit:

def store_utf8_to_utf16(cx, src, src_code_units):
  worst_case_size = 2 * src_code_units
  trap_if(worst_case_size > MAX_STRING_BYTE_LENGTH)
  ptr = cx.opts.realloc(0, 0, 2, worst_case_size)
  trap_if(ptr != align_to(ptr, 2))
  trap_if(ptr + worst_case_size > len(cx.opts.memory))
  encoded = src.encode('utf-16-le')
  cx.opts.memory[ptr : ptr+len(encoded)] = encoded
  if len(encoded) < worst_case_size:
    ptr = cx.opts.realloc(ptr, worst_case_size, 2, len(encoded))
    trap_if(ptr != align_to(ptr, 2))
    trap_if(ptr + len(encoded) > len(cx.opts.memory))
  code_units = int(len(encoded) / 2)
  return (ptr, code_units)

The next transcoding case handles latin1+utf16 encoding, where there general goal is to fit the incoming string into Latin-1 if possible based on the code points of the incoming string. The algorithm speculates that all code points do fit into Latin-1 and then falls back to a worst-case allocation size when a code point is found outside Latin-1. In this fallback case, the previously-copied Latin-1 bytes are inflated in place, inserting a 0 byte after every Latin-1 byte (iterating in reverse to avoid clobbering later bytes):

def store_string_to_latin1_or_utf16(cx, src, src_code_units):
  assert(src_code_units <= MAX_STRING_BYTE_LENGTH)
  ptr = cx.opts.realloc(0, 0, 2, src_code_units)
  trap_if(ptr != align_to(ptr, 2))
  trap_if(ptr + src_code_units > len(cx.opts.memory))
  dst_byte_length = 0
  for usv in src:
    if ord(usv) < (1 << 8):
      cx.opts.memory[ptr + dst_byte_length] = ord(usv)
      dst_byte_length += 1
    else:
      worst_case_size = 2 * src_code_units
      trap_if(worst_case_size > MAX_STRING_BYTE_LENGTH)
      ptr = cx.opts.realloc(ptr, src_code_units, 2, worst_case_size)
      trap_if(ptr != align_to(ptr, 2))
      trap_if(ptr + worst_case_size > len(cx.opts.memory))
      for j in range(dst_byte_length-1, -1, -1):
        cx.opts.memory[ptr + 2*j] = cx.opts.memory[ptr + j]
        cx.opts.memory[ptr + 2*j + 1] = 0
      encoded = src.encode('utf-16-le')
      cx.opts.memory[ptr+2*dst_byte_length : ptr+len(encoded)] = encoded[2*dst_byte_length : ]
      if worst_case_size > len(encoded):
        ptr = cx.opts.realloc(ptr, worst_case_size, 2, len(encoded))
        trap_if(ptr != align_to(ptr, 2))
        trap_if(ptr + len(encoded) > len(cx.opts.memory))
      tagged_code_units = int(len(encoded) / 2) | UTF16_TAG
      return (ptr, tagged_code_units)
  if dst_byte_length < src_code_units:
    ptr = cx.opts.realloc(ptr, src_code_units, 2, dst_byte_length)
    trap_if(ptr != align_to(ptr, 2))
    trap_if(ptr + dst_byte_length > len(cx.opts.memory))
  return (ptr, dst_byte_length)

The final transcoding case takes advantage of the extra heuristic information that the incoming UTF-16 bytes were intentionally chosen over Latin-1 by the producer, indicating that they probably contain code points outside Latin-1 and thus probably require inflation. Based on this information, the transcoding algorithm pessimistically allocates storage for UTF-16, deflating at the end if indeed no non-Latin-1 code points were encountered. This Latin-1 deflation ensures that if a group of components are all using latin1+utf16 and one component over-uses UTF-16, other components can recover the Latin-1 compression. (The Latin-1 check can be inexpensively fused with the UTF-16 validate+copy loop.)

