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 <h1>Kaleidoscope: Adding JIT and Optimizer Support</h1>
  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 <h2><a name="intro">Chapter 4 Introduction</a></h2>
  37 <!-- *********************************************************************** -->
  38
  39 <div>
  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 <h2><a name="trivialconstfold">Trivial Constant Folding</a></h2>
  52 <!-- *********************************************************************** -->
  53
  54 <div>
  55
  56 <p>
  57 Our demonstration for Chapter 3 is elegant and easy to extend.  Unfortunately,
  58 it does not produce wonderful code.  The IRBuilder, however, does give us
  59 obvious optimizations when compiling simple code:</p>
  60
  61 <div class="doc_code">
  62 <pre>
  63 ready&gt; <b>def test(x) 1+2+x;</b>
  64 Read function definition:
  65 define double @test(double %x) {
  66 entry:
  67         %addtmp = fadd double 3.000000e+00, %x
  68         ret double %addtmp
  69 }
  70 </pre>
  71 </div>
  72
  73 <p>This code is not a literal transcription of the AST built by parsing the
  74 input. That would be:
  75
  76 <div class="doc_code">
  77 <pre>
  78 ready&gt; <b>def test(x) 1+2+x;</b>
  79 Read function definition:
  80 define double @test(double %x) {
  81 entry:
  82         %addtmp = fadd double 2.000000e+00, 1.000000e+00
  83         %addtmp1 = fadd double %addtmp, %x
  84         ret double %addtmp1
  85 }
  86 </pre>
  87 </div>
  88
  89 <p>Constant folding, as seen above, in particular, is a very common and very
  90 important optimization: so much so that many language implementors implement
  91 constant folding support in their AST representation.</p>
  92
  93 <p>With LLVM, you don't need this support in the AST.  Since all calls to build
  94 LLVM IR go through the LLVM IR builder, the builder itself checked to see if
  95 there was a constant folding opportunity when you call it.  If so, it just does
  96 the constant fold and return the constant instead of creating an instruction.
  97
  98 <p>Well, that was easy :).  In practice, we recommend always using
  99 <tt>IRBuilder</tt> when generating code like this.  It has no
 100 "syntactic overhead" for its use (you don't have to uglify your compiler with
 101 constant checks everywhere) and it can dramatically reduce the amount of
 102 LLVM IR that is generated in some cases (particular for languages with a macro
 103 preprocessor or that use a lot of constants).</p>
 104
 105 <p>On the other hand, the <tt>IRBuilder</tt> is limited by the fact
 106 that it does all of its analysis inline with the code as it is built.  If you
 107 take a slightly more complex example:</p>
 108
 109 <div class="doc_code">
 110 <pre>
 111 ready&gt; <b>def test(x) (1+2+x)*(x+(1+2));</b>
 112 ready> Read function definition:
 113 define double @test(double %x) {
 114 entry:
 115         %addtmp = fadd double 3.000000e+00, %x
 116         %addtmp1 = fadd double %x, 3.000000e+00
 117         %multmp = fmul double %addtmp, %addtmp1
 118         ret double %multmp
 119 }
 120 </pre>
 121 </div>
 122
 123 <p>In this case, the LHS and RHS of the multiplication are the same value.  We'd
 124 really like to see this generate "<tt>tmp = x+3; result = tmp*tmp;</tt>" instead
 125 of computing "<tt>x+3</tt>" twice.</p>
 126
 127 <p>Unfortunately, no amount of local analysis will be able to detect and correct
 128 this.  This requires two transformations: reassociation of expressions (to
 129 make the add's lexically identical) and Common Subexpression Elimination (CSE)
 130 to  delete the redundant add instruction.  Fortunately, LLVM provides a broad
 131 range of optimizations that you can use, in the form of "passes".</p>
 132
 133 </div>
 134
 135 <!-- *********************************************************************** -->
 136 <h2><a name="optimizerpasses">LLVM Optimization Passes</a></h2>
 137 <!-- *********************************************************************** -->
 138
 139 <div>
 140
 141 <p>LLVM provides many optimization passes, which do many different sorts of
 142 things and have different tradeoffs.  Unlike other systems, LLVM doesn't hold
 143 to the mistaken notion that one set of optimizations is right for all languages
 144 and for all situations.  LLVM allows a compiler implementor to make complete
 145 decisions about what optimizations to use, in which order, and in what
 146 situation.</p>
 147
 148 <p>As a concrete example, LLVM supports both "whole module" passes, which look
 149 across as large of body of code as they can (often a whole file, but if run
 150 at link time, this can be a substantial portion of the whole program).  It also
 151 supports and includes "per-function" passes which just operate on a single
 152 function at a time, without looking at other functions.  For more information
 153 on passes and how they are run, see the <a href="../WritingAnLLVMPass.html">How
 154 to Write a Pass</a> document and the <a href="../Passes.html">List of LLVM
 155 Passes</a>.</p>
 156
 157 <p>For Kaleidoscope, we are currently generating functions on the fly, one at
 158 a time, as the user types them in.  We aren't shooting for the ultimate
 159 optimization experience in this setting, but we also want to catch the easy and
 160 quick stuff where possible.  As such, we will choose to run a few per-function
 161 optimizations as the user types the function in.  If we wanted to make a "static
 162 Kaleidoscope compiler", we would use exactly the code we have now, except that
 163 we would defer running the optimizer until the entire file has been parsed.</p>
 164
 165 <p>In order to get per-function optimizations going, we need to set up a
 166 <a href="../WritingAnLLVMPass.html#passmanager">FunctionPassManager</a> to hold and
 167 organize the LLVM optimizations that we want to run.  Once we have that, we can
 168 add a set of optimizations to run.  The code looks like this:</p>
 169
 170 <div class="doc_code">
 171 <pre>
 172   FunctionPassManager OurFPM(TheModule);
 173
 174   // Set up the optimizer pipeline.  Start with registering info about how the
 175   // target lays out data structures.
