docs/tutorial/LangImpl4.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: Adding JIT and Optimizer Support</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: Adding JIT and Optimizer Support</div>
  15
  16 <ul>
  17 <li><a href="index.html">Up to Tutorial Index</a></li>
  18 <li>Chapter 4
  19   <ol>
  20     <li><a href="#intro">Chapter 4 Introduction</a></li>
  21     <li><a href="#trivialconstfold">Trivial Constant Folding</a></li>
  22     <li><a href="#optimizerpasses">LLVM Optimization Passes</a></li>
  23     <li><a href="#jit">Adding a JIT Compiler</a></li>
  24     <li><a href="#code">Full Code Listing</a></li>
  25   </ol>
  26 </li>
  27 <li><a href="LangImpl5.html">Chapter 5</a>: Extending the Language: Control
  28 Flow</li>
  29 </ul>
  30
  31 <div class="doc_author">
  32   <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a></p>
  33 </div>
  34
  35 <!-- *********************************************************************** -->
  36 <div class="doc_section"><a name="intro">Chapter 4 Introduction</a></div>
  37 <!-- *********************************************************************** -->
  38
  39 <div class="doc_text">
  40
  41 <p>Welcome to Chapter 4 of the "<a href="index.html">Implementing a language
  42 with LLVM</a>" tutorial.  Chapters 1-3 described the implementation of a simple
  43 language and added support for generating LLVM IR.  This chapter describes
  44 two new techniques: adding optimizer support to your language, and adding JIT
  45 compiler support.  These additions will demonstrate how to get nice, efficient code
  46 for the Kaleidoscope language.</p>
  47
  48 </div>
  49
  50 <!-- *********************************************************************** -->
  51 <div class="doc_section"><a name="trivialconstfold">Trivial Constant
  52 Folding</a></div>
  53 <!-- *********************************************************************** -->
  54
  55 <div class="doc_text">
  56
  57 <p>
  58 Our demonstration for Chapter 3 is elegant and easy to extend.  Unfortunately,
  59 it does not produce wonderful code.  The IRBuilder, however, does give us
  60 obvious optimizations when compiling simple code:</p>
  61
  62 <div class="doc_code">
  63 <pre>
  64 ready&gt; <b>def test(x) 1+2+x;</b>
  65 Read function definition:
  66 define double @test(double %x) {
  67 entry:
  68         %addtmp = fadd double 3.000000e+00, %x
  69         ret double %addtmp
  70 }
  71 </pre>
  72 </div>
  73
  74 <p>This code is not a literal transcription of the AST built by parsing the
  75 input. That would be:
  76
  77 <div class="doc_code">
  78 <pre>
  79 ready&gt; <b>def test(x) 1+2+x;</b>
  80 Read function definition:
  81 define double @test(double %x) {
  82 entry:
  83         %addtmp = fadd double 2.000000e+00, 1.000000e+00
  84         %addtmp1 = fadd double %addtmp, %x
  85         ret double %addtmp1
  86 }
  87 </pre>
  88 </div>
  89
  90 <p>Constant folding, as seen above, in particular, is a very common and very
  91 important optimization: so much so that many language implementors implement
  92 constant folding support in their AST representation.</p>
  93
  94 <p>With LLVM, you don't need this support in the AST.  Since all calls to build
  95 LLVM IR go through the LLVM IR builder, the builder itself checked to see if
  96 there was a constant folding opportunity when you call it.  If so, it just does
  97 the constant fold and return the constant instead of creating an instruction.
  98
  99 <p>Well, that was easy :).  In practice, we recommend always using
 100 <tt>IRBuilder</tt> when generating code like this.  It has no
 101 "syntactic overhead" for its use (you don't have to uglify your compiler with
 102 constant checks everywhere) and it can dramatically reduce the amount of
 103 LLVM IR that is generated in some cases (particular for languages with a macro
 104 preprocessor or that use a lot of constants).</p>
 105
 106 <p>On the other hand, the <tt>IRBuilder</tt> is limited by the fact
 107 that it does all of its analysis inline with the code as it is built.  If you
 108 take a slightly more complex example:</p>
 109
 110 <div class="doc_code">
 111 <pre>
 112 ready&gt; <b>def test(x) (1+2+x)*(x+(1+2));</b>
 113 ready> Read function definition:
 114 define double @test(double %x) {
 115 entry:
 116         %addtmp = fadd double 3.000000e+00, %x
 117         %addtmp1 = fadd double %x, 3.000000e+00
 118         %multmp = fmul double %addtmp, %addtmp1
 119         ret double %multmp
 120 }
 121 </pre>
 122 </div>
 123
 124 <p>In this case, the LHS and RHS of the multiplication are the same value.  We'd
 125 really like to see this generate "<tt>tmp = x+3; result = tmp*tmp;</tt>" instead
 126 of computing "<tt>x+3</tt>" twice.</p>
 127
 128 <p>Unfortunately, no amount of local analysis will be able to detect and correct
 129 this.  This requires two transformations: reassociation of expressions (to
 130 make the add's lexically identical) and Common Subexpression Elimination (CSE)
 131 to  delete the redundant add instruction.  Fortunately, LLVM provides a broad
 132 range of optimizations that you can use, in the form of "passes".</p>
 133
 134 </div>
 135
 136 <!-- *********************************************************************** -->
 137 <div class="doc_section"><a name="optimizerpasses">LLVM Optimization
 138  Passes</a></div>
 139 <!-- *********************************************************************** -->
 140
 141 <div class="doc_text">
 142
 143 <p>LLVM provides many optimization passes, which do many different sorts of
 144 things and have different tradeoffs.  Unlike other systems, LLVM doesn't hold
 145 to the mistaken notion that one set of optimizations is right for all languages
 146 and for all situations.  LLVM allows a compiler implementor to make complete
 147 decisions about what optimizations to use, in which order, and in what
 148 situation.</p>
 149
 150 <p>As a concrete example, LLVM supports both "whole module" passes, which look
