Writing a basic x86-64 JIT compiler from scratch in stock Python

Shows how to JIT compile simple Python functions to native x86-64 machine code at runtime. Everything is done from scratch, using nothing but the built-in Python modules.

Transcript

1. Writing a basic x86-64 JIT
   compiler from scratch in
   stock Python
   Christian Stigen Larsen, 2018-01-23
   https://csl.name
   

   View Slide

2. View Slide

3. from jitcompiler import jit
   @jit
   def foo(a, b):
   return a*a - b*b
   

   View Slide

4. from jitcompiler import *
   @jit
   def foo(a, b):
   return a*a - b*b
   

   View Slide

5. from jitcompiler import *
   @jit
   def foo(a, b):
   return a*a - b*b
   

   View Slide

6. from jitcompiler import *
   @jit
   def foo(a, b):
   return a*a - b*b
   

   View Slide

7. from jitcompiler import *
   @jit
   def foo(a, b):
   return a*a - b*b
   >>> foo(2, 3)
   

   View Slide

8. from jitcompiler import *
   @jit
   def foo(a, b):
   return a*a - b*b
   >>> foo(2, 3)
   --- JIT-compiling
   

   View Slide

9. from jitcompiler import *
   @jit
   def foo(a, b):
   return a*a - b*b
   >>> foo(2, 3)
   --- JIT-compiling
   -5
   

   View Slide

10. from jitcompiler import *
    @jit
    def foo(a, b):
    return a*a - b*b
    >>> foo(2, 3)
    --- JIT-compiling
    -5
    >>> foo(3, 4)
    

    View Slide

11. from jitcompiler import *
    @jit
    def foo(a, b):
    return a*a - b*b
    >>> foo(2, 3)
    --- JIT-compiling
    -5
    >>> foo(3, 4)
    -7
    

    View Slide

12. View Slide

13. >>> print(disassemble(foo))
    0x100b1d000 48 89 fb mov rbx, rdi
    0x100b1d003 48 89 f8 mov rax, rdi
    0x100b1d006 48 0f af c3 imul rax, rbx
    0x100b1d00a 50 push rax
    0x100b1d00b 48 89 f3 mov rbx, rsi
    0x100b1d00e 48 89 f0 mov rax, rsi
    0x100b1d011 48 0f af c3 imul rax, rbx
    0x100b1d015 48 89 c3 mov rbx, rax
    0x100b1d018 58 pop rax
    0x100b1d019 48 29 d8 sub rax, rbx
    0x100b1d01c c3 ret
    

    View Slide

14. >>> print(disassemble(foo))
    0x100b1d000 48 89 fb mov rbx, rdi
    0x100b1d003 48 89 f8 mov rax, rdi
    0x100b1d006 48 0f af c3 imul rax, rbx
    0x100b1d00a 50 push rax
    0x100b1d00b 48 89 f3 mov rbx, rsi
    0x100b1d00e 48 89 f0 mov rax, rsi
    0x100b1d011 48 0f af c3 imul rax, rbx
    0x100b1d015 48 89 c3 mov rbx, rax
    0x100b1d018 58 pop rax
    0x100b1d019 48 29 d8 sub rax, rbx
    0x100b1d01c c3 ret
    

    View Slide

15. View Slide

16. Strategy
    Python Bound Name
    def foo
    

    View Slide

17. Python Function
    Strategy
    Bytecode
    Python Bound Name
    def foo
    

    View Slide

18. Python Function
    Strategy
    Python Bound Name
    Bytecode
    Bytecode
    def foo
    

    View Slide

19. Python Function
    Strategy
    Python Bound Name
    Bytecode
    Bytecode Bytecode
    def foo
    

    View Slide

20. Python Function
    Strategy
    Python Bound Name
    Bytecode
    Bytecode Bytecode
    IR
    def foo
    

    View Slide

21. Python Function
    Strategy
    Python Bound Name
    Bytecode
    Bytecode Bytecode
    IR
    Optimized IR
    def foo
    

    View Slide

22. Python Function
    Strategy
    Python Bound Name
    Bytecode
    Bytecode Bytecode
    IR
    Optimized IR
    x86-64 Machine Code
    def foo
    

    View Slide

23. Python Function
    Strategy
    Python Bound Name
    Memory Page
    Bytes …
    Read + Write
    Bytecode
    Bytecode Bytecode
    IR
    Optimized IR
    x86-64 Machine Code
    def foo
    

    View Slide

24. Python Function
    Strategy
    Python Bound Name
    Memory Page
    Bytes …
    Read + Write
    Bytecode
    x86-64 Machine Code
    def foo
    

