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="_static/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; }
17 Accurate Garbage Collection with LLVM
21 <li><a href="#introduction">Introduction</a>
23 <li><a href="#feature">Goals and non-goals</a></li>
27 <li><a href="#quickstart">Getting started</a>
29 <li><a href="#quickstart-compiler">In your compiler</a></li>
30 <li><a href="#quickstart-runtime">In your runtime library</a></li>
31 <li><a href="#shadow-stack">About the shadow stack</a></li>
35 <li><a href="#core">Core support</a>
37 <li><a href="#gcattr">Specifying GC code generation:
38 <tt>gc "..."</tt></a></li>
39 <li><a href="#gcroot">Identifying GC roots on the stack:
40 <tt>llvm.gcroot</tt></a></li>
41 <li><a href="#barriers">Reading and writing references in the heap</a>
43 <li><a href="#gcwrite">Write barrier: <tt>llvm.gcwrite</tt></a></li>
44 <li><a href="#gcread">Read barrier: <tt>llvm.gcread</tt></a></li>
50 <li><a href="#plugin">Compiler plugin interface</a>
52 <li><a href="#collector-algos">Overview of available features</a></li>
53 <li><a href="#stack-map">Computing stack maps</a></li>
54 <li><a href="#init-roots">Initializing roots to null:
55 <tt>InitRoots</tt></a></li>
56 <li><a href="#custom">Custom lowering of intrinsics: <tt>CustomRoots</tt>,
57 <tt>CustomReadBarriers</tt>, and <tt>CustomWriteBarriers</tt></a></li>
58 <li><a href="#safe-points">Generating safe points:
59 <tt>NeededSafePoints</tt></a></li>
60 <li><a href="#assembly">Emitting assembly code:
61 <tt>GCMetadataPrinter</tt></a></li>
65 <li><a href="#runtime-impl">Implementing a collector runtime</a>
67 <li><a href="#gcdescriptors">Tracing GC pointers from heap
72 <li><a href="#references">References</a></li>
76 <div class="doc_author">
77 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a> and
81 <!-- *********************************************************************** -->
83 <a name="introduction">Introduction</a>
85 <!-- *********************************************************************** -->
89 <p>Garbage collection is a widely used technique that frees the programmer from
90 having to know the lifetimes of heap objects, making software easier to produce
91 and maintain. Many programming languages rely on garbage collection for
92 automatic memory management. There are two primary forms of garbage collection:
93 conservative and accurate.</p>
95 <p>Conservative garbage collection often does not require any special support
96 from either the language or the compiler: it can handle non-type-safe
97 programming languages (such as C/C++) and does not require any special
98 information from the compiler. The
99 <a href="http://www.hpl.hp.com/personal/Hans_Boehm/gc/">Boehm collector</a> is
100 an example of a state-of-the-art conservative collector.</p>
102 <p>Accurate garbage collection requires the ability to identify all pointers in
103 the program at run-time (which requires that the source-language be type-safe in
104 most cases). Identifying pointers at run-time requires compiler support to
105 locate all places that hold live pointer variables at run-time, including the
106 <a href="#gcroot">processor stack and registers</a>.</p>
108 <p>Conservative garbage collection is attractive because it does not require any
109 special compiler support, but it does have problems. In particular, because the
110 conservative garbage collector cannot <i>know</i> that a particular word in the
111 machine is a pointer, it cannot move live objects in the heap (preventing the
112 use of compacting and generational GC algorithms) and it can occasionally suffer
113 from memory leaks due to integer values that happen to point to objects in the
114 program. In addition, some aggressive compiler transformations can break
115 conservative garbage collectors (though these seem rare in practice).</p>
117 <p>Accurate garbage collectors do not suffer from any of these problems, but
118 they can suffer from degraded scalar optimization of the program. In particular,
119 because the runtime must be able to identify and update all pointers active in
120 the program, some optimizations are less effective. In practice, however, the
121 locality and performance benefits of using aggressive garbage collection
122 techniques dominates any low-level losses.</p>
124 <p>This document describes the mechanisms and interfaces provided by LLVM to
125 support accurate garbage collection.</p>
127 <!-- ======================================================================= -->
129 <a name="feature">Goals and non-goals</a>
134 <p>LLVM's intermediate representation provides <a href="#intrinsics">garbage
135 collection intrinsics</a> that offer support for a broad class of
136 collector models. For instance, the intrinsics permit:</p>
139 <li>semi-space collectors</li>
140 <li>mark-sweep collectors</li>
141 <li>generational collectors</li>
142 <li>reference counting</li>
143 <li>incremental collectors</li>
144 <li>concurrent collectors</li>
145 <li>cooperative collectors</li>
148 <p>We hope that the primitive support built into the LLVM IR is sufficient to
149 support a broad class of garbage collected languages including Scheme, ML, Java,
150 C#, Perl, Python, Lua, Ruby, other scripting languages, and more.</p>
152 <p>However, LLVM does not itself provide a garbage collector—this should
153 be part of your language's runtime library. LLVM provides a framework for
154 compile time <a href="#plugin">code generation plugins</a>. The role of these
155 plugins is to generate code and data structures which conforms to the <em>binary
156 interface</em> specified by the <em>runtime library</em>. This is similar to the
157 relationship between LLVM and DWARF debugging info, for example. The
