blob: ddd609210d8d9ff1a237f269b3e819269f8ad415 [file] [log] [blame]
Chris Lattnerc0b42e92007-10-23 06:27:55 +00001<!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
Chris Lattner128eb862007-11-05 19:06:59 +000016<ul>
Chris Lattner0e555b12007-11-05 20:04:56 +000017<li><a href="index.html">Up to Tutorial Index</a></li>
Chris Lattner128eb862007-11-05 19:06:59 +000018<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>
Chris Lattner0e555b12007-11-05 20:04:56 +000027<li><a href="LangImpl5.html">Chapter 5</a>: Extending the Language: Control
28Flow</li>
Chris Lattner128eb862007-11-05 19:06:59 +000029</ul>
30
Chris Lattnerc0b42e92007-10-23 06:27:55 +000031<div class="doc_author">
32 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a></p>
33</div>
34
35<!-- *********************************************************************** -->
Chris Lattner128eb862007-11-05 19:06:59 +000036<div class="doc_section"><a name="intro">Chapter 4 Introduction</a></div>
Chris Lattnerc0b42e92007-10-23 06:27:55 +000037<!-- *********************************************************************** -->
38
39<div class="doc_text">
40
Chris Lattner128eb862007-11-05 19:06:59 +000041<p>Welcome to Chapter 4 of the "<a href="index.html">Implementing a language
Chris Lattnera54c2012007-11-07 05:28:43 +000042with LLVM</a>" tutorial. Chapters 1-3 described the implementation of a simple
43language and added support for generating LLVM IR. This chapter describes
Chris Lattner128eb862007-11-05 19:06:59 +000044two new techniques: adding optimizer support to your language, and adding JIT
Chris Lattner41fcea32007-11-13 07:06:30 +000045compiler support. These additions will demonstrate how to get nice, efficient code
46for the Kaleidoscope language.</p>
Chris Lattnerc0b42e92007-10-23 06:27:55 +000047
48</div>
49
50<!-- *********************************************************************** -->
Chris Lattner118749e2007-10-25 06:23:36 +000051<div class="doc_section"><a name="trivialconstfold">Trivial Constant
52Folding</a></div>
Chris Lattnerc0b42e92007-10-23 06:27:55 +000053<!-- *********************************************************************** -->
54
55<div class="doc_text">
56
57<p>
Chris Lattner118749e2007-10-25 06:23:36 +000058Our demonstration for Chapter 3 is elegant and easy to extend. Unfortunately,
Duncan Sands89f6d882008-04-13 06:22:09 +000059it does not produce wonderful code. The IRBuilder, however, does give us
60obvious optimizations when compiling simple code:</p>
Chris Lattner118749e2007-10-25 06:23:36 +000061
62<div class="doc_code">
63<pre>
64ready&gt; <b>def test(x) 1+2+x;</b>
65Read function definition:
66define double @test(double %x) {
67entry:
68 %addtmp = add double 3.000000e+00, %x
69 ret double %addtmp
70}
71</pre>
72</div>
73
Duncan Sands89f6d882008-04-13 06:22:09 +000074<p>This code is not a literal transcription of the AST built by parsing the
75input. That would be:
76
77<div class="doc_code">
78<pre>
79ready&gt; <b>def test(x) 1+2+x;</b>
80Read function definition:
81define double @test(double %x) {
82entry:
83 %addtmp = add double 2.000000e+00, 1.000000e+00
84 %addtmp1 = add double %addtmp, %x
85 ret double %addtmp1
86}
87</pre>
88</div>
89
90Constant folding, as seen above, in particular, is a very common and very
91important optimization: so much so that many language implementors implement
92constant 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
95LLVM IR go through the LLVM IR builder, the builder itself checked to see if
96there was a constant folding opportunity when you call it. If so, it just does
97the constant fold and return the constant instead of creating an instruction.
98
Chris Lattnera54c2012007-11-07 05:28:43 +000099<p>Well, that was easy :). In practice, we recommend always using
Duncan Sands89f6d882008-04-13 06:22:09 +0000100<tt>IRBuilder</tt> when generating code like this. It has no
Chris Lattner118749e2007-10-25 06:23:36 +0000101"syntactic overhead" for its use (you don't have to uglify your compiler with
102constant checks everywhere) and it can dramatically reduce the amount of
103LLVM IR that is generated in some cases (particular for languages with a macro
104preprocessor or that use a lot of constants).</p>
105
Duncan Sands89f6d882008-04-13 06:22:09 +0000106<p>On the other hand, the <tt>IRBuilder</tt> is limited by the fact
Chris Lattner118749e2007-10-25 06:23:36 +0000107that it does all of its analysis inline with the code as it is built. If you
108take a slightly more complex example:</p>
109
110<div class="doc_code">
111<pre>
112ready&gt; <b>def test(x) (1+2+x)*(x+(1+2));</b>
113ready> Read function definition:
114define double @test(double %x) {
115entry:
116 %addtmp = add double 3.000000e+00, %x
117 %addtmp1 = add double %x, 3.000000e+00
118 %multmp = mul 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
125really like to see this generate "<tt>tmp = x+3; result = tmp*tmp;</tt>" instead
Chris Lattner1ace67c2008-04-15 16:59:22 +0000126of computing "<tt>x+3</tt>" twice.</p>
Chris Lattner118749e2007-10-25 06:23:36 +0000127
128<p>Unfortunately, no amount of local analysis will be able to detect and correct
129this. This requires two transformations: reassociation of expressions (to
