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">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 <!-- *********************************************************************** -->
82 <div class="doc_section">
83 <a name="introduction">Introduction</a>
85 <!-- *********************************************************************** -->
87 <div class="doc_text">
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>
129 <!-- ======================================================================= -->
130 <div class="doc_subsection">
131 <a name="feature">Goals and non-goals</a>
134 <div class="doc_text">
136 <p>LLVM's intermediate representation provides <a href="#intrinsics">garbage
137 collection intrinsics</a> that offer support for a broad class of
138 collector models. For instance, the intrinsics permit:</p>
141 <li>semi-space collectors</li>
142 <li>mark-sweep collectors</li>
143 <li>generational collectors</li>
144 <li>reference counting</li>
145 <li>incremental collectors</li>
146 <li>concurrent collectors</li>
147 <li>cooperative collectors</li>
150 <p>We hope that the primitive support built into the LLVM IR is sufficient to
151 support a broad class of garbage collected languages including Scheme, ML, Java,
152 C#, Perl, Python, Lua, Ruby, other scripting languages, and more.</p>
154 <p>However, LLVM does not itself provide a garbage collector—this should
155 be part of your language's runtime library. LLVM provides a framework for
156 compile time <a href="#plugin">code generation plugins</a>. The role of these
157 plugins is to generate code and data structures which conforms to the <em>binary
158 interface</em> specified by the <em>runtime library</em>. This is similar to the
159 relationship between LLVM and DWARF debugging info, for example. The
160 difference primarily lies in the lack of an established standard in the domain
161 of garbage collection—thus the plugins.</p>
163 <p>The aspects of the binary interface with which LLVM's GC support is
167 <li>Creation of GC-safe points within code where collection is allowed to
169 <li>Computation of the stack map. For each safe point in the code, object
170 references within the stack frame must be identified so that the
171 collector may traverse and perhaps update them.</li>
172 <li>Write barriers when storing object references to the heap. These are
173 commonly used to optimize incremental scans in generational
175 <li>Emission of read barriers when loading object references. These are
176 useful for interoperating with concurrent collectors.</li>
179 <p>There are additional areas that LLVM does not directly address:</p>
182 <li>Registration of global roots with the runtime.</li>
183 <li>Registration of stack map entries with the runtime.</li>
184 <li>The functions used by the program to allocate memory, trigger a
185 collection, etc.</li>
186 <li>Computation or compilation of type maps, or registration of them with
187 the runtime. These are used to crawl the heap for object
191 <p>In general, LLVM's support for GC does not include features which can be
192 adequately addressed with other features of the IR and does not specify a
193 particular binary interface. On the plus side, this means that you should be
194 able to integrate LLVM with an existing runtime. On the other hand, it leaves
195 a lot of work for the developer of a novel language. However, it's easy to get
196 started quickly and scale up to a more sophisticated implementation as your
197 compiler matures.</p>
201 <!-- *********************************************************************** -->
202 <div class="doc_section">
203 <a name="quickstart">Getting started</a>
205 <!-- *********************************************************************** -->
207 <div class="doc_text">
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>
251 <!-- ======================================================================= -->
252 <div class="doc_subsection">
253 <a name="quickstart-compiler">In your compiler</a>
256 <div class="doc_text">
258 <p>To turn the shadow stack on for your functions, first call:</p>
260 <div class="doc_code"><pre
261 >F.setGC("shadow-stack");</pre></div>
263 <p>for each function your compiler emits. Since the shadow stack is built into
264 LLVM, you do not need to load a plugin.</p>
266 <p>Your compiler must also use <tt>@llvm.gcroot</tt> as documented.
267 Don't forget to create a root for each intermediate value that is generated
268 when evaluating an expression. In <tt>h(f(), g())</tt>, the result of
269 <tt>f()</tt> could easily be collected if evaluating <tt>g()</tt> triggers a
272 <p>There's no need to use <tt>@llvm.gcread</tt> and <tt>@llvm.gcwrite</tt> over
273 plain <tt>load</tt> and <tt>store</tt> for now. You will need them when
274 switching to a more advanced GC.</p>
278 <!-- ======================================================================= -->
279 <div class="doc_subsection">
280 <a name="quickstart-runtime">In your runtime</a>
283 <div class="doc_text">
285 <p>The shadow stack doesn't imply a memory allocation algorithm. A semispace
286 collector or building atop <tt>malloc</tt> are great places to start, and can
287 be implemented with very little code.</p>
289 <p>When it comes time to collect, however, your runtime needs to traverse the
290 stack roots, and for this it needs to integrate with the shadow stack. Luckily,
291 doing so is very simple. (This code is heavily commented to help you
292 understand the data structure, but there are only 20 lines of meaningful
297 <div class="doc_code"><pre
298 >/// @brief The map for a single function's stack frame. One of these is
299 /// compiled as constant data into the executable for each function.
