docs/tutorial/LangImpl3.html

   1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
   2                       "http://www.w3.org/TR/html4/strict.dtd">
   3
   4 <html>
   5 <head>
   6   <title>Kaleidoscope: Implementing code generation to LLVM IR</title>
   7   <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
   8   <meta name="author" content="Chris Lattner">
   9   <link rel="stylesheet" href="../llvm.css" type="text/css">
  10 </head>
  11
  12 <body>
  13
  14 <div class="doc_title">Kaleidoscope: Code generation to LLVM IR</div>
  15
  16 <ul>
  17 <li><a href="index.html">Up to Tutorial Index</a></li>
  18 <li>Chapter 3
  19   <ol>
  20     <li><a href="#intro">Chapter 3 Introduction</a></li>
  21     <li><a href="#basics">Code Generation Setup</a></li>
  22     <li><a href="#exprs">Expression Code Generation</a></li>
  23     <li><a href="#funcs">Function Code Generation</a></li>
  24     <li><a href="#driver">Driver Changes and Closing Thoughts</a></li>
  25     <li><a href="#code">Full Code Listing</a></li>
  26   </ol>
  27 </li>
  28 <li><a href="LangImpl4.html">Chapter 4</a>: Adding JIT and Optimizer
  29 Support</li>
  30 </ul>
  31
  32 <div class="doc_author">
  33   <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a></p>
  34 </div>
  35
  36 <!-- *********************************************************************** -->
  37 <div class="doc_section"><a name="intro">Chapter 3 Introduction</a></div>
  38 <!-- *********************************************************************** -->
  39
  40 <div class="doc_text">
  41
  42 <p>Welcome to Chapter 3 of the "<a href="index.html">Implementing a language
  43 with LLVM</a>" tutorial.  This chapter shows you how to transform the <a
  44 href="LangImpl2.html">Abstract Syntax Tree</a>, built in Chapter 2, into LLVM IR.
  45 This will teach you a little bit about how LLVM does things, as well as
  46 demonstrate how easy it is to use.  It's much more work to build a lexer and
  47 parser than it is to generate LLVM IR code. :)
  48 </p>
  49
  50 <p><b>Please note</b>: the code in this chapter and later require LLVM 2.2 or
  51 later.  LLVM 2.1 and before will not work with it.  Also note that you need
  52 to use a version of this tutorial that matches your LLVM release: If you are
  53 using an official LLVM release, use the version of the documentation included
  54 with your release or on the <a href="http://llvm.org/releases/">llvm.org
  55 releases page</a>.</p>
  56
  57 </div>
  58
  59 <!-- *********************************************************************** -->
  60 <div class="doc_section"><a name="basics">Code Generation Setup</a></div>
  61 <!-- *********************************************************************** -->
  62
  63 <div class="doc_text">
  64
  65 <p>
  66 In order to generate LLVM IR, we want some simple setup to get started.  First
  67 we define virtual code generation (codegen) methods in each AST class:</p>
  68
  69 <div class="doc_code">
  70 <pre>
  71 /// ExprAST - Base class for all expression nodes.
  72 class ExprAST {
  73 public:
  74   virtual ~ExprAST() {}
  75   <b>virtual Value *Codegen() = 0;</b>
  76 };
  77
  78 /// NumberExprAST - Expression class for numeric literals like "1.0".
  79 class NumberExprAST : public ExprAST {
  80   double Val;
  81 public:
  82   explicit NumberExprAST(double val) : Val(val) {}
  83   <b>virtual Value *Codegen();</b>
  84 };
  85 ...
  86 </pre>
  87 </div>
  88
  89 <p>The Codegen() method says to emit IR for that AST node along with all the things it
  90 depends on, and they all return an LLVM Value object.
  91 "Value" is the class used to represent a "<a
  92 href="http://en.wikipedia.org/wiki/Static_single_assignment_form">Static Single
  93 Assignment (SSA)</a> register" or "SSA value" in LLVM.  The most distinct aspect
  94 of SSA values is that their value is computed as the related instruction
  95 executes, and it does not get a new value until (and if) the instruction
  96 re-executes.  In other words, there is no way to "change" an SSA value.  For
  97 more information, please read up on <a
  98 href="http://en.wikipedia.org/wiki/Static_single_assignment_form">Static Single
  99 Assignment</a> - the concepts are really quite natural once you grok them.</p>
 100
 101 <p>Note that instead of adding virtual methods to the ExprAST class hierarchy,
 102 it could also make sense to use a <a
 103 href="http://en.wikipedia.org/wiki/Visitor_pattern">visitor pattern</a> or some
 104 other way to model this.  Again, this tutorial won't dwell on good software
 105 engineering practices: for our purposes, adding a virtual method is
 106 simplest.</p>
 107
 108 <p>The
 109 second thing we want is an "Error" method like we used for the parser, which will
 110 be used to report errors found during code generation (for example, use of an
 111 undeclared parameter):</p>
 112
 113 <div class="doc_code">
 114 <pre>
 115 Value *ErrorV(const char *Str) { Error(Str); return 0; }
 116
 117 static Module *TheModule;
 118 static IRBuilder&lt;&gt; Builder;
 119 static std::map&lt;std::string, Value*&gt; NamedValues;
 120 </pre>
 121 </div>
 122
 123 <p>The static variables will be used during code generation.  <tt>TheModule</tt>
 124 is the LLVM construct that contains all of the functions and global variables in
 125 a chunk of code.  In many ways, it is the top-level structure that the LLVM IR
 126 uses to contain code.</p>
 127
 128 <p>The <tt>Builder</tt> object is a helper object that makes it easy to generate
 129 LLVM instructions.  Instances of the <a
 130 href="http://llvm.org/doxygen/IRBuilder_8h-source.html"><tt>IRBuilder</tt></a>
 131 class template keep track of the current place to insert instructions and has
 132 methods to create new instructions.</p>
 133
 134 <p>The <tt>NamedValues</tt> map keeps track of which values are defined in the
 135 current scope and what their LLVM representation is.  (In other words, it is a
 136 symbol table for the code).  In this form of Kaleidoscope, the only things that
 137 can be referenced are function parameters.  As such, function parameters will
 138 be in this map when generating code for their function body.</p>
 139
 140 <p>
 141 With these basics in place, we can start talking about how to generate code for
 142 each expression.  Note that this assumes that the <tt>Builder</tt> has been set
 143 up to generate code <em>into</em> something.  For now, we'll assume that this
 144 has already been done, and we'll just use it to emit code.