 176   OurFPM.add(new TargetData(*TheExecutionEngine->getTargetData()));
 177   // Provide basic AliasAnalysis support for GVN.
 178   OurFPM.add(createBasicAliasAnalysisPass());
 179   // Do simple "peephole" optimizations and bit-twiddling optzns.
 180   OurFPM.add(createInstructionCombiningPass());
 181   // Reassociate expressions.
 182   OurFPM.add(createReassociatePass());
 183   // Eliminate Common SubExpressions.
 184   OurFPM.add(createGVNPass());
 185   // Simplify the control flow graph (deleting unreachable blocks, etc).
 186   OurFPM.add(createCFGSimplificationPass());
 187
 188   OurFPM.doInitialization();
 189
 190   // Set the global so the code gen can use this.
 191   TheFPM = &amp;OurFPM;
 192
 193   // Run the main "interpreter loop" now.
 194   MainLoop();
 195 </pre>
 196 </div>
 197
 198 <p>This code defines a <tt>FunctionPassManager</tt>, "<tt>OurFPM</tt>".  It
 199 requires a pointer to the <tt>Module</tt> to construct itself.  Once it is set
 200 up, we use a series of "add" calls to add a bunch of LLVM passes.  The first
 201 pass is basically boilerplate, it adds a pass so that later optimizations know
 202 how the data structures in the program are laid out.  The
 203 "<tt>TheExecutionEngine</tt>" variable is related to the JIT, which we will get
 204 to in the next section.</p>
 205
 206 <p>In this case, we choose to add 4 optimization passes.  The passes we chose
 207 here are a pretty standard set of "cleanup" optimizations that are useful for
 208 a wide variety of code.  I won't delve into what they do but, believe me,
 209 they are a good starting place :).</p>
 210
 211 <p>Once the PassManager is set up, we need to make use of it.  We do this by
 212 running it after our newly created function is constructed (in
 213 <tt>FunctionAST::Codegen</tt>), but before it is returned to the client:</p>
 214
 215 <div class="doc_code">
 216 <pre>
 217   if (Value *RetVal = Body->Codegen()) {
 218     // Finish off the function.
 219     Builder.CreateRet(RetVal);
 220
 221     // Validate the generated code, checking for consistency.
 222     verifyFunction(*TheFunction);
 223
 224     <b>// Optimize the function.
 225     TheFPM-&gt;run(*TheFunction);</b>
 226
 227     return TheFunction;
 228   }
 229 </pre>
 230 </div>
 231
 232 <p>As you can see, this is pretty straightforward.  The
 233 <tt>FunctionPassManager</tt> optimizes and updates the LLVM Function* in place,
 234 improving (hopefully) its body.  With this in place, we can try our test above
 235 again:</p>
 236
 237 <div class="doc_code">
 238 <pre>
 239 ready&gt; <b>def test(x) (1+2+x)*(x+(1+2));</b>
 240 ready> Read function definition:
 241 define double @test(double %x) {
 242 entry:
 243         %addtmp = fadd double %x, 3.000000e+00
 244         %multmp = fmul double %addtmp, %addtmp
 245         ret double %multmp
 246 }
 247 </pre>
 248 </div>
 249
 250 <p>As expected, we now get our nicely optimized code, saving a floating point
 251 add instruction from every execution of this function.</p>
 252
 253 <p>LLVM provides a wide variety of optimizations that can be used in certain
 254 circumstances.  Some <a href="../Passes.html">documentation about the various
 255 passes</a> is available, but it isn't very complete.  Another good source of
 256 ideas can come from looking at the passes that <tt>llvm-gcc</tt> or
 257 <tt>llvm-ld</tt> run to get started.  The "<tt>opt</tt>" tool allows you to
 258 experiment with passes from the command line, so you can see if they do
 259 anything.</p>
 260
 261 <p>Now that we have reasonable code coming out of our front-end, lets talk about
 262 executing it!</p>
 263
 264 </div>
 265
 266 <!-- *********************************************************************** -->
 267 <h2><a name="jit">Adding a JIT Compiler</a></h2>
 268 <!-- *********************************************************************** -->
 269
 270 <div>
 271
 272 <p>Code that is available in LLVM IR can have a wide variety of tools
 273 applied to it.  For example, you can run optimizations on it (as we did above),
 274 you can dump it out in textual or binary forms, you can compile the code to an
 275 assembly file (.s) for some target, or you can JIT compile it.  The nice thing
 276 about the LLVM IR representation is that it is the "common currency" between
 277 many different parts of the compiler.
 278 </p>
 279
 280 <p>In this section, we'll add JIT compiler support to our interpreter.  The
 281 basic idea that we want for Kaleidoscope is to have the user enter function
 282 bodies as they do now, but immediately evaluate the top-level expressions they
 283 type in.  For example, if they type in "1 + 2;", we should evaluate and print
 284 out 3.  If they define a function, they should be able to call it from the
 285 command line.</p>
 286
 287 <p>In order to do this, we first declare and initialize the JIT.  This is done
 288 by adding a global variable and a call in <tt>main</tt>:</p>
 289
 290 <div class="doc_code">
 291 <pre>
 292 <b>static ExecutionEngine *TheExecutionEngine;</b>
 293 ...