 151 across as large of body of code as they can (often a whole file, but if run
 152 at link time, this can be a substantial portion of the whole program).  It also
 153 supports and includes "per-function" passes which just operate on a single
 154 function at a time, without looking at other functions.  For more information
 155 on passes and how they are run, see the <a href="../WritingAnLLVMPass.html">How
 156 to Write a Pass</a> document and the <a href="../Passes.html">List of LLVM
 157 Passes</a>.</p>
 158
 159 <p>For Kaleidoscope, we are currently generating functions on the fly, one at
 160 a time, as the user types them in.  We aren't shooting for the ultimate
 161 optimization experience in this setting, but we also want to catch the easy and
 162 quick stuff where possible.  As such, we will choose to run a few per-function
 163 optimizations as the user types the function in.  If we wanted to make a "static
 164 Kaleidoscope compiler", we would use exactly the code we have now, except that
 165 we would defer running the optimizer until the entire file has been parsed.</p>
 166
 167 <p>In order to get per-function optimizations going, we need to set up a
 168 <a href="../WritingAnLLVMPass.html#passmanager">FunctionPassManager</a> to hold and
 169 organize the LLVM optimizations that we want to run.  Once we have that, we can
 170 add a set of optimizations to run.  The code looks like this:</p>
 171
 172 <div class="doc_code">
 173 <pre>
 174   FunctionPassManager OurFPM(TheModule);
 175
 176   // Set up the optimizer pipeline.  Start with registering info about how the
 177   // target lays out data structures.
 178   OurFPM.add(new TargetData(*TheExecutionEngine->getTargetData()));
 179   // Provide basic AliasAnalysis support for GVN.
 180   OurFPM.add(createBasicAliasAnalysisPass());
 181   // Do simple "peephole" optimizations and bit-twiddling optzns.
 182   OurFPM.add(createInstructionCombiningPass());
 183   // Reassociate expressions.
 184   OurFPM.add(createReassociatePass());
 185   // Eliminate Common SubExpressions.
 186   OurFPM.add(createGVNPass());
 187   // Simplify the control flow graph (deleting unreachable blocks, etc).
 188   OurFPM.add(createCFGSimplificationPass());
 189
 190   OurFPM.doInitialization();
 191
 192   // Set the global so the code gen can use this.
 193   TheFPM = &amp;OurFPM;
 194
 195   // Run the main "interpreter loop" now.
 196   MainLoop();
 197 </pre>
 198 </div>
 199
 200 <p>This code defines a <tt>FunctionPassManager</tt>, "<tt>OurFPM</tt>".  It
 201 requires a pointer to the <tt>Module</tt> to construct itself.  Once it is set
 202 up, we use a series of "add" calls to add a bunch of LLVM passes.  The first
 203 pass is basically boilerplate, it adds a pass so that later optimizations know
 204 how the data structures in the program are laid out.  The
 205 "<tt>TheExecutionEngine</tt>" variable is related to the JIT, which we will get
 206 to in the next section.</p>
 207
 208 <p>In this case, we choose to add 4 optimization passes.  The passes we chose
 209 here are a pretty standard set of "cleanup" optimizations that are useful for
 210 a wide variety of code.  I won't delve into what they do but, believe me,
 211 they are a good starting place :).</p>
 212
 213 <p>Once the PassManager is set up, we need to make use of it.  We do this by
 214 running it after our newly created function is constructed (in
 215 <tt>FunctionAST::Codegen</tt>), but before it is returned to the client:</p>
 216
 217 <div class="doc_code">
 218 <pre>
 219   if (Value *RetVal = Body->Codegen()) {
 220     // Finish off the function.
 221     Builder.CreateRet(RetVal);
 222
 223     // Validate the generated code, checking for consistency.
 224     verifyFunction(*TheFunction);
 225
 226     <b>// Optimize the function.
 227     TheFPM-&gt;run(*TheFunction);</b>
 228
 229     return TheFunction;
 230   }
 231 </pre>
 232 </div>
 233
 234 <p>As you can see, this is pretty straightforward.  The
 235 <tt>FunctionPassManager</tt> optimizes and updates the LLVM Function* in place,
 236 improving (hopefully) its body.  With this in place, we can try our test above
 237 again:</p>
 238
 239 <div class="doc_code">
 240 <pre>
 241 ready&gt; <b>def test(x) (1+2+x)*(x+(1+2));</b>
 242 ready> Read function definition:
 243 define double @test(double %x) {
 244 entry:
 245         %addtmp = fadd double %x, 3.000000e+00
 246         %multmp = fmul double %addtmp, %addtmp
 247         ret double %multmp
 248 }
 249 </pre>
 250 </div>
 251
 252 <p>As expected, we now get our nicely optimized code, saving a floating point
 253 add instruction from every execution of this function.</p>
 254
 255 <p>LLVM provides a wide variety of optimizations that can be used in certain
 256 circumstances.  Some <a href="../Passes.html">documentation about the various
 257 passes</a> is available, but it isn't very complete.  Another good source of
 258 ideas can come from looking at the passes that <tt>llvm-gcc</tt> or
 259 <tt>llvm-ld</tt> run to get started.  The "<tt>opt</tt>" tool allows you to
 260 experiment with passes from the command line, so you can see if they do
 261 anything.</p>
 262
 263 <p>Now that we have reasonable code coming out of our front-end, lets talk about
 264 executing it!</p>
 265
 266 </div>
 267
 268 <!-- *********************************************************************** -->
 269 <div class="doc_section"><a name="jit">Adding a JIT Compiler</a></div>
 270 <!-- *********************************************************************** -->
 271
 272 <div class="doc_text">
 273
 274 <p>Code that is available in LLVM IR can have a wide variety of tools
 275 applied to it.  For example, you can run optimizations on it (as we did above),
 276 you can dump it out in textual or binary forms, you can compile the code to an
 277 assembly file (.s) for some target, or you can JIT compile it.  The nice thing
 278 about the LLVM IR representation is that it is the "common currency" between
 279 many different parts of the compiler.