 145 </p>
 146
 147 </div>
 148
 149 <!-- *********************************************************************** -->
 150 <div class="doc_section"><a name="exprs">Expression Code Generation</a></div>
 151 <!-- *********************************************************************** -->
 152
 153 <div class="doc_text">
 154
 155 <p>Generating LLVM code for expression nodes is very straightforward: less
 156 than 45 lines of commented code for all four of our expression nodes.  First
 157 we'll do numeric literals:</p>
 158
 159 <div class="doc_code">
 160 <pre>
 161 Value *NumberExprAST::Codegen() {
 162   return ConstantFP::get(APFloat(Val));
 163 }
 164 </pre>
 165 </div>
 166
 167 <p>In the LLVM IR, numeric constants are represented with the
 168 <tt>ConstantFP</tt> class, which holds the numeric value in an <tt>APFloat</tt>
 169 internally (<tt>APFloat</tt> has the capability of holding floating point
 170 constants of <em>A</em>rbitrary <em>P</em>recision).  This code basically just
 171 creates and returns a <tt>ConstantFP</tt>.  Note that in the LLVM IR
 172 that constants are all uniqued together and shared.  For this reason, the API
 173 uses "the foo::get(..)" idiom instead of "new foo(..)" or "foo::Create(..)".</p>
 174
 175 <div class="doc_code">
 176 <pre>
 177 Value *VariableExprAST::Codegen() {
 178   // Look this variable up in the function.
 179   Value *V = NamedValues[Name];
 180   return V ? V : ErrorV("Unknown variable name");
 181 }
 182 </pre>
 183 </div>
 184
 185 <p>References to variables are also quite simple using LLVM.  In the simple version
 186 of Kaleidoscope, we assume that the variable has already been emited somewhere
 187 and its value is available.  In practice, the only values that can be in the
 188 <tt>NamedValues</tt> map are function arguments.  This
 189 code simply checks to see that the specified name is in the map (if not, an
 190 unknown variable is being referenced) and returns the value for it.  In future
 191 chapters, we'll add support for <a href="LangImpl5.html#for">loop induction
 192 variables</a> in the symbol table, and for <a
 193 href="LangImpl7.html#localvars">local variables</a>.</p>
 194
 195 <div class="doc_code">
 196 <pre>
 197 Value *BinaryExprAST::Codegen() {
 198   Value *L = LHS-&gt;Codegen();
 199   Value *R = RHS-&gt;Codegen();
 200   if (L == 0 || R == 0) return 0;
 201
 202   switch (Op) {
 203   case '+': return Builder.CreateAdd(L, R, "addtmp");
 204   case '-': return Builder.CreateSub(L, R, "subtmp");
 205   case '*': return Builder.CreateMul(L, R, "multmp");
 206   case '&lt;':
 207     L = Builder.CreateFCmpULT(L, R, "cmptmp");
 208     // Convert bool 0/1 to double 0.0 or 1.0
 209     return Builder.CreateUIToFP(L, Type::DoubleTy, "booltmp");
 210   default: return ErrorV("invalid binary operator");
 211   }
 212 }
 213 </pre>
 214 </div>
 215
 216 <p>Binary operators start to get more interesting.  The basic idea here is that
 217 we recursively emit code for the left-hand side of the expression, then the
 218 right-hand side, then we compute the result of the binary expression.  In this
 219 code, we do a simple switch on the opcode to create the right LLVM instruction.
 220 </p>
 221
 222 <p>In the example above, the LLVM builder class is starting to show its value.
 223 IRBuilder knows where to insert the newly created instruction, all you have to
 224 do is specify what instruction to create (e.g. with <tt>CreateAdd</tt>), which
 225 operands to use (<tt>L</tt> and <tt>R</tt> here) and optionally provide a name
 226 for the generated instruction.</p>
 227
 228 <p>One nice thing about LLVM is that the name is just a hint.  For instance, if
 229 the code above emits multiple "addtmp" variables, LLVM will automatically
 230 provide each one with an increasing, unique numeric suffix.  Local value names
 231 for instructions are purely optional, but it makes it much easier to read the
 232 IR dumps.</p>
 233
 234 <p><a href="../LangRef.html#instref">LLVM instructions</a> are constrained by
 235 strict rules: for example, the Left and Right operators of
 236 an <a href="../LangRef.html#i_add">add instruction</a> must have the same
 237 type, and the result type of the add must match the operand types.  Because
 238 all values in Kaleidoscope are doubles, this makes for very simple code for add,
 239 sub and mul.</p>
 240
 241 <p>On the other hand, LLVM specifies that the <a
 242 href="../LangRef.html#i_fcmp">fcmp instruction</a> always returns an 'i1' value
 243 (a one bit integer).  The problem with this is that Kaleidoscope wants the value to be a 0.0 or 1.0 value.  In order to get these semantics, we combine the fcmp instruction with
 244 a <a href="../LangRef.html#i_uitofp">uitofp instruction</a>.  This instruction
 245 converts its input integer into a floating point value by treating the input
 246 as an unsigned value.  In contrast, if we used the <a
 247 href="../LangRef.html#i_sitofp">sitofp instruction</a>, the Kaleidoscope '&lt;'
 248 operator would return 0.0 and -1.0, depending on the input value.</p>
 249
 250 <div class="doc_code">
 251 <pre>
 252 Value *CallExprAST::Codegen() {
 253   // Look up the name in the global module table.
 254   Function *CalleeF = TheModule-&gt;getFunction(Callee);
 255   if (CalleeF == 0)
 256     return ErrorV("Unknown function referenced");
 257
 258   // If argument mismatch error.