    View Slide

25. Python Function
    Strategy
    Python Bound Name
    Memory Page
    Bytes …
    Read + Write
    Bytecode
    x86-64 Machine Code
    Read + Execute
    def foo
    

    View Slide

26. Strategy
    Python Bound Name
    Memory Page
    Bytes …
    Read + Write
    x86-64 Machine Code
    Read + Execute
    FFI
    def foo
    

    View Slide

27. Intermission: Why JIT?
    

    View Slide

28. Intermission: Why JIT?
    Because we can, and it’s awesome.
    

    View Slide

29. View Slide

30. Python Bytecode
    def foo(a, b):
    return a*a - b*b
    

    View Slide

31. Python Bytecode
    def foo(a, b):
    return a*a - b*b
    >>> import dis
    >>> dis.dis(foo)
    

    View Slide

32. Python Bytecode
    def foo(a, b):
    return a*a - b*b
    >>> import dis
    >>> dis.dis(foo)
    2 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    

    View Slide

33. Python Bytecode
    def foo(a, b):
    return a*a - b*b
    >>> import dis
    >>> dis.dis(foo)
    2 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    

    View Slide

34. Python Bytecode
    def foo(a, b):
    return a*a - b*b
    >>> import dis
    >>> dis.dis(foo)
    2 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    

    View Slide

35. Python Bytecode
    def foo(a, b):
    return a*a - b*b
    >>> import dis
    >>> dis.dis(foo)
    2 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    

    View Slide

36. Python Bytecode
    def foo(a, b):
    return a*a - b*b
    >>> import dis
    >>> dis.dis(foo)
    2 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    

    View Slide

37. Python Bytecode
    def foo(a, b):
    return a*a - b*b
    >>> import dis
    >>> dis.dis(foo)
    2 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    

    View Slide

38. Python Bytecode
    def foo(a, b):
    return a*a - b*b
    >>> import dis
    >>> dis.dis(foo)
    2 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    

    View Slide

39. Python Bytecode
    def foo(a, b):
    return a*a - b*b
    >>> import dis
    >>> dis.dis(foo)
    2 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    

    View Slide

40. 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    

    View Slide

41. 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    

    View Slide

42. 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    a
    

    View Slide

43. 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    a a
    

    View Slide

44. 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    a*a
    

    View Slide

45. 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    a*a b
    

    View Slide

46. 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    a*a b b
    

    View Slide

47. 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    a*a b*b
    

    View Slide

48. 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    a*a-b*b
    

    View Slide

49. 0 LOAD_FAST 0 (a)
    3 LOAD_FAST 0 (a)
    6 BINARY_MULTIPLY
    7 LOAD_FAST 1 (b)
    10 LOAD_FAST 1 (b)
    13 BINARY_MULTIPLY
    14 BINARY_SUBTRACT
    15 RETURN_VALUE
    a*a-b*b
    

    View Slide

50. 0 LOAD_FAST 0 (a)
    

    View Slide

51. 0 LOAD_FAST 0 (a)
    >>> foo.__code__.co_code
    ‘|\x00\x00|\x00\x00\x14|\x01\x00|\x01\x00\x14\x185’
    

    View Slide

52. 0 LOAD_FAST 0 (a)
    >>> foo.__code__.co_code
    ‘|\x00\x00|\x00\x00\x14|\x01\x00|\x01\x00\x14\x185’
    

    View Slide

53. 0 LOAD_FAST 0 (a)
    >>> foo.__code__.co_code
    ‘|\x00\x00|\x00\x00\x14|\x01\x00|\x01\x00\x14\x185’
    

    View Slide

54. 0 LOAD_FAST 0 (a)
    >>> foo.__code__.co_code
    ‘|\x00\x00|\x00\x00\x14|\x01\x00|\x01\x00\x14\x185’
    

    View Slide

55. 0 LOAD_FAST 0 (a)
    >>> foo.__code__.co_code
    ‘|\x00\x00|\x00\x00\x14|\x01\x00|\x01\x00\x14\x185’
    >>> bytecode = map(ord, foo.__code__.co_code)
    

    View Slide

56. 0 LOAD_FAST 0 (a)
    >>> foo.__code__.co_code
    ‘|\x00\x00|\x00\x00\x14|\x01\x00|\x01\x00\x14\x185’
    >>> bytecode = map(ord, foo.__code__.co_code)
    >>> opcode = bytecode[0]
    

    View Slide

57. 0 LOAD_FAST 0 (a)
    >>> foo.__code__.co_code
    ‘|\x00\x00|\x00\x00\x14|\x01\x00|\x01\x00\x14\x185’
    >>> bytecode = map(ord, foo.__code__.co_code)
    >>> opcode = bytecode[0]
    >>> opcode
    124
    