301 /// Storage of metadata values is elided if the %metadata parameter to
302 /// @llvm.gcroot is null.
304 int32_t NumRoots; //< Number of roots in stack frame.
305 int32_t NumMeta; //< Number of metadata entries. May be < NumRoots.
306 const void *Meta[0]; //< Metadata for each root.
309 /// @brief A link in the dynamic shadow stack. One of these is embedded in the
310 /// stack frame of each function on the call stack.
312 StackEntry *Next; //< Link to next stack entry (the caller's).
313 const FrameMap *Map; //< Pointer to constant FrameMap.
314 void *Roots[0]; //< Stack roots (in-place array).
317 /// @brief The head of the singly-linked list of StackEntries. Functions push
318 /// and pop onto this in their prologue and epilogue.
320 /// Since there is only a global list, this technique is not threadsafe.
321 StackEntry *llvm_gc_root_chain;
323 /// @brief Calls Visitor(root, meta) for each GC root on the stack.
324 /// root and meta are exactly the values passed to
325 /// <tt>@llvm.gcroot</tt>.
327 /// Visitor could be a function to recursively mark live objects. Or it
328 /// might copy them to another heap or generation.
330 /// @param Visitor A function to invoke for every GC root on the stack.
331 void visitGCRoots(void (*Visitor)(void **Root, const void *Meta)) {
332 for (StackEntry *R = llvm_gc_root_chain; R; R = R->Next) {
335 // For roots [0, NumMeta), the metadata pointer is in the FrameMap.
336 for (unsigned e = R->Map->NumMeta; i != e; ++i)
337 Visitor(&R->Roots[i], R->Map->Meta[i]);
339 // For roots [NumMeta, NumRoots), the metadata pointer is null.
340 for (unsigned e = R->Map->NumRoots; i != e; ++i)
341 Visitor(&R->Roots[i], NULL);
345 <!-- ======================================================================= -->
346 <div class="doc_subsection">
347 <a name="shadow-stack">About the shadow stack</a>
350 <div class="doc_text">
352 <p>Unlike many GC algorithms which rely on a cooperative code generator to
353 compile stack maps, this algorithm carefully maintains a linked list of stack
354 roots [<a href="#henderson02">Henderson2002</a>]. This so-called "shadow stack"
355 mirrors the machine stack. Maintaining this data structure is slower than using
356 a stack map compiled into the executable as constant data, but has a significant
357 portability advantage because it requires no special support from the target
358 code generator, and does not require tricky platform-specific code to crawl
359 the machine stack.</p>
361 <p>The tradeoff for this simplicity and portability is:</p>
364 <li>High overhead per function call.</li>
365 <li>Not thread-safe.</li>
368 <p>Still, it's an easy way to get started. After your compiler and runtime are
369 up and running, writing a <a href="#plugin">plugin</a> will allow you to take
370 advantage of <a href="#collector-algos">more advanced GC features</a> of LLVM
371 in order to improve performance.</p>
375 <!-- *********************************************************************** -->
376 <div class="doc_section">
377 <a name="core">IR features</a><a name="intrinsics"></a>
379 <!-- *********************************************************************** -->
381 <div class="doc_text">
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
395 <!-- ======================================================================= -->
396 <div class="doc_subsection">
397 <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>(...) <u>gc "<i>name</i>"</u> { ...
404 <div class="doc_text">
406 <p>The <tt>gc</tt> function attribute is used to specify the desired GC style
407 to the compiler. Its programmatic equivalent is the <tt>setGC</tt> method of
408 <tt>Function</tt>.</p>
410 <p>Setting <tt>gc "<i>name</i>"</tt> on a function triggers a search for a
411 matching code generation plugin "<i>name</i>"; it is that plugin which defines
412 the exact nature of the code generated to support GC. If none is found, the
413 compiler will raise an error.</p>
415 <p>Specifying the GC style on a per-function basis allows LLVM to link together
416 programs that use different garbage collection algorithms (or none at all).</p>
420 <!-- ======================================================================= -->
421 <div class="doc_subsection">
422 <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 <div class="doc_text">
431 <p>The <tt>llvm.gcroot</tt> intrinsic is used to inform LLVM that a stack
432 variable references an object on the heap and is to be tracked for garbage
433 collection. The exact impact on generated code is specified by a <a
434 href="#plugin">compiler plugin</a>.</p>
436 <p>A compiler which uses mem2reg to raise imperative code using <tt>alloca</tt>
437 into SSA form need only add a call to <tt>@llvm.gcroot</tt> for those variables
438 which a pointers into the GC heap.</p>
440 <p>It is also important to mark intermediate values with <tt>llvm.gcroot</tt>.