 259   if (CalleeF-&gt;arg_size() != Args.size())
 260     return ErrorV("Incorrect # arguments passed");
 261
 262   std::vector&lt;Value*&gt; ArgsV;
 263   for (unsigned i = 0, e = Args.size(); i != e; ++i) {
 264     ArgsV.push_back(Args[i]-&gt;Codegen());
 265     if (ArgsV.back() == 0) return 0;
 266   }
 267
 268   return Builder.CreateCall(CalleeF, ArgsV.begin(), ArgsV.end(), "calltmp");
 269 }
 270 </pre>
 271 </div>
 272
 273 <p>Code generation for function calls is quite straightforward with LLVM.  The
 274 code above initially does a function name lookup in the LLVM Module's symbol
 275 table.  Recall that the LLVM Module is the container that holds all of the
 276 functions we are JIT'ing.  By giving each function the same name as what the
 277 user specifies, we can use the LLVM symbol table to resolve function names for
 278 us.</p>
 279
 280 <p>Once we have the function to call, we recursively codegen each argument that
 281 is to be passed in, and create an LLVM <a href="../LangRef.html#i_call">call
 282 instruction</a>.  Note that LLVM uses the native C calling conventions by
 283 default, allowing these calls to also call into standard library functions like
 284 "sin" and "cos", with no additional effort.</p>
 285
 286 <p>This wraps up our handling of the four basic expressions that we have so far
 287 in Kaleidoscope.  Feel free to go in and add some more.  For example, by
 288 browsing the <a href="../LangRef.html">LLVM language reference</a> you'll find
 289 several other interesting instructions that are really easy to plug into our
 290 basic framework.</p>
 291
 292 </div>
 293
 294 <!-- *********************************************************************** -->
 295 <div class="doc_section"><a name="funcs">Function Code Generation</a></div>
 296 <!-- *********************************************************************** -->
 297
 298 <div class="doc_text">
 299
 300 <p>Code generation for prototypes and functions must handle a number of
 301 details, which make their code less beautiful than expression code
 302 generation, but allows us to  illustrate some important points.  First, lets
 303 talk about code generation for prototypes: they are used both for function
 304 bodies and external function declarations.  The code starts with:</p>
 305
 306 <div class="doc_code">
 307 <pre>
 308 Function *PrototypeAST::Codegen() {
 309   // Make the function type:  double(double,double) etc.
 310   std::vector&lt;const Type*&gt; Doubles(Args.size(), Type::DoubleTy);
 311   FunctionType *FT = FunctionType::get(Type::DoubleTy, Doubles, false);
 312
 313   Function *F = Function::Create(FT, Function::ExternalLinkage, Name, TheModule);
 314 </pre>
 315 </div>
 316
 317 <p>This code packs a lot of power into a few lines.  Note first that this
 318 function returns a "Function*" instead of a "Value*".  Because a "prototype"
 319 really talks about the external interface for a function (not the value computed
 320 by an expression), it makes sense for it to return the LLVM Function it
 321 corresponds to when codegen'd.</p>
 322
 323 <p>The call to <tt>FunctionType::get</tt> creates
 324 the <tt>FunctionType</tt> that should be used for a given Prototype.  Since all
 325 function arguments in Kaleidoscope are of type double, the first line creates
 326 a vector of "N" LLVM double types.  It then uses the <tt>FunctionType::get</tt>
 327 method to create a function type that takes "N" doubles as arguments, returns
 328 one double as a result, and that is not vararg (the false parameter indicates
 329 this).  Note that Types in LLVM are uniqued just like Constants are, so you
 330 don't "new" a type, you "get" it.</p>
 331
 332 <p>The final line above actually creates the function that the prototype will
 333 correspond to.  This indicates the type, linkage and name to use, as well as which
 334 module to insert into.  "<a href="../LangRef.html#linkage">external linkage</a>"
 335 means that the function may be defined outside the current module and/or that it
 336 is callable by functions outside the module.  The Name passed in is the name the
 337 user specified: since "<tt>TheModule</tt>" is specified, this name is registered
 338 in "<tt>TheModule</tt>"s symbol table, which is used by the function call code
 339 above.</p>
 340
 341 <div class="doc_code">
 342 <pre>
 343   // If F conflicted, there was already something named 'Name'.  If it has a
 344   // body, don't allow redefinition or reextern.
 345   if (F-&gt;getName() != Name) {
 346     // Delete the one we just made and get the existing one.
 347     F-&gt;eraseFromParent();
 348     F = TheModule-&gt;getFunction(Name);
 349 </pre>
 350 </div>
 351
 352 <p>The Module symbol table works just like the Function symbol table when it
 353 comes to name conflicts: if a new function is created with a name was previously
 354 added to the symbol table, it will get implicitly renamed when added to the
 355 Module.  The code above exploits this fact to determine if there was a previous
 356 definition of this function.</p>
 357
 358 <p>In Kaleidoscope, I choose to allow redefinitions of functions in two cases:
 359 first, we want to allow 'extern'ing a function more than once, as long as the
 360 prototypes for the externs match (since all arguments have the same type, we
 361 just have to check that the number of arguments match).  Second, we want to
 362 allow 'extern'ing a function and then definining a body for it.  This is useful
 363 when defining mutually recursive functions.</p>
 364
 365 <p>In order to implement this, the code above first checks to see if there is
 366 a collision on the name of the function.  If so, it deletes the function we just
 367 created (by calling <tt>eraseFromParent</tt>) and then calling
 368 <tt>getFunction</tt> to get the existing function with the specified name.  Note
 369 that many APIs in LLVM have "erase" forms and "remove" forms.  The "remove" form
 370 unlinks the object from its parent (e.g. a Function from a Module) and returns
 371 it.  The "erase" form unlinks the object and then deletes it.</p>
 372
 373 <div class="doc_code">
 374 <pre>
 375     // If F already has a body, reject this.
 376     if (!F-&gt;empty()) {
 377       ErrorF("redefinition of function");
 378       return 0;
 379     }
 380
 381     // If F took a different number of args, reject.
 382     if (F-&gt;arg_size() != Args.size()) {
 383       ErrorF("redefinition of function with different # args");
 384       return 0;
 385     }
 386   }
 387 </pre>
 388 </div>
 389
 390 <p>In order to verify the logic above, we first check to see if the pre-existing
 391 function is "empty".  In this case, empty means that it has no basic blocks in
 392 it, which means it has no body.  If it has no body, it is a forward
 393 declaration.  Since we don't allow anything after a full definition of the
 394 function, the code rejects this case.  If the previous reference to a function
 395 was an 'extern', we simply verify that the number of arguments for that
 396 definition and this one match up.  If not, we emit an error.</p>
 397
 398 <div class="doc_code">
 399 <pre>
 400   // Set names for all arguments.
