1 ============================================== 2 Kaleidoscope: Adding JIT and Optimizer Support 3 ============================================== 4 5 .. contents:: 6 :local: 7 8 Chapter 4 Introduction 9 ====================== 10 11 Welcome to Chapter 4 of the "`Implementing a language with 12 LLVM <index.html>`_" tutorial. Chapters 1-3 described the implementation 13 of a simple language and added support for generating LLVM IR. This 14 chapter describes two new techniques: adding optimizer support to your 15 language, and adding JIT compiler support. These additions will 16 demonstrate how to get nice, efficient code for the Kaleidoscope 17 language. 18 19 Trivial Constant Folding 20 ======================== 21 22 Our demonstration for Chapter 3 is elegant and easy to extend. 23 Unfortunately, it does not produce wonderful code. The IRBuilder, 24 however, does give us obvious optimizations when compiling simple code: 25 26 :: 27 28 ready> def test(x) 1+2+x; 29 Read function definition: 30 define double @test(double %x) { 31 entry: 32 %addtmp = fadd double 3.000000e+00, %x 33 ret double %addtmp 34 } 35 36 This code is not a literal transcription of the AST built by parsing the 37 input. That would be: 38 39 :: 40 41 ready> def test(x) 1+2+x; 42 Read function definition: 43 define double @test(double %x) { 44 entry: 45 %addtmp = fadd double 2.000000e+00, 1.000000e+00 46 %addtmp1 = fadd double %addtmp, %x 47 ret double %addtmp1 48 } 49 50 Constant folding, as seen above, in particular, is a very common and 51 very important optimization: so much so that many language implementors 52 implement constant folding support in their AST representation. 53 54 With LLVM, you don't need this support in the AST. Since all calls to 55 build LLVM IR go through the LLVM IR builder, the builder itself checked 56 to see if there was a constant folding opportunity when you call it. If 57 so, it just does the constant fold and return the constant instead of 58 creating an instruction. 59 60 Well, that was easy :). In practice, we recommend always using 61 ``IRBuilder`` when generating code like this. It has no "syntactic 62 overhead" for its use (you don't have to uglify your compiler with 63 constant checks everywhere) and it can dramatically reduce the amount of 64 LLVM IR that is generated in some cases (particular for languages with a 65 macro preprocessor or that use a lot of constants). 66 67 On the other hand, the ``IRBuilder`` is limited by the fact that it does 68 all of its analysis inline with the code as it is built. If you take a 69 slightly more complex example: 70 71 :: 72 73 ready> def test(x) (1+2+x)*(x+(1+2)); 74 ready> Read function definition: 75 define double @test(double %x) { 76 entry: 77 %addtmp = fadd double 3.000000e+00, %x 78 %addtmp1 = fadd double %x, 3.000000e+00 79 %multmp = fmul double %addtmp, %addtmp1 80 ret double %multmp 81 } 82 83 In this case, the LHS and RHS of the multiplication are the same value. 84 We'd really like to see this generate "``tmp = x+3; result = tmp*tmp;``" 85 instead of computing "``x+3``" twice. 86 87 Unfortunately, no amount of local analysis will be able to detect and 88 correct this. This requires two transformations: reassociation of 89 expressions (to make the add's lexically identical) and Common 90 Subexpression Elimination (CSE) to delete the redundant add instruction. 91 Fortunately, LLVM provides a broad range of optimizations that you can 92 use, in the form of "passes". 93 94 LLVM Optimization Passes 95 ======================== 96 97 LLVM provides many optimization passes, which do many different sorts of 98 things and have different tradeoffs. Unlike other systems, LLVM doesn't 99 hold to the mistaken notion that one set of optimizations is right for 100 all languages and for all situations. LLVM allows a compiler implementor 101 to make complete decisions about what optimizations to use, in which 102 order, and in what situation. 103 104 As a concrete example, LLVM supports both "whole module" passes, which 105 look across as large of body of code as they can (often a whole file, 106 but if run at link time, this can be a substantial portion of the whole 107 program). It also supports and includes "per-function" passes which just 108 operate on a single function at a time, without looking at other 109 functions. For more information on passes and how they are run, see the 110 `How to Write a Pass <../WritingAnLLVMPass.html>`_ document and the 111 `List of LLVM Passes <../Passes.html>`_. 