441 For example, consider <tt>h(f(), g())</tt>. Beware leaking the result of
442 <tt>f()</tt> in the case that <tt>g()</tt> triggers a collection.</p>
444 <p>The first argument <b>must</b> be a value referring to an alloca instruction
445 or a bitcast of an alloca. The second contains a pointer to metadata that
446 should be associated with the pointer, and <b>must</b> be a constant or global
447 value address. If your target collector uses tags, use a null pointer for
450 <p>The <tt>%metadata</tt> argument can be used to avoid requiring heap objects
451 to have 'isa' pointers or tag bits. [<a href="#appel89">Appel89</a>, <a
452 href="#goldberg91">Goldberg91</a>, <a href="#tolmach94">Tolmach94</a>] If
453 specified, its value will be tracked along with the location of the pointer in
456 <p>Consider the following fragment of Java code:</p>
460 Object X; // A null-initialized reference to an object
465 <p>This block (which may be located in the middle of a function or in a loop
466 nest), could be compiled to this LLVM code:</p>
470 ;; In the entry block for the function, allocate the
471 ;; stack space for X, which is an LLVM pointer.
474 ;; Tell LLVM that the stack space is a stack root.
475 ;; Java has type-tags on objects, so we pass null as metadata.
476 %tmp = bitcast %Object** %X to i8**
477 call void @llvm.gcroot(i8** %X, i8* null)
480 ;; "CodeBlock" is the block corresponding to the start
481 ;; of the scope above.
483 ;; Java null-initializes pointers.
484 store %Object* null, %Object** %X
488 ;; As the pointer goes out of scope, store a null value into
489 ;; it, to indicate that the value is no longer live.
490 store %Object* null, %Object** %X
496 <!-- ======================================================================= -->
497 <div class="doc_subsection">
498 <a name="barriers">Reading and writing references in the heap</a>
501 <div class="doc_text">
503 <p>Some collectors need to be informed when the mutator (the program that needs
504 garbage collection) either reads a pointer from or writes a pointer to a field
505 of a heap object. The code fragments inserted at these points are called
506 <em>read barriers</em> and <em>write barriers</em>, respectively. The amount of
507 code that needs to be executed is usually quite small and not on the critical
508 path of any computation, so the overall performance impact of the barrier is
511 <p>Barriers often require access to the <em>object pointer</em> rather than the
512 <em>derived pointer</em> (which is a pointer to the field within the
513 object). Accordingly, these intrinsics take both pointers as separate arguments
514 for completeness. In this snippet, <tt>%object</tt> is the object pointer, and
515 <tt>%derived</tt> is the derived pointer:</p>
519 %class.Array = type { %class.Object, i32, [0 x %class.Object*] }
522 ;; Load the object pointer from a gcroot.
523 %object = load %class.Array** %object_addr
525 ;; Compute the derived pointer.
526 %derived = getelementptr %object, i32 0, i32 2, i32 %n</pre></blockquote>
528 <p>LLVM does not enforce this relationship between the object and derived
529 pointer (although a <a href="#plugin">plugin</a> might). However, it would be
530 an unusual collector that violated it.</p>
532 <p>The use of these intrinsics is naturally optional if the target GC does
533 require the corresponding barrier. Such a GC plugin will replace the intrinsic
534 calls with the corresponding <tt>load</tt> or <tt>store</tt> instruction if they
539 <!-- ======================================================================= -->
540 <div class="doc_subsubsection">
541 <a name="gcwrite">Write barrier: <tt>llvm.gcwrite</tt></a>
544 <div class="doc_code"><tt>
545 void @llvm.gcwrite(i8* %value, i8* %object, i8** %derived)
548 <div class="doc_text">
550 <p>For write barriers, LLVM provides the <tt>llvm.gcwrite</tt> intrinsic
551 function. It has exactly the same semantics as a non-volatile <tt>store</tt> to
552 the derived pointer (the third argument). The exact code generated is specified
553 by a <a href="#plugin">compiler plugin</a>.</p>
555 <p>Many important algorithms require write barriers, including generational
556 and concurrent collectors. Additionally, write barriers could be used to
557 implement reference counting.</p>
561 <!-- ======================================================================= -->
562 <div class="doc_subsubsection">
563 <a name="gcread">Read barrier: <tt>llvm.gcread</tt></a>
566 <div class="doc_code"><tt>
567 i8* @llvm.gcread(i8* %object, i8** %derived)<br>
570 <div class="doc_text">
572 <p>For read barriers, LLVM provides the <tt>llvm.gcread</tt> intrinsic function.