 401   unsigned Idx = 0;
 402   for (Function::arg_iterator AI = F-&gt;arg_begin(); Idx != Args.size();
 403        ++AI, ++Idx) {
 404     AI-&gt;setName(Args[Idx]);
 405
 406     // Add arguments to variable symbol table.
 407     NamedValues[Args[Idx]] = AI;
 408   }
 409   return F;
 410 }
 411 </pre>
 412 </div>
 413
 414 <p>The last bit of code for prototypes loops over all of the arguments in the
 415 function, setting the name of the LLVM Argument objects to match, and registering
 416 the arguments in the <tt>NamedValues</tt> map for future use by the
 417 <tt>VariableExprAST</tt> AST node.  Once this is set up, it returns the Function
 418 object to the caller.  Note that we don't check for conflicting
 419 argument names here (e.g. "extern foo(a b a)").  Doing so would be very
 420 straight-forward with the mechanics we have already used above.</p>
 421
 422 <div class="doc_code">
 423 <pre>
 424 Function *FunctionAST::Codegen() {
 425   NamedValues.clear();
 426
 427   Function *TheFunction = Proto-&gt;Codegen();
 428   if (TheFunction == 0)
 429     return 0;
 430 </pre>
 431 </div>
 432
 433 <p>Code generation for function definitions starts out simply enough: we just
 434 codegen the prototype (Proto) and verify that it is ok.  We then clear out the
 435 <tt>NamedValues</tt> map to make sure that there isn't anything in it from the
 436 last function we compiled.  Code generation of the prototype ensures that there
 437 is an LLVM Function object that is ready to go for us.</p>
 438
 439 <div class="doc_code">
 440 <pre>
 441   // Create a new basic block to start insertion into.
 442   BasicBlock *BB = BasicBlock::Create("entry", TheFunction);
 443   Builder.SetInsertPoint(BB);
 444
 445   if (Value *RetVal = Body-&gt;Codegen()) {
 446 </pre>
 447 </div>
 448
 449 <p>Now we get to the point where the <tt>Builder</tt> is set up.  The first
 450 line creates a new <a href="http://en.wikipedia.org/wiki/Basic_block">basic
 451 block</a> (named "entry"), which is inserted into <tt>TheFunction</tt>.  The
 452 second line then tells the builder that new instructions should be inserted into
 453 the end of the new basic block.  Basic blocks in LLVM are an important part
 454 of functions that define the <a
 455 href="http://en.wikipedia.org/wiki/Control_flow_graph">Control Flow Graph</a>.
 456 Since we don't have any control flow, our functions will only contain one
 457 block at this point.  We'll fix this in <a href="LangImpl5.html">Chapter 5</a> :).</p>
 458
 459 <div class="doc_code">
 460 <pre>
 461   if (Value *RetVal = Body-&gt;Codegen()) {
 462     // Finish off the function.
 463     Builder.CreateRet(RetVal);
 464
 465     // Validate the generated code, checking for consistency.
 466     verifyFunction(*TheFunction);
 467     return TheFunction;
 468   }
 469 </pre>
 470 </div>
 471
 472 <p>Once the insertion point is set up, we call the <tt>CodeGen()</tt> method for
 473 the root expression of the function.  If no error happens, this emits code to
 474 compute the expression into the entry block and returns the value that was
 475 computed.  Assuming no error, we then create an LLVM <a
 476 href="../LangRef.html#i_ret">ret instruction</a>, which completes the function.
 477 Once the function is built, we call <tt>verifyFunction</tt>, which
 478 is provided by LLVM.  This function does a variety of consistency checks on the
 479 generated code, to determine if our compiler is doing everything right.  Using
 480 this is important: it can catch a lot of bugs.  Once the function is finished
 481 and validated, we return it.</p>
 482
 483 <div class="doc_code">
 484 <pre>
 485   // Error reading body, remove function.
 486   TheFunction-&gt;eraseFromParent();
 487   return 0;
 488 }
 489 </pre>
 490 </div>
 491
 492 <p>The only piece left here is handling of the error case.  For simplicity, we
 493 handle this by merely deleting the function we produced with the
 494 <tt>eraseFromParent</tt> method.  This allows the user to redefine a function
 495 that they incorrectly typed in before: if we didn't delete it, it would live in
 496 the symbol table, with a body, preventing future redefinition.</p>
 497
 498 <p>This code does have a bug, though.  Since the <tt>PrototypeAST::Codegen</tt>
 499 can return a previously defined forward declaration, our code can actually delete
 500 a forward declaration.  There are a number of ways to fix this bug, see what you
 501 can come up with!  Here is a testcase:</p>
 502
 503 <div class="doc_code">
 504 <pre>
 505 extern foo(a b);     # ok, defines foo.
 506 def foo(a b) c;      # error, 'c' is invalid.
