1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <meta http-equiv="Content-Type" Content="text/html; charset=UTF-8" >
6 <title>Accurate Garbage Collection with LLVM</title>
7 <link rel="stylesheet" href="llvm.css" type="text/css">
8 <style type="text/css">
9 .rowhead { text-align: left; background: inherit; }
10 .indent { padding-left: 1em; }
11 .optl { color: #BFBFBF; }
16 <div class="doc_title">
17 Accurate Garbage Collection with LLVM
21 <li><a href="#introduction">Introduction</a>
23 <li><a href="#feature">GC features provided and algorithms
28 <li><a href="#usage">Using the collectors</a>
30 <li><a href="#shadow-stack">ShadowStack -
31 A highly portable collector</a></li>
32 <li><a href="#semispace">SemiSpace -
33 A simple copying collector runtime</a></li>
34 <li><a href="#ocaml">Ocaml -
35 An Objective Caml-compatible collector</a></li>
39 <li><a href="#core">Core support</a>
41 <li><a href="#gcattr">Specifying GC code generation:
42 <tt>gc "..."</tt></a></li>
43 <li><a href="#gcroot">Identifying GC roots on the stack:
44 <tt>llvm.gcroot</tt></a></li>
45 <li><a href="#barriers">Reading and writing references in the heap</a>
47 <li><a href="#gcwrite">Write barrier: <tt>llvm.gcwrite</tt></a></li>
48 <li><a href="#gcread">Read barrier: <tt>llvm.gcread</tt></a></li>
54 <li><a href="#runtime">Recommended runtime interface</a>
56 <li><a href="#initialize">Garbage collector startup and
57 initialization</a></li>
58 <li><a href="#allocate">Allocating memory from the GC</a></li>
59 <li><a href="#explicit">Explicit invocation of the garbage
61 <li><a href="#traceroots">Tracing GC pointers from the program
63 <li><a href="#staticroots">Tracing GC pointers from static roots</a></li>
67 <li><a href="#plugin">Implementing a collector plugin</a>
69 <li><a href="#collector-algos">Overview of available features</a></li>
70 <li><a href="#stack-map">Computing stack maps</a></li>
71 <li><a href="#init-roots">Initializing roots to null:
72 <tt>InitRoots</tt></a></li>
73 <li><a href="#custom">Custom lowering of intrinsics: <tt>CustomRoots</tt>,
74 <tt>CustomReadBarriers</tt>, and <tt>CustomWriteBarriers</tt></a></li>
75 <li><a href="#safe-points">Generating safe points:
76 <tt>NeededSafePoints</tt></a></li>
77 <li><a href="#assembly">Emitting assembly code:
78 <tt>beginAssembly</tt> and <tt>finishAssembly</tt></a></li>
82 <li><a href="#runtime-impl">Implementing a collector runtime</a>
84 <li><a href="#gcdescriptors">Tracing GC pointers from heap
89 <li><a href="#references">References</a></li>
93 <div class="doc_author">
94 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a> and
98 <!-- *********************************************************************** -->
99 <div class="doc_section">
100 <a name="introduction">Introduction</a>
102 <!-- *********************************************************************** -->
104 <div class="doc_text">
106 <p>Garbage collection is a widely used technique that frees the programmer from
107 having to know the lifetimes of heap objects, making software easier to produce
108 and maintain. Many programming languages rely on garbage collection for
109 automatic memory management. There are two primary forms of garbage collection:
110 conservative and accurate.</p>
112 <p>Conservative garbage collection often does not require any special support
113 from either the language or the compiler: it can handle non-type-safe
114 programming languages (such as C/C++) and does not require any special
115 information from the compiler. The
116 <a href="http://www.hpl.hp.com/personal/Hans_Boehm/gc/">Boehm collector</a> is
117 an example of a state-of-the-art conservative collector.</p>
119 <p>Accurate garbage collection requires the ability to identify all pointers in
120 the program at run-time (which requires that the source-language be type-safe in
121 most cases). Identifying pointers at run-time requires compiler support to
122 locate all places that hold live pointer variables at run-time, including the
123 <a href="#gcroot">processor stack and registers</a>.</p>
125 <p>Conservative garbage collection is attractive because it does not require any
126 special compiler support, but it does have problems. In particular, because the
127 conservative garbage collector cannot <i>know</i> that a particular word in the
128 machine is a pointer, it cannot move live objects in the heap (preventing the
129 use of compacting and generational GC algorithms) and it can occasionally suffer
130 from memory leaks due to integer values that happen to point to objects in the
131 program. In addition, some aggressive compiler transformations can break
132 conservative garbage collectors (though these seem rare in practice).</p>
134 <p>Accurate garbage collectors do not suffer from any of these problems, but
135 they can suffer from degraded scalar optimization of the program. In particular,
136 because the runtime must be able to identify and update all pointers active in
137 the program, some optimizations are less effective. In practice, however, the
138 locality and performance benefits of using aggressive garbage allocation
139 techniques dominates any low-level losses.</p>
141 <p>This document describes the mechanisms and interfaces provided by LLVM to
142 support accurate garbage collection.</p>
146 <!-- ======================================================================= -->
147 <div class="doc_subsection">
148 <a name="feature">GC features provided and algorithms supported</a>