    View Slide

58. 0 LOAD_FAST 0 (a)
    >>> foo.__code__.co_code
    ‘|\x00\x00|\x00\x00\x14|\x01\x00|\x01\x00\x14\x185’
    >>> bytecode = map(ord, foo.__code__.co_code)
    >>> opcode = bytecode[0]
    >>> opcode
    124
    >>> dis.opname[opcode]
    

    View Slide

59. 0 LOAD_FAST 0 (a)
    >>> foo.__code__.co_code
    ‘|\x00\x00|\x00\x00\x14|\x01\x00|\x01\x00\x14\x185’
    >>> bytecode = map(ord, foo.__code__.co_code)
    >>> opcode = bytecode[0]
    >>> opcode
    124
    >>> dis.opname[opcode]
    ‘LOAD_FAST’
    

    View Slide

60. 0 LOAD_FAST 0 (a)
    >>> foo.__code__.co_code
    ‘|\x00\x00|\x00\x00\x14|\x01\x00|\x01\x00\x14\x185’
    >>> bytecode = map(ord, foo.__code__.co_code)
    >>> opcode = bytecode[0]
    >>> opcode
    124
    >>> dis.opname[opcode]
    ‘LOAD_FAST’
    >>> arg = bytecode[1] | bytecode[2] << 8
    

    View Slide

61. 0 LOAD_FAST 0 (a)
    >>> foo.__code__.co_code
    ‘|\x00\x00|\x00\x00\x14|\x01\x00|\x01\x00\x14\x185’
    >>> bytecode = map(ord, foo.__code__.co_code)
    >>> opcode = bytecode[0]
    >>> opcode
    124
    >>> dis.opname[opcode]
    ‘LOAD_FAST’
    >>> arg = bytecode[1] | bytecode[2] << 8
    >>> arg
    0
    

    View Slide

62. 0 LOAD_FAST 0 (a)
    >>> foo.__code__.co_code
    ‘|\x00\x00|\x00\x00\x14|\x01\x00|\x01\x00\x14\x185’
    >>> bytecode = map(ord, foo.__code__.co_code)
    >>> opcode = bytecode[0]
    >>> opcode
    124
    >>> dis.opname[opcode]
    ‘LOAD_FAST’
    >>> arg = bytecode[1] | bytecode[2] << 8
    >>> arg
    0
    >>> foo.__code__.co_varnames[arg]
    ‘a’
    

    View Slide

63. View Slide

64. >>> def bar(): return 123
    

    View Slide

65. >>> dis.dis(bar)
    1 0 LOAD_CONST 1 (123)
    3 RETURN_VALUE
    >>> def bar(): return 123
    

    View Slide

66. >>> dis.dis(bar)
    1 0 LOAD_CONST 1 (123)
    3 RETURN_VALUE
    >>> def bar(): return 123
    

    View Slide

67. >>> dis.dis(bar)
    1 0 LOAD_CONST 1 (123)
    3 RETURN_VALUE
    >>> def bar(): return 123
    >>> bar.__code__.co_consts[1]
    123
    

    View Slide

68. >>> dis.dis(bar)
    1 0 LOAD_CONST 1 (123)
    3 RETURN_VALUE
    >>> def bar(): return 123
    >>> bar.__code__.co_consts[1]
    123
    

    View Slide

69. View Slide

70. class Compiler(object):
    # …
    

    View Slide

71. class Compiler(object):
    # …
    def fetch(self):
    byte = self.bytecode[self.index]
    

    View Slide

72. class Compiler(object):
    # …
    def fetch(self):
    byte = self.bytecode[self.index]
    self.index += 1
    return byte
    

    View Slide

73. class Compiler(object):
    # …
    def fetch(self):
    byte = self.bytecode[self.index]
    self.index += 1
    return byte
    def decode(self):
    opcode = self.fetch()
    opname = dis.opname[opcode]
    

    View Slide

74. class Compiler(object):
    # …
    def fetch(self):
    byte = self.bytecode[self.index]
    self.index += 1
    return byte
    def decode(self):
    opcode = self.fetch()
    opname = dis.opname[opcode]
    if takes_arg(opname):
    arg = self.fetch() | self.fetch() << 8
    else:
    arg = None
    

    View Slide

75. class Compiler(object):
    # …
    def fetch(self):
    byte = self.bytecode[self.index]
    self.index += 1
    return byte
    def decode(self):
    opcode = self.fetch()
    opname = dis.opname[opcode]
    if takes_arg(opname):
    arg = self.fetch() | self.fetch() << 8
    else:
    arg = None
    return opname, arg
    

    View Slide

76. View Slide

77. def compile(self):
    while self.index < len(self.bytecode):
    op, arg = self.decode()
    if op == “LOAD_FAST”:
    yield “push”, self.variable(arg), None
    elif …
    