 294 int main() {
 295   ..
 296   <b>// Create the JIT.  This takes ownership of the module.
 297   TheExecutionEngine = EngineBuilder(TheModule).create();</b>
 298   ..
 299 }
 300 </pre>
 301 </div>
 302
 303 <p>This creates an abstract "Execution Engine" which can be either a JIT
 304 compiler or the LLVM interpreter.  LLVM will automatically pick a JIT compiler
 305 for you if one is available for your platform, otherwise it will fall back to
 306 the interpreter.</p>
 307
 308 <p>Once the <tt>ExecutionEngine</tt> is created, the JIT is ready to be used.
 309 There are a variety of APIs that are useful, but the simplest one is the
 310 "<tt>getPointerToFunction(F)</tt>" method.  This method JIT compiles the
 311 specified LLVM Function and returns a function pointer to the generated machine
 312 code.  In our case, this means that we can change the code that parses a
 313 top-level expression to look like this:</p>
 314
 315 <div class="doc_code">
 316 <pre>
 317 static void HandleTopLevelExpression() {
 318   // Evaluate a top-level expression into an anonymous function.
 319   if (FunctionAST *F = ParseTopLevelExpr()) {
 320     if (Function *LF = F-&gt;Codegen()) {
 321       LF->dump();  // Dump the function for exposition purposes.
 322
 323       <b>// JIT the function, returning a function pointer.
 324       void *FPtr = TheExecutionEngine-&gt;getPointerToFunction(LF);
 325
 326       // Cast it to the right type (takes no arguments, returns a double) so we
 327       // can call it as a native function.
 328       double (*FP)() = (double (*)())(intptr_t)FPtr;
 329       fprintf(stderr, "Evaluated to %f\n", FP());</b>
 330     }
 331 </pre>
 332 </div>
 333
 334 <p>Recall that we compile top-level expressions into a self-contained LLVM
 335 function that takes no arguments and returns the computed double.  Because the
 336 LLVM JIT compiler matches the native platform ABI, this means that you can just
 337 cast the result pointer to a function pointer of that type and call it directly.
 338 This means, there is no difference between JIT compiled code and native machine
 339 code that is statically linked into your application.</p>
 340
 341 <p>With just these two changes, lets see how Kaleidoscope works now!</p>
 342
 343 <div class="doc_code">
 344 <pre>
 345 ready&gt; <b>4+5;</b>
 346 define double @""() {
 347 entry:
 348         ret double 9.000000e+00
 349 }
 350
 351 <em>Evaluated to 9.000000</em>
 352 </pre>
 353 </div>
 354
 355 <p>Well this looks like it is basically working.  The dump of the function
 356 shows the "no argument function that always returns double" that we synthesize
 357 for each top-level expression that is typed in.  This demonstrates very basic
 358 functionality, but can we do more?</p>
 359
 360 <div class="doc_code">
 361 <pre>
 362 ready&gt; <b>def testfunc(x y) x + y*2; </b>
 363 Read function definition:
 364 define double @testfunc(double %x, double %y) {
 365 entry:
 366         %multmp = fmul double %y, 2.000000e+00
 367         %addtmp = fadd double %multmp, %x
 368         ret double %addtmp
 369 }
 370
 371 ready&gt; <b>testfunc(4, 10);</b>
 372 define double @""() {
 373 entry:
 374         %calltmp = call double @testfunc(double 4.000000e+00, double 1.000000e+01)
 375         ret double %calltmp
 376 }
 377
 378 <em>Evaluated to 24.000000</em>
 379 </pre>
 380 </div>
 381
 382 <p>This illustrates that we can now call user code, but there is something a bit
 383 subtle going on here.  Note that we only invoke the JIT on the anonymous
 384 functions that <em>call testfunc</em>, but we never invoked it
 385 on <em>testfunc</em> itself.  What actually happened here is that the JIT
 386 scanned for all non-JIT'd functions transitively called from the anonymous
 387 function and compiled all of them before returning
 388 from <tt>getPointerToFunction()</tt>.</p>
 389
 390 <p>The JIT provides a number of other more advanced interfaces for things like
 391 freeing allocated machine code, rejit'ing functions to update them, etc.
 392 However, even with this simple code, we get some surprisingly powerful
 393 capabilities - check this out (I removed the dump of the anonymous functions,
 394 you should get the idea by now :) :</p>
 395
 396 <div class="doc_code">
 397 <pre>
 398 ready&gt; <b>extern sin(x);</b>
 399 Read extern:
 400 declare double @sin(double)
 401
 402 ready&gt; <b>extern cos(x);</b>
 403 Read extern:
 404 declare double @cos(double)
 405
 406 ready&gt; <b>sin(1.0);</b>
 407 <em>Evaluated to 0.841471</em>
 408
 409 ready&gt; <b>def foo(x) sin(x)*sin(x) + cos(x)*cos(x);</b>
 410 Read function definition:
 411 define double @foo(double %x) {
 412 entry:
 413         %calltmp = call double @sin(double %x)
 414         %multmp = fmul double %calltmp, %calltmp
 415         %calltmp2 = call double @cos(double %x)
 416         %multmp4 = fmul double %calltmp2, %calltmp2
 417         %addtmp = fadd double %multmp, %multmp4
 418         ret double %addtmp
 419 }
 420
 421 ready&gt; <b>foo(4.0);</b>
 422 <em>Evaluated to 1.000000</em>
 423 </pre>
 424 </div>
 425
 426 <p>Whoa, how does the JIT know about sin and cos?  The answer is surprisingly
 427 simple: in this
 428 example, the JIT started execution of a function and got to a function call.  It
 429 realized that the function was not yet JIT compiled and invoked the standard set
