File: PythonLangImpl4.html

package info (click to toggle)
llvm-py 0.6%2Bsvn105-2
links: PTS, VCS
area: main
in suites: wheezy
size: 2,060 kB
sloc: python: 3,844; ansic: 1,963; cpp: 508; pascal: 87; makefile: 9; sh: 1
file content (999 lines) | stat: -rw-r--r-- 34,142 bytes
parent folder | download | duplicates (3)
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
                      "http://www.w3.org/TR/html4/strict.dtd">

<html>
<head>
  <title>Kaleidoscope: Adding JIT and Optimizer Support</title>
  <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
  <meta name="author" content="Chris Lattner">
  <meta name="author" content="Max Shawabkeh">
  <link rel="stylesheet"
        href="http://www.llvm.org/docs/llvm.css"
        type="text/css">
</head>

<body>

<div class="doc_title">Kaleidoscope: Adding JIT and Optimizer Support</div>

<ul>
<li>
  <a href="http://www.llvm.org/docs/tutorial/index.html">
    Up to Tutorial Index
  </a>
</li>
<li>Chapter 4
  <ol>
    <li><a href="#intro">Chapter 4 Introduction</a></li>
    <li><a href="#trivialconstfold">Trivial Constant Folding</a></li>
    <li><a href="#optimizerpasses">LLVM Optimization Passes</a></li>
    <li><a href="#jit">Adding a JIT Compiler</a></li>
    <li><a href="#code">Full Code Listing</a></li>
  </ol>
</li>
<li><a href="PythonLangImpl5.html">Chapter 5</a>: Extending the Language:
Control Flow</li>
</ul>

<div class="doc_author">
  <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a>
    and <a href="http://max99x.com">Max Shawabkeh</a>
  </p>
</div>

<!-- *********************************************************************** -->
<div class="doc_section"><a name="intro">Chapter 4 Introduction</a></div>
<!-- *********************************************************************** -->

<div class="doc_text">

<p>Welcome to Chapter 4 of the
"<a href="http://www.llvm.org/docs/tutorial/index.html">Implementing a language
with LLVM</a>" tutorial.  Chapters 1-3 described the implementation of a simple
language and added support for generating LLVM IR.  This chapter describes
two new techniques: adding optimizer support to your language, and adding JIT
compiler support.  These additions will demonstrate how to get nice, efficient
code for the Kaleidoscope language.</p>

</div>

<!-- *********************************************************************** -->
<div class="doc_section"><a name="trivialconstfold">Trivial Constant
Folding</a></div>
<!-- *********************************************************************** -->

<div class="doc_text">

<p>
Our demonstration for Chapter 3 is elegant and easy to extend.  Unfortunately,
it does not produce wonderful code.  The LLVM Builder, however, does give us
obvious optimizations when compiling simple code:</p>

<div class="doc_code">
<pre>
ready&gt; <b>def test(x) 1+2+x</b>
Read function definition:
define double @test(double %x) {
entry:
        %addtmp = fadd double 3.000000e+00, %x
        ret double %addtmp
}
</pre>
</div>

<p>This code is not a literal transcription of the AST built by parsing the
input. That would be:

<div class="doc_code">
<pre>
ready&gt; <b>def test(x) 1+2+x</b>
Read function definition:
define double @test(double %x) {
entry:
        %addtmp = fadd double 2.000000e+00, 1.000000e+00
        %addtmp1 = fadd double %addtmp, %x
        ret double %addtmp1
}
</pre>
</div>

<p>Constant folding, as seen above, in particular, is a very common and very
important optimization: so much so that many language implementors implement
constant folding support in their AST representation.</p>

<p>With LLVM, you don't need this support in the AST.  Since all calls to build
LLVM IR go through the LLVM IR builder, the builder itself checked to see if
there was a constant folding opportunity when you call it.  If so, it just does
the constant fold and return the constant instead of creating an instruction.