573 It has exactly the same semantics as a non-volatile <tt>load</tt> from the
574 derived pointer (the second argument). The exact code generated is specified by
575 a <a href="#plugin">compiler plugin</a>.</p>
577 <p>Read barriers are needed by fewer algorithms than write barriers, and may
578 have a greater performance impact since pointer reads are more frequent than
583 <!-- *********************************************************************** -->
584 <div class="doc_section">
585 <a name="plugin">Implementing a collector plugin</a>
587 <!-- *********************************************************************** -->
589 <div class="doc_text">
591 <p>User code specifies which GC code generation to use with the <tt>gc</tt>
592 function attribute or, equivalently, with the <tt>setGC</tt> method of
593 <tt>Function</tt>.</p>
595 <p>To implement a GC plugin, it is necessary to subclass
596 <tt>llvm::GCStrategy</tt>, which can be accomplished in a few lines of
597 boilerplate code. LLVM's infrastructure provides access to several important
598 algorithms. For an uncontroversial collector, all that remains may be to
599 compile LLVM's computed stack map to assembly code (using the binary
600 representation expected by the runtime library). This can be accomplished in
601 about 100 lines of code.</p>
603 <p>This is not the appropriate place to implement a garbage collected heap or a
604 garbage collector itself. That code should exist in the language's runtime
605 library. The compiler plugin is responsible for generating code which
606 conforms to the binary interface defined by library, most essentially the
607 <a href="#stack-map">stack map</a>.</p>
609 <p>To subclass <tt>llvm::GCStrategy</tt> and register it with the compiler:</p>
611 <blockquote><pre>// lib/MyGC/MyGC.cpp - Example LLVM GC plugin
613 #include "llvm/CodeGen/GCStrategy.h"
614 #include "llvm/CodeGen/GCMetadata.h"
615 #include "llvm/Support/Compiler.h"
617 using namespace llvm;
620 class VISIBILITY_HIDDEN MyGC : public GCStrategy {
625 GCRegistry::Add<MyGC>
626 X("mygc", "My bespoke garbage collector.");
629 <p>This boilerplate collector does nothing. More specifically:</p>
632 <li><tt>llvm.gcread</tt> calls are replaced with the corresponding
633 <tt>load</tt> instruction.</li>
634 <li><tt>llvm.gcwrite</tt> calls are replaced with the corresponding
635 <tt>store</tt> instruction.</li>
636 <li>No safe points are added to the code.</li>
637 <li>The stack map is not compiled into the executable.</li>
640 <p>Using the LLVM makefiles (like the <a
641 href="http://llvm.org/viewvc/llvm-project/llvm/trunk/projects/sample/">sample
642 project</a>), this code can be compiled as a plugin using a simple
649 LIBRARYNAME = <var>MyGC</var>
652 include $(LEVEL)/Makefile.common</pre></blockquote>
654 <p>Once the plugin is compiled, code using it may be compiled using <tt>llc
655 -load=<var>MyGC.so</var></tt> (though <var>MyGC.so</var> may have some other
656 platform-specific extension):</p>
660 define void @f() gc "mygc" {
664 $ llvm-as < sample.ll | llc -load=MyGC.so</pre></blockquote>
666 <p>It is also possible to statically link the collector plugin into tools, such
667 as a language-specific compiler front-end.</p>
671 <!-- ======================================================================= -->
672 <div class="doc_subsection">
673 <a name="collector-algos">Overview of available features</a>
676 <div class="doc_text">
678 <p><tt>GCStrategy</tt> provides a range of features through which a plugin
679 may do useful work. Some of these are callbacks, some are algorithms that can
680 be enabled, disabled, or customized. This matrix summarizes the supported (and
681 planned) features and correlates them with the collection techniques which
682 typically require them.</p>
688 <th>shadow stack</th>
697 <th class="rowhead"><a href="#stack-map">stack map</a></th>
708 <th class="rowhead"><a href="#init-roots">initialize roots</a></th>
718 <tr class="doc_warning">
719 <th class="rowhead">derived pointers</th>
730 <th class="rowhead"><em><a href="#custom">custom lowering</a></em></th>
741 <th class="rowhead indent">gcroot</th>
752 <th class="rowhead indent">gcwrite</th>
763 <th class="rowhead indent">gcread</th>
774 <th class="rowhead"><em><a href="#safe-points">safe points</a></em></th>
785 <th class="rowhead indent">in calls</th>
