NOTE: This module has not been updated for many bytecode changes in recent Python 3.x versions, and at present is completely incompatible with Python 3.6 (to the point of causing core dumps). We suggest using the Python ast module instead.
This module was originally developed when Python 2.5 was new, and support for the ast module was still immature; the only real benefits it provides over Python's builtin AST are constant folding and the ability to generate hand-specified code not tied to Python's usual compilation process. (e.g. PEAK-Rules using it to create jump tables and common subexpression elimination in its rules interpreter.)
peak.util.assembler is a simple bytecode assembler module that handles most
low-level bytecode generation details like jump offsets, stack size tracking,
line number table generation, constant and variable name index tracking, etc.
That way, you can focus your attention on the desired semantics of your
bytecode instead of on these mechanical issues.
In addition to a low-level opcode-oriented API for directly generating specific Python bytecodes, this module also offers an extensible mini-AST framework for generating code from high-level specifications. This framework does most of the work needed to transform tree-like structures into linear bytecode instructions, and includes the ability to do compile-time constant folding.
Please see the BytecodeAssembler reference manual for more details.
Changes since version 0.6.1:
- Experimental Python 3 support, including emulation of restored
BINARY_DIVIDE,UNARY_CONVERT, andSLICE_#opcodes.
Changes since version 0.6:
- Fix bad stack calculations for BUILD_CLASS opcode
Changes since version 0.5.2:
- Symbolic disassembly with full emulation of backward-compatible
JUMP_IF_TRUEandJUMP_IF_FALSEopcodes on Python 2.7 -- tests now run clean on Python 2.7. - Support for backward emulation of Python 2.7's
JUMP_IF_TRUE_OR_POPandJUMP_IF_FALSE_OR_POPinstructions on earlier Python versions; these emulations are also used in BytecodeAssembler's internal code generation, for maximum performance on 2.7+ (with no change to performance on older versions).
Changes since version 0.5.1:
- Initial support for Python 2.7's new opcodes and semantics changes, mostly by emulating older versions' behavior with macros. (0.5.2 is really just a quick-fix release to allow packages using BytecodeAssembler to run on 2.7 without having to change any of their code generation; future releases will provide proper support for the new and changed opcodes, as well as a test suite that doesn't show spurious differences in the disassembly listings under Python 2.7.)
Changes since version 0.5:
- Fix incorrect stack size calculation for
MAKE_CLOSUREon Python 2.5+
Changes since version 0.3:
- New node types:
For(iterable, assign, body)-- define a "for" loop over iterableUnpackSequence(nodes)-- unpacks a sequence that'slen(nodes)long, and then generates the given nodes.LocalAssign(name)-- issues aSTORE_FAST,STORE_DEREForSTORE_LOCALas appropriate for the given name.Function(body, name='<lambda>', args=(), var=None, kw=None, defaults=())-- creates a nested function from body and puts it on the stack.If(cond, then_, else_=Pass)-- "if" statement analogueListComp(body)andLCAppend(value)-- implement list comprehensionsYieldStmt(value)-- generates aYIELD_VALUE(plus aPOP_TOPin Python 2.5+)
Codeobjects are now iterable, yielding(offset, op, arg)triples, where op is numeric and arg is either numeric orNone.Codeobjects'.code()method can now take a "parent"Codeobject, to link the child code's free variables to cell variables in the parent.- Added
Code.from_spec()classmethod, that initializes a code object from a name and argument spec. Codeobjects now have a.nested(name, args, var, kw)method, that creates a child code object with the sameco_filenameand the supplied name/arg spec.- Fixed incorrect stack tracking for the
FOR_ITERandYIELD_VALUEopcodes - Ensure that
CO_GENERATORflag is set ifYIELD_VALUEopcode is used - Change tests so that Python 2.3's broken line number handling in
dis.disand constant-folding optimizer don't generate spurious failures in this package's test suite.
Changes since version 0.2:
- Added
Suite,TryExcept, andTryFinallynode types - Added a
Getattrnode type that does static or dynamic attribute access and constant folding - Fixed
code.from_function()not copying theco_filenameattribute whencopy_linenowas specified. - The
repr()of AST nodes doesn't include a trailing comma for 1-argument node types any more. - Added a
Passsymbol that generates no code, aCompare()node type that does n-way comparisons, andAnd()andOr()node types for doing logical operations. - The
COMPARE_OP()method now accepts operator strings like"<=","not in","exception match", and so on, as well as numeric opcodes. See the standard library'sopcodemodule for a complete list of the strings accepted (in thecmp_optuple)."<>"is also accepted as an alias for"!=". - Added code to verify that forward jump offsets don't exceed a 64KB span, and support absolute backward jumps to locations >64KB.
Changes since version 0.1:
- Constant handling has been fixed so that it doesn't confuse equal values of
differing types (e.g.
1.0andTrue), or equal unhashable objects (e.g. two empty lists). - Removed
nil,ast_curry()andfolding_curry(), replacing them with thenodetype()decorator andfold_args(); please see the docs for more details. - Added stack tracking across jumps, globally verifying stack level prediction consistency and automatically rejecting attempts to generate dead code. It should now be virtually impossible to accidentally generate bytecode that can crash the interpreter. (If you find a way, let me know!)
Changes since version 0.0.1:
- Added massive quantities of new documentation and examples
- Full block, loop, and closure support
- High-level functional code generation from trees, with smart labels and blocks, constant folding, extensibility, smart local variable names, etc.
- The
.label()method was renamed to.here()to distinguish it from the new smartLabelobjects. - Docs and tests were moved to README.txt instead of assembler.txt
- Added a demo that implements a "switch"-like statement template that shows
how to extend the code generation system and how to abuse
END_FINALLYto implement a "computed goto" in bytecode. - Various bug fixes
There are a few features that aren't tested yet, and not all opcodes may be fully supported. Also note the following limitations:
- Jumps to as-yet-undefined labels cannot span a distance greater than 65,535 bytes.
- The
dis()function in Python 2.3 has a bug that makes it show incorrect line numbers when the difference between two adjacent line numbers is greater than 255. (To work around this, the test_suite uses a later version ofdis(), but do note that it may affect your own tests if you usedis()with Python 2.3 and use widely separated line numbers.)
