| blob | 0accd0c78a09796cb876e7eea47086c36e42a8b7 |
3 <html>
4 <head>
7 </head>
8 <body>
11 Accurate Garbage Collection with LLVM
12 </div>
14 <ol>
16 <ul>
18 </ul>
19 </li>
22 <ul>
27 </ul>
28 </li>
31 <ul>
32 <li><a href="#llvm_gc_readwrite">Implementing <tt>llvm_gc_read</tt> and <tt>llvm_gc_write</tt></a></li>
34 </ul>
35 </li>
37 <ul>
39 </ul>
40 </li>
42 <!--
43 <li><a href="#codegen">Implementing GC support in a code generator</a></li>
44 -->
45 </ol>
49 </div>
51 <!-- *********************************************************************** -->
54 </div>
55 <!-- *********************************************************************** -->
59 <p>Garbage collection is a widely used technique that frees the programmer from
60 having to know the life-times of heap objects, making software easier to produce
61 and maintain. Many programming languages rely on garbage collection for
62 automatic memory management. There are two primary forms of garbage collection:
63 conservative and accurate.</p>
65 <p>Conservative garbage collection often does not require any special support
66 from either the language or the compiler: it can handle non-type-safe
67 programming languages (such as C/C++) and does not require any special
68 information from the compiler. The
70 an example of a state-of-the-art conservative collector.</p>
72 <p>Accurate garbage collection requires the ability to identify all pointers in
73 the program at run-time (which requires that the source-language be type-safe in
74 most cases). Identifying pointers at run-time requires compiler support to
75 locate all places that hold live pointer variables at run-time, including the
78 <p>
79 Conservative garbage collection is attractive because it does not require any
80 special compiler support, but it does have problems. In particular, because the
82 machine is a pointer, it cannot move live objects in the heap (preventing the
83 use of compacting and generational GC algorithms) and it can occasionally suffer
84 from memory leaks due to integer values that happen to point to objects in the
85 program. In addition, some aggressive compiler transformations can break
86 conservative garbage collectors (though these seem rare in practice).
87 </p>
89 <p>
90 Accurate garbage collectors do not suffer from any of these problems, but they
91 can suffer from degraded scalar optimization of the program. In particular,
92 because the runtime must be able to identify and update all pointers active in
93 the program, some optimizations are less effective. In practice, however, the
94 locality and performance benefits of using aggressive garbage allocation
95 techniques dominates any low-level losses.
96 </p>
98 <p>
99 This document describes the mechanisms and interfaces provided by LLVM to
100 support accurate garbage collection.
101 </p>
103 </div>
105 <!-- ======================================================================= -->
108 </div>
112 <p>
113 LLVM provides support for a broad class of garbage collection algorithms,
114 including compacting semi-space collectors, mark-sweep collectors, generational
115 collectors, and even reference counting implementations. It includes support
118 collection). All LLVM code generators support garbage collection, including the
119 C backend.
120 </p>
122 <p>
123 We hope that the primitive support built into LLVM is sufficient to support a
124 broad class of garbage collected languages, including Scheme, ML, scripting
125 languages, Java, C#, etc. That said, the implemented garbage collectors may
126 need to be extended to support language-specific features such as finalization,
127 weak references, or other features. As these needs are identified and
128 implemented, they should be added to this specification.
129 </p>
131 <p>
132 LLVM does not currently support garbage collection of multi-threaded programs or
133 GC-safe points other than function calls, but these will be added in the future
134 as there is interest.
135 </p>
137 </div>
139 <!-- *********************************************************************** -->
142 </div>
143 <!-- *********************************************************************** -->
147 <p>This section describes the interfaces provided by LLVM and by the garbage
148 collector run-time that should be used by user programs. As such, this is the
149 interface that front-end authors should generate code for.
