Gitiles
Code Review Sign In
gerrit-public.fairphone.software / fp2-dev / platform / external / llvm / 128eb863dfb242aac982b505a9aab06dc35be8d1 / . / docs / tutorial / LangImpl3.html
blob: 2e961d2f610a2de050e84375ec1d2b2e204a9e7b [file] [log] [blame]
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]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: Implementing code generation to LLVM IR</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: Code generation to LLVM IR</div>
15
Chris Lattner128eb862007-11-05 19:06:59 +0000[diff] [blame^]16<ul>
17<li>Chapter 3
18 <ol>
19 <li><a href="#intro">Chapter 3 Introduction</a></li>
20 <li><a href="#basics">Code Generation setup</a></li>
21 <li><a href="#exprs">Expression Code Generation</a></li>
22 <li><a href="#funcs">Function Code Generation</a></li>
23 <li><a href="#driver">Driver Changes and Closing Thoughts</a></li>
24 <li><a href="#code">Full Code Listing</a></li>
25 </ol>
26</li>
27</ul>
28
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]29<div class="doc_author">
30 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a></p>
31</div>
32
33<!-- *********************************************************************** -->
Chris Lattner128eb862007-11-05 19:06:59 +0000[diff] [blame^]34<div class="doc_section"><a name="intro">Chapter 3 Introduction</a></div>
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]35<!-- *********************************************************************** -->
36
37<div class="doc_text">
38
Chris Lattner128eb862007-11-05 19:06:59 +0000[diff] [blame^]39<p>Welcome to Chapter 3 of the "<a href="index.html">Implementing a language
40with LLVM</a>" tutorial. This chapter shows you how to transform the <a
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]41href="LangImpl2.html">Abstract Syntax Tree built in Chapter 2</a> into LLVM IR.
42This will teach you a little bit about how LLVM does things, as well as
43demonstrate how easy it is to use. It's much more work to build a lexer and
44parser than it is to generate LLVM IR code.
45</p>
46
47</div>
48
49<!-- *********************************************************************** -->
50<div class="doc_section"><a name="basics">Code Generation setup</a></div>
51<!-- *********************************************************************** -->
52
53<div class="doc_text">
54
55<p>
56In order to generate LLVM IR, we want some simple setup to get started. First,
57we define virtual codegen methods in each AST class:</p>
58
59<div class="doc_code">
60<pre>
61/// ExprAST - Base class for all expression nodes.
62class ExprAST {
63public:
64 virtual ~ExprAST() {}
65 virtual Value *Codegen() = 0;
66};
67
68/// NumberExprAST - Expression class for numeric literals like "1.0".
69class NumberExprAST : public ExprAST {
70 double Val;
71public:
Chris Lattner28571ed2007-10-23 04:27:44 +0000[diff] [blame]72 explicit NumberExprAST(double val) : Val(val) {}
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]73 virtual Value *Codegen();
74};
75...
76</pre>
77</div>
78
Chris Lattner28571ed2007-10-23 04:27:44 +0000[diff] [blame]79<p>The Codegen() method says to emit IR for that AST node and all things it
80depends on, and they all return an LLVM Value object.
81"Value" is the class used to represent a "<a
82href="http://en.wikipedia.org/wiki/Static_single_assignment_form">Static Single
83Assignment (SSA)</a> register" or "SSA value" in LLVM. The most distinct aspect
84of SSA values is that their value is computed as the related instruction
85executes, and it does not get a new value until (and if) the instruction
86re-executes. In order words, there is no way to "change" an SSA value. For
87more information, please read up on <a
88href="http://en.wikipedia.org/wiki/Static_single_assignment_form">Static Single
89Assignment</a> - the concepts are really quite natural once you grok them.</p>
90
91<p>The
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]92second thing we want is an "Error" method like we used for parser, which will
93be used to report errors found during code generation (for example, use of an
94undeclared parameter):</p>
95
96<div class="doc_code">
97<pre>
98Value *ErrorV(const char *Str) { Error(Str); return 0; }
99
100static Module *TheModule;
101static LLVMBuilder Builder;
102static std::map&lt;std::string, Value*&gt; NamedValues;
103</pre>
104</div>
105
106<p>The static variables will be used during code generation. <tt>TheModule</tt>
107is the LLVM construct that contains all of the functions and global variables in
108a chunk of code. In many ways, it is the top-level structure that the LLVM IR
109uses to contain code.</p>
110
111<p>The <tt>Builder</tt> object is a helper object that makes it easy to generate
Chris Lattnerf6e53df2007-11-05 18:02:15 +0000[diff] [blame]112LLVM instructions. Instances of the <a
113href="http://llvm.org/doxygen/LLVMBuilder_8h-source.html"><tt>LLVMBuilder</tt>
114class</a> keeps track of the current place to
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]115insert instructions and has methods to create new instructions.</p>
116
117<p>The <tt>NamedValues</tt> map keeps track of which values are defined in the
118current scope and what their LLVM representation is. In this form of
119Kaleidoscope, the only things that can be referenced are function parameters.
120As such, function parameters will be in this map when generating code for their
121function body.</p>
122
123<p>
124With these basics in place, we can start talking about how to generate code for
125each expression. Note that this assumes that the <tt>Builder</tt> has been set
126up to generate code <em>into</em> something. For now, we'll assume that this
127has already been done, and we'll just use it to emit code.
