1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
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 <meta name="author" content="Erick Tryzelaar">
10 <link rel="stylesheet" href="../_static/llvm.css" type="text/css">
15 <h1>Kaleidoscope: Adding JIT and Optimizer Support</h1>
18 <li><a href="index.html">Up to Tutorial Index</a></li>
21 <li><a href="#intro">Chapter 4 Introduction</a></li>
22 <li><a href="#trivialconstfold">Trivial Constant Folding</a></li>
23 <li><a href="#optimizerpasses">LLVM Optimization Passes</a></li>
24 <li><a href="#jit">Adding a JIT Compiler</a></li>
25 <li><a href="#code">Full Code Listing</a></li>
28 <li><a href="OCamlLangImpl5.html">Chapter 5</a>: Extending the Language: Control
32 <div class="doc_author">
34 Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a>
35 and <a href="mailto:idadesub@users.sourceforge.net">Erick Tryzelaar</a>
39 <!-- *********************************************************************** -->
40 <h2><a name="intro">Chapter 4 Introduction</a></h2>
41 <!-- *********************************************************************** -->
45 <p>Welcome to Chapter 4 of the "<a href="index.html">Implementing a language
46 with LLVM</a>" tutorial. Chapters 1-3 described the implementation of a simple
47 language and added support for generating LLVM IR. This chapter describes
48 two new techniques: adding optimizer support to your language, and adding JIT
49 compiler support. These additions will demonstrate how to get nice, efficient code
50 for the Kaleidoscope language.</p>
54 <!-- *********************************************************************** -->
55 <h2><a name="trivialconstfold">Trivial Constant Folding</a></h2>
56 <!-- *********************************************************************** -->
60 <p><b>Note:</b> the default <tt>IRBuilder</tt> now always includes the constant
61 folding optimisations below.<p>
64 Our demonstration for Chapter 3 is elegant and easy to extend. Unfortunately,
65 it does not produce wonderful code. For example, when compiling simple code,
66 we don't get obvious optimizations:</p>
68 <div class="doc_code">
70 ready> <b>def test(x) 1+2+x;</b>
71 Read function definition:
72 define double @test(double %x) {
74 %addtmp = fadd double 1.000000e+00, 2.000000e+00
75 %addtmp1 = fadd double %addtmp, %x
81 <p>This code is a very, very literal transcription of the AST built by parsing
82 the input. As such, this transcription lacks optimizations like constant folding
83 (we'd like to get "<tt>add x, 3.0</tt>" in the example above) as well as other
84 more important optimizations. Constant folding, in particular, is a very common
85 and very important optimization: so much so that many language implementors
86 implement constant folding support in their AST representation.</p>
88 <p>With LLVM, you don't need this support in the AST. Since all calls to build
89 LLVM IR go through the LLVM builder, it would be nice if the builder itself
90 checked to see if there was a constant folding opportunity when you call it.
91 If so, it could just do the constant fold and return the constant instead of
92 creating an instruction. This is exactly what the <tt>LLVMFoldingBuilder</tt>
95 <p>All we did was switch from <tt>LLVMBuilder</tt> to
96 <tt>LLVMFoldingBuilder</tt>. Though we change no other code, we now have all of our
97 instructions implicitly constant folded without us having to do anything
98 about it. For example, the input above now compiles to:</p>
100 <div class="doc_code">
102 ready> <b>def test(x) 1+2+x;</b>
103 Read function definition:
104 define double @test(double %x) {
106 %addtmp = fadd double 3.000000e+00, %x
112 <p>Well, that was easy :). In practice, we recommend always using
113 <tt>LLVMFoldingBuilder</tt> when generating code like this. It has no
114 "syntactic overhead" for its use (you don't have to uglify your compiler with
115 constant checks everywhere) and it can dramatically reduce the amount of
116 LLVM IR that is generated in some cases (particular for languages with a macro
117 preprocessor or that use a lot of constants).</p>
119 <p>On the other hand, the <tt>LLVMFoldingBuilder</tt> is limited by the fact
120 that it does all of its analysis inline with the code as it is built. If you
121 take a slightly more complex example:</p>
123 <div class="doc_code">
125 ready> <b>def test(x) (1+2+x)*(x+(1+2));</b>
126 ready> Read function definition:
127 define double @test(double %x) {
129 %addtmp = fadd double 3.000000e+00, %x