130make the add's lexically identical) and Common Subexpression Elimination (CSE)
131to delete the redundant add instruction. Fortunately, LLVM provides a broad
132range 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
Chris Lattner41fcea32007-11-13 07:06:30 +0000143<p>LLVM provides many optimization passes, which do many different sorts of
Chris Lattner118749e2007-10-25 06:23:36 +0000144things and have different tradeoffs. Unlike other systems, LLVM doesn't hold
145to the mistaken notion that one set of optimizations is right for all languages
146and for all situations. LLVM allows a compiler implementor to make complete
147decisions about what optimizations to use, in which order, and in what
148situation.</p>
149
150<p>As a concrete example, LLVM supports both "whole module" passes, which look
151across as large of body of code as they can (often a whole file, but if run
152at link time, this can be a substantial portion of the whole program). It also
153supports and includes "per-function" passes which just operate on a single
154function at a time, without looking at other functions. For more information
Chris Lattner41fcea32007-11-13 07:06:30 +0000155on passes and how they are run, see the <a href="../WritingAnLLVMPass.html">How
Chris Lattnera54c2012007-11-07 05:28:43 +0000156to Write a Pass</a> document and the <a href="../Passes.html">List of LLVM
157Passes</a>.</p>
Chris Lattner118749e2007-10-25 06:23:36 +0000158
159<p>For Kaleidoscope, we are currently generating functions on the fly, one at
160a time, as the user types them in. We aren't shooting for the ultimate
161optimization experience in this setting, but we also want to catch the easy and
162quick stuff where possible. As such, we will choose to run a few per-function
163optimizations as the user types the function in. If we wanted to make a "static
164Kaleidoscope compiler", we would use exactly the code we have now, except that
165we 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
169organize the LLVM optimizations that we want to run. Once we have that, we can
170add a set of optimizations to run. The code looks like this:</p>
171
172<div class="doc_code">
173<pre>
174 ExistingModuleProvider OurModuleProvider(TheModule);
175 FunctionPassManager OurFPM(&amp;OurModuleProvider);
176
177 // Set up the optimizer pipeline. Start with registering info about how the
178 // target lays out data structures.
179 OurFPM.add(new TargetData(*TheExecutionEngine->getTargetData()));
180 // Do simple "peephole" optimizations and bit-twiddling optzns.
181 OurFPM.add(createInstructionCombiningPass());
182 // Reassociate expressions.
183 OurFPM.add(createReassociatePass());
184 // Eliminate Common SubExpressions.
185 OurFPM.add(createGVNPass());
186 // Simplify the control flow graph (deleting unreachable blocks, etc).
187 OurFPM.add(createCFGSimplificationPass());
188
189 // Set the global so the code gen can use this.
190 TheFPM = &amp;OurFPM;
191
192 // Run the main "interpreter loop" now.
193 MainLoop();
194</pre>
195</div>
196
Chris Lattner41fcea32007-11-13 07:06:30 +0000197<p>This code defines two objects, an <tt>ExistingModuleProvider</tt> and a
Chris Lattner118749e2007-10-25 06:23:36 +0000198<tt>FunctionPassManager</tt>. The former is basically a wrapper around our
199<tt>Module</tt> that the PassManager requires. It provides certain flexibility
Chris Lattner41fcea32007-11-13 07:06:30 +0000200that we're not going to take advantage of here, so I won't dive into any details
201about it.</p>
Chris Lattner118749e2007-10-25 06:23:36 +0000202
Chris Lattner41fcea32007-11-13 07:06:30 +0000203<p>The meat of the matter here, is the definition of "<tt>OurFPM</tt>". It
Chris Lattner118749e2007-10-25 06:23:36 +0000204requires a pointer to the <tt>Module</tt> (through the <tt>ModuleProvider</tt>)
205to construct itself. Once it is set up, we use a series of "add" calls to add
206a bunch of LLVM passes. The first pass is basically boilerplate, it adds a pass
207so that later optimizations know how the data structures in the program are
208layed out. The "<tt>TheExecutionEngine</tt>" variable is related to the JIT,
209which we will get to in the next section.</p>
210
211<p>In this case, we choose to add 4 optimization passes. The passes we chose
212here are a pretty standard set of "cleanup" optimizations that are useful for
Chris Lattner41fcea32007-11-13 07:06:30 +0000213a wide variety of code. I won't delve into what they do but, believe me,
Chris Lattnera54c2012007-11-07 05:28:43 +0000214they are a good starting place :).</p>
Chris Lattner118749e2007-10-25 06:23:36 +0000215
Chris Lattnera54c2012007-11-07 05:28:43 +0000216<p>Once the PassManager is set up, we need to make use of it. We do this by
Chris Lattner118749e2007-10-25 06:23:36 +0000217running it after our newly created function is constructed (in
218<tt>FunctionAST::Codegen</tt>), but before it is returned to the client:</p>
219
220<div class="doc_code">
221<pre>
222 if (Value *RetVal = Body->Codegen()) {
223 // Finish off the function.
224 Builder.CreateRet(RetVal);
225
226 // Validate the generated code, checking for consistency.
227 verifyFunction(*TheFunction);
228
Chris Lattnera54c2012007-11-07 05:28:43 +0000229 <b>// Optimize the function.