 280 </p>
 281
 282 <p>In this section, we'll add JIT compiler support to our interpreter.  The
 283 basic idea that we want for Kaleidoscope is to have the user enter function
 284 bodies as they do now, but immediately evaluate the top-level expressions they
 285 type in.  For example, if they type in "1 + 2;", we should evaluate and print
 286 out 3.  If they define a function, they should be able to call it from the
 287 command line.</p>
 288
 289 <p>In order to do this, we first declare and initialize the JIT.  This is done
 290 by adding a global variable and a call in <tt>main</tt>:</p>
 291
 292 <div class="doc_code">
 293 <pre>
 294 <b>static ExecutionEngine *TheExecutionEngine;</b>
 295 ...
 296 int main() {
 297   ..
 298   <b>// Create the JIT.  This takes ownership of the module.
 299   TheExecutionEngine = EngineBuilder(TheModule).create();</b>
 300   ..
 301 }
 302 </pre>
 303 </div>
 304
 305 <p>This creates an abstract "Execution Engine" which can be either a JIT
 306 compiler or the LLVM interpreter.  LLVM will automatically pick a JIT compiler
 307 for you if one is available for your platform, otherwise it will fall back to
 308 the interpreter.</p>
 309
 310 <p>Once the <tt>ExecutionEngine</tt> is created, the JIT is ready to be used.
 311 There are a variety of APIs that are useful, but the simplest one is the
 312 "<tt>getPointerToFunction(F)</tt>" method.  This method JIT compiles the
 313 specified LLVM Function and returns a function pointer to the generated machine
 314 code.  In our case, this means that we can change the code that parses a
 315 top-level expression to look like this:</p>
 316
 317 <div class="doc_code">
 318 <pre>
 319 static void HandleTopLevelExpression() {
 320   // Evaluate a top-level expression into an anonymous function.
 321   if (FunctionAST *F = ParseTopLevelExpr()) {
 322     if (Function *LF = F-&gt;Codegen()) {
 323       LF->dump();  // Dump the function for exposition purposes.
 324
 325       <b>// JIT the function, returning a function pointer.
 326       void *FPtr = TheExecutionEngine-&gt;getPointerToFunction(LF);
 327
 328       // Cast it to the right type (takes no arguments, returns a double) so we
 329       // can call it as a native function.
 330       double (*FP)() = (double (*)())(intptr_t)FPtr;
 331       fprintf(stderr, "Evaluated to %f\n", FP());</b>
 332     }
 333 </pre>
 334 </div>
 335
 336 <p>Recall that we compile top-level expressions into a self-contained LLVM
 337 function that takes no arguments and returns the computed double.  Because the
 338 LLVM JIT compiler matches the native platform ABI, this means that you can just
 339 cast the result pointer to a function pointer of that type and call it directly.
 340 This means, there is no difference between JIT compiled code and native machine
 341 code that is statically linked into your application.</p>
 342
 343 <p>With just these two changes, lets see how Kaleidoscope works now!</p>
 344
 345 <div class="doc_code">
 346 <pre>
 347 ready&gt; <b>4+5;</b>
 348 define double @""() {
 349 entry:
 350         ret double 9.000000e+00
 351 }
 352
 353 <em>Evaluated to 9.000000</em>
 354 </pre>
 355 </div>
 356
 357 <p>Well this looks like it is basically working.  The dump of the function
 358 shows the "no argument function that always returns double" that we synthesize
 359 for each top-level expression that is typed in.  This demonstrates very basic
 360 functionality, but can we do more?</p>
 361
 362 <div class="doc_code">
 363 <pre>
 364 ready&gt; <b>def testfunc(x y) x + y*2; </b>
 365 Read function definition:
 366 define double @testfunc(double %x, double %y) {
 367 entry:
 368         %multmp = fmul double %y, 2.000000e+00
 369         %addtmp = fadd double %multmp, %x
 370         ret double %addtmp
 371 }
 372
 373 ready&gt; <b>testfunc(4, 10);</b>
 374 define double @""() {
 375 entry:
 376         %calltmp = call double @testfunc(double 4.000000e+00, double 1.000000e+01)
 377         ret double %calltmp
 378 }
 379
 380 <em>Evaluated to 24.000000</em>
 381 </pre>
 382 </div>
 383
 384 <p>This illustrates that we can now call user code, but there is something a bit
 385 subtle going on here.  Note that we only invoke the JIT on the anonymous
 386 functions that <em>call testfunc</em>, but we never invoked it
 387 on <em>testfunc</em> itself.  What actually happened here is that the JIT
 388 scanned for all non-JIT'd functions transitively called from the anonymous
 389 function and compiled all of them before returning
 390 from <tt>getPointerToFunction()</tt>.</p>
 391
 392 <p>The JIT provides a number of other more advanced interfaces for things like
 393 freeing allocated machine code, rejit'ing functions to update them, etc.