130 %addtmp1 = fadd double %x, 3.000000e+00
131 %multmp = fmul double %addtmp, %addtmp1
137 <p>In this case, the LHS and RHS of the multiplication are the same value. We'd
138 really like to see this generate "<tt>tmp = x+3; result = tmp*tmp;</tt>" instead
139 of computing "<tt>x*3</tt>" twice.</p>
141 <p>Unfortunately, no amount of local analysis will be able to detect and correct
142 this. This requires two transformations: reassociation of expressions (to
143 make the add's lexically identical) and Common Subexpression Elimination (CSE)
144 to delete the redundant add instruction. Fortunately, LLVM provides a broad
145 range of optimizations that you can use, in the form of "passes".</p>
149 <!-- *********************************************************************** -->
150 <h2><a name="optimizerpasses">LLVM Optimization Passes</a></h2>
151 <!-- *********************************************************************** -->
155 <p>LLVM provides many optimization passes, which do many different sorts of
156 things and have different tradeoffs. Unlike other systems, LLVM doesn't hold
157 to the mistaken notion that one set of optimizations is right for all languages
158 and for all situations. LLVM allows a compiler implementor to make complete
159 decisions about what optimizations to use, in which order, and in what
162 <p>As a concrete example, LLVM supports both "whole module" passes, which look
163 across as large of body of code as they can (often a whole file, but if run
164 at link time, this can be a substantial portion of the whole program). It also
165 supports and includes "per-function" passes which just operate on a single
166 function at a time, without looking at other functions. For more information
167 on passes and how they are run, see the <a href="../WritingAnLLVMPass.html">How
168 to Write a Pass</a> document and the <a href="../Passes.html">List of LLVM
171 <p>For Kaleidoscope, we are currently generating functions on the fly, one at
172 a time, as the user types them in. We aren't shooting for the ultimate
173 optimization experience in this setting, but we also want to catch the easy and
174 quick stuff where possible. As such, we will choose to run a few per-function
175 optimizations as the user types the function in. If we wanted to make a "static
176 Kaleidoscope compiler", we would use exactly the code we have now, except that
177 we would defer running the optimizer until the entire file has been parsed.</p>
179 <p>In order to get per-function optimizations going, we need to set up a
180 <a href="../WritingAnLLVMPass.html#passmanager">Llvm.PassManager</a> to hold and
181 organize the LLVM optimizations that we want to run. Once we have that, we can
182 add a set of optimizations to run. The code looks like this:</p>
184 <div class="doc_code">
186 (* Create the JIT. *)
187 let the_execution_engine = ExecutionEngine.create Codegen.the_module in
188 let the_fpm = PassManager.create_function Codegen.the_module in
190 (* Set up the optimizer pipeline. Start with registering info about how the
191 * target lays out data structures. *)
192 DataLayout.add (ExecutionEngine.target_data the_execution_engine) the_fpm;
194 (* Do simple "peephole" optimizations and bit-twiddling optzn. *)
195 add_instruction_combining the_fpm;
197 (* reassociate expressions. *)
198 add_reassociation the_fpm;
200 (* Eliminate Common SubExpressions. *)
203 (* Simplify the control flow graph (deleting unreachable blocks, etc). *)
204 add_cfg_simplification the_fpm;
206 ignore (PassManager.initialize the_fpm);
208 (* Run the main "interpreter loop" now. *)
209 Toplevel.main_loop the_fpm the_execution_engine stream;
213 <p>The meat of the matter here, is the definition of "<tt>the_fpm</tt>". It
214 requires a pointer to the <tt>the_module</tt> to construct itself. Once it is
215 set up, we use a series of "add" calls to add a bunch of LLVM passes. The
216 first pass is basically boilerplate, it adds a pass so that later optimizations
217 know how the data structures in the program are laid out. The
218 "<tt>the_execution_engine</tt>" variable is related to the JIT, which we will
219 get to in the next section.</p>
221 <p>In this case, we choose to add 4 optimization passes. The passes we chose
222 here are a pretty standard set of "cleanup" optimizations that are useful for
223 a wide variety of code. I won't delve into what they do but, believe me,
224 they are a good starting place :).</p>
226 <p>Once the <tt>Llvm.PassManager.</tt> is set up, we need to make use of it.