 430 of routines to resolve the function.  In this case, there is no body defined
 431 for the function, so the JIT ended up calling "<tt>dlsym("sin")</tt>" on the
 432 Kaleidoscope process itself.
 433 Since "<tt>sin</tt>" is defined within the JIT's address space, it simply
 434 patches up calls in the module to call the libm version of <tt>sin</tt>
 435 directly.</p>
 436
 437 <p>The LLVM JIT provides a number of interfaces (look in the
 438 <tt>ExecutionEngine.h</tt> file) for controlling how unknown functions get
 439 resolved.  It allows you to establish explicit mappings between IR objects and
 440 addresses (useful for LLVM global variables that you want to map to static
 441 tables, for example), allows you to dynamically decide on the fly based on the
 442 function name, and even allows you to have the JIT compile functions lazily the
 443 first time they're called.</p>
 444
 445 <p>One interesting application of this is that we can now extend the language
 446 by writing arbitrary C++ code to implement operations.  For example, if we add:
 447 </p>
 448
 449 <div class="doc_code">
 450 <pre>
 451 /// putchard - putchar that takes a double and returns 0.
 452 extern "C"
 453 double putchard(double X) {
 454   putchar((char)X);
 455   return 0;
 456 }
 457 </pre>
 458 </div>
 459
 460 <p>Now we can produce simple output to the console by using things like:
 461 "<tt>extern putchard(x); putchard(120);</tt>", which prints a lowercase 'x' on
 462 the console (120 is the ASCII code for 'x').  Similar code could be used to
 463 implement file I/O, console input, and many other capabilities in
 464 Kaleidoscope.</p>
 465
 466 <p>This completes the JIT and optimizer chapter of the Kaleidoscope tutorial. At
 467 this point, we can compile a non-Turing-complete programming language, optimize
 468 and JIT compile it in a user-driven way.  Next up we'll look into <a
 469 href="LangImpl5.html">extending the language with control flow constructs</a>,
 470 tackling some interesting LLVM IR issues along the way.</p>
 471
 472 </div>
 473
 474 <!-- *********************************************************************** -->
 475 <h2><a name="code">Full Code Listing</a></h2>
 476 <!-- *********************************************************************** -->
 477
 478 <div>
 479
 480 <p>
 481 Here is the complete code listing for our running example, enhanced with the
 482 LLVM JIT and optimizer.  To build this example, use:
 483 </p>
 484
 485 <div class="doc_code">
 486 <pre>
 487    # Compile
 488    g++ -g toy.cpp `llvm-config --cppflags --ldflags --libs core jit native` -O3 -o toy
 489    # Run
 490    ./toy
 491 </pre>
 492 </div>
 493
 494 <p>
 495 If you are compiling this on Linux, make sure to add the "-rdynamic" option
 496 as well.  This makes sure that the external functions are resolved properly
 497 at runtime.</p>
 498
 499 <p>Here is the code:</p>
 500
 501 <div class="doc_code">
 502 <pre>
 503 #include "llvm/DerivedTypes.h"
 504 #include "llvm/ExecutionEngine/ExecutionEngine.h"
 505 #include "llvm/ExecutionEngine/JIT.h"
 506 #include "llvm/LLVMContext.h"
 507 #include "llvm/Module.h"
 508 #include "llvm/PassManager.h"
 509 #include "llvm/Analysis/Verifier.h"
 510 #include "llvm/Analysis/Passes.h"
 511 #include "llvm/Target/TargetData.h"
 512 #include "llvm/Target/TargetSelect.h"
 513 #include "llvm/Transforms/Scalar.h"
 514 #include "llvm/Support/IRBuilder.h"
 515 #include &lt;cstdio&gt;
 516 #include &lt;string&gt;
 517 #include &lt;map&gt;
 518 #include &lt;vector&gt;
 519 using namespace llvm;
 520
 521 //===----------------------------------------------------------------------===//
 522 // Lexer
 523 //===----------------------------------------------------------------------===//
 524
 525 // The lexer returns tokens [0-255] if it is an unknown character, otherwise one
 526 // of these for known things.
 527 enum Token {
 528   tok_eof = -1,
 529
 530   // commands
 531   tok_def = -2, tok_extern = -3,
 532
 533   // primary
 534   tok_identifier = -4, tok_number = -5
 535 };
 536
 537 static std::string IdentifierStr;  // Filled in if tok_identifier
 538 static double NumVal;              // Filled in if tok_number
 539
 540 /// gettok - Return the next token from standard input.
 541 static int gettok() {
 542   static int LastChar = ' ';
 543
 544   // Skip any whitespace.