128</p>
129
130</div>
131
132<!-- *********************************************************************** -->
133<div class="doc_section"><a name="exprs">Expression Code Generation</a></div>
134<!-- *********************************************************************** -->
135
136<div class="doc_text">
137
138<p>Generating LLVM code for expression nodes is very straight-forward: less
139than 45 lines of commented code for all four of our expression nodes. First,
140we'll do numeric literals:</p>
141
142<div class="doc_code">
143<pre>
144Value *NumberExprAST::Codegen() {
145 return ConstantFP::get(Type::DoubleTy, APFloat(Val));
146}
147</pre>
148</div>
149
Chris Lattnerd3f0cdd2007-10-23 04:51:30 +0000[diff] [blame]150<p>In the LLVM IR, numeric constants are represented with the
151<tt>ConstantFP</tt> class, which holds the numeric value in an <tt>APFloat</tt>
152internally (<tt>APFloat</tt> has the capability of holding floating point
153constants of <em>A</em>rbitrary <em>P</em>recision). This code basically just
154creates and returns a <tt>ConstantFP</tt>. Note that in the LLVM IR
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]155that constants are all uniqued together and shared. For this reason, the API
Chris Lattnerd3f0cdd2007-10-23 04:51:30 +0000[diff] [blame]156uses "the foo::get(..)" idiom instead of "new foo(..)" or "foo::create(..).</p>
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]157
158<div class="doc_code">
159<pre>
160Value *VariableExprAST::Codegen() {
161 // Look this variable up in the function.
162 Value *V = NamedValues[Name];
163 return V ? V : ErrorV("Unknown variable name");
164}
165</pre>
166</div>
167
Chris Lattnerd3f0cdd2007-10-23 04:51:30 +0000[diff] [blame]168<p>References to variables is also quite simple here. In the simple version
169of Kaleidoscope, we assume that the variable has already been emited somewhere
170and its value is available. In practice, the only values that can be in the
171<tt>NamedValues</tt> map are function arguments. This
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]172code simply checks to see that the specified name is in the map (if not, an
173unknown variable is being referenced) and returns the value for it.</p>
174
175<div class="doc_code">
176<pre>
177Value *BinaryExprAST::Codegen() {
178 Value *L = LHS-&gt;Codegen();
179 Value *R = RHS-&gt;Codegen();
180 if (L == 0 || R == 0) return 0;
181
182 switch (Op) {
183 case '+': return Builder.CreateAdd(L, R, "addtmp");
184 case '-': return Builder.CreateSub(L, R, "subtmp");
185 case '*': return Builder.CreateMul(L, R, "multmp");
186 case '&lt;':
187 L = Builder.CreateFCmpULT(L, R, "multmp");
188 // Convert bool 0/1 to double 0.0 or 1.0
189 return Builder.CreateUIToFP(L, Type::DoubleTy, "booltmp");
190 default: return ErrorV("invalid binary operator");
191 }
192}
193</pre>
194</div>
195
Chris Lattnerd3f0cdd2007-10-23 04:51:30 +0000[diff] [blame]196<p>Binary operators start to get more interesting. The basic idea here is that
197we recursively emit code for the left-hand side of the expression, then the
198right-hand side, then we compute the result of the binary expression. In this
199code, we do a simple switch on the opcode to create the right LLVM instruction.
200</p>
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]201
Chris Lattnerd3f0cdd2007-10-23 04:51:30 +0000[diff] [blame]202<p>In this example, the LLVM builder class is starting to show its value.
203Because it knows where to insert the newly created instruction, you just have to
204specificy what instruction to create (e.g. with <tt>CreateAdd</tt>), which
205operands to use (<tt>L</tt> and <tt>R</tt> here) and optionally provide a name
206for the generated instruction. One nice thing about LLVM is that the name is
207just a hint: if there are multiple additions in a single function, the first
208will be named "addtmp" and the second will be "autorenamed" by adding a suffix,
209giving it a name like "addtmp42". Local value names for instructions are purely
210optional, but it makes it much easier to read the IR dumps.</p>
211
212<p><a href="../LangRef.html#instref">LLVM instructions</a> are constrained to
213have very strict type properties: for example, the Left and Right operators of
214an <a href="../LangRef.html#i_add">add instruction</a> have to have the same
215type, and that the result of the add matches the operands. Because all values
216in Kaleidoscope are doubles, this makes for very simple code for add, sub and
217mul.</p>
218
219<p>On the other hand, LLVM specifies that the <a
220href="../LangRef.html#i_fcmp">fcmp instruction</a> always returns an 'i1' value
221(a one bit integer). However, Kaleidoscope wants the value to be a 0.0 or 1.0
222value. In order to get these semantics, we combine the fcmp instruction with
223a <a href="../LangRef.html#i_uitofp">uitofp instruction</a>. This instruction
224converts its input integer into a floating point value by treating the input
225as an unsigned value. In contrast, if we used the <a
226href="../LangRef.html#i_sitofp">sitofp instruction</a>, the Kaleidoscope '<'
227operator would return 0.0 and -1.0, depending on the input value.</p>
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]228
229<div class="doc_code">
230<pre>
231Value *CallExprAST::Codegen() {
232 // Look up the name in the global module table.
233 Function *CalleeF = TheModule-&gt;getFunction(Callee);
234 if (CalleeF == 0)
235 return ErrorV("Unknown function referenced");
236
237 // If argument mismatch error.