def store_probably_utf16_to_latin1_or_utf16(cx, src, src_code_units):
  src_byte_length = 2 * src_code_units
  trap_if(src_byte_length > MAX_STRING_BYTE_LENGTH)
  ptr = cx.opts.realloc(0, 0, 2, src_byte_length)
  trap_if(ptr != align_to(ptr, 2))
  trap_if(ptr + src_byte_length > len(cx.opts.memory))
  encoded = src.encode('utf-16-le')
  cx.opts.memory[ptr : ptr+len(encoded)] = encoded
  if any(ord(c) >= (1 << 8) for c in src):
    tagged_code_units = int(len(encoded) / 2) | UTF16_TAG
    return (ptr, tagged_code_units)
  latin1_size = int(len(encoded) / 2)
  for i in range(latin1_size):
    cx.opts.memory[ptr + i] = cx.opts.memory[ptr + 2*i]
  ptr = cx.opts.realloc(ptr, src_byte_length, 1, latin1_size)
  trap_if(ptr + latin1_size > len(cx.opts.memory))
  return (ptr, latin1_size)

Lists and records are stored by recursively storing their elements and are symmetric to the loading functions. Unlike strings, lists can simply allocate based on the up-front knowledge of length and static element size.

def store_list(cx, v, ptr, elem_type):
  begin, length = store_list_into_range(cx, v, elem_type)
  store_int(cx, begin, ptr, 4)
  store_int(cx, length, ptr + 4, 4)

def store_list_into_range(cx, v, elem_type):
  byte_length = len(v) * elem_size(elem_type)
  trap_if(byte_length >= (1 << 32))
  ptr = cx.opts.realloc(0, 0, alignment(elem_type), byte_length)
  trap_if(ptr != align_to(ptr, alignment(elem_type)))
  trap_if(ptr + byte_length > len(cx.opts.memory))
  for i,e in enumerate(v):
    store(cx, e, elem_type, ptr + i * elem_size(elem_type))
  return (ptr, len(v))

def store_record(cx, v, ptr, fields):
  for f in fields:
    ptr = align_to(ptr, alignment(f.t))
    store(cx, v[f.label], f.t, ptr)
    ptr += elem_size(f.t)

Variants are stored using the |-separated list of refines cases built by case_label_with_refinements (above) to iteratively find a matching case (which validation guarantees will succeed). While this code appears to do O(n) string matching, a normal implementation can statically fuse store_variant with its matching load_variant to ultimately build a dense array that maps producer's case indices to the consumer's case indices.

def store_variant(cx, v, ptr, cases):
  case_index, case_value = match_case(v, cases)
  disc_size = elem_size(discriminant_type(cases))
  store_int(cx, case_index, ptr, disc_size)
  ptr += disc_size
  ptr = align_to(ptr, max_case_alignment(cases))
  c = cases[case_index]
  if c.t is not None:
    store(cx, case_value, c.t, ptr)

def match_case(v, cases):
  assert(len(v.keys()) == 1)
  key = list(v.keys())[0]
  value = list(v.values())[0]
  for label in key.split('|'):
    case_index = find_case(label, cases)
    if case_index != -1:
      return (case_index, value)

Flags are converted from a dictionary to a bit-vector by iterating through the case-labels of the variant in the order they were listed in the type definition and OR-ing all the bits together. Flag lifting/lowering can be statically fused into array/integer operations (with a simple byte copy when the case lists are the same) to avoid any string operations in a similar manner to variants.

def store_flags(cx, v, ptr, labels):
  i = pack_flags_into_int(v, labels)
  store_int(cx, i, ptr, elem_size_flags(labels))

def pack_flags_into_int(v, labels):
  i = 0
  shift = 0
  for l in labels:
    i |= (int(bool(v[l])) << shift)
    shift += 1
  return i

Finally, own and borrow handles are lowered by initializing new handle elements in the current component instance's handle table:

def lower_own(cx, rep, t):
  h = HandleElem(rep, own=True)
  return cx.inst.handles.add(t.rt, h)

def lower_borrow(cx, rep, t):
  assert(isinstance(cx, ExportCall))
  if cx.inst is t.rt.impl:
    return rep
  h = HandleElem(rep, own=False, scope=cx)
  cx.create_borrow()
  return cx.inst.handles.add(t.rt, h)

The special case in lower_borrow is an optimization, recognizing that, when a borrowed handle is passed to the component that implemented the resource type, the only thing the borrowed handle is good for is calling resource.rep, so lowering might as well avoid the overhead of creating an intermediate borrow handle.