 394 However, even with this simple code, we get some surprisingly powerful
 395 capabilities - check this out (I removed the dump of the anonymous functions,
 396 you should get the idea by now :) :</p>
 397
 398 <div class="doc_code">
 399 <pre>
 400 ready&gt; <b>extern sin(x);</b>
 401 Read extern:
 402 declare double @sin(double)
 403
 404 ready&gt; <b>extern cos(x);</b>
 405 Read extern:
 406 declare double @cos(double)
 407
 408 ready&gt; <b>sin(1.0);</b>
 409 <em>Evaluated to 0.841471</em>
 410
 411 ready&gt; <b>def foo(x) sin(x)*sin(x) + cos(x)*cos(x);</b>
 412 Read function definition:
 413 define double @foo(double %x) {
 414 entry:
 415         %calltmp = call double @sin(double %x)
 416         %multmp = fmul double %calltmp, %calltmp
 417         %calltmp2 = call double @cos(double %x)
 418         %multmp4 = fmul double %calltmp2, %calltmp2
 419         %addtmp = fadd double %multmp, %multmp4
 420         ret double %addtmp
 421 }
 422
 423 ready&gt; <b>foo(4.0);</b>
 424 <em>Evaluated to 1.000000</em>
 425 </pre>
 426 </div>
 427
 428 <p>Whoa, how does the JIT know about sin and cos?  The answer is surprisingly
 429 simple: in this
 430 example, the JIT started execution of a function and got to a function call.  It
 431 realized that the function was not yet JIT compiled and invoked the standard set
 432 of routines to resolve the function.  In this case, there is no body defined
 433 for the function, so the JIT ended up calling "<tt>dlsym("sin")</tt>" on the
 434 Kaleidoscope process itself.
 435 Since "<tt>sin</tt>" is defined within the JIT's address space, it simply
 436 patches up calls in the module to call the libm version of <tt>sin</tt>
 437 directly.</p>
 438
 439 <p>The LLVM JIT provides a number of interfaces (look in the
 440 <tt>ExecutionEngine.h</tt> file) for controlling how unknown functions get
 441 resolved.  It allows you to establish explicit mappings between IR objects and
 442 addresses (useful for LLVM global variables that you want to map to static
 443 tables, for example), allows you to dynamically decide on the fly based on the
 444 function name, and even allows you to have the JIT compile functions lazily the
 445 first time they're called.</p>
 446
 447 <p>One interesting application of this is that we can now extend the language
 448 by writing arbitrary C++ code to implement operations.  For example, if we add:
 449 </p>
 450
 451 <div class="doc_code">
 452 <pre>
 453 /// putchard - putchar that takes a double and returns 0.
 454 extern "C"
 455 double putchard(double X) {
 456   putchar((char)X);
 457   return 0;
 458 }
 459 </pre>
 460 </div>
 461
 462 <p>Now we can produce simple output to the console by using things like:
 463 "<tt>extern putchard(x); putchard(120);</tt>", which prints a lowercase 'x' on
 464 the console (120 is the ASCII code for 'x').  Similar code could be used to
 465 implement file I/O, console input, and many other capabilities in
 466 Kaleidoscope.</p>
 467
 468 <p>This completes the JIT and optimizer chapter of the Kaleidoscope tutorial. At
 469 this point, we can compile a non-Turing-complete programming language, optimize
 470 and JIT compile it in a user-driven way.  Next up we'll look into <a
 471 href="LangImpl5.html">extending the language with control flow constructs</a>,
 472 tackling some interesting LLVM IR issues along the way.</p>
 473
 474 </div>
 475
 476 <!-- *********************************************************************** -->
 477 <div class="doc_section"><a name="code">Full Code Listing</a></div>
 478 <!-- *********************************************************************** -->
 479
 480 <div class="doc_text">
 481
 482 <p>
 483 Here is the complete code listing for our running example, enhanced with the
 484 LLVM JIT and optimizer.  To build this example, use:
 485 </p>
 486
 487 <div class="doc_code">
 488 <pre>
 489    # Compile
 490    g++ -g toy.cpp `llvm-config --cppflags --ldflags --libs core jit native` -O3 -o toy
 491    # Run
 492    ./toy
 493 </pre>
 494 </div>
 495
 496 <p>
 497 If you are compiling this on Linux, make sure to add the "-rdynamic" option
 498 as well.  This makes sure that the external functions are resolved properly
 499 at runtime.</p>
 500
 501 <p>Here is the code:</p>
 502
 503 <div class="doc_code">
 504 <pre>
 505 #include "llvm/DerivedTypes.h"
 506 #include "llvm/ExecutionEngine/ExecutionEngine.h"
 507 #include "llvm/ExecutionEngine/JIT.h"
 508 #include "llvm/LLVMContext.h"
 509 #include "llvm/Module.h"
 510 #include "llvm/PassManager.h"
 511 #include "llvm/Analysis/Verifier.h"
 512 #include "llvm/Analysis/Passes.h"
 513 #include "llvm/Target/TargetData.h"
 514 #include "llvm/Target/TargetSelect.h"
 515 #include "llvm/Transforms/Scalar.h"
 516 #include "llvm/Support/IRBuilder.h"
 517 #include &lt;cstdio&gt;
 518 #include &lt;string&gt;
 519 #include &lt;map&gt;
 520 #include &lt;vector&gt;
 521 using namespace llvm;
 522
 523 //===----------------------------------------------------------------------===//
 524 // Lexer
 525 //===----------------------------------------------------------------------===//
 526
 527 // The lexer returns tokens [0-255] if it is an unknown character, otherwise one
 528 // of these for known things.