<p>Well, that was easy :).  In practice, we recommend always using
<tt>llvm.core.Builder</tt> when generating code like this.  It has no
"syntactic overhead" for its use (you don't have to uglify your compiler with
constant checks everywhere) and it can dramatically reduce the amount of
LLVM IR that is generated in some cases (particular for languages with a macro
preprocessor or that use a lot of constants).</p>

<p>On the other hand, the <tt>Builder</tt> is limited by the fact that it does
all of its analysis inline with the code as it is built.  If you take a slightly
more complex example:</p>

<div class="doc_code">
<pre>
ready&gt; <b>def test(x) (1+2+x)*(x+(1+2))</b>
Read a function definition:
define double @test(double %x) {
entry:
  %addtmp = fadd double 3.000000e+00, %x          ; &lt;double&gt; [#uses=1]
  %addtmp1 = fadd double %x, 3.000000e+00         ; &lt;double&gt; [#uses=1]
  %multmp = fmul double %addtmp, %addtmp1         ; &lt;double&gt; [#uses=1]
  ret double %multmp
}
</pre>
</div>

<p>In this case, the LHS and RHS of the multiplication are the same value.  We'd
really like to see this generate "<tt>tmp = x+3; result = tmp*tmp;</tt>" instead
of computing "<tt>x+3</tt>" twice.</p>

<p>Unfortunately, no amount of local analysis will be able to detect and correct
this.  This requires two transformations: reassociation of expressions (to
make the add's lexically identical) and Common Subexpression Elimination (CSE)
to  delete the redundant add instruction.  Fortunately, LLVM provides a broad
range of optimizations that you can use, in the form of "passes".</p>

</div>

<!-- *********************************************************************** -->
<div class="doc_section"><a name="optimizerpasses">LLVM Optimization
 Passes</a></div>
<!-- *********************************************************************** -->

<div class="doc_text">

<p>LLVM provides many optimization passes, which do many different sorts of
things and have different tradeoffs.  Unlike other systems, LLVM doesn't hold
to the mistaken notion that one set of optimizations is right for all languages
and for all situations.  LLVM allows a compiler implementor to make complete
decisions about what optimizations to use, in which order, and in what
situation.</p>

<p>As a concrete example, LLVM supports both "whole module" passes, which look
across as large of body of code as they can (often a whole file, but if run
at link time, this can be a substantial portion of the whole program).  It also
supports and includes "per-function" passes which just operate on a single
function at a time, without looking at other functions.  For more information
on passes and how they are run, see the
<a href="http://www.llvm.org/docs/WritingAnLLVMPass.html">How to Write a
Pass</a> document and the <a href="http://www.llvm.org/docs/Passes.html">List of
LLVM Passes</a>.</p>

<p>For Kaleidoscope, we are currently generating functions on the fly, one at
a time, as the user types them in.  We aren't shooting for the ultimate
optimization experience in this setting, but we also want to catch the easy and
quick stuff where possible.  As such, we will choose to run a few per-function
optimizations as the user types the function in.  If we wanted to make a "static
Kaleidoscope compiler", we would use exactly the code we have now, except that
we would defer running the optimizer until the entire file has been parsed.</p>

<p>In order to get per-function optimizations going, we need to set up a
<a href="http://www.llvm.org/docs/WritingAnLLVMPass.html#passmanager">
FunctionPassManager</a> to hold and organize the LLVM optimizations that we want
to run.  Once we have that, we can add a set of optimizations to run.  The code
looks like this:</p>

<div class="doc_code">
<pre>
# The function optimization passes manager.
g_llvm_pass_manager = FunctionPassManager.new(g_llvm_module)

# The LLVM execution engine.
g_llvm_executor = ExecutionEngine.new(g_llvm_module)

...

def main():
  # Set up the optimizer pipeline. Start with registering info about how the
  # target lays out data structures.
  g_llvm_pass_manager.add(g_llvm_executor.target_data)
  # Do simple "peephole" optimizations and bit-twiddling optzns.
  g_llvm_pass_manager.add(PASS_INSTRUCTION_COMBINING)
  # Reassociate expressions.
  g_llvm_pass_manager.add(PASS_REASSOCIATE)
  # Eliminate Common SubExpressions.
  g_llvm_pass_manager.add(PASS_GVN)
  # Simplify the control flow graph (deleting unreachable blocks, etc).
  g_llvm_pass_manager.add(PASS_CFG_SIMPLIFICATION)

