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#passmanager>`_ 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. The code looks like 126 this: 127 128 .. code-block:: c++ 129 130 FunctionPassManager OurFPM(TheModule); 131 132 // Set up the optimizer pipeline. Start with registering info about how the 133 // target lays out data structures. 134 OurFPM.add(new DataLayout(*TheExecutionEngine->getDataLayout())); 135 // Provide basic AliasAnalysis support for GVN. 136 OurFPM.add(createBasicAliasAnalysisPass()); 137 // Do simple "peephole" optimizations and bit-twiddling optzns. 138 OurFPM.add(createInstructionCombiningPass()); 139 // Reassociate expressions. 140 OurFPM.add(createReassociatePass()); 141 // Eliminate Common SubExpressions. 142 OurFPM.add(createGVNPass()); 143 // Simplify the control flow graph (deleting unreachable blocks, etc). 144 OurFPM.add(createCFGSimplificationPass()); 145 146 OurFPM.doInitialization(); 147 148 // Set the global so the code gen can use this. 149 TheFPM = &OurFPM; 150 151 // Run the main "interpreter loop" now. 152 MainLoop(); 153 154 This code defines a ``FunctionPassManager``, "``OurFPM``". It requires a 155 pointer to the ``Module`` to construct itself. Once it is set up, we use 156 a series of "add" calls to add a bunch of LLVM passes. The first pass is 157 basically boilerplate, it adds a pass so that later optimizations know 158 how the data structures in the program are laid out. The 159 "``TheExecutionEngine``" variable is related to the JIT, which we will 160 get to in the next section. 161 162 In this case, we choose to add 4 optimization passes. The passes we 163 chose here are a pretty standard set of "cleanup" optimizations that are 164 useful for a wide variety of code. I won't delve into what they do but, 165 believe me, they are a good starting place :). 166 167 Once the PassManager is set up, we need to make use of it. We do this by 168 running it after our newly created function is constructed (in 169 ``FunctionAST::Codegen``), but before it is returned to the client: 170 171 .. code-block:: c++ 172 173 if (Value *RetVal = Body->Codegen()) { 174 // Finish off the function. 175 Builder.CreateRet(RetVal); 176 177 // Validate the generated code, checking for consistency. 178 verifyFunction(*TheFunction); 179 180 // Optimize the function. 181 TheFPM->run(*TheFunction); 182 183 return TheFunction; 184 } 185 186 As you can see, this is pretty straightforward. The 187 ``FunctionPassManager`` optimizes and updates the LLVM Function\* in 188 place, improving (hopefully) its body. With this in place, we can try 189 our test above again: 190 191 :: 192 193 ready> def test(x) (1+2+x)*(x+(1+2)); 194 ready> Read function definition: 195 define double @test(double %x) { 196 entry: 197 %addtmp = fadd double %x, 3.000000e+00 198 %multmp = fmul double %addtmp, %addtmp 199 ret double %multmp 200 } 201 202 As expected, we now get our nicely optimized code, saving a floating 203 point add instruction from every execution of this function. 204 205 LLVM provides a wide variety of optimizations that can be used in 206 certain circumstances. Some `documentation about the various 207 passes <../Passes.html>`_ is available, but it isn't very complete. 208 Another good source of ideas can come from looking at the passes that 209 ``Clang`` runs to get started. The "``opt``" tool allows you to 210 experiment with passes from the command line, so you can see if they do 211 anything. 212 213 Now that we have reasonable code coming out of our front-end, lets talk 214 about executing it! 215 216 Adding a JIT Compiler 217 ===================== 218 219 Code that is available in LLVM IR can have a wide variety of tools 220 applied to it. For example, you can run optimizations on it (as we did 221 above), you can dump it out in textual or binary forms, you can compile 222 the code to an assembly file (.s) for some target, or you can JIT 223 compile it. The nice thing about the LLVM IR representation is that it 224 is the "common currency" between many different parts of the compiler. 