1 =====================================
2 Garbage Collection with LLVM
3 =====================================
11 This document covers how to integrate LLVM into a compiler for a language which
12 supports garbage collection. **Note that LLVM itself does not provide a
13 garbage collector.** You must provide your own.
18 First, you should pick a collector strategy. LLVM includes a number of built
19 in ones, but you can also implement a loadable plugin with a custom definition.
20 Note that the collector strategy is a description of how LLVM should generate
21 code such that it interacts with your collector and runtime, not a description
22 of the collector itself.
24 Next, mark your generated functions as using your chosen collector strategy.
25 From c++, you can call:
29 F.setGC(<collector description name>);
32 This will produce IR like the following fragment:
36 define void @foo() gc "<collector description name>" { ... }
39 When generating LLVM IR for your functions, you will need to:
41 * Use ``@llvm.gcread`` and/or ``@llvm.gcwrite`` in place of standard load and
42 store instructions. These intrinsics are used to represent load and store
43 barriers. If you collector does not require such barriers, you can skip
46 * Use the memory allocation routines provided by your garbage collector's
49 * If your collector requires them, generate type maps according to your
50 runtime's binary interface. LLVM is not involved in the process. In
51 particular, the LLVM type system is not suitable for conveying such
52 information though the compiler.
54 * Insert any coordination code required for interacting with your collector.
55 Many collectors require running application code to periodically check a
56 flag and conditionally call a runtime function. This is often referred to
59 You will need to identify roots (i.e. references to heap objects your collector
60 needs to know about) in your generated IR, so that LLVM can encode them into
61 your final stack maps. Depending on the collector strategy chosen, this is
62 accomplished by using either the ``@llvm.gcroot`` intrinsics or an
63 ``gc.statepoint`` relocation sequence.
65 Don't forget to create a root for each intermediate value that is generated when
66 evaluating an expression. In ``h(f(), g())``, the result of ``f()`` could
67 easily be collected if evaluating ``g()`` triggers a collection.
69 Finally, you need to link your runtime library with the generated program
70 executable (for a static compiler) or ensure the appropriate symbols are
71 available for the runtime linker (for a JIT compiler).
77 What is Garbage Collection?
78 ---------------------------
80 Garbage collection is a widely used technique that frees the programmer from
81 having to know the lifetimes of heap objects, making software easier to produce
82 and maintain. Many programming languages rely on garbage collection for
83 automatic memory management. There are two primary forms of garbage collection:
84 conservative and accurate.
86 Conservative garbage collection often does not require any special support from
87 either the language or the compiler: it can handle non-type-safe programming
88 languages (such as C/C++) and does not require any special information from the
89 compiler. The `Boehm collector
90 <http://www.hpl.hp.com/personal/Hans_Boehm/gc/>`__ is an example of a
91 state-of-the-art conservative collector.
93 Accurate garbage collection requires the ability to identify all pointers in the
94 program at run-time (which requires that the source-language be type-safe in
95 most cases). Identifying pointers at run-time requires compiler support to
96 locate all places that hold live pointer variables at run-time, including the
97 :ref:`processor stack and registers <gcroot>`.
99 Conservative garbage collection is attractive because it does not require any
100 special compiler support, but it does have problems. In particular, because the
101 conservative garbage collector cannot *know* that a particular word in the
102 machine is a pointer, it cannot move live objects in the heap (preventing the
103 use of compacting and generational GC algorithms) and it can occasionally suffer
104 from memory leaks due to integer values that happen to point to objects in the
105 program. In addition, some aggressive compiler transformations can break
106 conservative garbage collectors (though these seem rare in practice).
108 Accurate garbage collectors do not suffer from any of these problems, but they
109 can suffer from degraded scalar optimization of the program. In particular,
110 because the runtime must be able to identify and update all pointers active in
111 the program, some optimizations are less effective. In practice, however, the
112 locality and performance benefits of using aggressive garbage collection
113 techniques dominates any low-level losses.
115 This document describes the mechanisms and interfaces provided by LLVM to
116 support accurate garbage collection.