158 difference primarily lies in the lack of an established standard in the domain
159 of garbage collection—thus the plugins.</p>
161 <p>The aspects of the binary interface with which LLVM's GC support is
165 <li>Creation of GC-safe points within code where collection is allowed to
167 <li>Computation of the stack map. For each safe point in the code, object
168 references within the stack frame must be identified so that the
169 collector may traverse and perhaps update them.</li>
170 <li>Write barriers when storing object references to the heap. These are
171 commonly used to optimize incremental scans in generational
173 <li>Emission of read barriers when loading object references. These are
174 useful for interoperating with concurrent collectors.</li>
177 <p>There are additional areas that LLVM does not directly address:</p>
180 <li>Registration of global roots with the runtime.</li>
181 <li>Registration of stack map entries with the runtime.</li>
182 <li>The functions used by the program to allocate memory, trigger a
183 collection, etc.</li>
184 <li>Computation or compilation of type maps, or registration of them with
185 the runtime. These are used to crawl the heap for object
189 <p>In general, LLVM's support for GC does not include features which can be
190 adequately addressed with other features of the IR and does not specify a
191 particular binary interface. On the plus side, this means that you should be
192 able to integrate LLVM with an existing runtime. On the other hand, it leaves
193 a lot of work for the developer of a novel language. However, it's easy to get
194 started quickly and scale up to a more sophisticated implementation as your
195 compiler matures.</p>
201 <!-- *********************************************************************** -->
203 <a name="quickstart">Getting started</a>
205 <!-- *********************************************************************** -->
209 <p>Using a GC with LLVM implies many things, for example:</p>
212 <li>Write a runtime library or find an existing one which implements a GC
214 <li>Implement a memory allocator.</li>
215 <li>Design a binary interface for the stack map, used to identify
216 references within a stack frame on the machine stack.*</li>
217 <li>Implement a stack crawler to discover functions on the call stack.*</li>
218 <li>Implement a registry for global roots.</li>
219 <li>Design a binary interface for type maps, used to identify references
220 within heap objects.</li>
221 <li>Implement a collection routine bringing together all of the above.</li>
223 <li>Emit compatible code from your compiler.<ul>
224 <li>Initialization in the main function.</li>
225 <li>Use the <tt>gc "..."</tt> attribute to enable GC code generation
226 (or <tt>F.setGC("...")</tt>).</li>
227 <li>Use <tt>@llvm.gcroot</tt> to mark stack roots.</li>
228 <li>Use <tt>@llvm.gcread</tt> and/or <tt>@llvm.gcwrite</tt> to
229 manipulate GC references, if necessary.</li>
230 <li>Allocate memory using the GC allocation routine provided by the
231 runtime library.</li>
232 <li>Generate type maps according to your runtime's binary interface.</li>
234 <li>Write a compiler plugin to interface LLVM with the runtime library.*<ul>
235 <li>Lower <tt>@llvm.gcread</tt> and <tt>@llvm.gcwrite</tt> to appropriate
236 code sequences.*</li>
237 <li>Compile LLVM's stack map to the binary form expected by the
240 <li>Load the plugin into the compiler. Use <tt>llc -load</tt> or link the
241 plugin statically with your language's compiler.*</li>
242 <li>Link program executables with the runtime.</li>
245 <p>To help with several of these tasks (those indicated with a *), LLVM
246 includes a highly portable, built-in ShadowStack code generator. It is compiled
247 into <tt>llc</tt> and works even with the interpreter and C backends.</p>
249 <!-- ======================================================================= -->
251 <a name="quickstart-compiler">In your compiler</a>
256 <p>To turn the shadow stack on for your functions, first call:</p>
258 <div class="doc_code"><pre
259 >F.setGC("shadow-stack");</pre></div>
261 <p>for each function your compiler emits. Since the shadow stack is built into
262 LLVM, you do not need to load a plugin.</p>
264 <p>Your compiler must also use <tt>@llvm.gcroot</tt> as documented.
265 Don't forget to create a root for each intermediate value that is generated
266 when evaluating an expression. In <tt>h(f(), g())</tt>, the result of
267 <tt>f()</tt> could easily be collected if evaluating <tt>g()</tt> triggers a
270 <p>There's no need to use <tt>@llvm.gcread</tt> and <tt>@llvm.gcwrite</tt> over
271 plain <tt>load</tt> and <tt>store</tt> for now. You will need them when
272 switching to a more advanced GC.</p>
276 <!-- ======================================================================= -->
278 <a name="quickstart-runtime">In your runtime</a>
283 <p>The shadow stack doesn't imply a memory allocation algorithm. A semispace
284 collector or building atop <tt>malloc</tt> are great places to start, and can
285 be implemented with very little code.</p>
287 <p>When it comes time to collect, however, your runtime needs to traverse the
288 stack roots, and for this it needs to integrate with the shadow stack. Luckily,
289 doing so is very simple. (This code is heavily commented to help you
290 understand the data structure, but there are only 20 lines of meaningful
293 <pre class="doc_code">
294 /// @brief The map for a single function's stack frame. One of these is
295 /// compiled as constant data into the executable for each function.
297 /// Storage of metadata values is elided if the %metadata parameter to
298 /// @llvm.gcroot is null.
300 int32_t NumRoots; //< Number of roots in stack frame.