If you find any other issues, please let me know.
Please also keep in mind that this is a work in progress, and the API may change if I come up with a better way to do something.
Questions and discussion regarding this software should be directed to the PEAK Mailing List.
Table of Contents
- Programmer API
- Code Objects
- Symbolic Disassembler
- Opcodes and Arguments
- High-Level Code Generation
- Constant Detection and Folding
- Custom Code Generation
- Setting the Code's Calling Signature
- Code Attributes
- Stack Size Tracking and Dead Code Detection
- Blocks, Loops, and Exception Handling
- Closures and Nested Functions
- Internals and Doctests
To generate bytecode, you create a Code instance and perform operations
on it. For example, here we create a Code object representing lines
15 and 16 of some input source:
>>> from peak.util.assembler import Code >>> c = Code() >>> c.set_lineno(15) # set the current line number (optional) >>> c.LOAD_CONST(42) >>> c.set_lineno(16) # set it as many times as you like >>> c.RETURN_VALUE()
You'll notice that most Code methods are named for a CPython bytecode
operation, but there also some other methods like .set_lineno() to let you
set the current line number. There's also a .code() method that returns
a Python code object, representing the current state of the Code you've
generated:
>>> from dis import dis >>> dis(c.code()) 15 0 LOAD_CONST 1 (42) 16 3 RETURN_VALUE
As you can see, Code instances automatically generate a line number table
that maps each set_lineno() to the corresponding position in the bytecode.
And of course, the resulting code objects can be run with eval() or
exec, or used with new.function/types.FunctionType to create a
function:
>>> eval(c.code()) 42 >>> exec(c.code()) # exec discards the return value, so no output here >>> try: ... from new import function ... except ImportError: # Python 3 workarounds ... from types import FunctionType as function ... long = int ... unicode = str >>> f = function(c.code(), globals()) >>> f() 42
Finally, code objects are also iterable, yielding (offset, opcode, arg)
tuples, where arg is None for opcodes with no arguments, and an integer
otherwise:
>>> import peak.util.assembler as op >>> list(c) == [ ... (0, op.LOAD_CONST, 1), ... (3, op.RETURN_VALUE, None) ... ] True
This can be useful for testing or otherwise inspecting code you've generated.
Python's built-in disassembler can be verbose and hard to read when inspecting complex generated code -- usually you don't care about bytecode offsets or line numbers as much as you care about labels, for example.
So, BytecodeAssembler provides its own, simplified disassembler, which we'll be using for more complex listings in this manual:
>>> from peak.util.assembler import dump
Some sample output, that also showcases some of BytecodeAssembler's High-Level Code Generation features:
>>> c = Code()
>>> from peak.util.assembler import Compare, Local
>>> 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
As you can see, the line numbers and bytecode offsets have been dropped,
making it esier to see where the jumps go. (This also makes doctests more
robust against Python version changes, as dump() has some extra code to
make conditional jumps appear consistent across the major changes that were
made to conditional jump instructions between Python 2.6 and 2.7.)
Code objects have methods for all of CPython's symbolic opcodes. Generally
speaking, each method accepts either zero or one argument, depending on whether
the opcode accepts an argument.
Python bytecode always encodes opcode arguments as 16 or 32-bit integers, but
sometimes these numbers are actually offsets into a sequence of names or
constants. Code objects take care of maintaining these sequences for you,
allowing you to just pass in a name or value directly, instead of needing to
keep track of what numbers map to what names or values.
The name or value you pass in to such methods will be looked up in the appropriate table (see Code Attributes below for a list), and if not found, it will be added:
>>> c = Code()
>>> c.co_consts, c.co_varnames, c.co_names
([None], [], [])
>>> c.LOAD_CONST(42)
>>> c.LOAD_FAST('x')
>>> c.LOAD_GLOBAL('y')
>>> c.LOAD_NAME('z')
>>> c.co_consts, c.co_varnames, c.co_names
([None, 42], ['x'], ['y', 'z'])
The one exception to this automatic addition feature is that opcodes referring to "free" or "cell" variables will not automatically add new names, because the names need to be defined first:
>>> c.LOAD_DEREF('q')
Traceback (most recent call last):
...
NameError: ('Undefined free or cell var', 'q')
In general, opcode methods take the same arguments as their Python bytecode equivalent. But there are a few special cases.
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)
<... '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}
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
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 (!=)
Typical real-life code generation use cases call for transforming tree-like
data structures into bytecode, rather than linearly outputting instructions.
Code objects provide for this using a simple but high-level transformation
API.
Code objects may be called, passing in one or more arguments. Each
argument will have bytecode generated for it, according to its type:
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, long(2), 3.0, 4j+5, "6", unicode("7"), False, None, c.code())
>>> dis(c.code())
0 0 LOAD_CONST 1 (1)
3 LOAD_CONST 2 (2...)
6 LOAD_CONST 3 (3.0)
9 LOAD_CONST 4 ((5+4j))
12 LOAD_CONST 5 ('6')
15 LOAD_CONST 6 (...'7')
18 LOAD_CONST 7 (False)
21 LOAD_CONST 0 (None)
24 LOAD_CONST 8 (<code object <lambda>...>)
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) # equal, but different types
>>> dis(c.code())
0 0 LOAD_CONST 1 (1)
3 LOAD_CONST 2 (True)
6 LOAD_CONST 3 (1.0)
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
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
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
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)
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'>)
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
>>> 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
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
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)
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
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!)