    View Slide

78. yield “push”, self.variable(arg), None
    

    View Slide

79. yield “push”, self.variable(arg), None
    

    View Slide

80. yield “push”, self.variable(arg), None
    IR & Assembly
    

    View Slide

81. yield “push”, self.variable(arg), None
    IR & Assembly
    rax
    rbx
    rcx
    …
    rsi
    rdi
    …
    Registers
    return address
    Stack
    

    View Slide

82. push
    yield “push”, self.variable(arg), None
    IR & Assembly
    rax
    rbx
    rcx
    …
    rsi
    rdi
    …
    Registers
    return address
    Stack
    

    View Slide

83. push
    yield “push”, self.variable(arg), None
    IR & Assembly
    rax
    rbx
    rcx
    …
    rsi
    rdi
    …
    Registers
    return address
    Stack
    

    View Slide

84. push
    yield “push”, self.variable(arg), None
    IR & Assembly
    rax
    rbx
    rcx
    …
    rsi
    rdi
    …
    Registers
    return address
    Stack
    

    View Slide

85. push
    yield “push”, self.variable(arg), None
    IR & Assembly
    rax
    rbx
    rcx
    …
    rsi
    rdi
    …
    Registers
    return address
    Stack
    

    View Slide

86. push
    yield “push”, self.variable(arg), None
    IR & Assembly
    rax
    rbx
    rcx
    …
    rsi
    rdi
    …
    Registers
    return address
    Stack
    def variable(self, index):
    passing_order = (“rdi”, “rsi”, “rdx”, “rcx”)
    return passing_order[index]
    

    View Slide

87. push
    yield “push”, self.variable(arg), None
    IR & Assembly
    rax
    rbx
    rcx
    …
    rsi
    rdi
    …
    Registers
    return address
    Stack
    

    View Slide

88. push
    yield “push”, self.variable(arg), None
    IR & Assembly
    rax
    rbx
    rcx
    …
    rsi
    rdi
    …
    Registers
    return address
    Stack
    rdi
    

    View Slide

89. View Slide

90. def baz(a, b): a = b
    

    View Slide

91. def baz(a, b): a = b
    LOAD_FAST 1 (b)
    STORE_FAST 0 (a)
    LOAD_CONST 0 (None)
    RETURN_VALUE
    

    View Slide

92. LOAD_FAST 1
    def baz(a, b): a = b
    LOAD_FAST 1 (b)
    STORE_FAST 0 (a)
    LOAD_CONST 0 (None)
    RETURN_VALUE
    

    View Slide

93. LOAD_FAST 1 push rsi
    def baz(a, b): a = b
    LOAD_FAST 1 (b)
    STORE_FAST 0 (a)
    LOAD_CONST 0 (None)
    RETURN_VALUE
    

    View Slide

94. LOAD_FAST 1 push rsi
    STORE_FAST 0
    def baz(a, b): a = b
    LOAD_FAST 1 (b)
    STORE_FAST 0 (a)
    LOAD_CONST 0 (None)
    RETURN_VALUE
    

    View Slide

95. LOAD_FAST 1 push rsi
    STORE_FAST 0 pop rax
    mov rdi, rax
    def baz(a, b): a = b
    LOAD_FAST 1 (b)
    STORE_FAST 0 (a)
    LOAD_CONST 0 (None)
    RETURN_VALUE
    

    View Slide

96. LOAD_FAST 1 push rsi
    STORE_FAST 0 pop rax
    mov rdi, rax
    def baz(a, b): a = b
    LOAD_FAST 1 (b)
    STORE_FAST 0 (a)
    LOAD_CONST 0 (None)
    RETURN_VALUE
    LOAD_CONST 0
    

    View Slide

97. LOAD_FAST 1 push rsi
    STORE_FAST 0 pop rax
    mov rdi, rax
    def baz(a, b): a = b
    LOAD_FAST 1 (b)
    STORE_FAST 0 (a)
    LOAD_CONST 0 (None)
    RETURN_VALUE
    LOAD_CONST 0 imm rax, 0
    push rax
    

    View Slide

98. LOAD_FAST 1 push rsi
    STORE_FAST 0 pop rax
    mov rdi, rax
    def baz(a, b): a = b
    LOAD_FAST 1 (b)
    STORE_FAST 0 (a)
    LOAD_CONST 0 (None)
    RETURN_VALUE
    LOAD_CONST 0 imm rax, 0
    push rax
    RETURN_VALUE
    