112 113 For Kaleidoscope, we are currently generating functions on the fly, one 114 at a time, as the user types them in. We aren't shooting for the 115 ultimate optimization experience in this setting, but we also want to 116 catch the easy and quick stuff where possible. As such, we will choose 117 to run a few per-function optimizations as the user types the function 118 in. If we wanted to make a "static Kaleidoscope compiler", we would use 119 exactly the code we have now, except that we would defer running the 120 optimizer until the entire file has been parsed. 121 122 In order to get per-function optimizations going, we need to set up a 123 `FunctionPassManager <../WritingAnLLVMPass.html#what-passmanager-doesr>`_ to hold 124 and organize the LLVM optimizations that we want to run. Once we have 125 that, we can add a set of optimizations to run. We'll need a new 126 FunctionPassManager for each module that we want to optimize, so we'll 127 write a function to create and initialize both the module and pass manager 128 for us: 129 130 .. code-block:: c++ 131 132 void InitializeModuleAndPassManager(void) { 133 // Open a new module. 134 TheModule = llvm::make_unique<Module>("my cool jit", getGlobalContext()); 135 TheModule->setDataLayout(TheJIT->getTargetMachine().createDataLayout()); 136 137 // Create a new pass manager attached to it. 138 TheFPM = llvm::make_unique<FunctionPassManager>(TheModule.get()); 139 140 // Provide basic AliasAnalysis support for GVN. 141 TheFPM.add(createBasicAliasAnalysisPass()); 142 // Do simple "peephole" optimizations and bit-twiddling optzns. 143 TheFPM.add(createInstructionCombiningPass()); 144 // Reassociate expressions. 145 TheFPM.add(createReassociatePass()); 146 // Eliminate Common SubExpressions. 147 TheFPM.add(createGVNPass()); 148 // Simplify the control flow graph (deleting unreachable blocks, etc). 149 TheFPM.add(createCFGSimplificationPass()); 150 151 TheFPM.doInitialization(); 152 } 153 154 This code initializes the global module ``TheModule``, and the function pass 155 manager ``TheFPM``, which is attached to ``TheModule``. Once the pass manager is 156 set up, we use a series of "add" calls to add a bunch of LLVM passes. 157 158 In this case, we choose to add five passes: one analysis pass (alias analysis), 159 and four optimization passes. The passes we choose here are a pretty standard set 160 of "cleanup" optimizations that are useful for a wide variety of code. I won't 161 delve into what they do but, believe me, they are a good starting place :). 162 163 Once the PassManager is set up, we need to make use of it. We do this by 164 running it after our newly created function is constructed (in 165 ``FunctionAST::codegen()``), but before it is returned to the client: 166 167 .. code-block:: c++ 168 169 if (Value *RetVal = Body->codegen()) { 170 // Finish off the function. 171 Builder.CreateRet(RetVal); 172 173 // Validate the generated code, checking for consistency. 174 verifyFunction(*TheFunction); 175 176 // Optimize the function. 177 TheFPM->run(*TheFunction); 178 179 return TheFunction; 180 } 181 182 As you can see, this is pretty straightforward. The 183 ``FunctionPassManager`` optimizes and updates the LLVM Function\* in 184 place, improving (hopefully) its body. With this in place, we can try 185 our test above again: 186 187 :: 188 189 ready> def test(x) (1+2+x)*(x+(1+2)); 190 ready> Read function definition: 191 define double @test(double %x) { 192 entry: 193 %addtmp = fadd double %x, 3.000000e+00 194 %multmp = fmul double %addtmp, %addtmp 195 ret double %multmp 196 } 197 198 As expected, we now get our nicely optimized code, saving a floating 199 point add instruction from every execution of this function. 200 201 LLVM provides a wide variety of optimizations that can be used in 202 certain circumstances. Some `documentation about the various 203 passes <../Passes.html>`_ is available, but it isn't very complete. 204 Another good source of ideas can come from looking at the passes that 205 ``Clang`` runs to get started. The "``opt``" tool allows you to 206 experiment with passes from the command line, so you can see if they do 207 anything. 208 209 Now that we have reasonable code coming out of our front-end, lets talk 210 about executing it! 211 212 Adding a JIT Compiler 213 ===================== 214 215 Code that is available in LLVM IR can have a wide variety of tools 216 applied to it. For example, you can run optimizations on it (as we did 217 above), you can dump it out in textual or binary forms, you can compile 218 the code to an assembly file (.s) for some target, or you can JIT 219 compile it. The nice thing about the LLVM IR representation is that it 220 is the "common currency" between many different parts of the compiler. 