// The binary form of a WebAssembly module: indices resolved, ready to encode.
//
// Vendored from Milky2018/wasm_core 0.5.0 (Apache-2.0) and then changed; see
// VENDORED.md for what and why. The short version: the model is upstream's, the
// type spine is ours.
//
// THE SEAM, from the other side. `wasm_types` is generic over `Idx` precisely so
// two instances of the type family exist without duplicating it: Wax names its
// types (`Idx = @ast.Ident`), the binary form numbers them (`Idx = Int`). This
// package is that second instance, so it supplies the four wrapper-carrying
// types -- functype, comptype, subtype, rectype -- that `wasm_types` leaves to
// each instance, and takes everything else from the shared spine.
//
// Upstream had its own flat `ValueType`: 31 constructors enumerating the ref
// types somebody happened to need (`RefStruct`, `RefNullFuncTyped`, ...). It
// cannot spell `(ref null $t)` for a `$t` that is not a func, struct or array,
// has no `exact` and no `cont`, and -- because nullability is baked into the
// constructor name rather than carried as a field -- it forced `ref.test` and
// `ref.test null` apart into two instructions that differ only in a flag. All
// of that goes away here.
///|
/// A value type, at the binary form's index instantiation.
pub type ValType = @wasm_types.ValType[Int]
///|
/// A reference type: nullability plus what it points at.
pub type RefType = @wasm_types.RefType[Int]
///|
/// What a reference can point at.
pub type HeapType = @wasm_types.HeapType[Int]
///|
/// What a struct field or array element stores.
pub type StorageType = @wasm_types.StorageType[Int]
///|
/// A struct field or array element type: a storage type plus mutability.
pub type FieldType = @wasm_types.FieldType[Int]
///|
/// A global's type: a value type plus mutability.
pub type GlobalType = @wasm_types.GlobalType[Int]
///|
/// The size bounds of a memory or table.
///
/// Upstream split this into a bare `{min, max}` plus `is_memory64`,
/// `page_size_log2` and `is_table64` fields hung off the memory and table types.
/// The shared spine already carries all of them, and carries `shared` besides,
/// which upstream cannot express at all -- so a shared memory was silently
/// unrepresentable there.
pub type Limits = @wasm_types.Limits
// ============================================================
// Defined types
// ============================================================
///|
/// A function type.
pub(all) struct FuncType {
params : Array[ValType]
results : Array[ValType]
} derive(Eq, Debug)
///|
/// A struct type.
pub(all) struct StructType {
fields : Array[FieldType]
} derive(Eq, Debug)
///|
/// An array type.
pub(all) struct ArrayType {
element : FieldType
} derive(Eq, Debug)
///|
/// What a defined type actually defines.
pub(all) enum CompositeType {
Func(FuncType)
Struct(StructType)
Array(ArrayType)
/// A continuation over the function type at this index -- the stack-switching
/// proposal's `(cont $ft)`. Upstream had the switching INSTRUCTIONS nowhere
/// and the type nowhere either; the instructions arrived first, but a module
/// that uses them has to be able to define what they act on.
Cont(Int)
} derive(Eq, Debug)
///|
/// A defined type, with its supertypes and whether it may be subtyped further.
///
/// `descriptor` and `describes` are the custom-descriptors proposal's two
/// clauses, and they come in pairs: `descriptor` names the type of this
/// struct's runtime descriptor, `describes` names the struct this one is the
/// descriptor OF. A type may carry either, both, or neither.
pub(all) struct SubType {
final_ : Bool
supertypes : Array[Int]
descriptor : Int?
describes : Int?
composite : CompositeType
} derive(Eq, Debug)
///|
/// A run of defined types emitted as one `rec (...)`.
///
/// Upstream carried this as one group id per type and then never read it: the
/// type section wrote every type flat, so a module whose types referred to each
/// other came out as a set of unrelated definitions.
///
/// `explicit` is not redundant with `len == 1`. A singleton `rec` is a
/// different type from the same definition written bare -- recursive type
/// identity is by group -- so the encoder cannot infer the prefix from the
/// length, and what the source said has to be recorded.
pub(all) struct RecGroup {
start : Int
len : Int
explicit : Bool
} derive(Eq, Debug)
///|
/// A subtype that is final and inherits from nothing -- the common case.
pub fn SubType::simple(composite : CompositeType) -> SubType {
{ final_: true, supertypes: [], descriptor: None, describes: None, composite }
}
///|
/// A final function type, from its parameters and results.
pub fn SubType::func(
params : Array[ValType],
results : Array[ValType],
) -> SubType {
SubType::simple(Func({ params, results }))
}
// ============================================================
// Instructions
// ============================================================
///|
pub(all) enum Instruction {
// Control instructions
Unreachable
Nop
Block(BlockType, Array[Instruction])
Loop(BlockType, Array[Instruction])
If(BlockType, Array[Instruction], Array[Instruction])
Br(Int) // label index
BrIf(Int)
BrTable(Array[Int], Int) // labels, default
Return
Call(Int) // function index
CallIndirect(Int, Int) // type index, table index
CallRef(Int) // type index (WasmGC)
ReturnCall(Int) // function index (tail call)
ReturnCallIndirect(Int, Int) // type index, table index (tail call)
ReturnCallRef(Int) // type index (tail call, WasmGC)
// Exception handling instructions
Throw(Int) // tag index - throw exception with tag
ThrowRef // throw exception from exnref on stack
TryTable(BlockType, Array[CatchHandler], Array[Instruction]) // block type, handlers, body
/// The DEPRECATED legacy handler: block type, body, per-tag catches, and an
/// optional catch-all. Unlike `try_table`, whose handlers are immediates that
/// branch out, these handlers are inline bodies that run in place.