227 We do this by running it after our newly created function is constructed (in
228 <tt>Codegen.codegen_func</tt>), but before it is returned to the client:</p>
230 <div class="doc_code">
232 let codegen_func the_fpm = function
235 let ret_val = codegen_expr body in
237 (* Finish off the function. *)
238 let _ = build_ret ret_val builder in
240 (* Validate the generated code, checking for consistency. *)
241 Llvm_analysis.assert_valid_function the_function;
243 (* Optimize the function. *)
244 let _ = PassManager.run_function the_function the_fpm in
250 <p>As you can see, this is pretty straightforward. The <tt>the_fpm</tt>
251 optimizes and updates the LLVM Function* in place, improving (hopefully) its
252 body. With this in place, we can try our test above again:</p>
254 <div class="doc_code">
256 ready> <b>def test(x) (1+2+x)*(x+(1+2));</b>
257 ready> Read function definition:
258 define double @test(double %x) {
260 %addtmp = fadd double %x, 3.000000e+00
261 %multmp = fmul double %addtmp, %addtmp
267 <p>As expected, we now get our nicely optimized code, saving a floating point
268 add instruction from every execution of this function.</p>
270 <p>LLVM provides a wide variety of optimizations that can be used in certain
271 circumstances. Some <a href="../Passes.html">documentation about the various
272 passes</a> is available, but it isn't very complete. Another good source of
273 ideas can come from looking at the passes that <tt>Clang</tt> runs to get
274 started. The "<tt>opt</tt>" tool allows you to experiment with passes from the
275 command line, so you can see if they do anything.</p>
277 <p>Now that we have reasonable code coming out of our front-end, lets talk about
282 <!-- *********************************************************************** -->
283 <h2><a name="jit">Adding a JIT Compiler</a></h2>
284 <!-- *********************************************************************** -->
288 <p>Code that is available in LLVM IR can have a wide variety of tools
289 applied to it. For example, you can run optimizations on it (as we did above),
290 you can dump it out in textual or binary forms, you can compile the code to an
291 assembly file (.s) for some target, or you can JIT compile it. The nice thing
292 about the LLVM IR representation is that it is the "common currency" between
293 many different parts of the compiler.
296 <p>In this section, we'll add JIT compiler support to our interpreter. The
297 basic idea that we want for Kaleidoscope is to have the user enter function
298 bodies as they do now, but immediately evaluate the top-level expressions they
299 type in. For example, if they type in "1 + 2;", we should evaluate and print
300 out 3. If they define a function, they should be able to call it from the
303 <p>In order to do this, we first declare and initialize the JIT. This is done
304 by adding a global variable and a call in <tt>main</tt>:</p>
306 <div class="doc_code">
311 <b>(* Create the JIT. *)
312 let the_execution_engine = ExecutionEngine.create Codegen.the_module in</b>
317 <p>This creates an abstract "Execution Engine" which can be either a JIT
318 compiler or the LLVM interpreter. LLVM will automatically pick a JIT compiler
319 for you if one is available for your platform, otherwise it will fall back to
322 <p>Once the <tt>Llvm_executionengine.ExecutionEngine.t</tt> is created, the JIT
323 is ready to be used. There are a variety of APIs that are useful, but the
324 simplest one is the "<tt>Llvm_executionengine.ExecutionEngine.run_function</tt>"
325 function. This method JIT compiles the specified LLVM Function and returns a
326 function pointer to the generated machine code. In our case, this means that we
327 can change the code that parses a top-level expression to look like this:</p>
329 <div class="doc_code">
331 (* Evaluate a top-level expression into an anonymous function. *)
332 let e = Parser.parse_toplevel stream in
333 print_endline "parsed a top-level expr";
334 let the_function = Codegen.codegen_func the_fpm e in
335 dump_value the_function;
337 (* JIT the function, returning a function pointer. *)
338 let result = ExecutionEngine.run_function the_function [||]
339 the_execution_engine in
341 print_string "Evaluated to ";
342 print_float (GenericValue.as_float Codegen.double_type result);
347 <p>Recall that we compile top-level expressions into a self-contained LLVM
348 function that takes no arguments and returns the computed double. Because the
349 LLVM JIT compiler matches the native platform ABI, this means that you can just
350 cast the result pointer to a function pointer of that type and call it directly.