    View Slide

99. LOAD_FAST 1 push rsi
    STORE_FAST 0 pop rax
    mov rdi, rax
    def baz(a, b): a = b
    LOAD_FAST 1 (b)
    STORE_FAST 0 (a)
    LOAD_CONST 0 (None)
    RETURN_VALUE
    LOAD_CONST 0 imm rax, 0
    push rax
    RETURN_VALUE pop rax
    ret
    

    View Slide

100. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     

     View Slide

101. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     

     View Slide

102. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax
     rsi
     rdi
     return address
     

     View Slide

103. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax
     rsi 22
     rdi 11
     return address
     

     View Slide

104. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax
     rsi 22
     rdi 11
     return address
     

     View Slide

105. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax
     rsi 22
     rdi 11
     return address
     

     View Slide

106. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax
     rsi 22
     rdi 11
     return address
     22
     

     View Slide

107. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax
     rsi 22
     rdi 11
     return address
     22
     

     View Slide

108. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 22
     rsi 22
     rdi 11
     return address
     

     View Slide

109. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 22
     rsi 22
     rdi 11
     return address
     

     View Slide

110. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 22
     rsi 22
     rdi 22
     return address
     

     View Slide

111. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 22
     rsi 22
     rdi 22
     return address
     

     View Slide

112. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     return address
     

     View Slide

113. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     return address
     

     View Slide

114. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     return address
     0
     

     View Slide

115. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     return address
     0
     

     View Slide

116. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     return address
     

     View Slide

117. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     return address
     

     View Slide

118. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     

     View Slide

119. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     

     View Slide

120. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     LOAD_FAST push
     STORE_FAST
     pop rax
     mov , rax
     LOAD_CONST
     imm rax
     push rax
     BINARY_MULTIPLY
     pop rax
     pop rbx
     imul rax, rbx
     push rax
     BINARY_ADD /
     INPLACE_ADD
     pop rax
     pop rbx
     add rax, rbx
     push rax
     BINARY_SUBTRACT /
     INPLACE_SUBTRACT
     pop rbx
     pop rax
     sub rax, rbx
     push rax
     UNARY_NEGATIVE
     pop rax
     neg rax
     push rax
     RETURN_VALUE
     pop rax
     ret
     

     View Slide

121. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     LOAD_FAST push
     STORE_FAST
     pop rax
     mov , rax
     LOAD_CONST
     imm rax
     push rax
     BINARY_MULTIPLY
     pop rax
     pop rbx
     imul rax, rbx
     push rax
     BINARY_ADD /
     INPLACE_ADD
     pop rax
     pop rbx
     add rax, rbx
     push rax
     BINARY_SUBTRACT /
     INPLACE_SUBTRACT
     pop rbx
     pop rax
     sub rax, rbx
     push rax
     UNARY_NEGATIVE
     pop rax
     neg rax
     push rax
     RETURN_VALUE
     pop rax
     ret
     

     View Slide

122. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     LOAD_FAST push
     STORE_FAST
     pop rax
     mov , rax
     LOAD_CONST
     imm rax
     push rax
     BINARY_MULTIPLY
     pop rax
     pop rbx
     imul rax, rbx
     push rax
     BINARY_ADD /
     INPLACE_ADD
     pop rax
     pop rbx
     add rax, rbx
     push rax
     BINARY_SUBTRACT /
     INPLACE_SUBTRACT
     pop rbx
     pop rax
     sub rax, rbx
     push rax
     UNARY_NEGATIVE
     pop rax
     neg rax
     push rax
     RETURN_VALUE
     pop rax
     ret
     

     View Slide

123. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     LOAD_FAST push
     STORE_FAST
     pop rax
     mov , rax
     LOAD_CONST
     imm rax
     push rax
     BINARY_MULTIPLY
     pop rax
     pop rbx
     imul rax, rbx
     push rax
     BINARY_ADD /
     INPLACE_ADD
     pop rax
     pop rbx
     add rax, rbx
     push rax
     BINARY_SUBTRACT /
     INPLACE_SUBTRACT
     pop rbx
     pop rax
     sub rax, rbx
     push rax
     UNARY_NEGATIVE
     pop rax
     neg rax
     push rax
     RETURN_VALUE
     pop rax
     ret
     

     View Slide

124. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     LOAD_FAST push
     STORE_FAST
     pop rax
     mov , rax
     LOAD_CONST
     imm rax
     push rax
     BINARY_MULTIPLY
     pop rax
     pop rbx
     imul rax, rbx
     push rax
     BINARY_ADD /
     INPLACE_ADD
     pop rax
     pop rbx
     add rax, rbx
     push rax
     BINARY_SUBTRACT /
     INPLACE_SUBTRACT
     pop rbx
     pop rax
     sub rax, rbx
     push rax
     UNARY_NEGATIVE
     pop rax
     neg rax
     push rax
     RETURN_VALUE
     pop rax
     ret
     