  g_llvm_pass_manager.initialize()
</pre>
</div>

<p>This code defines a <tt>FunctionPassManager</tt>,
<tt>g_llvm_pass_manager</tt>.  Once it is set up, we use a series of "add" calls
to add a bunch of LLVM passes.  The first pass is basically boilerplate, it adds
a pass so that later optimizations know how the data structures in the program
are laid out.  (The "<tt>g_llvm_executor</tt>" variable is related to the JIT,
which we will get to in the next section.)  In this case, we choose to add 4
optimization passes.  The passes we chose here are a pretty standard set of
"cleanup" optimizations that are useful for a wide variety of code.  I won't
delve into what they do but, believe me, they are a good starting place :).</p>

<p>Once the pass manager is set up, we need to make use of it.  We do this by
running it after our newly created function is constructed (in
<tt>FunctionNode.CodeGen</tt>), but before it is returned to the client:</p>

<div class="doc_code">
<pre>
      return_value = self.body.CodeGen()
      g_llvm_builder.ret(return_value)

      # Validate the generated code, checking for consistency.
      function.verify()

      <b># Optimize the function.
      g_llvm_pass_manager.run(function)</b>
</pre>
</div>

<p>As you can see, this is pretty straightforward.  The
<tt>FunctionPassManager</tt> optimizes and updates the LLVM Function in place,
improving (hopefully) its body.  With this in place, we can try our test above
again:</p>

<div class="doc_code">
<pre>
ready&gt; <b>def test(x) (1+2+x)*(x+(1+2))</b>
Read a function definition:
define double @test(double %x) {
entry:
  %addtmp = fadd double %x, 3.000000e+00          ; &lt;double&gt; [#uses=2]
  %multmp = fmul double %addtmp, %addtmp          ; &lt;double&gt; [#uses=1]
  ret double %multmp
}
</pre>
</div>

<p>As expected, we now get our nicely optimized code, saving a floating point
add instruction from every execution of this function.</p>

<p>LLVM provides a wide variety of optimizations that can be used in certain
circumstances.  Some
<a href="http://www.llvm.org/docs/Passes.html">documentation about the various
passes</a> is available, but it isn't very complete.  Another good source of
ideas can come from looking at the passes that <tt>llvm-gcc</tt> or
<tt>llvm-ld</tt> run to get started.  The "<tt>opt</tt>" tool allows you to
experiment with passes from the command line, so you can see if they do
anything.</p>

<p>Now that we have reasonable code coming out of our front-end, lets talk about
executing it!</p>

</div>

<!-- *********************************************************************** -->
<div class="doc_section"><a name="jit">Adding a JIT Compiler</a></div>
<!-- *********************************************************************** -->

<div class="doc_text">

<p>Code that is available in LLVM IR can have a wide variety of tools
applied to it.  For example, you can run optimizations on it (as we did above),
you can dump it out in textual or binary forms, you can compile the code to an
assembly file (.s) for some target, or you can JIT compile it.  The nice thing
about the LLVM IR representation is that it is the "common currency" between
many different parts of the compiler.
</p>

<p>In this section, we'll add JIT compiler support to our interpreter.  The
basic idea that we want for Kaleidoscope is to have the user enter function
bodies as they do now, but immediately evaluate the top-level expressions they
type in.  For example, if they type in "1 + 2", we should evaluate and print
out 3.  If they define a function, they should be able to call it from the
command line.</p>

<p>In order to do this, we first declare and initialize the JIT.  This is done
by adding and initializing a global variable:</p>

<div class="doc_code">
<pre>
# The LLVM execution engine.
g_llvm_executor = ExecutionEngine.new(g_llvm_module)
</pre>
</div>

<p>This creates an abstract "Execution Engine" which can be either a JIT
compiler or the LLVM interpreter.  LLVM will automatically pick a JIT compiler
for you if one is available for your platform, otherwise it will fall back to
the interpreter.</p>

<p>Once the <tt>ExecutionEngine</tt> is created, the JIT is ready to be used.
We can use the <tt>run_function</tt> method of the execution engine to execute
a compiled function and get its return value.  In our case, this means that we
can change the code that parses a top-level expression to look like this:</p>