301 int32_t NumMeta; //< Number of metadata entries. May be < NumRoots.
302 const void *Meta[0]; //< Metadata for each root.
305 /// @brief A link in the dynamic shadow stack. One of these is embedded in the
306 /// stack frame of each function on the call stack.
308 StackEntry *Next; //< Link to next stack entry (the caller's).
309 const FrameMap *Map; //< Pointer to constant FrameMap.
310 void *Roots[0]; //< Stack roots (in-place array).
313 /// @brief The head of the singly-linked list of StackEntries. Functions push
314 /// and pop onto this in their prologue and epilogue.
316 /// Since there is only a global list, this technique is not threadsafe.
317 StackEntry *llvm_gc_root_chain;
319 /// @brief Calls Visitor(root, meta) for each GC root on the stack.
320 /// root and meta are exactly the values passed to
321 /// <tt>@llvm.gcroot</tt>.
323 /// Visitor could be a function to recursively mark live objects. Or it
324 /// might copy them to another heap or generation.
326 /// @param Visitor A function to invoke for every GC root on the stack.
327 void visitGCRoots(void (*Visitor)(void **Root, const void *Meta)) {
328 for (StackEntry *R = llvm_gc_root_chain; R; R = R->Next) {
331 // For roots [0, NumMeta), the metadata pointer is in the FrameMap.
332 for (unsigned e = R->Map->NumMeta; i != e; ++i)
333 Visitor(&R->Roots[i], R->Map->Meta[i]);
335 // For roots [NumMeta, NumRoots), the metadata pointer is null.
336 for (unsigned e = R->Map->NumRoots; i != e; ++i)
337 Visitor(&R->Roots[i], NULL);
343 <!-- ======================================================================= -->
345 <a name="shadow-stack">About the shadow stack</a>
350 <p>Unlike many GC algorithms which rely on a cooperative code generator to
351 compile stack maps, this algorithm carefully maintains a linked list of stack
352 roots [<a href="#henderson02">Henderson2002</a>]. This so-called "shadow stack"
353 mirrors the machine stack. Maintaining this data structure is slower than using
354 a stack map compiled into the executable as constant data, but has a significant
355 portability advantage because it requires no special support from the target
356 code generator, and does not require tricky platform-specific code to crawl
357 the machine stack.</p>
359 <p>The tradeoff for this simplicity and portability is:</p>
362 <li>High overhead per function call.</li>
363 <li>Not thread-safe.</li>
366 <p>Still, it's an easy way to get started. After your compiler and runtime are
367 up and running, writing a <a href="#plugin">plugin</a> will allow you to take
368 advantage of <a href="#collector-algos">more advanced GC features</a> of LLVM
369 in order to improve performance.</p>
375 <!-- *********************************************************************** -->
377 <a name="core">IR features</a><a name="intrinsics"></a>
379 <!-- *********************************************************************** -->
383 <p>This section describes the garbage collection facilities provided by the
384 <a href="LangRef.html">LLVM intermediate representation</a>. The exact behavior
385 of these IR features is specified by the binary interface implemented by a
386 <a href="#plugin">code generation plugin</a>, not by this document.</p>
388 <p>These facilities are limited to those strictly necessary; they are not
389 intended to be a complete interface to any garbage collector. A program will
390 need to interface with the GC library using the facilities provided by that
393 <!-- ======================================================================= -->
395 <a name="gcattr">Specifying GC code generation: <tt>gc "..."</tt></a>
400 <div class="doc_code"><tt>
401 define <i>ty</i> @<i>name</i>(...) <span style="text-decoration: underline">gc "<i>name</i>"</span> { ...
404 <p>The <tt>gc</tt> function attribute is used to specify the desired GC style
405 to the compiler. Its programmatic equivalent is the <tt>setGC</tt> method of
406 <tt>Function</tt>.</p>
408 <p>Setting <tt>gc "<i>name</i>"</tt> on a function triggers a search for a
409 matching code generation plugin "<i>name</i>"; it is that plugin which defines
410 the exact nature of the code generated to support GC. If none is found, the
411 compiler will raise an error.</p>
413 <p>Specifying the GC style on a per-function basis allows LLVM to link together
414 programs that use different garbage collection algorithms (or none at all).</p>
418 <!-- ======================================================================= -->
420 <a name="gcroot">Identifying GC roots on the stack: <tt>llvm.gcroot</tt></a>
425 <div class="doc_code"><tt>
426 void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
429 <p>The <tt>llvm.gcroot</tt> intrinsic is used to inform LLVM that a stack
430 variable references an object on the heap and is to be tracked for garbage
431 collection. The exact impact on generated code is specified by a <a
432 href="#plugin">compiler plugin</a>. All calls to <tt>llvm.gcroot</tt> <b>must</b> reside
433 inside the first basic block.</p>
435 <p>A compiler which uses mem2reg to raise imperative code using <tt>alloca</tt>
436 into SSA form need only add a call to <tt>@llvm.gcroot</tt> for those variables
437 which a pointers into the GC heap.</p>
439 <p>It is also important to mark intermediate values with <tt>llvm.gcroot</tt>.