151 <div class="doc_text">
153 <p>LLVM's intermediate representation provides <a href="#intrinsics">garbage
154 collection intrinsics</a> that offer support for a broad class of
155 collector models. For instance, the intrinsics permit:</p>
158 <li>semi-space collectors</li>
159 <li>mark-sweep collectors</li>
160 <li>generational collectors</li>
161 <li>reference counting</li>
162 <li>incremental collectors</li>
163 <li>concurrent collectors</li>
164 <li>cooperative collectors</li>
167 <p>We hope that the primitive support built into the LLVM IR is sufficient to
168 support a broad class of garbage collected languages including Scheme, ML, Java,
169 C#, Perl, Python, Lua, Ruby, other scripting languages, and more.</p>
171 <p>However, LLVM does not itself implement a garbage collector. This is because
172 collectors are tightly coupled to object models, and LLVM is agnostic to object
173 models. Since LLVM is agnostic to object models, it would be inappropriate for
174 LLVM to dictate any particular collector. Instead, LLVM provides a framework for
175 garbage collector implementations in two manners:</p>
178 <li><b>At compile time</b> with <a href="#plugin">collector plugins</a> for
179 the compiler. Collector plugins have ready access to important garbage
180 collector algorithms. Leveraging these tools, it is straightforward to
181 emit type-accurate stack maps for your runtime in as little as ~100 lines of
184 <li><b>At runtime</b> with <a href="#runtime">suggested runtime
185 interfaces</a>, which allow front-end compilers to support a range of
186 collection runtimes.</li>
191 <!-- *********************************************************************** -->
192 <div class="doc_section">
193 <a name="usage">Using the collectors</a>
195 <!-- *********************************************************************** -->
197 <div class="doc_text">
199 <p>In general, using a collector implies:</p>
202 <li>Emitting compatible code, including initialization in the main
203 program if necessary.</li>
204 <li>Loading a compiler plugin if the collector is not statically linked with
205 your compiler. For <tt>llc</tt>, use the <tt>-load</tt> option.</li>
206 <li>Selecting the collection algorithm by applying the <tt>gc "..."</tt>
207 attribute to your garbage collected functions, or equivalently with
208 the <tt>setCollector</tt> method.</li>
209 <li>Linking your final executable with the garbage collector runtime.</li>
212 <p>This table summarizes the available runtimes.</p>
217 <th><tt>gc</tt> attribute</th>
219 <th><tt>gcroot</tt></th>
220 <th><tt>gcread</tt></th>
221 <th><tt>gcwrite</tt></th>
223 <tr valign="baseline">
224 <td><a href="#semispace">SemiSpace</a></td>
225 <td><tt>gc "shadow-stack"</tt></td>
231 <tr valign="baseline">
232 <td><a href="#ocaml">Ocaml</a></td>
233 <td><tt>gc "ocaml"</tt></td>
234 <td><i>provided by ocamlopt</i></td>
241 <p>The sections for <a href="#intrinsics">Collection intrinsics</a> and
242 <a href="#runtime">Recommended runtime interface</a> detail the interfaces that
243 collectors may require user programs to utilize.</p>
247 <!-- ======================================================================= -->
248 <div class="doc_subsection">
249 <a name="shadow-stack">ShadowStack - A highly portable collector</a>
252 <div class="doc_code"><tt>
253 Collector *llvm::createShadowStackCollector();
256 <div class="doc_text">
258 <p>The ShadowStack backend is invoked with the <tt>gc "shadow-stack"</tt>
260 Unlike many collectors which rely on a cooperative code generator to generate
261 stack maps, this algorithm carefully maintains a linked list of stack root
262 descriptors [<a href="#henderson02">Henderson2002</a>]. This so-called "shadow
263 stack" mirrors the machine stack. Maintaining this data structure is slower
264 than using stack maps, but has a significant portability advantage because it
265 requires no special support from the target code generator.</p>
267 <p>The ShadowStack collector does not use read or write barriers, so the user
268 program may use <tt>load</tt> and <tt>store</tt> instead of <tt>llvm.gcread</tt>
269 and <tt>llvm.gcwrite</tt>.</p>
271 <p>ShadowStack is a code generator plugin only. It must be paired with a
272 compatible runtime.</p>
276 <!-- ======================================================================= -->
277 <div class="doc_subsection">
278 <a name="semispace">SemiSpace - A simple copying collector runtime</a>
281 <div class="doc_text">
283 <p>The SemiSpace runtime implements the <a href="runtime">suggested
284 runtime interface</a> and is compatible with the ShadowStack backend.</p>
286 <p>SemiSpace is a very simple copying collector. When it starts up, it
287 allocates two blocks of memory for the heap. It uses a simple bump-pointer
288 allocator to allocate memory from the first block until it runs out of space.
289 When it runs out of space, it traces through all of the roots of the program,
290 copying blocks to the other half of the memory space.</p>
292 <p>This runtime is highly experimental and has not been used in a real project.
293 Enhancements would be welcomed.</p>
297 <!-- ======================================================================= -->
298 <div class="doc_subsection">
299 <a name="ocaml">Ocaml - An Objective Caml-compatible collector</a>
302 <div class="doc_code"><tt>
303 Collector *llvm::createOcamlCollector();
306 <div class="doc_text">
308 <p>The ocaml backend is invoked with the <tt>gc "ocaml"</tt> function attribute.