 545   while (isspace(LastChar))
 546     LastChar = getchar();
 547
 548   if (isalpha(LastChar)) { // identifier: [a-zA-Z][a-zA-Z0-9]*
 549     IdentifierStr = LastChar;
 550     while (isalnum((LastChar = getchar())))
 551       IdentifierStr += LastChar;
 552
 553     if (IdentifierStr == "def") return tok_def;
 554     if (IdentifierStr == "extern") return tok_extern;
 555     return tok_identifier;
 556   }
 557
 558   if (isdigit(LastChar) || LastChar == '.') {   // Number: [0-9.]+
 559     std::string NumStr;
 560     do {
 561       NumStr += LastChar;
 562       LastChar = getchar();
 563     } while (isdigit(LastChar) || LastChar == '.');
 564
 565     NumVal = strtod(NumStr.c_str(), 0);
 566     return tok_number;
 567   }
 568
 569   if (LastChar == '#') {
 570     // Comment until end of line.
 571     do LastChar = getchar();
 572     while (LastChar != EOF &amp;&amp; LastChar != '\n' &amp;&amp; LastChar != '\r');
 573
 574     if (LastChar != EOF)
 575       return gettok();
 576   }
 577
 578   // Check for end of file.  Don't eat the EOF.
 579   if (LastChar == EOF)
 580     return tok_eof;
 581
 582   // Otherwise, just return the character as its ascii value.
 583   int ThisChar = LastChar;
 584   LastChar = getchar();
 585   return ThisChar;
 586 }
 587
 588 //===----------------------------------------------------------------------===//
 589 // Abstract Syntax Tree (aka Parse Tree)
 590 //===----------------------------------------------------------------------===//
 591
 592 /// ExprAST - Base class for all expression nodes.
 593 class ExprAST {
 594 public:
 595   virtual ~ExprAST() {}
 596   virtual Value *Codegen() = 0;
 597 };
 598
 599 /// NumberExprAST - Expression class for numeric literals like "1.0".
 600 class NumberExprAST : public ExprAST {
 601   double Val;
 602 public:
 603   NumberExprAST(double val) : Val(val) {}
 604   virtual Value *Codegen();
 605 };
 606
 607 /// VariableExprAST - Expression class for referencing a variable, like "a".
 608 class VariableExprAST : public ExprAST {
 609   std::string Name;
 610 public:
 611   VariableExprAST(const std::string &amp;name) : Name(name) {}
 612   virtual Value *Codegen();
 613 };
 614
 615 /// BinaryExprAST - Expression class for a binary operator.
 616 class BinaryExprAST : public ExprAST {
 617   char Op;
 618   ExprAST *LHS, *RHS;
 619 public:
 620   BinaryExprAST(char op, ExprAST *lhs, ExprAST *rhs)
 621     : Op(op), LHS(lhs), RHS(rhs) {}
 622   virtual Value *Codegen();
 623 };
 624
 625 /// CallExprAST - Expression class for function calls.
 626 class CallExprAST : public ExprAST {
 627   std::string Callee;
 628   std::vector&lt;ExprAST*&gt; Args;
 629 public:
 630   CallExprAST(const std::string &amp;callee, std::vector&lt;ExprAST*&gt; &amp;args)
 631     : Callee(callee), Args(args) {}
 632   virtual Value *Codegen();
 633 };
 634
 635 /// PrototypeAST - This class represents the "prototype" for a function,
 636 /// which captures its name, and its argument names (thus implicitly the number
 637 /// of arguments the function takes).
 638 class PrototypeAST {
 639   std::string Name;
 640   std::vector&lt;std::string&gt; Args;
 641 public:
 642   PrototypeAST(const std::string &amp;name, const std::vector&lt;std::string&gt; &amp;args)
 643     : Name(name), Args(args) {}
 644
 645   Function *Codegen();
 646 };
 647
 648 /// FunctionAST - This class represents a function definition itself.
 649 class FunctionAST {
 650   PrototypeAST *Proto;
 651   ExprAST *Body;
 652 public:
 653   FunctionAST(PrototypeAST *proto, ExprAST *body)
 654     : Proto(proto), Body(body) {}
 655
 656   Function *Codegen();
 657 };
 658
 659 //===----------------------------------------------------------------------===//
 660 // Parser
 661 //===----------------------------------------------------------------------===//
 662
 663 /// CurTok/getNextToken - Provide a simple token buffer.  CurTok is the current
 664 /// token the parser is looking at.  getNextToken reads another token from the
 665 /// lexer and updates CurTok with its results.
 666 static int CurTok;
 667 static int getNextToken() {
 668   return CurTok = gettok();
 669 }
 670
 671 /// BinopPrecedence - This holds the precedence for each binary operator that is
 672 /// defined.
 673 static std::map&lt;char, int&gt; BinopPrecedence;
 674
 675 /// GetTokPrecedence - Get the precedence of the pending binary operator token.
 676 static int GetTokPrecedence() {
 677   if (!isascii(CurTok))
 678     return -1;
 679
 680   // Make sure it's a declared binop.
 681   int TokPrec = BinopPrecedence[CurTok];
 682   if (TokPrec &lt;= 0) return -1;
 683   return TokPrec;
 684 }
 685
 686 /// Error* - These are little helper functions for error handling.
 687 ExprAST *Error(const char *Str) { fprintf(stderr, "Error: %s\n", Str);return 0;}
 688 PrototypeAST *ErrorP(const char *Str) { Error(Str); return 0; }
 689 FunctionAST *ErrorF(const char *Str) { Error(Str); return 0; }
 690
 691 static ExprAST *ParseExpression();
 692
 693 /// identifierexpr
 694 ///   ::= identifier
 695 ///   ::= identifier '(' expression* ')'
 696 static ExprAST *ParseIdentifierExpr() {
 697   std::string IdName = IdentifierStr;
 698
 699   getNextToken();  // eat identifier.