150 </p>
152 </div>
154 <!-- ======================================================================= -->
157 </div>
163 </tt></div>
165 <p>
167 on the stack. The first argument contains the address of the variable on the
168 stack, and the second contains a pointer to metadata that should be associated
171 specified location to initialize the pointer.</p>
173 <p>
174 Consider the following fragment of Java code:
175 </p>
177 <pre>
178 {
179 Object X; // A null-initialized reference to an object
180 ...
181 }
182 </pre>
184 <p>
185 This block (which may be located in the middle of a function or in a loop nest),
186 could be compiled to this LLVM code:
187 </p>
189 <pre>
190 Entry:
191 ;; In the entry block for the function, allocate the
192 ;; stack space for X, which is an LLVM pointer.
193 %X = alloca %Object*
194 ...
196 ;; "CodeBlock" is the block corresponding to the start
197 ;; of the scope above.
198 CodeBlock:
199 ;; Initialize the object, telling LLVM that it is now live.
200 ;; Java has type-tags on objects, so it doesn't need any
201 ;; metadata.
202 call void %llvm.gcroot(%Object** %X, sbyte* null)
203 ...
205 ;; As the pointer goes out of scope, store a null value into
206 ;; it, to indicate that the value is no longer live.
207 store %Object* null, %Object** %X
208 ...
209 </pre>
211 </div>
213 <!-- ======================================================================= -->
216 </div>
221 sbyte *%llvm_gc_allocate(unsigned %Size)
222 </tt></div>
225 garbage collector implementation to allocate memory. It returns a
226 zeroed-out block of memory of the appropriate size.</p>
228 </div>
230 <!-- ======================================================================= -->
233 </div>
238 sbyte *%llvm.gcread(sbyte *, sbyte **)<br>
239 void %llvm.gcwrite(sbyte*, sbyte*, sbyte**)
240 </tt></div>
242 <p>Several of the more interesting garbage collectors (e.g., generational
243 collectors) need to be informed when the mutator (the program that needs garbage
244 collection) reads or writes object references into the heap. In the case of a
245 generational collector, it needs to keep track of which "old" generation objects
246 have references stored into them. The amount of code that typically needs to be
247 executed is usually quite small (and not on the critical path of any
248 computation), so the overall performance impact of the inserted code is
249 tolerable.</p>
251 <p>To support garbage collectors that use read or write barriers, LLVM provides
253 intrinsic has exactly the same semantics as a non-volatile LLVM load and the
254 second has the same semantics as a non-volatile LLVM store, with the
255 additions that they also take a pointer to the start of the memory
256 object as an argument. At code generation
257 time, these intrinsics are replaced with calls into the garbage collector
260 inlined into the code.
261 </p>
263 <p>
264 If you are writing a front-end for a garbage collected language, every load or
265 store of a reference from or to the heap should use these intrinsics instead of
266 normal LLVM loads/stores.</p>
268 </div>
270 <!-- ======================================================================= -->
273 </div>
278 void %llvm_gc_initialize(unsigned %InitialHeapSize)
279 </tt></div>
281 <p>
283 garbage collection functions are called. This gives the garbage collector the
284 chance to initialize itself and allocate the heap spaces. The initial heap size
285 to allocate should be specified as an argument.
286 </p>
288 </div>
290 <!-- ======================================================================= -->
293 </div>
298 void %llvm_gc_collect()
299 </tt></div>
301 <p>
303 implementations to provide a full collection, even when the heap is not
304 exhausted. This can be used by end-user code as a hint, and may be ignored by
305 the garbage collector.
306 </p>
308 </div>
311 <!-- *********************************************************************** -->
314 </div>
315 <!-- *********************************************************************** -->
319 <p>
320 Implementing a garbage collector for LLVM is fairly straight-forward. The LLVM
321 garbage collectors are provided in a form that makes them easy to link into the
322 language-specific runtime that a language front-end would use. They require
323 functionality from the language-specific runtime to get information about <a
325 </p>
327 <p>The
328 implementation must include the <a
335 implementations</a> available.