310 <a href="http://caml.inria.fr/">Objective Caml</a> language runtime by emitting
311 a type-accurate stack map in the form of an ocaml 3.10.0-compatible frametable.
312 The linkage requirements are satisfied automatically by the <tt>ocamlopt</tt>
313 compiler when linking an executable.</p>
315 <p>The ocaml collector does not use read or write barriers, so the user program
316 may use <tt>load</tt> and <tt>store</tt> instead of <tt>llvm.gcread</tt> and
317 <tt>llvm.gcwrite</tt>.</p>
322 <!-- *********************************************************************** -->
323 <div class="doc_section">
324 <a name="core">Core support</a><a name="intrinsics"></a>
326 <!-- *********************************************************************** -->
328 <div class="doc_text">
330 <p>This section describes the garbage collection facilities provided by the
331 <a href="LangRef.html">LLVM intermediate representation</a>.</p>
333 <p>These facilities are limited to those strictly necessary for compilation.
334 They are not intended to be a complete interface to any garbage collector.
335 Notably, heap allocation is not among the supplied primitives. A user program
336 will also need to interface with the runtime, using either the
337 <a href="#runtime">suggested runtime interface</a> or another interface
338 specified by the runtime.</p>
342 <!-- ======================================================================= -->
343 <div class="doc_subsection">
344 <a name="gcattr">Specifying GC code generation: <tt>gc "..."</tt></a>
347 <div class="doc_code"><tt>
348 define <i>ty</i> @<i>name</i>(...) <u>gc "<i>collector</i>"</u> { ...
351 <div class="doc_text">
353 <p>The <tt>gc</tt> function attribute is used to specify the desired collector
354 algorithm to the compiler. It is equivalent to specifying the collector name
355 programmatically using the <tt>setCollector</tt> method of
356 <tt>Function</tt>.</p>
358 <p>Specifying the collector on a per-function basis allows LLVM to link together
359 programs that use different garbage collection algorithms.</p>
363 <!-- ======================================================================= -->
364 <div class="doc_subsection">
365 <a name="gcroot">Identifying GC roots on the stack: <tt>llvm.gcroot</tt></a>
368 <div class="doc_code"><tt>
369 void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
372 <div class="doc_text">
374 <p>The <tt>llvm.gcroot</tt> intrinsic is used to inform LLVM of a pointer
375 variable on the stack. The first argument <b>must</b> be a value referring to an alloca instruction
376 or a bitcast of an alloca. The second contains a pointer to metadata that
377 should be associated with the pointer, and <b>must</b> be a constant or global
378 value address. If your target collector uses tags, use a null pointer for
381 <p>Consider the following fragment of Java code:</p>
385 Object X; // A null-initialized reference to an object
390 <p>This block (which may be located in the middle of a function or in a loop
391 nest), could be compiled to this LLVM code:</p>
395 ;; In the entry block for the function, allocate the
396 ;; stack space for X, which is an LLVM pointer.
399 ;; Tell LLVM that the stack space is a stack root.
400 ;; Java has type-tags on objects, so we pass null as metadata.
401 %tmp = bitcast %Object** %X to i8**
402 call void @llvm.gcroot(i8** %X, i8* null)
405 ;; "CodeBlock" is the block corresponding to the start
406 ;; of the scope above.
408 ;; Java null-initializes pointers.
409 store %Object* null, %Object** %X
413 ;; As the pointer goes out of scope, store a null value into
414 ;; it, to indicate that the value is no longer live.
415 store %Object* null, %Object** %X
421 <!-- ======================================================================= -->
422 <div class="doc_subsection">
423 <a name="barriers">Reading and writing references in the heap</a>
426 <div class="doc_text">
428 <p>Some collectors need to be informed when the mutator (the program that needs
429 garbage collection) either reads a pointer from or writes a pointer to a field
430 of a heap object. The code fragments inserted at these points are called
431 <em>read barriers</em> and <em>write barriers</em>, respectively. The amount of
432 code that needs to be executed is usually quite small and not on the critical
433 path of any computation, so the overall performance impact of the barrier is
436 <p>Barriers often require access to the <em>object pointer</em> rather than the
437 <em>derived pointer</em> (which is a pointer to the field within the
438 object). Accordingly, these intrinsics take both pointers as separate arguments
439 for completeness. In this snippet, <tt>%object</tt> is the object pointer, and
440 <tt>%derived</tt> is the derived pointer:</p>
444 %class.Array = type { %class.Object, i32, [0 x %class.Object*] }
447 ;; Load the object pointer from a gcroot.
448 %object = load %class.Array** %object_addr
450 ;; Compute the derived pointer.