230 TheFPM-&gt;run(*TheFunction);</b>
Chris Lattner118749e2007-10-25 06:23:36 +0000231
232 return TheFunction;
233 }
234</pre>
235</div>
236
Chris Lattner41fcea32007-11-13 07:06:30 +0000237<p>As you can see, this is pretty straightforward. The
Chris Lattner118749e2007-10-25 06:23:36 +0000238<tt>FunctionPassManager</tt> optimizes and updates the LLVM Function* in place,
239improving (hopefully) its body. With this in place, we can try our test above
240again:</p>
241
242<div class="doc_code">
243<pre>
244ready&gt; <b>def test(x) (1+2+x)*(x+(1+2));</b>
245ready> Read function definition:
246define double @test(double %x) {
247entry:
248 %addtmp = add double %x, 3.000000e+00
249 %multmp = mul double %addtmp, %addtmp
250 ret double %multmp
251}
252</pre>
253</div>
254
255<p>As expected, we now get our nicely optimized code, saving a floating point
Chris Lattnera54c2012007-11-07 05:28:43 +0000256add instruction from every execution of this function.</p>
Chris Lattner118749e2007-10-25 06:23:36 +0000257
258<p>LLVM provides a wide variety of optimizations that can be used in certain
Chris Lattner72714232007-10-25 17:52:39 +0000259circumstances. Some <a href="../Passes.html">documentation about the various
260passes</a> is available, but it isn't very complete. Another good source of
Chris Lattner41fcea32007-11-13 07:06:30 +0000261ideas can come from looking at the passes that <tt>llvm-gcc</tt> or
Chris Lattner118749e2007-10-25 06:23:36 +0000262<tt>llvm-ld</tt> run to get started. The "<tt>opt</tt>" tool allows you to
263experiment with passes from the command line, so you can see if they do
264anything.</p>
265
266<p>Now that we have reasonable code coming out of our front-end, lets talk about
267executing it!</p>
268
269</div>
270
271<!-- *********************************************************************** -->
272<div class="doc_section"><a name="jit">Adding a JIT Compiler</a></div>
273<!-- *********************************************************************** -->
274
275<div class="doc_text">
276
Chris Lattnera54c2012007-11-07 05:28:43 +0000277<p>Code that is available in LLVM IR can have a wide variety of tools
Chris Lattner118749e2007-10-25 06:23:36 +0000278applied to it. For example, you can run optimizations on it (as we did above),
279you can dump it out in textual or binary forms, you can compile the code to an
280assembly file (.s) for some target, or you can JIT compile it. The nice thing
Chris Lattnera54c2012007-11-07 05:28:43 +0000281about the LLVM IR representation is that it is the "common currency" between
282many different parts of the compiler.
Chris Lattner118749e2007-10-25 06:23:36 +0000283</p>
284
Chris Lattnera54c2012007-11-07 05:28:43 +0000285<p>In this section, we'll add JIT compiler support to our interpreter. The
Chris Lattner118749e2007-10-25 06:23:36 +0000286basic idea that we want for Kaleidoscope is to have the user enter function
287bodies as they do now, but immediately evaluate the top-level expressions they
288type in. For example, if they type in "1 + 2;", we should evaluate and print
289out 3. If they define a function, they should be able to call it from the
290command line.</p>
291
292<p>In order to do this, we first declare and initialize the JIT. This is done
293by adding a global variable and a call in <tt>main</tt>:</p>
294
295<div class="doc_code">
296<pre>
Chris Lattnera54c2012007-11-07 05:28:43 +0000297<b>static ExecutionEngine *TheExecutionEngine;</b>
Chris Lattner118749e2007-10-25 06:23:36 +0000298...
299int main() {
300 ..
Chris Lattnera54c2012007-11-07 05:28:43 +0000301 <b>// Create the JIT.
302 TheExecutionEngine = ExecutionEngine::create(TheModule);</b>
Chris Lattner118749e2007-10-25 06:23:36 +0000303 ..
304}
305</pre>
306</div>
307
308<p>This creates an abstract "Execution Engine" which can be either a JIT
309compiler or the LLVM interpreter. LLVM will automatically pick a JIT compiler
310for you if one is available for your platform, otherwise it will fall back to
311the interpreter.</p>
312
313<p>Once the <tt>ExecutionEngine</tt> is created, the JIT is ready to be used.
Chris Lattner41fcea32007-11-13 07:06:30 +0000314There are a variety of APIs that are useful, but the simplest one is the
Chris Lattner118749e2007-10-25 06:23:36 +0000315"<tt>getPointerToFunction(F)</tt>" method. This method JIT compiles the
316specified LLVM Function and returns a function pointer to the generated machine
317code. In our case, this means that we can change the code that parses a
318top-level expression to look like this:</p>
319
320<div class="doc_code">
321<pre>
322static void HandleTopLevelExpression() {
323 // Evaluate a top level expression into an anonymous function.
324 if (FunctionAST *F = ParseTopLevelExpr()) {
325 if (Function *LF = F-&gt;Codegen()) {
326 LF->dump(); // Dump the function for exposition purposes.
327
Chris Lattnera54c2012007-11-07 05:28:43 +0000328 <b>// JIT the function, returning a function pointer.
Chris Lattner118749e2007-10-25 06:23:36 +0000329 void *FPtr = TheExecutionEngine-&gt;getPointerToFunction(LF);
330
331 // Cast it to the right type (takes no arguments, returns a double) so we
332 // can call it as a native function.
333 double (*FP)() = (double (*)())FPtr;
Chris Lattnera54c2012007-11-07 05:28:43 +0000334 fprintf(stderr, "Evaluated to %f\n", FP());</b>
Chris Lattner118749e2007-10-25 06:23:36 +0000335 }
336</pre>
337</div>
338
339<p>Recall that we compile top-level expressions into a self-contained LLVM
340function that takes no arguments and returns the computed double. Because the
341LLVM JIT compiler matches the native platform ABI, this means that you can just
342cast the result pointer to a function pointer of that type and call it directly.