121 LLVM's intermediate representation provides :ref:`garbage collection intrinsics
122 <gc_intrinsics>` that offer support for a broad class of collector models. For
123 instance, the intrinsics permit:
125 * semi-space collectors
127 * mark-sweep collectors
129 * generational collectors
131 * incremental collectors
133 * concurrent collectors
135 * cooperative collectors
139 We hope that the support built into the LLVM IR is sufficient to support a
140 broad class of garbage collected languages including Scheme, ML, Java, C#,
141 Perl, Python, Lua, Ruby, other scripting languages, and more.
143 Note that LLVM **does not itself provide a garbage collector** --- this should
144 be part of your language's runtime library. LLVM provides a framework for
145 describing the garbage collectors requirements to the compiler. In particular,
146 LLVM provides support for generating stack maps at call sites, polling for a
147 safepoint, and emitting load and store barriers. You can also extend LLVM -
148 possibly through a loadable :ref:`code generation plugins <plugin>` - to
149 generate code and data structures which conforms to the *binary interface*
150 specified by the *runtime library*. This is similar to the relationship between
151 LLVM and DWARF debugging info, for example. The difference primarily lies in
152 the lack of an established standard in the domain of garbage collection --- thus
153 the need for a flexible extension mechanism.
155 The aspects of the binary interface with which LLVM's GC support is
158 * Creation of GC safepoints within code where collection is allowed to execute
161 * Computation of the stack map. For each safe point in the code, object
162 references within the stack frame must be identified so that the collector may
163 traverse and perhaps update them.
165 * Write barriers when storing object references to the heap. These are commonly
166 used to optimize incremental scans in generational collectors.
168 * Emission of read barriers when loading object references. These are useful
169 for interoperating with concurrent collectors.
171 There are additional areas that LLVM does not directly address:
173 * Registration of global roots with the runtime.
175 * Registration of stack map entries with the runtime.
177 * The functions used by the program to allocate memory, trigger a collection,
180 * Computation or compilation of type maps, or registration of them with the
181 runtime. These are used to crawl the heap for object references.
183 In general, LLVM's support for GC does not include features which can be
184 adequately addressed with other features of the IR and does not specify a
185 particular binary interface. On the plus side, this means that you should be
186 able to integrate LLVM with an existing runtime. On the other hand, it can
187 have the effect of leaving a lot of work for the developer of a novel
188 language. We try to mitigate this by providing built in collector strategy
189 descriptions that can work with many common collector designs and easy
190 extension points. If you don't already have a specific binary interface
191 you need to support, we recommend trying to use one of these built in collector
199 This section describes the garbage collection facilities provided by the
200 :doc:`LLVM intermediate representation <LangRef>`. The exact behavior of these
201 IR features is specified by the selected :ref:`GC strategy description
204 Specifying GC code generation: ``gc "..."``
205 -------------------------------------------
209 define <returntype> @name(...) gc "name" { ... }
211 The ``gc`` function attribute is used to specify the desired GC strategy to the
212 compiler. Its programmatic equivalent is the ``setGC`` method of ``Function``.
214 Setting ``gc "name"`` on a function triggers a search for a matching subclass
215 of GCStrategy. Some collector strategies are built in. You can add others
216 using either the loadable plugin mechanism, or by patching your copy of LLVM.
217 It is the selected GC strategy which defines the exact nature of the code
218 generated to support GC. If none is found, the compiler will raise an error.
220 Specifying the GC style on a per-function basis allows LLVM to link together
221 programs that use different garbage collection algorithms (or none at all).
225 Identifying GC roots on the stack
226 ----------------------------------
228 LLVM currently supports two different mechanisms for describing references in
229 compiled code at safepoints. ``llvm.gcroot`` is the older mechanism;
230 ``gc.statepoint`` has been added more recently. At the moment, you can choose
231 either implementation (on a per :ref:`GC strategy <plugin>` basis). Longer
232 term, we will probably either migrate away from ``llvm.gcroot`` entirely, or
233 substantially merge their implementations. Note that most new development
234 work is focused on ``gc.statepoint``.
236 Using ``gc.statepoint``
237 ^^^^^^^^^^^^^^^^^^^^^^^^
238 :doc:`This page <Statepoints>` contains detailed documentation for
241 Using ``llvm.gcwrite``
242 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
246 void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
248 The ``llvm.gcroot`` intrinsic is used to inform LLVM that a stack variable
249 references an object on the heap and is to be tracked for garbage collection.
250 The exact impact on generated code is specified by the Function's selected
251 :ref:`GC strategy <plugin>`. All calls to ``llvm.gcroot`` **must** reside
252 inside the first basic block.
254 The first argument **must** be a value referring to an alloca instruction or a
255 bitcast of an alloca. The second contains a pointer to metadata that should be
256 associated with the pointer, and **must** be a constant or global value
257 address. If your target collector uses tags, use a null pointer for metadata.
259 A compiler which performs manual SSA construction **must** ensure that SSA
260 values representing GC references are stored in to the alloca passed to the
261 respective ``gcroot`` before every call site and reloaded after every call.