451 %derived = getelementptr %object, i32 0, i32 2, i32 %n</pre></blockquote>
455 <!-- ======================================================================= -->
456 <div class="doc_subsubsection">
457 <a name="gcwrite">Write barrier: <tt>llvm.gcwrite</tt></a>
460 <div class="doc_code"><tt>
461 void @llvm.gcwrite(i8* %value, i8* %object, i8** %derived)
464 <div class="doc_text">
466 <p>For write barriers, LLVM provides the <tt>llvm.gcwrite</tt> intrinsic
467 function. It has exactly the same semantics as a non-volatile <tt>store</tt> to
468 the derived pointer (the third argument).</p>
470 <p>Many important algorithms require write barriers, including generational
471 and concurrent collectors. Additionally, write barriers could be used to
472 implement reference counting.</p>
474 <p>The use of this intrinsic is optional if the target collector does use
475 write barriers. If so, the collector will replace it with the corresponding
480 <!-- ======================================================================= -->
481 <div class="doc_subsubsection">
482 <a name="gcread">Read barrier: <tt>llvm.gcread</tt></a>
485 <div class="doc_code"><tt>
486 i8* @llvm.gcread(i8* %object, i8** %derived)<br>
489 <div class="doc_text">
491 <p>For read barriers, LLVM provides the <tt>llvm.gcread</tt> intrinsic function.
492 It has exactly the same semantics as a non-volatile <tt>load</tt> from the
493 derived pointer (the second argument).</p>
495 <p>Read barriers are needed by fewer algorithms than write barriers, and may
496 have a greater performance impact since pointer reads are more frequent than
499 <p>As with <tt>llvm.gcwrite</tt>, a target collector might not require the use
500 of this intrinsic.</p>
504 <!-- *********************************************************************** -->
505 <div class="doc_section">
506 <a name="runtime">Recommended runtime interface</a>
508 <!-- *********************************************************************** -->
510 <div class="doc_text">
512 <p>LLVM specifies the following recommended runtime interface to the garbage
513 collection at runtime. A program should use these interfaces to accomplish the
514 tasks not supported by the intrinsics.</p>
516 <p>Unlike the intrinsics, which are integral to LLVM's code generator, there is
517 nothing unique about these interfaces; a front-end compiler and runtime are free
518 to agree to a different specification.</p>
520 <p class="doc_warning">Note: This interface is a work in progress.</p>
524 <!-- ======================================================================= -->
525 <div class="doc_subsection">
526 <a name="initialize">Garbage collector startup and initialization</a>
529 <div class="doc_text">
531 <div class="doc_code"><tt>
532 void llvm_gc_initialize(unsigned InitialHeapSize);
536 The <tt>llvm_gc_initialize</tt> function should be called once before any other
537 garbage collection functions are called. This gives the garbage collector the
538 chance to initialize itself and allocate the heap. The initial heap size to
539 allocate should be specified as an argument.
544 <!-- ======================================================================= -->
545 <div class="doc_subsection">
546 <a name="allocate">Allocating memory from the GC</a>
549 <div class="doc_text">
551 <div class="doc_code"><tt>
552 void *llvm_gc_allocate(unsigned Size);
555 <p>The <tt>llvm_gc_allocate</tt> function is a global function defined by the
556 garbage collector implementation to allocate memory. It returns a
557 zeroed-out block of memory of the specified size, sufficiently aligned to store
562 <!-- ======================================================================= -->
563 <div class="doc_subsection">
564 <a name="explicit">Explicit invocation of the garbage collector</a>
567 <div class="doc_text">
569 <div class="doc_code"><tt>
570 void llvm_gc_collect();
574 The <tt>llvm_gc_collect</tt> function is exported by the garbage collector
575 implementations to provide a full collection, even when the heap is not
576 exhausted. This can be used by end-user code as a hint, and may be ignored by
577 the garbage collector.
582 <!-- ======================================================================= -->
583 <div class="doc_subsection">
584 <a name="traceroots">Tracing GC pointers from the program stack</a>
587 <div class="doc_text">
588 <div class="doc_code"><tt>
589 void llvm_cg_walk_gcroots(void (*FP)(void **Root, void *Meta));
593 The <tt>llvm_cg_walk_gcroots</tt> function is a function provided by the code
594 generator that iterates through all of the GC roots on the stack, calling the
595 specified function pointer with each record. For each GC root, the address of
596 the pointer and the meta-data (from the <a
597 href="#gcroot"><tt>llvm.gcroot</tt></a> intrinsic) are provided.