Flattening

With only the definitions above, the Canonical ABI would be forced to place all parameters and results in linear memory. While this is necessary in the general case, in many cases performance can be improved by passing small-enough values in registers by using core function parameters and results. To support this optimization, the Canonical ABI defines flatten to map component function types to core function types by attempting to decompose all the non-dynamically-sized component value types into core value types.

For a variety of practical reasons, we need to limit the total number of flattened parameters and results, falling back to storing everything in linear memory. The number of flattened results is currently limited to 1 due to various parts of the toolchain (notably the C ABI) not yet being able to express multi-value returns. Hopefully this limitation is temporary and can be lifted before the Component Model is fully standardized.

When there are too many flat values, in general, a single i32 pointer can be passed instead (pointing to a tuple in linear memory). When lowering into linear memory, this requires the Canonical ABI to call realloc (in lower below) to allocate space to put the tuple. As an optimization, when lowering the return value of an imported function (via canon lower), the caller can have already allocated space for the return value (e.g., efficiently on the stack), passing in an i32 pointer as an parameter instead of returning an i32 as a return value.

Given all this, the top-level definition of flatten is:

MAX_FLAT_PARAMS = 16
MAX_FLAT_RESULTS = 1

def flatten_functype(ft, context):
  flat_params = flatten_types(ft.param_types())
  if len(flat_params) > MAX_FLAT_PARAMS:
    flat_params = ['i32']

  flat_results = flatten_types(ft.result_types())
  if len(flat_results) > MAX_FLAT_RESULTS:
    match context:
      case 'lift':
        flat_results = ['i32']
      case 'lower':
        flat_params += ['i32']
        flat_results = []

  return CoreFuncType(flat_params, flat_results)

def flatten_types(ts):
  return [ft for t in ts for ft in flatten_type(t)]

Presenting the definition of flatten_type piecewise, we start with the top-level case analysis:

def flatten_type(t):
  match despecialize(t):
    case Bool()               : return ['i32']
    case U8() | U16() | U32() : return ['i32']
    case S8() | S16() | S32() : return ['i32']
    case S64() | U64()        : return ['i64']
    case F32()                : return ['f32']
    case F64()                : return ['f64']
    case Char()               : return ['i32']
    case String() | List(_)   : return ['i32', 'i32']
    case Record(fields)       : return flatten_record(fields)
    case Variant(cases)       : return flatten_variant(cases)
    case Flags(labels)        : return ['i32'] * num_i32_flags(labels)
    case Own(_) | Borrow(_)   : return ['i32']

Record flattening simply flattens each field in sequence.

def flatten_record(fields):
  flat = []
  for f in fields:
    flat += flatten_type(f.t)
  return flat

Variant flattening is more involved due to the fact that each case payload can have a totally different flattening. Rather than giving up when there is a type mismatch, the Canonical ABI relies on the fact that the 4 core value types can be easily bit-cast between each other and defines a join operator to pick the tightest approximation. What this means is that, regardless of the dynamic case, all flattened variants are passed with the same static set of core types, which may involve, e.g., reinterpreting an f32 as an i32 or zero-extending an i32 into an i64.

def flatten_variant(cases):
  flat = []
  for c in cases:
    if c.t is not None:
      for i,ft in enumerate(flatten_type(c.t)):
        if i < len(flat):
          flat[i] = join(flat[i], ft)
        else:
          flat.append(ft)
  return flatten_type(discriminant_type(cases)) + flat

def join(a, b):
  if a == b: return a
  if (a == 'i32' and b == 'f32') or (a == 'f32' and b == 'i32'): return 'i32'
  return 'i64'