238 if (CalleeF-&gt;arg_size() != Args.size())
239 return ErrorV("Incorrect # arguments passed");
240
241 std::vector&lt;Value*&gt; ArgsV;
242 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
243 ArgsV.push_back(Args[i]-&gt;Codegen());
244 if (ArgsV.back() == 0) return 0;
245 }
246
247 return Builder.CreateCall(CalleeF, ArgsV.begin(), ArgsV.end(), "calltmp");
248}
249</pre>
250</div>
251
Chris Lattnerd3f0cdd2007-10-23 04:51:30 +0000[diff] [blame]252<p>Code generation for function calls is quite straight-forward with LLVM. The
253code above first looks the name of the function up in the LLVM Module's symbol
254table. Recall that the LLVM Module is the container that holds all of the
255functions we are JIT'ing. By giving each function the same name as what the
256user specifies, we can use the LLVM symbol table to resolve function names for
257us.</p>
258
259<p>Once we have the function to call, we recursively codegen each argument that
260is to be passed in, and create an LLVM <a href="../LangRef.html#i_call">call
261instruction</a>. Note that LLVM uses the native C calling conventions by
262default, allowing these calls to call into standard library functions like
263"sin" and "cos" with no additional effort.</p>
264
265<p>This wraps up our handling of the four basic expressions that we have so far
266in Kaleidoscope. Feel free to go in and add some more. For example, by
267browsing the <a href="../LangRef.html">LLVM language reference</a> you'll find
268several other interesting instructions that are really easy to plug into our
269basic framework.</p>
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]270
271</div>
272
273<!-- *********************************************************************** -->
Chris Lattner35abbf52007-10-23 06:23:57 +0000[diff] [blame]274<div class="doc_section"><a name="funcs">Function Code Generation</a></div>
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]275<!-- *********************************************************************** -->
276
277<div class="doc_text">
278
Chris Lattner35abbf52007-10-23 06:23:57 +0000[diff] [blame]279<p>Code generation for prototypes and functions has to handle a number of
280details, which make their code less beautiful and elegant than expression code
281generation, but they illustrate some important points. First, lets talk about
282code generation for prototypes: this is used both for function bodies as well
283as external function declarations. The code starts with:</p>
284
285<div class="doc_code">
286<pre>
287Function *PrototypeAST::Codegen() {
288 // Make the function type: double(double,double) etc.
289 std::vector&lt;const Type*&gt; Doubles(Args.size(), Type::DoubleTy);
290 FunctionType *FT = FunctionType::get(Type::DoubleTy, Doubles, false);
291
292 Function *F = new Function(FT, Function::ExternalLinkage, Name, TheModule);
293</pre>
294</div>
295
296<p>This code packs a lot of power into a few lines. The first step is to create
297the <tt>FunctionType</tt> that should be used for a given Prototype. Since all
298function arguments in Kaleidoscope are of type double, the first line creates
299a vector of "N" LLVM Double types. It then uses the <tt>FunctionType::get</tt>
300method to create a function type that takes "N" doubles as arguments, returns
301one double as a result, and that is not vararg (the false parameter indicates
302this). Note that Types in LLVM are uniqued just like Constants are, so you
303don't "new" a type, you "get" it.</p>
304
305<p>The final line above actually creates the function that the prototype will
306correspond to. This indicates which type, linkage, and name to use, and which
307module to insert into. "<a href="LangRef.html#linkage">external linkage</a>"
308means that the function may be defined outside the current module and/or that it
309is callable by functions outside the module. The Name passed in is the name the
310user specified: since "<tt>TheModule</tt>" is specified, this name is registered
311in "<tt>TheModule</tt>"s symbol table, which is used by the function call code
312above.</p>
313
314<div class="doc_code">
315<pre>
316 // If F conflicted, there was already something named 'Name'. If it has a
317 // body, don't allow redefinition or reextern.
318 if (F-&gt;getName() != Name) {
319 // Delete the one we just made and get the existing one.
320 F-&gt;eraseFromParent();
321 F = TheModule-&gt;getFunction(Name);
322</pre>
323</div>
324
325<p>The Module symbol table works just like the Function symbol table when it
326comes to name conflicts: if a new function is created with a name was previously
327added to the symbol table, it will get implicitly renamed when added to the
328Module. The code above exploits this fact to tell if there was a previous
329definition of this function.</p>
330
331<p>In Kaleidoscope, I choose to allow redefinitions of functions in two cases:
332first, we want to allow 'extern'ing a function more than once, so long as the
333prototypes for the externs match (since all arguments have the same type, we
334just have to check that the number of arguments match). Second, we want to
335allow 'extern'ing a function and then definining a body for it. This is useful
336when defining mutually recursive functions.</p>
337
338<p>In order to implement this, the code above first checks to see if there is
339a collision on the name of the function. If so, it deletes the function we just
340created (by calling <tt>eraseFromParent</tt>) and then calling
341<tt>getFunction</tt> to get the existing function with the specified name. Note
342that many APIs in LLVM have "erase" forms and "remove" forms. The "remove" form
343unlinks the object from its parent (e.g. a Function from a Module) and returns
344it. The "erase" form unlinks the object and then deletes it.</p>
345
346<div class="doc_code">
347<pre>
348 // If F already has a body, reject this.
349 if (!F-&gt;empty()) {
350 ErrorF("redefinition of function");
351 return 0;
352 }
353
354 // If F took a different number of args, reject.
355 if (F-&gt;arg_size() != Args.size()) {
356 ErrorF("redefinition of function with different # args");
357 return 0;
358 }
359 }
360</pre>
361</div>
362
363<p>In order to verify the logic above, we first check to see if the preexisting
364function is "empty". In this case, empty means that it has no basic blocks in
365it, which means it has no body. If it has no body, this means its a forward
366declaration. Since we don't allow anything after a full definition of the
367function, the code rejects this case. If the previous reference to a function
368was an 'extern', we simply verify that the number of arguments for that
369definition and this one match up. If not, we emit an error.</p>
370
371<div class="doc_code">
372<pre>
373 // Set names for all arguments.
374 unsigned Idx = 0;
375 for (Function::arg_iterator AI = F-&gt;arg_begin(); Idx != Args.size();
376 ++AI, ++Idx) {
377 AI-&gt;setName(Args[Idx]);
378
379 // Add arguments to variable symbol table.
380 NamedValues[Args[Idx]] = AI;
381 }
382 return F;
383}
384</pre>
385</div>
386
387<p>The last bit of code for prototypes loops over all of the arguments in the
388function, setting the name of the LLVM Argument objects to match and registering
389the arguments in the <tt>NamedValues</tt> map for future use by the
390<tt>VariableExprAST</tt> AST node. Once this is set up, it returns the Function
391object to the caller. Note that we don't check for conflicting
392argument names here (e.g. "extern foo(a b a)"). Doing so would be very
393straight-forward.</p>
394
395<div class="doc_code">
396<pre>
397Function *FunctionAST::Codegen() {
398 NamedValues.clear();
399
400 Function *TheFunction = Proto-&gt;Codegen();
401 if (TheFunction == 0)
402 return 0;
403</pre>
404</div>
405
406<p>Code generation for function definitions starts out simply enough: first we
407codegen the prototype and verify that it is ok. We also clear out the
408<tt>NamedValues</tt> map to make sure that there isn't anything in it from the
409last function we compiled.</p>
410
411<div class="doc_code">
412<pre>
413 // Create a new basic block to start insertion into.