262 A compiler which uses mem2reg to raise imperative code using ``alloca`` into
263 SSA form need only add a call to ``@llvm.gcroot`` for those variables which
264 are pointers into the GC heap.
266 It is also important to mark intermediate values with ``llvm.gcroot``. For
267 example, consider ``h(f(), g())``. Beware leaking the result of ``f()`` in the
268 case that ``g()`` triggers a collection. Note, that stack variables must be
269 initialized and marked with ``llvm.gcroot`` in function's prologue.
271 The ``%metadata`` argument can be used to avoid requiring heap objects to have
272 'isa' pointers or tag bits. [Appel89_, Goldberg91_, Tolmach94_] If specified,
273 its value will be tracked along with the location of the pointer in the stack
276 Consider the following fragment of Java code:
281 Object X; // A null-initialized reference to an object
285 This block (which may be located in the middle of a function or in a loop nest),
286 could be compiled to this LLVM code:
291 ;; In the entry block for the function, allocate the
292 ;; stack space for X, which is an LLVM pointer.
295 ;; Tell LLVM that the stack space is a stack root.
296 ;; Java has type-tags on objects, so we pass null as metadata.
297 %tmp = bitcast %Object** %X to i8**
298 call void @llvm.gcroot(i8** %tmp, i8* null)
301 ;; "CodeBlock" is the block corresponding to the start
302 ;; of the scope above.
304 ;; Java null-initializes pointers.
305 store %Object* null, %Object** %X
309 ;; As the pointer goes out of scope, store a null value into
310 ;; it, to indicate that the value is no longer live.
311 store %Object* null, %Object** %X
314 Reading and writing references in the heap
315 ------------------------------------------
317 Some collectors need to be informed when the mutator (the program that needs
318 garbage collection) either reads a pointer from or writes a pointer to a field
319 of a heap object. The code fragments inserted at these points are called *read
320 barriers* and *write barriers*, respectively. The amount of code that needs to
321 be executed is usually quite small and not on the critical path of any
322 computation, so the overall performance impact of the barrier is tolerable.
324 Barriers often require access to the *object pointer* rather than the *derived
325 pointer* (which is a pointer to the field within the object). Accordingly,
326 these intrinsics take both pointers as separate arguments for completeness. In
327 this snippet, ``%object`` is the object pointer, and ``%derived`` is the derived
333 %class.Array = type { %class.Object, i32, [0 x %class.Object*] }
336 ;; Load the object pointer from a gcroot.
337 %object = load %class.Array** %object_addr
339 ;; Compute the derived pointer.
340 %derived = getelementptr %object, i32 0, i32 2, i32 %n
342 LLVM does not enforce this relationship between the object and derived pointer
343 (although a particular :ref:`collector strategy <plugin>` might). However, it
344 would be an unusual collector that violated it.
346 The use of these intrinsics is naturally optional if the target GC does not
347 require the corresponding barrier. The GC strategy used with such a collector
348 should replace the intrinsic calls with the corresponding ``load`` or
349 ``store`` instruction if they are used.
351 One known deficiency with the current design is that the barrier intrinsics do
352 not include the size or alignment of the underlying operation performed. It is
353 currently assumed that the operation is of pointer size and the alignment is
354 assumed to be the target machine's default alignment.
356 Write barrier: ``llvm.gcwrite``
357 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
361 void @llvm.gcwrite(i8* %value, i8* %object, i8** %derived)
363 For write barriers, LLVM provides the ``llvm.gcwrite`` intrinsic function. It
364 has exactly the same semantics as a non-volatile ``store`` to the derived
365 pointer (the third argument). The exact code generated is specified by the
366 Function's selected :ref:`GC strategy <plugin>`.
368 Many important algorithms require write barriers, including generational and
369 concurrent collectors. Additionally, write barriers could be used to implement
372 Read barrier: ``llvm.gcread``
373 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
377 i8* @llvm.gcread(i8* %object, i8** %derived)
379 For read barriers, LLVM provides the ``llvm.gcread`` intrinsic function. It has
380 exactly the same semantics as a non-volatile ``load`` from the derived pointer
381 (the second argument). The exact code generated is specified by the Function's
382 selected :ref:`GC strategy <plugin>`.
384 Read barriers are needed by fewer algorithms than write barriers, and may have a
385 greater performance impact since pointer reads are more frequent than writes.