Chris Lattner41fcea32007-11-13 07:06:30 +0000343This means, there is no difference between JIT compiled code and native machine
Chris Lattner118749e2007-10-25 06:23:36 +0000344code that is statically linked into your application.</p>
345
346<p>With just these two changes, lets see how Kaleidoscope works now!</p>
347
348<div class="doc_code">
349<pre>
350ready&gt; <b>4+5;</b>
351define double @""() {
352entry:
353 ret double 9.000000e+00
354}
355
356<em>Evaluated to 9.000000</em>
357</pre>
358</div>
359
360<p>Well this looks like it is basically working. The dump of the function
361shows the "no argument function that always returns double" that we synthesize
Chris Lattner41fcea32007-11-13 07:06:30 +0000362for each top level expression that is typed in. This demonstrates very basic
Chris Lattner118749e2007-10-25 06:23:36 +0000363functionality, but can we do more?</p>
364
365<div class="doc_code">
366<pre>
Chris Lattner2e89f3a2007-10-31 07:30:39 +0000367ready&gt; <b>def testfunc(x y) x + y*2; </b>
Chris Lattner118749e2007-10-25 06:23:36 +0000368Read function definition:
369define double @testfunc(double %x, double %y) {
370entry:
371 %multmp = mul double %y, 2.000000e+00
372 %addtmp = add double %multmp, %x
373 ret double %addtmp
374}
375
376ready&gt; <b>testfunc(4, 10);</b>
377define double @""() {
378entry:
379 %calltmp = call double @testfunc( double 4.000000e+00, double 1.000000e+01 )
380 ret double %calltmp
381}
382
383<em>Evaluated to 24.000000</em>
384</pre>
385</div>
386
Chris Lattner41fcea32007-11-13 07:06:30 +0000387<p>This illustrates that we can now call user code, but there is something a bit subtle
388going on here. Note that we only invoke the JIT on the anonymous functions
389that <em>call testfunc</em>, but we never invoked it on <em>testfunc
390</em>itself.</p>
Chris Lattner118749e2007-10-25 06:23:36 +0000391
Chris Lattner41fcea32007-11-13 07:06:30 +0000392<p>What actually happened here is that the anonymous function was
Chris Lattner118749e2007-10-25 06:23:36 +0000393JIT'd when requested. When the Kaleidoscope app calls through the function
394pointer that is returned, the anonymous function starts executing. It ends up
Chris Lattnera54c2012007-11-07 05:28:43 +0000395making the call to the "testfunc" function, and ends up in a stub that invokes
Chris Lattner118749e2007-10-25 06:23:36 +0000396the JIT, lazily, on testfunc. Once the JIT finishes lazily compiling testfunc,
Chris Lattnera54c2012007-11-07 05:28:43 +0000397it returns and the code re-executes the call.</p>
Chris Lattner118749e2007-10-25 06:23:36 +0000398
Chris Lattner41fcea32007-11-13 07:06:30 +0000399<p>In summary, the JIT will lazily JIT code, on the fly, as it is needed. The
Chris Lattner118749e2007-10-25 06:23:36 +0000400JIT provides a number of other more advanced interfaces for things like freeing
401allocated machine code, rejit'ing functions to update them, etc. However, even
402with this simple code, we get some surprisingly powerful capabilities - check
403this out (I removed the dump of the anonymous functions, you should get the idea
404by now :) :</p>
405
406<div class="doc_code">
407<pre>
408ready&gt; <b>extern sin(x);</b>
409Read extern:
410declare double @sin(double)
411
412ready&gt; <b>extern cos(x);</b>
413Read extern:
414declare double @cos(double)
415
416ready&gt; <b>sin(1.0);</b>
417<em>Evaluated to 0.841471</em>
Chris Lattner72714232007-10-25 17:52:39 +0000418
Chris Lattner118749e2007-10-25 06:23:36 +0000419ready&gt; <b>def foo(x) sin(x)*sin(x) + cos(x)*cos(x);</b>
420Read function definition:
421define double @foo(double %x) {
422entry:
423 %calltmp = call double @sin( double %x )
424 %multmp = mul double %calltmp, %calltmp
425 %calltmp2 = call double @cos( double %x )
426 %multmp4 = mul double %calltmp2, %calltmp2
427 %addtmp = add double %multmp, %multmp4
428 ret double %addtmp
429}
430
431ready&gt; <b>foo(4.0);</b>
432<em>Evaluated to 1.000000</em>
433</pre>
434</div>
435
Chris Lattnera54c2012007-11-07 05:28:43 +0000436<p>Whoa, how does the JIT know about sin and cos? The answer is surprisingly
437simple: in this
Chris Lattner118749e2007-10-25 06:23:36 +0000438example, the JIT started execution of a function and got to a function call. It
439realized that the function was not yet JIT compiled and invoked the standard set
440of routines to resolve the function. In this case, there is no body defined
Chris Lattnera54c2012007-11-07 05:28:43 +0000441for the function, so the JIT ended up calling "<tt>dlsym("sin")</tt>" on the
442Kaleidoscope process itself.