Flat Lifting

Values are lifted by iterating over a list of parameter or result Core WebAssembly values:

@dataclass
class CoreValueIter:
  values: list[int|float]
  i = 0
  def next(self, t):
    v = self.values[self.i]
    self.i += 1
    match t:
      case 'i32': assert(isinstance(v, int) and 0 <= v < 2**32)
      case 'i64': assert(isinstance(v, int) and 0 <= v < 2**64)
      case 'f32': assert(isinstance(v, (int,float)))
      case 'f64': assert(isinstance(v, (int,float)))
      case _    : assert(False)
    return v

The match is only used for spec-level assertions; no runtime typecase is required.

The lift_flat function defines how to convert a list of core values into a single high-level value of type t. Presenting the definition of lift_flat piecewise, we start with the top-level case analysis:

def lift_flat(cx, vi, t):
  match despecialize(t):
    case Bool()         : return convert_int_to_bool(vi.next('i32'))
    case U8()           : return lift_flat_unsigned(vi, 32, 8)
    case U16()          : return lift_flat_unsigned(vi, 32, 16)
    case U32()          : return lift_flat_unsigned(vi, 32, 32)
    case U64()          : return lift_flat_unsigned(vi, 64, 64)
    case S8()           : return lift_flat_signed(vi, 32, 8)
    case S16()          : return lift_flat_signed(vi, 32, 16)
    case S32()          : return lift_flat_signed(vi, 32, 32)
    case S64()          : return lift_flat_signed(vi, 64, 64)
    case F32()          : return canonicalize_nan32(vi.next('f32'))
    case F64()          : return canonicalize_nan64(vi.next('f64'))
    case Char()         : return convert_i32_to_char(cx, vi.next('i32'))
    case String()       : return lift_flat_string(cx, vi)
    case List(t)        : return lift_flat_list(cx, vi, t)
    case Record(fields) : return lift_flat_record(cx, vi, fields)
    case Variant(cases) : return lift_flat_variant(cx, vi, cases)
    case Flags(labels)  : return lift_flat_flags(vi, labels)
    case Own()          : return lift_own(cx, vi.next('i32'), t)
    case Borrow()       : return lift_borrow(cx, vi.next('i32'), t)

Integers are lifted from core i32 or i64 values using the signedness of the target type to interpret the high-order bit. When the target type is narrower than an i32, the Canonical ABI ignores the unused high bits (like load_int). The conversion logic here assumes that i32 values are always represented as unsigned Python ints and thus lifting to a signed type performs a manual 2s complement conversion in the Python (which would be a no-op in hardware).

def lift_flat_unsigned(vi, core_width, t_width):
  i = vi.next('i' + str(core_width))
  assert(0 <= i < (1 << core_width))
  return i % (1 << t_width)

def lift_flat_signed(vi, core_width, t_width):
  i = vi.next('i' + str(core_width))
  assert(0 <= i < (1 << core_width))
  i %= (1 << t_width)
  if i >= (1 << (t_width - 1)):
    return i - (1 << t_width)
  return i

The contents of strings and lists are always stored in memory so lifting these types is essentially the same as loading them from memory; the only difference is that the pointer and length come from i32 values instead of from linear memory:

def lift_flat_string(cx, vi):
  ptr = vi.next('i32')
  packed_length = vi.next('i32')
  return load_string_from_range(cx, ptr, packed_length)

def lift_flat_list(cx, vi, elem_type):
  ptr = vi.next('i32')
  length = vi.next('i32')
  return load_list_from_range(cx, ptr, length, elem_type)

Records are lifted by recursively lifting their fields:

def lift_flat_record(cx, vi, fields):
  record = {}
  for f in fields:
    record[f.label] = lift_flat(cx, vi, f.t)
  return record