 529 enum Token {
 530   tok_eof = -1,
 531
 532   // commands
 533   tok_def = -2, tok_extern = -3,
 534
 535   // primary
 536   tok_identifier = -4, tok_number = -5
 537 };
 538
 539 static std::string IdentifierStr;  // Filled in if tok_identifier
 540 static double NumVal;              // Filled in if tok_number
 541
 542 /// gettok - Return the next token from standard input.
 543 static int gettok() {
 544   static int LastChar = ' ';
 545
 546   // Skip any whitespace.
 547   while (isspace(LastChar))
 548     LastChar = getchar();
 549
 550   if (isalpha(LastChar)) { // identifier: [a-zA-Z][a-zA-Z0-9]*
 551     IdentifierStr = LastChar;
 552     while (isalnum((LastChar = getchar())))
 553       IdentifierStr += LastChar;
 554
 555     if (IdentifierStr == "def") return tok_def;
 556     if (IdentifierStr == "extern") return tok_extern;
 557     return tok_identifier;
 558   }
 559
 560   if (isdigit(LastChar) || LastChar == '.') {   // Number: [0-9.]+
 561     std::string NumStr;
 562     do {
 563       NumStr += LastChar;
 564       LastChar = getchar();
 565     } while (isdigit(LastChar) || LastChar == '.');
 566
 567     NumVal = strtod(NumStr.c_str(), 0);
 568     return tok_number;
 569   }
 570
 571   if (LastChar == '#') {
 572     // Comment until end of line.
 573     do LastChar = getchar();
 574     while (LastChar != EOF &amp;&amp; LastChar != '\n' &amp;&amp; LastChar != '\r');
 575
 576     if (LastChar != EOF)
 577       return gettok();
 578   }
 579
 580   // Check for end of file.  Don't eat the EOF.
 581   if (LastChar == EOF)
 582     return tok_eof;
 583
 584   // Otherwise, just return the character as its ascii value.
 585   int ThisChar = LastChar;
 586   LastChar = getchar();
 587   return ThisChar;
 588 }
 589
 590 //===----------------------------------------------------------------------===//
 591 // Abstract Syntax Tree (aka Parse Tree)
 592 //===----------------------------------------------------------------------===//
 593
 594 /// ExprAST - Base class for all expression nodes.
 595 class ExprAST {
 596 public:
 597   virtual ~ExprAST() {}
 598   virtual Value *Codegen() = 0;
 599 };
 600
 601 /// NumberExprAST - Expression class for numeric literals like "1.0".
 602 class NumberExprAST : public ExprAST {
 603   double Val;
 604 public:
 605   NumberExprAST(double val) : Val(val) {}
 606   virtual Value *Codegen();
 607 };
 608
 609 /// VariableExprAST - Expression class for referencing a variable, like "a".
 610 class VariableExprAST : public ExprAST {
 611   std::string Name;
 612 public:
 613   VariableExprAST(const std::string &amp;name) : Name(name) {}
 614   virtual Value *Codegen();
 615 };
 616
 617 /// BinaryExprAST - Expression class for a binary operator.
 618 class BinaryExprAST : public ExprAST {
 619   char Op;
 620   ExprAST *LHS, *RHS;
 621 public:
 622   BinaryExprAST(char op, ExprAST *lhs, ExprAST *rhs)
 623     : Op(op), LHS(lhs), RHS(rhs) {}
 624   virtual Value *Codegen();
 625 };
 626
 627 /// CallExprAST - Expression class for function calls.
 628 class CallExprAST : public ExprAST {
 629   std::string Callee;
 630   std::vector&lt;ExprAST*&gt; Args;
 631 public:
 632   CallExprAST(const std::string &amp;callee, std::vector&lt;ExprAST*&gt; &amp;args)
 633     : Callee(callee), Args(args) {}
 634   virtual Value *Codegen();
 635 };
 636
 637 /// PrototypeAST - This class represents the "prototype" for a function,
 638 /// which captures its name, and its argument names (thus implicitly the number
 639 /// of arguments the function takes).
 640 class PrototypeAST {
 641   std::string Name;
 642   std::vector&lt;std::string&gt; Args;
 643 public:
 644   PrototypeAST(const std::string &amp;name, const std::vector&lt;std::string&gt; &amp;args)
 645     : Name(name), Args(args) {}
 646
 647   Function *Codegen();
 648 };
 649
 650 /// FunctionAST - This class represents a function definition itself.
 651 class FunctionAST {
 652   PrototypeAST *Proto;
 653   ExprAST *Body;
 654 public:
 655   FunctionAST(PrototypeAST *proto, ExprAST *body)
 656     : Proto(proto), Body(body) {}
 657
 658   Function *Codegen();
 659 };
 660
 661 //===----------------------------------------------------------------------===//
 662 // Parser
 663 //===----------------------------------------------------------------------===//
 664
 665 /// CurTok/getNextToken - Provide a simple token buffer.  CurTok is the current
 666 /// token the parser is looking at.  getNextToken reads another token from the
 667 /// lexer and updates CurTok with its results.
 668 static int CurTok;
 669 static int getNextToken() {
 670   return CurTok = gettok();
 671 }
 672
 673 /// BinopPrecedence - This holds the precedence for each binary operator that is
 674 /// defined.
 675 static std::map&lt;char, int&gt; BinopPrecedence;
 676
 677 /// GetTokPrecedence - Get the precedence of the pending binary operator token.
 678 static int GetTokPrecedence() {
 679   if (!isascii(CurTok))
 680     return -1;
 681
 682   // Make sure it's a declared binop.