389 .. _builtin-gc-strategies:
391 Built In GC Strategies
392 ======================
394 LLVM includes built in support for several varieties of garbage collectors.
397 ----------------------
399 To use this collector strategy, mark your functions with:
403 F.setGC("shadow-stack");
405 Unlike many GC algorithms which rely on a cooperative code generator to compile
406 stack maps, this algorithm carefully maintains a linked list of stack roots
407 [:ref:`Henderson2002 <henderson02>`]. This so-called "shadow stack" mirrors the
408 machine stack. Maintaining this data structure is slower than using a stack map
409 compiled into the executable as constant data, but has a significant portability
410 advantage because it requires no special support from the target code generator,
411 and does not require tricky platform-specific code to crawl the machine stack.
413 The tradeoff for this simplicity and portability is:
415 * High overhead per function call.
419 Still, it's an easy way to get started. After your compiler and runtime are up
420 and running, writing a :ref:`plugin <plugin>` will allow you to take advantage
421 of :ref:`more advanced GC features <collector-algos>` of LLVM in order to
425 The shadow stack doesn't imply a memory allocation algorithm. A semispace
426 collector or building atop ``malloc`` are great places to start, and can be
427 implemented with very little code.
429 When it comes time to collect, however, your runtime needs to traverse the stack
430 roots, and for this it needs to integrate with the shadow stack. Luckily, doing
431 so is very simple. (This code is heavily commented to help you understand the
432 data structure, but there are only 20 lines of meaningful code.)
436 /// @brief The map for a single function's stack frame. One of these is
437 /// compiled as constant data into the executable for each function.
439 /// Storage of metadata values is elided if the %metadata parameter to
440 /// @llvm.gcroot is null.
442 int32_t NumRoots; //< Number of roots in stack frame.
443 int32_t NumMeta; //< Number of metadata entries. May be < NumRoots.
444 const void *Meta[0]; //< Metadata for each root.
447 /// @brief A link in the dynamic shadow stack. One of these is embedded in
448 /// the stack frame of each function on the call stack.
450 StackEntry *Next; //< Link to next stack entry (the caller's).
451 const FrameMap *Map; //< Pointer to constant FrameMap.
452 void *Roots[0]; //< Stack roots (in-place array).
455 /// @brief The head of the singly-linked list of StackEntries. Functions push
456 /// and pop onto this in their prologue and epilogue.
458 /// Since there is only a global list, this technique is not threadsafe.
459 StackEntry *llvm_gc_root_chain;
461 /// @brief Calls Visitor(root, meta) for each GC root on the stack.
462 /// root and meta are exactly the values passed to
465 /// Visitor could be a function to recursively mark live objects. Or it
466 /// might copy them to another heap or generation.
468 /// @param Visitor A function to invoke for every GC root on the stack.
469 void visitGCRoots(void (*Visitor)(void **Root, const void *Meta)) {
470 for (StackEntry *R = llvm_gc_root_chain; R; R = R->Next) {
473 // For roots [0, NumMeta), the metadata pointer is in the FrameMap.
474 for (unsigned e = R->Map->NumMeta; i != e; ++i)
475 Visitor(&R->Roots[i], R->Map->Meta[i]);
477 // For roots [NumMeta, NumRoots), the metadata pointer is null.
478 for (unsigned e = R->Map->NumRoots; i != e; ++i)
479 Visitor(&R->Roots[i], NULL);
484 The 'Erlang' and 'Ocaml' GCs
485 -----------------------------
487 LLVM ships with two example collectors which leverage the ``gcroot``
488 mechanisms. To our knowledge, these are not actually used by any language
489 runtime, but they do provide a reasonable starting point for someone interested
490 in writing an ``gcroot`` compatible GC plugin. In particular, these are the
491 only in tree examples of how to produce a custom binary stack map format using
492 a ``gcroot`` strategy.
494 As there names imply, the binary format produced is intended to model that
495 used by the Erlang and OCaml compilers respectively.
498 The Statepoint Example GC
499 -------------------------
503 F.setGC("statepoint-example");
505 This GC provides an example of how one might use the infrastructure provided
506 by ``gc.statepoint``. This example GC is compatible with the
507 :ref:`PlaceSafepoints` and :ref:`RewriteStatepointsForGC` utility passes
508 which simplify ``gc.statepoint`` sequence insertion. If you need to build a
509 custom GC strategy around the ``gc.statepoints`` mechanisms, it is recommended
510 that you use this one as a starting point.
512 This GC strategy does not support read or write barriers. As a result, these
513 intrinsics are lowered to normal loads and stores.
