NIR: An immediate representation for Nim. WIP

compiler/ic/bitabs.nim

@@ -13,6 +13,8 @@ type
     vals: seq[T] # indexed by LitId
     keys: seq[LitId]  # indexed by hash(val)
 
+proc initBiTable*[T](): BiTable[T] = BiTable[T](vals: @[], keys: @[])
+
 proc nextTry(h, maxHash: Hash): Hash {.inline.} =
   result = (h + 1) and maxHash
 

compiler/nir/cir.nim

@@ -0,0 +1,78 @@
+#
+#
+#           The Nim Compiler
+#        (c) Copyright 2023 Andreas Rumpf
+#
+#    See the file "copying.txt", included in this
+#    distribution, for details about the copyright.
+#
+
+# We produce C code as a list of tokens.
+
+import std / assertions
+import .. / ic / bitabs
+
+type
+  Token = LitId # indexing into the tokens BiTable[string]
+
+  PredefinedToken = enum
+    IgnoreMe = "<unused>"
+    EmptyToken = ""
+    DeclPrefix = "" # the next token is the name of a definition
+    CurlyLe = "{"
+    CurlyRi = "}"
+    ParLe = "("
+    ParRi = ")"
+    BracketLe = "["
+    BracketRi = "]"
+    NewLine = "\n"
+    Semicolon = ";"
+    Comma = ", "
+    Space = " "
+    Colon = ":"
+    Dot = "."
+    Arrow = "->"
+    Star = "*"
+    Amp = "&"
+    AsgnOpr = " = "
+    ScopeOpr = "::"
+    ConstKeyword = "const "
+    StaticKeyword = "static "
+    NimString = "NimString"
+    StrLitPrefix = "(NimChar*)"
+    StrLitNamePrefix = "Qstr"
+    LoopKeyword = "while (true) "
+    WhileKeyword = "while ("
+    IfKeyword = "if ("
+    ElseKeyword = "else "
+    SwitchKeyword = "switch ("
+    CaseKeyword = "case "
+    DefaultKeyword = "default:"
+    BreakKeyword = "break"
+    NullPtr = "nullptr"
+    IfNot = "if (!("
+    ReturnKeyword = "return "
+
+const
+  ModulePrefix = Token(int(ReturnKeyword)+1)
+
+proc fillTokenTable(tab: var BiTable[string]) =
+  for e in EmptyToken..high(PredefinedToken):
+    let id = tab.getOrIncl $e
+    assert id == LitId(e)
+
+type
+  GeneratedCode* = object
+    code: seq[LitId]
+    tokens: BiTable[string]
+
+proc initGeneratedCode*(): GeneratedCode =
+  result = GeneratedCode(code: @[], tokens: initBiTable[string]())
+  fillTokenTable(result.tokens)
+
+proc add*(g: var GeneratedCode; t: PredefinedToken) {.inline.} =
+  g.code.add Token(t)
+
+proc add*(g: var GeneratedCode; s: string) {.inline.} =
+  g.code.add g.tokens.getOrIncl(s)
+

compiler/nir/nir.nim

@@ -0,0 +1,22 @@
+#
+#
+#           The Nim Compiler
+#        (c) Copyright 2023 Andreas Rumpf
+#
+#    See the file "copying.txt", included in this
+#    distribution, for details about the copyright.
+#
+
+## Nim Intermediate Representation, designed to capture all of Nim's semantics without losing too much
+## precious information. Can easily be translated into C. And to JavaScript, hopefully.
+
+import nirtypes, nirinsts
+
+type
+  Module* = object
+    types: TypeGraph
+    data: seq[Tree]
+    init: seq[Tree]
+    procs: seq[Tree]
+
+