LegacyTry(
BlockType,
Array[Instruction],
Array[(Int, Array[Instruction])],
Array[Instruction]?
)
// Parametric instructions
Drop
Select
SelectTyped(Array[ValType]) // select with explicit result type
// Variable instructions
LocalGet(Int)
LocalSet(Int)
LocalTee(Int)
GlobalGet(Int)
GlobalSet(Int)
// Table instructions
TableGet(Int) // table index
TableSet(Int) // table index
TableSize(Int) // table index
TableGrow(Int) // table index
TableFill(Int) // table index
TableCopy(Int, Int) // dest table index, src table index
/// The TABLE index first, then the element segment -- and, as for
/// `MemoryInit`, the encoding writes them the other way round.
TableInit(Int, Int)
// Memory instructions - (memidx, align, offset)
// offset is Int64 to support memory64 64-bit offsets
I32Load(Int, Int, Int64) // memidx, align, offset
I64Load(Int, Int, Int64)
F32Load(Int, Int, Int64)
F64Load(Int, Int, Int64)
I32Load8S(Int, Int, Int64)
I32Load8U(Int, Int, Int64)
I32Load16S(Int, Int, Int64)
I32Load16U(Int, Int, Int64)
I64Load8S(Int, Int, Int64)
I64Load8U(Int, Int, Int64)
I64Load16S(Int, Int, Int64)
I64Load16U(Int, Int, Int64)
I64Load32S(Int, Int, Int64)
I64Load32U(Int, Int, Int64)
I32Store(Int, Int, Int64)
I64Store(Int, Int, Int64)
F32Store(Int, Int, Int64)
F64Store(Int, Int, Int64)
I32Store8(Int, Int, Int64)
I32Store16(Int, Int, Int64)
I64Store8(Int, Int, Int64)
I64Store16(Int, Int, Int64)
I64Store32(Int, Int, Int64)
MemorySize(Int) // memidx
MemoryGrow(Int) // memidx
/// The MEMORY index first, then the data segment. Note the encoding writes
/// them the other way round -- segment, then memory -- which is exactly why
/// the order is stated here rather than left to be inferred from the bytes.
MemoryInit(Int, Int)
DataDrop(Int) // data segment index
MemoryCopy(Int, Int) // dest memidx, src memidx
MemoryFill(Int) // memidx
ElemDrop(Int) // element segment index
// Atomic instructions (0xFE prefix)
/// The 0xFE sub-opcode FIRST, then the memarg: memory index, alignment
/// exponent, offset. The sub-opcode leads because that is the order the
/// bytes go out in, and naming it second invited reading the memory index
/// as the opcode.
Atomic(Int, Int, Int, Int64)
/// `atomic.fence`, which unlike every other atomic has no memory operand.
AtomicFence
// Stack switching. Absent from upstream entirely -- the model had no way to
// say any of this -- though the Wax AST has had all seven since Phase 2.
ContNew(Int) // type index
ContBind(Int, Int) // source type index, target type index
Suspend(Int) // tag index
Resume(Int, Array[OnClause]) // type index, resume table
ResumeThrow(Int, Int, Array[OnClause]) // type index, tag index, resume table
ResumeThrowRef(Int, Array[OnClause]) // type index, resume table
Switch(Int, Int) // type index, tag index
/// An instruction with compilation hints on it.
///
/// A wrapper rather than a field on every instruction: hints are rare, the
/// enum is 500 constructors wide, and the encoder needs exactly one place to
/// notice them. It emits no opcode of its own.
Hinted(InstrHints, Instruction)
/// A constant, with the text the source wrote it as.
///
/// A literal is not its value: `0x4` and `4` are the same i32 and two
/// different programs to read, and `0x1.4p+3` does not come back from the
/// bits of the float it denotes. The binary encodes the value and forgets
/// the spelling; the text format writes what was written, so the spelling
/// rides along on the instruction the way a branch hint does.
///
/// Carries no opcode of its own -- the encoder unwraps it.
Spelled(String, Instruction)
/// An array construction that the source wrote as a STRING literal.
///
/// The literal lowers to one constant per byte and an `array.new_fixed` over
/// them; the text format writes `(@string "...")` and lets the reader assume
/// the rest. Nothing in a run of constants says it was ever text, so the
/// bytes are carried here.
///
/// Carries no opcode of its own either.
FromString(Bytes, Instruction)
/// An `i32.const` the source wrote as a CHARACTER literal, carrying that
/// character's UTF-8 bytes. Its sibling above, for the other literal whose
/// value says nothing about how it was written.
