blob: e39e85ce3cd22b09ae33641e74f9cee6a2531e23 [file] [log] [blame]
Chris Lattner0d8c2db2004-05-23 21:02:20 +00001<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
3<html>
4<head>
5 <title>Accurate Garbage Collection with LLVM</title>
6 <link rel="stylesheet" href="llvm.css" type="text/css">
7</head>
8<body>
9
10<div class="doc_title">
11 Accurate Garbage Collection with LLVM
12</div>
13
14<ol>
15 <li><a href="#introduction">Introduction</a>
16 <ul>
17 <li><a href="#feature">GC features provided and algorithms supported</a></li>
18 </ul>
19 </li>
20
21 <li><a href="#interfaces">Interfaces for user programs</a>
22 <ul>
23 <li><a href="#roots">Identifying GC roots on the stack: <tt>llvm.gcroot</tt></a></li>
Chris Lattner0d8c2db2004-05-23 21:02:20 +000024 <li><a href="#allocate">Allocating memory from the GC</a></li>
25 <li><a href="#barriers">Reading and writing references to the heap</a></li>
26 <li><a href="#explicit">Explicit invocation of the garbage collector</a></li>
27 </ul>
28 </li>
29
30 <li><a href="#gcimpl">Implementing a garbage collector</a>
31 <ul>
32 <li><a href="#llvm_gc_readwrite">Implementing <tt>llvm_gc_read</tt> and <tt>llvm_gc_write</tt></a></li>
Chris Lattner9b2a1842004-05-27 05:52:10 +000033 <li><a href="#callbacks">Callback functions used to implement the garbage collector</a></li>
Chris Lattner0d8c2db2004-05-23 21:02:20 +000034 </ul>
35 </li>
Chris Lattner9b2a1842004-05-27 05:52:10 +000036 <li><a href="#gcimpls">GC implementations available</a>
37 <ul>
38 <li><a href="#semispace">SemiSpace - A simple copying garbage collector</a></li>
Chris Lattner0b02dbc2004-07-09 05:03:54 +000039 </ul>
Chris Lattner9b2a1842004-05-27 05:52:10 +000040 </li>
Chris Lattner0d8c2db2004-05-23 21:02:20 +000041
42<!--
43 <li><a href="#codegen">Implementing GC support in a code generator</a></li>
44-->
45</ol>
46
47<div class="doc_author">
48 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a></p>
49</div>
50
51<!-- *********************************************************************** -->
52<div class="doc_section">
53 <a name="introduction">Introduction</a>
54</div>
55<!-- *********************************************************************** -->
56
57<div class="doc_text">
58
59<p>Garbage collection is a widely used technique that frees the programmer from
60having to know the life-times of heap objects, making software easier to produce
61and maintain. Many programming languages rely on garbage collection for
62automatic memory management. There are two primary forms of garbage collection:
63conservative and accurate.</p>
64
65<p>Conservative garbage collection often does not require any special support
66from either the language or the compiler: it can handle non-type-safe
67programming languages (such as C/C++) and does not require any special
Jeff Cohen65fc36b2007-04-18 17:26:14 +000068information from the compiler. The
69<a href="http://www.hpl.hp.com/personal/Hans_Boehm/gc/">Boehm collector</a> is
70an example of a state-of-the-art conservative collector.</p>
Chris Lattner0d8c2db2004-05-23 21:02:20 +000071
72<p>Accurate garbage collection requires the ability to identify all pointers in
73the program at run-time (which requires that the source-language be type-safe in
74most cases). Identifying pointers at run-time requires compiler support to
75locate all places that hold live pointer variables at run-time, including the
76<a href="#roots">processor stack and registers</a>.</p>
77
78<p>
79Conservative garbage collection is attractive because it does not require any
80special compiler support, but it does have problems. In particular, because the
81conservative garbage collector cannot <i>know</i> that a particular word in the
82machine is a pointer, it cannot move live objects in the heap (preventing the
83use of compacting and generational GC algorithms) and it can occasionally suffer
84from memory leaks due to integer values that happen to point to objects in the
85program. In addition, some aggressive compiler transformations can break
86conservative garbage collectors (though these seem rare in practice).