 700
 701   if (CurTok != '(') // Simple variable ref.
 702     return new VariableExprAST(IdName);
 703
 704   // Call.
 705   getNextToken();  // eat (
 706   std::vector&lt;ExprAST*&gt; Args;
 707   if (CurTok != ')') {
 708     while (1) {
 709       ExprAST *Arg = ParseExpression();
 710       if (!Arg) return 0;
 711       Args.push_back(Arg);
 712
 713       if (CurTok == ')') break;
 714
 715       if (CurTok != ',')
 716         return Error("Expected ')' or ',' in argument list");
 717       getNextToken();
 718     }
 719   }
 720
 721   // Eat the ')'.
 722   getNextToken();
 723
 724   return new CallExprAST(IdName, Args);
 725 }
 726
 727 /// numberexpr ::= number
 728 static ExprAST *ParseNumberExpr() {
 729   ExprAST *Result = new NumberExprAST(NumVal);
 730   getNextToken(); // consume the number
 731   return Result;
 732 }
 733
 734 /// parenexpr ::= '(' expression ')'
 735 static ExprAST *ParseParenExpr() {
 736   getNextToken();  // eat (.
 737   ExprAST *V = ParseExpression();
 738   if (!V) return 0;
 739
 740   if (CurTok != ')')
 741     return Error("expected ')'");
 742   getNextToken();  // eat ).
 743   return V;
 744 }
 745
 746 /// primary
 747 ///   ::= identifierexpr
 748 ///   ::= numberexpr
 749 ///   ::= parenexpr
 750 static ExprAST *ParsePrimary() {
 751   switch (CurTok) {
 752   default: return Error("unknown token when expecting an expression");
 753   case tok_identifier: return ParseIdentifierExpr();
 754   case tok_number:     return ParseNumberExpr();
 755   case '(':            return ParseParenExpr();
 756   }
 757 }
 758
 759 /// binoprhs
 760 ///   ::= ('+' primary)*
 761 static ExprAST *ParseBinOpRHS(int ExprPrec, ExprAST *LHS) {
 762   // If this is a binop, find its precedence.
 763   while (1) {
 764     int TokPrec = GetTokPrecedence();
 765
 766     // If this is a binop that binds at least as tightly as the current binop,
 767     // consume it, otherwise we are done.
 768     if (TokPrec &lt; ExprPrec)
 769       return LHS;
 770
 771     // Okay, we know this is a binop.
 772     int BinOp = CurTok;
 773     getNextToken();  // eat binop
 774
 775     // Parse the primary expression after the binary operator.
 776     ExprAST *RHS = ParsePrimary();
 777     if (!RHS) return 0;
 778
 779     // If BinOp binds less tightly with RHS than the operator after RHS, let
 780     // the pending operator take RHS as its LHS.
 781     int NextPrec = GetTokPrecedence();
 782     if (TokPrec &lt; NextPrec) {
 783       RHS = ParseBinOpRHS(TokPrec+1, RHS);
 784       if (RHS == 0) return 0;
 785     }
 786
 787     // Merge LHS/RHS.
 788     LHS = new BinaryExprAST(BinOp, LHS, RHS);
 789   }
 790 }
 791
 792 /// expression
 793 ///   ::= primary binoprhs
 794 ///
 795 static ExprAST *ParseExpression() {
 796   ExprAST *LHS = ParsePrimary();
 797   if (!LHS) return 0;
 798
 799   return ParseBinOpRHS(0, LHS);
 800 }
 801
 802 /// prototype
 803 ///   ::= id '(' id* ')'
 804 static PrototypeAST *ParsePrototype() {
 805   if (CurTok != tok_identifier)
 806     return ErrorP("Expected function name in prototype");
 807
 808   std::string FnName = IdentifierStr;
 809   getNextToken();
 810
 811   if (CurTok != '(')
 812     return ErrorP("Expected '(' in prototype");
 813
 814   std::vector&lt;std::string&gt; ArgNames;
 815   while (getNextToken() == tok_identifier)
 816     ArgNames.push_back(IdentifierStr);
 817   if (CurTok != ')')
 818     return ErrorP("Expected ')' in prototype");
 819
 820   // success.
 821   getNextToken();  // eat ')'.
 822
 823   return new PrototypeAST(FnName, ArgNames);
 824 }
 825
 826 /// definition ::= 'def' prototype expression
 827 static FunctionAST *ParseDefinition() {
 828   getNextToken();  // eat def.
 829   PrototypeAST *Proto = ParsePrototype();
 830   if (Proto == 0) return 0;
 831
 832   if (ExprAST *E = ParseExpression())
 833     return new FunctionAST(Proto, E);
 834   return 0;
 835 }
 836
 837 /// toplevelexpr ::= expression
 838 static FunctionAST *ParseTopLevelExpr() {
 839   if (ExprAST *E = ParseExpression()) {
 840     // Make an anonymous proto.
 841     PrototypeAST *Proto = new PrototypeAST("", std::vector&lt;std::string&gt;());
 842     return new FunctionAST(Proto, E);
 843   }
 844   return 0;
 845 }
 846
 847 /// external ::= 'extern' prototype
 848 static PrototypeAST *ParseExtern() {
 849   getNextToken();  // eat extern.