 683   int TokPrec = BinopPrecedence[CurTok];
 684   if (TokPrec &lt;= 0) return -1;
 685   return TokPrec;
 686 }
 687
 688 /// Error* - These are little helper functions for error handling.
 689 ExprAST *Error(const char *Str) { fprintf(stderr, "Error: %s\n", Str);return 0;}
 690 PrototypeAST *ErrorP(const char *Str) { Error(Str); return 0; }
 691 FunctionAST *ErrorF(const char *Str) { Error(Str); return 0; }
 692
 693 static ExprAST *ParseExpression();
 694
 695 /// identifierexpr
 696 ///   ::= identifier
 697 ///   ::= identifier '(' expression* ')'
 698 static ExprAST *ParseIdentifierExpr() {
 699   std::string IdName = IdentifierStr;
 700
 701   getNextToken();  // eat identifier.
 702
 703   if (CurTok != '(') // Simple variable ref.
 704     return new VariableExprAST(IdName);
 705
 706   // Call.
 707   getNextToken();  // eat (
 708   std::vector&lt;ExprAST*&gt; Args;
 709   if (CurTok != ')') {
 710     while (1) {
 711       ExprAST *Arg = ParseExpression();
 712       if (!Arg) return 0;
 713       Args.push_back(Arg);
 714
 715       if (CurTok == ')') break;
 716
 717       if (CurTok != ',')
 718         return Error("Expected ')' or ',' in argument list");
 719       getNextToken();
 720     }
 721   }
 722
 723   // Eat the ')'.
 724   getNextToken();
 725
 726   return new CallExprAST(IdName, Args);
 727 }
 728
 729 /// numberexpr ::= number
 730 static ExprAST *ParseNumberExpr() {
 731   ExprAST *Result = new NumberExprAST(NumVal);
 732   getNextToken(); // consume the number
 733   return Result;
 734 }
 735
 736 /// parenexpr ::= '(' expression ')'
 737 static ExprAST *ParseParenExpr() {
 738   getNextToken();  // eat (.
 739   ExprAST *V = ParseExpression();
 740   if (!V) return 0;
 741
 742   if (CurTok != ')')
 743     return Error("expected ')'");
 744   getNextToken();  // eat ).
 745   return V;
 746 }
 747
 748 /// primary
 749 ///   ::= identifierexpr
 750 ///   ::= numberexpr
 751 ///   ::= parenexpr
 752 static ExprAST *ParsePrimary() {
 753   switch (CurTok) {
 754   default: return Error("unknown token when expecting an expression");
 755   case tok_identifier: return ParseIdentifierExpr();
 756   case tok_number:     return ParseNumberExpr();
 757   case '(':            return ParseParenExpr();
 758   }
 759 }
 760
 761 /// binoprhs
 762 ///   ::= ('+' primary)*
 763 static ExprAST *ParseBinOpRHS(int ExprPrec, ExprAST *LHS) {
 764   // If this is a binop, find its precedence.
 765   while (1) {
 766     int TokPrec = GetTokPrecedence();
 767
 768     // If this is a binop that binds at least as tightly as the current binop,
 769     // consume it, otherwise we are done.
 770     if (TokPrec &lt; ExprPrec)
 771       return LHS;
 772
 773     // Okay, we know this is a binop.
 774     int BinOp = CurTok;
 775     getNextToken();  // eat binop
 776
 777     // Parse the primary expression after the binary operator.
 778     ExprAST *RHS = ParsePrimary();
 779     if (!RHS) return 0;
 780
 781     // If BinOp binds less tightly with RHS than the operator after RHS, let
 782     // the pending operator take RHS as its LHS.
 783     int NextPrec = GetTokPrecedence();
 784     if (TokPrec &lt; NextPrec) {
 785       RHS = ParseBinOpRHS(TokPrec+1, RHS);
 786       if (RHS == 0) return 0;
 787     }
 788
 789     // Merge LHS/RHS.
 790     LHS = new BinaryExprAST(BinOp, LHS, RHS);
 791   }
 792 }
 793
 794 /// expression
 795 ///   ::= primary binoprhs
 796 ///
 797 static ExprAST *ParseExpression() {
 798   ExprAST *LHS = ParsePrimary();
 799   if (!LHS) return 0;
 800
 801   return ParseBinOpRHS(0, LHS);
 802 }
 803
 804 /// prototype
 805 ///   ::= id '(' id* ')'
 806 static PrototypeAST *ParsePrototype() {
 807   if (CurTok != tok_identifier)
 808     return ErrorP("Expected function name in prototype");
 809
 810   std::string FnName = IdentifierStr;
 811   getNextToken();
 812
 813   if (CurTok != '(')
 814     return ErrorP("Expected '(' in prototype");
 815
 816   std::vector&lt;std::string&gt; ArgNames;
 817   while (getNextToken() == tok_identifier)
 818     ArgNames.push_back(IdentifierStr);
 819   if (CurTok != ')')
 820     return ErrorP("Expected ')' in prototype");
 821
 822   // success.
 823   getNextToken();  // eat ')'.
 824
 825   return new PrototypeAST(FnName, ArgNames);
 826 }
 827
 828 /// definition ::= 'def' prototype expression
 829 static FunctionAST *ParseDefinition() {
 830   getNextToken();  // eat def.
 831   PrototypeAST *Proto = ParsePrototype();
 832   if (Proto == 0) return 0;
 833
 834   if (ExprAST *E = ParseExpression())
 835     return new FunctionAST(Proto, E);
 836   return 0;
 837 }
 838
 839 /// toplevelexpr ::= expression
 840 static FunctionAST *ParseTopLevelExpr() {
 841   if (ExprAST *E = ParseExpression()) {
 842     // Make an anonymous proto.