351 This means, there is no difference between JIT compiled code and native machine
352 code that is statically linked into your application.</p>
354 <p>With just these two changes, lets see how Kaleidoscope works now!</p>
356 <div class="doc_code">
358 ready> <b>4+5;</b>
359 define double @""() {
361 ret double 9.000000e+00
364 <em>Evaluated to 9.000000</em>
368 <p>Well this looks like it is basically working. The dump of the function
369 shows the "no argument function that always returns double" that we synthesize
370 for each top level expression that is typed in. This demonstrates very basic
371 functionality, but can we do more?</p>
373 <div class="doc_code">
375 ready> <b>def testfunc(x y) x + y*2; </b>
376 Read function definition:
377 define double @testfunc(double %x, double %y) {
379 %multmp = fmul double %y, 2.000000e+00
380 %addtmp = fadd double %multmp, %x
384 ready> <b>testfunc(4, 10);</b>
385 define double @""() {
387 %calltmp = call double @testfunc(double 4.000000e+00, double 1.000000e+01)
391 <em>Evaluated to 24.000000</em>
395 <p>This illustrates that we can now call user code, but there is something a bit
396 subtle going on here. Note that we only invoke the JIT on the anonymous
397 functions that <em>call testfunc</em>, but we never invoked it
398 on <em>testfunc</em> itself. What actually happened here is that the JIT
399 scanned for all non-JIT'd functions transitively called from the anonymous
400 function and compiled all of them before returning
401 from <tt>run_function</tt>.</p>
403 <p>The JIT provides a number of other more advanced interfaces for things like
404 freeing allocated machine code, rejit'ing functions to update them, etc.
405 However, even with this simple code, we get some surprisingly powerful
406 capabilities - check this out (I removed the dump of the anonymous functions,
407 you should get the idea by now :) :</p>
409 <div class="doc_code">
411 ready> <b>extern sin(x);</b>
413 declare double @sin(double)
415 ready> <b>extern cos(x);</b>
417 declare double @cos(double)
419 ready> <b>sin(1.0);</b>
420 <em>Evaluated to 0.841471</em>
422 ready> <b>def foo(x) sin(x)*sin(x) + cos(x)*cos(x);</b>
423 Read function definition:
424 define double @foo(double %x) {
426 %calltmp = call double @sin(double %x)
427 %multmp = fmul double %calltmp, %calltmp
428 %calltmp2 = call double @cos(double %x)
429 %multmp4 = fmul double %calltmp2, %calltmp2
430 %addtmp = fadd double %multmp, %multmp4
434 ready> <b>foo(4.0);</b>
435 <em>Evaluated to 1.000000</em>
439 <p>Whoa, how does the JIT know about sin and cos? The answer is surprisingly
440 simple: in this example, the JIT started execution of a function and got to a
441 function call. It realized that the function was not yet JIT compiled and
442 invoked the standard set of routines to resolve the function. In this case,
443 there is no body defined for the function, so the JIT ended up calling
444 "<tt>dlsym("sin")</tt>" on the Kaleidoscope process itself. Since
445 "<tt>sin</tt>" is defined within the JIT's address space, it simply patches up
446 calls in the module to call the libm version of <tt>sin</tt> directly.</p>
448 <p>The LLVM JIT provides a number of interfaces (look in the
449 <tt>llvm_executionengine.mli</tt> file) for controlling how unknown functions
450 get resolved. It allows you to establish explicit mappings between IR objects
451 and addresses (useful for LLVM global variables that you want to map to static
452 tables, for example), allows you to dynamically decide on the fly based on the
453 function name, and even allows you to have the JIT compile functions lazily the
454 first time they're called.</p>
456 <p>One interesting application of this is that we can now extend the language
457 by writing arbitrary C code to implement operations. For example, if we add:
460 <div class="doc_code">
462 /* putchard - putchar that takes a double and returns 0. */
464 double putchard(double X) {
471 <p>Now we can produce simple output to the console by using things like:
472 "<tt>extern putchard(x); putchard(120);</tt>", which prints a lowercase 'x' on
473 the console (120 is the ASCII code for 'x'). Similar code could be used to
474 implement file I/O, console input, and many other capabilities in
477 <p>This completes the JIT and optimizer chapter of the Kaleidoscope tutorial. At
478 this point, we can compile a non-Turing-complete programming language, optimize
479 and JIT compile it in a user-driven way. Next up we'll look into <a
480 href="OCamlLangImpl5.html">extending the language with control flow
481 constructs</a>, tackling some interesting LLVM IR issues along the way.</p>
485 <!-- *********************************************************************** -->
486 <h2><a name="code">Full Code Listing</a></h2>
487 <!-- *********************************************************************** -->
492 Here is the complete code listing for our running example, enhanced with the
493 LLVM JIT and optimizer. To build this example, use:
496 <div class="doc_code">
505 <p>Here is the code:</p>
509 <dd class="doc_code">
511 <{lexer,parser}.ml>: use_camlp4, pp(camlp4of)
512 <*.{byte,native}>: g++, use_llvm, use_llvm_analysis
513 <*.{byte,native}>: use_llvm_executionengine, use_llvm_target
514 <*.{byte,native}>: use_llvm_scalar_opts, use_bindings
518 <dt>myocamlbuild.ml:</dt>
519 <dd class="doc_code">
521 open Ocamlbuild_plugin;;
523 ocaml_lib ~extern:true "llvm";;
524 ocaml_lib ~extern:true "llvm_analysis";;
525 ocaml_lib ~extern:true "llvm_executionengine";;
526 ocaml_lib ~extern:true "llvm_target";;
527 ocaml_lib ~extern:true "llvm_scalar_opts";;
529 flag ["link"; "ocaml"; "g++"] (S[A"-cc"; A"g++"]);;
530 dep ["link"; "ocaml"; "use_bindings"] ["bindings.o"];;
535 <dd class="doc_code">
537 (*===----------------------------------------------------------------------===
539 *===----------------------------------------------------------------------===*)
541 (* The lexer returns these 'Kwd' if it is an unknown character, otherwise one of
542 * these others for known things. *)
548 | Ident of string | Number of float
556 <dd class="doc_code">
558 (*===----------------------------------------------------------------------===
560 *===----------------------------------------------------------------------===*)
563 (* Skip any whitespace. *)
564 | [< ' (' ' | '\n' | '\r' | '\t'); stream >] -> lex stream
566 (* identifier: [a-zA-Z][a-zA-Z0-9] *)
567 | [< ' ('A' .. 'Z' | 'a' .. 'z' as c); stream >] ->
568 let buffer = Buffer.create 1 in
569 Buffer.add_char buffer c;
570 lex_ident buffer stream
572 (* number: [0-9.]+ *)
573 | [< ' ('0' .. '9' as c); stream >] ->
574 let buffer = Buffer.create 1 in
575 Buffer.add_char buffer c;
576 lex_number buffer stream
578 (* Comment until end of line. *)
579 | [< ' ('#'); stream >] ->
582 (* Otherwise, just return the character as its ascii value. *)
583 | [< 'c; stream >] ->
584 [< 'Token.Kwd c; lex stream >]
587 | [< >] -> [< >]
589 and lex_number buffer = parser
590 | [< ' ('0' .. '9' | '.' as c); stream >] ->
591 Buffer.add_char buffer c;
592 lex_number buffer stream
593 | [< stream=lex >] ->
594 [< 'Token.Number (float_of_string (Buffer.contents buffer)); stream >]
596 and lex_ident buffer = parser
597 | [< ' ('A' .. 'Z' | 'a' .. 'z' | '0' .. '9' as c); stream >] ->
598 Buffer.add_char buffer c;
599 lex_ident buffer stream
600 | [< stream=lex >] ->
601 match Buffer.contents buffer with
602 | "def" -> [< 'Token.Def; stream >]
603 | "extern" -> [< 'Token.Extern; stream >]
604 | id -> [< 'Token.Ident id; stream >]
606 and lex_comment = parser
607 | [< ' ('\n'); stream=lex >] -> stream