87</p>
88
89<p>
90Accurate garbage collectors do not suffer from any of these problems, but they
91can suffer from degraded scalar optimization of the program. In particular,
92because the runtime must be able to identify and update all pointers active in
93the program, some optimizations are less effective. In practice, however, the
94locality and performance benefits of using aggressive garbage allocation
95techniques dominates any low-level losses.
96</p>
97
98<p>
99This document describes the mechanisms and interfaces provided by LLVM to
100support accurate garbage collection.
101</p>
102
103</div>
104
105<!-- ======================================================================= -->
106<div class="doc_subsection">
107 <a name="feature">GC features provided and algorithms supported</a>
108</div>
109
110<div class="doc_text">
111
112<p>
113LLVM provides support for a broad class of garbage collection algorithms,
114including compacting semi-space collectors, mark-sweep collectors, generational
115collectors, and even reference counting implementations. It includes support
116for <a href="#barriers">read and write barriers</a>, and associating <a
117href="#roots">meta-data with stack objects</a> (used for tagless garbage
118collection). All LLVM code generators support garbage collection, including the
119C backend.
120</p>
121
122<p>
123We hope that the primitive support built into LLVM is sufficient to support a
124broad class of garbage collected languages, including Scheme, ML, scripting
125languages, Java, C#, etc. That said, the implemented garbage collectors may
126need to be extended to support language-specific features such as finalization,
127weak references, or other features. As these needs are identified and
128implemented, they should be added to this specification.
129</p>
130
131<p>
132LLVM does not currently support garbage collection of multi-threaded programs or
133GC-safe points other than function calls, but these will be added in the future
134as there is interest.
135</p>
136
137</div>
138
139<!-- *********************************************************************** -->
140<div class="doc_section">
141 <a name="interfaces">Interfaces for user programs</a>
142</div>
143<!-- *********************************************************************** -->
144
145<div class="doc_text">
146
147<p>This section describes the interfaces provided by LLVM and by the garbage
148collector run-time that should be used by user programs. As such, this is the
149interface that front-end authors should generate code for.
150</p>
151
152</div>
153
154<!-- ======================================================================= -->
155<div class="doc_subsection">
156 <a name="roots">Identifying GC roots on the stack: <tt>llvm.gcroot</tt></a>
157</div>
158
159<div class="doc_text">
160
161<div class="doc_code"><tt>
162 void %llvm.gcroot(&lt;ty&gt;** %ptrloc, &lt;ty2&gt;* %metadata)
163</tt></div>
164
165<p>
166The <tt>llvm.gcroot</tt> intrinsic is used to inform LLVM of a pointer variable
167on the stack. The first argument contains the address of the variable on the
168stack, and the second contains a pointer to metadata that should be associated
Chris Lattner36597a52007-09-12 17:53:10 +0000169with the pointer (which <b>must</b> be a constant or global value address).</p>
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000170
171<p>
172Consider the following fragment of Java code:
173</p>
174
175<pre>
176 {
177 Object X; // A null-initialized reference to an object
178 ...
179 }
180</pre>
181
182<p>
183This block (which may be located in the middle of a function or in a loop nest),
184could be compiled to this LLVM code:
185</p>
186
187<pre>
188Entry:
189 ;; In the entry block for the function, allocate the
190 ;; stack space for X, which is an LLVM pointer.
191 %X = alloca %Object*
192 ...
193
Chris Lattner36597a52007-09-12 17:53:10 +0000194 ;; Java null-initializes pointers.
195 store %Object* null, %Object** %X
196
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000197 ;; "CodeBlock" is the block corresponding to the start
Reid Spencer03d186a2004-05-25 08:45:31 +0000198 ;; of the scope above.
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000199CodeBlock:
200 ;; Initialize the object, telling LLVM that it is now live.
201 ;; Java has type-tags on objects, so it doesn't need any
202 ;; metadata.
203 call void %llvm.gcroot(%Object** %X, sbyte* null)
204 ...
205
206 ;; As the pointer goes out of scope, store a null value into
207 ;; it, to indicate that the value is no longer live.
208 store %Object* null, %Object** %X
209 ...