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
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)
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]
The const_value() function can be used to check if an expression tree has
a constant value, and to obtain that value. Simple constants are returned
as-is:
>>> from peak.util.assembler import const_value
>>> simple_values = [1, long(2), 3.0, 4j+5, "6", unicode("7"), False, None, c.code()]
>>> list(map(const_value, simple_values))
[1, 2..., 3.0, (5+4j), '6', ...'7', False, None, <code object <lambda>...>]
Values wrapped in a Const() are also returned as-is:
>>> list(map(const_value, map(Const, simple_values))) [1, 2..., 3.0, (5+4j), '6', ...'7', False, None, <code object <lambda>...>]
But no other node types produce constant values; instead, NotAConstant is
raised:
>>> const_value(Local('x'))
Traceback (most recent call last):
...
peak.util.assembler.NotAConstant: Local('x')
Tuples of constants are recursively replaced by constant tuples:
>>> const_value( (1,2) ) (1, 2) >>> const_value( (1, (2, Const(3))) ) (1, (2, 3))
But any non-constant values anywhere in the structure cause an error:
>>> const_value( (1,Global('y')) )
Traceback (most recent call last):
...
peak.util.assembler.NotAConstant: Global('y')
As do any types not previously described here:
>>> const_value([1,2]) Traceback (most recent call last): ... peak.util.assembler.NotAConstant: [1, 2]
Unless of course they're wrapped with Const:
>>> const_value(Const([1,2])) [1, 2]
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(<... '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(<... '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 (<... 'int'>)
>>> c = Code()
>>> c( Call(Const(type), [1], fold=False) )
>>> dis(c.code())
0 0 LOAD_CONST 1 (<... '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.
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
Code generation is extensible: you can use any callable as a code-generation target. It will be called with exactly one argument: the code object. It can then perform whatever operations are desired.
In the most trivial case, you can use any unbound Code method as a code
generation target, e.g.:
>>> c = Code()
>>> c.LOAD_GLOBAL('foo')
>>> c(Call(Code.DUP_TOP, ()))
>>> dis(c.code())
0 0 LOAD_GLOBAL 0 (foo)
3 DUP_TOP
4 CALL_FUNCTION 0
As you can see, the Code.DUP_TOP() is called on the code instance, causing
a DUP_TOP opcode to be output. This is sometimes a handy trick for
accessing values that are already on the stack. More commonly, however, you'll
want to implement more sophisticated callables.
To make it easy to create diverse target types, a nodetype() decorator is
provided:
>>> from peak.util.assembler import nodetype
It allows you to create code generation target types using functions. Your
function should take one or more arguments, with a code=None optional
argument in the last position. It should check whether code is None when
called, and if so, return a tuple of the preceding arguments. If code
is not None, then it should do whatever code generating tasks are required.
For example:
>>> def TryFinally(block1, block2, code=None): ... if code is None: ... return block1, block2 ... code( ... Code.SETUP_FINALLY, ... block1, ... Code.POP_BLOCK, ... block2, ... Code.END_FINALLY ... ) >>> TryFinally = nodetype()(TryFinally)
Note: although the nodetype() generator can be used above the function definition in either Python 2.3 or 2.4, it cannot be done in a doctest under Python 2.3, so this document doesn't attempt to demonstrate that. Under 2.4, you would do something like this:
@nodetype() def TryFinally(...):
and code that needs to also work under 2.3 should do something like this:
nodetype() def TryFinally(...):
But to keep the examples here working with doctest, we'll be doing our
nodetype() calls after the end of the function definitions, e.g.:
>>> def ExprStmt(value, code=None):
... if code is None:
... return value,
... code( value, Code.POP_TOP )
>>> ExprStmt = nodetype()(ExprStmt)
>>> 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
The nodetype() decorator is virtually identical to the struct()
decorator in the DecoratorTools package, except that it does not support
*args, does not create a field for the code argument, and generates a
__call__() method that reinvokes the wrapped function to do the actual
code generation.
Among the benefits of this decorator are:
It gives your node types a great debugging format:
>>> tf = TryFinally(ExprStmt(1), ExprStmt(2)) >>> tf TryFinally(ExprStmt(1), ExprStmt(2))
It makes named fields accessible:
>>> tf.block1 ExprStmt(1) >>> tf.block2 ExprStmt(2)
Hashing and comparison work as expected (handy for algorithms that require comparing or caching AST subtrees, such as common subexpression elimination):
>>> ExprStmt(1) == ExprStmt(1) True >>> ExprStmt(1) == ExprStmt(2) False
Please see the struct decorator documentation for info on how to customize node types further.
Note: hashing only works if all the values you return in your argument tuple
are hashable, so you should try to convert them if possible. For example, if
an argument accepts any sequence, you should probably convert it to a tuple
before returning it. Most of the examples in this document, and the node types
supplied by peak.util.assembler itself do this.
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__')) <... '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
The simplest way to set up the calling signature for a Code instance is
to clone an existing function or code object's signature, using the
Code.from_function() or Code.from_code() classmethods. These methods
create a new Code instance whose calling signature (number and names of
arguments) matches that of the original function or code objects:
>>> def f1(a,b,*c,**d): ... pass >>> c = Code.from_function(f1) >>> f2 = function(c.code(), globals()) >>> import inspect >>> tuple(inspect.getargspec(f1)) (['a', 'b'], 'c', 'd', None) >>> tuple(inspect.getargspec(f2)) (['a', 'b'], 'c', 'd', None)
Note that these constructors do not copy any actual code from the code
or function objects. They simply copy the signature, and, if you set the
copy_lineno keyword argument to a true value, they will also set the
created code object's co_firstlineno to match that of the original code or
function object:
>>> c1 = Code.from_function(f1, copy_lineno=True) >>> c1.co_firstlineno 1 >>> c1.co_filename is f1.func_code.co_filename True
If you create a Code instance from a function that has nested positional
arguments, the returned code object will include a prologue to unpack the
arguments properly:
>>> def f3(a, (b,c), (d,(e,f))):
... pass
>>> f4 = function(Code.from_function(f3).code(), globals())
>>> dis(f4)
0 0 LOAD_FAST 1 (.1)
3 UNPACK_SEQUENCE 2
6 STORE_FAST 3 (b)
9 STORE_FAST 4 (c)
12 LOAD_FAST 2 (.2)
15 UNPACK_SEQUENCE 2
18 STORE_FAST 5 (d)
21 UNPACK_SEQUENCE 2
24 STORE_FAST 6 (e)
27 STORE_FAST 7 (f)
This is roughly the same code that Python would generate to do the same
unpacking process, and is designed so that the inspect module will
recognize it as an argument unpacking prologue:
>>> tuple(inspect.getargspec(f3)) (['a', ['b', 'c'], ['d', ['e', 'f']]], None, None, None) >>> tuple(inspect.getargspec(f4)) (['a', ['b', 'c'], ['d', ['e', 'f']]], None, None, None)
You can also use the from_spec(name='<lambda>', args=(), var=None, kw=None)
classmethod to explicitly set a name and argument spec for a new code object:
>>> c = Code.from_spec('a', ('b', ('c','d'), 'e'), 'f', 'g')
>>> c.co_name
'a'
>>> c.co_varnames
['b', '.1', 'e', 'f', 'g', 'c', 'd']
>>> c.co_argcount
3
>>> tuple(inspect.getargs(c.code()))
(['b', ['c', 'd'], 'e'], 'f', 'g')
Code instances have a variety of attributes corresponding to either the
attributes of the Python code objects they generate, or to the current state
of code generation.