336 </p>
337 </div>
340 <!-- ======================================================================= -->
343 </div>
347 void *llvm_gc_read(void*, void **)<br>
348 void llvm_gc_write(void*, void *, void**)
349 </tt></div>
351 <p>
353 they do not need read/write barriers. In this case, just load or store the
354 pointer, then return.
355 </p>
357 <p>
358 If an actual read or write barrier is needed, it should be straight-forward to
359 implement it.
360 </p>
362 </div>
364 <!-- ======================================================================= -->
367 </div>
370 <p>
371 Garbage collector implementations make use of call-back functions that are
372 implemented by other parts of the LLVM system.
373 </p>
374 </div>
376 <!--_________________________________________________________________________-->
379 </div>
383 void llvm_cg_walk_gcroots(void (*FP)(void **Root, void *Meta));
384 </tt></div>
386 <p>
388 generator that iterates through all of the GC roots on the stack, calling the
389 specified function pointer with each record. For each GC root, the address of
390 the pointer and the meta-data (from the <a
392 </p>
393 </div>
395 <!--_________________________________________________________________________-->
398 </div>
401 TODO
402 </div>
405 <!--_________________________________________________________________________-->
408 </div>
411 <p>
412 The three most common ways to keep track of where pointers live in heap objects
413 are (listed in order of space overhead required):</p>
415 <ol>
416 <li>In languages with polymorphic objects, pointers from an object header are
417 usually used to identify the GC pointers in the heap object. This is common for
418 object-oriented languages like Self, Smalltalk, Java, or C#.</li>
421 determined from the roots of the heap or from some other meta-data [<a
424 propagate the information around from meta data stored with the roots. This
425 often eliminates the need to have a header on objects in the heap. This is
426 common in the ML family.</li>
428 <li>If all heap objects have pointers in the same locations, or pointers can be
429 distinguished just by looking at them (e.g., the low order bit is clear), no
430 book-keeping is needed at all. This is common for Lisp-like languages.</li>
431 </ol>
433 <p>The LLVM garbage collectors are capable of supporting all of these styles of
434 language, including ones that mix various implementations. To do this, it
435 allows the source-language to associate meta-data with the <a
437 information. In addition, LLVM allows the front-end to extract GC information
438 from in any form from a specific object pointer (this supports situations #1 and
439 #3).
440 </p>
446 </div>
450 <!-- *********************************************************************** -->
453 </div>
454 <!-- *********************************************************************** -->
458 <p>
459 To make this more concrete, the currently implemented LLVM garbage collectors
461 If you are interested in implementing an algorithm, there are many interesting
462 possibilities (mark/sweep, a generational collector, a reference counting
463 collector, etc), or you could choose to improve one of the existing algorithms.
464 </p>
466 </div>
468 <!-- ======================================================================= -->
471 </div>
474 <p>
475 SemiSpace is a very simple copying collector. When it starts up, it allocates
476 two blocks of memory for the heap. It uses a simple bump-pointer allocator to
477 allocate memory from the first block until it runs out of space. When it runs
478 out of space, it traces through all of the roots of the program, copying blocks
479 to the other half of the memory space.
480 </p>
482 </div>
484 <!--_________________________________________________________________________-->
486 Possible Improvements
487 </div>
491 <p>
492 If a collection cycle happens and the heap is not compacted very much (say less
493 than 25% of the allocated memory was freed), the memory regions should be
494 doubled in size.</p>
496 </div>
498 <!-- *********************************************************************** -->
501 </div>
502 <!-- *********************************************************************** -->
510 strongly typed programming languages. Benjamin Goldberg. ACM SIGPLAN
514 explicit type parameters. Andrew Tolmach. Proceedings of the 1994 ACM
515 conference on LISP and functional programming.</p>
517 </div>
519 <!-- *********************************************************************** -->
521 <hr>
522 <address>
530 Last modified: $Date$
531 </address>
533 </body>
534 </html>