210</pre>
211
212</div>
213
214<!-- ======================================================================= -->
215<div class="doc_subsection">
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000216 <a name="allocate">Allocating memory from the GC</a>
217</div>
218
219<div class="doc_text">
220
221<div class="doc_code"><tt>
222 sbyte *%llvm_gc_allocate(unsigned %Size)
223</tt></div>
224
225<p>The <tt>llvm_gc_allocate</tt> function is a global function defined by the
Chris Lattneraab3aff2004-06-09 03:59:05 +0000226garbage collector implementation to allocate memory. It returns a
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000227zeroed-out block of memory of the appropriate size.</p>
228
229</div>
230
231<!-- ======================================================================= -->
232<div class="doc_subsection">
233 <a name="barriers">Reading and writing references to the heap</a>
234</div>
235
236<div class="doc_text">
237
238<div class="doc_code"><tt>
Chris Lattner728f03f2004-07-22 05:49:38 +0000239 sbyte *%llvm.gcread(sbyte *, sbyte **)<br>
240 void %llvm.gcwrite(sbyte*, sbyte*, sbyte**)
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000241</tt></div>
242
243<p>Several of the more interesting garbage collectors (e.g., generational
244collectors) need to be informed when the mutator (the program that needs garbage
245collection) reads or writes object references into the heap. In the case of a
246generational collector, it needs to keep track of which "old" generation objects
247have references stored into them. The amount of code that typically needs to be
Chris Lattneraab3aff2004-06-09 03:59:05 +0000248executed is usually quite small (and not on the critical path of any
249computation), so the overall performance impact of the inserted code is
250tolerable.</p>
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000251
252<p>To support garbage collectors that use read or write barriers, LLVM provides
253the <tt>llvm.gcread</tt> and <tt>llvm.gcwrite</tt> intrinsics. The first
254intrinsic has exactly the same semantics as a non-volatile LLVM load and the
Chris Lattner728f03f2004-07-22 05:49:38 +0000255second has the same semantics as a non-volatile LLVM store, with the
256additions that they also take a pointer to the start of the memory
257object as an argument. At code generation
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000258time, these intrinsics are replaced with calls into the garbage collector
259(<tt><a href="#llvm_gc_readwrite">llvm_gc_read</a></tt> and <tt><a
260href="#llvm_gc_readwrite">llvm_gc_write</a></tt> respectively), which are then
261inlined into the code.
262</p>
263
264<p>
265If you are writing a front-end for a garbage collected language, every load or
266store of a reference from or to the heap should use these intrinsics instead of
267normal LLVM loads/stores.</p>
268
269</div>
270
271<!-- ======================================================================= -->
272<div class="doc_subsection">
273 <a name="initialize">Garbage collector startup and initialization</a>
274</div>
275
276<div class="doc_text">
277
278<div class="doc_code"><tt>
Chris Lattner9b2a1842004-05-27 05:52:10 +0000279 void %llvm_gc_initialize(unsigned %InitialHeapSize)
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000280</tt></div>
281
282<p>
283The <tt>llvm_gc_initialize</tt> function should be called once before any other
284garbage collection functions are called. This gives the garbage collector the
Chris Lattner9b2a1842004-05-27 05:52:10 +0000285chance to initialize itself and allocate the heap spaces. The initial heap size
286to allocate should be specified as an argument.
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000287</p>
288
289</div>
290
291<!-- ======================================================================= -->
292<div class="doc_subsection">
293 <a name="explicit">Explicit invocation of the garbage collector</a>
294</div>
295
296<div class="doc_text">
297
298<div class="doc_code"><tt>
299 void %llvm_gc_collect()
300</tt></div>
301
302<p>
303The <tt>llvm_gc_collect</tt> function is exported by the garbage collector
304implementations to provide a full collection, even when the heap is not
305exhausted. This can be used by end-user code as a hint, and may be ignored by
306the garbage collector.
307</p>
308
309</div>
310
311
312<!-- *********************************************************************** -->
313<div class="doc_section">
314 <a name="gcimpl">Implementing a garbage collector</a>
315</div>
316<!-- *********************************************************************** -->
317
318<div class="doc_text">
319
320<p>
Chris Lattner9b2a1842004-05-27 05:52:10 +0000321Implementing a garbage collector for LLVM is fairly straight-forward. The LLVM
322garbage collectors are provided in a form that makes them easy to link into the
323language-specific runtime that a language front-end would use. They require
324functionality from the language-specific runtime to get information about <a
325href="#gcdescriptors">where pointers are located in heap objects</a>.