<div class="doc_code">
<pre>
  def HandleTopLevelExpression(self):
    try:
      function = self.ParseTopLevelExpr().CodeGen()
      result = g_llvm_executor.run_function(function, [])
      print 'Evaluated to:', result.as_real(Type.double())
    except Exception, e:
      print 'Error:', e
      try:
        self.Next() # Skip for error recovery.
      except:
        pass
</pre>
</div>

<p>Recall that we compile top-level expressions into a self-contained LLVM
function that takes no arguments and returns the computed double.</p>

<p>With just these two changes, lets see how Kaleidoscope works now!</p>

<div class="doc_code">
<pre>
ready&gt; <b>4+5</b>
Read a top level expression:
define double @0() {
entry:
  ret double 9.000000e+00
}

Evaluated to: 9.0
</pre>
</div>

<p>Well this looks like it is basically working.  The dump of the function
shows the "no argument function that always returns double" that we synthesize
for each top-level expression that is typed in.  This demonstrates very basic
functionality, but can we do more?</p>

<div class="doc_code">
<pre>
ready&gt; <b>def testfunc(x y) x + y*2</b>
Read a function definition:
define double @testfunc(double %x, double %y) {
entry:
  %multmp = fmul double %y, 2.000000e+00          ; &lt;double&gt; [#uses=1]
  %addtmp = fadd double %multmp, %x               ; &lt;double&gt; [#uses=1]
  ret double %addtmp
}

ready&gt; <b>testfunc(4, 10)</b>
Read a top level expression:
define double @0() {
entry:
  %calltmp = call double @testfunc(double 4.000000e+00, double 1.000000e+01) ; &lt;double&gt; [#uses=1]
  ret double %calltmp
}

<em>Evaluated to: 24.0</em>
</pre>
</div>

<p>This illustrates that we can now call user code, but there is something a bit
subtle going on here.  Note that we only invoke the JIT on the anonymous
functions that <em>call testfunc</em>, but we never invoked it
on <em>testfunc</em> itself.  What actually happened here is that the JIT
scanned for all non-JIT'd functions transitively called from the anonymous
function and compiled all of them before returning from <tt>run_function()</tt>.
</p>

<p>The JIT provides a number of other more advanced interfaces for things like
freeing allocated machine code, rejit'ing functions to update them, etc.
However, even with this simple code, we get some surprisingly powerful
capabilities - check this out (I removed the dump of the anonymous functions,
you should get the idea by now :) :</p>

<div class="doc_code">
<pre>
ready&gt; <b>extern sin(x)</b>
Read an extern:
declare double @sin(double)

ready&gt; <b>extern cos(x)</b>
Read an extern:
declare double @cos(double)

ready&gt; <b>sin(1.0)</b>
<em>Evaluated to: 0.841470984808</em>

ready&gt; <b>def foo(x) sin(x)*sin(x) + cos(x)*cos(x)</b>
Read a function definition:
define double @foo(double %x) {
entry:
  %calltmp = call double @sin(double %x)          ; &lt;double&gt; [#uses=1]
  %calltmp1 = call double @sin(double %x)         ; &lt;double&gt; [#uses=1]
  %multmp = fmul double %calltmp, %calltmp1       ; &lt;double&gt; [#uses=1]
  %calltmp2 = call double @cos(double %x)         ; &lt;double&gt; [#uses=1]
  %calltmp3 = call double @cos(double %x)         ; &lt;double&gt; [#uses=1]
  %multmp4 = fmul double %calltmp2, %calltmp3     ; &lt;double&gt; [#uses=1]
  %addtmp = fadd double %multmp, %multmp4         ; &lt;double&gt; [#uses=1]
  ret double %addtmp
}

ready&gt; <b>foo(4.0)</b>
<em>Evaluated to: 1.000000</em>
</pre>
</div>

<p>Whoa, how does the JIT know about sin and cos?  The answer is surprisingly
simple: in this example, the JIT started execution of a function and got to a
function call.  It realized that the function was not yet JIT compiled and
invoked the standard set of routines to resolve the function.  In this case,
there is no body defined for the function, so the JIT ended up calling
"<tt>dlsym("sin")</tt>" on the Python process that is hosting our Kaleidoscope
prompt. Since "<tt>sin</tt>" is defined within the JIT's address space, it
simply patches up calls in the module to call the libm version of <tt>sin</tt>
directly.</p>