 850   return ParsePrototype();
 851 }
 852
 853 //===----------------------------------------------------------------------===//
 854 // Code Generation
 855 //===----------------------------------------------------------------------===//
 856
 857 static Module *TheModule;
 858 static IRBuilder&lt;&gt; Builder(getGlobalContext());
 859 static std::map&lt;std::string, Value*&gt; NamedValues;
 860 static FunctionPassManager *TheFPM;
 861
 862 Value *ErrorV(const char *Str) { Error(Str); return 0; }
 863
 864 Value *NumberExprAST::Codegen() {
 865   return ConstantFP::get(getGlobalContext(), APFloat(Val));
 866 }
 867
 868 Value *VariableExprAST::Codegen() {
 869   // Look this variable up in the function.
 870   Value *V = NamedValues[Name];
 871   return V ? V : ErrorV("Unknown variable name");
 872 }
 873
 874 Value *BinaryExprAST::Codegen() {
 875   Value *L = LHS-&gt;Codegen();
 876   Value *R = RHS-&gt;Codegen();
 877   if (L == 0 || R == 0) return 0;
 878
 879   switch (Op) {
 880   case '+': return Builder.CreateFAdd(L, R, "addtmp");
 881   case '-': return Builder.CreateFSub(L, R, "subtmp");
 882   case '*': return Builder.CreateFMul(L, R, "multmp");
 883   case '&lt;':
 884     L = Builder.CreateFCmpULT(L, R, "cmptmp");
 885     // Convert bool 0/1 to double 0.0 or 1.0
 886     return Builder.CreateUIToFP(L, Type::getDoubleTy(getGlobalContext()),
 887                                 "booltmp");
 888   default: return ErrorV("invalid binary operator");
 889   }
 890 }
 891
 892 Value *CallExprAST::Codegen() {
 893   // Look up the name in the global module table.
 894   Function *CalleeF = TheModule-&gt;getFunction(Callee);
 895   if (CalleeF == 0)
 896     return ErrorV("Unknown function referenced");
 897
 898   // If argument mismatch error.
 899   if (CalleeF-&gt;arg_size() != Args.size())
 900     return ErrorV("Incorrect # arguments passed");
 901
 902   std::vector&lt;Value*&gt; ArgsV;
 903   for (unsigned i = 0, e = Args.size(); i != e; ++i) {
 904     ArgsV.push_back(Args[i]-&gt;Codegen());
 905     if (ArgsV.back() == 0) return 0;
 906   }
 907
 908   return Builder.CreateCall(CalleeF, ArgsV.begin(), ArgsV.end(), "calltmp");
 909 }
 910
 911 Function *PrototypeAST::Codegen() {
 912   // Make the function type:  double(double,double) etc.
 913   std::vector&lt;const Type*&gt; Doubles(Args.size(),
 914                                    Type::getDoubleTy(getGlobalContext()));
 915   FunctionType *FT = FunctionType::get(Type::getDoubleTy(getGlobalContext()),
 916                                        Doubles, false);
 917
 918   Function *F = Function::Create(FT, Function::ExternalLinkage, Name, TheModule);
 919
 920   // If F conflicted, there was already something named 'Name'.  If it has a
 921   // body, don't allow redefinition or reextern.
 922   if (F-&gt;getName() != Name) {
 923     // Delete the one we just made and get the existing one.
 924     F-&gt;eraseFromParent();
 925     F = TheModule-&gt;getFunction(Name);
 926
 927     // If F already has a body, reject this.
 928     if (!F-&gt;empty()) {
 929       ErrorF("redefinition of function");
 930       return 0;
 931     }
 932
 933     // If F took a different number of args, reject.
 934     if (F-&gt;arg_size() != Args.size()) {
 935       ErrorF("redefinition of function with different # args");
 936       return 0;
 937     }
 938   }
 939
 940   // Set names for all arguments.
 941   unsigned Idx = 0;
 942   for (Function::arg_iterator AI = F-&gt;arg_begin(); Idx != Args.size();
 943        ++AI, ++Idx) {
 944     AI-&gt;setName(Args[Idx]);
 945
 946     // Add arguments to variable symbol table.
 947     NamedValues[Args[Idx]] = AI;
 948   }
 949
 950   return F;
 951 }
 952
 953 Function *FunctionAST::Codegen() {
 954   NamedValues.clear();
 955
 956   Function *TheFunction = Proto-&gt;Codegen();
 957   if (TheFunction == 0)
 958     return 0;
 959
 960   // Create a new basic block to start insertion into.
 961   BasicBlock *BB = BasicBlock::Create(getGlobalContext(), "entry", TheFunction);
 962   Builder.SetInsertPoint(BB);
 963
 964   if (Value *RetVal = Body-&gt;Codegen()) {
 965     // Finish off the function.
 966     Builder.CreateRet(RetVal);
 967
 968     // Validate the generated code, checking for consistency.
 969     verifyFunction(*TheFunction);
 970
 971     // Optimize the function.
 972     TheFPM-&gt;run(*TheFunction);
 973
 974     return TheFunction;
 975   }
 976
 977   // Error reading body, remove function.