Variants are also lifted recursively. Lifting a variant must carefully follow the definition of flatten_variant above, consuming the exact same core types regardless of the dynamic case payload being lifted. Because of the join performed by flatten_variant, we need a more-permissive value iterator that reinterprets between the different types appropriately and also traps if the high bits of an i64 are set for a 32-bit type:

def lift_flat_variant(cx, vi, cases):
  flat_types = flatten_variant(cases)
  assert(flat_types.pop(0) == 'i32')
  case_index = vi.next('i32')
  trap_if(case_index >= len(cases))
  class CoerceValueIter:
    def next(self, want):
      have = flat_types.pop(0)
      x = vi.next(have)
      match (have, want):
        case ('i32', 'f32') : return decode_i32_as_float(x)
        case ('i64', 'i32') : return wrap_i64_to_i32(x)
        case ('i64', 'f32') : return decode_i32_as_float(wrap_i64_to_i32(x))
        case ('i64', 'f64') : return decode_i64_as_float(x)
        case _              : assert(have == want); return x
  c = cases[case_index]
  if c.t is None:
    v = None
  else:
    v = lift_flat(cx, CoerceValueIter(), c.t)
  for have in flat_types:
    _ = vi.next(have)
  return { case_label_with_refinements(c, cases): v }

def wrap_i64_to_i32(i):
  assert(0 <= i < (1 << 64))
  return i % (1 << 32)

Finally, flags are lifted by OR-ing together all the flattened i32 values and then lifting to a record the same way as when loading flags from linear memory.

def lift_flat_flags(vi, labels):
  i = 0
  shift = 0
  for _ in range(num_i32_flags(labels)):
    i |= (vi.next('i32') << shift)
    shift += 32
  return unpack_flags_from_int(i, labels)

Flat Lowering

The lower_flat function defines how to convert a value v of a given type t into zero or more core values. Presenting the definition of lower_flat piecewise, we start with the top-level case analysis:

def lower_flat(cx, v, t):
  match despecialize(t):
    case Bool()         : return [int(v)]
    case U8()           : return [v]
    case U16()          : return [v]
    case U32()          : return [v]
    case U64()          : return [v]
    case S8()           : return lower_flat_signed(v, 32)
    case S16()          : return lower_flat_signed(v, 32)
    case S32()          : return lower_flat_signed(v, 32)
    case S64()          : return lower_flat_signed(v, 64)
    case F32()          : return [maybe_scramble_nan32(v)]
    case F64()          : return [maybe_scramble_nan64(v)]
    case Char()         : return [char_to_i32(v)]
    case String()       : return lower_flat_string(cx, v)
    case List(t)        : return lower_flat_list(cx, v, t)
    case Record(fields) : return lower_flat_record(cx, v, fields)
    case Variant(cases) : return lower_flat_variant(cx, v, cases)
    case Flags(labels)  : return lower_flat_flags(v, labels)
    case Own()          : return [lower_own(cx, v, t)]
    case Borrow()       : return [lower_borrow(cx, v, t)]

Since component-level values are assumed in-range and, as previously stated, core i32 values are always internally represented as unsigned ints, unsigned integer values need no extra conversion. Signed integer values are converted to unsigned core i32s by 2s complement arithmetic (which again would be a no-op in hardware):

def lower_flat_signed(i, core_bits):
  if i < 0:
    i += (1 << core_bits)
  return [i]

Since strings and lists are stored in linear memory, lifting can reuse the previous definitions; only the resulting pointers are returned differently (as i32 values instead of as a pair in linear memory):

def lower_flat_string(cx, v):
  ptr, packed_length = store_string_into_range(cx, v)
  return [ptr, packed_length]

def lower_flat_list(cx, v, elem_type):
  (ptr, length) = store_list_into_range(cx, v, elem_type)
  return [ptr, length]

Records are lowered by recursively lowering their fields:

def lower_flat_record(cx, v, fields):
  flat = []
  for f in fields:
    flat += lower_flat(cx, v[f.label], f.t)
  return flat