For example, the co_argcount and co_varnames attributes
correspond to those used in creating the code for a Python function. If you
want your code to be a function, you can set them as follows:
>>> c = Code() >>> c.co_argcount = 3 >>> c.co_varnames = ['a','b','c'] >>> c.LOAD_CONST(42) >>> c.RETURN_VALUE() >>> f = function(c.code(), globals()) >>> f(1,2,3) 42 >>> import inspect >>> tuple(inspect.getargspec(f)) (['a', 'b', 'c'], None, None, None)
Although Python code objects want co_varnames to be a tuple, Code
instances use a list, so that names can be added during code generation. The
.code() method automatically creates tuples where necessary.
Here are all of the Code attributes you may want to read or write:
- co_filename
- A string representing the source filename for this code. If it's an actual
filename, then tracebacks that pass through the generated code will display
lines from the file. The default value is
'<generated code>'. - co_name
- The name of the function, class, or other block that this code represents.
The default value is
'<lambda>'. - co_argcount
- Number of positional arguments a function accepts; defaults to 0
- co_varnames
- A list of strings naming the code's local variables, beginning with its
positional argument names, followed by its
*and**argument names, if applicable, followed by any other local variable names. These names are used by theLOAD_FASTandSTORE_FASTopcodes, and invoking the.LOAD_FAST(name)and.STORE_FAST(name)methods of a code object will automatically add the given name to this list, if it's not already present. - co_flags
The flags for the Python code object. This defaults to
CO_OPTIMIZED | CO_NEWLOCALS, which is the correct value for a function using "fast" locals. This value is automatically or-ed withCO_NOFREEwhen generating a code object, if theco_cellvarsandco_freevarsattributes are empty. And if you use theLOAD_NAME(),STORE_NAME(), orDELETE_NAME()methods, theCO_OPTIMIZEDbit is automatically reset, since these opcodes can only be used when the code is running with a real (i.e. not virtualized)locals()dictionary.If you need to change any other flag bits besides the above, you'll need to set or clear them manually. For your convenience, the
peak.util.assemblermodule exports all theCO_constants used by Python. For example, you can useCO_VARARGSandCO_VARKEYWORDSto indicate whether a function accepts*or**arguments, as long as you extend theco_varnameslist accordingly. (Assuming you don't have an existing function or code object with the desired signature, in which case you could just use thefrom_function()orfrom_code()classmethods instead of messing with these low-level attributes and flags.)- stack_size
The predicted height of the runtime value stack, as of the current opcode. Its value is automatically updated by most opcodes, but if you are doing something sufficiently tricky (as in the
Switchdemo, below) you may need to explicitly set it.The
stack_sizeautomatically becomesNoneafter any unconditional jump operations, such asJUMP_FORWARD,BREAK_LOOP, orRETURN_VALUE. When the stack size isNone, the only operations that can be performed are the resolving of forward references (which will set the stack size to what it was when the reference was created), or manually setting the stack size.- co_freevars
- A tuple of strings naming a function's "free" variables. Defaults to an empty tuple. A function's free variables are the variables it "inherits" from its surrounding scope. If you're going to use this, you should set it only once, before generating any code that references any free or cell variables.
- co_cellvars
- A tuple of strings naming a function's "cell" variables. Defaults to an empty tuple. A function's cell variables are the variables that are "inherited" by one or more of its nested functions. If you're going to use this, you should set it only once, before generating any code that references any free or cell variables.
These other attributes are automatically generated and maintained, so you'll probably never have a reason to change them:
- co_consts
- A list of constants used by the code; the first (zeroth?) constant is
always
None. Normally, this is automatically maintained; the.LOAD_CONST(value)method checks to see if the constant is already present in this list, and adds it if it is not there. - co_names
- A list of non-optimized or global variable names. It's automatically updated whenever you invoke a method to generate an opcode that uses such names.
- co_code
- A byte array containing the generated code. Don't mess with this.
- co_firstlineno
- The first line number of the generated code. It automatically gets set
if you call
.set_lineno()before generating any code; otherwise it defaults to zero. - co_lnotab
- A byte array containing a generated line number table. It's automatically generated, so don't mess with it.
- co_stacksize
- The maximum amount of stack space the code will require to run. This
value is updated automatically as you generate code or change
the
stack_sizeattribute.
Code objects automatically track the predicted stack size as code is
generated, by updating the stack_size attribute as each operation occurs.
A history is kept so that backward jumps can be checked to ensure that the
current stack height is the same as at the jump's target. Similarly, when
forward jumps are resolved, the stack size at the jump target is checked
against the stack size at the jump's origin. If there are multiple jumps to
the same location, they must all have the same stack size at the origin and
the destination.
In addition, whenever any unconditional jump code is generated (i.e.
JUMP_FORWARD, BREAK_LOOP, CONTINUE_LOOP, JUMP_ABSOLUTE, or
RETURN_VALUE), the predicted stack_size is set to None. This
means that the Code object does not know what the stack size will be at
the current location. You cannot issue any instructions when the predicted
stack size is None, as you will receive an AssertionError:
>>> c = Code() >>> fwd = c.JUMP_FORWARD() >>> print(c.stack_size) # forward jump marks stack size as unknown None >>> c.LOAD_CONST(42) Traceback (most recent call last): ... AssertionError: Unknown stack size at this location
Instead, you must resolve a forward reference (or define a previously-jumped to label). This will propagate the stack size at the source of the jump to the current location, updating the stack size:
>>> fwd() >>> c.stack_size 0
Note, by the way, that this means it is impossible for you to generate static
"dead code". In other words, you cannot generate code that isn't reachable.