326</p>
327
328<p>The
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000329implementation must include the <a
330href="#allocate"><tt>llvm_gc_allocate</tt></a> and <a
331href="#explicit"><tt>llvm_gc_collect</tt></a> functions, and it must implement
332the <a href="#llvm_gc_readwrite">read/write barrier</a> functions as well. To
333do this, it will probably have to <a href="#traceroots">trace through the roots
334from the stack</a> and understand the <a href="#gcdescriptors">GC descriptors
335for heap objects</a>. Luckily, there are some <a href="#gcimpls">example
336implementations</a> available.
337</p>
338</div>
339
340
341<!-- ======================================================================= -->
342<div class="doc_subsection">
343 <a name="llvm_gc_readwrite">Implementing <tt>llvm_gc_read</tt> and <tt>llvm_gc_write</tt></a>
344</div>
345
346<div class="doc_text">
347 <div class="doc_code"><tt>
Chris Lattner728f03f2004-07-22 05:49:38 +0000348 void *llvm_gc_read(void*, void **)<br>
349 void llvm_gc_write(void*, void *, void**)
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000350 </tt></div>
351
352<p>
353These functions <i>must</i> be implemented in every garbage collector, even if
354they do not need read/write barriers. In this case, just load or store the
355pointer, then return.
356</p>
357
358<p>
359If an actual read or write barrier is needed, it should be straight-forward to
Chris Lattner728f03f2004-07-22 05:49:38 +0000360implement it.
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000361</p>
362
363</div>
364
365<!-- ======================================================================= -->
366<div class="doc_subsection">
Chris Lattner0b02dbc2004-07-09 05:03:54 +0000367 <a name="callbacks">Callback functions used to implement the garbage collector</a>
Chris Lattner9b2a1842004-05-27 05:52:10 +0000368</div>
369
Chris Lattner0b02dbc2004-07-09 05:03:54 +0000370<div class="doc_text">
371<p>
Chris Lattner9b2a1842004-05-27 05:52:10 +0000372Garbage collector implementations make use of call-back functions that are
373implemented by other parts of the LLVM system.
Chris Lattner0b02dbc2004-07-09 05:03:54 +0000374</p>
375</div>
Chris Lattner9b2a1842004-05-27 05:52:10 +0000376
377<!--_________________________________________________________________________-->
378<div class="doc_subsubsection">
379 <a name="traceroots">Tracing GC pointers from the program stack</a>
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000380</div>
381
382<div class="doc_text">
383 <div class="doc_code"><tt>
384 void llvm_cg_walk_gcroots(void (*FP)(void **Root, void *Meta));
385 </tt></div>
386
387<p>
388The <tt>llvm_cg_walk_gcroots</tt> function is a function provided by the code
389generator that iterates through all of the GC roots on the stack, calling the
390specified function pointer with each record. For each GC root, the address of
391the pointer and the meta-data (from the <a
Duncan Sands8036ca42007-03-30 12:22:09 +0000392href="#roots"><tt>llvm.gcroot</tt></a> intrinsic) are provided.
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000393</p>
394</div>
395
Chris Lattner9b2a1842004-05-27 05:52:10 +0000396<!--_________________________________________________________________________-->
397<div class="doc_subsubsection">
398 <a name="staticroots">Tracing GC pointers from static roots</a>
399</div>
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000400
Chris Lattner9b2a1842004-05-27 05:52:10 +0000401<div class="doc_text">
402TODO
403</div>
404
405
406<!--_________________________________________________________________________-->
407<div class="doc_subsubsection">
408 <a name="gcdescriptors">Tracing GC pointers from heap objects</a>
409</div>
410
411<div class="doc_text">
412<p>
413The three most common ways to keep track of where pointers live in heap objects
414are (listed in order of space overhead required):</p>
415
416<ol>
417<li>In languages with polymorphic objects, pointers from an object header are
418usually used to identify the GC pointers in the heap object. This is common for
419object-oriented languages like Self, Smalltalk, Java, or C#.</li>
420
421<li>If heap objects are not polymorphic, often the "shape" of the heap can be
422determined from the roots of the heap or from some other meta-data [<a
423href="#appel89">Appel89</a>, <a href="#goldberg91">Goldberg91</a>, <a
424href="#tolmach94">Tolmach94</a>]. In this case, the garbage collector can
425propagate the information around from meta data stored with the roots. This
426often eliminates the need to have a header on objects in the heap. This is
427common in the ML family.</li>
428
429<li>If all heap objects have pointers in the same locations, or pointers can be
430distinguished just by looking at them (e.g., the low order bit is clear), no
431book-keeping is needed at all. This is common for Lisp-like languages.</li>
432</ol>
433
434<p>The LLVM garbage collectors are capable of supporting all of these styles of
435language, including ones that mix various implementations. To do this, it
436allows the source-language to associate meta-data with the <a
437href="#roots">stack roots</a>, and the heap tracing routines can propagate the
438information. In addition, LLVM allows the front-end to extract GC information
439from in any form from a specific object pointer (this supports situations #1 and
440#3).
