Call Arguments

First, the CALL_FUNCTION(), CALL_FUNCTION_VAR(), CALL_FUNCTION_KW(), and CALL_FUNCTION_VAR_KW() methods all take two arguments, both of which are optional. (The _VAR and _KW suffixes in the method names indicate whether or not a *args or **kwargs or both are also present on the stack, in addition to the explicit positional and keyword arguments.)

The first argument of each of these methods, is the number of positional arguments on the stack, and the second is the number of keyword/value pairs on the stack (to be used as keyword arguments). Both default to zero if not supplied:

>>> c = Code()
>>> c.LOAD_CONST(type)
>>> c.LOAD_CONST(27)
>>> c.CALL_FUNCTION(1)      # 1 positional, no keywords
>>> c.RETURN_VALUE()

>>> eval(c.code())          # computes type(27)
<type 'int'>

>>> c = Code()
>>> c.LOAD_CONST(dict)
>>> c.LOAD_CONST('x')
>>> c.LOAD_CONST(42)
>>> c.CALL_FUNCTION(0,1)    # no positional, 1 keyword
>>> c.RETURN_VALUE()

>>> eval(c.code())          # computes dict(x=42)
{'x': 42}

Jump Targets

Opcodes that perform jumps or refer to addresses can be invoked in one of two ways. First, if you are jumping backwards (e.g. with JUMP_ABSOLUTE or CONTINUE_LOOP), you can obtain the target bytecode offset using the .here() method, and then later pass that offset into the appropriate method:

>>> c = Code()
>>> c.LOAD_CONST(42)
>>> where = c.here()         # get a location near the start of the code
>>> c.DUP_TOP()
>>> c.POP_TOP()
>>> c.JUMP_ABSOLUTE(where)   # now jump back to it

>>> dump(c.code())
                LOAD_CONST               1 (42)
        L1:     DUP_TOP
                POP_TOP
                JUMP_ABSOLUTE            L1

But if you are jumping forward, you will need to call the jump or setup method without any arguments. The return value will be a "forward reference" object that can be called later to indicate that the desired jump target has been reached:

>>> c = Code()
>>> c.LOAD_CONST(99)
>>> forward = c.JUMP_IF_TRUE() # create a jump and a forward reference

>>> c.LOAD_CONST(42)            # this is what we want to skip over
>>> c.POP_TOP()

>>> forward()   # calling the reference changes the jump to point here
>>> c.LOAD_CONST(23)
>>> c.RETURN_VALUE()

>>> dump(c.code())
                LOAD_CONST               1 (99)
                JUMP_IF_TRUE             L1
                LOAD_CONST               2 (42)
                POP_TOP
        L1:     LOAD_CONST               3 (23)
                RETURN_VALUE

>>> eval(c.code())
23

Other Special Opcodes

The MAKE_CLOSURE method takes an argument for the number of default values on the stack, just like the "real" Python opcode. However, it also has an an additional required argument: the number of closure cells on the stack. The Python interpreter normally gets this number from a code object that's on the stack, but Code objects need this value in order to update the current stack size, for purposes of computing the required total stack size:

>>> def x(a,b):     # a simple closure example
...     def y():
...         return a+b
...     return y

>>> c = Code()
>>> c.co_cellvars = ('a','b')

>>> import sys
>>> c.LOAD_CLOSURE('a')
>>> c.LOAD_CLOSURE('b')
>>> if sys.version>='2.5':
...     c.BUILD_TUPLE(2) # In Python 2.5+, free vars must be in a tuple
>>> c.LOAD_CONST(None)   # in real code, this'd be a Python code constant
>>> c.MAKE_CLOSURE(0,2)  # no defaults, 2 free vars in the new function

>>> c.stack_size         # This will be 1, no matter what Python version
1

The COMPARE_OP method takes an argument which can be a valid comparison integer constant, or a string containing a Python operator, e.g.:

>>> c = Code()
>>> c.LOAD_CONST(1)
>>> c.LOAD_CONST(2)
>>> c.COMPARE_OP('not in')
>>> dis(c.code())
  0           0 LOAD_CONST               1 (1)
              3 LOAD_CONST               2 (2)
              6 COMPARE_OP               7 (not in)

The full list of valid operator strings can be found in the standard library's opcode module. "<>" is also accepted as an alias for "!=":

>>> c.LOAD_CONST(3)
>>> c.COMPARE_OP('<>')
>>> dis(c.code())
  0           0 LOAD_CONST               1 (1)
              3 LOAD_CONST               2 (2)
              6 COMPARE_OP               7 (not in)
              9 LOAD_CONST               3 (3)
             12 COMPARE_OP               3 (!=)

Simple Constants

If an argument is an integer, long, float, complex, string, unicode, boolean, None, or Python code object, it is treated as though it was passed to the LOAD_CONST method directly:

>>> c = Code()
>>> c(1, 2L, 3.0, 4j+5, "6", u"7", False, None, c.code())
>>> dis(c.code())
  0           0 LOAD_CONST               1 (1)
              3 LOAD_CONST               2 (2L)
              6 LOAD_CONST               3 (3.0)
              9 LOAD_CONST               4 ((5+4j))
             12 LOAD_CONST               5 ('6')
             15 LOAD_CONST               6 (u'7')
             18 LOAD_CONST               7 (False)
             21 LOAD_CONST               0 (None)
             24 LOAD_CONST               8 (<code object <lambda> at ...>)

Note that although some values of different types may compare equal to each other, Code objects will not substitute a value of a different type than the one you requested:

>>> c = Code()
>>> c(1, True, 1.0, 1L)     # equal, but different types
>>> dis(c.code())
  0           0 LOAD_CONST               1 (1)
              3 LOAD_CONST               2 (True)
              6 LOAD_CONST               3 (1.0)
              9 LOAD_CONST               4 (1L)

Simple Containers

If an argument is a tuple, list, or dictionary, code is generated to reconstruct the given data, recursively:

>>> c = Code()
>>> c({1:(2,"3"), 4:[5,6]})
>>> dis(c.code())
  0           0 BUILD_MAP                0
              3 DUP_TOP
              4 LOAD_CONST               1 (1)
              7 LOAD_CONST               2 (2)
             10 LOAD_CONST               3 ('3')
             13 BUILD_TUPLE              2
             16 ROT_THREE
             17 STORE_SUBSCR
             18 DUP_TOP
             19 LOAD_CONST               4 (4)
             22 LOAD_CONST               5 (5)
             25 LOAD_CONST               6 (6)
             28 BUILD_LIST               2
             31 ROT_THREE
             32 STORE_SUBSCR

Arbitrary Constants

The Const wrapper allows you to treat any object as a literal constant, regardless of its type:

>>> from peak.util.assembler import Const

>>> c = Code()
>>> c( Const( (1,2,3) ) )
>>> dis(c.code())
  0           0 LOAD_CONST               1 ((1, 2, 3))

As you can see, the above creates code that references an actual tuple as a constant, rather than generating code to recreate the tuple using a series of LOAD_CONST operations followed by a BUILD_TUPLE.

If the value wrapped in a Const is not hashable, it is compared by identity rather than value. This prevents equal mutable values from being reused by accident, e.g. if you plan to mutate the "constant" values later:

>>> c = Code()
>>> c(Const([]), Const([]))     # equal, but not the same object!
>>> dis(c.code())
  0           0 LOAD_CONST               1 ([])
              3 LOAD_CONST               2 ([])

Thus, although Const objects hash and compare based on equality for hashable types:

>>> hash(Const(3)) == hash(3)
True
>>> Const(3)==Const(3)
True

They hash and compare based on object identity for non-hashable types:

>>> c = Const([])
>>> hash(c) == hash(id(c.value))
True
>>> c == Const(c.value)     # compares equal if same object
True
>>> c == Const([])          # but is not equal to a merely equal object
False

Suite and Pass

On occasion, it's helpful to be able to group a sequence of opcodes, expressions, or statements together, to be passed as an argument to other node types. The Suite node type accomplishes this:

>>> from peak.util.assembler import Suite, Pass

>>> c = Code()
>>> c.return_(Suite([Const(42), Code.DUP_TOP, Code.POP_TOP]))
>>> dis(c.code())
  0           0 LOAD_CONST               1 (42)
              3 DUP_TOP
              4 POP_TOP
              5 RETURN_VALUE    

And Pass is a shortcut for an empty Suite, that generates nothing:

>>> Suite([])
Pass

>>> c = Code()
>>> c(Pass)
>>> c.return_(None)
>>> dis(c.code())
  0           0 LOAD_CONST               0 (None)
              3 RETURN_VALUE    

Local and Global Names

The Local and Global wrappers take a name, and load either a local or global variable, respectively:

>>> from peak.util.assembler import Global, Local

>>> c = Code()
>>> c( Local('x'), Global('y') )
>>> dis(c.code())
  0           0 LOAD_FAST                0 (x)
              3 LOAD_GLOBAL              0 (y)

As with simple constants and Const wrappers, these objects can be used to construct more complex expressions, like {a:(b,c)}:

>>> c = Code()
>>> c( {Local('a'): (Local('b'), Local('c'))} )
>>> dis(c.code())
  0           0 BUILD_MAP                0
              3 DUP_TOP
              4 LOAD_FAST                0 (a)
              7 LOAD_FAST                1 (b)
             10 LOAD_FAST                2 (c)
             13 BUILD_TUPLE              2
             16 ROT_THREE
             17 STORE_SUBSCR

The LocalAssign node type takes a name, and stores a value in a local variable:

>>> from peak.util.assembler import LocalAssign
>>> c = Code()
>>> c(42, LocalAssign('x'))
>>> dis(c.code())
  0           0 LOAD_CONST               1 (42)
              3 STORE_FAST               0 (x)

If the code object is not using "fast locals" (i.e. CO_OPTIMIZED isn't set), local variables will be referenced using LOAD_NAME and STORE_NAME instead of LOAD_FAST and STORE_FAST, and if the referenced local name is a "cell" or "free" variable, LOAD_DEREF and STORE_DEREF are used instead:

>>> from peak.util.assembler import CO_OPTIMIZED
>>> c = Code()
>>> c.co_flags &= ~CO_OPTIMIZED
>>> c.co_cellvars = ('y',)
>>> c.co_freevars = ('z',)
>>> c( Local('x'), Local('y'), Local('z') )
>>> c( LocalAssign('x'), LocalAssign('y'), LocalAssign('z') )
>>> dis(c.code())
  0           0 LOAD_NAME                0 (x)
              3 LOAD_DEREF               0 (y)
              6 LOAD_DEREF               1 (z)
              9 STORE_NAME               0 (x)
             12 STORE_DEREF              0 (y)
             15 STORE_DEREF              1 (z)

Obtaining Attributes

The Getattr node type takes an expression and an attribute name. The attribute name can be a constant string, in which case a LOAD_ATTR opcode is used, and constant folding is done if possible:

>>> from peak.util.assembler import Getattr

>>> c = Code()
>>> c(Getattr(Local('x'), '__class__'))
>>> dis(c.code())
  0           0 LOAD_FAST                0 (x)
              3 LOAD_ATTR                0 (__class__)


>>> Getattr(Const(object), '__class__') # const expression, const result
Const(<type 'type'>)

Or the attribute name can be an expression, in which case a getattr() call is compiled instead:

>>> c = Code()
>>> c(Getattr(Local('x'), Local('y')))
>>> dis(c.code())
  0           0 LOAD_CONST               1 (<built-in function getattr>)
              3 LOAD_FAST                0 (x)
              6 LOAD_FAST                1 (y)
              9 CALL_FUNCTION            2