796 <th class="rowhead indent">before calls</th>
806 <tr class="doc_warning">
807 <th class="rowhead indent">for loops</th>
818 <th class="rowhead indent">before escape</th>
828 <tr class="doc_warning">
829 <th class="rowhead">emit code at safe points</th>
840 <th class="rowhead"><em>output</em></th>
851 <th class="rowhead indent"><a href="#assembly">assembly</a></th>
861 <tr class="doc_warning">
862 <th class="rowhead indent">JIT</th>
866 <td class="optl">✘</td>
867 <td class="optl">✘</td>
868 <td class="optl">✘</td>
869 <td class="optl">✘</td>
870 <td class="optl">✘</td>
872 <tr class="doc_warning">
873 <th class="rowhead indent">obj</th>
877 <td class="optl">✘</td>
878 <td class="optl">✘</td>
879 <td class="optl">✘</td>
880 <td class="optl">✘</td>
881 <td class="optl">✘</td>
883 <tr class="doc_warning">
884 <th class="rowhead">live analysis</th>
888 <td class="optl">✘</td>
889 <td class="optl">✘</td>
890 <td class="optl">✘</td>
891 <td class="optl">✘</td>
892 <td class="optl">✘</td>
894 <tr class="doc_warning">
895 <th class="rowhead">register map</th>
899 <td class="optl">✘</td>
900 <td class="optl">✘</td>
901 <td class="optl">✘</td>
902 <td class="optl">✘</td>
903 <td class="optl">✘</td>
907 <div><span class="doc_warning">*</span> Derived pointers only pose a
908 hazard to copying collectors.</div>
909 <div><span class="optl">✘</span> in gray denotes a feature which
910 could be utilized if available.</div>
915 <p>To be clear, the collection techniques above are defined as:</p>
918 <dt>Shadow Stack</dt>
919 <dd>The mutator carefully maintains a linked list of stack roots.</dd>
920 <dt>Reference Counting</dt>
921 <dd>The mutator maintains a reference count for each object and frees an
922 object when its count falls to zero.</dd>
924 <dd>When the heap is exhausted, the collector marks reachable objects starting
925 from the roots, then deallocates unreachable objects in a sweep
928 <dd>As reachability analysis proceeds, the collector copies objects from one
929 heap area to another, compacting them in the process. Copying collectors
930 enable highly efficient "bump pointer" allocation and can improve locality
933 <dd>(Including generational collectors.) Incremental collectors generally have
934 all the properties of a copying collector (regardless of whether the
935 mature heap is compacting), but bring the added complexity of requiring
938 <dd>Denotes a multithreaded mutator; the collector must still stop the mutator
939 ("stop the world") before beginning reachability analysis. Stopping a
940 multithreaded mutator is a complicated problem. It generally requires
941 highly platform specific code in the runtime, and the production of
942 carefully designed machine code at safe points.</dd>
944 <dd>In this technique, the mutator and the collector run concurrently, with
945 the goal of eliminating pause times. In a <em>cooperative</em> collector,
946 the mutator further aids with collection should a pause occur, allowing
947 collection to take advantage of multiprocessor hosts. The "stop the world"
948 problem of threaded collectors is generally still present to a limited
949 extent. Sophisticated marking algorithms are necessary. Read barriers may
953 <p>As the matrix indicates, LLVM's garbage collection infrastructure is already
954 suitable for a wide variety of collectors, but does not currently extend to
955 multithreaded programs. This will be added in the future as there is
960 <!-- ======================================================================= -->
961 <div class="doc_subsection">
962 <a name="stack-map">Computing stack maps</a>
965 <div class="doc_text">
967 <p>LLVM automatically computes a stack map. One of the most important features
968 of a <tt>GCStrategy</tt> is to compile this information into the executable in
969 the binary representation expected by the runtime library.</p>
971 <p>The stack map consists of the location and identity of each GC root in the
972 each function in the module. For each root:</p>
975 <li><tt>RootNum</tt>: The index of the root.</li>
976 <li><tt>StackOffset</tt>: The offset of the object relative to the frame
978 <li><tt>RootMetadata</tt>: The value passed as the <tt>%metadata</tt>
979 parameter to the <a href="#gcroot"><tt>@llvm.gcroot</tt></a> intrinsic.</li>
982 <p>Also, for the function as a whole:</p>