 507 def bar() foo(1, 2); # error, unknown function "foo"
 508 </pre>
 509 </div>
 510
 511 </div>
 512
 513 <!-- *********************************************************************** -->
 514 <div class="doc_section"><a name="driver">Driver Changes and
 515 Closing Thoughts</a></div>
 516 <!-- *********************************************************************** -->
 517
 518 <div class="doc_text">
 519
 520 <p>
 521 For now, code generation to LLVM doesn't really get us much, except that we can
 522 look at the pretty IR calls.  The sample code inserts calls to Codegen into the
 523 "<tt>HandleDefinition</tt>", "<tt>HandleExtern</tt>" etc functions, and then
 524 dumps out the LLVM IR.  This gives a nice way to look at the LLVM IR for simple
 525 functions.  For example:
 526 </p>
 527
 528 <div class="doc_code">
 529 <pre>
 530 ready> <b>4+5</b>;
 531 Read top-level expression:
 532 define double @""() {
 533 entry:
 534         %addtmp = add double 4.000000e+00, 5.000000e+00
 535         ret double %addtmp
 536 }
 537 </pre>
 538 </div>
 539
 540 <p>Note how the parser turns the top-level expression into anonymous functions
 541 for us.  This will be handy when we add <a href="LangImpl4.html#jit">JIT
 542 support</a> in the next chapter.  Also note that the code is very literally
 543 transcribed, no optimizations are being performed.  We will
 544 <a href="LangImpl4.html#trivialconstfold">add optimizations</a> explicitly in
 545 the next chapter.</p>
 546
 547 <div class="doc_code">
 548 <pre>
 549 ready&gt; <b>def foo(a b) a*a + 2*a*b + b*b;</b>
 550 Read function definition:
 551 define double @foo(double %a, double %b) {
 552 entry:
 553         %multmp = mul double %a, %a
 554         %multmp1 = mul double 2.000000e+00, %a
 555         %multmp2 = mul double %multmp1, %b
 556         %addtmp = add double %multmp, %multmp2
 557         %multmp3 = mul double %b, %b
 558         %addtmp4 = add double %addtmp, %multmp3
 559         ret double %addtmp4
 560 }
 561 </pre>
 562 </div>
 563
 564 <p>This shows some simple arithmetic. Notice the striking similarity to the
 565 LLVM builder calls that we use to create the instructions.</p>
 566
 567 <div class="doc_code">
 568 <pre>
 569 ready&gt; <b>def bar(a) foo(a, 4.0) + bar(31337);</b>
 570 Read function definition:
 571 define double @bar(double %a) {
 572 entry:
 573         %calltmp = call double @foo( double %a, double 4.000000e+00 )
 574         %calltmp1 = call double @bar( double 3.133700e+04 )
 575         %addtmp = add double %calltmp, %calltmp1
 576         ret double %addtmp
 577 }
 578 </pre>
 579 </div>
 580
 581 <p>This shows some function calls.  Note that this function will take a long
 582 time to execute if you call it.  In the future we'll add conditional control
 583 flow to actually make recursion useful :).</p>
 584
 585 <div class="doc_code">
 586 <pre>
 587 ready&gt; <b>extern cos(x);</b>
 588 Read extern:
 589 declare double @cos(double)
 590
 591 ready&gt; <b>cos(1.234);</b>
 592 Read top-level expression:
 593 define double @""() {
 594 entry:
 595         %calltmp = call double @cos( double 1.234000e+00 )
 596         ret double %calltmp
 597 }
 598 </pre>
 599 </div>
 600
 601 <p>This shows an extern for the libm "cos" function, and a call to it.</p>
 602
 603
 604 <div class="doc_code">
 605 <pre>
 606 ready&gt; <b>^D</b>
 607 ; ModuleID = 'my cool jit'
 608
 609 define double @""() {
 610 entry:
 611         %addtmp = add double 4.000000e+00, 5.000000e+00
 612         ret double %addtmp
 613 }
 614
 615 define double @foo(double %a, double %b) {
 616 entry:
 617         %multmp = mul double %a, %a
 618         %multmp1 = mul double 2.000000e+00, %a
 619         %multmp2 = mul double %multmp1, %b
 620         %addtmp = add double %multmp, %multmp2
 621         %multmp3 = mul double %b, %b
 622         %addtmp4 = add double %addtmp, %multmp3
 623         ret double %addtmp4
 624 }
 625
 626 define double @bar(double %a) {
 627 entry:
 628         %calltmp = call double @foo( double %a, double 4.000000e+00 )
 629         %calltmp1 = call double @bar( double 3.133700e+04 )
 630         %addtmp = add double %calltmp, %calltmp1
 631         ret double %addtmp
 632 }
 633
 634 declare double @cos(double)
 635
 636 define double @""() {
 637 entry:
 638         %calltmp = call double @cos( double 1.234000e+00 )
 639         ret double %calltmp
 640 }
 641 </pre>
 642 </div>
 643
 644 <p>When you quit the current demo, it dumps out the IR for the entire module
 645 generated.  Here you can see the big picture with all the functions referencing
 646 each other.</p>
 647
 648 <p>This wraps up the third chapter of the Kaleidoscope tutorial.  Up next, we'll
 649 describe how to <a href="LangImpl4.html">add JIT codegen and optimizer
 650 support</a> to this so we can actually start running code!</p>
 651
 652 </div>
 653
 654
 655 <!-- *********************************************************************** -->
 656 <div class="doc_section"><a name="code">Full Code Listing</a></div>
 657 <!-- *********************************************************************** -->
 658
 659 <div class="doc_text">
 660
 661 <p>
 662 Here is the complete code listing for our running example, enhanced with the
 663 LLVM code generator.    Because this uses the LLVM libraries, we need to link
 664 them in.  To do this, we use the <a
 665 href="http://llvm.org/cmds/llvm-config.html">llvm-config</a> tool to inform
 666 our makefile/command line about which options to use:</p>
 667
 668 <div class="doc_code">
 669 <pre>
 670    # Compile
 671    g++ -g -O3 toy.cpp `llvm-config --cppflags --ldflags --libs core` -o toy
 672    # Run
 673    ./toy
 674 </pre>
 675 </div>
 676
 677 <p>Here is the code:</p>
 678
 679 <div class="doc_code">
 680 <pre>
 681 // To build this:
 682 // See example below.
 683
 684 #include "llvm/DerivedTypes.h"
 685 #include "llvm/Module.h"
 686 #include "llvm/Analysis/Verifier.h"
 687 #include "llvm/Support/IRBuilder.h"
 688 #include &lt;cstdio&gt;
 689 #include &lt;string&gt;
 690 #include &lt;map&gt;
 691 #include &lt;vector&gt;
 692 using namespace llvm;
 693
 694 //===----------------------------------------------------------------------===//
 695 // Lexer
 696 //===----------------------------------------------------------------------===//
 697
 698 // The lexer returns tokens [0-255] if it is an unknown character, otherwise one
 699 // of these for known things.
 700 enum Token {
 701   tok_eof = -1,
 702
 703   // commands
 704   tok_def = -2, tok_extern = -3,
 705
 706   // primary
 707   tok_identifier = -4, tok_number = -5,
 708 };
 709
 710 static std::string IdentifierStr;  // Filled in if tok_identifier
 711 static double NumVal;              // Filled in if tok_number
 712
 713 /// gettok - Return the next token from standard input.
 714 static int gettok() {
 715   static int LastChar = ' ';
 716
 717   // Skip any whitespace.
 718   while (isspace(LastChar))
 719     LastChar = getchar();
 720
 721   if (isalpha(LastChar)) { // identifier: [a-zA-Z][a-zA-Z0-9]*
 722     IdentifierStr = LastChar;
 723     while (isalnum((LastChar = getchar())))
 724       IdentifierStr += LastChar;
 725
 726     if (IdentifierStr == "def") return tok_def;
 727     if (IdentifierStr == "extern") return tok_extern;
 728     return tok_identifier;
 729   }
 730
 731   if (isdigit(LastChar) || LastChar == '.') {   // Number: [0-9.]+
 732     std::string NumStr;
 733     do {
 734       NumStr += LastChar;
 735       LastChar = getchar();
 736     } while (isdigit(LastChar) || LastChar == '.');
 737
 738     NumVal = strtod(NumStr.c_str(), 0);
 739     return tok_number;
 740   }
 741
 742   if (LastChar == '#') {
 743     // Comment until end of line.