You should therefore check if stack_size is None before generating
code that might be unreachable. For example, consider this If
implementation:
>>> def If(cond, then, else_=Pass, code=None): ... if code is None: ... return cond, then, else_ ... else_clause = Label() ... end_if = Label() ... code(cond, else_clause.JUMP_IF_FALSE_OR_POP, then) ... code(end_if.JUMP_FORWARD, else_clause, Code.POP_TOP, else_) ... code(end_if) >>> If = nodetype()(If)
It works okay if there's no dead code:
>>> c = Code()
>>> c( 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)
But it breaks if you end the "then" block with a return:
>>> c = Code() >>> c( If(23, Return(42), 55) ) Traceback (most recent call last): ... AssertionError: Unknown stack size at this location
What we need is something like this instead:
>>> def If(cond, then, else_=Pass, code=None): ... if code is None: ... return cond, then, else_ ... else_clause = Label() ... end_if = Label() ... code(cond, else_clause.JUMP_IF_FALSE_OR_POP, then) ... if code.stack_size is not None: ... end_if.JUMP_FORWARD(code) ... code(else_clause, Code.POP_TOP, else_, end_if) >>> If = nodetype()(If)
As you can see, the dead code is now eliminated:
>>> c = Code()
>>> c( If(Local('a'), Return(42), 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)
The Python SETUP_FINALLY, SETUP_EXCEPT, and SETUP_LOOP opcodes
all create "blocks" that go on the frame's "block stack" at runtime. Each of
these opcodes must be matched with exactly one POP_BLOCK opcode -- no
more, and no less. Code objects enforce this using an internal block stack
that matches each setup with its corresponding POP_BLOCK. Trying to pop
a nonexistent block, or trying to generate code when unclosed blocks exist is
an error:
>>> c = Code() >>> c.POP_BLOCK() Traceback (most recent call last): ... AssertionError: Not currently in a block >>> c.SETUP_FINALLY() >>> c.code() Traceback (most recent call last): ... AssertionError: 1 unclosed block(s) >>> c.POP_BLOCK() >>> c.code() <code object <lambda>...>
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)
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 (<...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 (<...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
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
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
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
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(list), (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.
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]]
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.
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)
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 ...>)
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 ...>)
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 ...>)
... 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.)
Line number tracking:
>>> def simple_code(flno, slno, consts=1, ):
... c = Code()
... c.set_lineno(flno)
... for i in range(consts): c.LOAD_CONST(None)
... c.set_lineno(slno)
... c.RETURN_VALUE()
... return c.code()
>>> dis(simple_code(1,1))
1 0 LOAD_CONST 0 (None)
3 RETURN_VALUE
>>> simple_code(1,1).co_stacksize
1
>>> dis(simple_code(13,414))
13 0 LOAD_CONST 0 (None)
414 3 RETURN_VALUE
>>> dis(simple_code(13,14,100))
13 0 LOAD_CONST 0 (None)
3 LOAD_CONST 0 (None)
...
14 300 RETURN_VALUE
>>> simple_code(13,14,100).co_stacksize
100
>>> dis(simple_code(13,572,120))
13 0 LOAD_CONST 0 (None)
3 LOAD_CONST 0 (None)
...
572 360 RETURN_VALUE
Stack size tracking:
>>> c = Code() # 0 >>> c.LOAD_CONST(1) # 1 >>> c.POP_TOP() # 0 >>> c.LOAD_CONST(2) # 1 >>> c.LOAD_CONST(3) # 2 >>> c.co_stacksize 2 >>> c.stack_history [0, ..., 1, 0, ..., 1] >>> c.BINARY_ADD() # 1 >>> c.LOAD_CONST(4) # 2 >>> c.co_stacksize 2 >>> c.stack_history [0, ..., 1, 0, 1, ..., 2, ..., 1] >>> c.LOAD_CONST(5) >>> c.LOAD_CONST(6) >>> c.co_stacksize 4 >>> c.POP_TOP() >>> c.stack_size 3 >>> c.BUILD_CLASS() >>> c.stack_size 1
Stack underflow detection/recovery, and global/local variable names:
>>> c = Code()
>>> c.LOAD_GLOBAL('foo')
>>> c.stack_size
1
>>> c.STORE_ATTR('bar') # drops stack by 2
Traceback (most recent call last):
...
AssertionError: Stack underflow
>>> c.co_names # 'bar' isn't added unless success
['foo']
>>> c.LOAD_ATTR('bar')
>>> c.co_names
['foo', 'bar']
>>> c.DELETE_FAST('baz')
>>> c.co_varnames
['baz']
>>> dis(c.code())
0 0 LOAD_GLOBAL 0 (foo)
3 LOAD_ATTR 1 (bar)
6 DELETE_FAST 0 (baz)
Code iteration:
>>> c.DUP_TOP() >>> c.return_(Code.POP_TOP) >>> list(c) == [ ... (0, op.LOAD_GLOBAL, 0), ... (3, op.LOAD_ATTR, 1), ... (6, op.DELETE_FAST, 0), ... (9, op.DUP_TOP, None), ... (10, op.POP_TOP, None), ... (11, op.RETURN_VALUE, None) ... ] True
Code patching:
>>> c = Code()
>>> c.LOAD_CONST(42)
>>> c.STORE_FAST('x')
>>> c.LOAD_FAST('x')
>>> c.DELETE_FAST('x')
>>> c.RETURN_VALUE()
>>> dis(c.code())
0 0 LOAD_CONST 1 (42)
3 STORE_FAST 0 (x)
6 LOAD_FAST 0 (x)
9 DELETE_FAST 0 (x)
12 RETURN_VALUE
>>> c.co_varnames
['x']
>>> c.co_varnames.append('y')
>>> c._patch(
... {op.LOAD_FAST: op.LOAD_FAST,
... op.STORE_FAST: op.STORE_FAST,
... op.DELETE_FAST: op.DELETE_FAST},
... {0: 1}
... )
>>> dis(c.code())
0 0 LOAD_CONST 1 (42)
3 STORE_FAST 1 (y)
6 LOAD_FAST 1 (y)
9 DELETE_FAST 1 (y)
12 RETURN_VALUE
>>> c._patch({op.RETURN_VALUE: op.POP_TOP})
>>> dis(c.code())
0 0 LOAD_CONST 1 (42)
3 STORE_FAST 1 (y)
6 LOAD_FAST 1 (y)
9 DELETE_FAST 1 (y)
12 POP_TOP
Converting locals to free/cell vars:
>>> c = Code()
>>> c.LOAD_CONST(42)
>>> c.STORE_FAST('x')
>>> c.LOAD_FAST('x')
>>> dis(c.code())
0 0 LOAD_CONST 1 (42)
3 STORE_FAST 0 (x)
6 LOAD_FAST 0 (x)
>>> c.co_freevars = 'y', 'x'
>>> c.co_cellvars = 'z',
>>> c._locals_to_cells()
>>> dis(c.code())
0 0 LOAD_CONST 1 (42)
3 STORE_DEREF 2 (x)
6 LOAD_DEREF 2 (x)
>>> c.DELETE_FAST('x')
>>> c._locals_to_cells()
Traceback (most recent call last):
...