Chris Lattner118749e2007-10-25 06:23:36 +0000443Since "<tt>sin</tt>" is defined within the JIT's address space, it simply
444patches up calls in the module to call the libm version of <tt>sin</tt>
445directly.</p>
446
447<p>The LLVM JIT provides a number of interfaces (look in the
448<tt>ExecutionEngine.h</tt> file) for controlling how unknown functions get
449resolved. It allows you to establish explicit mappings between IR objects and
450addresses (useful for LLVM global variables that you want to map to static
451tables, for example), allows you to dynamically decide on the fly based on the
452function name, and even allows you to have the JIT abort itself if any lazy
453compilation is attempted.</p>
454
Chris Lattner72714232007-10-25 17:52:39 +0000455<p>One interesting application of this is that we can now extend the language
456by writing arbitrary C++ code to implement operations. For example, if we add:
457</p>
458
459<div class="doc_code">
460<pre>
461/// putchard - putchar that takes a double and returns 0.
462extern "C"
463double putchard(double X) {
464 putchar((char)X);
465 return 0;
466}
467</pre>
468</div>
469
470<p>Now we can produce simple output to the console by using things like:
471"<tt>extern putchard(x); putchard(120);</tt>", which prints a lowercase 'x' on
Chris Lattnera54c2012007-11-07 05:28:43 +0000472the console (120 is the ASCII code for 'x'). Similar code could be used to
Chris Lattner72714232007-10-25 17:52:39 +0000473implement file I/O, console input, and many other capabilities in
474Kaleidoscope.</p>
475
Chris Lattner118749e2007-10-25 06:23:36 +0000476<p>This completes the JIT and optimizer chapter of the Kaleidoscope tutorial. At
477this point, we can compile a non-Turing-complete programming language, optimize
478and JIT compile it in a user-driven way. Next up we'll look into <a
479href="LangImpl5.html">extending the language with control flow constructs</a>,
480tackling some interesting LLVM IR issues along the way.</p>
481
482</div>
483
484<!-- *********************************************************************** -->
485<div class="doc_section"><a name="code">Full Code Listing</a></div>
486<!-- *********************************************************************** -->
487
488<div class="doc_text">
489
490<p>
491Here is the complete code listing for our running example, enhanced with the
492LLVM JIT and optimizer. To build this example, use:
493</p>
494
495<div class="doc_code">
496<pre>
497 # Compile
498 g++ -g toy.cpp `llvm-config --cppflags --ldflags --libs core jit native` -O3 -o toy
499 # Run
500 ./toy
501</pre>
502</div>
503
504<p>Here is the code:</p>
505
506<div class="doc_code">
507<pre>
508#include "llvm/DerivedTypes.h"
509#include "llvm/ExecutionEngine/ExecutionEngine.h"
510#include "llvm/Module.h"
511#include "llvm/ModuleProvider.h"
512#include "llvm/PassManager.h"
513#include "llvm/Analysis/Verifier.h"
514#include "llvm/Target/TargetData.h"
515#include "llvm/Transforms/Scalar.h"
Duncan Sands89f6d882008-04-13 06:22:09 +0000516#include "llvm/Support/IRBuilder.h"
Chris Lattner118749e2007-10-25 06:23:36 +0000517#include &lt;cstdio&gt;
518#include &lt;string&gt;
519#include &lt;map&gt;
520#include &lt;vector&gt;
521using 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.
529enum 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
539static std::string IdentifierStr; // Filled in if tok_identifier
540static double NumVal; // Filled in if tok_number
541
542/// gettok - Return the next token from standard input.
543static 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();
Chris Lattnerc80c23f2007-12-02 22:46:01 +0000574 while (LastChar != EOF &amp;&amp; LastChar != '\n' &amp;&amp; LastChar != '\r');
Chris Lattner118749e2007-10-25 06:23:36 +0000575
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
Chris Lattnerc0b42e92007-10-23 06:27:55 +0000594/// ExprAST - Base class for all expression nodes.
595class ExprAST {
596public:
597 virtual ~ExprAST() {}
598 virtual Value *Codegen() = 0;
599};
600
601/// NumberExprAST - Expression class for numeric literals like "1.0".
602class NumberExprAST : public ExprAST {
603 double Val;
604public:
Chris Lattner118749e2007-10-25 06:23:36 +0000605 NumberExprAST(double val) : Val(val) {}
Chris Lattnerc0b42e92007-10-23 06:27:55 +0000606 virtual Value *Codegen();
607};
Chris Lattner118749e2007-10-25 06:23:36 +0000608
609/// VariableExprAST - Expression class for referencing a variable, like "a".
610class VariableExprAST : public ExprAST {
611 std::string Name;
612public:
613 VariableExprAST(const std::string &amp;name) : Name(name) {}
614 virtual Value *Codegen();
615};
616
617/// BinaryExprAST - Expression class for a binary operator.
618class BinaryExprAST : public ExprAST {
619 char Op;
620 ExprAST *LHS, *RHS;
621public:
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.
628class CallExprAST : public ExprAST {
629 std::string Callee;
630 std::vector&lt;ExprAST*&gt; Args;
631public:
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 argument names as well as if it is an operator.
639class PrototypeAST {
640 std::string Name;
641 std::vector&lt;std::string&gt; Args;
642public:
643 PrototypeAST(const std::string &amp;name, const std::vector&lt;std::string&gt; &amp;args)
644 : Name(name), Args(args) {}
645
646 Function *Codegen();
647};
648
649/// FunctionAST - This class represents a function definition itself.