Calling Functions and Methods

>>> from peak.util.assembler import Call

The Call wrapper takes 1-4 arguments: the expression to be called, a sequence of positional arguments, a sequence of keyword/value pairs for explicit keyword arguments, an "*" argument, and a "**" argument. To omit any of the optional arguments, just pass in an empty sequence in its place:

>>> c = Code()
>>> c( Call(Global('type'), [Const(27)]) )

>>> dis(c.code())   # type(27)
  0           0 LOAD_GLOBAL              0 (type)
              3 LOAD_CONST               1 (27)
              6 CALL_FUNCTION            1

>>> c = Code()
>>> c(Call(Global('dict'), (), [('x', 42)]))

>>> dis(c.code())   # dict(x=42)
  0           0 LOAD_GLOBAL              0 (dict)
              3 LOAD_CONST               1 ('x')
              6 LOAD_CONST               2 (42)
              9 CALL_FUNCTION            256

>>> c = Code()
>>> c(Call(Global('foo'), (), (), Local('args'), Local('kw')))

>>> dis(c.code())   # foo(*args, **kw)
  0           0 LOAD_GLOBAL              0 (foo)
              3 LOAD_FAST                0 (args)
              6 LOAD_FAST                1 (kw)
              9 CALL_FUNCTION_VAR_KW     0

Returning Values

The Return(target) wrapper generates code for its target, followed by a RETURN_VALUE opcode:

>>> from peak.util.assembler import Return

>>> c = Code()
>>> c( Return(1) )
>>> dis(c.code())
  0           0 LOAD_CONST               1 (1)
              3 RETURN_VALUE

Code objects also have a return_() method that provides a more compact spelling of the same thing:

>>> c = Code()
>>> c.return_((1,2))
>>> dis(c.code())
  0           0 LOAD_CONST               1 (1)
              3 LOAD_CONST               2 (2)
              6 BUILD_TUPLE              2
              9 RETURN_VALUE

Both Return and return_() can be used with no argument, in which case None is returned:

>>> c = Code()
>>> c.return_()
>>> dis(c.code())
  0           0 LOAD_CONST               0 (None)
              3 RETURN_VALUE

>>> c = Code()
>>> c( Return() )
>>> dis(c.code())
  0           0 LOAD_CONST               0 (None)
              3 RETURN_VALUE

If Conditions

The If() node type generates conditional code, roughly equivalent to a Python if/else statement:

>>> from peak.util.assembler import If
>>> c = Code()
>>> c( If(Local('a'), Return(42), Return(55)) )
>>> dump(c.code())
                LOAD_FAST                0 (a)
                JUMP_IF_FALSE            L1
                POP_TOP
                LOAD_CONST               1 (42)
                RETURN_VALUE
        L1:     POP_TOP
                LOAD_CONST               2 (55)
                RETURN_VALUE

However, it can also be used like a Python 2.5+ conditional expression (regardless of the targeted Python version):

>>> c = Code()
>>> c( Return(If(Local('a'), 42, 55)) )
>>> dump(c.code())
                LOAD_FAST                0 (a)
                JUMP_IF_FALSE            L1
                POP_TOP
                LOAD_CONST               1 (42)
                JUMP_FORWARD             L2
        L1:     POP_TOP
                LOAD_CONST               2 (55)
        L2:     RETURN_VALUE

Note that If() does not do constant-folding on its condition; even if the condition is a constant, it will be tested at runtime. This avoids issues with using mutable constants, e.g.:

>>> c = Code()
>>> c(If(Const([]), 42, 55))
>>> dump(c.code())
                LOAD_CONST               1 ([])
                JUMP_IF_FALSE            L1
                POP_TOP
                LOAD_CONST               2 (42)
                JUMP_FORWARD             L2
        L1:     POP_TOP
                LOAD_CONST               3 (55)

Labels and Jump Targets

The forward reference callbacks returned by jump operations are also usable as code generation values, indicating that the jump should go to the current location. For example:

>>> c = Code()
>>> c.LOAD_CONST(99)
>>> forward = c.JUMP_IF_FALSE()
>>> c( 1, Code.POP_TOP, forward, Return(3) )
>>> dump(c.code())
                LOAD_CONST               1 (99)
                JUMP_IF_FALSE            L1
                LOAD_CONST               2 (1)
                POP_TOP
        L1:     LOAD_CONST               3 (3)
                RETURN_VALUE

However, there's an easier way to do the same thing, using Label objects:

>>> from peak.util.assembler import Label
>>> c = Code()
>>> skip = Label()

>>> c(99, skip.JUMP_IF_FALSE, 1, Code.POP_TOP, skip, Return(3))
>>> dump(c.code())
                LOAD_CONST               1 (99)
                JUMP_IF_FALSE            L1
                LOAD_CONST               2 (1)
                POP_TOP
        L1:     LOAD_CONST               3 (3)
                RETURN_VALUE

This approach has the advantage of being easy to use in complex trees. Label objects have attributes corresponding to every opcode that uses a bytecode address argument. Generating code for these attributes emits the the corresponding opcode, and generating code for the label itself defines where the previous opcodes will jump to. Labels can have multiple jumps targeting them, either before or after they are defined. But they can't be defined more than once:

>>> c(skip)
Traceback (most recent call last):
  ...
AssertionError: Label previously defined

More Conditional Jump Instructions

In Python 2.7, the traditional JUMP_IF_TRUE and JUMP_IF_FALSE instructions were replaced with four new instructions that either conditionally or unconditionally pop the value being tested. This was done to improve performance, since virtually all conditional jumps in Python code pop the value on one branch or the other.

To provide better cross-version compatibility, BytecodeAssembler emulates the old instructions on Python 2.7 by emitting a DUP_TOP followed by a POP_JUMP_IF_FALSE or POP_JUMP_IF_TRUE instruction.