AssertionError: Can't delete local 'x' used in nested scope
>>> c = Code()
>>> c.LOAD_CONST(42)
>>> c.STORE_FAST('x')
>>> c.LOAD_FAST('x')
>>> c.co_freevars
()
>>> c.makefree(['x'])
>>> c.co_freevars
('x',)
>>> dis(c.code())
0 0 LOAD_CONST 1 (42)
3 STORE_DEREF 0 (x)
6 LOAD_DEREF 0 (x)
>>> c = Code()
>>> c.LOAD_CONST(42)
>>> c.STORE_FAST('x')
>>> c.LOAD_FAST('x')
>>> c.makecells(['x'])
>>> c.co_freevars
()
>>> c.co_cellvars
('x',)
>>> dis(c.code())
0 0 LOAD_CONST 1 (42)
3 STORE_DEREF 0 (x)
6 LOAD_DEREF 0 (x)
>>> c = Code()
>>> c.LOAD_CONST(42)
>>> c.STORE_FAST('x')
>>> c.LOAD_FAST('x')
>>> c.makefree('x')
>>> c.makecells(['y'])
>>> c.co_freevars
('x',)
>>> c.co_cellvars
('y',)
>>> dis(c.code())
0 0 LOAD_CONST 1 (42)
3 STORE_DEREF 1 (x)
6 LOAD_DEREF 1 (x)
>>> c = Code()
>>> c.co_flags &= ~op.CO_OPTIMIZED
>>> c.makecells(['q'])
Traceback (most recent call last):
...
AssertionError: Can't use cellvars in unoptimized scope
Auto-free promotion with code parent:
>>> p = Code() >>> c = Code() >>> c.LOAD_FAST('x') >>> dis(c.code(p)) 0 0 LOAD_DEREF 0 (x) >>> p.co_cellvars ('x',)>>> p = Code() >>> c = Code.from_function(lambda x,y,z=2: None) >>> c.LOAD_FAST('x') >>> c.LOAD_FAST('y') >>> c.LOAD_FAST('z')>>> dis(c.code(p)) 0 0 LOAD_FAST 0 (x) 3 LOAD_FAST 1 (y) 6 LOAD_FAST 2 (z) >>> p.co_cellvars ()>>> c.LOAD_FAST('q') >>> dis(c.code(p)) 0 0 LOAD_FAST 0 (x) 3 LOAD_FAST 1 (y) 6 LOAD_FAST 2 (z) 9 LOAD_DEREF 0 (q) >>> p.co_cellvars ('q',)>>> p = Code() >>> c = Code.from_function(lambda x,*y,**z: None) >>> c.LOAD_FAST('q') >>> c.LOAD_FAST('x') >>> c.LOAD_FAST('y') >>> c.LOAD_FAST('z') >>> dis(c.code(p)) 0 0 LOAD_DEREF 0 (q) 3 LOAD_FAST 0 (x) 6 LOAD_FAST 1 (y) 9 LOAD_FAST 2 (z) >>> p.co_cellvars ('q',)>>> p = Code() >>> c = Code.from_function(lambda x,*y: None) >>> c.LOAD_FAST('x') >>> c.LOAD_FAST('y') >>> c.LOAD_FAST('z') >>> dis(c.code(p)) 0 0 LOAD_FAST 0 (x) 3 LOAD_FAST 1 (y) 6 LOAD_DEREF 0 (z) >>> p.co_cellvars ('z',)>>> p = Code() >>> c = Code.from_function(lambda x,**y: None) >>> c.LOAD_FAST('x') >>> c.LOAD_FAST('y') >>> c.LOAD_FAST('z') >>> dis(c.code(p)) 0 0 LOAD_FAST 0 (x) 3 LOAD_FAST 1 (y) 6 LOAD_DEREF 0 (z) >>> p.co_cellvars ('z',)
Stack tracking on jumps:
>>> c = Code()
>>> else_ = Label()
>>> end = Label()
>>> c(99, else_.JUMP_IF_TRUE_OR_POP, end.JUMP_FORWARD)
>>> c(else_, Code.POP_TOP, end)
>>> dump(c.code())
LOAD_CONST 1 (99)
JUMP_IF_TRUE L1
POP_TOP
JUMP_FORWARD L2
L1: POP_TOP
>>> c.stack_size
0
>>> if sys.version>='2.7':
... print(c.stack_history == [0, 1, 1, 1, 0, 0, 0, None, None, 1])
... else:
... print(c.stack_history == [0, 1, 1, 1, 1, 1, 1, 0, None, None, 1])
True
>>> c = Code()
>>> fwd = c.JUMP_FORWARD()
>>> c.LOAD_CONST(42) # forward jump marks stack size unknown
Traceback (most recent call last):
...
AssertionError: Unknown stack size at this location
>>> c.stack_size = 0
>>> c.LOAD_CONST(42)
>>> fwd()
Traceback (most recent call last):
...