440 For example, consider <tt>h(f(), g())</tt>. Beware leaking the result of
441 <tt>f()</tt> in the case that <tt>g()</tt> triggers a collection. Note, that
442 stack variables must be initialized and marked with <tt>llvm.gcroot</tt> in
443 function's prologue.</p>
445 <p>The first argument <b>must</b> be a value referring to an alloca instruction
446 or a bitcast of an alloca. The second contains a pointer to metadata that
447 should be associated with the pointer, and <b>must</b> be a constant or global
448 value address. If your target collector uses tags, use a null pointer for
451 <p>The <tt>%metadata</tt> argument can be used to avoid requiring heap objects
452 to have 'isa' pointers or tag bits. [<a href="#appel89">Appel89</a>, <a
453 href="#goldberg91">Goldberg91</a>, <a href="#tolmach94">Tolmach94</a>] If
454 specified, its value will be tracked along with the location of the pointer in
457 <p>Consider the following fragment of Java code:</p>
459 <pre class="doc_code">
461 Object X; // A null-initialized reference to an object
466 <p>This block (which may be located in the middle of a function or in a loop
467 nest), could be compiled to this LLVM code:</p>
469 <pre class="doc_code">
471 ;; In the entry block for the function, allocate the
472 ;; stack space for X, which is an LLVM pointer.
475 ;; Tell LLVM that the stack space is a stack root.
476 ;; Java has type-tags on objects, so we pass null as metadata.
477 %tmp = bitcast %Object** %X to i8**
478 call void @llvm.gcroot(i8** %tmp, i8* null)
481 ;; "CodeBlock" is the block corresponding to the start
482 ;; of the scope above.
484 ;; Java null-initializes pointers.
485 store %Object* null, %Object** %X
489 ;; As the pointer goes out of scope, store a null value into
490 ;; it, to indicate that the value is no longer live.
491 store %Object* null, %Object** %X
497 <!-- ======================================================================= -->
499 <a name="barriers">Reading and writing references in the heap</a>
504 <p>Some collectors need to be informed when the mutator (the program that needs
505 garbage collection) either reads a pointer from or writes a pointer to a field
506 of a heap object. The code fragments inserted at these points are called
507 <em>read barriers</em> and <em>write barriers</em>, respectively. The amount of
508 code that needs to be executed is usually quite small and not on the critical
509 path of any computation, so the overall performance impact of the barrier is
512 <p>Barriers often require access to the <em>object pointer</em> rather than the
513 <em>derived pointer</em> (which is a pointer to the field within the
514 object). Accordingly, these intrinsics take both pointers as separate arguments
515 for completeness. In this snippet, <tt>%object</tt> is the object pointer, and
516 <tt>%derived</tt> is the derived pointer:</p>
520 %class.Array = type { %class.Object, i32, [0 x %class.Object*] }
523 ;; Load the object pointer from a gcroot.
524 %object = load %class.Array** %object_addr
526 ;; Compute the derived pointer.
527 %derived = getelementptr %object, i32 0, i32 2, i32 %n</pre></blockquote>
529 <p>LLVM does not enforce this relationship between the object and derived
530 pointer (although a <a href="#plugin">plugin</a> might). However, it would be
531 an unusual collector that violated it.</p>
533 <p>The use of these intrinsics is naturally optional if the target GC does
534 require the corresponding barrier. Such a GC plugin will replace the intrinsic
535 calls with the corresponding <tt>load</tt> or <tt>store</tt> instruction if they
538 <!-- ======================================================================= -->
540 <a name="gcwrite">Write barrier: <tt>llvm.gcwrite</tt></a>
545 <div class="doc_code"><tt>
546 void @llvm.gcwrite(i8* %value, i8* %object, i8** %derived)
549 <p>For write barriers, LLVM provides the <tt>llvm.gcwrite</tt> intrinsic
550 function. It has exactly the same semantics as a non-volatile <tt>store</tt> to
551 the derived pointer (the third argument). The exact code generated is specified
552 by a <a href="#plugin">compiler plugin</a>.</p>
554 <p>Many important algorithms require write barriers, including generational
555 and concurrent collectors. Additionally, write barriers could be used to
556 implement reference counting.</p>
560 <!-- ======================================================================= -->
562 <a name="gcread">Read barrier: <tt>llvm.gcread</tt></a>
567 <div class="doc_code"><tt>
568 i8* @llvm.gcread(i8* %object, i8** %derived)<br>
571 <p>For read barriers, LLVM provides the <tt>llvm.gcread</tt> intrinsic function.