However, since this decreases performance, BytecodeAssembler also emulates Python 2.7's JUMP_IF_FALSE_OR_POP and JUMP_IF_FALSE_OR_TRUE opcodes on older Pythons:

>>> c = Code()
>>> l1, l2 = Label(), Label()
>>> c(Local('a'), l1.JUMP_IF_FALSE_OR_POP, Return(27), l1)
>>> c(l2.JUMP_IF_TRUE_OR_POP, Return(42), l2, Code.RETURN_VALUE)
>>> dump(c.code())
                LOAD_FAST                0 (a)
                JUMP_IF_FALSE           L1
                POP_TOP
                LOAD_CONST               1 (27)
                RETURN_VALUE   
        L1:     JUMP_IF_TRUE            L2
                POP_TOP
                LOAD_CONST               2 (42)
                RETURN_VALUE
        L2:     RETURN_VALUE

This means that you can immediately begin using the "or-pop" variations, in place of a jump followed by a pop, and BytecodeAssembler will use the faster single instruction automatically on Python 2.7+.

BytecodeAssembler also supports using Python 2.7's conditional jumps that do unconditional pops, but currently cannot emulate them on older Python versions, so at the moment you should use them only when your code requires Python 2.7.

(Note: for ease in doctesting across Python versions, the dump() function always shows the code as if it were generated for Python 2.6 or lower, so if you need to check the actual bytecodes generated, you must use Python's dis.dis() function instead!)

N-Way Comparisons

You can generate N-way comparisons using the Compare() node type:

>>> from peak.util.assembler import Compare

>>> c = Code()
>>> c(Compare(Local('a'), [('<', Local('b'))]))
>>> dis(c.code())
  0           0 LOAD_FAST                0 (a)
              3 LOAD_FAST                1 (b)
              6 COMPARE_OP               0 (<)

3-way comparisons generate code that's a bit more complex. Here's a three-way comparison (a<b<c):

>>> c = Code()
>>> c.return_(Compare(Local('a'), [('<', Local('b')), ('<', Local('c'))]))
>>> dump(c.code())
                LOAD_FAST                0 (a)
                LOAD_FAST                1 (b)
                DUP_TOP
                ROT_THREE
                COMPARE_OP               0 (<)
                JUMP_IF_FALSE           L1
                POP_TOP
                LOAD_FAST                2 (c)
                COMPARE_OP               0 (<)
                JUMP_FORWARD            L2
        L1:     ROT_TWO
                POP_TOP
        L2:     RETURN_VALUE

And a four-way (a<b>c!=d):

>>> c = Code()
>>> c.return_(
...     Compare( Local('a'), [
...         ('<', Local('b')), ('>', Local('c')), ('!=', Local('d'))
...     ])
... )
>>> dump(c.code())
                LOAD_FAST                0 (a)
                LOAD_FAST                1 (b)
                DUP_TOP
                ROT_THREE
                COMPARE_OP               0 (<)
                JUMP_IF_FALSE           L1
                POP_TOP
                LOAD_FAST                2 (c)
                DUP_TOP
                ROT_THREE
                COMPARE_OP               4 (>)
                JUMP_IF_FALSE           L1
                POP_TOP
                LOAD_FAST                3 (d)
                COMPARE_OP               3 (!=)
                JUMP_FORWARD            L2
        L1:     ROT_TWO
                POP_TOP
        L2:     RETURN_VALUE

Sequence Unpacking

The UnpackSequence node type takes a sequence of code generation targets, and generates an UNPACK_SEQUENCE of the correct length, followed by the targets:

>>> from peak.util.assembler import UnpackSequence
>>> c = Code()  
>>> c((1,2), UnpackSequence([LocalAssign('x'), LocalAssign('y')]))
>>> dis(c.code())   # x, y = 1, 2
  0           0 LOAD_CONST               1 (1)
              3 LOAD_CONST               2 (2)
              6 BUILD_TUPLE              2
              9 UNPACK_SEQUENCE          2
             12 STORE_FAST               0 (x)
             15 STORE_FAST               1 (y)

Yield Statements

The YieldStmt node type generates the necessary opcode(s) for a yield statement, based on the target Python version. (In Python 2.5+, a POP_TOP must be generated after a YIELD_VALUE in order to create a yield statement, as opposed to a yield expression.) It also sets the code flags needed to make the resulting code object a generator:

>>> from peak.util.assembler import YieldStmt
>>> c = Code()
>>> c(YieldStmt(1), YieldStmt(2), Return(None))
>>> list(eval(c.code()))
[1, 2]

Folding Function Calls

The Call wrapper can also do simple constant folding, if all of its input parameters are constants. (Actually, the args and kwargs arguments must be sequences of constants and 2-tuples of constants, respectively.)

If a Call can thus compute its value in advance, it does so, returning a Const node instead of a Call node:

>>> Call( Const(type), [1] )
Const(<type 'int'>)

Thus, you can also take the const_value() of such calls:

>>> const_value( Call( Const(dict), [], [('x',27)] ) )
{'x': 27}

Which means that constant folding can propagate up an AST if the result is passed in to another Call:

>>> Call(Const(type), [Call( Const(dict), [], [('x',27)] )])
Const(<type 'dict'>)

Notice that this folding takes place eagerly, during AST construction. If you want to implement delayed folding after constant propagation or variable substitution, you'll need to recreate the tree, or use your own custom AST types. (See Custom Code Generation, below.)

Note that you can disable folding using the fold=False keyword argument to Call, if you want to ensure that even compile-time constants are computed at runtime. Compare:

>>> c = Code()
>>> c( Call(Const(type), [1]) )
>>> dis(c.code())
  0           0 LOAD_CONST               1 (<type 'int'>)

>>> c = Code()
>>> c( Call(Const(type), [1], fold=False) )
>>> dis(c.code())
  0           0 LOAD_CONST               1 (<type 'type'>)
              3 LOAD_CONST               2 (1)
              6 CALL_FUNCTION            1

Folding is also automatically disabled for calls with no arguments of any kind (such as globals() or locals()), whose values are much more likely to change dynamically at runtime:

>>> c = Code()
>>> c( Call(Const(locals)) )
>>> dis(c.code())
  0           0 LOAD_CONST               1 (<built-in function locals>)
              3 CALL_FUNCTION            0

Note, however, that folding is disabled for any zero-argument call, regardless of the thing being called. It is not specific to locals() and globals(), in other words.