225 226 In this section, we'll add JIT compiler support to our interpreter. The 227 basic idea that we want for Kaleidoscope is to have the user enter 228 function bodies as they do now, but immediately evaluate the top-level 229 expressions they type in. For example, if they type in "1 + 2;", we 230 should evaluate and print out 3. If they define a function, they should 231 be able to call it from the command line. 232 233 In order to do this, we first declare and initialize the JIT. This is 234 done by adding a global variable and a call in ``main``: 235 236 .. code-block:: c++ 237 238 static ExecutionEngine *TheExecutionEngine; 239 ... 240 int main() { 241 .. 242 // Create the JIT. This takes ownership of the module. 243 TheExecutionEngine = EngineBuilder(TheModule).create(); 244 .. 245 } 246 247 This creates an abstract "Execution Engine" which can be either a JIT 248 compiler or the LLVM interpreter. LLVM will automatically pick a JIT 249 compiler for you if one is available for your platform, otherwise it 250 will fall back to the interpreter. 251 252 Once the ``ExecutionEngine`` is created, the JIT is ready to be used. 253 There are a variety of APIs that are useful, but the simplest one is the 254 "``getPointerToFunction(F)``" method. This method JIT compiles the 255 specified LLVM Function and returns a function pointer to the generated 256 machine code. In our case, this means that we can change the code that 257 parses a top-level expression to look like this: 258 259 .. code-block:: c++ 260 261 static void HandleTopLevelExpression() { 262 // Evaluate a top-level expression into an anonymous function. 263 if (FunctionAST *F = ParseTopLevelExpr()) { 264 if (Function *LF = F->Codegen()) { 265 LF->dump(); // Dump the function for exposition purposes. 266 267 // JIT the function, returning a function pointer. 268 void *FPtr = TheExecutionEngine->getPointerToFunction(LF); 269 270 // Cast it to the right type (takes no arguments, returns a double) so we 271 // can call it as a native function. 272 double (*FP)() = (double (*)())(intptr_t)FPtr; 273 fprintf(stderr, "Evaluated to %f\n", FP()); 274 } 275 276 Recall that we compile top-level expressions into a self-contained LLVM 277 function that takes no arguments and returns the computed double. 278 Because the LLVM JIT compiler matches the native platform ABI, this 279 means that you can just cast the result pointer to a function pointer of 280 that type and call it directly. This means, there is no difference 281 between JIT compiled code and native machine code that is statically 282 linked into your application. 283 284 With just these two changes, lets see how Kaleidoscope works now! 285 286 :: 287 288 ready> 4+5; 289 Read top-level expression: 290 define double @0() { 291 entry: 292 ret double 9.000000e+00 293 } 294 295 Evaluated to 9.000000 296 297 Well this looks like it is basically working. The dump of the function 298 shows the "no argument function that always returns double" that we 299 synthesize for each top-level expression that is typed in. This 300 demonstrates very basic functionality, but can we do more? 301 302 :: 303 304 ready> def testfunc(x y) x + y*2; 305 Read function definition: 306 define double @testfunc(double %x, double %y) { 307 entry: 308 %multmp = fmul double %y, 2.000000e+00 309 %addtmp = fadd double %multmp, %x 310 ret double %addtmp 311 } 312 313 ready> testfunc(4, 10); 314 Read top-level expression: 315 define double @1() { 316 entry: 317 %calltmp = call double @testfunc(double 4.000000e+00, double 1.000000e+01) 318 ret double %calltmp 319 } 320 321 Evaluated to 24.000000 322 323 This illustrates that we can now call user code, but there is something 324 a bit subtle going on here. Note that we only invoke the JIT on the 325 anonymous functions that *call testfunc*, but we never invoked it on 326 *testfunc* itself. What actually happened here is that the JIT scanned 327 for all non-JIT'd functions transitively called from the anonymous 328 function and compiled all of them before returning from 329 ``getPointerToFunction()``. 