601 <!-- ======================================================================= -->
602 <div class="doc_subsection">
603 <a name="staticroots">Tracing GC pointers from static roots</a>
606 <div class="doc_text">
611 <!-- *********************************************************************** -->
612 <div class="doc_section">
613 <a name="plugin">Implementing a collector plugin</a>
615 <!-- *********************************************************************** -->
617 <div class="doc_text">
619 <p>User code specifies which collector plugin to use with the <tt>gc</tt>
620 function attribute or, equivalently, with the <tt>setCollector</tt> method of
621 <tt>Function</tt>.</p>
623 <p>To implement a collector plugin, it is necessary to subclass
624 <tt>llvm::Collector</tt>, which can be accomplished in a few lines of
625 boilerplate code. LLVM's infrastructure provides access to several important
626 algorithms. For an uncontroversial collector, all that remains may be to emit
627 the assembly code for the collector's unique stack map data structure, which
628 might be accomplished in as few as 100 LOC.</p>
630 <p>To subclass <tt>llvm::Collector</tt> and register a collector:</p>
632 <blockquote><pre>// lib/MyGC/MyGC.cpp - Example LLVM collector plugin
634 #include "llvm/CodeGen/Collector.h"
635 #include "llvm/CodeGen/Collectors.h"
636 #include "llvm/CodeGen/CollectorMetadata.h"
637 #include "llvm/Support/Compiler.h"
639 using namespace llvm;
642 class VISIBILITY_HIDDEN MyCollector : public Collector {
647 CollectorRegistry::Add<MyCollector>
648 X("mygc", "My bespoke garbage collector.");
651 <p>Using the LLVM makefiles (like the <a
652 href="http://llvm.org/viewvc/llvm-project/llvm/trunk/projects/sample/">sample
653 project</a>), this can be built into a plugin using a simple makefile:</p>
659 LIBRARYNAME = <var>MyGC</var>
662 include $(LEVEL)/Makefile.common</pre></blockquote>
664 <p>Once the plugin is compiled, code using it may be compiled using <tt>llc
665 -load=<var>MyGC.so</var></tt> (though <var>MyGC.so</var> may have some other
666 platform-specific extension):</p>
670 define void @f() gc "mygc" {
674 $ llvm-as < sample.ll | llc -load=MyGC.so</pre></blockquote>
676 <p>It is also possible to statically link the collector plugin into tools, such
677 as a language-specific compiler front-end.</p>
681 <!-- ======================================================================= -->
682 <div class="doc_subsection">
683 <a name="collector-algos">Overview of available features</a>
686 <div class="doc_text">
688 <p>The boilerplate collector above does nothing. More specifically:</p>
691 <li><tt>llvm.gcread</tt> calls are replaced with the corresponding
692 <tt>load</tt> instruction.</li>
693 <li><tt>llvm.gcwrite</tt> calls are replaced with the corresponding
694 <tt>store</tt> instruction.</li>
695 <li>No stack map is emitted, and no safe points are added.</li>
698 <p><tt>Collector</tt> provides a range of features through which a plugin
699 collector may do useful work. This matrix summarizes the supported (and planned)
700 features and correlates them with the collection techniques which typically
707 <th>shadow stack</th>
716 <th class="rowhead"><a href="#stack-map">stack map</a></th>
727 <th class="rowhead"><a href="#init-roots">initialize roots</a></th>
737 <tr class="doc_warning">
738 <th class="rowhead">derived pointers</th>
749 <th class="rowhead"><em><a href="#custom">custom lowering</a></em></th>
760 <th class="rowhead indent">gcroot</th>
771 <th class="rowhead indent">gcwrite</th>
782 <th class="rowhead indent">gcread</th>
793 <th class="rowhead"><em><a href="#safe-points">safe points</a></em></th>
804 <th class="rowhead indent">in calls</th>
815 <th class="rowhead indent">before calls</th>
825 <tr class="doc_warning">
826 <th class="rowhead indent">for loops</th>
837 <th class="rowhead indent">before escape</th>
847 <tr class="doc_warning">
848 <th class="rowhead">emit code at safe points</th>
859 <th class="rowhead"><em>output</em></th>
870 <th class="rowhead indent"><a href="#assembly">assembly</a></th>
880 <tr class="doc_warning">
881 <th class="rowhead indent">JIT</th>
885 <td class="optl">✘</td>
886 <td class="optl">✘</td>
887 <td class="optl">✘</td>
888 <td class="optl">✘</td>
889 <td class="optl">✘</td>
891 <tr class="doc_warning">
892 <th class="rowhead indent">obj</th>
896 <td class="optl">✘</td>
897 <td class="optl">✘</td>
898 <td class="optl">✘</td>
899 <td class="optl">✘</td>
900 <td class="optl">✘</td>
902 <tr class="doc_warning">
903 <th class="rowhead">live analysis</th>
907 <td class="optl">✘</td>
908 <td class="optl">✘</td>
909 <td class="optl">✘</td>
910 <td class="optl">✘</td>
911 <td class="optl">✘</td>
913 <tr class="doc_warning">
914 <th class="rowhead">register map</th>
918 <td class="optl">✘</td>
919 <td class="optl">✘</td>
920 <td class="optl">✘</td>
921 <td class="optl">✘</td>
922 <td class="optl">✘</td>
926 <div><span class="doc_warning">*</span> Derived pointers only pose a
927 hazard to copying collectors.</div>
928 <div><span class="optl">✘</span> in gray denotes a feature which
929 could be utilized if available.</div>
934 <p>To be clear, the collection techniques above are defined as:</p>
937 <dt>Shadow Stack</dt>
938 <dd>The mutator carefully maintains a linked list of stack root
940 <dt>Reference Counting</dt>
941 <dd>The mutator maintains a reference count for each object and frees an
942 object when its count falls to zero.</dd>
944 <dd>When the heap is exhausted, the collector marks reachable objects starting
945 from the roots, then deallocates unreachable objects in a sweep
948 <dd>As reachability analysis proceeds, the collector copies objects from one