985 <li><tt>getFrameSize()</tt>: The overall size of the function's initial
986 stack frame, not accounting for any dynamic allocation.</li>
987 <li><tt>roots_size()</tt>: The count of roots in the function.</li>
990 <p>To access the stack map, use <tt>GCFunctionMetadata::roots_begin()</tt> and
991 -<tt>end()</tt> from the <tt><a
992 href="#assembly">GCMetadataPrinter</a></tt>:</p>
995 >for (iterator I = begin(), E = end(); I != E; ++I) {
996 GCFunctionInfo *FI = *I;
997 unsigned FrameSize = FI->getFrameSize();
998 size_t RootCount = FI->roots_size();
1000 for (GCFunctionInfo::roots_iterator RI = FI->roots_begin(),
1001 RE = FI->roots_end();
1003 int RootNum = RI->Num;
1004 int RootStackOffset = RI->StackOffset;
1005 Constant *RootMetadata = RI->Metadata;
1007 }</pre></blockquote>
1009 <p>If the <tt>llvm.gcroot</tt> intrinsic is eliminated before code generation by
1010 a custom lowering pass, LLVM will compute an empty stack map. This may be useful
1011 for collector plugins which implement reference counting or a shadow stack.</p>
1016 <!-- ======================================================================= -->
1017 <div class="doc_subsection">
1018 <a name="init-roots">Initializing roots to null: <tt>InitRoots</tt></a>
1021 <div class="doc_text">
1026 }</pre></blockquote>
1028 <p>When set, LLVM will automatically initialize each root to <tt>null</tt> upon
1029 entry to the function. This prevents the GC's sweep phase from visiting
1030 uninitialized pointers, which will almost certainly cause it to crash. This
1031 initialization occurs before custom lowering, so the two may be used
1034 <p>Since LLVM does not yet compute liveness information, there is no means of
1035 distinguishing an uninitialized stack root from an initialized one. Therefore,
1036 this feature should be used by all GC plugins. It is enabled by default.</p>
1041 <!-- ======================================================================= -->
1042 <div class="doc_subsection">
1043 <a name="custom">Custom lowering of intrinsics: <tt>CustomRoots</tt>,
1044 <tt>CustomReadBarriers</tt>, and <tt>CustomWriteBarriers</tt></a>
1047 <div class="doc_text">
1049 <p>For GCs which use barriers or unusual treatment of stack roots, these
1050 flags allow the collector to perform arbitrary transformations of the LLVM
1054 >class MyGC : public GCStrategy {
1058 CustomReadBarriers = true;
1059 CustomWriteBarriers = true;
1062 virtual bool initializeCustomLowering(Module &M);
1063 virtual bool performCustomLowering(Function &F);
1064 };</pre></blockquote>
1066 <p>If any of these flags are set, then LLVM suppresses its default lowering for
1067 the corresponding intrinsics and instead calls
1068 <tt>performCustomLowering</tt>.</p>
1070 <p>LLVM's default action for each intrinsic is as follows:</p>
1073 <li><tt>llvm.gcroot</tt>: Leave it alone. The code generator must see it
1074 or the stack map will not be computed.</li>
1075 <li><tt>llvm.gcread</tt>: Substitute a <tt>load</tt> instruction.</li>
1076 <li><tt>llvm.gcwrite</tt>: Substitute a <tt>store</tt> instruction.</li>
1079 <p>If <tt>CustomReadBarriers</tt> or <tt>CustomWriteBarriers</tt> are specified,
1080 then <tt>performCustomLowering</tt> <strong>must</strong> eliminate the
1081 corresponding barriers.</p>
1083 <p><tt>performCustomLowering</tt> must comply with the same restrictions as <a
1084 href="WritingAnLLVMPass.html#runOnFunction"><tt
1085 >FunctionPass::runOnFunction</tt></a>.
1086 Likewise, <tt>initializeCustomLowering</tt> has the same semantics as <a
1087 href="WritingAnLLVMPass.html#doInitialization_mod"><tt
1088 >Pass::doInitialization(Module&)</tt></a>.</p>
1090 <p>The following can be used as a template:</p>
1093 >#include "llvm/Module.h"
1094 #include "llvm/IntrinsicInst.h"
1096 bool MyGC::initializeCustomLowering(Module &M) {
1100 bool MyGC::performCustomLowering(Function &F) {
1101 bool MadeChange = false;
1103 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1104 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; )
1105 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(II++))
1106 if (Function *F = CI->getCalledFunction())
1107 switch (F->getIntrinsicID()) {
1108 case Intrinsic::gcwrite:
1109 // Handle llvm.gcwrite.
1110 CI->eraseFromParent();
1113 case Intrinsic::gcread:
1114 // Handle llvm.gcread.