compiler/nir/nirinsts.nim

@@ -0,0 +1,148 @@
+#
+#
+#           The Nim Compiler
+#        (c) Copyright 2023 Andreas Rumpf
+#
+#    See the file "copying.txt", included in this
+#    distribution, for details about the copyright.
+#
+
+## NIR instructions. Somewhat inspired by LLVM's instructions.
+
+import std / assertions
+import .. / ic / bitabs
+import nirlineinfos
+
+type
+  SymId* = distinct int
+  InstKind* = enum
+    Nop,
+    ImmediateVal,
+    IntVal,
+    StrVal,
+    SymDef,
+    SymUse,
+    ModuleId, # module ID
+    ModuleSymUse, # `module.x`
+    Label,
+    Goto,
+    GotoBack,
+    Typed,   # with type ID
+    NilVal,  # last atom
+
+    ArrayConstr,
+    ObjConstr,
+    Ret,
+    Yld,
+
+    Select,
+    SummonGlobal,
+    SummonThreadLocal,
+    Summon, # x = Summon Typed <Type ID>; x begins to live
+    Kill, # `Kill x`: scope end for `x`
+    Load,
+    Store,
+    ArrayAt, # addr(a[i])
+    FieldAt, # addr(obj.field)
+
+    Call,
+    IndirectCall,
+    CheckedCall, # call that can raise
+    CheckedIndirectCall, # call that can raise
+    CheckedAdd, # with overflow checking etc.
+    CheckedSub,
+    CheckedMul,
+    CheckedDiv,
+    CheckedMod,
+    Add,
+    Sub,
+    Mul,
+    Div,
+    Mod,
+    BitShl,
+    BitShr,
+    BitAnd,
+    BitOr,
+    BitXor,
+    BitNot,
+    Eq,
+    Le,
+    Lt,
+    Emit,
+    ProcDecl
+
+const
+  LastAtomicValue = NilVal
+
+  InstKindBits = 8'u32
+  InstKindMask = (1'u32 shl InstKindBits) - 1'u32
+
+type
+  Instr* = object     # 8 bytes
+    x: uint32
+    info: PackedLineInfo
+
+template kind*(n: Instr): InstKind = InstKind(n.x and InstKindMask)
+template operand(n: Instr): uint32 = (n.x shr InstKindBits)
+
+template toX(k: InstKind; operand: uint32): uint32 =
+  uint32(k) or (operand shl InstKindBits)
+
+template toX(k: InstKind; operand: LitId): uint32 =
+  uint32(k) or (operand.uint32 shl InstKindBits)
+
+type
+  Tree* = object
+    nodes: seq[Instr]
+
+  Values* = object
+    numbers: BiTable[int64]
+    strings: BiTable[string]
+
+type
+  PatchPos* = distinct int
+  NodePos* = distinct int
+
+const
+  InvalidPatchPos* = PatchPos(-1)
+
+proc isValid(p: PatchPos): bool {.inline.} = p.int != -1
+
+proc prepare(tree: var Tree; kind: InstKind): PatchPos =
+  result = PatchPos tree.nodes.len
+  tree.nodes.add Instr(x: toX(kind, 1'u32))
+
+proc isAtom(tree: Tree; pos: int): bool {.inline.} = tree.nodes[pos].kind <= LastAtomicValue
+proc isAtom(tree: Tree; pos: NodePos): bool {.inline.} = tree.nodes[pos.int].kind <= LastAtomicValue
+
+proc patch(tree: var Tree; pos: PatchPos) =
+  let pos = pos.int
+  let k = tree.nodes[pos].kind
+  assert k > LastAtomicValue
+  let distance = int32(tree.nodes.len - pos)
+  assert distance > 0
+  tree.nodes[pos].x = toX(k, cast[uint32](distance))
+
+proc len*(tree: Tree): int {.inline.} = tree.nodes.len
+
+template rawSpan(n: Instr): int = int(operand(n))
+
+proc nextChild(tree: Tree; pos: var int) {.inline.} =
+  if tree.nodes[pos].kind > LastAtomicValue:
+    assert tree.nodes[pos].operand > 0'u32
+    inc pos, tree.nodes[pos].rawSpan
+  else:
+    inc pos
+
+iterator sons*(tree: Tree; n: NodePos): NodePos =
+  var pos = n.int
+  assert tree.nodes[pos].kind > LastAtomicValue
+  let last = pos + tree.nodes[pos].rawSpan
+  inc pos
+  while pos < last:
+    yield NodePos pos
+    nextChild tree, pos
+
+template `[]`*(t: Tree; n: NodePos): Instr = t.nodes[n.int]
+
+