949 heap area to another, compacting them in the process. Copying collectors
950 enable highly efficient "bump pointer" allocation and can improve locality
953 <dd>(Including generational collectors.) Incremental collectors generally have
954 all the properties of a copying collector (regardless of whether the
955 mature heap is compacting), but bring the added complexity of requiring
958 <dd>Denotes a multithreaded mutator; the collector must still stop the mutator
959 ("stop the world") before beginning reachability analysis. Stopping a
960 multithreaded mutator is a complicated problem. It generally requires
961 highly platform specific code in the runtime, and the production of
962 carefully designed machine code at safe points.</dd>
964 <dd>In this technique, the mutator and the collector run concurrently, with
965 the goal of eliminating pause times. In a <em>cooperative</em> collector,
966 the mutator further aids with collection should a pause occur, allowing
967 collection to take advantage of multiprocessor hosts. The "stop the world"
968 problem of threaded collectors is generally still present to a limited
969 extent. Sophisticated marking algorithms are necessary. Read barriers may
973 <p>As the matrix indicates, LLVM's garbage collection infrastructure is already
974 suitable for a wide variety of collectors, but does not currently extend to
975 multithreaded programs. This will be added in the future as there is
980 <!-- ======================================================================= -->
981 <div class="doc_subsection">
982 <a name="stack-map">Computing stack maps</a>
985 <div class="doc_text">
988 >for (iterator I = begin(), E = end(); I != E; ++I) {
989 CollectorMetadata *MD = *I;
990 unsigned FrameSize = MD->getFrameSize();
991 size_t RootCount = MD->roots_size();
993 for (CollectorMetadata::roots_iterator RI = MD->roots_begin(),
994 RE = MD->roots_end();
996 int RootNum = RI->Num;
997 int RootStackOffset = RI->StackOffset;
998 Constant *RootMetadata = RI->Metadata;
1000 }</pre></blockquote>
1002 <p>LLVM automatically computes a stack map. All a <tt>Collector</tt> needs to do
1003 is access it using <tt>CollectorMetadata::roots_begin()</tt> and
1004 -<tt>end()</tt>. If the <tt>llvm.gcroot</tt> intrinsic is eliminated before code
1005 generation by a custom lowering pass, LLVM's stack map will be empty.</p>
1010 <!-- ======================================================================= -->
1011 <div class="doc_subsection">
1012 <a name="init-roots">Initializing roots to null: <tt>InitRoots</tt></a>
1015 <div class="doc_text">
1018 >MyCollector::MyCollector() {
1020 }</pre></blockquote>
1022 <p>When set, LLVM will automatically initialize each root to <tt>null</tt> upon
1023 entry to the function. This prevents the reachability analysis from finding
1024 uninitialized values in stack roots at runtime, which will almost certainly
1025 cause it to segfault. This initialization occurs before custom lowering, so the
1026 two may be used together.</p>
1028 <p>Since LLVM does not yet compute liveness information, this feature should be
1029 used by all collectors which do not custom lower <tt>llvm.gcroot</tt>, and even
1035 <!-- ======================================================================= -->
1036 <div class="doc_subsection">
1037 <a name="custom">Custom lowering of intrinsics: <tt>CustomRoots</tt>,
1038 <tt>CustomReadBarriers</tt>, and <tt>CustomWriteBarriers</tt></a>
1041 <div class="doc_text">
1043 <p>For collectors with barriers or unusual treatment of stack roots, these
1044 flags allow the collector to perform any required transformation on the LLVM
1048 >class MyCollector : public Collector {
1052 CustomReadBarriers = true;
1053 CustomWriteBarriers = true;
1056 virtual bool initializeCustomLowering(Module &M);
1057 virtual bool performCustomLowering(Function &F);
1058 };</pre></blockquote>
1060 <p>If any of these flags are set, then LLVM suppresses its default lowering for
1061 the corresponding intrinsics and instead passes them on to a custom lowering
1062 pass specified by the collector.</p>
1064 <p>LLVM's default action for each intrinsic is as follows:</p>
1067 <li><tt>llvm.gcroot</tt>: Pass through to the code generator to generate a
1069 <li><tt>llvm.gcread</tt>: Substitute a <tt>load</tt> instruction.</li>
1070 <li><tt>llvm.gcwrite</tt>: Substitute a <tt>store</tt> instruction.</li>
1073 <p>If <tt>CustomReadBarriers</tt> or <tt>CustomWriteBarriers</tt> are specified,
1074 then <tt>performCustomLowering</tt> <strong>must</strong> eliminate the
1075 corresponding barriers.</p>
1077 <p><tt>performCustomLowering</tt>, must comply with the same restrictions as <a
1078 href="WritingAnLLVMPass.html#runOnFunction"><tt>runOnFunction</tt></a>, and
1079 that <tt>initializeCustomLowering</tt> has the same semantics as <a
1080 href="WritingAnLLVMPass.html#doInitialization_mod"><tt>doInitialization(Module
1081 &)</tt></a>.</p>
1083 <p>The following can be used as a template:</p>
1086 >#include "llvm/Module.h"
1087 #include "llvm/IntrinsicInst.h"
1089 bool MyCollector::initializeCustomLowering(Module &M) {
1093 bool MyCollector::performCustomLowering(Function &F) {
1094 bool MadeChange = false;
1096 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1097 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; )
1098 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(II++))
1099 if (Function *F = CI->getCalledFunction())
1100 switch (F->getIntrinsicID()) {
1101 case Intrinsic::gcwrite:
1102 // Handle llvm.gcwrite.
1103 CI->eraseFromParent();
1106 case Intrinsic::gcread:
1107 // Handle llvm.gcread.
1108 CI->eraseFromParent();
1111 case Intrinsic::gcroot:
1112 // Handle llvm.gcroot.