515 The stack map format generated by this GC strategy can be found in the
516 :ref:`stackmap-section` using a format documented :ref:`here
517 <statepoint-stackmap-format>`. This format is intended to be the standard
518 format supported by LLVM going forward.
524 If none of the built in GC strategy descriptions met your needs above, you will
525 need to define a custom GCStrategy and possibly, a custom LLVM pass to perform
526 lowering. Your best example of where to start defining a custom GCStrategy
527 would be to look at one of the built in strategies.
529 You may be able to structure this additional code as a loadable plugin library.
530 Loadable plugins are sufficient if all you need is to enable a different
531 combination of built in functionality, but if you need to provide a custom
532 lowering pass, you will need to build a patched version of LLVM. If you think
533 you need a patched build, please ask for advice on llvm-dev. There may be an
534 easy way we can extend the support to make it work for your use case without
535 requiring a custom build.
537 Collector Requirements
538 ----------------------
540 You should be able to leverage any existing collector library that includes the following elements:
542 #. A memory allocator which exposes an allocation function your compiled
545 #. A binary format for the stack map. A stack map describes the location
546 of references at a safepoint and is used by precise collectors to identify
547 references within a stack frame on the machine stack. Note that collectors
548 which conservatively scan the stack don't require such a structure.
550 #. A stack crawler to discover functions on the call stack, and enumerate the
551 references listed in the stack map for each call site.
553 #. A mechanism for identifying references in global locations (e.g. global
556 #. If you collector requires them, an LLVM IR implementation of your collectors
557 load and store barriers. Note that since many collectors don't require
558 barriers at all, LLVM defaults to lowering such barriers to normal loads
559 and stores unless you arrange otherwise.
562 Implementing a collector plugin
563 -------------------------------
565 User code specifies which GC code generation to use with the ``gc`` function
566 attribute or, equivalently, with the ``setGC`` method of ``Function``.
568 To implement a GC plugin, it is necessary to subclass ``llvm::GCStrategy``,
569 which can be accomplished in a few lines of boilerplate code. LLVM's
570 infrastructure provides access to several important algorithms. For an
571 uncontroversial collector, all that remains may be to compile LLVM's computed
572 stack map to assembly code (using the binary representation expected by the
573 runtime library). This can be accomplished in about 100 lines of code.
575 This is not the appropriate place to implement a garbage collected heap or a
576 garbage collector itself. That code should exist in the language's runtime
577 library. The compiler plugin is responsible for generating code which conforms
578 to the binary interface defined by library, most essentially the :ref:`stack map
581 To subclass ``llvm::GCStrategy`` and register it with the compiler:
585 // lib/MyGC/MyGC.cpp - Example LLVM GC plugin
587 #include "llvm/CodeGen/GCStrategy.h"
588 #include "llvm/CodeGen/GCMetadata.h"
589 #include "llvm/Support/Compiler.h"
591 using namespace llvm;
594 class LLVM_LIBRARY_VISIBILITY MyGC : public GCStrategy {
599 GCRegistry::Add<MyGC>
600 X("mygc", "My bespoke garbage collector.");
603 This boilerplate collector does nothing. More specifically:
605 * ``llvm.gcread`` calls are replaced with the corresponding ``load``
608 * ``llvm.gcwrite`` calls are replaced with the corresponding ``store``
611 * No safe points are added to the code.
613 * The stack map is not compiled into the executable.
615 Using the LLVM makefiles, this code
616 can be compiled as a plugin using a simple makefile:
626 include $(LEVEL)/Makefile.common
628 Once the plugin is compiled, code using it may be compiled using ``llc
629 -load=MyGC.so`` (though MyGC.so may have some other platform-specific
635 define void @f() gc "mygc" {
639 $ llvm-as < sample.ll | llc -load=MyGC.so
641 It is also possible to statically link the collector plugin into tools, such as
642 a language-specific compiler front-end.