608 | [< 'c; e=lex_comment >] -> e
609 | [< >] -> [< >]
614 <dd class="doc_code">
616 (*===----------------------------------------------------------------------===
617 * Abstract Syntax Tree (aka Parse Tree)
618 *===----------------------------------------------------------------------===*)
620 (* expr - Base type for all expression nodes. *)
622 (* variant for numeric literals like "1.0". *)
625 (* variant for referencing a variable, like "a". *)
628 (* variant for a binary operator. *)
629 | Binary of char * expr * expr
631 (* variant for function calls. *)
632 | Call of string * expr array
634 (* proto - This type represents the "prototype" for a function, which captures
635 * its name, and its argument names (thus implicitly the number of arguments the
636 * function takes). *)
637 type proto = Prototype of string * string array
639 (* func - This type represents a function definition itself. *)
640 type func = Function of proto * expr
645 <dd class="doc_code">
647 (*===---------------------------------------------------------------------===
649 *===---------------------------------------------------------------------===*)
651 (* binop_precedence - This holds the precedence for each binary operator that is
653 let binop_precedence:(char, int) Hashtbl.t = Hashtbl.create 10
655 (* precedence - Get the precedence of the pending binary operator token. *)
656 let precedence c = try Hashtbl.find binop_precedence c with Not_found -> -1
662 let rec parse_primary = parser
663 (* numberexpr ::= number *)
664 | [< 'Token.Number n >] -> Ast.Number n
666 (* parenexpr ::= '(' expression ')' *)
667 | [< 'Token.Kwd '('; e=parse_expr; 'Token.Kwd ')' ?? "expected ')'" >] -> e
671 * ::= identifier '(' argumentexpr ')' *)
672 | [< 'Token.Ident id; stream >] ->
673 let rec parse_args accumulator = parser
674 | [< e=parse_expr; stream >] ->
676 | [< 'Token.Kwd ','; e=parse_args (e :: accumulator) >] -> e
677 | [< >] -> e :: accumulator
679 | [< >] -> accumulator
681 let rec parse_ident id = parser
683 | [< 'Token.Kwd '(';
685 'Token.Kwd ')' ?? "expected ')'">] ->
686 Ast.Call (id, Array.of_list (List.rev args))
688 (* Simple variable ref. *)
689 | [< >] -> Ast.Variable id
691 parse_ident id stream
693 | [< >] -> raise (Stream.Error "unknown token when expecting an expression.")
696 * ::= ('+' primary)* *)
697 and parse_bin_rhs expr_prec lhs stream =
698 match Stream.peek stream with
699 (* If this is a binop, find its precedence. *)
700 | Some (Token.Kwd c) when Hashtbl.mem binop_precedence c ->
701 let token_prec = precedence c in
703 (* If this is a binop that binds at least as tightly as the current binop,
704 * consume it, otherwise we are done. *)
705 if token_prec < expr_prec then lhs else begin
709 (* Parse the primary expression after the binary operator. *)
710 let rhs = parse_primary stream in
712 (* Okay, we know this is a binop. *)
714 match Stream.peek stream with
715 | Some (Token.Kwd c2) ->
716 (* If BinOp binds less tightly with rhs than the operator after
717 * rhs, let the pending operator take rhs as its lhs. *)
718 let next_prec = precedence c2 in
719 if token_prec < next_prec
720 then parse_bin_rhs (token_prec + 1) rhs stream
726 let lhs = Ast.Binary (c, lhs, rhs) in
727 parse_bin_rhs expr_prec lhs stream
732 * ::= primary binoprhs *)
733 and parse_expr = parser
734 | [< lhs=parse_primary; stream >] -> parse_bin_rhs 0 lhs stream
737 * ::= id '(' id* ')' *)
738 let parse_prototype =
739 let rec parse_args accumulator = parser
740 | [< 'Token.Ident id; e=parse_args (id::accumulator) >] -> e
741 | [< >] -> accumulator
745 | [< 'Token.Ident id;
746 'Token.Kwd '(' ?? "expected '(' in prototype";
748 'Token.Kwd ')' ?? "expected ')' in prototype" >] ->
750 Ast.Prototype (id, Array.of_list (List.rev args))
753 raise (Stream.Error "expected function name in prototype")