 744     do LastChar = getchar();
 745     while (LastChar != EOF &amp;&amp; LastChar != '\n' &amp;&amp; LastChar != '\r');
 746
 747     if (LastChar != EOF)
 748       return gettok();
 749   }
 750
 751   // Check for end of file.  Don't eat the EOF.
 752   if (LastChar == EOF)
 753     return tok_eof;
 754
 755   // Otherwise, just return the character as its ascii value.
 756   int ThisChar = LastChar;
 757   LastChar = getchar();
 758   return ThisChar;
 759 }
 760
 761 //===----------------------------------------------------------------------===//
 762 // Abstract Syntax Tree (aka Parse Tree)
 763 //===----------------------------------------------------------------------===//
 764
 765 /// ExprAST - Base class for all expression nodes.
 766 class ExprAST {
 767 public:
 768   virtual ~ExprAST() {}
 769   virtual Value *Codegen() = 0;
 770 };
 771
 772 /// NumberExprAST - Expression class for numeric literals like "1.0".
 773 class NumberExprAST : public ExprAST {
 774   double Val;
 775 public:
 776   explicit NumberExprAST(double val) : Val(val) {}
 777   virtual Value *Codegen();
 778 };
 779
 780 /// VariableExprAST - Expression class for referencing a variable, like "a".
 781 class VariableExprAST : public ExprAST {
 782   std::string Name;
 783 public:
 784   explicit VariableExprAST(const std::string &amp;name) : Name(name) {}
 785   virtual Value *Codegen();
 786 };
 787
 788 /// BinaryExprAST - Expression class for a binary operator.
 789 class BinaryExprAST : public ExprAST {
 790   char Op;
 791   ExprAST *LHS, *RHS;
 792 public:
 793   BinaryExprAST(char op, ExprAST *lhs, ExprAST *rhs)
 794     : Op(op), LHS(lhs), RHS(rhs) {}
 795   virtual Value *Codegen();
 796 };
 797
 798 /// CallExprAST - Expression class for function calls.
 799 class CallExprAST : public ExprAST {
 800   std::string Callee;
 801   std::vector&lt;ExprAST*&gt; Args;
 802 public:
 803   CallExprAST(const std::string &amp;callee, std::vector&lt;ExprAST*&gt; &amp;args)
 804     : Callee(callee), Args(args) {}
 805   virtual Value *Codegen();
 806 };
 807
 808 /// PrototypeAST - This class represents the "prototype" for a function,
 809 /// which captures its argument names as well as if it is an operator.
 810 class PrototypeAST {
 811   std::string Name;
 812   std::vector&lt;std::string&gt; Args;
 813 public:
 814   PrototypeAST(const std::string &amp;name, const std::vector&lt;std::string&gt; &amp;args)
 815     : Name(name), Args(args) {}
 816
 817   Function *Codegen();
 818 };
 819
 820 /// FunctionAST - This class represents a function definition itself.
 821 class FunctionAST {
 822   PrototypeAST *Proto;
 823   ExprAST *Body;
 824 public:
 825   FunctionAST(PrototypeAST *proto, ExprAST *body)
 826     : Proto(proto), Body(body) {}
 827
 828   Function *Codegen();
 829 };
 830
 831 //===----------------------------------------------------------------------===//
 832 // Parser
 833 //===----------------------------------------------------------------------===//
 834
 835 /// CurTok/getNextToken - Provide a simple token buffer.  CurTok is the current
 836 /// token the parser it looking at.  getNextToken reads another token from the
 837 /// lexer and updates CurTok with its results.
 838 static int CurTok;
 839 static int getNextToken() {
 840   return CurTok = gettok();
 841 }
 842
 843 /// BinopPrecedence - This holds the precedence for each binary operator that is
 844 /// defined.
 845 static std::map&lt;char, int&gt; BinopPrecedence;
 846
 847 /// GetTokPrecedence - Get the precedence of the pending binary operator token.
 848 static int GetTokPrecedence() {
 849   if (!isascii(CurTok))
 850     return -1;
 851
 852   // Make sure it's a declared binop.
 853   int TokPrec = BinopPrecedence[CurTok];
 854   if (TokPrec &lt;= 0) return -1;
 855   return TokPrec;
 856 }
 857
 858 /// Error* - These are little helper functions for error handling.
 859 ExprAST *Error(const char *Str) { fprintf(stderr, "Error: %s\n", Str);return 0;}
 860 PrototypeAST *ErrorP(const char *Str) { Error(Str); return 0; }
 861 FunctionAST *ErrorF(const char *Str) { Error(Str); return 0; }
 862
 863 static ExprAST *ParseExpression();
 864
 865 /// identifierexpr
 866 ///   ::= identifier
 867 ///   ::= identifier '(' expression* ')'
 868 static ExprAST *ParseIdentifierExpr() {
 869   std::string IdName = IdentifierStr;
 870
 871   getNextToken();  // eat identifier.
 872
 873   if (CurTok != '(') // Simple variable ref.
 874     return new VariableExprAST(IdName);
 875
 876   // Call.
 877   getNextToken();  // eat (
 878   std::vector&lt;ExprAST*&gt; Args;
 879   if (CurTok != ')') {
 880     while (1) {
 881       ExprAST *Arg = ParseExpression();
 882       if (!Arg) return 0;
 883       Args.push_back(Arg);
 884
 885       if (CurTok == ')') break;
 886
 887       if (CurTok != ',')
 888         return Error("Expected ')' or ',' in argument list");
 889       getNextToken();
 890     }
 891   }
 892
 893   // Eat the ')'.
 894   getNextToken();
 895
 896   return new CallExprAST(IdName, Args);
 897 }
 898
 899 /// numberexpr ::= number
 900 static ExprAST *ParseNumberExpr() {
 901   ExprAST *Result = new NumberExprAST(NumVal);
 902   getNextToken(); // consume the number
 903   return Result;
 904 }
 905
 906 /// parenexpr ::= '(' expression ')'
 907 static ExprAST *ParseParenExpr() {
 908   getNextToken();  // eat (.