FromChar(Bytes, Instruction)
// Reference type instructions
RefNull(HeapType) // ref.null: push a null reference to this heap type
RefIsNull // ref.is_null: test if reference is null
RefFunc(Int) // ref.func: push reference to function by index
RefAsNonNull // ref.as_non_null: convert nullable ref to non-null ref
RefEqInstr // ref.eq: compare two references for equality
BrOnNull(Int) // br_on_null: branch if reference is null
BrOnNonNull(Int) // br_on_non_null: branch if reference is non-null
// Numeric instructions - Constants
I32Const(Int)
I64Const(Int64)
F32Const(Float)
F64Const(Double)
// Numeric instructions - i32
I32Eqz
I32Eq
I32Ne
I32LtS
I32LtU
I32GtS
I32GtU
I32LeS
I32LeU
I32GeS
I32GeU
I32Clz
I32Ctz
I32Popcnt
I32Add
I32Sub
I32Mul
I32DivS
I32DivU
I32RemS
I32RemU
I32And
I32Or
I32Xor
I32Shl
I32ShrS
I32ShrU
I32Rotl
I32Rotr
// Sign-extension operators
I32Extend8S
I32Extend16S
// Numeric instructions - i64
I64Eqz
I64Eq
I64Ne
I64LtS
I64LtU
I64GtS
I64GtU
I64LeS
I64LeU
I64GeS
I64GeU
I64Clz
I64Ctz
I64Popcnt
I64Add
I64Sub
I64Mul
I64MulWideS
I64MulWideU
I64Add128
I64Sub128
I64DivS
I64DivU
I64RemS
I64RemU
I64And
I64Or
I64Xor
I64Shl
I64ShrS
I64ShrU
I64Rotl
I64Rotr
// Sign-extension operators
I64Extend8S
I64Extend16S
I64Extend32S
// Numeric instructions - f32
F32Eq
F32Ne
F32Lt
F32Gt
F32Le
F32Ge
F32Abs
F32Neg
F32Ceil
F32Floor
F32Trunc
F32Nearest
F32Sqrt
F32Add
F32Sub
F32Mul
F32Div
F32Min
F32Max
F32Copysign
// Numeric instructions - f64
F64Eq
F64Ne
F64Lt
F64Gt
F64Le
F64Ge
F64Abs
F64Neg
F64Ceil
F64Floor
F64Trunc
F64Nearest
F64Sqrt
F64Add
F64Sub
F64Mul
F64Div
F64Min
F64Max
F64Copysign
// Conversion instructions
I32WrapI64
I32TruncF32S
I32TruncF32U
I32TruncF64S
I32TruncF64U
I64ExtendI32S
I64ExtendI32U
I64TruncF32S
I64TruncF32U
I64TruncF64S
I64TruncF64U
F32ConvertI32S
F32ConvertI32U
F32ConvertI64S
F32ConvertI64U
F32DemoteF64
F64ConvertI32S
F64ConvertI32U
F64ConvertI64S
F64ConvertI64U
F64PromoteF32
I32ReinterpretF32
I64ReinterpretF64
F32ReinterpretI32
F64ReinterpretI64
// Saturating truncation instructions (nontrapping float-to-int)
I32TruncSatF32S
I32TruncSatF32U
I32TruncSatF64S
I32TruncSatF64U
I64TruncSatF32S
I64TruncSatF32U
I64TruncSatF64S
I64TruncSatF64U
// GC instructions - struct operations
/// The custom-descriptors forms: like `struct.new` and its default sibling,
/// but taking the runtime DESCRIPTOR as a further operand, pushed last --
/// above the field values, because that is the order the instruction reads.
StructNewDesc(Int)
StructNewDefaultDesc(Int)
/// A struct's own descriptor, read back off it.
RefGetDesc(Int)
/// A cast that tests a value against a DESCRIPTOR rather than against a type
/// immediate: the descriptor is an operand, so the same instruction can test
/// against a descriptor only known at run time.
RefCastDescEq(RefType)
/// The branching forms of the same test. Both reference types are immediates
/// and their nullability rides in one byte before the label.
BrOnCastDescEq(Int, RefType, RefType)
BrOnCastDescEqFail(Int, RefType, RefType)
StructNew(Int) // type index
StructNewDefault(Int) // type index
StructGet(Int, Int) // type index, field index
StructGetS(Int, Int) // type index, field index (signed packed)
StructGetU(Int, Int) // type index, field index (unsigned packed)
StructSet(Int, Int) // type index, field index
// GC instructions - array operations
ArrayNew(Int) // type index
ArrayNewDefault(Int) // type index
ArrayNewFixed(Int, Int) // type index, length
ArrayNewData(Int, Int) // type index, data index
ArrayNewElem(Int, Int) // type index, elem index
ArrayGet(Int) // type index
ArrayGetS(Int) // type index (signed packed)
ArrayGetU(Int) // type index (unsigned packed)
ArraySet(Int) // type index
ArrayLen // get array length
ArrayFill(Int) // type index
ArrayCopy(Int, Int) // dest type index, src type index
ArrayInitData(Int, Int) // type index, data index
ArrayInitElem(Int, Int) // type index, elem index
// GC instructions - reference casting
RefTest(RefType) // ref.test: does the reference match this type?