1113 CI->eraseFromParent();
1119 }</pre></blockquote>
1124 <!-- ======================================================================= -->
1125 <div class="doc_subsection">
1126 <a name="safe-points">Generating safe points: <tt>NeededSafePoints</tt></a>
1129 <div class="doc_text">
1131 <p>LLVM can compute four kinds of safe points:</p>
1135 /// PointKind - The type of a collector-safe point.
1138 Loop, //< Instr is a loop (backwards branch).
1139 Return, //< Instr is a return instruction.
1140 PreCall, //< Instr is a call instruction.
1141 PostCall //< Instr is the return address of a call.
1143 }</pre></blockquote>
1145 <p>A collector can request any combination of the four by setting the
1146 <tt>NeededSafePoints</tt> mask:</p>
1149 >MyCollector::MyCollector() {
1150 NeededSafePoints = 1 << GC::Loop
1151 | 1 << GC::Return
1152 | 1 << GC::PreCall
1153 | 1 << GC::PostCall;
1154 }</pre></blockquote>
1156 <p>It can then use the following routines to access safe points.</p>
1159 >for (iterator I = begin(), E = end(); I != E; ++I) {
1160 CollectorMetadata *MD = *I;
1161 size_t PointCount = MD->size();
1163 for (CollectorMetadata::iterator PI = MD->begin(),
1164 PE = MD->end(); PI != PE; ++PI) {
1165 GC::PointKind PointKind = PI->Kind;
1166 unsigned PointNum = PI->Num;
1171 <p>Almost every collector requires <tt>PostCall</tt> safe points, since these
1172 correspond to the moments when the function is suspended during a call to a
1175 <p>Threaded programs generally require <tt>Loop</tt> safe points to guarantee
1176 that the application will reach a safe point within a bounded amount of time,
1177 even if it is executing a long-running loop which contains no function
1180 <p>Threaded collectors may also require <tt>Return</tt> and <tt>PreCall</tt>
1181 safe points to implement "stop the world" techniques using self-modifying code,
1182 where it is important that the program not exit the function without reaching a
1183 safe point (because only the topmost function has been patched).</p>
1188 <!-- ======================================================================= -->
1189 <div class="doc_subsection">
1190 <a name="assembly">Emitting assembly code:
1191 <tt>beginAssembly</tt> and <tt>finishAssembly</tt></a>
1194 <div class="doc_text">
1196 <p>LLVM allows a collector to print arbitrary assembly code before and after
1197 the rest of a module's assembly code. From the latter callback, the collector
1198 can print stack maps built by the code generator.</p>
1200 <p>Note that LLVM does not currently have analogous APIs to support code
1201 generation in the JIT, nor using the object writers.</p>
1204 >class MyCollector : public Collector {
1206 virtual void beginAssembly(std::ostream &OS, AsmPrinter &AP,
1207 const TargetAsmInfo &TAI);
1209 virtual void finishAssembly(std::ostream &OS, AsmPrinter &AP,
1210 const TargetAsmInfo &TAI);
1211 }</pre></blockquote>
1213 <p>The collector should use <tt>AsmPrinter</tt> and <tt>TargetAsmInfo</tt> to
1214 print portable assembly code to the <tt>std::ostream</tt>. The collector itself
1215 contains the stack map for the entire module, and may access the
1216 <tt>CollectorMetadata</tt> using its own <tt>begin()</tt> and <tt>end()</tt>
1217 methods. Here's a realistic example:</p>
1220 >#include "llvm/CodeGen/AsmPrinter.h"
1221 #include "llvm/Function.h"
1222 #include "llvm/Target/TargetMachine.h"
1223 #include "llvm/Target/TargetData.h"
1224 #include "llvm/Target/TargetAsmInfo.h"
1226 void MyCollector::beginAssembly(std::ostream &OS, AsmPrinter &AP,
1227 const TargetAsmInfo &TAI) {
1231 void MyCollector::finishAssembly(std::ostream &OS, AsmPrinter &AP,
1232 const TargetAsmInfo &TAI) {
1233 // Set up for emitting addresses.
1234 const char *AddressDirective;
1235 int AddressAlignLog;
1236 if (AP.TM.getTargetData()->getPointerSize() == sizeof(int32_t)) {
1237 AddressDirective = TAI.getData32bitsDirective();
1238 AddressAlignLog = 2;
1240 AddressDirective = TAI.getData64bitsDirective();
1241 AddressAlignLog = 3;
1244 // Put this in the data section.
1245 AP.SwitchToDataSection(TAI.getDataSection());
1247 // For each function...
1248 for (iterator FI = begin(), FE = end(); FI != FE; ++FI) {
1249 CollectorMetadata &MD = **FI;
1251 // Emit this data structure:
1254 // int32_t PointCount;
1256 // void *SafePointAddress;
1257 // int32_t LiveCount;
1258 // int32_t LiveOffsets[LiveCount];
1259 // } Points[PointCount];
1260 // } __gcmap_<FUNCTIONNAME>;
1262 // Align to address width.
1263 AP.EmitAlignment(AddressAlignLog);
1265 // Emit the symbol by which the stack map can be found.
1267 Symbol += TAI.getGlobalPrefix();
1268 Symbol += "__gcmap_";
1269 Symbol += MD.getFunction().getName();
1270 if (const char *GlobalDirective = TAI.getGlobalDirective())
1271 OS << GlobalDirective << Symbol << "\n";
1272 OS << TAI.getGlobalPrefix() << Symbol << ":\n";
1275 AP.EmitInt32(MD.size());
1276 AP.EOL("safe point count");
1278 // And each safe point...