compiler/nir/nirlineinfos.nim

@@ -0,0 +1,78 @@
+#
+#
+#           The Nim Compiler
+#        (c) Copyright 2023 Andreas Rumpf
+#
+#    See the file "copying.txt", included in this
+#    distribution, for details about the copyright.
+#
+
+# For the line information we use 32 bits. They are used as follows:
+# Bit 0 (AsideBit): If we have inline line information or not. If not, the
+# remaining 31 bits are used as an index into a seq[(LitId, int, int)].
+#
+# We use 10 bits for the "file ID", this means a program can consist of as much
+# as 1024 different files. (If it uses more files than that, the overflow bit
+# would be set.)
+# This means we have 21 bits left to encode the (line, col) pair. We use 7 bits for the column
+# so 128 is the limit and 14 bits for the line number.
+# The packed representation supports files with up to 16384 lines.
+# Keep in mind that whenever any limit is reached the AsideBit is set and the real line
+# information is kept in a side channel.
+
+import std / assertions
+
+const
+  AsideBit = 1
+  FileBits = 10
+  LineBits = 14
+  ColBits = 7
+  FileMax = (1 shl FileBits) - 1
+  LineMax = (1 shl LineBits) - 1
+  ColMax = (1 shl ColBits) - 1
+
+static:
+  assert AsideBit + FileBits + LineBits + ColBits == 32
+
+import .. / ic / bitabs # for LitId
+
+type
+  PackedLineInfo* = distinct uint32
+
+  LineInfoManager* = object
+    aside*: seq[(LitId, int32, int32)]
+
+proc pack*(m: var LineInfoManager; file: LitId; line, col: int32): PackedLineInfo =
+  if file.uint32 <= FileMax.uint32 and line <= LineMax and col <= ColMax:
+    let col = if col < 0'i32: 0'u32 else: col.uint32
+    let line = if line < 0'i32: 0'u32 else: line.uint32
+    # use inline representation:
+    result = PackedLineInfo((file.uint32 shl 1'u32) or (line shl uint32(AsideBit + FileBits)) or
+      (col shl uint32(AsideBit + FileBits + LineBits)))
+  else:
+    result = PackedLineInfo((m.aside.len shl 1) or AsideBit)
+    m.aside.add (file, line, col)
+
+proc unpack*(m: LineInfoManager; i: PackedLineInfo): (LitId, int32, int32) =
+  let i = i.uint32
+  if (i and 1'u32) == 0'u32:
+    # inline representation:
+    result = (LitId((i shr 1'u32) and FileMax.uint32),
+      int32((i shr uint32(AsideBit + FileBits)) and LineMax.uint32),
+      int32((i shr uint32(AsideBit + FileBits + LineBits)) and ColMax.uint32))
+  else:
+    result = m.aside[int(i shr 1'u32)]
+
+proc getFileId*(m: LineInfoManager; i: PackedLineInfo): LitId =
+  result = unpack(m, i)[0]
+
+when isMainModule:
+  var m = LineInfoManager(aside: @[])
+  for i in 0'i32..<16388'i32:
+    for col in 0'i32..<100'i32:
+      let packed = pack(m, LitId(1023), i, col)
+      let u = unpack(m, packed)
+      assert u[0] == LitId(1023)
+      assert u[1] == i
+      assert u[2] == col
+  echo m.aside.len