755 (* definition ::= 'def' prototype expression *)
756 let parse_definition = parser
757 | [< 'Token.Def; p=parse_prototype; e=parse_expr >] ->
760 (* toplevelexpr ::= expression *)
761 let parse_toplevel = parser
762 | [< e=parse_expr >] ->
763 (* Make an anonymous proto. *)
764 Ast.Function (Ast.Prototype ("", [||]), e)
766 (* external ::= 'extern' prototype *)
767 let parse_extern = parser
768 | [< 'Token.Extern; e=parse_prototype >] -> e
773 <dd class="doc_code">
775 (*===----------------------------------------------------------------------===
777 *===----------------------------------------------------------------------===*)
781 exception Error of string
783 let context = global_context ()
784 let the_module = create_module context "my cool jit"
785 let builder = builder context
786 let named_values:(string, llvalue) Hashtbl.t = Hashtbl.create 10
787 let double_type = double_type context
789 let rec codegen_expr = function
790 | Ast.Number n -> const_float double_type n
791 | Ast.Variable name ->
792 (try Hashtbl.find named_values name with
793 | Not_found -> raise (Error "unknown variable name"))
794 | Ast.Binary (op, lhs, rhs) ->
795 let lhs_val = codegen_expr lhs in
796 let rhs_val = codegen_expr rhs in
799 | '+' -> build_add lhs_val rhs_val "addtmp" builder
800 | '-' -> build_sub lhs_val rhs_val "subtmp" builder
801 | '*' -> build_mul lhs_val rhs_val "multmp" builder
803 (* Convert bool 0/1 to double 0.0 or 1.0 *)
804 let i = build_fcmp Fcmp.Ult lhs_val rhs_val "cmptmp" builder in
805 build_uitofp i double_type "booltmp" builder
806 | _ -> raise (Error "invalid binary operator")
808 | Ast.Call (callee, args) ->
809 (* Look up the name in the module table. *)
811 match lookup_function callee the_module with
812 | Some callee -> callee
813 | None -> raise (Error "unknown function referenced")
815 let params = params callee in
817 (* If argument mismatch error. *)
818 if Array.length params == Array.length args then () else
819 raise (Error "incorrect # arguments passed");
820 let args = Array.map codegen_expr args in
821 build_call callee args "calltmp" builder
823 let codegen_proto = function
824 | Ast.Prototype (name, args) ->
825 (* Make the function type: double(double,double) etc. *)
826 let doubles = Array.make (Array.length args) double_type in
827 let ft = function_type double_type doubles in
829 match lookup_function name the_module with
830 | None -> declare_function name ft the_module
832 (* If 'f' conflicted, there was already something named 'name'. If it
833 * has a body, don't allow redefinition or reextern. *)
835 (* If 'f' already has a body, reject this. *)
836 if block_begin f <> At_end f then
837 raise (Error "redefinition of function");
839 (* If 'f' took a different number of arguments, reject. *)
840 if element_type (type_of f) <> ft then
841 raise (Error "redefinition of function with different # args");
845 (* Set names for all arguments. *)
846 Array.iteri (fun i a ->
849 Hashtbl.add named_values n a;
853 let codegen_func the_fpm = function
854 | Ast.Function (proto, body) ->
855 Hashtbl.clear named_values;
856 let the_function = codegen_proto proto in
858 (* Create a new basic block to start insertion into. *)
859 let bb = append_block context "entry" the_function in
860 position_at_end bb builder;
863 let ret_val = codegen_expr body in
865 (* Finish off the function. *)
866 let _ = build_ret ret_val builder in
868 (* Validate the generated code, checking for consistency. *)
869 Llvm_analysis.assert_valid_function the_function;
871 (* Optimize the function. *)
872 let _ = PassManager.run_function the_function the_fpm in
876 delete_function the_function;
881 <dt>toplevel.ml:</dt>
882 <dd class="doc_code">
884 (*===----------------------------------------------------------------------===