572 It has exactly the same semantics as a non-volatile <tt>load</tt> from the
573 derived pointer (the second argument). The exact code generated is specified by
574 a <a href="#plugin">compiler plugin</a>.</p>
576 <p>Read barriers are needed by fewer algorithms than write barriers, and may
577 have a greater performance impact since pointer reads are more frequent than
586 <!-- *********************************************************************** -->
588 <a name="plugin">Implementing a collector plugin</a>
590 <!-- *********************************************************************** -->
594 <p>User code specifies which GC code generation to use with the <tt>gc</tt>
595 function attribute or, equivalently, with the <tt>setGC</tt> method of
596 <tt>Function</tt>.</p>
598 <p>To implement a GC plugin, it is necessary to subclass
599 <tt>llvm::GCStrategy</tt>, which can be accomplished in a few lines of
600 boilerplate code. LLVM's infrastructure provides access to several important
601 algorithms. For an uncontroversial collector, all that remains may be to
602 compile LLVM's computed stack map to assembly code (using the binary
603 representation expected by the runtime library). This can be accomplished in
604 about 100 lines of code.</p>
606 <p>This is not the appropriate place to implement a garbage collected heap or a
607 garbage collector itself. That code should exist in the language's runtime
608 library. The compiler plugin is responsible for generating code which
609 conforms to the binary interface defined by library, most essentially the
610 <a href="#stack-map">stack map</a>.</p>
612 <p>To subclass <tt>llvm::GCStrategy</tt> and register it with the compiler:</p>
614 <blockquote><pre>// lib/MyGC/MyGC.cpp - Example LLVM GC plugin
616 #include "llvm/CodeGen/GCStrategy.h"
617 #include "llvm/CodeGen/GCMetadata.h"
618 #include "llvm/Support/Compiler.h"
620 using namespace llvm;
623 class LLVM_LIBRARY_VISIBILITY MyGC : public GCStrategy {
628 GCRegistry::Add<MyGC>
629 X("mygc", "My bespoke garbage collector.");
632 <p>This boilerplate collector does nothing. More specifically:</p>
635 <li><tt>llvm.gcread</tt> calls are replaced with the corresponding
636 <tt>load</tt> instruction.</li>
637 <li><tt>llvm.gcwrite</tt> calls are replaced with the corresponding
638 <tt>store</tt> instruction.</li>
639 <li>No safe points are added to the code.</li>
640 <li>The stack map is not compiled into the executable.</li>
643 <p>Using the LLVM makefiles (like the <a
644 href="http://llvm.org/viewvc/llvm-project/llvm/trunk/projects/sample/">sample
645 project</a>), this code can be compiled as a plugin using a simple
652 LIBRARYNAME = <var>MyGC</var>
655 include $(LEVEL)/Makefile.common</pre></blockquote>
657 <p>Once the plugin is compiled, code using it may be compiled using <tt>llc
658 -load=<var>MyGC.so</var></tt> (though <var>MyGC.so</var> may have some other
659 platform-specific extension):</p>
663 define void @f() gc "mygc" {
667 $ llvm-as < sample.ll | llc -load=MyGC.so</pre></blockquote>
669 <p>It is also possible to statically link the collector plugin into tools, such
670 as a language-specific compiler front-end.</p>
672 <!-- ======================================================================= -->
674 <a name="collector-algos">Overview of available features</a>
679 <p><tt>GCStrategy</tt> provides a range of features through which a plugin
680 may do useful work. Some of these are callbacks, some are algorithms that can
681 be enabled, disabled, or customized. This matrix summarizes the supported (and
682 planned) features and correlates them with the collection techniques which
683 typically require them.</p>
689 <th>shadow stack</th>
698 <th class="rowhead"><a href="#stack-map">stack map</a></th>
709 <th class="rowhead"><a href="#init-roots">initialize roots</a></th>
719 <tr class="doc_warning">
720 <th class="rowhead">derived pointers</th>
731 <th class="rowhead"><em><a href="#custom">custom lowering</a></em></th>
742 <th class="rowhead indent">gcroot</th>
753 <th class="rowhead indent">gcwrite</th>
764 <th class="rowhead indent">gcread</th>
775 <th class="rowhead"><em><a href="#safe-points">safe points</a></em></th>
786 <th class="rowhead indent">in calls</th>
797 <th class="rowhead indent">before calls</th>
807 <tr class="doc_warning">
808 <th class="rowhead indent">for loops</th>
819 <th class="rowhead indent">before escape</th>
829 <tr class="doc_warning">
830 <th class="rowhead">emit code at safe points</th>
841 <th class="rowhead"><em>output</em></th>
852 <th class="rowhead indent"><a href="#assembly">assembly</a></th>
862 <tr class="doc_warning">
863 <th class="rowhead indent">JIT</th>
867 <td class="optl">✘</td>
868 <td class="optl">✘</td>
869 <td class="optl">✘</td>
870 <td class="optl">✘</td>
871 <td class="optl">✘</td>
873 <tr class="doc_warning">
874 <th class="rowhead indent">obj</th>
878 <td class="optl">✘</td>
879 <td class="optl">✘</td>
880 <td class="optl">✘</td>
881 <td class="optl">✘</td>
882 <td class="optl">✘</td>
884 <tr class="doc_warning">
885 <th class="rowhead">live analysis</th>
889 <td class="optl">✘</td>
890 <td class="optl">✘</td>
891 <td class="optl">✘</td>
892 <td class="optl">✘</td>
893 <td class="optl">✘</td>
895 <tr class="doc_warning">
896 <th class="rowhead">register map</th>
900 <td class="optl">✘</td>
901 <td class="optl">✘</td>
902 <td class="optl">✘</td>
903 <td class="optl">✘</td>
904 <td class="optl">✘</td>
908 <div><span class="doc_warning">*</span> Derived pointers only pose a
909 hazard to copying collectors.</div>
910 <div><span class="optl">✘</span> in gray denotes a feature which
911 could be utilized if available.</div>
916 <p>To be clear, the collection techniques above are defined as:</p>
919 <dt>Shadow Stack</dt>
920 <dd>The mutator carefully maintains a linked list of stack roots.</dd>
921 <dt>Reference Counting</dt>
922 <dd>The mutator maintains a reference count for each object and frees an