646 Overview of available features
647 ------------------------------
649 ``GCStrategy`` provides a range of features through which a plugin may do useful
650 work. Some of these are callbacks, some are algorithms that can be enabled,
651 disabled, or customized. This matrix summarizes the supported (and planned)
652 features and correlates them with the collection techniques which typically
655 .. |v| unicode:: 0x2714
658 .. |x| unicode:: 0x2718
661 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
662 | Algorithm | Done | Shadow | refcount | mark- | copying | incremental | threaded | concurrent |
663 | | | stack | | sweep | | | | |
664 +============+======+========+==========+=======+=========+=============+==========+============+
665 | stack map | |v| | | | |x| | |x| | |x| | |x| | |x| |
666 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
667 | initialize | |v| | |x| | |x| | |x| | |x| | |x| | |x| | |x| |
668 | roots | | | | | | | | |
669 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
670 | derived | NO | | | | | | **N**\* | **N**\* |
671 | pointers | | | | | | | | |
672 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
673 | **custom | |v| | | | | | | | |
674 | lowering** | | | | | | | | |
675 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
676 | *gcroot* | |v| | |x| | |x| | | | | | |
677 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
678 | *gcwrite* | |v| | | |x| | | | |x| | | |x| |
679 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
680 | *gcread* | |v| | | | | | | | |x| |
681 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
682 | **safe | | | | | | | | |
683 | points** | | | | | | | | |
684 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
685 | *in | |v| | | | |x| | |x| | |x| | |x| | |x| |
686 | calls* | | | | | | | | |
687 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
688 | *before | |v| | | | | | | |x| | |x| |
689 | calls* | | | | | | | | |
690 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
691 | *for | NO | | | | | | **N** | **N** |
692 | loops* | | | | | | | | |
693 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
694 | *before | |v| | | | | | | |x| | |x| |
695 | escape* | | | | | | | | |
696 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
697 | emit code | NO | | | | | | **N** | **N** |
698 | at safe | | | | | | | | |
699 | points | | | | | | | | |
700 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
701 | **output** | | | | | | | | |
702 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
703 | *assembly* | |v| | | | |x| | |x| | |x| | |x| | |x| |
704 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
705 | *JIT* | NO | | | **?** | **?** | **?** | **?** | **?** |
706 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
707 | *obj* | NO | | | **?** | **?** | **?** | **?** | **?** |
708 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
709 | live | NO | | | **?** | **?** | **?** | **?** | **?** |
710 | analysis | | | | | | | | |
711 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
712 | register | NO | | | **?** | **?** | **?** | **?** | **?** |
713 | map | | | | | | | | |
714 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
715 | \* Derived pointers only pose a hasard to copying collections. |
716 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
717 | **?** denotes a feature which could be utilized if available. |
718 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
720 To be clear, the collection techniques above are defined as:
723 The mutator carefully maintains a linked list of stack roots.
726 The mutator maintains a reference count for each object and frees an object
727 when its count falls to zero.
730 When the heap is exhausted, the collector marks reachable objects starting
731 from the roots, then deallocates unreachable objects in a sweep phase.
734 As reachability analysis proceeds, the collector copies objects from one heap
735 area to another, compacting them in the process. Copying collectors enable
736 highly efficient "bump pointer" allocation and can improve locality of
740 (Including generational collectors.) Incremental collectors generally have all
741 the properties of a copying collector (regardless of whether the mature heap
742 is compacting), but bring the added complexity of requiring write barriers.
745 Denotes a multithreaded mutator; the collector must still stop the mutator
746 ("stop the world") before beginning reachability analysis. Stopping a
747 multithreaded mutator is a complicated problem. It generally requires highly
748 platform-specific code in the runtime, and the production of carefully
749 designed machine code at safe points.
752 In this technique, the mutator and the collector run concurrently, with the
753 goal of eliminating pause times. In a *cooperative* collector, the mutator
754 further aids with collection should a pause occur, allowing collection to take
755 advantage of multiprocessor hosts. The "stop the world" problem of threaded
756 collectors is generally still present to a limited extent. Sophisticated
757 marking algorithms are necessary. Read barriers may be necessary.
759 As the matrix indicates, LLVM's garbage collection infrastructure is already
760 suitable for a wide variety of collectors, but does not currently extend to
761 multithreaded programs. This will be added in the future as there is
769 LLVM automatically computes a stack map. One of the most important features
770 of a ``GCStrategy`` is to compile this information into the executable in
771 the binary representation expected by the runtime library.
773 The stack map consists of the location and identity of each GC root in the
774 each function in the module. For each root:
776 * ``RootNum``: The index of the root.
778 * ``StackOffset``: The offset of the object relative to the frame pointer.
780 * ``RootMetadata``: The value passed as the ``%metadata`` parameter to the
781 ``@llvm.gcroot`` intrinsic.
783 Also, for the function as a whole:
785 * ``getFrameSize()``: The overall size of the function's initial stack frame,
786 not accounting for any dynamic allocation.
788 * ``roots_size()``: The count of roots in the function.