RefCast(RefType) // ref.cast: cast, trapping on failure
BrOnCast(Int, RefType, RefType) // label, source type, target type
BrOnCastFail(Int, RefType, RefType) // label, source type, target type
// GC instructions - i31
RefI31 // create i31ref from i32
I31GetS // get signed i32 from i31ref
I31GetU // get unsigned i32 from i31ref
// GC instructions - type conversion
AnyConvertExtern // convert externref to anyref
ExternConvertAny // convert anyref to externref
// ============================================================
// SIMD Instructions (128-bit packed SIMD)
// ============================================================
// SIMD constants
V128Const(Bytes) // 16-byte constant
// SIMD load/store - (memidx, align, offset)
V128Load(Int, Int, Int64)
V128Load8x8S(Int, Int, Int64)
V128Load8x8U(Int, Int, Int64)
V128Load16x4S(Int, Int, Int64)
V128Load16x4U(Int, Int, Int64)
V128Load32x2S(Int, Int, Int64)
V128Load32x2U(Int, Int, Int64)
V128Load8Splat(Int, Int, Int64)
V128Load16Splat(Int, Int, Int64)
V128Load32Splat(Int, Int, Int64)
V128Load64Splat(Int, Int, Int64)
V128Load32Zero(Int, Int, Int64)
V128Load64Zero(Int, Int, Int64)
V128Store(Int, Int, Int64)
// SIMD load/store lane - (memidx, align, offset, lane)
V128Load8Lane(Int, Int, Int64, Int)
V128Load16Lane(Int, Int, Int64, Int)
V128Load32Lane(Int, Int, Int64, Int)
V128Load64Lane(Int, Int, Int64, Int)
V128Store8Lane(Int, Int, Int64, Int)
V128Store16Lane(Int, Int, Int64, Int)
V128Store32Lane(Int, Int, Int64, Int)
V128Store64Lane(Int, Int, Int64, Int)
// SIMD shuffle/swizzle
I8x16Shuffle(FixedArray[Int]) // 16 lane indices
I8x16Swizzle
// SIMD splat (scalar -> vector)
I8x16Splat
I16x8Splat
I32x4Splat
I64x2Splat
F32x4Splat
F64x2Splat
// SIMD extract lane (vector -> scalar)
I8x16ExtractLaneS(Int)
I8x16ExtractLaneU(Int)
I16x8ExtractLaneS(Int)
I16x8ExtractLaneU(Int)
I32x4ExtractLane(Int)
I64x2ExtractLane(Int)
F32x4ExtractLane(Int)
F64x2ExtractLane(Int)
// SIMD replace lane (vector, scalar -> vector)
I8x16ReplaceLane(Int)
I16x8ReplaceLane(Int)
I32x4ReplaceLane(Int)
I64x2ReplaceLane(Int)
F32x4ReplaceLane(Int)
F64x2ReplaceLane(Int)
// i8x16 comparisons
I8x16Eq
I8x16Ne
I8x16LtS
I8x16LtU
I8x16GtS
I8x16GtU
I8x16LeS
I8x16LeU
I8x16GeS
I8x16GeU
// i16x8 comparisons
I16x8Eq
I16x8Ne
I16x8LtS
I16x8LtU
I16x8GtS
I16x8GtU
I16x8LeS
I16x8LeU
I16x8GeS
I16x8GeU
// i32x4 comparisons
I32x4Eq
I32x4Ne
I32x4LtS
I32x4LtU
I32x4GtS
I32x4GtU
I32x4LeS
I32x4LeU
I32x4GeS
I32x4GeU
// i64x2 comparisons
I64x2Eq
I64x2Ne
I64x2LtS
I64x2GtS
I64x2LeS
I64x2GeS
// f32x4 comparisons
F32x4Eq
F32x4Ne
F32x4Lt
F32x4Gt
F32x4Le
F32x4Ge
// f64x2 comparisons
F64x2Eq
F64x2Ne
F64x2Lt
F64x2Gt
F64x2Le
F64x2Ge
// v128 bitwise operations
V128Not
V128And
V128AndNot
V128Or
V128Xor
V128Bitselect
V128AnyTrue
// i8x16 operations
I8x16Abs
I8x16Neg
I8x16Popcnt
I8x16AllTrue
I8x16Bitmask
I8x16NarrowI16x8S
I8x16NarrowI16x8U
I8x16Shl
I8x16ShrS
I8x16ShrU
I8x16Add
I8x16AddSatS
I8x16AddSatU
I8x16Sub
I8x16SubSatS
I8x16SubSatU
I8x16MinS
I8x16MinU
I8x16MaxS
I8x16MaxU
I8x16AvgrU
// i16x8 operations
I16x8ExtAddPairwiseI8x16S
I16x8ExtAddPairwiseI8x16U
I16x8Abs