Variants are also lowered recursively. Symmetric to lift_flat_variant above, lower_flat_variant must consume all flattened types of flatten_variant, manually coercing the otherwise-incompatible type pairings allowed by join:

def lower_flat_variant(cx, v, cases):
  case_index, case_value = match_case(v, cases)
  flat_types = flatten_variant(cases)
  assert(flat_types.pop(0) == 'i32')
  c = cases[case_index]
  if c.t is None:
    payload = []
  else:
    payload = lower_flat(cx, case_value, c.t)
    for i,(fv,have) in enumerate(zip(payload, flatten_type(c.t))):
      want = flat_types.pop(0)
      match (have, want):
        case ('f32', 'i32') : payload[i] = encode_float_as_i32(fv)
        case ('i32', 'i64') : payload[i] = fv
        case ('f32', 'i64') : payload[i] = encode_float_as_i32(fv)
        case ('f64', 'i64') : payload[i] = encode_float_as_i64(fv)
        case _              : assert(have == want)
  for _ in flat_types:
    payload.append(0)
  return [case_index] + payload

Finally, flags are lowered by slicing the bit vector into i32 chunks:

def lower_flat_flags(v, labels):
  i = pack_flags_into_int(v, labels)
  flat = []
  for _ in range(num_i32_flags(labels)):
    flat.append(i & 0xffffffff)
    i >>= 32
  assert(i == 0)
  return flat

Lifting and Lowering Values

The lift_values function defines how to lift a list of at most max_flat core parameters or results given by the CoreValueIter vi into a tuple of values with types ts:

def lift_values(cx, max_flat, vi, ts):
  flat_types = flatten_types(ts)
  if len(flat_types) > max_flat:
    ptr = vi.next('i32')
    tuple_type = Tuple(ts)
    trap_if(ptr != align_to(ptr, alignment(tuple_type)))
    trap_if(ptr + elem_size(tuple_type) > len(cx.opts.memory))
    return list(load(cx, ptr, tuple_type).values())
  else:
    return [ lift_flat(cx, vi, t) for t in ts ]

The lower_values function defines how to lower a list of component-level values vs of types ts into a list of at most max_flat core values. As already described for flatten above, lowering handles the greater-than-max_flat case by either allocating storage with realloc or accepting a caller-allocated buffer as an out-param:

def lower_values(cx, max_flat, vs, ts, out_param = None):
  inst = cx.inst
  assert(inst.may_leave)
  inst.may_leave = False

  flat_types = flatten_types(ts)
  if len(flat_types) > max_flat:
    tuple_type = Tuple(ts)
    tuple_value = {str(i): v for i,v in enumerate(vs)}
    if out_param is None:
      ptr = cx.opts.realloc(0, 0, alignment(tuple_type), elem_size(tuple_type))
    else:
      ptr = out_param.next('i32')
    trap_if(ptr != align_to(ptr, alignment(tuple_type)))
    trap_if(ptr + elem_size(tuple_type) > len(cx.opts.memory))
    store(cx, tuple_value, tuple_type, ptr)
    flat_vales = [ptr]
  else:
    flat_vals = []
    for i in range(len(vs)):
      flat_vals += lower_flat(cx, vs[i], ts[i])

  inst.may_leave = True
  return flat_vals

The may_leave flag is used by canon_lower below to prevent a component from calling out of the component while in the middle of lowering, ensuring that the relative ordering of the side effects of lift_values and lower_values cannot be observed and thus an implementation may reliably fuse lift_values with lower_values when making a cross-component call, avoiding any intermediate copy.