650class FunctionAST {
651 PrototypeAST *Proto;
652 ExprAST *Body;
653public:
654 FunctionAST(PrototypeAST *proto, ExprAST *body)
655 : Proto(proto), Body(body) {}
656
657 Function *Codegen();
658};
659
660//===----------------------------------------------------------------------===//
661// Parser
662//===----------------------------------------------------------------------===//
663
664/// CurTok/getNextToken - Provide a simple token buffer. CurTok is the current
665/// token the parser it looking at. getNextToken reads another token from the
666/// lexer and updates CurTok with its results.
667static int CurTok;
668static int getNextToken() {
669 return CurTok = gettok();
670}
671
672/// BinopPrecedence - This holds the precedence for each binary operator that is
673/// defined.
674static std::map&lt;char, int&gt; BinopPrecedence;
675
676/// GetTokPrecedence - Get the precedence of the pending binary operator token.
677static int GetTokPrecedence() {
678 if (!isascii(CurTok))
679 return -1;
680
681 // Make sure it's a declared binop.
682 int TokPrec = BinopPrecedence[CurTok];
683 if (TokPrec &lt;= 0) return -1;
684 return TokPrec;
685}
686
687/// Error* - These are little helper functions for error handling.
688ExprAST *Error(const char *Str) { fprintf(stderr, "Error: %s\n", Str);return 0;}
689PrototypeAST *ErrorP(const char *Str) { Error(Str); return 0; }
690FunctionAST *ErrorF(const char *Str) { Error(Str); return 0; }
691
692static ExprAST *ParseExpression();
693
694/// identifierexpr
Chris Lattner20a0c802007-11-05 17:54:34 +0000695/// ::= identifier
696/// ::= identifier '(' expression* ')'
Chris Lattner118749e2007-10-25 06:23:36 +0000697static ExprAST *ParseIdentifierExpr() {
698 std::string IdName = IdentifierStr;
699
Chris Lattner20a0c802007-11-05 17:54:34 +0000700 getNextToken(); // eat identifier.
Chris Lattner118749e2007-10-25 06:23:36 +0000701
702 if (CurTok != '(') // Simple variable ref.
703 return new VariableExprAST(IdName);
704
705 // Call.
706 getNextToken(); // eat (
707 std::vector&lt;ExprAST*&gt; Args;
Chris Lattner71155212007-11-06 01:39:12 +0000708 if (CurTok != ')') {
709 while (1) {
710 ExprAST *Arg = ParseExpression();
711 if (!Arg) return 0;
712 Args.push_back(Arg);
Chris Lattner118749e2007-10-25 06:23:36 +0000713
Chris Lattner71155212007-11-06 01:39:12 +0000714 if (CurTok == ')') break;
Chris Lattner118749e2007-10-25 06:23:36 +0000715
Chris Lattner71155212007-11-06 01:39:12 +0000716 if (CurTok != ',')
Chris Lattner6c4be9c2008-04-14 16:44:41 +0000717 return Error("Expected ')' or ',' in argument list");
Chris Lattner71155212007-11-06 01:39:12 +0000718 getNextToken();
719 }
Chris Lattner118749e2007-10-25 06:23:36 +0000720 }
721
722 // Eat the ')'.
723 getNextToken();
724
725 return new CallExprAST(IdName, Args);
726}
727
728/// numberexpr ::= number
729static ExprAST *ParseNumberExpr() {
730 ExprAST *Result = new NumberExprAST(NumVal);
731 getNextToken(); // consume the number
732 return Result;
733}
734
735/// parenexpr ::= '(' expression ')'
736static ExprAST *ParseParenExpr() {
737 getNextToken(); // eat (.
738 ExprAST *V = ParseExpression();
739 if (!V) return 0;
740
741 if (CurTok != ')')
742 return Error("expected ')'");
743 getNextToken(); // eat ).
744 return V;
745}
746
747/// primary
748/// ::= identifierexpr
749/// ::= numberexpr
750/// ::= parenexpr
751static ExprAST *ParsePrimary() {
752 switch (CurTok) {
753 default: return Error("unknown token when expecting an expression");
754 case tok_identifier: return ParseIdentifierExpr();
755 case tok_number: return ParseNumberExpr();
756 case '(': return ParseParenExpr();
757 }
758}
759
760/// binoprhs
761/// ::= ('+' primary)*
762static ExprAST *ParseBinOpRHS(int ExprPrec, ExprAST *LHS) {
763 // If this is a binop, find its precedence.
764 while (1) {
765 int TokPrec = GetTokPrecedence();
766
767 // If this is a binop that binds at least as tightly as the current binop,
768 // consume it, otherwise we are done.
769 if (TokPrec &lt; ExprPrec)
770 return LHS;
771
772 // Okay, we know this is a binop.
773 int BinOp = CurTok;
774 getNextToken(); // eat binop
775
776 // Parse the primary expression after the binary operator.
777 ExprAST *RHS = ParsePrimary();
778 if (!RHS) return 0;
779
780 // If BinOp binds less tightly with RHS than the operator after RHS, let
781 // the pending operator take RHS as its LHS.
782 int NextPrec = GetTokPrecedence();
783 if (TokPrec &lt; NextPrec) {
784 RHS = ParseBinOpRHS(TokPrec+1, RHS);
785 if (RHS == 0) return 0;
786 }
787
788 // Merge LHS/RHS.