I16x8Neg
I16x8Q15MulrSatS
I16x8AllTrue
I16x8Bitmask
I16x8NarrowI32x4S
I16x8NarrowI32x4U
I16x8ExtendLowI8x16S
I16x8ExtendHighI8x16S
I16x8ExtendLowI8x16U
I16x8ExtendHighI8x16U
I16x8Shl
I16x8ShrS
I16x8ShrU
I16x8Add
I16x8AddSatS
I16x8AddSatU
I16x8Sub
I16x8SubSatS
I16x8SubSatU
I16x8Mul
I16x8MinS
I16x8MinU
I16x8MaxS
I16x8MaxU
I16x8AvgrU
I16x8ExtMulLowI8x16S
I16x8ExtMulHighI8x16S
I16x8ExtMulLowI8x16U
I16x8ExtMulHighI8x16U
// i32x4 operations
I32x4ExtAddPairwiseI16x8S
I32x4ExtAddPairwiseI16x8U
I32x4Abs
I32x4Neg
I32x4AllTrue
I32x4Bitmask
I32x4ExtendLowI16x8S
I32x4ExtendHighI16x8S
I32x4ExtendLowI16x8U
I32x4ExtendHighI16x8U
I32x4Shl
I32x4ShrS
I32x4ShrU
I32x4Add
I32x4Sub
I32x4Mul
I32x4MinS
I32x4MinU
I32x4MaxS
I32x4MaxU
I32x4DotI16x8S
I32x4ExtMulLowI16x8S
I32x4ExtMulHighI16x8S
I32x4ExtMulLowI16x8U
I32x4ExtMulHighI16x8U
// i64x2 operations
I64x2Abs
I64x2Neg
I64x2AllTrue
I64x2Bitmask
I64x2ExtendLowI32x4S
I64x2ExtendHighI32x4S
I64x2ExtendLowI32x4U
I64x2ExtendHighI32x4U
I64x2Shl
I64x2ShrS
I64x2ShrU
I64x2Add
I64x2Sub
I64x2Mul
I64x2ExtMulLowI32x4S
I64x2ExtMulHighI32x4S
I64x2ExtMulLowI32x4U
I64x2ExtMulHighI32x4U
// f32x4 operations
F32x4Ceil
F32x4Floor
F32x4Trunc
F32x4Nearest
F32x4Abs
F32x4Neg
F32x4Sqrt
F32x4Add
F32x4Sub
F32x4Mul
F32x4Div
F32x4Min
F32x4Max
F32x4Pmin
F32x4Pmax
// f64x2 operations
F64x2Ceil
F64x2Floor
F64x2Trunc
F64x2Nearest
F64x2Abs
F64x2Neg
F64x2Sqrt
F64x2Add
F64x2Sub
F64x2Mul
F64x2Div
F64x2Min
F64x2Max
F64x2Pmin
F64x2Pmax
// SIMD conversions
I32x4TruncSatF32x4S
I32x4TruncSatF32x4U
F32x4ConvertI32x4S
F32x4ConvertI32x4U
I32x4TruncSatF64x2SZero
I32x4TruncSatF64x2UZero
F64x2ConvertLowI32x4S
F64x2ConvertLowI32x4U
F32x4DemoteF64x2Zero
F64x2PromoteLowF32x4
// Relaxed SIMD instructions
I8x16RelaxedSwizzle
I32x4RelaxedTruncF32x4S
I32x4RelaxedTruncF32x4U
I32x4RelaxedTruncF64x2SZero
I32x4RelaxedTruncF64x2UZero
F32x4RelaxedMadd
F32x4RelaxedNmadd
F64x2RelaxedMadd
F64x2RelaxedNmadd
I8x16RelaxedLaneselect
I16x8RelaxedLaneselect
I32x4RelaxedLaneselect
I64x2RelaxedLaneselect
F32x4RelaxedMin
F32x4RelaxedMax
F64x2RelaxedMin
F64x2RelaxedMax
I16x8RelaxedQ15mulrS
I16x8RelaxedDotI8x16I7x16S
I32x4RelaxedDotI8x16I7x16AddS
} derive(Eq, Debug)
///|
/// What a block, loop or `if` consumes and produces.
pub(all) enum BlockType {
Empty
Value(ValType)
MultiValue(Array[ValType])
InlineType(Array[ValType], Array[ValType])
TypeIndex(Int)
} derive(Eq, Debug)
///|
/// The compilation hints attached to a single instruction.
///
/// These do not change what the instruction does; they are advice, carried out
/// of line in the `metadata.code.*` sections, keyed by the instruction's byte
/// offset within its function body. Which is why they hang off the instruction
/// here: the offset is only knowable while encoding, and only the encoder knows
/// it.
pub(all) struct InstrHints {
/// `#[likely]` / `#[unlikely]` on a branch.
branch : Bool?
/// `#[freq]`: a saturating base-2 logarithm of executions per call, already
/// reduced to the byte the section stores.
freq : Int?
/// `#[targets]`: likely call targets, as (function index, percentage).
targets : Array[(Int, Int)]?