compiler/nir/nirtypes.nim

@@ -0,0 +1,308 @@
+#
+#
+#           The Nim Compiler
+#        (c) Copyright 2023 Andreas Rumpf
+#
+#    See the file "copying.txt", included in this
+#    distribution, for details about the copyright.
+#
+
+## Type system for NIR. Close to C's type system but without its quirks.
+
+import std / assertions
+import .. / ic / bitabs
+
+type
+  NirTypeKind* = enum
+    VoidTy, IntTy, UIntTy, FloatTy, BoolTy, CharTy, NameVal, IntVal,
+    VarargsTy, # the `...` in a C prototype; also the last "atom"
+    APtrTy, # pointer to aliasable memory
+    UPtrTy, # pointer to unique/unaliasable memory
+    AArrayPtrTy, # pointer to array of aliasable memory
+    UArrayPtrTy, # pointer to array of unique/unaliasable memory
+    ArrayTy,
+    LastArrayTy, # array of unspecified size as a last field inside an object
+    ObjectTy,
+    UnionTy,
+    ProcTy,
+    ObjectDecl,
+    UnionDecl,
+    FieldDecl
+
+const
+  TypeKindBits = 8'u32
+  TypeKindMask = (1'u32 shl TypeKindBits) - 1'u32
+
+type
+  TypeNode* = object     # 4 bytes
+    x: uint32
+
+template kind*(n: TypeNode): NirTypeKind = NirTypeKind(n.x and TypeKindMask)
+template operand(n: TypeNode): uint32 = (n.x shr TypeKindBits)
+
+template toX(k: NirTypeKind; operand: uint32): uint32 =
+  uint32(k) or (operand shl TypeKindBits)
+
+template toX(k: NirTypeKind; operand: LitId): uint32 =
+  uint32(k) or (operand.uint32 shl TypeKindBits)
+
+type
+  TypeId* = distinct int
+  TypeGraph* = object
+    nodes: seq[TypeNode]
+    names: BiTable[string]
+    numbers: BiTable[uint64]
+
+const
+  VoidId* = TypeId 0
+  Bool8Id* = TypeId 1
+  Char8Id* = TypeId 2
+  Int8Id* = TypeId 3
+  Int16Id* = TypeId 4
+  Int32Id* = TypeId 5
+  Int64Id* = TypeId 6
+  UInt8Id* = TypeId 7
+  UInt16Id* = TypeId 8
+  UInt32Id* = TypeId 9
+  UInt64Id* = TypeId 10
+  Float32Id* = TypeId 11
+  Float64Id* = TypeId 12
+  LastBuiltinId* = 12
+
+proc initTypeGraph*(): TypeGraph =
+  result = TypeGraph(nodes: @[
+    TypeNode(x: toX(VoidTy, 0'u32)),
+    TypeNode(x: toX(BoolTy, 8'u32)),
+    TypeNode(x: toX(CharTy, 8'u32)),
+    TypeNode(x: toX(IntTy, 8'u32)),
+    TypeNode(x: toX(IntTy, 16'u32)),
+    TypeNode(x: toX(IntTy, 32'u32)),
+    TypeNode(x: toX(IntTy, 64'u32)),
+    TypeNode(x: toX(UIntTy, 8'u32)),
+    TypeNode(x: toX(UIntTy, 16'u32)),
+    TypeNode(x: toX(UIntTy, 32'u32)),
+    TypeNode(x: toX(UIntTy, 64'u32)),
+    TypeNode(x: toX(FloatTy, 32'u32)),
+    TypeNode(x: toX(FloatTy, 64'u32))
+  ])
+  assert result.nodes.len == LastBuiltinId+1
+
+type
+  TypePatchPos* = distinct int
+
+const
+  InvalidTypePatchPos* = TypePatchPos(-1)
+  LastAtomicValue = VarargsTy
+
+proc isValid(p: TypePatchPos): bool {.inline.} = p.int != -1
+
+proc prepare(tree: var TypeGraph; kind: NirTypeKind): TypePatchPos =
+  result = TypePatchPos tree.nodes.len
+  tree.nodes.add TypeNode(x: toX(kind, 1'u32))
+
+proc isAtom(tree: TypeGraph; pos: int): bool {.inline.} = tree.nodes[pos].kind <= LastAtomicValue