790 To access the stack map, use ``GCFunctionMetadata::roots_begin()`` and
791 -``end()`` from the :ref:`GCMetadataPrinter <assembly>`:
795 for (iterator I = begin(), E = end(); I != E; ++I) {
796 GCFunctionInfo *FI = *I;
797 unsigned FrameSize = FI->getFrameSize();
798 size_t RootCount = FI->roots_size();
800 for (GCFunctionInfo::roots_iterator RI = FI->roots_begin(),
801 RE = FI->roots_end();
803 int RootNum = RI->Num;
804 int RootStackOffset = RI->StackOffset;
805 Constant *RootMetadata = RI->Metadata;
809 If the ``llvm.gcroot`` intrinsic is eliminated before code generation by a
810 custom lowering pass, LLVM will compute an empty stack map. This may be useful
811 for collector plugins which implement reference counting or a shadow stack.
815 Initializing roots to null: ``InitRoots``
816 -----------------------------------------
824 When set, LLVM will automatically initialize each root to ``null`` upon entry to
825 the function. This prevents the GC's sweep phase from visiting uninitialized
826 pointers, which will almost certainly cause it to crash. This initialization
827 occurs before custom lowering, so the two may be used together.
829 Since LLVM does not yet compute liveness information, there is no means of
830 distinguishing an uninitialized stack root from an initialized one. Therefore,
831 this feature should be used by all GC plugins. It is enabled by default.
833 Custom lowering of intrinsics: ``CustomRoots``, ``CustomReadBarriers``, and ``CustomWriteBarriers``
834 ---------------------------------------------------------------------------------------------------
836 For GCs which use barriers or unusual treatment of stack roots, these
837 flags allow the collector to perform arbitrary transformations of the
842 class MyGC : public GCStrategy {
846 CustomReadBarriers = true;
847 CustomWriteBarriers = true;
851 If any of these flags are set, LLVM suppresses its default lowering for
852 the corresponding intrinsics. Instead, you must provide a custom Pass
853 which lowers the intrinsics as desired. If you have opted in to custom
854 lowering of a particular intrinsic your pass **must** eliminate all
855 instances of the corresponding intrinsic in functions which opt in to
856 your GC. The best example of such a pass is the ShadowStackGC and it's
857 ShadowStackGCLowering pass.
859 There is currently no way to register such a custom lowering pass
860 without building a custom copy of LLVM.
864 Generating safe points: ``NeededSafePoints``
865 --------------------------------------------
867 LLVM can compute four kinds of safe points:
872 /// PointKind - The type of a collector-safe point.
875 Loop, //< Instr is a loop (backwards branch).
876 Return, //< Instr is a return instruction.
877 PreCall, //< Instr is a call instruction.
878 PostCall //< Instr is the return address of a call.
882 A collector can request any combination of the four by setting the
883 ``NeededSafePoints`` mask:
888 NeededSafePoints = 1 << GC::Loop
894 It can then use the following routines to access safe points.
898 for (iterator I = begin(), E = end(); I != E; ++I) {
899 GCFunctionInfo *MD = *I;
900 size_t PointCount = MD->size();
902 for (GCFunctionInfo::iterator PI = MD->begin(),
903 PE = MD->end(); PI != PE; ++PI) {
904 GC::PointKind PointKind = PI->Kind;
905 unsigned PointNum = PI->Num;
909 Almost every collector requires ``PostCall`` safe points, since these correspond
910 to the moments when the function is suspended during a call to a subroutine.
912 Threaded programs generally require ``Loop`` safe points to guarantee that the
913 application will reach a safe point within a bounded amount of time, even if it
914 is executing a long-running loop which contains no function calls.
916 Threaded collectors may also require ``Return`` and ``PreCall`` safe points to
917 implement "stop the world" techniques using self-modifying code, where it is
918 important that the program not exit the function without reaching a safe point
919 (because only the topmost function has been patched).
923 Emitting assembly code: ``GCMetadataPrinter``
924 ---------------------------------------------
926 LLVM allows a plugin to print arbitrary assembly code before and after the rest
927 of a module's assembly code. At the end of the module, the GC can compile the
928 LLVM stack map into assembly code. (At the beginning, this information is not
931 Since AsmWriter and CodeGen are separate components of LLVM, a separate abstract
932 base class and registry is provided for printing assembly code, the
933 ``GCMetadaPrinter`` and ``GCMetadataPrinterRegistry``. The AsmWriter will look
934 for such a subclass if the ``GCStrategy`` sets ``UsesMetadata``:
942 This separation allows JIT-only clients to be smaller.
944 Note that LLVM does not currently have analogous APIs to support code generation
945 in the JIT, nor using the object writers.