} derive(Eq, Debug)
///|
/// No hints.
pub fn InstrHints::none() -> InstrHints {
{ branch: None, freq: None, targets: None }
}
///|
pub fn InstrHints::is_empty(self : InstrHints) -> Bool {
self.branch is None && self.freq is None && self.targets is None
}
///|
/// `#[priority]`: advice about the function as a whole rather than about one
/// instruction, which is why it lives on the code entry and is recorded at
/// offset 0 -- the position that means "the function itself".
pub(all) struct Priority {
compilation : Int
optimization : Int?
} derive(Eq, Debug)
///|
/// One arm of a resume table: what happens when the continuation suspends with
/// a given tag.
pub(all) enum OnClause {
/// Branch to this label.
OnLabel(Int, Int)
/// Switch to the continuation on the stack.
OnSwitch(Int)
} derive(Eq, Debug)
///|
/// One arm of a `try_table`.
pub(all) enum CatchHandler {
/// Catch this tag, branching to this label.
Catch(Int, Int)
/// As `Catch`, but the exception reference is pushed too.
CatchRef(Int, Int)
/// Catch anything, branching to this label.
CatchAll(Int)
/// As `CatchAll`, but the exception reference is pushed too.
CatchAllRef(Int)
} derive(Eq, Debug)
// ============================================================
// Module structure
// ============================================================
///|
/// A memory's type.
pub(all) struct MemoryType {
limits : Limits
} derive(Eq, Debug)
///|
/// A table's type. The element type is a REFERENCE type, not a value type:
/// nothing else can go in a table.
pub(all) struct TableType {
elem_type : RefType
limits : Limits
} derive(Eq, Debug)
///|
/// A defined table, with the initialiser every element starts at.
pub(all) struct Table {
type_ : TableType
init : Array[Instruction]?
init_spans : Array[Span]
} derive(Eq, Debug)
///|
/// An exception tag: an index into the type section, naming a function type
/// whose parameters are the exception's payload and whose results are empty.
pub(all) struct TagType {
type_idx : Int
} derive(Eq, Debug)
///|
/// What an import brings in.
pub(all) enum ImportDesc {
/// A function's type index, and whether the import is EXACT -- a `ref.func`
/// on it then yields an exact reference rather than a plain one. The flag
/// rides in the kind byte, not beside the index: 0x20 instead of 0x00.
Func(Int, Bool)
Table(TableType)
Memory(MemoryType)
Global(GlobalType)
Tag(Int)
} derive(Eq, Debug)
///|
pub(all) struct Import {
mod_name : Bytes
name : Bytes
desc : ImportDesc
/// The rest of a COMPACT group, when this entry heads one.
///
/// The compact-import-section proposal writes a module name once for a run
/// of imports from it. The marker sits where an externtype kind byte would,
/// and neither marker is a valid kind, so a plain import stays unambiguous.
group : ImportGroup?
} derive(Eq, Debug)
///|
/// The two compact forms: per-item descriptors, or one shared descriptor and a
/// list of names.
pub(all) enum ImportGroup {
/// Marker 0x7F: each item carries its own name and descriptor.
Heterogeneous(Array[(Bytes, ImportDesc)])
/// Marker 0x7E: one descriptor, then the names that share it.
Homogeneous(Array[Bytes])
} derive(Eq, Debug)
///|
/// What an export names, as an index into the matching space.
pub(all) enum ExportDesc {
Func(Int)
Table(Int)
Memory(Int)
Global(Int)
Tag(Int)
} derive(Eq, Debug)
///|
pub(all) struct Export {
name : Bytes
desc : ExportDesc
} derive(Eq, Debug)
///|
/// A function body: its locals, then its instructions.
/// One node's run in a body: where it starts, where its OWN emission begins
/// (everything before that is its operands), where it ends, and where the
/// source wrote it.
pub(all) struct Span {
start : Int
head : Int
end : Int
loc : @basic.Location
} derive(Eq, Debug)
///|
pub(all) struct FunctionCode {
locals : Array[ValType]
body : Array[Instruction]
/// Where each source instruction's run starts and ends in `body`, in the
/// order the runs COMPLETE.
///
/// Not part of the format and not encoded: the binary writes the body flat
/// and the text writes it FOLDED, and this is the nesting the text needs.
/// It is recorded where the body was built rather than re-derived from an
/// arity table, because a second derivation could disagree with the first
/// and nothing would say which output was wrong.
spans : Array[Span]
/// The same, for each NESTED body -- a block's, a loop's, each arm of an
/// `if` -- in the order those bodies completed while the function was
/// lowered: deepest first, then left to right. A nested body is a separate
/// array with its own indices, so its spans cannot live in the list above;
/// that order is what pairs them up again.
nested_spans : Array[Array[Span]]
/// The `#[if(..)]` groups written INSIDE this body, matched to the statements
/// they hold by where each was written -- the same way the module's are.
conditionals : Array[CondGroup]
/// `#[priority]` on the function, if any.
priority : Priority?