Canonical Definitions

Using the above supporting definitions, we can describe the static and dynamic semantics of component-level canon definitions. The following subsections cover each of these canon cases.

canon lift

For a canonical definition:

(canon lift $callee:<funcidx> $opts:<canonopt>* (func $f (type $ft)))

validation specifies:

  • $callee must have type flatten_functype($ft, 'lift')
  • $f is given type $ft
  • a memory is present if required by lifting and is a subtype of (memory 1)
  • a realloc is present if required by lifting and has type (func (param i32 i32 i32 i32) (result i32))
  • if a post-return is present, it has type (func (param flatten_functype($ft).results))

When instantiating component instance $inst:

  • Define $f to be the partially-bound closure canon_lift($opts, $inst, $callee, $ft)

Thus, $f captures $opts, $inst, $callee and $ft in a closure which can be subsequently exported or passed into a child instance (via with). If $f ends up being called by the host, the host is responsible for, in a host-defined manner, conjuring up component values suitable for passing into lower and, conversely, consuming the component values produced by lift. For example, if the host is a native JS runtime, the JavaScript embedding would specify how native JavaScript values are converted to and from component values. Alternatively, if the host is a Unix CLI that invokes component exports directly from the command line, the CLI could choose to automatically parse argv into component-level values according to the declared types of the export. In any case, canon lift specifies how these variously-produced values are consumed as parameters (and produced as results) by a single host-agnostic component.

Given the above closure arguments, canon_lift is defined:

def canon_lift(opts, inst, callee, ft, start_thunk, return_thunk):
  export_call = ExportCall(opts, inst)
  trap_if(not inst.may_enter)

  flat_args = lower_values(export_call, MAX_FLAT_PARAMS, start_thunk(), ft.param_types())
  flat_results = call_and_trap_on_throw(callee, flat_args)
  return_thunk(lift_values(export_call, MAX_FLAT_RESULTS, CoreValueIter(flat_results), ft.result_types()))

  if opts.post_return is not None:
    call_and_trap_on_throw(opts.post_return, flat_results)

  export_call.exit()

def call_and_trap_on_throw(callee, args):
  try:
    return callee(args)
  except CoreWebAssemblyException:
    trap()

Uncaught Core WebAssembly exceptions result in a trap at component boundaries. Thus, if a component wishes to signal an error, it must use some sort of explicit type such as result (whose error case particular language bindings may choose to map to and from exceptions).

The start_thunk and return_thunk are used to model the interleaving of reading arguments out of the caller's stack and memory and writing results back into the caller's stack and memory. After the results have been copied from the callee's memory into the caller's memory, the callee's post_return function is called to allow the callee to reclaim any memory.

canon lower

For a canonical definition:

(canon lower $callee:<funcidx> $opts:<canonopt>* (core func $f))

where $callee has type $ft, validation specifies:

  • $f is given type flatten_functype($ft, 'lower')
  • a memory is present if required by lifting and is a subtype of (memory 1)
  • a realloc is present if required by lifting and has type (func (param i32 i32 i32 i32) (result i32))
  • there is no post-return in $opts

When instantiating component instance $inst:

  • Define $f to be the partially-bound closure: canon_lower($opts, $inst, $callee, $ft)

where canon_lower is defined:

def canon_lower(opts, inst, callee, calling_import, ft, flat_args):
  import_call = ImportCall(opts, inst)
  trap_if(not inst.may_leave)

  assert(inst.may_enter)
  if calling_import:
    inst.may_enter = False

  flat_args = CoreValueIter(flat_args)
  flat_results = None

  def start_thunk():
    return lift_values(import_call, MAX_FLAT_PARAMS, flat_args, ft.param_types())

  def return_thunk(results):
    nonlocal flat_results
    flat_results = lower_values(import_call, MAX_FLAT_RESULTS, results, ft.result_types(), flat_args)

  callee(start_thunk, return_thunk)

  if calling_import:
    inst.may_enter = True

  import_call.exit()
  return flat_results

Since any cross-component call necessarily transits through a statically-known canon_lower+canon_lift call pair, an AOT compiler can fuse canon_lift and canon_lower into a single, efficient trampoline. In the future this may allow efficient compilation of permissive subtyping between components (including the elimination of string operations on the labels of records and variants) as well as post-MVP adapter functions.

By clearing may_enter for the duration of calls to imports, the may_enter guard in canon_lift ensures that components cannot be externally reentered, which is part of the component invariants. The calling_import condition allows a parent component to call into a child component (which is, by definition, not a call to an import) and for the child to then reenter the parent through a function the parent explicitly supplied to the child's instantiate. This form of internal reentrance allows the parent to fully virtualize the child's imports.