Logical And/Or

You can evaluate logical and/or expressions using the And and Or node types:

>>> from peak.util.assembler import And, Or

>>> c = Code()
>>> c.return_( And([Local('x'), Local('y')]) )
>>> dump(c.code())
                LOAD_FAST                0 (x)
                JUMP_IF_FALSE           L1        
                POP_TOP
                LOAD_FAST                1 (y)
        L1:     RETURN_VALUE

>>> c = Code()
>>> c.return_( Or([Local('x'), Local('y')]) )
>>> dump(c.code())
                LOAD_FAST                0 (x)
                JUMP_IF_TRUE            L1
                POP_TOP
                LOAD_FAST                1 (y)
        L1:     RETURN_VALUE

True or false constants are folded automatically, avoiding code generation for intermediate values that will never be used in the result:

>>> c = Code()
>>> c.return_( And([1, 2, Local('y')]) )
>>> dis(c.code())
  0           0 LOAD_FAST                0 (y)
              3 RETURN_VALUE

>>> c = Code()
>>> c.return_( And([1, 2, Local('y'), 0]) )
>>> dump(c.code())
                LOAD_FAST                0 (y)
                JUMP_IF_FALSE           L1
                POP_TOP
                LOAD_CONST               1 (0)
        L1:     RETURN_VALUE

>>> c = Code()
>>> c.return_( Or([1, 2, Local('y')]) )
>>> dis(c.code())
  0           0 LOAD_CONST               1 (1)
              3 RETURN_VALUE

>>> c = Code()
>>> c.return_( Or([False, Local('y'), 3]) )
>>> dump(c.code())
                LOAD_FAST                0 (y)
                JUMP_IF_TRUE            L1
                POP_TOP
                LOAD_CONST               1 (3)
        L1:     RETURN_VALUE

Constant Folding in Custom Targets

If you want to incorporate constant-folding into your AST nodes, you can do so by checking for constant values and folding them at either construction or code generation time. For example, this And node type (a simpler version of the one included in peak.util.assembler) folds constants during code generation, by not generating unnecessary branches when it can prove which way a branch will go:

>>> from peak.util.assembler import NotAConstant

>>> def And(values, code=None):
...     if code is None:
...         return tuple(values),
...     end = Label()
...     for value in values[:-1]:
...         try:
...             if const_value(value):
...                 continue        # true constants can be skipped
...         except NotAConstant:    # but non-constants require code
...             code(value, end.JUMP_IF_FALSE_OR_POP)
...         else:       # and false constants end the chain right away
...             return code(value, end)
...     code(values[-1], end)
>>> And = nodetype()(And)

>>> c = Code()
>>> c.return_( And([1, 2]) )
>>> dis(c.code())
  0           0 LOAD_CONST               1 (2)
              3 RETURN_VALUE

>>> c = Code()
>>> c.return_( And([1, 2, Local('x')]) )
>>> dis(c.code())
  0           0 LOAD_FAST                0 (x)
              3 RETURN_VALUE

>>> c = Code()
>>> c.return_( And([Local('x'), False, 27]) )
>>> dump(c.code())
                LOAD_FAST                0 (x)
                JUMP_IF_FALSE           L1
                POP_TOP
                LOAD_CONST               1 (False)
        L1:     RETURN_VALUE

The above example only folds constants at code generation time, however. You can also do constant folding at AST construction time, using the fold_args() function. For example:

>>> from peak.util.assembler import fold_args

>>> def Getattr(ob, name, code=None):
...     try:
...         name = const_value(name)
...     except NotAConstant:
...         return Call(Const(getattr), [ob, name])
...     if code is None:
...         return fold_args(Getattr, ob, name)
...     code(ob)
...     code.LOAD_ATTR(name)
>>> Getattr = nodetype()(Getattr)

>>> const_value(Getattr(1, '__class__'))
<type 'int'>

The fold_args() function tries to evaluate the node immediately, if all of its arguments are constants, by creating a temporary Code object, and running the supplied function against it, then doing an eval() on the generated code and wrapping the result in a Const. However, if any of the arguments are non-constant, the original arguments (less the function) are returned. This causes a normal node instance to be created instead of a Const.

This isn't a very fast way of doing partial evaluation, but it makes it really easy to define new code generation targets without writing custom constant-folding code for each one. Just return fold_args(ThisType, *args) instead of return args, if you want your node constructor to be able to do eager evaluation. If you need to, you can check your parameters in order to decide whether to call fold_args() or not; this is in fact how Call implements its fold argument and the suppression of folding when the call has no arguments.

(By the way, this same Getattr node type is also available

Exception Stack Size Adjustment

When you issue a SETUP_EXCEPT or SETUP_FINALLY, the code's maximum stack size is raised to ensure that it's at least 3 items higher than the current stack size. That way, there will be room for the items that Python puts on the stack when jumping to a block's exception handling code:

>>> c = Code()
>>> c.SETUP_FINALLY()
>>> c.stack_size, c.co_stacksize
(0, 3)

As you can see, the current stack size is unchanged, but the maximum stack size has increased. This increase is relative to the current stack size, though; it's not an absolute increase:

>>> c = Code()
>>> c(1,2,3,4, *[Code.POP_TOP]*4)   # push 4 things, then pop 'em
>>> c.SETUP_FINALLY()
>>> c.stack_size, c.co_stacksize
(0, 4)

And this stack adjustment doesn't happen for loops, because they don't have exception handlers:

>>> c = Code()
>>> c.SETUP_LOOP()
>>> c.stack_size, c.co_stacksize
(0, 0)