     View Slide

125. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     LOAD_FAST push
     STORE_FAST
     pop rax
     mov , rax
     LOAD_CONST
     imm rax
     push rax
     BINARY_MULTIPLY
     pop rax
     pop rbx
     imul rax, rbx
     push rax
     BINARY_ADD /
     INPLACE_ADD
     pop rax
     pop rbx
     add rax, rbx
     push rax
     BINARY_SUBTRACT /
     INPLACE_SUBTRACT
     pop rbx
     pop rax
     sub rax, rbx
     push rax
     UNARY_NEGATIVE
     pop rax
     neg rax
     push rax
     RETURN_VALUE
     pop rax
     ret
     

     View Slide

126. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     LOAD_FAST push
     STORE_FAST
     pop rax
     mov , rax
     LOAD_CONST
     imm rax
     push rax
     BINARY_MULTIPLY
     pop rax
     pop rbx
     imul rax, rbx
     push rax
     BINARY_ADD /
     INPLACE_ADD
     pop rax
     pop rbx
     add rax, rbx
     push rax
     BINARY_SUBTRACT /
     INPLACE_SUBTRACT
     pop rbx
     pop rax
     sub rax, rbx
     push rax
     UNARY_NEGATIVE
     pop rax
     neg rax
     push rax
     RETURN_VALUE
     pop rax
     ret
     

     View Slide

127. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     LOAD_FAST push
     STORE_FAST
     pop rax
     mov , rax
     LOAD_CONST
     imm rax
     push rax
     BINARY_MULTIPLY
     pop rax
     pop rbx
     imul rax, rbx
     push rax
     BINARY_ADD /
     INPLACE_ADD
     pop rax
     pop rbx
     add rax, rbx
     push rax
     BINARY_SUBTRACT /
     INPLACE_SUBTRACT
     pop rbx
     pop rax
     sub rax, rbx
     push rax
     UNARY_NEGATIVE
     pop rax
     neg rax
     push rax
     RETURN_VALUE
     pop rax
     ret
     

     View Slide

128. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     LOAD_FAST push
     STORE_FAST
     pop rax
     mov , rax
     LOAD_CONST
     imm rax
     push rax
     BINARY_MULTIPLY
     pop rax
     pop rbx
     imul rax, rbx
     push rax
     BINARY_ADD /
     INPLACE_ADD
     pop rax
     pop rbx
     add rax, rbx
     push rax
     BINARY_SUBTRACT /
     INPLACE_SUBTRACT
     pop rbx
     pop rax
     sub rax, rbx
     push rax
     UNARY_NEGATIVE
     pop rax
     neg rax
     push rax
     RETURN_VALUE
     pop rax
     ret
     

     View Slide

129. push rsi
     pop rax
     mov rdi, rax
     def baz(a, b): a = b
     imm rax, 0
     push rax
     pop rax
     ret
     >>> baz(11, 22)
     rax 0
     rsi 22
     rdi 22
     

     View Slide

130. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     

     View Slide

131. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     

     View Slide

132. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     

     View Slide

133. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     

     View Slide

134. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     mov rax, rsi
     

     View Slide

135. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     mov rax, rsi
     

     View Slide

136. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     mov rax, rsi
     Peephole Optimization
     

     View Slide

137. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     mov rax, rsi
     Peephole Optimization
     

     View Slide

138. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     mov rax, rsi
     mov rdi, rax
     Peephole Optimization
     

     View Slide

139. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     mov rax, rsi
     mov rdi, rax
     Peephole Optimization
     

     View Slide

140. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     mov rax, rsi
     mov rdi, rax
     imm rax, 0
     Peephole Optimization
     

     View Slide

141. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     mov rax, rsi
     mov rdi, rax
     imm rax, 0
     Peephole Optimization
     

     View Slide

142. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     mov rax, rsi
     mov rdi, rax
     imm rax, 0
     Peephole Optimization
     

     View Slide

143. push rsi
     pop rax
     mov rdi, rax
     imm rax, 0
     push rax
     pop rax
     ret
     mov rax, rsi
     mov rdi, rax
     imm rax, 0
     ret
     Peephole Optimization
     

     View Slide

144. mov rax, rsi
     mov rdi, rax
     imm rax, 0
     ret
     Peephole Optimization
     

     View Slide

145. mov rdi, rsi
     imm rax, 0
     ret
     Peephole Optimization
     

     View Slide

146. mov rdi, rsi
     imm rax, 0
     ret
     Peephole Optimization
     

     View Slide

147. mov rdi, rsi
     imm rax, 0
     ret
     Peephole Optimization
     

     View Slide

148. View Slide

149. class Assembler(object):
     # …
     def emit(self, *args):
     for code in args:
     self.block[self.index] = code
     self.index += 1
     Machine Code Generation
     