885 * Top-Level parsing and JIT Driver
886 *===----------------------------------------------------------------------===*)
889 open Llvm_executionengine
891 (* top ::= definition | external | expression | ';' *)
892 let rec main_loop the_fpm the_execution_engine stream =
893 match Stream.peek stream with
896 (* ignore top-level semicolons. *)
897 | Some (Token.Kwd ';') ->
899 main_loop the_fpm the_execution_engine stream
905 let e = Parser.parse_definition stream in
906 print_endline "parsed a function definition.";
907 dump_value (Codegen.codegen_func the_fpm e);
909 let e = Parser.parse_extern stream in
910 print_endline "parsed an extern.";
911 dump_value (Codegen.codegen_proto e);
913 (* Evaluate a top-level expression into an anonymous function. *)
914 let e = Parser.parse_toplevel stream in
915 print_endline "parsed a top-level expr";
916 let the_function = Codegen.codegen_func the_fpm e in
917 dump_value the_function;
919 (* JIT the function, returning a function pointer. *)
920 let result = ExecutionEngine.run_function the_function [||]
921 the_execution_engine in
923 print_string "Evaluated to ";
924 print_float (GenericValue.as_float Codegen.double_type result);
926 with Stream.Error s | Codegen.Error s ->
927 (* Skip token for error recovery. *)
931 print_string "ready> "; flush stdout;
932 main_loop the_fpm the_execution_engine stream
937 <dd class="doc_code">
939 (*===----------------------------------------------------------------------===
941 *===----------------------------------------------------------------------===*)
944 open Llvm_executionengine
946 open Llvm_scalar_opts
949 ignore (initialize_native_target ());
951 (* Install standard binary operators.
952 * 1 is the lowest precedence. *)
953 Hashtbl.add Parser.binop_precedence '<' 10;
954 Hashtbl.add Parser.binop_precedence '+' 20;
955 Hashtbl.add Parser.binop_precedence '-' 20;
956 Hashtbl.add Parser.binop_precedence '*' 40; (* highest. *)
958 (* Prime the first token. *)
959 print_string "ready> "; flush stdout;
960 let stream = Lexer.lex (Stream.of_channel stdin) in
962 (* Create the JIT. *)
963 let the_execution_engine = ExecutionEngine.create Codegen.the_module in
964 let the_fpm = PassManager.create_function Codegen.the_module in
966 (* Set up the optimizer pipeline. Start with registering info about how the
967 * target lays out data structures. *)
968 DataLayout.add (ExecutionEngine.target_data the_execution_engine) the_fpm;
970 (* Do simple "peephole" optimizations and bit-twiddling optzn. *)
971 add_instruction_combination the_fpm;
973 (* reassociate expressions. *)
974 add_reassociation the_fpm;
976 (* Eliminate Common SubExpressions. *)
979 (* Simplify the control flow graph (deleting unreachable blocks, etc). *)
980 add_cfg_simplification the_fpm;
982 ignore (PassManager.initialize the_fpm);
984 (* Run the main "interpreter loop" now. *)
985 Toplevel.main_loop the_fpm the_execution_engine stream;
987 (* Print out all the generated code. *)
988 dump_module Codegen.the_module
996 <dd class="doc_code">
998 #include <stdio.h>
1000 /* putchard - putchar that takes a double and returns 0. */
1001 extern double putchard(double X) {
1009 <a href="OCamlLangImpl5.html">Next: Extending the language: control flow</a>
1012 <!-- *********************************************************************** -->
1015 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1016 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
1017 <a href="http://validator.w3.org/check/referer"><img
1018 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!"></a>
1020 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1021 <a href="mailto:idadesub@users.sourceforge.net">Erick Tryzelaar</a><br>
1022 <a href="http://llvm.org/">The LLVM Compiler Infrastructure</a><br>
1023 Last modified: $Date$