Try/Except Blocks

In the case of SETUP_EXCEPT, the current stack size is increased by 3 after a POP_BLOCK, because the code that follows will be an exception handler and will thus always have exception items on the stack:

>>> c = Code()
>>> c.SETUP_EXCEPT()
>>> else_ = c.POP_BLOCK()
>>> c.stack_size, c.co_stacksize
(3, 3)

When a POP_BLOCK() is matched with a SETUP_EXCEPT, it automatically emits a JUMP_FORWARD and returns a forward reference that should be called back when the "else" clause or end of the entire try/except statement is reached:

>>> c.POP_TOP()     # get rid of exception info
>>> c.POP_TOP()
>>> c.POP_TOP()
>>> else_()
>>> c.return_()
>>> dump(c.code())
                SETUP_EXCEPT            L1
                POP_BLOCK
                JUMP_FORWARD            L2
        L1:     POP_TOP
                POP_TOP
                POP_TOP
        L2:     LOAD_CONST               0 (None)
                RETURN_VALUE

In the example above, an empty block executes with an exception handler that begins at offset 7. When the block is done, it jumps forward to the end of the try/except construct at offset 10. The exception handler does nothing but remove the exception information from the stack before it falls through to the end.

Note, by the way, that it's usually easier to use labels to define blocks like this:

>>> c = Code()
>>> done = Label()
>>> c(
...     done.SETUP_EXCEPT,
...     done.POP_BLOCK,
...         Code.POP_TOP, Code.POP_TOP, Code.POP_TOP,
...     done,
...     Return()
... )

>>> dump(c.code())
                SETUP_EXCEPT             L1
                POP_BLOCK
                JUMP_FORWARD             L2
        L1:     POP_TOP
                POP_TOP
                POP_TOP
        L2:     LOAD_CONST               0 (None)
                RETURN_VALUE

(Labels have a POP_BLOCK attribute that you can pass in when generating code.)

And, for generating typical try/except blocks, you can use the TryExcept node type, which takes a body, a sequence of exception-type/handler pairs, and an optional "else" clause:

>>> from peak.util.assembler import TryExcept
>>> c = Code()
>>> c.return_(
...     TryExcept(
...         Return(1),                                      # body
...         [(Const(KeyError),2), (Const(TypeError),3)],    # handlers
...         Return(4)                                       # else clause
...     )
... )

>>> dump(c.code())
                SETUP_EXCEPT            L1
                LOAD_CONST               1 (1)
                RETURN_VALUE
                POP_BLOCK
                JUMP_FORWARD            L4
        L1:     DUP_TOP
                LOAD_CONST               2 (<...exceptions.KeyError...>)
                COMPARE_OP              10 (exception match)
                JUMP_IF_FALSE           L2
                POP_TOP
                POP_TOP
                POP_TOP
                POP_TOP
                LOAD_CONST               3 (2)
                JUMP_FORWARD            L5
        L2:     POP_TOP
                DUP_TOP
                LOAD_CONST               4 (<...exceptions.TypeError...>)
                COMPARE_OP              10 (exception match)
                JUMP_IF_FALSE           L3
                POP_TOP
                POP_TOP
                POP_TOP
                POP_TOP
                LOAD_CONST               5 (3)
                JUMP_FORWARD            L5
        L3:     POP_TOP
                END_FINALLY
        L4:     LOAD_CONST               6 (4)
                RETURN_VALUE
        L5:     RETURN_VALUE

Try/Finally Blocks

When a POP_BLOCK() is matched with a SETUP_FINALLY, it automatically emits a LOAD_CONST(None), so that when the corresponding END_FINALLY is reached, it will know that the "try" block exited normally. Thus, the normal pattern for producing a try/finally construct is as follows:

>>> c = Code()
>>> c.SETUP_FINALLY()
>>> # "try" suite goes here
>>> c.POP_BLOCK()
>>> # "finally" suite goes here
>>> c.END_FINALLY()

And it produces code that looks like this:

>>> dump(c.code())
                SETUP_FINALLY           L1
                POP_BLOCK
                LOAD_CONST               0 (None)
        L1:     END_FINALLY

The END_FINALLY opcode will remove 1, 2, or 3 values from the stack at runtime, depending on how the "try" block was exited. In the case of simply "falling off the end" of the "try" block, however, the inserted LOAD_CONST(None) puts one value on the stack, and that one value is popped off by the END_FINALLY. For that reason, Code objects treat END_FINALLY as if it always popped exactly one value from the stack, even though at runtime this may vary. This means that the estimated stack levels within the "finally" clause may not be accurate -- which is why POP_BLOCK() adjusts the maximum expected stack size to accomodate up to three values being put on the stack by the Python interpreter for exception handling.

For your convenience, the TryFinally node type can also be used to generate try/finally blocks:

>>> from peak.util.assembler import TryFinally
>>> c = Code()
>>> c( TryFinally(ExprStmt(1), ExprStmt(2)) )
>>> dump(c.code())
                SETUP_FINALLY           L1
                LOAD_CONST               1 (1)
                POP_TOP
                POP_BLOCK
                LOAD_CONST               0 (None)
        L1:     LOAD_CONST               2 (2)
                POP_TOP
                END_FINALLY

Loops

The POP_BLOCK for a loop marks the end of the loop body, and the beginning of the "else" clause, if there is one. It returns a forward reference that should be called back either at the end of the "else" clause, or immediately if there is no "else". Any BREAK_LOOP opcodes that appear in the loop body will jump ahead to the point at which the forward reference is resolved.