} derive(Eq, Debug)
///|
pub(all) struct Global {
type_ : GlobalType
init : Array[Instruction]
/// The spans of that initialiser, so the text form can fold it. A constant
/// expression is its own body with its own indices, so its spans are its
/// own too.
init_spans : Array[Span]
} derive(Eq, Debug)
///|
/// How an element segment reaches its table.
pub(all) enum ElemMode {
/// Copied into a table at instantiation, at this offset.
Active(Int, Array[Instruction])
/// Left for `table.init`.
Passive
/// Present only so the references in it count as declared.
Declarative
} derive(Eq, Debug)
///|
pub(all) struct Element {
mode : ElemMode
type_ : RefType
init : Array[Array[Instruction]]
/// One span list per initialiser, and one for the active offset.
init_spans : Array[Array[Span]]
offset_spans : Array[Span]
} derive(Eq, Debug)
///|
/// How a data segment reaches its memory.
///
/// Upstream had no such type: it hung a bare `memory_idx` off the segment and
/// signalled "passive" with a NEGATIVE index. That makes an unrepresentable
/// state representable -- an active segment with an empty offset expression --
/// and it decided the encoding from `offset.is_empty()`, so such a segment was
/// silently written as active-at-an-explicit-memory instead.
pub(all) enum DataMode {
/// Copied into this memory at instantiation, at this offset.
Active(Int, Array[Instruction])
/// Left for `memory.init`.
Passive
} derive(Eq, Debug)
///|
pub(all) struct Data {
mode : DataMode
init : Bytes
/// The spans of the active offset expression.
offset_spans : Array[Span]
/// How the source wrote those bytes.
///
/// A segment is a run of bytes to the binary and a list of PIECES to the
/// text: `"hdr" (f32 0.2 0.3 0.4) (i8 1 2 3)` flattens to bytes and does not
/// come back from them. Empty when nothing recorded it, and then the bytes
/// are written as one string.
spelling : Array[DataPiece]
} derive(Eq, Debug)
///|
/// One piece of a data segment, as the source wrote it.
pub(all) enum DataPiece {
/// A byte string.
PieceStr(Bytes)
/// A numeric run: the element keyword, then the literals as written. A
/// vector run keeps its shapes among them, because `(v128 i32x4 1 2 3 4
/// f64x2 1.0 2.0)` is one run of two vectors written two different ways.
PieceRun(String, Array[String])
} derive(Eq, Debug)
///|
/// The `name` custom section: what everything in the module is called.
///
/// Upstream had a single `func_names : Map[Int, String]` on the module, which
/// the encoder never read -- so every name was dropped. There are twelve name
/// spaces, not one, and `wasm_output.ml` emits all of them.
///
/// Maps rather than arrays, mirroring the reference's `IntMap`: the format
/// wants each vector in ascending index order, and making that the encoder's
/// job rather than the caller's is one fewer way to differ by a byte.
pub(all) struct Names {
mut module_ : Bytes?
functions : Map[Int, Bytes]
/// Locals, labels and fields are indexed twice: by the function or type they
/// belong to, then within it.
locals : Map[Int, Map[Int, Bytes]]
labels : Map[Int, Map[Int, Bytes]]
types : Map[Int, Bytes]
tables : Map[Int, Bytes]
memories : Map[Int, Bytes]
globals : Map[Int, Bytes]
elem : Map[Int, Bytes]
data : Map[Int, Bytes]
fields : Map[Int, Map[Int, Bytes]]
tags : Map[Int, Bytes]
} derive(Debug)
///|
/// No names at all, which is also what makes the section disappear.
pub fn Names::empty() -> Names {
{
module_: None,
functions: {},
locals: {},
labels: {},
types: {},
tables: {},
memories: {},
globals: {},
elem: {},
data: {},
fields: {},
tags: {},
}
}
///|
/// A whole module, in the form the encoder writes out.
pub(all) struct Module {
/// Every defined type, flat, because that is the space indices refer to.
types : Array[SubType]
/// How those types are grouped into `rec (...)`. The groups partition
/// `types` in order, so `rec_groups` is a view of the same array rather than
/// a second copy of it.
rec_groups : Array[RecGroup]
imports : Array[Import]
/// One type index per defined function, parallel to `codes`.
funcs : Array[Int]
tables : Array[Table]
memories : Array[MemoryType]
globals : Array[Global]
exports : Array[Export]
mut start : Int?
elems : Array[Element]
codes : Array[FunctionCode]
datas : Array[Data]
tags : Array[TagType]
/// What the TEXT form needs and the binary drops.
///
/// Everything else here is encoded. These are not: they are the source
/// facts a binary has no room for -- where a field stood, what a
/// declaration called its parameters -- kept together so that reading this
/// model for the encoding does not mean reading past them.
text : TextView
/// The `name` custom section.
names : Names
/// The `target_features` custom section: (prefix byte, feature name) pairs,
/// emitted verbatim and in order, because entries from other producers pass
/// through untouched.
target_features : Array[(Int, Bytes)]
} derive(Debug)
///|
/// The source facts the text form needs and the binary has no room for.