923 object when its count falls to zero.</dd>
925 <dd>When the heap is exhausted, the collector marks reachable objects starting
926 from the roots, then deallocates unreachable objects in a sweep
929 <dd>As reachability analysis proceeds, the collector copies objects from one
930 heap area to another, compacting them in the process. Copying collectors
931 enable highly efficient "bump pointer" allocation and can improve locality
934 <dd>(Including generational collectors.) Incremental collectors generally have
935 all the properties of a copying collector (regardless of whether the
936 mature heap is compacting), but bring the added complexity of requiring
939 <dd>Denotes a multithreaded mutator; the collector must still stop the mutator
940 ("stop the world") before beginning reachability analysis. Stopping a
941 multithreaded mutator is a complicated problem. It generally requires
942 highly platform specific code in the runtime, and the production of
943 carefully designed machine code at safe points.</dd>
945 <dd>In this technique, the mutator and the collector run concurrently, with
946 the goal of eliminating pause times. In a <em>cooperative</em> collector,
947 the mutator further aids with collection should a pause occur, allowing
948 collection to take advantage of multiprocessor hosts. The "stop the world"
949 problem of threaded collectors is generally still present to a limited
950 extent. Sophisticated marking algorithms are necessary. Read barriers may
954 <p>As the matrix indicates, LLVM's garbage collection infrastructure is already
955 suitable for a wide variety of collectors, but does not currently extend to
956 multithreaded programs. This will be added in the future as there is
961 <!-- ======================================================================= -->
963 <a name="stack-map">Computing stack maps</a>
968 <p>LLVM automatically computes a stack map. One of the most important features
969 of a <tt>GCStrategy</tt> is to compile this information into the executable in
970 the binary representation expected by the runtime library.</p>
972 <p>The stack map consists of the location and identity of each GC root in the
973 each function in the module. For each root:</p>
976 <li><tt>RootNum</tt>: The index of the root.</li>
977 <li><tt>StackOffset</tt>: The offset of the object relative to the frame
979 <li><tt>RootMetadata</tt>: The value passed as the <tt>%metadata</tt>
980 parameter to the <a href="#gcroot"><tt>@llvm.gcroot</tt></a> intrinsic.</li>
983 <p>Also, for the function as a whole:</p>
986 <li><tt>getFrameSize()</tt>: The overall size of the function's initial
987 stack frame, not accounting for any dynamic allocation.</li>
988 <li><tt>roots_size()</tt>: The count of roots in the function.</li>
991 <p>To access the stack map, use <tt>GCFunctionMetadata::roots_begin()</tt> and
992 -<tt>end()</tt> from the <tt><a
993 href="#assembly">GCMetadataPrinter</a></tt>:</p>
996 >for (iterator I = begin(), E = end(); I != E; ++I) {
997 GCFunctionInfo *FI = *I;
998 unsigned FrameSize = FI->getFrameSize();
999 size_t RootCount = FI->roots_size();
1001 for (GCFunctionInfo::roots_iterator RI = FI->roots_begin(),
1002 RE = FI->roots_end();
1004 int RootNum = RI->Num;
1005 int RootStackOffset = RI->StackOffset;
1006 Constant *RootMetadata = RI->Metadata;
1008 }</pre></blockquote>
1010 <p>If the <tt>llvm.gcroot</tt> intrinsic is eliminated before code generation by
1011 a custom lowering pass, LLVM will compute an empty stack map. This may be useful
1012 for collector plugins which implement reference counting or a shadow stack.</p>
1017 <!-- ======================================================================= -->
1019 <a name="init-roots">Initializing roots to null: <tt>InitRoots</tt></a>
1027 }</pre></blockquote>
1029 <p>When set, LLVM will automatically initialize each root to <tt>null</tt> upon
1030 entry to the function. This prevents the GC's sweep phase from visiting
1031 uninitialized pointers, which will almost certainly cause it to crash. This
1032 initialization occurs before custom lowering, so the two may be used
1035 <p>Since LLVM does not yet compute liveness information, there is no means of
1036 distinguishing an uninitialized stack root from an initialized one. Therefore,
1037 this feature should be used by all GC plugins. It is enabled by default.</p>
1042 <!-- ======================================================================= -->
1044 <a name="custom">Custom lowering of intrinsics: <tt>CustomRoots</tt>,
1045 <tt>CustomReadBarriers</tt>, and <tt>CustomWriteBarriers</tt></a>
1050 <p>For GCs which use barriers or unusual treatment of stack roots, these
1051 flags allow the collector to perform arbitrary transformations of the LLVM
1055 >class MyGC : public GCStrategy {
1059 CustomReadBarriers = true;
1060 CustomWriteBarriers = true;
1063 virtual bool initializeCustomLowering(Module &M);
1064 virtual bool performCustomLowering(Function &F);
1065 };</pre></blockquote>
1067 <p>If any of these flags are set, then LLVM suppresses its default lowering for
1068 the corresponding intrinsics and instead calls
1069 <tt>performCustomLowering</tt>.</p>
1071 <p>LLVM's default action for each intrinsic is as follows:</p>
1074 <li><tt>llvm.gcroot</tt>: Leave it alone. The code generator must see it
1075 or the stack map will not be computed.</li>
1076 <li><tt>llvm.gcread</tt>: Substitute a <tt>load</tt> instruction.</li>
1077 <li><tt>llvm.gcwrite</tt>: Substitute a <tt>store</tt> instruction.</li>
1080 <p>If <tt>CustomReadBarriers</tt> or <tt>CustomWriteBarriers</tt> are specified,
1081 then <tt>performCustomLowering</tt> <strong>must</strong> eliminate the
1082 corresponding barriers.</p>
1084 <p><tt>performCustomLowering</tt> must comply with the same restrictions as <a
1085 href="WritingAnLLVMPass.html#runOnFunction"><tt
1086 >FunctionPass::runOnFunction</tt></a>.