789 LHS = new BinaryExprAST(BinOp, LHS, RHS);
790 }
791}
792
793/// expression
794/// ::= primary binoprhs
795///
796static ExprAST *ParseExpression() {
797 ExprAST *LHS = ParsePrimary();
798 if (!LHS) return 0;
799
800 return ParseBinOpRHS(0, LHS);
801}
802
803/// prototype
804/// ::= id '(' id* ')'
805static PrototypeAST *ParsePrototype() {
806 if (CurTok != tok_identifier)
807 return ErrorP("Expected function name in prototype");
808
809 std::string FnName = IdentifierStr;
810 getNextToken();
811
812 if (CurTok != '(')
813 return ErrorP("Expected '(' in prototype");
814
815 std::vector&lt;std::string&gt; ArgNames;
816 while (getNextToken() == tok_identifier)
817 ArgNames.push_back(IdentifierStr);
818 if (CurTok != ')')
819 return ErrorP("Expected ')' in prototype");
820
821 // success.
822 getNextToken(); // eat ')'.
823
824 return new PrototypeAST(FnName, ArgNames);
825}
826
827/// definition ::= 'def' prototype expression
828static FunctionAST *ParseDefinition() {
829 getNextToken(); // eat def.
830 PrototypeAST *Proto = ParsePrototype();
831 if (Proto == 0) return 0;
832
833 if (ExprAST *E = ParseExpression())
834 return new FunctionAST(Proto, E);
835 return 0;
836}
837
838/// toplevelexpr ::= expression
839static FunctionAST *ParseTopLevelExpr() {
840 if (ExprAST *E = ParseExpression()) {
841 // Make an anonymous proto.
842 PrototypeAST *Proto = new PrototypeAST("", std::vector&lt;std::string&gt;());
843 return new FunctionAST(Proto, E);
844 }
845 return 0;
846}
847
848/// external ::= 'extern' prototype
849static PrototypeAST *ParseExtern() {
850 getNextToken(); // eat extern.
851 return ParsePrototype();
852}
853
854//===----------------------------------------------------------------------===//
855// Code Generation
856//===----------------------------------------------------------------------===//
857
858static Module *TheModule;
Duncan Sands89f6d882008-04-13 06:22:09 +0000859static IRBuilder Builder;
Chris Lattner118749e2007-10-25 06:23:36 +0000860static std::map&lt;std::string, Value*&gt; NamedValues;
861static FunctionPassManager *TheFPM;
862
863Value *ErrorV(const char *Str) { Error(Str); return 0; }
864
865Value *NumberExprAST::Codegen() {
866 return ConstantFP::get(Type::DoubleTy, APFloat(Val));
867}
868
869Value *VariableExprAST::Codegen() {
870 // Look this variable up in the function.
871 Value *V = NamedValues[Name];
872 return V ? V : ErrorV("Unknown variable name");
873}
874
875Value *BinaryExprAST::Codegen() {
876 Value *L = LHS-&gt;Codegen();
877 Value *R = RHS-&gt;Codegen();
878 if (L == 0 || R == 0) return 0;
879
880 switch (Op) {
881 case '+': return Builder.CreateAdd(L, R, "addtmp");
882 case '-': return Builder.CreateSub(L, R, "subtmp");
883 case '*': return Builder.CreateMul(L, R, "multmp");
884 case '&lt;':
Chris Lattner71155212007-11-06 01:39:12 +0000885 L = Builder.CreateFCmpULT(L, R, "cmptmp");
Chris Lattner118749e2007-10-25 06:23:36 +0000886 // Convert bool 0/1 to double 0.0 or 1.0
887 return Builder.CreateUIToFP(L, Type::DoubleTy, "booltmp");
888 default: return ErrorV("invalid binary operator");
889 }
890}
891
892Value *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
911Function *PrototypeAST::Codegen() {
912 // Make the function type: double(double,double) etc.
913 std::vector&lt;const Type*&gt; Doubles(Args.size(), Type::DoubleTy);
914 FunctionType *FT = FunctionType::get(Type::DoubleTy, Doubles, false);
915
916 Function *F = new Function(FT, Function::ExternalLinkage, Name, TheModule);
917
918 // If F conflicted, there was already something named 'Name'. If it has a
919 // body, don't allow redefinition or reextern.
920 if (F-&gt;getName() != Name) {
921 // Delete the one we just made and get the existing one.
922 F-&gt;eraseFromParent();
923 F = TheModule-&gt;getFunction(Name);
924
925 // If F already has a body, reject this.
926 if (!F-&gt;empty()) {
927 ErrorF("redefinition of function");
928 return 0;
929 }
930
931 // If F took a different number of args, reject.
932 if (F-&gt;arg_size() != Args.size()) {
933 ErrorF("redefinition of function with different # args");
934 return 0;
935 }
936 }
937
938 // Set names for all arguments.
939 unsigned Idx = 0;
940 for (Function::arg_iterator AI = F-&gt;arg_begin(); Idx != Args.size();
941 ++AI, ++Idx) {
942 AI-&gt;setName(Args[Idx]);
943
944 // Add arguments to variable symbol table.
945 NamedValues[Args[Idx]] = AI;
946 }
947
948 return F;
949}
950
951Function *FunctionAST::Codegen() {
952 NamedValues.clear();
953
954 Function *TheFunction = Proto-&gt;Codegen();
955 if (TheFunction == 0)
956 return 0;
957
958 // Create a new basic block to start insertion into.
959 BasicBlock *BB = new BasicBlock("entry", TheFunction);
960 Builder.SetInsertPoint(BB);
961
962 if (Value *RetVal = Body-&gt;Codegen()) {
963 // Finish off the function.
964 Builder.CreateRet(RetVal);
965
966 // Validate the generated code, checking for consistency.
967 verifyFunction(*TheFunction);
968
969 // Optimize the function.