 909   ExprAST *V = ParseExpression();
 910   if (!V) return 0;
 911
 912   if (CurTok != ')')
 913     return Error("expected ')'");
 914   getNextToken();  // eat ).
 915   return V;
 916 }
 917
 918 /// primary
 919 ///   ::= identifierexpr
 920 ///   ::= numberexpr
 921 ///   ::= parenexpr
 922 static ExprAST *ParsePrimary() {
 923   switch (CurTok) {
 924   default: return Error("unknown token when expecting an expression");
 925   case tok_identifier: return ParseIdentifierExpr();
 926   case tok_number:     return ParseNumberExpr();
 927   case '(':            return ParseParenExpr();
 928   }
 929 }
 930
 931 /// binoprhs
 932 ///   ::= ('+' primary)*
 933 static ExprAST *ParseBinOpRHS(int ExprPrec, ExprAST *LHS) {
 934   // If this is a binop, find its precedence.
 935   while (1) {
 936     int TokPrec = GetTokPrecedence();
 937
 938     // If this is a binop that binds at least as tightly as the current binop,
 939     // consume it, otherwise we are done.
 940     if (TokPrec &lt; ExprPrec)
 941       return LHS;
 942
 943     // Okay, we know this is a binop.
 944     int BinOp = CurTok;
 945     getNextToken();  // eat binop
 946
 947     // Parse the primary expression after the binary operator.
 948     ExprAST *RHS = ParsePrimary();
 949     if (!RHS) return 0;
 950
 951     // If BinOp binds less tightly with RHS than the operator after RHS, let
 952     // the pending operator take RHS as its LHS.
 953     int NextPrec = GetTokPrecedence();
 954     if (TokPrec &lt; NextPrec) {
 955       RHS = ParseBinOpRHS(TokPrec+1, RHS);
 956       if (RHS == 0) return 0;
 957     }
 958
 959     // Merge LHS/RHS.
 960     LHS = new BinaryExprAST(BinOp, LHS, RHS);
 961   }
 962 }
 963
 964 /// expression
 965 ///   ::= primary binoprhs
 966 ///
 967 static ExprAST *ParseExpression() {
 968   ExprAST *LHS = ParsePrimary();
 969   if (!LHS) return 0;
 970
 971   return ParseBinOpRHS(0, LHS);
 972 }
 973
 974 /// prototype
 975 ///   ::= id '(' id* ')'
 976 static PrototypeAST *ParsePrototype() {
 977   if (CurTok != tok_identifier)
 978     return ErrorP("Expected function name in prototype");
 979
 980   std::string FnName = IdentifierStr;
 981   getNextToken();
 982
 983   if (CurTok != '(')
 984     return ErrorP("Expected '(' in prototype");
 985
 986   std::vector&lt;std::string&gt; ArgNames;
 987   while (getNextToken() == tok_identifier)
 988     ArgNames.push_back(IdentifierStr);
 989   if (CurTok != ')')
 990     return ErrorP("Expected ')' in prototype");
 991
 992   // success.
 993   getNextToken();  // eat ')'.
 994
 995   return new PrototypeAST(FnName, ArgNames);
 996 }
 997
 998 /// definition ::= 'def' prototype expression
 999 static FunctionAST *ParseDefinition() {
1000   getNextToken();  // eat def.
1001   PrototypeAST *Proto = ParsePrototype();
1002   if (Proto == 0) return 0;
1003
1004   if (ExprAST *E = ParseExpression())
1005     return new FunctionAST(Proto, E);
1006   return 0;
1007 }
1008
1009 /// toplevelexpr ::= expression
1010 static FunctionAST *ParseTopLevelExpr() {
1011   if (ExprAST *E = ParseExpression()) {
1012     // Make an anonymous proto.
1013     PrototypeAST *Proto = new PrototypeAST("", std::vector&lt;std::string&gt;());
1014     return new FunctionAST(Proto, E);
1015   }
1016   return 0;
1017 }
1018
1019 /// external ::= 'extern' prototype
1020 static PrototypeAST *ParseExtern() {
1021   getNextToken();  // eat extern.
1022   return ParsePrototype();
1023 }
1024
1025 //===----------------------------------------------------------------------===//
1026 // Code Generation
1027 //===----------------------------------------------------------------------===//
1028
1029 static Module *TheModule;
1030 static IRBuilder&lt;&gt; Builder;
1031 static std::map&lt;std::string, Value*&gt; NamedValues;
1032
1033 Value *ErrorV(const char *Str) { Error(Str); return 0; }
1034
1035 Value *NumberExprAST::Codegen() {
1036   return ConstantFP::get(APFloat(Val));
1037 }
1038
1039 Value *VariableExprAST::Codegen() {
1040   // Look this variable up in the function.
1041   Value *V = NamedValues[Name];
1042   return V ? V : ErrorV("Unknown variable name");
1043 }
1044
1045 Value *BinaryExprAST::Codegen() {
1046   Value *L = LHS-&gt;Codegen();
1047   Value *R = RHS-&gt;Codegen();
1048   if (L == 0 || R == 0) return 0;
1049
1050   switch (Op) {
1051   case '+': return Builder.CreateAdd(L, R, "addtmp");
1052   case '-': return Builder.CreateSub(L, R, "subtmp");
1053   case '*': return Builder.CreateMul(L, R, "multmp");
1054   case '&lt;':
1055     L = Builder.CreateFCmpULT(L, R, "cmptmp");
1056     // Convert bool 0/1 to double 0.0 or 1.0
1057     return Builder.CreateUIToFP(L, Type::DoubleTy, "booltmp");
1058   default: return ErrorV("invalid binary operator");
1059   }
1060 }
1061
1062 Value *CallExprAST::Codegen() {
1063   // Look up the name in the global module table.
1064   Function *CalleeF = TheModule-&gt;getFunction(Callee);
1065   if (CalleeF == 0)
1066     return ErrorV("Unknown function referenced");
1067
1068   // If argument mismatch error.