221 222 In this section, we'll add JIT compiler support to our interpreter. The 223 basic idea that we want for Kaleidoscope is to have the user enter 224 function bodies as they do now, but immediately evaluate the top-level 225 expressions they type in. For example, if they type in "1 + 2;", we 226 should evaluate and print out 3. If they define a function, they should 227 be able to call it from the command line. 228 229 In order to do this, we first declare and initialize the JIT. This is 230 done by adding a global variable ``TheJIT``, and initializing it in 231 ``main``: 232 233 .. code-block:: c++ 234 235 static std::unique_ptr<KaleidoscopeJIT> TheJIT; 236 ... 237 int main() { 238 .. 239 TheJIT = llvm::make_unique<KaleidoscopeJIT>(); 240 241 // Run the main "interpreter loop" now. 242 MainLoop(); 243 244 return 0; 245 } 246 247 The KaleidoscopeJIT class is a simple JIT built specifically for these 248 tutorials. In later chapters we will look at how it works and extend it with 249 new features, but for now we will take it as given. Its API is very simple:: 250 ``addModule`` adds an LLVM IR module to the JIT, making its functions 251 available for execution; ``removeModule`` removes a module, freeing any 252 memory associated with the code in that module; and ``findSymbol`` allows us 253 to look up pointers to the compiled code. 254 255 We can take this simple API and change our code that parses top-level expressions to 256 look like this: 257 258 .. code-block:: c++ 259 260 static void HandleTopLevelExpression() { 261 // Evaluate a top-level expression into an anonymous function. 262 if (auto FnAST = ParseTopLevelExpr()) { 263 if (FnAST->codegen()) { 264 265 // JIT the module containing the anonymous expression, keeping a handle so 266 // we can free it later. 267 auto H = TheJIT->addModule(std::move(TheModule)); 268 InitializeModuleAndPassManager(); 269 270 // Search the JIT for the __anon_expr symbol. 271 auto ExprSymbol = TheJIT->findSymbol("__anon_expr"); 272 assert(ExprSymbol && "Function not found"); 273 274 // Get the symbol's address and cast it to the right type (takes no 275 // arguments, returns a double) so we can call it as a native function. 276 double (*FP)() = (double (*)())(intptr_t)ExprSymbol.getAddress(); 277 fprintf(stderr, "Evaluated to %f\n", FP()); 278 279 // Delete the anonymous expression module from the JIT. 280 TheJIT->removeModule(H); 281 } 282 283 If parsing and codegen succeeed, the next step is to add the module containing 284 the top-level expression to the JIT. We do this by calling addModule, which 285 triggers code generation for all the functions in the module, and returns a 286 handle that can be used to remove the module from the JIT later. Once the module 287 has been added to the JIT it can no longer be modified, so we also open a new 288 module to hold subsequent code by calling ``InitializeModuleAndPassManager()``. 289 290 Once we've added the module to the JIT we need to get a pointer to the final 291 generated code. We do this by calling the JIT's findSymbol method, and passing 292 the name of the top-level expression function: ``__anon_expr``. Since we just 293 added this function, we assert that findSymbol returned a result. 294 295 Next, we get the in-memory address of the ``__anon_expr`` function by calling 296 ``getAddress()`` on the symbol. Recall that we compile top-level expressions 297 into a self-contained LLVM function that takes no arguments and returns the 298 computed double. Because the LLVM JIT compiler matches the native platform ABI, 299 this means that you can just cast the result pointer to a function pointer of 300 that type and call it directly. This means, there is no difference between JIT 301 compiled code and native machine code that is statically linked into your 302 application. 303 304 Finally, since we don't support re-evaluation of top-level expressions, we 305 remove the module from the JIT when we're done to free the associated memory. 306 Recall, however, that the module we created a few lines earlier (via 307 ``InitializeModuleAndPassManager``) is still open and waiting for new code to be 308 added. 309 310 With just these two changes, lets see how Kaleidoscope works now! 311 312 :: 313 314 ready> 4+5; 315 Read top-level expression: 316 define double @0() { 317 entry: 318 ret double 9.000000e+00 319 } 320 321 Evaluated to 9.000000 322 323 Well this looks like it is basically working. The dump of the function 324 shows the "no argument function that always returns double" that we 325 synthesize for each top-level expression that is typed in. This 326 demonstrates very basic functionality, but can we do more? 327 328 :: 329 330 ready> def testfunc(x y) x + y*2; 331 Read function definition: 332 define double @testfunc(double %x, double %y) { 333 entry: 334 %multmp = fmul double %y, 2.000000e+00 335 %addtmp = fadd double %multmp, %x 336 ret double %addtmp 337 } 338 339 ready> testfunc(4, 10); 340 Read top-level expression: 341 define double @1() { 342 entry: 343 %calltmp = call double @testfunc(double 4.000000e+00, double 1.000000e+01) 344 ret double %calltmp 345 } 346 347 Evaluated to 24.000000 348 349 ready> testfunc(5, 10); 350 ready> LLVM ERROR: Program used external function 'testfunc' which could not be resolved! 