+proc isAtom(tree: TypeGraph; pos: TypeId): bool {.inline.} = tree.nodes[pos.int].kind <= LastAtomicValue
+
+proc patch(tree: var TypeGraph; pos: TypePatchPos) =
+  let pos = pos.int
+  let k = tree.nodes[pos].kind
+  assert k > LastAtomicValue
+  let distance = int32(tree.nodes.len - pos)
+  assert distance > 0
+  tree.nodes[pos].x = toX(k, cast[uint32](distance))
+
+proc len*(tree: TypeGraph): int {.inline.} = tree.nodes.len
+
+template rawSpan(n: TypeNode): int = int(operand(n))
+
+proc nextChild(tree: TypeGraph; pos: var int) {.inline.} =
+  if tree.nodes[pos].kind > LastAtomicValue:
+    assert tree.nodes[pos].operand > 0'u32
+    inc pos, tree.nodes[pos].rawSpan
+  else:
+    inc pos
+
+iterator sons*(tree: TypeGraph; n: TypeId): TypeId =
+  var pos = n.int
+  assert tree.nodes[pos].kind > LastAtomicValue
+  let last = pos + tree.nodes[pos].rawSpan
+  inc pos
+  while pos < last:
+    yield TypeId pos
+    nextChild tree, pos
+
+template `[]`*(t: TypeGraph; n: TypeId): TypeNode = t.nodes[n.int]
+
+proc elementType*(tree: TypeGraph; n: TypeId): TypeId {.inline.} =
+  assert tree[n].kind in {APtrTy, UPtrTy, AArrayPtrTy, UArrayPtrTy, ArrayTy, LastArrayTy}
+  result = TypeId(n.int+1)
+
+proc kind*(tree: TypeGraph; n: TypeId): NirTypeKind {.inline.} = tree[n].kind
+
+proc span(tree: TypeGraph; pos: int): int {.inline.} =
+  if tree.nodes[pos].kind <= LastAtomicValue: 1 else: int(tree.nodes[pos].operand)
+
+proc sons2(tree: TypeGraph; n: TypeId): (TypeId, TypeId) =
+  assert(not isAtom(tree, n.int))
+  let a = n.int+1
+  let b = a + span(tree, a)
+  result = (TypeId a, TypeId b)
+
+proc sons3(tree: TypeGraph; n: TypeId): (TypeId, TypeId, TypeId) =
+  assert(not isAtom(tree, n.int))
+  let a = n.int+1
+  let b = a + span(tree, a)
+  let c = b + span(tree, b)
+  result = (TypeId a, TypeId b, TypeId c)
+
+proc arrayLen*(tree: TypeGraph; n: TypeId): BiggestUInt =
+  assert tree[n].kind == ArrayTy
+  result = tree.numbers[LitId tree[n].operand]
+
+proc openType*(tree: var TypeGraph; kind: NirTypeKind): TypePatchPos =
+  assert kind in {APtrTy, UPtrTy, AArrayPtrTy, UArrayPtrTy,
+    ArrayTy, LastArrayTy, ProcTy, ObjectDecl, UnionDecl,
+    FieldDecl}
+  result = prepare(tree, kind)
+
+proc sealType*(tree: var TypeGraph; p: TypePatchPos): TypeId =
+  # TODO: Search for an existing instance of this type in
+  # order to reduce memory consumption.
+  result = TypeId(p)
+  patch tree, p
+
+proc nominalType*(tree: var TypeGraph; kind: NirTypeKind; name: string): TypeId =
+  assert kind in {ObjectTy, UnionTy}
+  result = TypeId tree.nodes.len
+  tree.nodes.add TypeNode(x: toX(kind, tree.names.getOrIncl(name)))
+
+proc addBuiltinType*(g: var TypeGraph; id: TypeId) =
+  g.nodes.add g[id]
+
+proc addType*(g: var TypeGraph; t: TypeId) =
+  let pos = t.int
+  let L = span(g, pos)
+  let d = g.nodes.len
+  g.nodes.setLen(d + L)
+  assert L > 0
+  for i in 0..<L:
+    g.nodes[d+i] = g.nodes[pos+i]