1087 Likewise, <tt>initializeCustomLowering</tt> has the same semantics as <a
1088 href="WritingAnLLVMPass.html#doInitialization_mod"><tt
1089 >Pass::doInitialization(Module&)</tt></a>.</p>
1091 <p>The following can be used as a template:</p>
1094 >#include "llvm/Module.h"
1095 #include "llvm/IntrinsicInst.h"
1097 bool MyGC::initializeCustomLowering(Module &M) {
1101 bool MyGC::performCustomLowering(Function &F) {
1102 bool MadeChange = false;
1104 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1105 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; )
1106 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(II++))
1107 if (Function *F = CI->getCalledFunction())
1108 switch (F->getIntrinsicID()) {
1109 case Intrinsic::gcwrite:
1110 // Handle llvm.gcwrite.
1111 CI->eraseFromParent();
1114 case Intrinsic::gcread:
1115 // Handle llvm.gcread.
1116 CI->eraseFromParent();
1119 case Intrinsic::gcroot:
1120 // Handle llvm.gcroot.
1121 CI->eraseFromParent();
1127 }</pre></blockquote>
1132 <!-- ======================================================================= -->
1134 <a name="safe-points">Generating safe points: <tt>NeededSafePoints</tt></a>
1139 <p>LLVM can compute four kinds of safe points:</p>
1143 /// PointKind - The type of a collector-safe point.
1146 Loop, //< Instr is a loop (backwards branch).
1147 Return, //< Instr is a return instruction.
1148 PreCall, //< Instr is a call instruction.
1149 PostCall //< Instr is the return address of a call.
1151 }</pre></blockquote>
1153 <p>A collector can request any combination of the four by setting the
1154 <tt>NeededSafePoints</tt> mask:</p>
1158 NeededSafePoints = 1 << GC::Loop
1159 | 1 << GC::Return
1160 | 1 << GC::PreCall
1161 | 1 << GC::PostCall;
1162 }</pre></blockquote>
1164 <p>It can then use the following routines to access safe points.</p>
1167 >for (iterator I = begin(), E = end(); I != E; ++I) {
1168 GCFunctionInfo *MD = *I;
1169 size_t PointCount = MD->size();
1171 for (GCFunctionInfo::iterator PI = MD->begin(),
1172 PE = MD->end(); PI != PE; ++PI) {
1173 GC::PointKind PointKind = PI->Kind;
1174 unsigned PointNum = PI->Num;
1179 <p>Almost every collector requires <tt>PostCall</tt> safe points, since these
1180 correspond to the moments when the function is suspended during a call to a
1183 <p>Threaded programs generally require <tt>Loop</tt> safe points to guarantee
1184 that the application will reach a safe point within a bounded amount of time,
1185 even if it is executing a long-running loop which contains no function
1188 <p>Threaded collectors may also require <tt>Return</tt> and <tt>PreCall</tt>
1189 safe points to implement "stop the world" techniques using self-modifying code,
1190 where it is important that the program not exit the function without reaching a
1191 safe point (because only the topmost function has been patched).</p>
1196 <!-- ======================================================================= -->
1198 <a name="assembly">Emitting assembly code: <tt>GCMetadataPrinter</tt></a>
1203 <p>LLVM allows a plugin to print arbitrary assembly code before and after the
1204 rest of a module's assembly code. At the end of the module, the GC can compile
1205 the LLVM stack map into assembly code. (At the beginning, this information is not
1208 <p>Since AsmWriter and CodeGen are separate components of LLVM, a separate
1209 abstract base class and registry is provided for printing assembly code, the
1210 <tt>GCMetadaPrinter</tt> and <tt>GCMetadataPrinterRegistry</tt>. The AsmWriter
1211 will look for such a subclass if the <tt>GCStrategy</tt> sets
1212 <tt>UsesMetadata</tt>:</p>
1216 UsesMetadata = true;
1217 }</pre></blockquote>
1219 <p>This separation allows JIT-only clients to be smaller.</p>
1221 <p>Note that LLVM does not currently have analogous APIs to support code
1222 generation in the JIT, nor using the object writers.</p>
1225 >// lib/MyGC/MyGCPrinter.cpp - Example LLVM GC printer
1227 #include "llvm/CodeGen/GCMetadataPrinter.h"
1228 #include "llvm/Support/Compiler.h"
1230 using namespace llvm;
1233 class LLVM_LIBRARY_VISIBILITY MyGCPrinter : public GCMetadataPrinter {
1235 virtual void beginAssembly(std::ostream &OS, AsmPrinter &AP,
1236 const TargetAsmInfo &TAI);
1238 virtual void finishAssembly(std::ostream &OS, AsmPrinter &AP,
1239 const TargetAsmInfo &TAI);
1242 GCMetadataPrinterRegistry::Add<MyGCPrinter>
1243 X("mygc", "My bespoke garbage collector.");
1244 }</pre></blockquote>
1246 <p>The collector should use <tt>AsmPrinter</tt> and <tt>TargetAsmInfo</tt> to
1247 print portable assembly code to the <tt>std::ostream</tt>. The collector itself
1248 contains the stack map for the entire module, and may access the
1249 <tt>GCFunctionInfo</tt> using its own <tt>begin()</tt> and <tt>end()</tt>
1250 methods. Here's a realistic example:</p>
1253 >#include "llvm/CodeGen/AsmPrinter.h"
1254 #include "llvm/Function.h"
1255 #include "llvm/Target/TargetMachine.h"
1256 #include "llvm/Target/TargetData.h"
1257 #include "llvm/Target/TargetAsmInfo.h"
1259 void MyGCPrinter::beginAssembly(std::ostream &OS, AsmPrinter &AP,
1260 const TargetAsmInfo &TAI) {
1264 void MyGCPrinter::finishAssembly(std::ostream &OS, AsmPrinter &AP,
1265 const TargetAsmInfo &TAI) {
1266 // Set up for emitting addresses.