     View Slide

150. class Assembler(object):
     # …
     def emit(self, *args):
     for code in args:
     self.block[self.index] = code
     self.index += 1
     Machine Code Generation
     def push(self, a, dummy):
     self.emit(0x50 | self.registers(a))
     

     View Slide

151. class Assembler(object):
     # …
     def emit(self, *args):
     for code in args:
     self.block[self.index] = code
     self.index += 1
     Machine Code Generation
     def push(self, a, dummy):
     self.emit(0x50 | self.registers(a))
     

     View Slide

152. class Assembler(object):
     # …
     def emit(self, *args):
     for code in args:
     self.block[self.index] = code
     self.index += 1
     Machine Code Generation
     def push(self, a, dummy):
     self.emit(0x50 | self.registers(a))
     def registers(self, a):
     order = (“rax”, “rcx”, …)
     return order.index(a)
     

     View Slide

153. class Assembler(object):
     # …
     def emit(self, *args):
     for code in args:
     self.block[self.index] = code
     self.index += 1
     Machine Code Generation
     def push(self, a, dummy):
     self.emit(0x50 | self.registers(a))
     def registers(self, a):
     order = (“rax”, “rcx”, …)
     return order.index(a)
     

     View Slide

154. Machine Code Generation
     

     View Slide

155. def add(self, a, b):
     self.emit(0x48,
     0x01,
     0xc0 | self.registers(a, b))
     def immediate(self, a, constant):
     self.emit(0x48,
     0xb8 | self.registers(a),
     *self.little_endian(constant))
     Machine Code Generation
     

     View Slide

156. def add(self, a, b):
     self.emit(0x48,
     0x01,
     0xc0 | self.registers(a, b))
     def immediate(self, a, constant):
     self.emit(0x48,
     0xb8 | self.registers(a),
     *self.little_endian(constant))
     Machine Code Generation
     

     View Slide

157. def add(self, a, b):
     self.emit(0x48,
     0x01,
     0xc0 | self.registers(a, b))
     def immediate(self, a, constant):
     self.emit(0x48,
     0xb8 | self.registers(a),
     *self.little_endian(constant))
     Machine Code Generation
     

     View Slide

158. def add(self, a, b):
     self.emit(0x48,
     0x01,
     0xc0 | self.registers(a, b))
     def immediate(self, a, constant):
     self.emit(0x48,
     0xb8 | self.registers(a),
     *self.little_endian(constant))
     Machine Code Generation
     

     View Slide

159. def add(self, a, b):
     self.emit(0x48,
     0x01,
     0xc0 | self.registers(a, b))
     def immediate(self, a, constant):
     self.emit(0x48,
     0xb8 | self.registers(a),
     *self.little_endian(constant))
     Machine Code Generation
     

     View Slide

160. Machine Code Generation
     

     View Slide

161. def ret(self, a, b):
     self.emit(0xc3)
     def push(self, a, _):
     self.emit(0x50 | self.registers(a))
     def pop(self, a, _):
     self.emit(0x58 | self.registers(a))
     def imul(self, a, b):
     self.emit(0x48, 0x0f, 0xaf, 0xc0 | self.registers(a, b))
     def add(self, a, b):
     self.emit(0x48, 0x01, 0xc0 | self.registers(b, a))
     def sub(self, a, b):
     self.emit(0x48, 0x29, 0xc0 | self.registers(b, a))
     def neg(self, a, _):
     self.emit(0x48, 0xf7, 0xd8 | self.register(a))
     def mov(self, a, b):
     self.emit(0x48, 0x89, 0xc0 | self.registers(b, a))
     def immediate(self, a, number):
     self.emit(0x48, 0xb8 | self.registers(a), *self.little_endian(number))
     Machine Code Generation
     

     View Slide

162. View Slide

163. Memory Management
     

     View Slide

164. Memory Management
     • Allocate one page of memory using mmap
     

     View Slide

165. Memory Management
     • Allocate one page of memory using mmap
     • Use mprotect to change from R+W to R+X
     

     View Slide

166. Memory Management
     

     View Slide

167. Memory Management
     def create_block(size):
     ptr = mmap(0, size, MMAP.PROT_WRIT
     | MMAP.PROT_READ
     | MMAP.MAP_ANONYMOUS, 0, 0)
     if ptr == MAP_FAILED:
     raise RuntimeError(…)
     return ptr
     