Here, we'll generate a loop that counts down from 5 to 0, with an "else" clause that returns 42. Three labels are needed: one to mark the end of the overall block, one that's looped back to, and one that marks the "else" clause:

>>> c = Code()
>>> block = Label()
>>> loop = Label()
>>> else_ = Label()
>>> c(
...     block.SETUP_LOOP,
...         5,      # initial setup - this could be a GET_ITER instead
...     loop,
...         else_.JUMP_IF_FALSE,        # while x:
...         1, Code.BINARY_SUBTRACT,    #     x -= 1
...         loop.CONTINUE_LOOP,
...     else_,                          # else:
...         Code.POP_TOP,
...     block.POP_BLOCK,
...         Return(42),                 #     return 42
...     block,
...     Return()
... )

>>> dump(c.code())
                SETUP_LOOP              L3
                LOAD_CONST               1 (5)
        L1:     JUMP_IF_FALSE           L2
                LOAD_CONST               2 (1)
                BINARY_SUBTRACT
                JUMP_ABSOLUTE           L1
        L2:     POP_TOP
                POP_BLOCK
                LOAD_CONST               3 (42)
                RETURN_VALUE
        L3:     LOAD_CONST               0 (None)
                RETURN_VALUE

>>> eval(c.code())
42

Break and Continue

The BREAK_LOOP and CONTINUE_LOOP opcodes can only be used inside of an active loop:

>>> c = Code()
>>> c.BREAK_LOOP()
Traceback (most recent call last):
  ...
AssertionError: Not inside a loop

>>> c.CONTINUE_LOOP(c.here())
Traceback (most recent call last):
  ...
AssertionError: Not inside a loop

And CONTINUE_LOOP is automatically replaced with a JUMP_ABSOLUTE if it occurs directly inside a loop block:

>>> c.LOAD_CONST(57)
>>> c.SETUP_LOOP()
>>> fwd = c.JUMP_IF_TRUE()
>>> c.CONTINUE_LOOP(c.here())
>>> fwd()
>>> c.BREAK_LOOP()
>>> c.POP_BLOCK()()
>>> dump(c.code())
                LOAD_CONST               1 (57)
                SETUP_LOOP              L3
                JUMP_IF_TRUE            L2
        L1:     JUMP_ABSOLUTE           L1
        L2:     BREAK_LOOP
                POP_BLOCK

In other words, CONTINUE_LOOP only really emits a CONTINUE_LOOP opcode if it's inside some other kind of block within the loop, e.g. a "try" clause:

>>> c = Code()
>>> c.LOAD_CONST(57)
>>> c.SETUP_LOOP()
>>> loop = c.here()
>>> c.SETUP_FINALLY()
>>> fwd = c.JUMP_IF_TRUE()
>>> c.CONTINUE_LOOP(loop)
>>> fwd()
>>> c.POP_BLOCK()
>>> c.END_FINALLY()
>>> c.POP_BLOCK()()
>>> dump(c.code())
                LOAD_CONST               1 (57)
                SETUP_LOOP              L4
        L1:     SETUP_FINALLY           L3
                JUMP_IF_TRUE            L2
                CONTINUE_LOOP           L1
        L2:     POP_BLOCK
                LOAD_CONST               0 (None)
        L3:     END_FINALLY
                POP_BLOCK

for Loops

There is a For() node type available for generating simple loops (without break/continue support). It takes an iterable expression, an assignment clause, and a loop body:

>>> from peak.util.assembler import For
>>> y = Call(Const(range), (3,))
>>> x = LocalAssign('x')
>>> body = Suite([Local('x'), Code.PRINT_EXPR])

>>> c = Code()
>>> c(For(y, x, body))  # for x in range(3): print x
>>> c.return_()
>>> dump(c.code())
                LOAD_CONST               1 ([0, 1, 2])
                GET_ITER
        L1:     FOR_ITER                L2
                STORE_FAST               0 (x)
                LOAD_FAST                0 (x)
                PRINT_EXPR
                JUMP_ABSOLUTE           L1
        L2:     LOAD_CONST               0 (None)
                RETURN_VALUE

The arguments are given in execution order: first the "in" value of the loop, then the assignment to a loop variable, and finally the body of the loop. The distinction between the assignment and body, however, is only for clarity and convenience (to avoid needing to glue the assignment to the body with a Suite). If you already have a suite or only need one node for the entire loop body, you can do the same thing with only two arguments:

>>> c = Code()
>>> c(For(y, Code.PRINT_EXPR))
>>> c.return_()
>>> dump(c.code())
                LOAD_CONST               1 ([0, 1, 2])
                GET_ITER
        L1:     FOR_ITER                L2
                PRINT_EXPR
                JUMP_ABSOLUTE           L1
        L2:     LOAD_CONST               0 (None)
                RETURN_VALUE

Notice, by the way, that For() does NOT set up a loop block for you, so if you want to be able to use break and continue, you'll need to wrap the loop in a labelled SETUP_LOOP/POP_BLOCK pair, as described in the preceding sections.

List Comprehensions

In order to generate correct list comprehension code for the target Python version, you must use the ListComp() and LCAppend() node types. This is because Python versions 2.4 and up store the list being built in a temporary variable, and use a special LIST_APPEND opcode to append values, while 2.3 stores the list's append() method in the temporary variable, and calls it to append values.

The ListComp() node wraps a code body (usually a For() loop) and manages the creation and destruction of a temporary variable (e.g. _[1], _[2], etc.). The LCAppend() node type wraps a value or expression to be appended to the innermost active ListComp() in progress:

>>> from peak.util.assembler import ListComp, LCAppend
>>> c = Code()
>>> simple = ListComp(For(y, x, LCAppend(Local('x'))))
>>> c.return_(simple)
>>> eval(c.code())
[0, 1, 2]

>>> c = Code()
>>> c.return_(ListComp(For(y, x, LCAppend(simple))))
>>> eval(c.code())
[[0, 1, 2], [0, 1, 2], [0, 1, 2]]

Free and Cell Variables

To implement closures and nested scopes, your code objects must use "free" or "cell" variables in place of regular "fast locals". A "free" variable is one that is defined in an outer scope, and a "cell" variable is one that's defined in the current scope, but will also be used by nested functions.

