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: ``llvm.gcroot``
226 --------------------------------------------------
230 void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
232 The ``llvm.gcroot`` intrinsic is used to inform LLVM that a stack variable
233 references an object on the heap and is to be tracked for garbage collection.
234 The exact impact on generated code is specified by a :ref:`compiler plugin
235 <plugin>`. All calls to ``llvm.gcroot`` **must** reside inside the first basic
238 A compiler which uses mem2reg to raise imperative code using ``alloca`` into SSA
239 form need only add a call to ``@llvm.gcroot`` for those variables which a
240 pointers into the GC heap.
242 It is also important to mark intermediate values with ``llvm.gcroot``. For
243 example, consider ``h(f(), g())``. Beware leaking the result of ``f()`` in the
244 case that ``g()`` triggers a collection. Note, that stack variables must be
245 initialized and marked with ``llvm.gcroot`` in function's prologue.
247 The first argument **must** be a value referring to an alloca instruction or a
248 bitcast of an alloca. The second contains a pointer to metadata that should be
249 associated with the pointer, and **must** be a constant or global value
250 address. If your target collector uses tags, use a null pointer for metadata.
252 The ``%metadata`` argument can be used to avoid requiring heap objects to have
253 'isa' pointers or tag bits. [Appel89_, Goldberg91_, Tolmach94_] If specified,
254 its value will be tracked along with the location of the pointer in the stack
257 Consider the following fragment of Java code:
262 Object X; // A null-initialized reference to an object
266 This block (which may be located in the middle of a function or in a loop nest),
267 could be compiled to this LLVM code:
272 ;; In the entry block for the function, allocate the
273 ;; stack space for X, which is an LLVM pointer.
276 ;; Tell LLVM that the stack space is a stack root.
277 ;; Java has type-tags on objects, so we pass null as metadata.
278 %tmp = bitcast %Object** %X to i8**
279 call void @llvm.gcroot(i8** %tmp, i8* null)
282 ;; "CodeBlock" is the block corresponding to the start
283 ;; of the scope above.
285 ;; Java null-initializes pointers.
286 store %Object* null, %Object** %X
290 ;; As the pointer goes out of scope, store a null value into
291 ;; it, to indicate that the value is no longer live.
292 store %Object* null, %Object** %X
295 Reading and writing references in the heap
296 ------------------------------------------
298 Some collectors need to be informed when the mutator (the program that needs
299 garbage collection) either reads a pointer from or writes a pointer to a field
300 of a heap object. The code fragments inserted at these points are called *read
301 barriers* and *write barriers*, respectively. The amount of code that needs to
302 be executed is usually quite small and not on the critical path of any
303 computation, so the overall performance impact of the barrier is tolerable.
305 Barriers often require access to the *object pointer* rather than the *derived
306 pointer* (which is a pointer to the field within the object). Accordingly,
307 these intrinsics take both pointers as separate arguments for completeness. In
308 this snippet, ``%object`` is the object pointer, and ``%derived`` is the derived
314 %class.Array = type { %class.Object, i32, [0 x %class.Object*] }
317 ;; Load the object pointer from a gcroot.
318 %object = load %class.Array** %object_addr
320 ;; Compute the derived pointer.
321 %derived = getelementptr %object, i32 0, i32 2, i32 %n
323 LLVM does not enforce this relationship between the object and derived pointer
324 (although a particular :ref:`collector strategy <plugin>` might). However, it
325 would be an unusual collector that violated it.
327 The use of these intrinsics is naturally optional if the target GC does not
328 require the corresponding barrier. The GC strategy used with such a collector
329 should replace the intrinsic calls with the corresponding ``load`` or
330 ``store`` instruction if they are used.
332 One known deficiency with the current design is that the barrier intrinsics do
333 not include the size or alignment of the underlying operation performed. It is
334 currently assumed that the operation is of pointer size and the alignment is
335 assumed to be the target machine's default alignment.