970 TheFPM-&gt;run(*TheFunction);
971
972 return TheFunction;
973 }
974
975 // Error reading body, remove function.
976 TheFunction-&gt;eraseFromParent();
977 return 0;
978}
979
980//===----------------------------------------------------------------------===//
981// Top-Level parsing and JIT Driver
982//===----------------------------------------------------------------------===//
983
984static ExecutionEngine *TheExecutionEngine;
985
986static void HandleDefinition() {
987 if (FunctionAST *F = ParseDefinition()) {
988 if (Function *LF = F-&gt;Codegen()) {
989 fprintf(stderr, "Read function definition:");
990 LF-&gt;dump();
991 }
992 } else {
993 // Skip token for error recovery.
994 getNextToken();
995 }
996}
997
998static void HandleExtern() {
999 if (PrototypeAST *P = ParseExtern()) {
1000 if (Function *F = P-&gt;Codegen()) {
1001 fprintf(stderr, "Read extern: ");
1002 F-&gt;dump();
1003 }
1004 } else {
1005 // Skip token for error recovery.
1006 getNextToken();
1007 }
1008}
1009
1010static void HandleTopLevelExpression() {
1011 // Evaluate a top level expression into an anonymous function.
1012 if (FunctionAST *F = ParseTopLevelExpr()) {
1013 if (Function *LF = F-&gt;Codegen()) {
1014 // JIT the function, returning a function pointer.
1015 void *FPtr = TheExecutionEngine-&gt;getPointerToFunction(LF);
1016
1017 // Cast it to the right type (takes no arguments, returns a double) so we
1018 // can call it as a native function.
1019 double (*FP)() = (double (*)())FPtr;
1020 fprintf(stderr, "Evaluated to %f\n", FP());
1021 }
1022 } else {
1023 // Skip token for error recovery.
1024 getNextToken();
1025 }
1026}
1027
1028/// top ::= definition | external | expression | ';'
1029static void MainLoop() {
1030 while (1) {
1031 fprintf(stderr, "ready&gt; ");
1032 switch (CurTok) {
1033 case tok_eof: return;
1034 case ';': getNextToken(); break; // ignore top level semicolons.
1035 case tok_def: HandleDefinition(); break;
1036 case tok_extern: HandleExtern(); break;
1037 default: HandleTopLevelExpression(); break;
1038 }
1039 }
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.
1049extern "C"
1050double putchard(double X) {
1051 putchar((char)X);
1052 return 0;
1053}
1054
1055//===----------------------------------------------------------------------===//
1056// Main driver code.
1057//===----------------------------------------------------------------------===//
1058
1059int main() {
1060 // Install standard binary operators.
1061 // 1 is lowest precedence.
1062 BinopPrecedence['&lt;'] = 10;
1063 BinopPrecedence['+'] = 20;
1064 BinopPrecedence['-'] = 20;
1065 BinopPrecedence['*'] = 40; // highest.
1066
1067 // Prime the first token.
1068 fprintf(stderr, "ready&gt; ");
1069 getNextToken();
1070
1071 // Make the module, which holds all the code.
1072 TheModule = new Module("my cool jit");
1073
1074 // Create the JIT.
1075 TheExecutionEngine = ExecutionEngine::create(TheModule);
1076
1077 {
1078 ExistingModuleProvider OurModuleProvider(TheModule);
1079 FunctionPassManager OurFPM(&amp;OurModuleProvider);
1080
1081 // Set up the optimizer pipeline. Start with registering info about how the
1082 // target lays out data structures.
1083 OurFPM.add(new TargetData(*TheExecutionEngine-&gt;getTargetData()));
1084 // Do simple "peephole" optimizations and bit-twiddling optzns.
1085 OurFPM.add(createInstructionCombiningPass());
1086 // Reassociate expressions.
1087 OurFPM.add(createReassociatePass());
1088 // Eliminate Common SubExpressions.
1089 OurFPM.add(createGVNPass());
1090 // Simplify the control flow graph (deleting unreachable blocks, etc).
1091 OurFPM.add(createCFGSimplificationPass());
1092
1093 // Set the global so the code gen can use this.
1094 TheFPM = &amp;OurFPM;
1095
1096 // Run the main "interpreter loop" now.
1097 MainLoop();
1098
1099 TheFPM = 0;
Chris Lattner515686b2008-02-05 06:18:42 +00001100
1101 // Print out all of the generated code.
1102 TheModule-&gt;dump();
1103 } // Free module provider (and thus the module) and pass manager.
Chris Lattner118749e2007-10-25 06:23:36 +00001104
Chris Lattner118749e2007-10-25 06:23:36 +00001105 return 0;
1106}
Chris Lattnerc0b42e92007-10-23 06:27:55 +00001107</pre>
1108</div>
1109
Chris Lattner729eb142008-02-10 19:11:04 +00001110<a href="LangImpl5.html">Next: Extending the language: control flow</a>
Chris Lattnerc0b42e92007-10-23 06:27:55 +00001111</div>
1112
1113<!-- *********************************************************************** -->
1114<hr>
1115<address>
1116 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1117 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
1118 <a href="http://validator.w3.org/check/referer"><img
Chris Lattner8eef4b22007-10-23 06:30:50 +00001119 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!"></a>
Chris Lattnerc0b42e92007-10-23 06:27:55 +00001120
1121 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1122 <a href="http://llvm.org">The LLVM Compiler Infrastructure</a><br>
1123 Last modified: $Date: 2007-10-17 11:05:13 -0700 (Wed, 17 Oct 2007) $
1124</address>
1125</body>
1126</html>