330 331 The JIT provides a number of other more advanced interfaces for things 332 like freeing allocated machine code, rejit'ing functions to update them, 333 etc. However, even with this simple code, we get some surprisingly 334 powerful capabilities - check this out (I removed the dump of the 335 anonymous functions, you should get the idea by now :) : 336 337 :: 338 339 ready> extern sin(x); 340 Read extern: 341 declare double @sin(double) 342 343 ready> extern cos(x); 344 Read extern: 345 declare double @cos(double) 346 347 ready> sin(1.0); 348 Read top-level expression: 349 define double @2() { 350 entry: 351 ret double 0x3FEAED548F090CEE 352 } 353 354 Evaluated to 0.841471 355 356 ready> def foo(x) sin(x)*sin(x) + cos(x)*cos(x); 357 Read function definition: 358 define double @foo(double %x) { 359 entry: 360 %calltmp = call double @sin(double %x) 361 %multmp = fmul double %calltmp, %calltmp 362 %calltmp2 = call double @cos(double %x) 363 %multmp4 = fmul double %calltmp2, %calltmp2 364 %addtmp = fadd double %multmp, %multmp4 365 ret double %addtmp 366 } 367 368 ready> foo(4.0); 369 Read top-level expression: 370 define double @3() { 371 entry: 372 %calltmp = call double @foo(double 4.000000e+00) 373 ret double %calltmp 374 } 375 376 Evaluated to 1.000000 377 378 Whoa, how does the JIT know about sin and cos? The answer is 379 surprisingly simple: in this example, the JIT started execution of a 380 function and got to a function call. It realized that the function was 381 not yet JIT compiled and invoked the standard set of routines to resolve 382 the function. In this case, there is no body defined for the function, 383 so the JIT ended up calling "``dlsym("sin")``" on the Kaleidoscope 384 process itself. Since "``sin``" is defined within the JIT's address 385 space, it simply patches up calls in the module to call the libm version 386 of ``sin`` directly. 387 388 The LLVM JIT provides a number of interfaces (look in the 389 ``ExecutionEngine.h`` file) for controlling how unknown functions get 390 resolved. It allows you to establish explicit mappings between IR 391 objects and addresses (useful for LLVM global variables that you want to 392 map to static tables, for example), allows you to dynamically decide on 393 the fly based on the function name, and even allows you to have the JIT 394 compile functions lazily the first time they're called. 395 396 One interesting application of this is that we can now extend the 397 language by writing arbitrary C++ code to implement operations. For 398 example, if we add: 399 400 .. code-block:: c++ 401 402 /// putchard - putchar that takes a double and returns 0. 403 extern "C" 404 double putchard(double X) { 405 putchar((char)X); 406 return 0; 407 } 408 409 Now we can produce simple output to the console by using things like: 410 "``extern putchard(x); putchard(120);``", which prints a lowercase 'x' 411 on the console (120 is the ASCII code for 'x'). Similar code could be 412 used to implement file I/O, console input, and many other capabilities 413 in Kaleidoscope. 414 415 This completes the JIT and optimizer chapter of the Kaleidoscope 416 tutorial. At this point, we can compile a non-Turing-complete 417 programming language, optimize and JIT compile it in a user-driven way. 418 Next up we'll look into `extending the language with control flow 419 constructs <LangImpl5.html>`_, tackling some interesting LLVM IR issues 420 along the way. 421 422 Full Code Listing 423 ================= 424 425 Here is the complete code listing for our running example, enhanced with 426 the LLVM JIT and optimizer. To build this example, use: 427 428 .. code-block:: bash 429 430 # Compile 431 clang++ -g toy.cpp `llvm-config --cppflags --ldflags --libs core jit native` -O3 -o toy 432 # Run 433 ./toy 434 435 If you are compiling this on Linux, make sure to add the "-rdynamic" 436 option as well. This makes sure that the external functions are resolved 437 properly at runtime. 438 439 Here is the code: 440 441 .. literalinclude:: ../../examples/Kaleidoscope/Chapter4/toy.cpp 442 :language: c++ 443 444 `Next: Extending the language: control flow <LangImpl5.html>`_ 445 446