<p>One interesting application of this is that we can now extend the language
by writing arbitrary C++ code to implement operations.  For example, we can
create a C file with the following simple function:
</p>

<div class="doc_code">
<pre>
#include &lt;stdio.h&gt;

double putchard(double x) {
  putchar((char)x);
  return 0;
}
</pre>
</div>

<p>We can then compile this into a shared library with GCC:</p>

<div class="doc_code">
<pre>
gcc -shared -fPIC -o putchard.so  putchard.c
</pre>
</div>

<p>Now we can load this library into the Python process using
<tt>llvm.core.load_library_permanently</tt> and access it from Kaleidoscope to
produce simple output to the console:</p>

<div class="doc_code">
<pre>
>>> <b>import llvm.core</b>
>>> <b>llvm.core.load_library_permanently('/home/max/llvm-py-tutorial/putchard.so')</b>
>>> <b>import kaleidoscope</b>
>>> <b>kaleidoscope.main()</b>
ready&gt; <b>extern putchard(x)</b>
Read an extern:
declare double @putchard(double)

ready&gt; <b>putchard(65) + putchard(66) + putchard(67) + putchard(10)</b>
<em>ABC</em>
Evaluated to: 0.0
</pre>
</div>

<p>Similar code could be used to implement file I/O, console input, and many
other capabilities in Kaleidoscope.</p>

<p>This completes the JIT and optimizer chapter of the Kaleidoscope tutorial. At
this point, we can compile a non-Turing-complete programming language, optimize
and JIT compile it in a user-driven way.  Next up we'll look into <a
href="PythonLangImpl5.html">extending the language with control flow
constructs</a>, tackling some interesting LLVM IR issues along the way.</p>

</div>

<!-- *********************************************************************** -->
<div class="doc_section"><a name="code">Full Code Listing</a></div>
<!-- *********************************************************************** -->

<div class="doc_text">

<p>
Here is the complete code listing for our running example, enhanced with the
LLVM JIT and optimizer:
</p>

<div class="doc_code">
<pre>
#!/usr/bin/env python

import re
from llvm.core import Module, Constant, Type, Function, Builder, FCMP_ULT
from llvm.ee import ExecutionEngine, TargetData
from llvm.passes import FunctionPassManager
from llvm.passes import (PASS_INSTRUCTION_COMBINING,
                         PASS_REASSOCIATE,
                         PASS_GVN,
                         PASS_CFG_SIMPLIFICATION)

################################################################################
## Globals
################################################################################

# The LLVM module, which holds all the IR code.
g_llvm_module = Module.new('my cool jit')

# The LLVM instruction builder. Created whenever a new function is entered.
g_llvm_builder = None

# A dictionary that keeps track of which values are defined in the current scope
# and what their LLVM representation is.
g_named_values = {}

# The function optimization passes manager.
g_llvm_pass_manager = FunctionPassManager.new(g_llvm_module)

# The LLVM execution engine.
g_llvm_executor = ExecutionEngine.new(g_llvm_module)

################################################################################
## Lexer
################################################################################

# The lexer yields one of these types for each token.
class EOFToken(object):
  pass

class DefToken(object):
  pass

class ExternToken(object):
  pass

class IdentifierToken(object):
  def __init__(self, name): self.name = name

class NumberToken(object):
  def __init__(self, value): self.value = value

class CharacterToken(object):
  def __init__(self, char): self.char = char
  def __eq__(self, other):
    return isinstance(other, CharacterToken) and self.char == other.char
  def __ne__(self, other): return not self == other

# Regular expressions that tokens and comments of our language.
REGEX_NUMBER = re.compile('[0-9]+(?:\.[0-9]+)?')
REGEX_IDENTIFIER = re.compile('[a-zA-Z][a-zA-Z0-9]*')
REGEX_COMMENT = re.compile('#.*')

def Tokenize(string):
  while string:
    # Skip whitespace.
    if string[0].isspace():
      string = string[1:]
      continue