The simplest way to set up free or cell variables is to use a code object's makefree(names) and makecells(names) methods:

>>> c = Code()
>>> c.co_cellvars
()
>>> c.co_freevars
()

>>> c.makefree(['x', 'y'])
>>> c.makecells(['z'])

>>> c.co_cellvars
('z',)
>>> c.co_freevars
('x', 'y')

When a name has been defined as a free or cell variable, the _DEREF opcode variants are used to generate Local() and LocalAssign() nodes:

>>> c((Local('x'), Local('y')), LocalAssign('z'))
>>> dis(c.code())
  0           0 LOAD_DEREF               1 (x)
              3 LOAD_DEREF               2 (y)
              6 BUILD_TUPLE              2
              9 STORE_DEREF              0 (z)

If you have already written code in a code object that operates on the relevant locals, the code is retroactively patched to use the _DEREF opcodes:

>>> c = Code()
>>> c((Local('x'), Local('y')), LocalAssign('z'))
>>> dis(c.code())
  0           0 LOAD_FAST                0 (x)
              3 LOAD_FAST                1 (y)
              6 BUILD_TUPLE              2
              9 STORE_FAST               2 (z)

>>> c.makefree(['x', 'y'])
>>> c.makecells(['z'])

>>> dis(c.code())
  0           0 LOAD_DEREF               1 (x)
              3 LOAD_DEREF               2 (y)
              6 BUILD_TUPLE              2
              9 STORE_DEREF              0 (z)

This means that you can defer the decision of which locals are free/cell variables until the code is ready to be generated. In fact, by passing in a "parent" code object to the .code() method, you can get BytecodeAssembler to automatically call makefree() and makecells() for the correct variable names in the child and parent code objects, as we'll see in the next section.

Nested Code Objects

To create a code object for use in a nested scope, you can use the parent code object's nested() method. It works just like the from_spec() classmethod, except that the co_filename of the parent is copied to the child:

>>> p = Code()
>>> p.co_filename = 'testname'

>>> c = p.nested('sub', ['a','b'], 'c', 'd')

>>> c.co_name
'sub'

>>> c.co_filename
'testname'

>>> tuple(inspect.getargs(c.code(p)))
(['a', 'b'], 'c', 'd')

Notice that you must pass the parent code object to the child's .code() method to ensure that free/cell variables are properly set up. When the code() method is given another code object as a parameter, it automatically converts any locally-read (but not written) to "free" variables in the child code, and ensures that those same variables become "cell" variables in the supplied parent code object:

>>> p.LOAD_CONST(42)
>>> p(LocalAssign('a'))
>>> dis(p.code())
  0           0 LOAD_CONST               1 (42)
              3 STORE_FAST               0 (a)

>>> c = p.nested()
>>> c(Local('a'))

>>> dis(c.code(p))
  0           0 LOAD_DEREF               0 (a)

>>> dis(p.code())
  0           0 LOAD_CONST               1 (42)
              3 STORE_DEREF              0 (a)

Notice that the STORE_FAST in the parent code object was automatically patched to a STORE_DEREF, with an updated offset if applicable. Any future use of Local('a') or LocalAssign('a') in the parent or child code objects will now refer to the free/cell variable, rather than the "local" variable:

>>> p(Local('a'))
>>> dis(p.code())
  0           0 LOAD_CONST               1 (42)
              3 STORE_DEREF              0 (a)
              6 LOAD_DEREF               0 (a)

>>> c(LocalAssign('a'))
>>> dis(c.code(p))
  0           0 LOAD_DEREF               0 (a)
              3 STORE_DEREF              0 (a)

Function()

The Function(body, name='<lambda>', args=(), var=None, kw=None, defaults=()) node type creates a function object from the specified body and the optional name, argument specs, and defaults. The Function() node generates code to create the function object with the appropriate defaults and closure (if applicable), and any needed free/cell variables are automatically set up in the parent and child code objects. The newly generated function will be on top of the stack at the end of the generated code:

>>> from peak.util.assembler import Function
>>> c = Code()
>>> c.co_filename = '<string>'
>>> c.return_(Function(Return(Local('a')), 'f', ['a'], defaults=[42]))
>>> dis(c.code())
  0           0 LOAD_CONST               1 (42)
              3 LOAD_CONST               2 (<... f ..., file "<string>", line -1>)
              6 MAKE_FUNCTION            1
              9 RETURN_VALUE

Now that we've generated the code for a function returning a function, let's run it, to get the function we defined:

>>> f = eval(c.code())
>>> f
<function f at ...>

>>> tuple(inspect.getargspec(f))
(['a'], None, None, (42,))

>>> f()
42

>>> f(99)
99

Now let's create a doubly nested function, with some extras:

>>> c = Code()
>>> c.co_filename = '<string>'
>>> c.return_(
...     Function(Return(Function(Return(Local('a')))),
...     'f', ['a', 'b'], 'c', 'd', [99, 66])
... )
>>> dis(c.code())
  0           0 LOAD_CONST               1 (99)
              3 LOAD_CONST               2 (66)
              6 LOAD_CONST               3 (<... f ..., file "<string>", line -1>)
              9 MAKE_FUNCTION            2
             12 RETURN_VALUE

>>> f = eval(c.code())
>>> f
<function f at ...>

>>> tuple(inspect.getargspec(f))
(['a', 'b'], 'c', 'd', (99, 66))

>>> dis(f)
  0           0 LOAD_CLOSURE             0 (a)
              ... LOAD_CONST               1 (<... <lambda> ..., file "<string>", line -1>)
              ... MAKE_CLOSURE             0
              ... RETURN_VALUE

>>> dis(f())
  0           0 LOAD_DEREF               0 (a)
              3 RETURN_VALUE    

>>> f(42)()
42

>>> f()()
99

As you can see, Function() not only takes care of setting up free/cell variables in all the relevant scopes, it also chooses whether to use MAKE_FUNCTION or MAKE_CLOSURE, and generates code for the defaults.

(Note, by the way, that the defaults argument should be a sequence of generatable expressions; in the examples here, we used numbers, but they could have been arbitrary expression nodes.)