1115 CI->eraseFromParent();
1118 case Intrinsic::gcroot:
1119 // Handle llvm.gcroot.
1120 CI->eraseFromParent();
1126 }</pre></blockquote>
1131 <!-- ======================================================================= -->
1132 <div class="doc_subsection">
1133 <a name="safe-points">Generating safe points: <tt>NeededSafePoints</tt></a>
1136 <div class="doc_text">
1138 <p>LLVM can compute four kinds of safe points:</p>
1142 /// PointKind - The type of a collector-safe point.
1145 Loop, //< Instr is a loop (backwards branch).
1146 Return, //< Instr is a return instruction.
1147 PreCall, //< Instr is a call instruction.
1148 PostCall //< Instr is the return address of a call.
1150 }</pre></blockquote>
1152 <p>A collector can request any combination of the four by setting the
1153 <tt>NeededSafePoints</tt> mask:</p>
1157 NeededSafePoints = 1 << GC::Loop
1158 | 1 << GC::Return
1159 | 1 << GC::PreCall
1160 | 1 << GC::PostCall;
1161 }</pre></blockquote>
1163 <p>It can then use the following routines to access safe points.</p>
1166 >for (iterator I = begin(), E = end(); I != E; ++I) {
1167 GCFunctionInfo *MD = *I;
1168 size_t PointCount = MD->size();
1170 for (GCFunctionInfo::iterator PI = MD->begin(),
1171 PE = MD->end(); PI != PE; ++PI) {
1172 GC::PointKind PointKind = PI->Kind;
1173 unsigned PointNum = PI->Num;
1178 <p>Almost every collector requires <tt>PostCall</tt> safe points, since these
1179 correspond to the moments when the function is suspended during a call to a
1182 <p>Threaded programs generally require <tt>Loop</tt> safe points to guarantee
1183 that the application will reach a safe point within a bounded amount of time,
1184 even if it is executing a long-running loop which contains no function
1187 <p>Threaded collectors may also require <tt>Return</tt> and <tt>PreCall</tt>
1188 safe points to implement "stop the world" techniques using self-modifying code,
1189 where it is important that the program not exit the function without reaching a
1190 safe point (because only the topmost function has been patched).</p>
1195 <!-- ======================================================================= -->
1196 <div class="doc_subsection">
1197 <a name="assembly">Emitting assembly code: <tt>GCMetadataPrinter</tt></a>
1200 <div class="doc_text">
1202 <p>LLVM allows a plugin to print arbitrary assembly code before and after the
1203 rest of a module's assembly code. At the end of the module, the GC can compile
1204 the LLVM stack map into assembly code. (At the beginning, this information is not
1207 <p>Since AsmWriter and CodeGen are separate components of LLVM, a separate
1208 abstract base class and registry is provided for printing assembly code, the
1209 <tt>GCMetadaPrinter</tt> and <tt>GCMetadataPrinterRegistry</tt>. The AsmWriter
1210 will look for such a subclass if the <tt>GCStrategy</tt> sets
1211 <tt>UsesMetadata</tt>:</p>
1215 UsesMetadata = true;
1216 }</pre></blockquote>
1218 <p>This separation allows JIT-only clients to be smaller.</p>
1220 <p>Note that LLVM does not currently have analogous APIs to support code
1221 generation in the JIT, nor using the object writers.</p>
1224 >// lib/MyGC/MyGCPrinter.cpp - Example LLVM GC printer
1226 #include "llvm/CodeGen/GCMetadataPrinter.h"
1227 #include "llvm/Support/Compiler.h"
1229 using namespace llvm;
1232 class VISIBILITY_HIDDEN MyGCPrinter : public GCMetadataPrinter {
1234 virtual void beginAssembly(std::ostream &OS, AsmPrinter &AP,
1235 const TargetAsmInfo &TAI);
1237 virtual void finishAssembly(std::ostream &OS, AsmPrinter &AP,
1238 const TargetAsmInfo &TAI);
1241 GCMetadataPrinterRegistry::Add<MyGCPrinter>
1242 X("mygc", "My bespoke garbage collector.");
1243 }</pre></blockquote>
1245 <p>The collector should use <tt>AsmPrinter</tt> and <tt>TargetAsmInfo</tt> to
1246 print portable assembly code to the <tt>std::ostream</tt>. The collector itself
1247 contains the stack map for the entire module, and may access the
1248 <tt>GCFunctionInfo</tt> using its own <tt>begin()</tt> and <tt>end()</tt>
1249 methods. Here's a realistic example:</p>
1252 >#include "llvm/CodeGen/AsmPrinter.h"
1253 #include "llvm/Function.h"
1254 #include "llvm/Target/TargetMachine.h"
1255 #include "llvm/Target/TargetData.h"
1256 #include "llvm/Target/TargetAsmInfo.h"
1258 void MyGCPrinter::beginAssembly(std::ostream &OS, AsmPrinter &AP,
1259 const TargetAsmInfo &TAI) {
1263 void MyGCPrinter::finishAssembly(std::ostream &OS, AsmPrinter &AP,
1264 const TargetAsmInfo &TAI) {
1265 // Set up for emitting addresses.