 843     PrototypeAST *Proto = new PrototypeAST("", std::vector&lt;std::string&gt;());
 844     return new FunctionAST(Proto, E);
 845   }
 846   return 0;
 847 }
 848
 849 /// external ::= 'extern' prototype
 850 static PrototypeAST *ParseExtern() {
 851   getNextToken();  // eat extern.
 852   return ParsePrototype();
 853 }
 854
 855 //===----------------------------------------------------------------------===//
 856 // Code Generation
 857 //===----------------------------------------------------------------------===//
 858
 859 static Module *TheModule;
 860 static IRBuilder&lt;&gt; Builder(getGlobalContext());
 861 static std::map&lt;std::string, Value*&gt; NamedValues;
 862 static FunctionPassManager *TheFPM;
 863
 864 Value *ErrorV(const char *Str) { Error(Str); return 0; }
 865
 866 Value *NumberExprAST::Codegen() {
 867   return ConstantFP::get(getGlobalContext(), APFloat(Val));
 868 }
 869
 870 Value *VariableExprAST::Codegen() {
 871   // Look this variable up in the function.
 872   Value *V = NamedValues[Name];
 873   return V ? V : ErrorV("Unknown variable name");
 874 }
 875
 876 Value *BinaryExprAST::Codegen() {
 877   Value *L = LHS-&gt;Codegen();
 878   Value *R = RHS-&gt;Codegen();
 879   if (L == 0 || R == 0) return 0;
 880
 881   switch (Op) {
 882   case '+': return Builder.CreateFAdd(L, R, "addtmp");
 883   case '-': return Builder.CreateFSub(L, R, "subtmp");
 884   case '*': return Builder.CreateFMul(L, R, "multmp");
 885   case '&lt;':
 886     L = Builder.CreateFCmpULT(L, R, "cmptmp");
 887     // Convert bool 0/1 to double 0.0 or 1.0
 888     return Builder.CreateUIToFP(L, Type::getDoubleTy(getGlobalContext()),
 889                                 "booltmp");
 890   default: return ErrorV("invalid binary operator");
 891   }
 892 }
 893
 894 Value *CallExprAST::Codegen() {
 895   // Look up the name in the global module table.
 896   Function *CalleeF = TheModule-&gt;getFunction(Callee);
 897   if (CalleeF == 0)
 898     return ErrorV("Unknown function referenced");
 899
 900   // If argument mismatch error.
 901   if (CalleeF-&gt;arg_size() != Args.size())
 902     return ErrorV("Incorrect # arguments passed");
 903
 904   std::vector&lt;Value*&gt; ArgsV;
 905   for (unsigned i = 0, e = Args.size(); i != e; ++i) {
 906     ArgsV.push_back(Args[i]-&gt;Codegen());
 907     if (ArgsV.back() == 0) return 0;
 908   }
 909
 910   return Builder.CreateCall(CalleeF, ArgsV.begin(), ArgsV.end(), "calltmp");
 911 }
 912
 913 Function *PrototypeAST::Codegen() {
 914   // Make the function type:  double(double,double) etc.
 915   std::vector&lt;const Type*&gt; Doubles(Args.size(),
 916                                    Type::getDoubleTy(getGlobalContext()));
 917   FunctionType *FT = FunctionType::get(Type::getDoubleTy(getGlobalContext()),
 918                                        Doubles, false);
 919
 920   Function *F = Function::Create(FT, Function::ExternalLinkage, Name, TheModule);
 921
 922   // If F conflicted, there was already something named 'Name'.  If it has a
 923   // body, don't allow redefinition or reextern.
 924   if (F-&gt;getName() != Name) {
 925     // Delete the one we just made and get the existing one.
 926     F-&gt;eraseFromParent();
 927     F = TheModule-&gt;getFunction(Name);
 928
 929     // If F already has a body, reject this.
 930     if (!F-&gt;empty()) {
 931       ErrorF("redefinition of function");
 932       return 0;
 933     }
 934
 935     // If F took a different number of args, reject.
 936     if (F-&gt;arg_size() != Args.size()) {
 937       ErrorF("redefinition of function with different # args");
 938       return 0;
 939     }
 940   }
 941
 942   // Set names for all arguments.
 943   unsigned Idx = 0;
 944   for (Function::arg_iterator AI = F-&gt;arg_begin(); Idx != Args.size();
 945        ++AI, ++Idx) {
 946     AI-&gt;setName(Args[Idx]);
 947
 948     // Add arguments to variable symbol table.
 949     NamedValues[Args[Idx]] = AI;
 950   }
 951
 952   return F;
 953 }
 954
 955 Function *FunctionAST::Codegen() {
 956   NamedValues.clear();
 957
 958   Function *TheFunction = Proto-&gt;Codegen();
 959   if (TheFunction == 0)
 960     return 0;
 961
 962   // Create a new basic block to start insertion into.
 963   BasicBlock *BB = BasicBlock::Create(getGlobalContext(), "entry", TheFunction);
 964   Builder.SetInsertPoint(BB);
 965
 966   if (Value *RetVal = Body-&gt;Codegen()) {
 967     // Finish off the function.
 968     Builder.CreateRet(RetVal);
 969
 970     // Validate the generated code, checking for consistency.
 971     verifyFunction(*TheFunction);
 972
 973     // Optimize the function.
 974     TheFPM-&gt;run(*TheFunction);
 975
 976     return TheFunction;
 977   }
 978
 979   // Error reading body, remove function.