351 352 353 Function definitions and calls also work, but something went very wrong on that 354 last line. The call looks valid, so what happened? As you may have guessed from 355 the the API a Module is a unit of allocation for the JIT, and testfunc was part 356 of the same module that contained anonymous expression. When we removed that 357 module from the JIT to free the memory for the anonymous expression, we deleted 358 the definition of ``testfunc`` along with it. Then, when we tried to call 359 testfunc a second time, the JIT could no longer find it. 360 361 The easiest way to fix this is to put the anonymous expression in a separate 362 module from the rest of the function definitions. The JIT will happily resolve 363 function calls across module boundaries, as long as each of the functions called 364 has a prototype, and is added to the JIT before it is called. By putting the 365 anonymous expression in a different module we can delete it without affecting 366 the rest of the functions. 367 368 In fact, we're going to go a step further and put every function in its own 369 module. Doing so allows us to exploit a useful property of the KaleidoscopeJIT 370 that will make our environment more REPL-like: Functions can be added to the 371 JIT more than once (unlike a module where every function must have a unique 372 definition). When you look up a symbol in KaleidoscopeJIT it will always return 373 the most recent definition: 374 375 :: 376 377 ready> def foo(x) x + 1; 378 Read function definition: 379 define double @foo(double %x) { 380 entry: 381 %addtmp = fadd double %x, 1.000000e+00 382 ret double %addtmp 383 } 384 385 ready> foo(2); 386 Evaluated to 3.000000 387 388 ready> def foo(x) x + 2; 389 define double @foo(double %x) { 390 entry: 391 %addtmp = fadd double %x, 2.000000e+00 392 ret double %addtmp 393 } 394 395 ready> foo(2); 396 Evaluated to 4.000000 397 398 399 To allow each function to live in its own module we'll need a way to 400 re-generate previous function declarations into each new module we open: 401 402 .. code-block:: c++ 403 404 static std::unique_ptr<KaleidoscopeJIT> TheJIT; 405 406 ... 407 408 Function *getFunction(std::string Name) { 409 // First, see if the function has already been added to the current module. 410 if (auto *F = TheModule->getFunction(Name)) 411 return F; 412 413 // If not, check whether we can codegen the declaration from some existing 414 // prototype. 415 auto FI = FunctionProtos.find(Name); 416 if (FI != FunctionProtos.end()) 417 return FI->second->codegen(); 418 419 // If no existing prototype exists, return null. 420 return nullptr; 421 } 422 423 ... 424 425 Value *CallExprAST::codegen() { 426 // Look up the name in the global module table. 427 Function *CalleeF = getFunction(Callee); 428 429 ... 430 431 Function *FunctionAST::codegen() { 432 // Transfer ownership of the prototype to the FunctionProtos map, but keep a 433 // reference to it for use below. 434 auto &P = *Proto; 435 FunctionProtos[Proto->getName()] = std::move(Proto); 436 Function *TheFunction = getFunction(P.getName()); 437 if (!TheFunction) 438 return nullptr; 439 440 441 To enable this, we'll start by adding a new global, ``FunctionProtos``, that 442 holds the most recent prototype for each function. We'll also add a convenience 443 method, ``getFunction()``, to replace calls to ``TheModule->getFunction()``. 444 Our convenience method searches ``TheModule`` for an existing function 445 declaration, falling back to generating a new declaration from FunctionProtos if 446 it doesn't find one. In ``CallExprAST::codegen()`` we just need to replace the 447 call to ``TheModule->getFunction()``. In ``FunctionAST::codegen()`` we need to 448 update the FunctionProtos map first, then call ``getFunction()``. With this 449 done, we can always obtain a function declaration in the current module for any 450 previously declared function. 451 452 We also need to update HandleDefinition and HandleExtern: 453 454 .. code-block:: c++ 455 456 static void HandleDefinition() { 457 if (auto FnAST = ParseDefinition()) { 458 if (auto *FnIR = FnAST->codegen()) { 459 fprintf(stderr, "Read function definition:"); 460 FnIR->dump(); 461 TheJIT->addModule(std::move(TheModule)); 462 InitializeModuleAndPassManager(); 463 } 464 } else { 465 // Skip token for error recovery. 