337 Write barrier: ``llvm.gcwrite``
338 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
342 void @llvm.gcwrite(i8* %value, i8* %object, i8** %derived)
344 For write barriers, LLVM provides the ``llvm.gcwrite`` intrinsic function. It
345 has exactly the same semantics as a non-volatile ``store`` to the derived
346 pointer (the third argument). The exact code generated is specified by the
347 Function's selected :ref:`GC strategy <plugin>`.
349 Many important algorithms require write barriers, including generational and
350 concurrent collectors. Additionally, write barriers could be used to implement
353 Read barrier: ``llvm.gcread``
354 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
358 i8* @llvm.gcread(i8* %object, i8** %derived)
360 For read barriers, LLVM provides the ``llvm.gcread`` intrinsic function. It has
361 exactly the same semantics as a non-volatile ``load`` from the derived pointer
362 (the second argument). The exact code generated is specified by the Function's
363 selected :ref:`GC strategy <plugin>`.
365 Read barriers are needed by fewer algorithms than write barriers, and may have a
366 greater performance impact since pointer reads are more frequent than writes.
373 LLVM includes built in support for several varieties of garbage collectors.
376 ----------------------
378 To use this collector strategy, mark your functions with:
382 F.setGC("shadow-stack");
384 Unlike many GC algorithms which rely on a cooperative code generator to compile
385 stack maps, this algorithm carefully maintains a linked list of stack roots
386 [:ref:`Henderson2002 <henderson02>`]. This so-called "shadow stack" mirrors the
387 machine stack. Maintaining this data structure is slower than using a stack map
388 compiled into the executable as constant data, but has a significant portability
389 advantage because it requires no special support from the target code generator,
390 and does not require tricky platform-specific code to crawl the machine stack.
392 The tradeoff for this simplicity and portability is:
394 * High overhead per function call.
398 Still, it's an easy way to get started. After your compiler and runtime are up
399 and running, writing a :ref:`plugin <plugin>` will allow you to take advantage
400 of :ref:`more advanced GC features <collector-algos>` of LLVM in order to
404 The shadow stack doesn't imply a memory allocation algorithm. A semispace
405 collector or building atop ``malloc`` are great places to start, and can be
406 implemented with very little code.
408 When it comes time to collect, however, your runtime needs to traverse the stack
409 roots, and for this it needs to integrate with the shadow stack. Luckily, doing
410 so is very simple. (This code is heavily commented to help you understand the
411 data structure, but there are only 20 lines of meaningful code.)
415 /// @brief The map for a single function's stack frame. One of these is
416 /// compiled as constant data into the executable for each function.
418 /// Storage of metadata values is elided if the %metadata parameter to
419 /// @llvm.gcroot is null.
421 int32_t NumRoots; //< Number of roots in stack frame.
422 int32_t NumMeta; //< Number of metadata entries. May be < NumRoots.
423 const void *Meta[0]; //< Metadata for each root.
426 /// @brief A link in the dynamic shadow stack. One of these is embedded in
427 /// the stack frame of each function on the call stack.
429 StackEntry *Next; //< Link to next stack entry (the caller's).
430 const FrameMap *Map; //< Pointer to constant FrameMap.
431 void *Roots[0]; //< Stack roots (in-place array).
434 /// @brief The head of the singly-linked list of StackEntries. Functions push
435 /// and pop onto this in their prologue and epilogue.
437 /// Since there is only a global list, this technique is not threadsafe.
438 StackEntry *llvm_gc_root_chain;
440 /// @brief Calls Visitor(root, meta) for each GC root on the stack.
441 /// root and meta are exactly the values passed to
444 /// Visitor could be a function to recursively mark live objects. Or it
445 /// might copy them to another heap or generation.
447 /// @param Visitor A function to invoke for every GC root on the stack.
448 void visitGCRoots(void (*Visitor)(void **Root, const void *Meta)) {
449 for (StackEntry *R = llvm_gc_root_chain; R; R = R->Next) {
452 // For roots [0, NumMeta), the metadata pointer is in the FrameMap.