414 BasicBlock *BB = new BasicBlock("entry", TheFunction);
415 Builder.SetInsertPoint(BB);
416
417 if (Value *RetVal = Body-&gt;Codegen()) {
Chris Lattner35abbf52007-10-23 06:23:57 +0000[diff] [blame]418</pre>
419</div>
420
421<p>Now we get to the point where the <tt>Builder</tt> is set up. The first
422line creates a new <a href="http://en.wikipedia.org/wiki/Basic_block">basic
423block</a> (named "entry"), which is inserted into <tt>TheFunction</tt>. The
424second line then tells the builder that new instructions should be inserted into
425the end of the new basic block. Basic blocks in LLVM are an important part
426of functions that define the <a
427href="http://en.wikipedia.org/wiki/Control_flow_graph">Control Flow Graph</a>.
428Since we don't have any control flow, our functions will only contain one
429block so far. We'll fix this in a future installment :).</p>
430
Chris Lattnerd9b86162007-10-25 04:30:35 +0000[diff] [blame]431<div class="doc_code">
432<pre>
433 if (Value *RetVal = Body-&gt;Codegen()) {
434 // Finish off the function.
435 Builder.CreateRet(RetVal);
436
437 // Validate the generated code, checking for consistency.
438 verifyFunction(*TheFunction);
439 return TheFunction;
440 }
441</pre>
442</div>
443
Chris Lattner35abbf52007-10-23 06:23:57 +0000[diff] [blame]444<p>Once the insertion point is set up, we call the <tt>CodeGen()</tt> method for
445the root expression of the function. If no error happens, this emits code to
446compute the expression into the entry block and returns the value that was
447computed. Assuming no error, we then create an LLVM <a
Chris Lattnerd9b86162007-10-25 04:30:35 +0000[diff] [blame]448href="../LangRef.html#i_ret">ret instruction</a>, which completes the function.
449Once the function is built, we call the <tt>verifyFunction</tt> function, which
450is provided by LLVM. This function does a variety of consistency checks on the
451generated code, to determine if our compiler is doing everything right. Using
452this is important: it can catch a lot of bugs. Once the function is finished
453and validated, we return it.</p>
Chris Lattner35abbf52007-10-23 06:23:57 +0000[diff] [blame]454
455<div class="doc_code">
456<pre>
457 // Error reading body, remove function.
458 TheFunction-&gt;eraseFromParent();
459 return 0;
460}
461</pre>
462</div>
463
464<p>The only piece left here is handling of the error case. For simplicity, we
465simply handle this by deleting the function we produced with the
466<tt>eraseFromParent</tt> method. This allows the user to redefine a function
467that they incorrectly typed in before: if we didn't delete it, it would live in
468the symbol table, with a body, preventing future redefinition.</p>
469
470<p>This code does have a bug though. Since the <tt>PrototypeAST::Codegen</tt>
471can return a previously defined forward declaration, this can actually delete
472a forward declaration. There are a number of ways to fix this bug, see what you
473can come up with! Here is a testcase:</p>
474
475<div class="doc_code">
476<pre>
477extern foo(a b); # ok, defines foo.
478def foo(a b) c; # error, 'c' is invalid.
479def bar() foo(1, 2); # error, unknown function "foo"
480</pre>
481</div>
482
483</div>
484
485<!-- *********************************************************************** -->
486<div class="doc_section"><a name="driver">Driver Changes and
487Closing Thoughts</a></div>
488<!-- *********************************************************************** -->
489
490<div class="doc_text">
491
492<p>
493For now, code generation to LLVM doesn't really get us much, except that we can
494look at the pretty IR calls. The sample code inserts calls to Codegen into the
495"<tt>HandleDefinition</tt>", "<tt>HandleExtern</tt>" etc functions, and then
496dumps out the LLVM IR. This gives a nice way to look at the LLVM IR for simple
497functions. For example:
498</p>
499
500<div class="doc_code">
501<pre>
502ready> <b>4+5</b>;
503ready> Read top-level expression:
504define double @""() {
505entry:
506 %addtmp = add double 4.000000e+00, 5.000000e+00
507 ret double %addtmp
508}
509</pre>
510</div>
511
512<p>Note how the parser turns the top-level expression into anonymous functions
513for us. This will be handy when we add JIT support in the next chapter. Also
514note that the code is very literally transcribed, no optimizations are being
515performed. We will add optimizations explicitly in the next chapter.</p>
516
517<div class="doc_code">
518<pre>
519ready&gt; <b>def foo(a b) a*a + 2*a*b + b*b;</b>
520ready&gt; Read function definition:
521define double @foo(double %a, double %b) {
522entry:
523 %multmp = mul double %a, %a
524 %multmp1 = mul double 2.000000e+00, %a
525 %multmp2 = mul double %multmp1, %b
526 %addtmp = add double %multmp, %multmp2
527 %multmp3 = mul double %b, %b
528 %addtmp4 = add double %addtmp, %multmp3
529 ret double %addtmp4
530}
531</pre>
532</div>
533
534<p>This shows some simple arithmetic. Notice the striking similarity to the
535LLVM builder calls that we use to create the instructions.</p>
536
537<div class="doc_code">
538<pre>
539ready&gt; <b>def bar(a) foo(a, 4.0) + bar(31337);</b>
540ready&gt; Read function definition:
541define double @bar(double %a) {
542entry:
543 %calltmp = call double @foo( double %a, double 4.000000e+00 )
544 %calltmp1 = call double @bar( double 3.133700e+04 )
545 %addtmp = add double %calltmp, %calltmp1
546 ret double %addtmp
547}
548</pre>
549</div>