1267 const char *AddressDirective;
1268 int AddressAlignLog;
1269 if (AP.TM.getTargetData()->getPointerSize() == sizeof(int32_t)) {
1270 AddressDirective = TAI.getData32bitsDirective();
1271 AddressAlignLog = 2;
1273 AddressDirective = TAI.getData64bitsDirective();
1274 AddressAlignLog = 3;
1277 // Put this in the data section.
1278 AP.SwitchToDataSection(TAI.getDataSection());
1280 // For each function...
1281 for (iterator FI = begin(), FE = end(); FI != FE; ++FI) {
1282 GCFunctionInfo &MD = **FI;
1284 // Emit this data structure:
1287 // int32_t PointCount;
1289 // void *SafePointAddress;
1290 // int32_t LiveCount;
1291 // int32_t LiveOffsets[LiveCount];
1292 // } Points[PointCount];
1293 // } __gcmap_<FUNCTIONNAME>;
1295 // Align to address width.
1296 AP.EmitAlignment(AddressAlignLog);
1298 // Emit the symbol by which the stack map entry can be found.
1300 Symbol += TAI.getGlobalPrefix();
1301 Symbol += "__gcmap_";
1302 Symbol += MD.getFunction().getName();
1303 if (const char *GlobalDirective = TAI.getGlobalDirective())
1304 OS << GlobalDirective << Symbol << "\n";
1305 OS << TAI.getGlobalPrefix() << Symbol << ":\n";
1308 AP.EmitInt32(MD.size());
1309 AP.EOL("safe point count");
1311 // And each safe point...
1312 for (GCFunctionInfo::iterator PI = MD.begin(),
1313 PE = MD.end(); PI != PE; ++PI) {
1314 // Align to address width.
1315 AP.EmitAlignment(AddressAlignLog);
1317 // Emit the address of the safe point.
1318 OS << AddressDirective
1319 << TAI.getPrivateGlobalPrefix() << "label" << PI->Num;
1320 AP.EOL("safe point address");
1322 // Emit the stack frame size.
1323 AP.EmitInt32(MD.getFrameSize());
1324 AP.EOL("stack frame size");
1326 // Emit the number of live roots in the function.
1327 AP.EmitInt32(MD.live_size(PI));
1328 AP.EOL("live root count");
1330 // And for each live root...
1331 for (GCFunctionInfo::live_iterator LI = MD.live_begin(PI),
1332 LE = MD.live_end(PI);
1334 // Print its offset within the stack frame.
1335 AP.EmitInt32(LI->StackOffset);
1336 AP.EOL("stack offset");
1347 <!-- *********************************************************************** -->
1349 <a name="references">References</a>
1351 <!-- *********************************************************************** -->
1355 <p><a name="appel89">[Appel89]</a> Runtime Tags Aren't Necessary. Andrew
1356 W. Appel. Lisp and Symbolic Computation 19(7):703-705, July 1989.</p>
1358 <p><a name="goldberg91">[Goldberg91]</a> Tag-free garbage collection for
1359 strongly typed programming languages. Benjamin Goldberg. ACM SIGPLAN
1362 <p><a name="tolmach94">[Tolmach94]</a> Tag-free garbage collection using
1363 explicit type parameters. Andrew Tolmach. Proceedings of the 1994 ACM
1364 conference on LISP and functional programming.</p>
1366 <p><a name="henderson02">[Henderson2002]</a> <a
1367 href="http://citeseer.ist.psu.edu/henderson02accurate.html">
1368 Accurate Garbage Collection in an Uncooperative Environment</a>.
1369 Fergus Henderson. International Symposium on Memory Management 2002.</p>
1374 <!-- *********************************************************************** -->
1378 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1379 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
1380 <a href="http://validator.w3.org/check/referer"><img
1381 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
1383 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1384 <a href="http://llvm.org/">LLVM Compiler Infrastructure</a><br>
1385 Last modified: $Date$