441</p>
442
443<p><b>Making this efficient</b></p>
444
445
446
447</div>
448
449
450
451<!-- *********************************************************************** -->
452<div class="doc_section">
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000453 <a name="gcimpls">GC implementations available</a>
454</div>
Chris Lattner9b2a1842004-05-27 05:52:10 +0000455<!-- *********************************************************************** -->
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000456
457<div class="doc_text">
458
459<p>
460To make this more concrete, the currently implemented LLVM garbage collectors
Chris Lattner9b2a1842004-05-27 05:52:10 +0000461all live in the <tt>llvm/runtime/GC/*</tt> directories in the LLVM source-base.
462If you are interested in implementing an algorithm, there are many interesting
463possibilities (mark/sweep, a generational collector, a reference counting
464collector, etc), or you could choose to improve one of the existing algorithms.
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000465</p>
466
467</div>
468
Chris Lattner9b2a1842004-05-27 05:52:10 +0000469<!-- ======================================================================= -->
470<div class="doc_subsection">
Chris Lattner0b02dbc2004-07-09 05:03:54 +0000471 <a name="semispace">SemiSpace - A simple copying garbage collector</a>
Chris Lattner9b2a1842004-05-27 05:52:10 +0000472</div>
473
474<div class="doc_text">
475<p>
476SemiSpace is a very simple copying collector. When it starts up, it allocates
477two blocks of memory for the heap. It uses a simple bump-pointer allocator to
478allocate memory from the first block until it runs out of space. When it runs
479out of space, it traces through all of the roots of the program, copying blocks
480to the other half of the memory space.
481</p>
482
483</div>
484
485<!--_________________________________________________________________________-->
486<div class="doc_subsubsection">
487 Possible Improvements
488</div>
489
490<div class="doc_text">
491
492<p>
493If a collection cycle happens and the heap is not compacted very much (say less
494than 25% of the allocated memory was freed), the memory regions should be
495doubled in size.</p>
496
497</div>
498
499<!-- *********************************************************************** -->
500<div class="doc_section">
501 <a name="references">References</a>
502</div>
503<!-- *********************************************************************** -->
504
505<div class="doc_text">
506
507<p><a name="appel89">[Appel89]</a> Runtime Tags Aren't Necessary. Andrew
508W. Appel. Lisp and Symbolic Computation 19(7):703-705, July 1989.</p>
509
510<p><a name="goldberg91">[Goldberg91]</a> Tag-free garbage collection for
511strongly typed programming languages. Benjamin Goldberg. ACM SIGPLAN
512PLDI'91.</p>
513
514<p><a name="tolmach94">[Tolmach94]</a> Tag-free garbage collection using
515explicit type parameters. Andrew Tolmach. Proceedings of the 1994 ACM
516conference on LISP and functional programming.</p>
517
518</div>
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000519
520<!-- *********************************************************************** -->
521
522<hr>
523<address>
524 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
525 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
526 <a href="http://validator.w3.org/check/referer"><img
527 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!"></a>
528
529 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
Reid Spencer05fe4b02006-03-14 05:39:39 +0000530 <a href="http://llvm.org">LLVM Compiler Infrastructure</a><br>
Chris Lattner0d8c2db2004-05-23 21:02:20 +0000531 Last modified: $Date$
532</address>
533
534</body>
535</html>