550
Chris Lattnerfc60ab02007-11-05 17:39:26 +0000[diff] [blame]551<p>This shows some function calls. Note that this function will take a long
552time to execute if you call it. In the future we'll add conditional control
553flow to make recursion actually be useful :).</p>
Chris Lattner35abbf52007-10-23 06:23:57 +0000[diff] [blame]554
555<div class="doc_code">
556<pre>
557ready&gt; <b>extern cos(x);</b>
558ready&gt; Read extern:
559declare double @cos(double)
560
561ready&gt; <b>cos(1.234);</b>
562ready&gt; Read top-level expression:
563define double @""() {
564entry:
Chris Lattner8eef4b22007-10-23 06:30:50 +0000[diff] [blame]565 %calltmp = call double @cos( double 1.234000e+00 )
Chris Lattner35abbf52007-10-23 06:23:57 +0000[diff] [blame]566 ret double %calltmp
567}
568</pre>
569</div>
570
571<p>This shows an extern for the libm "cos" function, and a call to it.</p>
572
573
574<div class="doc_code">
575<pre>
576ready&gt; <b>^D</b>
577; ModuleID = 'my cool jit'
578
579define double @""() {
580entry:
581 %addtmp = add double 4.000000e+00, 5.000000e+00
582 ret double %addtmp
583}
584
585define double @foo(double %a, double %b) {
586entry:
587 %multmp = mul double %a, %a
588 %multmp1 = mul double 2.000000e+00, %a
589 %multmp2 = mul double %multmp1, %b
590 %addtmp = add double %multmp, %multmp2
591 %multmp3 = mul double %b, %b
592 %addtmp4 = add double %addtmp, %multmp3
593 ret double %addtmp4
594}
595
596define double @bar(double %a) {
597entry:
598 %calltmp = call double @foo( double %a, double 4.000000e+00 )
599 %calltmp1 = call double @bar( double 3.133700e+04 )
600 %addtmp = add double %calltmp, %calltmp1
601 ret double %addtmp
602}
603
604declare double @cos(double)
605
606define double @""() {
607entry:
608 %calltmp = call double @cos( double 1.234000e+00 )
609 ret double %calltmp
610}
611</pre>
612</div>
613
614<p>When you quit the current demo, it dumps out the IR for the entire module
615generated. Here you can see the big picture with all the functions referencing
616each other.</p>
617
618<p>This wraps up this chapter of the Kaleidoscope tutorial. Up next we'll
619describe how to <a href="LangImpl4.html">add JIT codegen and optimizer
620support</a> to this so we can actually start running code!</p>
621
622</div>
623
624
625<!-- *********************************************************************** -->
626<div class="doc_section"><a name="code">Full Code Listing</a></div>
627<!-- *********************************************************************** -->
628
629<div class="doc_text">
630
631<p>
632Here is the complete code listing for our running example, enhanced with the
633LLVM code generator. Because this uses the LLVM libraries, we need to link
634them in. To do this, we use the <a
635href="http://llvm.org/cmds/llvm-config.html">llvm-config</a> tool to inform
636our makefile/command line about which options to use:</p>
637
638<div class="doc_code">
639<pre>
640 # Compile
Chris Lattner9ac0ca02007-10-24 05:09:48 +0000[diff] [blame]641 g++ -g toy.cpp `llvm-config --cppflags --ldflags --libs core` -o toy
Chris Lattner35abbf52007-10-23 06:23:57 +0000[diff] [blame]642 # Run
643 ./toy
644</pre>
645</div>
646
647<p>Here is the code:</p>
648
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]649<div class="doc_code">
650<pre>
651// To build this:
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]652// See example below.
653
654#include "llvm/DerivedTypes.h"
655#include "llvm/Module.h"
Chris Lattnerd9b86162007-10-25 04:30:35 +0000[diff] [blame]656#include "llvm/Analysis/Verifier.h"
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]657#include "llvm/Support/LLVMBuilder.h"
658#include &lt;cstdio&gt;
659#include &lt;string&gt;
660#include &lt;map&gt;
661#include &lt;vector&gt;
662using namespace llvm;
663
664//===----------------------------------------------------------------------===//
665// Lexer
666//===----------------------------------------------------------------------===//
667
668// The lexer returns tokens [0-255] if it is an unknown character, otherwise one
669// of these for known things.
670enum Token {
671 tok_eof = -1,
672
673 // commands
674 tok_def = -2, tok_extern = -3,
675
676 // primary
677 tok_identifier = -4, tok_number = -5,
678};
679
680static std::string IdentifierStr; // Filled in if tok_identifier
681static double NumVal; // Filled in if tok_number
682
683/// gettok - Return the next token from standard input.
684static int gettok() {
685 static int LastChar = ' ';
686
687 // Skip any whitespace.
688 while (isspace(LastChar))
689 LastChar = getchar();
690
691 if (isalpha(LastChar)) { // identifier: [a-zA-Z][a-zA-Z0-9]*
692 IdentifierStr = LastChar;
693 while (isalnum((LastChar = getchar())))
694 IdentifierStr += LastChar;
695
696 if (IdentifierStr == "def") return tok_def;
697 if (IdentifierStr == "extern") return tok_extern;
698 return tok_identifier;
699 }
700
701 if (isdigit(LastChar) || LastChar == '.') { // Number: [0-9.]+
702 std::string NumStr;
703 do {
704 NumStr += LastChar;
705 LastChar = getchar();
706 } while (isdigit(LastChar) || LastChar == '.');
707
708 NumVal = strtod(NumStr.c_str(), 0);
709 return tok_number;
710 }
711
712 if (LastChar == '#') {
713 // Comment until end of line.
714 do LastChar = getchar();
715 while (LastChar != EOF &amp;&amp; LastChar != '\n' &amp; LastChar != '\r');
716
717 if (LastChar != EOF)
718 return gettok();
719 }
720
721 // Check for end of file. Don't eat the EOF.