 980   TheFunction-&gt;eraseFromParent();
 981   return 0;
 982 }
 983
 984 //===----------------------------------------------------------------------===//
 985 // Top-Level parsing and JIT Driver
 986 //===----------------------------------------------------------------------===//
 987
 988 static ExecutionEngine *TheExecutionEngine;
 989
 990 static void HandleDefinition() {
 991   if (FunctionAST *F = ParseDefinition()) {
 992     if (Function *LF = F-&gt;Codegen()) {
 993       fprintf(stderr, "Read function definition:");
 994       LF-&gt;dump();
 995     }
 996   } else {
 997     // Skip token for error recovery.
 998     getNextToken();
 999   }
1000 }
1001
1002 static void HandleExtern() {
1003   if (PrototypeAST *P = ParseExtern()) {
1004     if (Function *F = P-&gt;Codegen()) {
1005       fprintf(stderr, "Read extern: ");
1006       F-&gt;dump();
1007     }
1008   } else {
1009     // Skip token for error recovery.
1010     getNextToken();
1011   }
1012 }
1013
1014 static void HandleTopLevelExpression() {
1015   // Evaluate a top-level expression into an anonymous function.
1016   if (FunctionAST *F = ParseTopLevelExpr()) {
1017     if (Function *LF = F-&gt;Codegen()) {
1018       // JIT the function, returning a function pointer.
1019       void *FPtr = TheExecutionEngine-&gt;getPointerToFunction(LF);
1020
1021       // Cast it to the right type (takes no arguments, returns a double) so we
1022       // can call it as a native function.
1023       double (*FP)() = (double (*)())(intptr_t)FPtr;
1024       fprintf(stderr, "Evaluated to %f\n", FP());
1025     }
1026   } else {
1027     // Skip token for error recovery.
1028     getNextToken();
1029   }
1030 }
1031
1032 /// top ::= definition | external | expression | ';'
1033 static void MainLoop() {
1034   while (1) {
1035     fprintf(stderr, "ready&gt; ");
1036     switch (CurTok) {
1037     case tok_eof:    return;
1038     case ';':        getNextToken(); break;  // ignore top-level semicolons.
1039     case tok_def:    HandleDefinition(); break;
1040     case tok_extern: HandleExtern(); break;
1041     default:         HandleTopLevelExpression(); break;
1042     }
1043   }
1044 }
1045
1046 //===----------------------------------------------------------------------===//
1047 // "Library" functions that can be "extern'd" from user code.
1048 //===----------------------------------------------------------------------===//
1049
1050 /// putchard - putchar that takes a double and returns 0.
1051 extern "C"
1052 double putchard(double X) {
1053   putchar((char)X);
1054   return 0;
1055 }
1056
1057 //===----------------------------------------------------------------------===//
1058 // Main driver code.
1059 //===----------------------------------------------------------------------===//
1060
1061 int main() {
1062   InitializeNativeTarget();
1063   LLVMContext &amp;Context = getGlobalContext();
1064
1065   // Install standard binary operators.
1066   // 1 is lowest precedence.
1067   BinopPrecedence['&lt;'] = 10;
1068   BinopPrecedence['+'] = 20;
1069   BinopPrecedence['-'] = 20;
1070   BinopPrecedence['*'] = 40;  // highest.
1071
1072   // Prime the first token.
1073   fprintf(stderr, "ready&gt; ");
1074   getNextToken();
1075
1076   // Make the module, which holds all the code.
1077   TheModule = new Module("my cool jit", Context);
1078
1079   // Create the JIT.  This takes ownership of the module.
1080   std::string ErrStr;
1081   TheExecutionEngine = EngineBuilder(TheModule).setErrorStr(&ErrStr).create();
1082   if (!TheExecutionEngine) {
1083     fprintf(stderr, "Could not create ExecutionEngine: %s\n", ErrStr.c_str());
1084     exit(1);
1085   }
1086
1087   FunctionPassManager OurFPM(TheModule);
1088
1089   // Set up the optimizer pipeline.  Start with registering info about how the
1090   // target lays out data structures.
1091   OurFPM.add(new TargetData(*TheExecutionEngine-&gt;getTargetData()));
1092   // Provide basic AliasAnalysis support for GVN.
1093   OurFPM.add(createBasicAliasAnalysisPass());
1094   // Do simple "peephole" optimizations and bit-twiddling optzns.
1095   OurFPM.add(createInstructionCombiningPass());
1096   // Reassociate expressions.
1097   OurFPM.add(createReassociatePass());
1098   // Eliminate Common SubExpressions.
1099   OurFPM.add(createGVNPass());
1100   // Simplify the control flow graph (deleting unreachable blocks, etc).
1101   OurFPM.add(createCFGSimplificationPass());
1102
1103   OurFPM.doInitialization();
1104
1105   // Set the global so the code gen can use this.
1106   TheFPM = &amp;OurFPM;
1107
1108   // Run the main "interpreter loop" now.
1109   MainLoop();
1110
1111   TheFPM = 0;
1112
1113   // Print out all of the generated code.
1114   TheModule-&gt;dump();
1115
1116   return 0;
1117 }
1118 </pre>
1119 </div>
1120
1121 <a href="LangImpl5.html">Next: Extending the language: control flow</a>
1122 </div>
1123
1124 <!-- *********************************************************************** -->
1125 <hr>
1126 <address>
1127   <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1128   src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
1129   <a href="http://validator.w3.org/check/referer"><img
1130   src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!"></a>
1131
1132   <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1133   <a href="http://llvm.org">The LLVM Compiler Infrastructure</a><br>
1134   Last modified: $Date$
1135 </address>
1136 </body>
1137 </html>