///
/// Recorded as the lowering goes rather than recovered afterwards: there is
/// nothing in the sections to recover them from, and a second derivation could
/// disagree with the first with nothing to say which output is wrong.
pub(all) struct TextView {
/// The order the SOURCE wrote the fields in.
///
/// The sections group items by kind, which is not the order they were
/// written, and the text format writes each field where it stood. Nothing in
/// the sections records that order, so the lowering notes it as it goes --
/// the same way it notes the instruction spans, and for the same reason: a
/// second derivation could disagree with the first.
///
/// Only the binary encoder's input is authoritative for the bytes; this is a
/// view for the text form, and is empty when nothing recorded it.
field_order : Array[FieldRef]
/// Where each of those fields stood in the source, parallel to
/// `field_order`.
///
/// The binary has no notion of a source position; the text format needs one,
/// because a comment attaches to the field it was written against and there
/// is nothing else to attach it to.
field_locs : Array[@basic.Location]
/// The parameter names a DECLARATION wrote, for the declarations whose
/// names the binary does not carry.
///
/// A name section can hold local names for an imported function, and the
/// reference does not write them -- so these cannot ride in `names.locals`
/// without changing the bytes. The text format shows them, so they are kept
/// apart from the section that is encoded.
decl_param_names : Map[ParamOwner, Map[Int, Bytes]]
/// How each declaration wrote its type.
///
/// `tag t: ft;` writes `(type $ft)`, `tag t(i32);` spells the signature out,
/// and `fn f: ft(i32);` writes BOTH -- they are independent clauses. The
/// binary keeps only the index, which is the same in all three cases, so
/// which was written has to be recorded.
decl_typeuse : Map[ParamOwner, TypeUse]
/// The exports written as a FIELD rather than as a clause, by export index.
///
/// Two reasons an export cannot be a clause on the thing it exports. A
/// GUARDED one has nowhere to put its condition. And an import inside a
/// compact group is not written as a form of its own at all -- the group
/// writes one name per item and one descriptor for all of them -- so there
/// is nothing for a clause to hang off.
standalone_exports : Map[Int, Bool]
/// The types something referred to BY NAME, by type index.
///
/// A type the source did not declare is written as a field only when
/// something names it: `(ref $)` in a global's type makes the entry
/// a field, and an entry nothing reaches is spelled inline wherever it is
/// used -- writing it too would say the same signature twice.
named_types : Map[Int, Bool]
/// The conditional groups the source wrote.
///
/// A conditional has no binary form -- that is the whole point of it, and
/// `-f wasm` refuses a module that still has one. The text keeps it, so the
/// condition and the extent of each branch are recorded here and the fields
/// are grouped back under them by where they stood.
conditionals : Array[CondGroup]
} derive(Debug)
///|
/// One `#[if(..)]` group: the condition as the TEXT format writes it, and
/// where each branch stood in the source.
pub(all) struct CondGroup {
cond : String
loc : @basic.Location
then_ : @basic.Location
else_ : @basic.Location?
} derive(Eq, Debug)
///|
/// An empty view, for a module nothing recorded one for.
pub fn TextView::empty() -> TextView {
{
field_order: [],
field_locs: [],
decl_param_names: Map([]),
decl_typeuse: Map([]),
standalone_exports: Map([]),
named_types: Map([]),
conditionals: [],
}
}
///|
/// Which of the two clauses a declaration wrote for its type. Both are
/// optional and they are independent: `fn f: ft(i32)` wrote both.
pub(all) struct TypeUse {
named : Bool
spelled : Bool
} derive(Eq, Debug)
///|
/// Whose parameters a name list belongs to. Imported functions and tags sit in
/// different index spaces, so the index alone does not say.
pub(all) enum ParamOwner {
OwnerFunc(Int)
OwnerTag(Int)
/// A declared `type t = fn(x: i32)`, by type index. Its parameter names are
/// no more encodable than an import's.
OwnerType(Int)
} derive(Eq, Hash, Debug)
///|
/// One field of a module, named by where its item landed.
pub(all) enum FieldRef {
/// The type-section entries one `type` or `rec` field declared.
FTypes(Array[Int])
/// A `#![feature = "name"]` inner attribute, which the text format keeps as
/// a leading `(@feature "name")` so the emitted module still says what it
/// needs. There is no binary section for it; the text is where it survives.
FFeature(Bytes)
FImport(Int)
FFunc(Int)
/// `(start $f)`. Written where the `#[start]` attribute was, because that is
/// what a conditional around it guards.
FStart(Int)
/// A standalone `(export "n" (func $f))`, by export index. Only a GUARDED
/// export takes this form; the rest are clauses on what they export.
FExport(Int)
FTable(Int)
FMemory(Int)
FGlobal(Int)
FTag(Int)
FElem(Int)
FData(Int)
} derive(Eq, Debug)
///|
/// A module with nothing in it.
pub fn Module::empty() -> Module {
{
types: [],
rec_groups: [],
text: TextView::empty(),
imports: [],
funcs: [],
tables: [],
memories: [],
globals: [],
exports: [],
start: None,
elems: [],
codes: [],
datas: [],
tags: [],
names: Names::empty(),
target_features: [],
}
}