722 if (LastChar == EOF)
723 return tok_eof;
724
725 // Otherwise, just return the character as its ascii value.
726 int ThisChar = LastChar;
727 LastChar = getchar();
728 return ThisChar;
729}
730
731//===----------------------------------------------------------------------===//
732// Abstract Syntax Tree (aka Parse Tree)
733//===----------------------------------------------------------------------===//
734
735/// ExprAST - Base class for all expression nodes.
736class ExprAST {
737public:
738 virtual ~ExprAST() {}
739 virtual Value *Codegen() = 0;
740};
741
742/// NumberExprAST - Expression class for numeric literals like "1.0".
743class NumberExprAST : public ExprAST {
744 double Val;
745public:
Chris Lattner28571ed2007-10-23 04:27:44 +0000[diff] [blame]746 explicit NumberExprAST(double val) : Val(val) {}
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]747 virtual Value *Codegen();
748};
749
750/// VariableExprAST - Expression class for referencing a variable, like "a".
751class VariableExprAST : public ExprAST {
752 std::string Name;
753public:
Chris Lattner28571ed2007-10-23 04:27:44 +0000[diff] [blame]754 explicit VariableExprAST(const std::string &amp;name) : Name(name) {}
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]755 virtual Value *Codegen();
756};
757
758/// BinaryExprAST - Expression class for a binary operator.
759class BinaryExprAST : public ExprAST {
760 char Op;
761 ExprAST *LHS, *RHS;
762public:
763 BinaryExprAST(char op, ExprAST *lhs, ExprAST *rhs)
764 : Op(op), LHS(lhs), RHS(rhs) {}
765 virtual Value *Codegen();
766};
767
768/// CallExprAST - Expression class for function calls.
769class CallExprAST : public ExprAST {
770 std::string Callee;
771 std::vector&lt;ExprAST*&gt; Args;
772public:
773 CallExprAST(const std::string &amp;callee, std::vector&lt;ExprAST*&gt; &amp;args)
774 : Callee(callee), Args(args) {}
775 virtual Value *Codegen();
776};
777
778/// PrototypeAST - This class represents the "prototype" for a function,
779/// which captures its argument names as well as if it is an operator.
780class PrototypeAST {
781 std::string Name;
782 std::vector&lt;std::string&gt; Args;
783public:
784 PrototypeAST(const std::string &amp;name, const std::vector&lt;std::string&gt; &amp;args)
785 : Name(name), Args(args) {}
786
787 Function *Codegen();
788};
789
790/// FunctionAST - This class represents a function definition itself.
791class FunctionAST {
792 PrototypeAST *Proto;
793 ExprAST *Body;
794public:
795 FunctionAST(PrototypeAST *proto, ExprAST *body)
796 : Proto(proto), Body(body) {}
797
798 Function *Codegen();
799};
800
801//===----------------------------------------------------------------------===//
802// Parser
803//===----------------------------------------------------------------------===//
804
805/// CurTok/getNextToken - Provide a simple token buffer. CurTok is the current
806/// token the parser it looking at. getNextToken reads another token from the
807/// lexer and updates CurTok with its results.
808static int CurTok;
809static int getNextToken() {
810 return CurTok = gettok();
811}
812
813/// BinopPrecedence - This holds the precedence for each binary operator that is
814/// defined.
815static std::map&lt;char, int&gt; BinopPrecedence;
816
817/// GetTokPrecedence - Get the precedence of the pending binary operator token.
818static int GetTokPrecedence() {
819 if (!isascii(CurTok))
820 return -1;
821
822 // Make sure it's a declared binop.
823 int TokPrec = BinopPrecedence[CurTok];
824 if (TokPrec &lt;= 0) return -1;
825 return TokPrec;
826}
827
828/// Error* - These are little helper functions for error handling.
829ExprAST *Error(const char *Str) { fprintf(stderr, "Error: %s\n", Str);return 0;}
830PrototypeAST *ErrorP(const char *Str) { Error(Str); return 0; }
831FunctionAST *ErrorF(const char *Str) { Error(Str); return 0; }
832
833static ExprAST *ParseExpression();
834
835/// identifierexpr
Chris Lattner20a0c802007-11-05 17:54:34 +0000[diff] [blame]836/// ::= identifier
837/// ::= identifier '(' expression* ')'
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]838static ExprAST *ParseIdentifierExpr() {
839 std::string IdName = IdentifierStr;
840
Chris Lattner20a0c802007-11-05 17:54:34 +0000[diff] [blame]841 getNextToken(); // eat identifier.
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]842
843 if (CurTok != '(') // Simple variable ref.
844 return new VariableExprAST(IdName);
845
846 // Call.
847 getNextToken(); // eat (
848 std::vector&lt;ExprAST*&gt; Args;
849 while (1) {
850 ExprAST *Arg = ParseExpression();
851 if (!Arg) return 0;
852 Args.push_back(Arg);
853
854 if (CurTok == ')') break;
855
856 if (CurTok != ',')
857 return Error("Expected ')'");
858 getNextToken();
859 }
860
861 // Eat the ')'.
862 getNextToken();
863
864 return new CallExprAST(IdName, Args);
865}
866
867/// numberexpr ::= number
868static ExprAST *ParseNumberExpr() {
869 ExprAST *Result = new NumberExprAST(NumVal);
870 getNextToken(); // consume the number
871 return Result;
872}
873
874/// parenexpr ::= '(' expression ')'
875static ExprAST *ParseParenExpr() {
876 getNextToken(); // eat (.
877 ExprAST *V = ParseExpression();
878 if (!V) return 0;
879
880 if (CurTok != ')')
881 return Error("expected ')'");
882 getNextToken(); // eat ).