466 getNextToken(); 467 } 468 } 469 470 static void HandleExtern() { 471 if (auto ProtoAST = ParseExtern()) { 472 if (auto *FnIR = ProtoAST->codegen()) { 473 fprintf(stderr, "Read extern: "); 474 FnIR->dump(); 475 FunctionProtos[ProtoAST->getName()] = std::move(ProtoAST); 476 } 477 } else { 478 // Skip token for error recovery. 479 getNextToken(); 480 } 481 } 482 483 In HandleDefinition, we add two lines to transfer the newly defined function to 484 the JIT and open a new module. In HandleExtern, we just need to add one line to 485 add the prototype to FunctionProtos. 486 487 With these changes made, lets try our REPL again (I removed the dump of the 488 anonymous functions this time, you should get the idea by now :) : 489 490 :: 491 492 ready> def foo(x) x + 1; 493 ready> foo(2); 494 Evaluated to 3.000000 495 496 ready> def foo(x) x + 2; 497 ready> foo(2); 498 Evaluated to 4.000000 499 500 It works! 501 502 Even with this simple code, we get some surprisingly powerful capabilities - 503 check this out: 504 505 :: 506 507 ready> extern sin(x); 508 Read extern: 509 declare double @sin(double) 510 511 ready> extern cos(x); 512 Read extern: 513 declare double @cos(double) 514 515 ready> sin(1.0); 516 Read top-level expression: 517 define double @2() { 518 entry: 519 ret double 0x3FEAED548F090CEE 520 } 521 522 Evaluated to 0.841471 523 524 ready> def foo(x) sin(x)*sin(x) + cos(x)*cos(x); 525 Read function definition: 526 define double @foo(double %x) { 527 entry: 528 %calltmp = call double @sin(double %x) 529 %multmp = fmul double %calltmp, %calltmp 530 %calltmp2 = call double @cos(double %x) 531 %multmp4 = fmul double %calltmp2, %calltmp2 532 %addtmp = fadd double %multmp, %multmp4 533 ret double %addtmp 534 } 535 536 ready> foo(4.0); 537 Read top-level expression: 538 define double @3() { 539 entry: 540 %calltmp = call double @foo(double 4.000000e+00) 541 ret double %calltmp 542 } 543 544 Evaluated to 1.000000 545 546 Whoa, how does the JIT know about sin and cos? The answer is surprisingly 547 simple: The KaleidoscopeJIT has a straightforward symbol resolution rule that 548 it uses to find symbols that aren't available in any given module: First 549 it searches all the modules that have already been added to the JIT, from the 550 most recent to the oldest, to find the newest definition. If no definition is 551 found inside the JIT, it falls back to calling "``dlsym("sin")``" on the 552 Kaleidoscope process itself. Since "``sin``" is defined within the JIT's 553 address space, it simply patches up calls in the module to call the libm 554 version of ``sin`` directly. 555 556 In the future we'll see how tweaking this symbol resolution rule can be used to 557 enable all sorts of useful features, from security (restricting the set of 558 symbols available to JIT'd code), to dynamic code generation based on symbol 559 names, and even lazy compilation. 560 561 One immediate benefit of the symbol resolution rule is that we can now extend 562 the language by writing arbitrary C++ code to implement operations. For example, 563 if we add: 564 565 .. code-block:: c++ 566 567 /// putchard - putchar that takes a double and returns 0. 568 extern "C" double putchard(double X) { 569 fputc((char)X, stderr); 570 return 0; 571 } 572 573 Now we can produce simple output to the console by using things like: 574 "``extern putchard(x); putchard(120);``", which prints a lowercase 'x' 575 on the console (120 is the ASCII code for 'x'). Similar code could be 576 used to implement file I/O, console input, and many other capabilities 577 in Kaleidoscope. 578 579 This completes the JIT and optimizer chapter of the Kaleidoscope 580 tutorial. At this point, we can compile a non-Turing-complete 581 programming language, optimize and JIT compile it in a user-driven way. 582 Next up we'll look into `extending the language with control flow 583 constructs <LangImpl5.html>`_, tackling some interesting LLVM IR issues 584 along the way. 585 586 Full Code Listing 587 ================= 588 589 Here is the complete code listing for our running example, enhanced with 590 the LLVM JIT and optimizer. To build this example, use: 591 592 .. code-block:: bash 593 594 # Compile 595 clang++ -g toy.cpp `llvm-config --cxxflags --ldflags --system-libs --libs core mcjit native` -O3 -o toy 596 # Run 597 ./toy 598 599 If you are compiling this on Linux, make sure to add the "-rdynamic" 600 option as well. This makes sure that the external functions are resolved 601 properly at runtime. 602 603 Here is the code: 604 605 .. literalinclude:: ../../examples/Kaleidoscope/Chapter4/toy.cpp 606 :language: c++ 607 608 `Next: Extending the language: control flow <LangImpl5.html>`_ 609 610