AssertionError: Stack level mismatch: actual=1 expected=0
>>> from peak.util.assembler import For
>>> c = Code()
>>> c(For((), Code.POP_TOP, Pass))
>>> c.return_()
>>> dump(c.code())
BUILD_TUPLE 0
GET_ITER
L1: FOR_ITER L2
POP_TOP
JUMP_ABSOLUTE L1
L2: LOAD_CONST 0 (None)
RETURN_VALUE
>>> c.stack_history
[0, 1, 1, 1, 1, 2, 2, 2, 1, None, None, 0, 1, 1, 1]
Yield value:
>>> import sys >>> from peak.util.assembler import CO_GENERATOR >>> c = Code() >>> c.co_flags & CO_GENERATOR 0 >>> c(42, Code.YIELD_VALUE) >>> c.stack_size == int(sys.version>='2.5') True >>> (c.co_flags & CO_GENERATOR) == CO_GENERATOR True
Sequence operators and stack tracking:
Function calls and raise:
>>> c = Code()
>>> c.LOAD_GLOBAL('locals')
>>> c.CALL_FUNCTION() # argc/kwargc default to 0
>>> c.POP_TOP()
>>> c.LOAD_GLOBAL('foo')
>>> c.LOAD_CONST(1)
>>> c.LOAD_CONST('x')
>>> c.LOAD_CONST(2)
>>> c.CALL_FUNCTION(1,1) # argc, kwargc
>>> c.POP_TOP()
>>> dis(c.code())
0 0 LOAD_GLOBAL 0 (locals)
3 CALL_FUNCTION 0
6 POP_TOP
7 LOAD_GLOBAL 1 (foo)
10 LOAD_CONST 1 (1)
13 LOAD_CONST 2 ('x')
16 LOAD_CONST 3 (2)
19 CALL_FUNCTION 257
22 POP_TOP
>>> c = Code()
>>> c.LOAD_GLOBAL('foo')
>>> c.LOAD_CONST(1)
>>> c.LOAD_CONST('x')
>>> c.LOAD_CONST(2)
>>> c.BUILD_MAP(0)
>>> c.stack_size
5
>>> c.CALL_FUNCTION_KW(1,1)
>>> c.POP_TOP()
>>> c.stack_size
0
>>> c = Code()
>>> c.LOAD_GLOBAL('foo')
>>> c.LOAD_CONST(1)
>>> c.LOAD_CONST('x')
>>> c.LOAD_CONST(1)
>>> c.BUILD_TUPLE(1)
>>> c.CALL_FUNCTION_VAR(0,1)
>>> c.POP_TOP()
>>> c.stack_size
0
>>> c = Code()
>>> c.LOAD_GLOBAL('foo')
>>> c.LOAD_CONST(1)
>>> c.LOAD_CONST('x')
>>> c.LOAD_CONST(1)
>>> c.BUILD_TUPLE(1)
>>> c.BUILD_MAP(0)
>>> c.CALL_FUNCTION_VAR_KW(0,1)
>>> c.POP_TOP()
>>> c.stack_size
0
>>> c = Code()
>>> c.RAISE_VARARGS(0)
>>> c.RAISE_VARARGS(1)
Traceback (most recent call last):
...
AssertionError: Stack underflow
>>> c.LOAD_CONST(1)
>>> c.RAISE_VARARGS(1)
>>> dis(c.code())
0 0 RAISE_VARARGS 0
3 LOAD_CONST 1 (1)
6 RAISE_VARARGS 1
Sequence building, unpacking, dup'ing:
>>> c = Code()
>>> c.LOAD_CONST(1)
>>> c.LOAD_CONST(2)
>>> c.BUILD_TUPLE(3)
Traceback (most recent call last):
...
AssertionError: Stack underflow
>>> c.BUILD_LIST(3)
Traceback (most recent call last):
...
AssertionError: Stack underflow
>>> c.BUILD_TUPLE(2)
>>> c.stack_size
1
>>> c.UNPACK_SEQUENCE(2)
>>> c.stack_size
2
>>> c.DUP_TOPX(3)
Traceback (most recent call last):
...
AssertionError: Stack underflow
>>> c.DUP_TOPX(2)
>>> c.stack_size
4
>>> c.LOAD_CONST(3)
>>> c.BUILD_LIST(5)
>>> c.stack_size
1
>>> c.UNPACK_SEQUENCE(5)
>>> c.BUILD_SLICE(3)
>>> c.stack_size
3
>>> c.BUILD_SLICE(3)
>>> c.stack_size
1
>>> c.BUILD_SLICE(2)
Traceback (most recent call last):
...