883 return V;
884}
885
886/// primary
887/// ::= identifierexpr
888/// ::= numberexpr
889/// ::= parenexpr
890static ExprAST *ParsePrimary() {
891 switch (CurTok) {
892 default: return Error("unknown token when expecting an expression");
893 case tok_identifier: return ParseIdentifierExpr();
894 case tok_number: return ParseNumberExpr();
895 case '(': return ParseParenExpr();
896 }
897}
898
899/// binoprhs
900/// ::= ('+' primary)*
901static ExprAST *ParseBinOpRHS(int ExprPrec, ExprAST *LHS) {
902 // If this is a binop, find its precedence.
903 while (1) {
904 int TokPrec = GetTokPrecedence();
905
906 // If this is a binop that binds at least as tightly as the current binop,
907 // consume it, otherwise we are done.
908 if (TokPrec &lt; ExprPrec)
909 return LHS;
910
911 // Okay, we know this is a binop.
912 int BinOp = CurTok;
913 getNextToken(); // eat binop
914
915 // Parse the primary expression after the binary operator.
916 ExprAST *RHS = ParsePrimary();
917 if (!RHS) return 0;
918
919 // If BinOp binds less tightly with RHS than the operator after RHS, let
920 // the pending operator take RHS as its LHS.
921 int NextPrec = GetTokPrecedence();
922 if (TokPrec &lt; NextPrec) {
923 RHS = ParseBinOpRHS(TokPrec+1, RHS);
924 if (RHS == 0) return 0;
925 }
926
927 // Merge LHS/RHS.
928 LHS = new BinaryExprAST(BinOp, LHS, RHS);
929 }
930}
931
932/// expression
933/// ::= primary binoprhs
934///
935static ExprAST *ParseExpression() {
936 ExprAST *LHS = ParsePrimary();
937 if (!LHS) return 0;
938
939 return ParseBinOpRHS(0, LHS);
940}
941
942/// prototype
943/// ::= id '(' id* ')'
944static PrototypeAST *ParsePrototype() {
945 if (CurTok != tok_identifier)
946 return ErrorP("Expected function name in prototype");
947
948 std::string FnName = IdentifierStr;
949 getNextToken();
950
951 if (CurTok != '(')
952 return ErrorP("Expected '(' in prototype");
953
954 std::vector&lt;std::string&gt; ArgNames;
955 while (getNextToken() == tok_identifier)
956 ArgNames.push_back(IdentifierStr);
957 if (CurTok != ')')
958 return ErrorP("Expected ')' in prototype");
959
960 // success.
961 getNextToken(); // eat ')'.
962
963 return new PrototypeAST(FnName, ArgNames);
964}
965
966/// definition ::= 'def' prototype expression
967static FunctionAST *ParseDefinition() {
968 getNextToken(); // eat def.
969 PrototypeAST *Proto = ParsePrototype();
970 if (Proto == 0) return 0;
971
972 if (ExprAST *E = ParseExpression())
973 return new FunctionAST(Proto, E);
974 return 0;
975}
976
977/// toplevelexpr ::= expression
978static FunctionAST *ParseTopLevelExpr() {
979 if (ExprAST *E = ParseExpression()) {
980 // Make an anonymous proto.
981 PrototypeAST *Proto = new PrototypeAST("", std::vector&lt;std::string&gt;());
982 return new FunctionAST(Proto, E);
983 }
984 return 0;
985}
986
987/// external ::= 'extern' prototype
988static PrototypeAST *ParseExtern() {
989 getNextToken(); // eat extern.
990 return ParsePrototype();
991}
992
993//===----------------------------------------------------------------------===//
994// Code Generation
995//===----------------------------------------------------------------------===//
996
997static Module *TheModule;
998static LLVMBuilder Builder;
999static std::map&lt;std::string, Value*&gt; NamedValues;
1000
1001Value *ErrorV(const char *Str) { Error(Str); return 0; }
1002
1003Value *NumberExprAST::Codegen() {
1004 return ConstantFP::get(Type::DoubleTy, APFloat(Val));
1005}
1006
1007Value *VariableExprAST::Codegen() {
1008 // Look this variable up in the function.
1009 Value *V = NamedValues[Name];
1010 return V ? V : ErrorV("Unknown variable name");
1011}
1012
1013Value *BinaryExprAST::Codegen() {
1014 Value *L = LHS-&gt;Codegen();
1015 Value *R = RHS-&gt;Codegen();
1016 if (L == 0 || R == 0) return 0;
1017
1018 switch (Op) {
1019 case '+': return Builder.CreateAdd(L, R, "addtmp");
1020 case '-': return Builder.CreateSub(L, R, "subtmp");
1021 case '*': return Builder.CreateMul(L, R, "multmp");
1022 case '&lt;':
1023 L = Builder.CreateFCmpULT(L, R, "multmp");
1024 // Convert bool 0/1 to double 0.0 or 1.0
1025 return Builder.CreateUIToFP(L, Type::DoubleTy, "booltmp");
1026 default: return ErrorV("invalid binary operator");
1027 }
1028}
1029
1030Value *CallExprAST::Codegen() {
1031 // Look up the name in the global module table.
1032 Function *CalleeF = TheModule-&gt;getFunction(Callee);
1033 if (CalleeF == 0)
1034 return ErrorV("Unknown function referenced");
1035
1036 // If argument mismatch error.
1037 if (CalleeF-&gt;arg_size() != Args.size())
1038 return ErrorV("Incorrect # arguments passed");
1039
1040 std::vector&lt;Value*&gt; ArgsV;
1041 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
1042 ArgsV.push_back(Args[i]-&gt;Codegen());
1043 if (ArgsV.back() == 0) return 0;
1044 }
1045
1046 return Builder.CreateCall(CalleeF, ArgsV.begin(), ArgsV.end(), "calltmp");
1047}
1048
1049Function *PrototypeAST::Codegen() {
1050 // Make the function type: double(double,double) etc.