+
+proc addArrayLen*(g: var TypeGraph; len: uint64) =
+  g.nodes.add TypeNode(x: toX(IntVal, g.numbers.getOrIncl(len)))
+
+proc addName*(g: var TypeGraph; name: string) =
+  g.nodes.add TypeNode(x: toX(NameVal, g.names.getOrIncl(name)))
+
+proc addField*(g: var TypeGraph; name: string; typ: TypeId) =
+  let f = g.openType FieldDecl
+  g.addType typ
+  g.addName name
+  discard sealType(g, f)
+
+proc toString*(dest: var string; g: TypeGraph; i: TypeId) =
+  case g[i].kind
+  of VoidTy: dest.add "void"
+  of IntTy:
+    dest.add "i"
+    dest.addInt g[i].operand
+  of UIntTy:
+    dest.add "u"
+    dest.addInt g[i].operand
+  of FloatTy:
+    dest.add "f"
+    dest.addInt g[i].operand
+  of BoolTy:
+    dest.add "b"
+    dest.addInt g[i].operand
+  of CharTy:
+    dest.add "c"
+    dest.addInt g[i].operand
+  of NameVal:
+    dest.add g.names[LitId g[i].operand]
+  of IntVal:
+    dest.add $g.numbers[LitId g[i].operand]
+  of VarargsTy:
+    dest.add "..."
+  of APtrTy:
+    dest.add "aptr["
+    toString(dest, g, g.elementType(i))
+    dest.add "]"
+  of UPtrTy:
+    dest.add "uptr["
+    toString(dest, g, g.elementType(i))
+    dest.add "]"
+  of AArrayPtrTy:
+    dest.add "aArrayPtr["
+    toString(dest, g, g.elementType(i))
+    dest.add "]"
+  of UArrayPtrTy:
+    dest.add "uArrayPtr["
+    toString(dest, g, g.elementType(i))
+    dest.add "]"
+  of ArrayTy:
+    dest.add "Array["
+    let (elems, len) = g.sons2(i)
+    toString(dest, g, elems)
+    dest.add ", "
+    toString(dest, g, len)
+    dest.add "]"
+  of LastArrayTy:
+    # array of unspecified size as a last field inside an object
+    dest.add "LastArrayTy["
+    toString(dest, g, g.elementType(i))
+    dest.add "]"
+  of ObjectTy:
+    dest.add "object "
+    dest.add g.names[LitId g[i].operand]
+  of UnionTy:
+    dest.add "union "
+    dest.add g.names[LitId g[i].operand]
+  of ProcTy:
+    dest.add "proc["
+    for t in sons(g, i): toString(dest, g, t)
+    dest.add "]"
+  of ObjectDecl:
+    dest.add "object["
+    for t in sons(g, i):
+      toString(dest, g, t)
+      dest.add '\n'
+    dest.add "]"
+  of UnionDecl:
+    dest.add "union["
+    for t in sons(g, i):
+      toString(dest, g, t)
+      dest.add '\n'
+    dest.add "]"
+  of FieldDecl:
+    let (typ, name) = g.sons2(i)
+    toString(dest, g, typ)
+    dest.add ' '
+    toString(dest, g, name)
+
+proc toString*(dest: var string; g: TypeGraph) =
+  var i = 0
+  while i < g.len:
+    toString(dest, g, TypeId i)
+    dest.add '\n'
+    nextChild g, i
+
+proc `$`(g: TypeGraph): string =
+  result = ""
+  toString(result, g)
+
+when isMainModule:
+  var g = initTypeGraph()
+
+  let a = g.openType ArrayTy
+  g.addBuiltinType Int8Id
+  g.addArrayLen 5'u64
+  let finalArrayType = sealType(g, a)
+
+  let obj = g.openType ObjectDecl
+  g.nodes.add TypeNode(x: toX(NameVal, g.names.getOrIncl("MyType")))
+
+  g.addField "p", finalArrayType
+  discard sealType(g, obj)
+
+  echo g