    # Run regexes.
    comment_match = REGEX_COMMENT.match(string)
    number_match = REGEX_NUMBER.match(string)
    identifier_match = REGEX_IDENTIFIER.match(string)

    # Check if any of the regexes matched and yield the appropriate result.
    if comment_match:
      comment = comment_match.group(0)
      string = string[len(comment):]
    elif number_match:
      number = number_match.group(0)
      yield NumberToken(float(number))
      string = string[len(number):]
    elif identifier_match:
      identifier = identifier_match.group(0)
      # Check if we matched a keyword.
      if identifier == 'def':
        yield DefToken()
      elif identifier == 'extern':
        yield ExternToken()
      else:
        yield IdentifierToken(identifier)
      string = string[len(identifier):]
    else:
      # Yield the ASCII value of the unknown character.
      yield CharacterToken(string[0])
      string = string[1:]

  yield EOFToken()

################################################################################
## Abstract Syntax Tree (aka Parse Tree)
################################################################################

# Base class for all expression nodes.
class ExpressionNode(object):
  pass

# Expression class for numeric literals like "1.0".
class NumberExpressionNode(ExpressionNode):

  def __init__(self, value):
    self.value = value

  def CodeGen(self):
    return Constant.real(Type.double(), self.value)

# Expression class for referencing a variable, like "a".
class VariableExpressionNode(ExpressionNode):

  def __init__(self, name):
    self.name = name

  def CodeGen(self):
    if self.name in g_named_values:
      return g_named_values[self.name]
    else:
      raise RuntimeError('Unknown variable name: ' + self.name)

# Expression class for a binary operator.
class BinaryOperatorExpressionNode(ExpressionNode):

  def __init__(self, operator, left, right):
    self.operator = operator
    self.left = left
    self.right = right

  def CodeGen(self):
    left = self.left.CodeGen()
    right = self.right.CodeGen()

    if self.operator == '+':
      return g_llvm_builder.fadd(left, right, 'addtmp')
    elif self.operator == '-':
      return g_llvm_builder.fsub(left, right, 'subtmp')
    elif self.operator == '*':
      return g_llvm_builder.fmul(left, right, 'multmp')
    elif self.operator == '&lt;':
      result = g_llvm_builder.fcmp(FCMP_ULT, left, right, 'cmptmp')
      # Convert bool 0 or 1 to double 0.0 or 1.0.
      return g_llvm_builder.uitofp(result, Type.double(), 'booltmp')
    else:
      raise RuntimeError('Unknown binary operator.')

# Expression class for function calls.
class CallExpressionNode(ExpressionNode):

  def __init__(self, callee, args):
    self.callee = callee
    self.args = args

  def CodeGen(self):
    # Look up the name in the global module table.
    callee = g_llvm_module.get_function_named(self.callee)

    # Check for argument mismatch error.
    if len(callee.args) != len(self.args):
      raise RuntimeError('Incorrect number of arguments passed.')

    arg_values = [i.CodeGen() for i in self.args]

    return g_llvm_builder.call(callee, arg_values, 'calltmp')

# This class represents the "prototype" for a function, which captures its name,
# and its argument names (thus implicitly the number of arguments the function
# takes).
class PrototypeNode(object):

  def __init__(self, name, args):
    self.name = name
    self.args = args

  def CodeGen(self):
    # Make the function type, eg. double(double,double).
    funct_type = Type.function(
      Type.double(), [Type.double()] * len(self.args), False)

    function = Function.new(g_llvm_module, funct_type, self.name)

    # If the name conflicted, there was already something with the same name.
    # If it has a body, don't allow redefinition or reextern.
    if function.name != self.name:
      function.delete()
      function = g_llvm_module.get_function_named(self.name)

      # If the function already has a body, reject this.
      if not function.is_declaration:
        raise RuntimeError('Redefinition of function.')

      # If F took a different number of args, reject.
      if len(callee.args) != len(self.args):
        raise RuntimeError('Redeclaration of a function with different number '
                           'of args.')