1069   if (CalleeF-&gt;arg_size() != Args.size())
1070     return ErrorV("Incorrect # arguments passed");
1071
1072   std::vector&lt;Value*&gt; ArgsV;
1073   for (unsigned i = 0, e = Args.size(); i != e; ++i) {
1074     ArgsV.push_back(Args[i]-&gt;Codegen());
1075     if (ArgsV.back() == 0) return 0;
1076   }
1077
1078   return Builder.CreateCall(CalleeF, ArgsV.begin(), ArgsV.end(), "calltmp");
1079 }
1080
1081 Function *PrototypeAST::Codegen() {
1082   // Make the function type:  double(double,double) etc.
1083   std::vector&lt;const Type*&gt; Doubles(Args.size(), Type::DoubleTy);
1084   FunctionType *FT = FunctionType::get(Type::DoubleTy, Doubles, false);
1085
1086   Function *F = Function::Create(FT, Function::ExternalLinkage, Name, TheModule);
1087
1088   // If F conflicted, there was already something named 'Name'.  If it has a
1089   // body, don't allow redefinition or reextern.
1090   if (F-&gt;getName() != Name) {
1091     // Delete the one we just made and get the existing one.
1092     F-&gt;eraseFromParent();
1093     F = TheModule-&gt;getFunction(Name);
1094
1095     // If F already has a body, reject this.
1096     if (!F-&gt;empty()) {
1097       ErrorF("redefinition of function");
1098       return 0;
1099     }
1100
1101     // If F took a different number of args, reject.
1102     if (F-&gt;arg_size() != Args.size()) {
1103       ErrorF("redefinition of function with different # args");
1104       return 0;
1105     }
1106   }
1107
1108   // Set names for all arguments.
1109   unsigned Idx = 0;
1110   for (Function::arg_iterator AI = F-&gt;arg_begin(); Idx != Args.size();
1111        ++AI, ++Idx) {
1112     AI-&gt;setName(Args[Idx]);
1113
1114     // Add arguments to variable symbol table.
1115     NamedValues[Args[Idx]] = AI;
1116   }
1117
1118   return F;
1119 }
1120
1121 Function *FunctionAST::Codegen() {
1122   NamedValues.clear();
1123
1124   Function *TheFunction = Proto-&gt;Codegen();
1125   if (TheFunction == 0)
1126     return 0;
1127
1128   // Create a new basic block to start insertion into.
1129   BasicBlock *BB = BasicBlock::Create("entry", TheFunction);
1130   Builder.SetInsertPoint(BB);
1131
1132   if (Value *RetVal = Body-&gt;Codegen()) {
1133     // Finish off the function.
1134     Builder.CreateRet(RetVal);
1135
1136     // Validate the generated code, checking for consistency.
1137     verifyFunction(*TheFunction);
1138     return TheFunction;
1139   }
1140
1141   // Error reading body, remove function.
1142   TheFunction-&gt;eraseFromParent();
1143   return 0;
1144 }
1145
1146 //===----------------------------------------------------------------------===//
1147 // Top-Level parsing and JIT Driver
1148 //===----------------------------------------------------------------------===//
1149
1150 static void HandleDefinition() {
1151   if (FunctionAST *F = ParseDefinition()) {
1152     if (Function *LF = F-&gt;Codegen()) {
1153       fprintf(stderr, "Read function definition:");
1154       LF-&gt;dump();
1155     }
1156   } else {
1157     // Skip token for error recovery.
1158     getNextToken();
1159   }
1160 }
1161
1162 static void HandleExtern() {
1163   if (PrototypeAST *P = ParseExtern()) {
1164     if (Function *F = P-&gt;Codegen()) {
1165       fprintf(stderr, "Read extern: ");
1166       F-&gt;dump();
1167     }
1168   } else {
1169     // Skip token for error recovery.
1170     getNextToken();
1171   }
1172 }
1173
1174 static void HandleTopLevelExpression() {
1175   // Evaluate a top level expression into an anonymous function.
1176   if (FunctionAST *F = ParseTopLevelExpr()) {
1177     if (Function *LF = F-&gt;Codegen()) {
1178       fprintf(stderr, "Read top-level expression:");
1179       LF-&gt;dump();
1180     }
1181   } else {
1182     // Skip token for error recovery.
1183     getNextToken();
1184   }
1185 }
1186
1187 /// top ::= definition | external | expression | ';'
1188 static void MainLoop() {
1189   while (1) {
1190     fprintf(stderr, "ready&gt; ");
1191     switch (CurTok) {
1192     case tok_eof:    return;
1193     case ';':        getNextToken(); break;  // ignore top level semicolons.
1194     case tok_def:    HandleDefinition(); break;
1195     case tok_extern: HandleExtern(); break;
1196     default:         HandleTopLevelExpression(); break;
1197     }
1198   }
1199 }
1200
1201
1202
1203 //===----------------------------------------------------------------------===//
1204 // "Library" functions that can be "extern'd" from user code.
1205 //===----------------------------------------------------------------------===//
1206
1207 /// putchard - putchar that takes a double and returns 0.
1208 extern "C"
1209 double putchard(double X) {
1210   putchar((char)X);
1211   return 0;
1212 }
1213
1214 //===----------------------------------------------------------------------===//
1215 // Main driver code.
1216 //===----------------------------------------------------------------------===//
1217
1218 int main() {
1219   TheModule = new Module("my cool jit");
1220
1221   // Install standard binary operators.
1222   // 1 is lowest precedence.
1223   BinopPrecedence['&lt;'] = 10;
1224   BinopPrecedence['+'] = 20;
1225   BinopPrecedence['-'] = 20;
1226   BinopPrecedence['*'] = 40;  // highest.
1227
1228   // Prime the first token.
1229   fprintf(stderr, "ready&gt; ");
1230   getNextToken();
1231
1232   MainLoop();
1233   TheModule-&gt;dump();
1234   return 0;
1235 }
1236 </pre>
1237 </div>
1238 <a href="LangImpl4.html">Next: Adding JIT and Optimizer Support</a>
1239 </div>
1240
1241 <!-- *********************************************************************** -->
1242 <hr>
1243 <address>
1244   <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1245   src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
1246   <a href="http://validator.w3.org/check/referer"><img
1247   src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!"></a>
1248
1249   <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1250   <a href="http://llvm.org">The LLVM Compiler Infrastructure</a><br>
1251   Last modified: $Date: 2007-10-17 11:05:13 -0700 (Wed, 17 Oct 2007) $
1252 </address>
1253 </body>
1254 </html>