949 // lib/MyGC/MyGCPrinter.cpp - Example LLVM GC printer
951 #include "llvm/CodeGen/GCMetadataPrinter.h"
952 #include "llvm/Support/Compiler.h"
954 using namespace llvm;
957 class LLVM_LIBRARY_VISIBILITY MyGCPrinter : public GCMetadataPrinter {
959 virtual void beginAssembly(AsmPrinter &AP);
961 virtual void finishAssembly(AsmPrinter &AP);
964 GCMetadataPrinterRegistry::Add<MyGCPrinter>
965 X("mygc", "My bespoke garbage collector.");
968 The collector should use ``AsmPrinter`` to print portable assembly code. The
969 collector itself contains the stack map for the entire module, and may access
970 the ``GCFunctionInfo`` using its own ``begin()`` and ``end()`` methods. Here's
975 #include "llvm/CodeGen/AsmPrinter.h"
976 #include "llvm/IR/Function.h"
977 #include "llvm/IR/DataLayout.h"
978 #include "llvm/Target/TargetAsmInfo.h"
979 #include "llvm/Target/TargetMachine.h"
981 void MyGCPrinter::beginAssembly(AsmPrinter &AP) {
985 void MyGCPrinter::finishAssembly(AsmPrinter &AP) {
986 MCStreamer &OS = AP.OutStreamer;
987 unsigned IntPtrSize = AP.TM.getSubtargetImpl()->getDataLayout()->getPointerSize();
989 // Put this in the data section.
990 OS.SwitchSection(AP.getObjFileLowering().getDataSection());
992 // For each function...
993 for (iterator FI = begin(), FE = end(); FI != FE; ++FI) {
994 GCFunctionInfo &MD = **FI;
996 // A compact GC layout. Emit this data structure:
999 // int32_t PointCount;
1000 // void *SafePointAddress[PointCount];
1001 // int32_t StackFrameSize; // in words
1002 // int32_t StackArity;
1003 // int32_t LiveCount;
1004 // int32_t LiveOffsets[LiveCount];
1005 // } __gcmap_<FUNCTIONNAME>;
1007 // Align to address width.
1008 AP.EmitAlignment(IntPtrSize == 4 ? 2 : 3);
1011 OS.AddComment("safe point count");
1012 AP.EmitInt32(MD.size());
1014 // And each safe point...
1015 for (GCFunctionInfo::iterator PI = MD.begin(),
1016 PE = MD.end(); PI != PE; ++PI) {
1017 // Emit the address of the safe point.
1018 OS.AddComment("safe point address");
1019 MCSymbol *Label = PI->Label;
1020 AP.EmitLabelPlusOffset(Label/*Hi*/, 0/*Offset*/, 4/*Size*/);
1023 // Stack information never change in safe points! Only print info from the
1025 GCFunctionInfo::iterator PI = MD.begin();
1027 // Emit the stack frame size.
1028 OS.AddComment("stack frame size (in words)");
1029 AP.EmitInt32(MD.getFrameSize() / IntPtrSize);
1031 // Emit stack arity, i.e. the number of stacked arguments.
1032 unsigned RegisteredArgs = IntPtrSize == 4 ? 5 : 6;
1033 unsigned StackArity = MD.getFunction().arg_size() > RegisteredArgs ?
1034 MD.getFunction().arg_size() - RegisteredArgs : 0;
1035 OS.AddComment("stack arity");
1036 AP.EmitInt32(StackArity);
1038 // Emit the number of live roots in the function.
1039 OS.AddComment("live root count");
1040 AP.EmitInt32(MD.live_size(PI));
1042 // And for each live root...
1043 for (GCFunctionInfo::live_iterator LI = MD.live_begin(PI),
1044 LE = MD.live_end(PI);
1046 // Emit live root's offset within the stack frame.
1047 OS.AddComment("stack index (offset / wordsize)");
1048 AP.EmitInt32(LI->StackOffset);
1058 [Appel89] Runtime Tags Aren't Necessary. Andrew W. Appel. Lisp and Symbolic
1059 Computation 19(7):703-705, July 1989.
1063 [Goldberg91] Tag-free garbage collection for strongly typed programming
1064 languages. Benjamin Goldberg. ACM SIGPLAN PLDI'91.
1068 [Tolmach94] Tag-free garbage collection using explicit type parameters. Andrew
1069 Tolmach. Proceedings of the 1994 ACM conference on LISP and functional
1074 [Henderson2002] `Accurate Garbage Collection in an Uncooperative Environment
1075 <http://citeseer.ist.psu.edu/henderson02accurate.html>`__