     View Slide

168. Memory Management
     def create_block(size):
     ptr = mmap(0, size, MMAP.PROT_WRIT
     | MMAP.PROT_READ
     | MMAP.MAP_ANONYMOUS, 0, 0)
     if ptr == MAP_FAILED:
     raise RuntimeError(…)
     return ptr
     

     View Slide

169. Memory Management
     def create_block(size):
     ptr = mmap(0, size, MMAP.PROT_WRIT
     | MMAP.PROT_READ
     | MMAP.MAP_ANONYMOUS, 0, 0)
     if ptr == MAP_FAILED:
     raise RuntimeError(…)
     return ptr
     

     View Slide

170. Memory Management
     def create_block(size):
     ptr = mmap(0, size, MMAP.PROT_WRIT
     | MMAP.PROT_READ
     | MMAP.MAP_ANONYMOUS, 0, 0)
     if ptr == MAP_FAILED:
     raise RuntimeError(…)
     return ptr
     

     View Slide

171. Memory Management
     

     View Slide

172. Memory Management
     def make_executable(block, size):
     if mprotect(block, size, MMAP.PROT_READ
     | MMAP.PROT_EXEC) != 0:
     raise RuntimeError(…)
     

     View Slide

173. Memory Management
     def make_executable(block, size):
     if mprotect(block, size, MMAP.PROT_READ
     | MMAP.PROT_EXEC) != 0:
     raise RuntimeError(…)
     

     View Slide

174. View Slide

175. FFI
     

     View Slide

176. FFI
     • Use ctypes as FFI
     

     View Slide

177. FFI
     • Use ctypes as FFI
     • Grab machine code address, create signature and bind to
     function.
     

     View Slide

178. FFI
     • Use ctypes as FFI
     • Grab machine code address, create signature and bind to
     function.
     • Ctypes also seem to save and restore registers
     

     View Slide

179. FFI
     • Use ctypes as FFI
     • Grab machine code address, create signature and bind to
     function.
     • Ctypes also seem to save and restore registers
     • More overhead: Marshalling arguments
     

     View Slide

180. FFI
     

     View Slide

181. FFI
     signature = ctypes.CFUNCTYPE([ctypes.c_int64,
     ctypes.c_int64])
     signature.restype = ctypes.c_int64
     func = signature(assembler.address)
     print(func(1, 2))
     

     View Slide

182. FFI
     signature = ctypes.CFUNCTYPE([ctypes.c_int64,
     ctypes.c_int64])
     signature.restype = ctypes.c_int64
     func = signature(assembler.address)
     print(func(1, 2))
     

     View Slide

183. FFI
     signature = ctypes.CFUNCTYPE([ctypes.c_int64,
     ctypes.c_int64])
     signature.restype = ctypes.c_int64
     func = signature(assembler.address)
     print(func(1, 2))
     

     View Slide

184. FFI
     signature = ctypes.CFUNCTYPE([ctypes.c_int64,
     ctypes.c_int64])
     signature.restype = ctypes.c_int64
     func = signature(assembler.address)
     print(func(1, 2))
     

     View Slide

185. FFI
     signature = ctypes.CFUNCTYPE([ctypes.c_int64,
     ctypes.c_int64])
     signature.restype = ctypes.c_int64
     func = signature(assembler.address)
     print(func(1, 2))
     

     View Slide

186. View Slide

187. The “@jit” Decorator
     def jit(func):
     def front(*args, **kw):
     if not hasattr(front, “func”):
     front.func = compile_native(func)
     return front.func(*args, **kw)
     return front
     @jit
     def foo(a, b):
     return a*a - b*b
     

     View Slide

188. The “@jit” Decorator
     def jit(func):
     def front(*args, **kw):
     if not hasattr(front, “func”):
     front.func = compile_native(func)
     return front.func(*args, **kw)
     return front
     @jit
     def foo(a, b):
     return a*a - b*b
     

     View Slide

189. Demo
     

     View Slide

190. Things we haven’t touched on
     

     View Slide

191. Things we haven’t touched on
     • Code generation: SSA, branching
     

     View Slide

192. Things we haven’t touched on
     • Code generation: SSA, branching
     • Optimization: TAC, register allocation
     

     View Slide

193. Things we haven’t touched on
     • Code generation: SSA, branching
     • Optimization: TAC, register allocation
     • Compatibility: Python object system
     

     View Slide

194. Things we haven’t touched on
     • Code generation: SSA, branching
     • Optimization: TAC, register allocation
     • Compatibility: Python object system
     • Performance gotchas
     

     View Slide

195. Pointers
     • Full details on blog post at

     https://csl.name/post/python-compiler/
     • PeachPy
     • Truffle + Graal, LLVM, libjit, etc.
     • NUMBA
     

     View Slide

196. Thanks y’all!
     

     View Slide