AssertionError: Stack underflow
>>> dis(c.code())
0 0 LOAD_CONST 1 (1)
3 LOAD_CONST 2 (2)
6 BUILD_TUPLE 2
9 UNPACK_SEQUENCE 2
12 DUP_TOPX 2
15 LOAD_CONST 3 (3)
18 BUILD_LIST 5
21 UNPACK_SEQUENCE 5
24 BUILD_SLICE 3
27 BUILD_SLICE 3
Stack levels for MAKE_FUNCTION/MAKE_CLOSURE:
>>> c = Code() >>> c.MAKE_FUNCTION(0) Traceback (most recent call last): ... AssertionError: Stack underflow >>> c.LOAD_CONST(1) >>> c.LOAD_CONST(2) # simulate being a function >>> c.MAKE_FUNCTION(1) >>> c.stack_size 1 >>> c = Code() >>> c.MAKE_CLOSURE(0, 0) Traceback (most recent call last): ... AssertionError: Stack underflow >>> c = Code() >>> c.LOAD_CONST(1) # closure >>> if sys.version>='2.5': c.BUILD_TUPLE(1) >>> c.LOAD_CONST(2) # default >>> c.LOAD_CONST(3) # simulate being a function >>> c.MAKE_CLOSURE(1, 1) >>> c.stack_size 1 >>> c = Code() >>> c.LOAD_CONST(1) >>> c.LOAD_CONST(2) >>> if sys.version>='2.5': c.BUILD_TUPLE(2) >>> c.LOAD_CONST(3) # simulate being a function >>> c.MAKE_CLOSURE(0, 2) >>> c.stack_size 1
Labels and backpatching forward references:
>>> c = Code() >>> where = c.here() >>> c.LOAD_CONST(1) >>> c.JUMP_FORWARD(where) Traceback (most recent call last): ... AssertionError: Relative jumps can't go backwards
"Call" combinations:
>>> c = Code()
>>> c.set_lineno(1)
>>> c(Call(Global('foo'), [Local('q')],
... [('x',Const(1))], Local('starargs'))
... )
>>> c.RETURN_VALUE()
>>> dis(c.code())
1 0 LOAD_GLOBAL 0 (foo)
3 LOAD_FAST 0 (q)
6 LOAD_CONST 1 ('x')
9 LOAD_CONST 2 (1)
12 LOAD_FAST 1 (starargs)
15 CALL_FUNCTION_VAR 257
18 RETURN_VALUE
>>> c = Code()
>>> c.set_lineno(1)
>>> c(Call(Global('foo'), [Local('q')], [('x',Const(1))],
... None, Local('kwargs'))
... )
>>> c.RETURN_VALUE()
>>> dis(c.code())
1 0 LOAD_GLOBAL 0 (foo)
3 LOAD_FAST 0 (q)
6 LOAD_CONST 1 ('x')
9 LOAD_CONST 2 (1)
12 LOAD_FAST 1 (kwargs)
15 CALL_FUNCTION_KW 257
18 RETURN_VALUE
Cloning:
>>> c = Code.from_function(lambda (x,y):1, True)
>>> dis(c.code())
1 0 LOAD_FAST 0 (.0)
3 UNPACK_SEQUENCE 2
6 STORE_FAST 1 (x)
9 STORE_FAST 2 (y)
>>> c = Code.from_function(lambda x,(y,(z,a,b)):1, True)
>>> dis(c.code())
1 0 LOAD_FAST 1 (.1)
3 UNPACK_SEQUENCE 2
6 STORE_FAST 2 (y)
9 UNPACK_SEQUENCE 3
12 STORE_FAST 3 (z)
15 STORE_FAST 4 (a)
18 STORE_FAST 5 (b)
Constant folding for *args and **kw:
>>> c = Code()
>>> c.return_(Call(Const(type), [], [], (1,)))
>>> dis(c.code())
0 0 LOAD_CONST 1 (<... 'int'>)
3 RETURN_VALUE
>>> c = Code()
>>> c.return_(Call(Const(dict), [], [], [], Const({'x':1})))
>>> dis(c.code())
0 0 LOAD_CONST 1 ({'x': 1})
3 RETURN_VALUE
Try/Except stack level tracking:
>>> def class_or_type_of(expr):
... return Suite([expr, TryExcept(
... Suite([Getattr(Code.DUP_TOP, '__class__'), Code.ROT_TWO]),
... [(Const(AttributeError), Call(Const(type), (Code.ROT_TWO,)))]
... )])
>>> def type_or_class(x): pass
>>> c = Code.from_function(type_or_class)
>>> c.return_(class_or_type_of(Local('x')))
>>> dump(c.code())
LOAD_FAST 0 (x)
SETUP_EXCEPT L1
DUP_TOP
LOAD_ATTR 0 (__class__)
ROT_TWO
POP_BLOCK
JUMP_FORWARD L3
L1: DUP_TOP
LOAD_CONST 1 (<...AttributeError...>)
COMPARE_OP 10 (exception match)
JUMP_IF_FALSE L2
POP_TOP
POP_TOP
POP_TOP
POP_TOP...
LOAD_CONST 2 (<... 'type'>)
ROT_TWO
CALL_FUNCTION 1
JUMP_FORWARD L3
L2: POP_TOP
END_FINALLY
L3: RETURN_VALUE
>>> type_or_class.__code__ = type_or_class.func_code = c.code()
>>> type_or_class(23)
<... 'int'>
Finally, to give an example of a creative way to abuse Python bytecode, here is an implementation of a simple "switch/case/else" structure:
>>> from peak.util.assembler import LOAD_CONST, POP_BLOCK
>>> import sys
>>> WHY_CONTINUE = {'2.3':5}.get(sys.version[:3], 32)
>>> def Switch(expr, cases, default=Pass, code=None):
... if code is None:
... return expr, tuple(cases), default
...
... d = {}
... else_block = Label()
... cleanup = Label()
... end_switch = Label()
...
... code(
... end_switch.SETUP_LOOP,
... Call(Const(d.get), [expr]),
... else_block.JUMP_IF_FALSE,
... WHY_CONTINUE, Code.END_FINALLY
... )
...
... cursize = code.stack_size - 1 # adjust for removed WHY_CONTINUE
... for key, value in cases:
... d[const_value(key)] = code.here()
... code.stack_size = cursize
... code(value)
... if code.stack_size is not None: # if the code can fall through,
... code(cleanup.JUMP_FORWARD) # jump forward to the cleanup
...
... code(
... else_block,
... Code.POP_TOP, default,
... cleanup,
... Code.POP_BLOCK,
... end_switch
... )
>>> Switch = nodetype()(Switch)
>>> c = Code()
>>> c.co_argcount=1
>>> c(Switch(Local('x'), [(1,Return(42)),(2,Return("foo"))], Return(27)))
>>> c.return_()
>>> f = function(c.code(), globals())
>>> f(1)
42
>>> f(2)
'foo'
>>> f(3)
27
>>> dump(c.code())
SETUP_LOOP L2
LOAD_CONST 1 (<...method ...get of ...>)
LOAD_FAST 0 (x)
CALL_FUNCTION 1
JUMP_IF_FALSE L1
LOAD_CONST 2 (...)
END_FINALLY
LOAD_CONST 3 (42)
RETURN_VALUE
LOAD_CONST 4 ('foo')
RETURN_VALUE
L1: POP_TOP
LOAD_CONST 5 (27)
RETURN_VALUE
POP_BLOCK
L2: LOAD_CONST 0 (None)
RETURN_VALUE
- Test NAME vs. FAST operators flag checks/sets
- Test code flags generation/cloning
- Exhaustive tests of all opcodes' stack history effects
- Test wide jumps and wide argument generation in general
- Remove/renumber local variables when a local is converted to free/cell