Chris Lattner35abbf52007-10-23 06:23:57 +0000[diff] [blame]1051 std::vector&lt;const Type*&gt; Doubles(Args.size(), Type::DoubleTy);
1052 FunctionType *FT = FunctionType::get(Type::DoubleTy, Doubles, false);
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]1053
1054 Function *F = new Function(FT, Function::ExternalLinkage, Name, TheModule);
1055
1056 // If F conflicted, there was already something named 'Name'. If it has a
1057 // body, don't allow redefinition or reextern.
1058 if (F-&gt;getName() != Name) {
1059 // Delete the one we just made and get the existing one.
1060 F-&gt;eraseFromParent();
1061 F = TheModule-&gt;getFunction(Name);
1062
1063 // If F already has a body, reject this.
1064 if (!F-&gt;empty()) {
1065 ErrorF("redefinition of function");
1066 return 0;
1067 }
1068
1069 // If F took a different number of args, reject.
1070 if (F-&gt;arg_size() != Args.size()) {
1071 ErrorF("redefinition of function with different # args");
1072 return 0;
1073 }
1074 }
1075
1076 // Set names for all arguments.
1077 unsigned Idx = 0;
1078 for (Function::arg_iterator AI = F-&gt;arg_begin(); Idx != Args.size();
1079 ++AI, ++Idx) {
1080 AI-&gt;setName(Args[Idx]);
1081
1082 // Add arguments to variable symbol table.
1083 NamedValues[Args[Idx]] = AI;
1084 }
1085
1086 return F;
1087}
1088
1089Function *FunctionAST::Codegen() {
1090 NamedValues.clear();
1091
1092 Function *TheFunction = Proto-&gt;Codegen();
1093 if (TheFunction == 0)
1094 return 0;
1095
1096 // Create a new basic block to start insertion into.
Chris Lattner35abbf52007-10-23 06:23:57 +0000[diff] [blame]1097 BasicBlock *BB = new BasicBlock("entry", TheFunction);
1098 Builder.SetInsertPoint(BB);
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]1099
1100 if (Value *RetVal = Body-&gt;Codegen()) {
1101 // Finish off the function.
1102 Builder.CreateRet(RetVal);
Chris Lattnerd9b86162007-10-25 04:30:35 +0000[diff] [blame]1103
1104 // Validate the generated code, checking for consistency.
1105 verifyFunction(*TheFunction);
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]1106 return TheFunction;
1107 }
1108
1109 // Error reading body, remove function.
1110 TheFunction-&gt;eraseFromParent();
1111 return 0;
1112}
1113
1114//===----------------------------------------------------------------------===//
1115// Top-Level parsing and JIT Driver
1116//===----------------------------------------------------------------------===//
1117
1118static void HandleDefinition() {
1119 if (FunctionAST *F = ParseDefinition()) {
1120 if (Function *LF = F-&gt;Codegen()) {
1121 fprintf(stderr, "Read function definition:");
1122 LF-&gt;dump();
1123 }
1124 } else {
1125 // Skip token for error recovery.
1126 getNextToken();
1127 }
1128}
1129
1130static void HandleExtern() {
1131 if (PrototypeAST *P = ParseExtern()) {
1132 if (Function *F = P-&gt;Codegen()) {
1133 fprintf(stderr, "Read extern: ");
1134 F-&gt;dump();
1135 }
1136 } else {
1137 // Skip token for error recovery.
1138 getNextToken();
1139 }
1140}
1141
1142static void HandleTopLevelExpression() {
1143 // Evaluate a top level expression into an anonymous function.
1144 if (FunctionAST *F = ParseTopLevelExpr()) {
1145 if (Function *LF = F-&gt;Codegen()) {
1146 fprintf(stderr, "Read top-level expression:");
1147 LF-&gt;dump();
1148 }
1149 } else {
1150 // Skip token for error recovery.
1151 getNextToken();
1152 }
1153}
1154
1155/// top ::= definition | external | expression | ';'
1156static void MainLoop() {
1157 while (1) {
1158 fprintf(stderr, "ready&gt; ");
1159 switch (CurTok) {
1160 case tok_eof: return;
1161 case ';': getNextToken(); break; // ignore top level semicolons.
1162 case tok_def: HandleDefinition(); break;
1163 case tok_extern: HandleExtern(); break;
1164 default: HandleTopLevelExpression(); break;
1165 }
1166 }
1167}
1168
1169
1170
1171//===----------------------------------------------------------------------===//
1172// "Library" functions that can be "extern'd" from user code.
1173//===----------------------------------------------------------------------===//
1174
1175/// putchard - putchar that takes a double and returns 0.
1176extern "C"
1177double putchard(double X) {
1178 putchar((char)X);
1179 return 0;
1180}
1181
1182//===----------------------------------------------------------------------===//
1183// Main driver code.
1184//===----------------------------------------------------------------------===//
1185
1186int main() {
1187 TheModule = new Module("my cool jit");
1188
1189 // Install standard binary operators.
1190 // 1 is lowest precedence.
1191 BinopPrecedence['&lt;'] = 10;
1192 BinopPrecedence['+'] = 20;
1193 BinopPrecedence['-'] = 20;
1194 BinopPrecedence['*'] = 40; // highest.
1195
1196 // Prime the first token.
1197 fprintf(stderr, "ready&gt; ");
1198 getNextToken();
1199
1200 MainLoop();
1201 TheModule-&gt;dump();
1202 return 0;
1203}
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]1204</pre>
1205</div>
1206</div>
1207
1208<!-- *********************************************************************** -->
1209<hr>
1210<address>
1211 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1212 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
1213 <a href="http://validator.w3.org/check/referer"><img
Chris Lattner8eef4b22007-10-23 06:30:50 +0000[diff] [blame]1214 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!"></a>
Chris Lattner2e902042007-10-22 07:01:42 +0000[diff] [blame]1215
1216 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1217 <a href="http://llvm.org">The LLVM Compiler Infrastructure</a><br>
1218 Last modified: $Date: 2007-10-17 11:05:13 -0700 (Wed, 17 Oct 2007) $
1219</address>
1220</body>
1221</html>