    # Set names for all arguments and add them to the variables symbol table.
    for arg, arg_name in zip(function.args, self.args):
      arg.name = arg_name
      # Add arguments to variable symbol table.
      g_named_values[arg_name] = arg

    return function

# This class represents a function definition itself.
class FunctionNode(object):

  def __init__(self, prototype, body):
    self.prototype = prototype
    self.body = body

  def CodeGen(self):
    # Clear scope.
    g_named_values.clear()

    # Create a function object.
    function = self.prototype.CodeGen()

    # Create a new basic block to start insertion into.
    block = function.append_basic_block('entry')
    global g_llvm_builder
    g_llvm_builder = Builder.new(block)

    # Finish off the function.
    try:
      return_value = self.body.CodeGen()
      g_llvm_builder.ret(return_value)

      # Validate the generated code, checking for consistency.
      function.verify()

      # Optimize the function.
      g_llvm_pass_manager.run(function)
    except:
      function.delete()
      raise

    return function


################################################################################
## Parser
################################################################################

class Parser(object):

  def __init__(self, tokens, binop_precedence):
    self.tokens = tokens
    self.binop_precedence = binop_precedence
    self.Next()

  # Provide a simple token buffer. Parser.current is the current token the
  # parser is looking at. Parser.Next() reads another token from the lexer and
  # updates Parser.current with its results.
  def Next(self):
    self.current = self.tokens.next()

  # Gets the precedence of the current token, or -1 if the token is not a binary
  # operator.
  def GetCurrentTokenPrecedence(self):
    if isinstance(self.current, CharacterToken):
      return self.binop_precedence.get(self.current.char, -1)
    else:
      return -1

  # identifierexpr ::= identifier | identifier '(' expression* ')'
  def ParseIdentifierExpr(self):
    identifier_name = self.current.name
    self.Next()  # eat identifier.

    if self.current != CharacterToken('('):  # Simple variable reference.
      return VariableExpressionNode(identifier_name)

    # Call.
    self.Next()  # eat '('.
    args = []
    if self.current != CharacterToken(')'):
      while True:
        args.append(self.ParseExpression())
        if self.current == CharacterToken(')'):
          break
        elif self.current != CharacterToken(','):
          raise RuntimeError('Expected ")" or "," in argument list.')
        self.Next()

    self.Next()  # eat ')'.
    return CallExpressionNode(identifier_name, args)

  # numberexpr ::= number
  def ParseNumberExpr(self):
    result = NumberExpressionNode(self.current.value)
    self.Next()  # consume the number.
    return result

  # parenexpr ::= '(' expression ')'
  def ParseParenExpr(self):
    self.Next()  # eat '('.

    contents = self.ParseExpression()

    if self.current != CharacterToken(')'):
      raise RuntimeError('Expected ")".')
    self.Next()  # eat ')'.

    return contents

  # primary ::= identifierexpr | numberexpr | parenexpr
  def ParsePrimary(self):
    if isinstance(self.current, IdentifierToken):
      return self.ParseIdentifierExpr()
    elif isinstance(self.current, NumberToken):
      return self.ParseNumberExpr()
    elif self.current == CharacterToken('('):
      return self.ParseParenExpr()
    else:
      raise RuntimeError('Unknown token when expecting an expression.')

  # binoprhs ::= (operator primary)*
  def ParseBinOpRHS(self, left, left_precedence):
    # If this is a binary operator, find its precedence.
    while True:
      precedence = self.GetCurrentTokenPrecedence()

      # If this is a binary operator that binds at least as tightly as the
      # current one, consume it; otherwise we are done.
      if precedence &lt; left_precedence:
        return left

      binary_operator = self.current.char
      self.Next()  # eat the operator.

      # Parse the primary expression after the binary operator.
      right = self.ParsePrimary()

      # If binary_operator binds less tightly with right than the operator after
      # right, let the pending operator take right as its left.
      next_precedence = self.GetCurrentTokenPrecedence()
      if precedence &lt; next_precedence:
        right = self.ParseBinOpRHS(right, precedence + 1)

      # Merge left/right.
      left = BinaryOperatorExpressionNode(binary_operator, left, right)

  # expression ::= primary binoprhs
  def ParseExpression(self):
    left = self.ParsePrimary()
    return self.ParseBinOpRHS(left, 0)

  # prototype ::= id '(' id* ')'
  def ParsePrototype(self):
    if not isinstance(self.current, IdentifierToken):
      raise RuntimeError('Expected function name in prototype.')

    function_name = self.current.name
    self.Next()  # eat function name.

    if self.current != CharacterToken('('):
      raise RuntimeError('Expected "(" in prototype.')
    self.Next()  # eat '('.

    arg_names = []
    while isinstance(self.current, IdentifierToken):
      arg_names.append(self.current.name)
      self.Next()

    if self.current != CharacterToken(')'):
      raise RuntimeError('Expected ")" in prototype.')