1266 const char *AddressDirective;
1267 int AddressAlignLog;
1268 if (AP.TM.getTargetData()->getPointerSize() == sizeof(int32_t)) {
1269 AddressDirective = TAI.getData32bitsDirective();
1270 AddressAlignLog = 2;
1272 AddressDirective = TAI.getData64bitsDirective();
1273 AddressAlignLog = 3;
1276 // Put this in the data section.
1277 AP.SwitchToDataSection(TAI.getDataSection());
1279 // For each function...
1280 for (iterator FI = begin(), FE = end(); FI != FE; ++FI) {
1281 GCFunctionInfo &MD = **FI;
1283 // Emit this data structure:
1286 // int32_t PointCount;
1288 // void *SafePointAddress;
1289 // int32_t LiveCount;
1290 // int32_t LiveOffsets[LiveCount];
1291 // } Points[PointCount];
1292 // } __gcmap_<FUNCTIONNAME>;
1294 // Align to address width.
1295 AP.EmitAlignment(AddressAlignLog);
1297 // Emit the symbol by which the stack map entry can be found.
1299 Symbol += TAI.getGlobalPrefix();
1300 Symbol += "__gcmap_";
1301 Symbol += MD.getFunction().getName();
1302 if (const char *GlobalDirective = TAI.getGlobalDirective())
1303 OS << GlobalDirective << Symbol << "\n";
1304 OS << TAI.getGlobalPrefix() << Symbol << ":\n";
1307 AP.EmitInt32(MD.size());
1308 AP.EOL("safe point count");
1310 // And each safe point...
1311 for (GCFunctionInfo::iterator PI = MD.begin(),
1312 PE = MD.end(); PI != PE; ++PI) {
1313 // Align to address width.
1314 AP.EmitAlignment(AddressAlignLog);
1316 // Emit the address of the safe point.
1317 OS << AddressDirective
1318 << TAI.getPrivateGlobalPrefix() << "label" << PI->Num;
1319 AP.EOL("safe point address");
1321 // Emit the stack frame size.
1322 AP.EmitInt32(MD.getFrameSize());
1323 AP.EOL("stack frame size");
1325 // Emit the number of live roots in the function.
1326 AP.EmitInt32(MD.live_size(PI));
1327 AP.EOL("live root count");
1329 // And for each live root...
1330 for (GCFunctionInfo::live_iterator LI = MD.live_begin(PI),
1331 LE = MD.live_end(PI);
1333 // Print its offset within the stack frame.
1334 AP.EmitInt32(LI->StackOffset);
1335 AP.EOL("stack offset");
1345 <!-- *********************************************************************** -->
1346 <div class="doc_section">
1347 <a name="references">References</a>
1349 <!-- *********************************************************************** -->
1351 <div class="doc_text">
1353 <p><a name="appel89">[Appel89]</a> Runtime Tags Aren't Necessary. Andrew
1354 W. Appel. Lisp and Symbolic Computation 19(7):703-705, July 1989.</p>
1356 <p><a name="goldberg91">[Goldberg91]</a> Tag-free garbage collection for
1357 strongly typed programming languages. Benjamin Goldberg. ACM SIGPLAN
1360 <p><a name="tolmach94">[Tolmach94]</a> Tag-free garbage collection using
1361 explicit type parameters. Andrew Tolmach. Proceedings of the 1994 ACM
1362 conference on LISP and functional programming.</p>
1364 <p><a name="henderson02">[Henderson2002]</a> <a
1365 href="http://citeseer.ist.psu.edu/henderson02accurate.html">
1366 Accurate Garbage Collection in an Uncooperative Environment</a>.
1367 Fergus Henderson. International Symposium on Memory Management 2002.</p>
1372 <!-- *********************************************************************** -->
1376 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1377 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
1378 <a href="http://validator.w3.org/check/referer"><img
1379 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
1381 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1382 <a href="http://llvm.org">LLVM Compiler Infrastructure</a><br>
1383 Last modified: $Date$