 978   TheFunction-&gt;eraseFromParent();
 979   return 0;
 980 }
 981
 982 //===----------------------------------------------------------------------===//
 983 // Top-Level parsing and JIT Driver
 984 //===----------------------------------------------------------------------===//
 985
 986 static ExecutionEngine *TheExecutionEngine;
 987
 988 static void HandleDefinition() {
 989   if (FunctionAST *F = ParseDefinition()) {
 990     if (Function *LF = F-&gt;Codegen()) {
 991       fprintf(stderr, "Read function definition:");
 992       LF-&gt;dump();
 993     }
 994   } else {
 995     // Skip token for error recovery.
 996     getNextToken();
 997   }
 998 }
 999
1000 static void HandleExtern() {
1001   if (PrototypeAST *P = ParseExtern()) {
1002     if (Function *F = P-&gt;Codegen()) {
1003       fprintf(stderr, "Read extern: ");
1004       F-&gt;dump();
1005     }
1006   } else {
1007     // Skip token for error recovery.
1008     getNextToken();
1009   }
1010 }
1011
1012 static void HandleTopLevelExpression() {
1013   // Evaluate a top-level expression into an anonymous function.
1014   if (FunctionAST *F = ParseTopLevelExpr()) {
1015     if (Function *LF = F-&gt;Codegen()) {
1016       // JIT the function, returning a function pointer.
1017       void *FPtr = TheExecutionEngine-&gt;getPointerToFunction(LF);
1018
1019       // Cast it to the right type (takes no arguments, returns a double) so we
1020       // can call it as a native function.
1021       double (*FP)() = (double (*)())(intptr_t)FPtr;
1022       fprintf(stderr, "Evaluated to %f\n", FP());
1023     }
1024   } else {
1025     // Skip token for error recovery.
1026     getNextToken();
1027   }
1028 }
1029
1030 /// top ::= definition | external | expression | ';'
1031 static void MainLoop() {
1032   while (1) {
1033     fprintf(stderr, "ready&gt; ");
1034     switch (CurTok) {
1035     case tok_eof:    return;
1036     case ';':        getNextToken(); break;  // ignore top-level semicolons.
1037     case tok_def:    HandleDefinition(); break;
1038     case tok_extern: HandleExtern(); break;
1039     default:         HandleTopLevelExpression(); break;
1040     }
1041   }
1042 }
1043
1044 //===----------------------------------------------------------------------===//
1045 // "Library" functions that can be "extern'd" from user code.
1046 //===----------------------------------------------------------------------===//
1047
1048 /// putchard - putchar that takes a double and returns 0.
1049 extern "C"
1050 double putchard(double X) {
1051   putchar((char)X);
1052   return 0;
1053 }
1054
1055 //===----------------------------------------------------------------------===//
1056 // Main driver code.
1057 //===----------------------------------------------------------------------===//
1058
1059 int main() {
1060   InitializeNativeTarget();
1061   LLVMContext &amp;Context = getGlobalContext();
1062
1063   // Install standard binary operators.
1064   // 1 is lowest precedence.
1065   BinopPrecedence['&lt;'] = 10;
1066   BinopPrecedence['+'] = 20;
1067   BinopPrecedence['-'] = 20;
1068   BinopPrecedence['*'] = 40;  // highest.
1069
1070   // Prime the first token.
1071   fprintf(stderr, "ready&gt; ");
1072   getNextToken();
1073
1074   // Make the module, which holds all the code.
1075   TheModule = new Module("my cool jit", Context);
1076
1077   // Create the JIT.  This takes ownership of the module.
1078   std::string ErrStr;
1079 TheExecutionEngine = EngineBuilder(TheModule).setErrorStr(&amp;ErrStr).create();
1080   if (!TheExecutionEngine) {
1081     fprintf(stderr, "Could not create ExecutionEngine: %s\n", ErrStr.c_str());
1082     exit(1);
1083   }
1084
1085   FunctionPassManager OurFPM(TheModule);
1086
1087   // Set up the optimizer pipeline.  Start with registering info about how the
1088   // target lays out data structures.
1089   OurFPM.add(new TargetData(*TheExecutionEngine-&gt;getTargetData()));
1090   // Provide basic AliasAnalysis support for GVN.
1091   OurFPM.add(createBasicAliasAnalysisPass());
1092   // Do simple "peephole" optimizations and bit-twiddling optzns.
1093   OurFPM.add(createInstructionCombiningPass());
1094   // Reassociate expressions.
1095   OurFPM.add(createReassociatePass());
1096   // Eliminate Common SubExpressions.
1097   OurFPM.add(createGVNPass());
1098   // Simplify the control flow graph (deleting unreachable blocks, etc).
1099   OurFPM.add(createCFGSimplificationPass());
1100
1101   OurFPM.doInitialization();
1102
1103   // Set the global so the code gen can use this.
1104   TheFPM = &amp;OurFPM;
1105
1106   // Run the main "interpreter loop" now.
1107   MainLoop();
1108
1109   TheFPM = 0;
1110
1111   // Print out all of the generated code.
1112   TheModule-&gt;dump();
1113
1114   return 0;
1115 }
1116 </pre>
1117 </div>
1118
1119 <a href="LangImpl5.html">Next: Extending the language: control flow</a>
1120 </div>
1121
1122 <!-- *********************************************************************** -->
1123 <hr>
1124 <address>
1125   <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1126   src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
1127   <a href="http://validator.w3.org/check/referer"><img
1128   src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!"></a>
1129
1130   <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1131   <a href="http://llvm.org/">The LLVM Compiler Infrastructure</a><br>
1132   Last modified: $Date$
1133 </address>
1134 </body>
1135 </html>