Because may_enter is not cleared on the exceptional exit path taken by trap(), if there is a trap during Core WebAssembly execution of lifting or lowering, the component is left permanently un-enterable, ensuring the lockdown-after-trap component invariant.

canon resource.new

For a canonical definition:

(canon resource.new $rt (core func $f))

validation specifies:

  • $rt must refer to locally-defined (not imported) resource type
  • $f is given type (func (param $rt.rep) (result i32)), where $rt.rep is currently fixed to be i32.

Calling $f invokes the following function, which adds an owning handle containing the given resource representation in the current component instance's handle table:

def canon_resource_new(inst, rt, rep):
  h = HandleElem(rep, own=True)
  i = inst.handles.add(rt, h)
  return [i]

canon resource.drop

For a canonical definition:

(canon resource.drop $rt (core func $f))

validation specifies:

  • $rt must refer to resource type
  • $f is given type (func (param i32))

Calling $f invokes the following function, which removes the handle from the current component instance's handle table and, if the handle was owning, calls the resource's destructor.

def canon_resource_drop(inst, rt, i):
  h = inst.handles.remove(rt, i)
  if h.own:
    assert(h.scope is None)
    trap_if(h.lend_count != 0)
    trap_if(inst is not rt.impl and not rt.impl.may_enter)
    if rt.dtor:
      rt.dtor(h.rep)
  else:
    h.scope.drop_borrow()
  return []

The may_enter guard ensures the non-reentrance component invariant, since a destructor call is analogous to a call to an export.

canon resource.rep

For a canonical definition:

(canon resource.rep $rt (core func $f))

validation specifies:

  • $rt must refer to a locally-defined (not imported) resource type
  • $f is given type (func (param i32) (result $rt.rep)), where $rt.rep is currently fixed to be i32.

Calling $f invokes the following function, which extracts the resource representation from the handle.

def canon_resource_rep(inst, rt, i):
  h = inst.handles.get(rt, i)
  return [h.rep]

Note that the "locally-defined" requirement above ensures that only the component instance defining a resource can access its representation.

🧵 canon thread.spawn

For a canonical definition:

(canon thread.spawn (type $ft) (core func $st))

validation specifies:

  • $ft must refer to a shared function type; initially, only the type (func shared (param $c i32)) is allowed (see explanation below)
  • $st is given type (func (param $f (ref null $ft)) (param $c i32) (result $e i32)).

Note: ideally, a thread could be spawned with arbitrary thread parameters. Currently, that would require additional work in the toolchain to support so, for simplicity, the current proposal simply fixes a single i32 parameter type. However, thread.spawn could be extended to allow arbitrary thread parameters in the future, once it's concretely beneficial to the toolchain. The inclusion of $ft ensures backwards compatibility for when arbitrary parameters are allowed.

Calling $st checks that the reference $f is not null. Then, it spawns a thread which:

  • invokes $f with $c
  • executes $f until completion or trap in a shared context as described by the shared-everything threads proposal.

In pseudocode, $st looks like:

def canon_thread_spawn(f, c):
  trap_if(f is None)
  if DETERMINISTIC_PROFILE:
    return [-1]

  def thread_start():
    try:
      f(c)
    except CoreWebAssemblyException:
      trap()

  if spawn(thread_start):
    return [0]
  else:
    return [-1]

🧵 canon thread.hw_concurrency

For a canonical definition:

(canon thread.hw_concurrency (core func $f))

validation specifies:

  • $f is given type (func shared (result i32)).

Calling $f returns the number of threads the underlying hardware can be expected to execute concurrently. This value can be artificially limited by engine configuration and is not allowed to change over the lifetime of a component instance.

def canon_thread_hw_concurrency():
  if DETERMINISTIC_PROFILE:
    return [1]
  else:
    return [NUM_ALLOWED_THREADS]