    # Success.
    self.Next()  # eat ')'.

    return PrototypeNode(function_name, arg_names)

  # definition ::= 'def' prototype expression
  def ParseDefinition(self):
    self.Next()  # eat def.
    proto = self.ParsePrototype()
    body = self.ParseExpression()
    return FunctionNode(proto, body)

  # toplevelexpr ::= expression
  def ParseTopLevelExpr(self):
    proto = PrototypeNode('', [])
    return FunctionNode(proto, self.ParseExpression())

  # external ::= 'extern' prototype
  def ParseExtern(self):
    self.Next()  # eat extern.
    return self.ParsePrototype()

  # Top-Level parsing
  def HandleDefinition(self):
    self.Handle(self.ParseDefinition, 'Read a function definition:')

  def HandleExtern(self):
    self.Handle(self.ParseExtern, 'Read an extern:')

  def HandleTopLevelExpression(self):
    try:
      function = self.ParseTopLevelExpr().CodeGen()
      result = g_llvm_executor.run_function(function, [])
      print 'Evaluated to:', result.as_real(Type.double())
    except Exception, e:
      print 'Error:', e
      try:
        self.Next() # Skip for error recovery.
      except:
        pass

  def Handle(self, function, message):
    try:
      print message, function().CodeGen()
    except Exception, e:
      print 'Error:', e
      try:
        self.Next() # Skip for error recovery.
      except:
        pass

################################################################################
## Main driver code.
################################################################################

def main():
  # Set up the optimizer pipeline. Start with registering info about how the
  # target lays out data structures.
  g_llvm_pass_manager.add(g_llvm_executor.target_data)
  # Do simple "peephole" optimizations and bit-twiddling optzns.
  g_llvm_pass_manager.add(PASS_INSTRUCTION_COMBINING)
  # Reassociate expressions.
  g_llvm_pass_manager.add(PASS_REASSOCIATE)
  # Eliminate Common SubExpressions.
  g_llvm_pass_manager.add(PASS_GVN)
  # Simplify the control flow graph (deleting unreachable blocks, etc).
  g_llvm_pass_manager.add(PASS_CFG_SIMPLIFICATION)

  g_llvm_pass_manager.initialize()

  # Install standard binary operators.
  # 1 is lowest possible precedence. 40 is the highest.
  operator_precedence = {
    '&lt;': 10,
    '+': 20,
    '-': 20,
    '*': 40
  }

  # Run the main "interpreter loop".
  while True:
    print 'ready&gt;',
    try:
      raw = raw_input()
    except KeyboardInterrupt:
      break

    parser = Parser(Tokenize(raw), operator_precedence)
    while True:
      # top ::= definition | external | expression | EOF
      if isinstance(parser.current, EOFToken):
        break
      if isinstance(parser.current, DefToken):
        parser.HandleDefinition()
      elif isinstance(parser.current, ExternToken):
        parser.HandleExtern()
      else:
        parser.HandleTopLevelExpression()

  # Print out all of the generated code.
  print '\n', g_llvm_module

if __name__ == '__main__':
  main()
</pre>
</div>

<a href="PythonLangImpl5.html">Next: Extending the language: control flow</a>
</div>

<!-- *********************************************************************** -->
<hr>
<address>
  <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
  src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
  <a href="http://validator.w3.org/check/referer"><img
  src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!"></a>

  <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
  <a href="http://max99x.com">Max Shawabkeh</a><br>
  <a href="http://llvm.org">The LLVM Compiler Infrastructure</a><br>
  Last modified: $Date$
</address>
</body>
</html>