1279 for (CollectorMetadata::iterator PI = MD.begin(),
1280 PE = MD.end(); PI != PE; ++PI) {
1281 // Align to address width.
1282 AP.EmitAlignment(AddressAlignLog);
1284 // Emit the address of the safe point.
1285 OS << AddressDirective
1286 << TAI.getPrivateGlobalPrefix() << "label" << PI->Num;
1287 AP.EOL("safe point address");
1289 // Emit the stack frame size.
1290 AP.EmitInt32(MD.getFrameSize());
1291 AP.EOL("stack frame size");
1293 // Emit the number of live roots in the function.
1294 AP.EmitInt32(MD.live_size(PI));
1295 AP.EOL("live root count");
1297 // And for each live root...
1298 for (CollectorMetadata::live_iterator LI = MD.live_begin(PI),
1299 LE = MD.live_end(PI);
1301 // Print its offset within the stack frame.
1302 AP.EmitInt32(LI->StackOffset);
1303 AP.EOL("stack offset");
1313 <!-- *********************************************************************** -->
1314 <div class="doc_section">
1315 <a name="runtime-impl">Implementing a collector runtime</a>
1317 <!-- *********************************************************************** -->
1319 <div class="doc_text">
1321 <p>Implementing a garbage collector for LLVM is fairly straightforward. The
1322 LLVM garbage collectors are provided in a form that makes them easy to link into
1323 the language-specific runtime that a language front-end would use. They require
1324 functionality from the language-specific runtime to get information about <a
1325 href="#gcdescriptors">where pointers are located in heap objects</a>.</p>
1327 <p>The implementation must include the
1328 <a href="#allocate"><tt>llvm_gc_allocate</tt></a> and
1329 <a href="#explicit"><tt>llvm_gc_collect</tt></a> functions. To do this, it will
1330 probably have to <a href="#traceroots">trace through the roots
1331 from the stack</a> and understand the <a href="#gcdescriptors">GC descriptors
1332 for heap objects</a>. Luckily, there are some <a href="#usage">example
1333 implementations</a> available.
1338 <!-- ======================================================================= -->
1339 <div class="doc_subsection">
1340 <a name="gcdescriptors">Tracing GC pointers from heap objects</a>
1343 <div class="doc_text">
1345 The three most common ways to keep track of where pointers live in heap objects
1346 are (listed in order of space overhead required):</p>
1349 <li>In languages with polymorphic objects, pointers from an object header are
1350 usually used to identify the GC pointers in the heap object. This is common for
1351 object-oriented languages like Self, Smalltalk, Java, or C#.</li>
1353 <li>If heap objects are not polymorphic, often the "shape" of the heap can be
1354 determined from the roots of the heap or from some other meta-data [<a
1355 href="#appel89">Appel89</a>, <a href="#goldberg91">Goldberg91</a>, <a
1356 href="#tolmach94">Tolmach94</a>]. In this case, the garbage collector can
1357 propagate the information around from meta data stored with the roots. This
1358 often eliminates the need to have a header on objects in the heap. This is
1359 common in the ML family.</li>
1361 <li>If all heap objects have pointers in the same locations, or pointers can be
1362 distinguished just by looking at them (e.g., the low order bit is clear), no
1363 book-keeping is needed at all. This is common for Lisp-like languages.</li>
1366 <p>The LLVM garbage collectors are capable of supporting all of these styles of
1367 language, including ones that mix various implementations. To do this, it
1368 allows the source-language to associate meta-data with the <a
1369 href="#gcroot">stack roots</a>, and the heap tracing routines can propagate the
1370 information. In addition, LLVM allows the front-end to extract GC information
1371 in any form from a specific object pointer (this supports situations #1 and #3).
1377 <!-- *********************************************************************** -->
1378 <div class="doc_section">
1379 <a name="references">References</a>
1381 <!-- *********************************************************************** -->
1383 <div class="doc_text">
1385 <p><a name="appel89">[Appel89]</a> Runtime Tags Aren't Necessary. Andrew
1386 W. Appel. Lisp and Symbolic Computation 19(7):703-705, July 1989.</p>
1388 <p><a name="goldberg91">[Goldberg91]</a> Tag-free garbage collection for
1389 strongly typed programming languages. Benjamin Goldberg. ACM SIGPLAN
1392 <p><a name="tolmach94">[Tolmach94]</a> Tag-free garbage collection using
1393 explicit type parameters. Andrew Tolmach. Proceedings of the 1994 ACM
1394 conference on LISP and functional programming.</p>
1396 <p><a name="henderson02">[Henderson2002]</a> <a
1397 href="http://citeseer.ist.psu.edu/henderson02accurate.html">
1398 Accurate Garbage Collection in an Uncooperative Environment</a>.
1399 Fergus Henderson. International Symposium on Memory Management 2002.</p>
1404 <!-- *********************************************************************** -->
1408 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1409 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
1410 <a href="http://validator.w3.org/check/referer"><img
1411 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!"></a>
1413 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1414 <a href="http://llvm.org">LLVM Compiler Infrastructure</a><br>
1415 Last modified: $Date$