453 for (unsigned e = R->Map->NumMeta; i != e; ++i)
454 Visitor(&R->Roots[i], R->Map->Meta[i]);
456 // For roots [NumMeta, NumRoots), the metadata pointer is null.
457 for (unsigned e = R->Map->NumRoots; i != e; ++i)
458 Visitor(&R->Roots[i], NULL);
463 The 'Erlang' and 'Ocaml' GCs
464 -----------------------------
466 LLVM ships with two example collectors which leverage the ''gcroot''
467 mechanisms. To our knowledge, these are not actually used by any language
468 runtime, but they do provide a reasonable starting point for someone interested
469 in writing an ''gcroot' compatible GC plugin. In particular, these are the
470 only in tree examples of how to produce a custom binary stack map format using
471 a ''gcroot'' strategy.
473 As there names imply, the binary format produced is intended to model that
474 used by the Erlang and OCaml compilers respectively.
477 The Statepoint Example GC
478 -------------------------
482 F.setGC("statepoint-example");
484 This GC provides an example of how one might use the infrastructure provided
485 by ''gc.statepoint''.
491 If none of the built in GC strategy descriptions met your needs above, you will
492 need to define a custom GCStrategy and possibly, a custom LLVM pass to perform
493 lowering. Your best example of where to start defining a custom GCStrategy
494 would be to look at one of the built in strategies.
496 You may be able to structure this additional code as a loadable plugin library.
497 Loadable plugins are sufficient if all you need is to enable a different
498 combination of built in functionality, but if you need to provide a custom
499 lowering pass, you will need to build a patched version of LLVM. If you think
500 you need a patched build, please ask for advice on llvm-dev. There may be an
501 easy way we can extend the support to make it work for your use case without
502 requiring a custom build.
504 Collector Requirements
505 ----------------------
507 You should be able to leverage any existing collector library that includes the following elements:
509 #. A memory allocator which exposes an allocation function your compiled
512 #. A binary format for the stack map. A stack map describes the location
513 of references at a safepoint and is used by precise collectors to identify
514 references within a stack frame on the machine stack. Note that collectors
515 which conservatively scan the stack don't require such a structure.
517 #. A stack crawler to discover functions on the call stack, and enumerate the
518 references listed in the stack map for each call site.
520 #. A mechanism for identifying references in global locations (e.g. global
523 #. If you collector requires them, an LLVM IR implementation of your collectors
524 load and store barriers. Note that since many collectors don't require
525 barriers at all, LLVM defaults to lowering such barriers to normal loads
526 and stores unless you arrange otherwise.
529 Implementing a collector plugin
530 -------------------------------
532 User code specifies which GC code generation to use with the ``gc`` function
533 attribute or, equivalently, with the ``setGC`` method of ``Function``.
535 To implement a GC plugin, it is necessary to subclass ``llvm::GCStrategy``,
536 which can be accomplished in a few lines of boilerplate code. LLVM's
537 infrastructure provides access to several important algorithms. For an
538 uncontroversial collector, all that remains may be to compile LLVM's computed
539 stack map to assembly code (using the binary representation expected by the
540 runtime library). This can be accomplished in about 100 lines of code.
542 This is not the appropriate place to implement a garbage collected heap or a
543 garbage collector itself. That code should exist in the language's runtime
544 library. The compiler plugin is responsible for generating code which conforms
545 to the binary interface defined by library, most essentially the :ref:`stack map
548 To subclass ``llvm::GCStrategy`` and register it with the compiler:
552 // lib/MyGC/MyGC.cpp - Example LLVM GC plugin
554 #include "llvm/CodeGen/GCStrategy.h"
555 #include "llvm/CodeGen/GCMetadata.h"
556 #include "llvm/Support/Compiler.h"
558 using namespace llvm;
561 class LLVM_LIBRARY_VISIBILITY MyGC : public GCStrategy {
566 GCRegistry::Add<MyGC>
567 X("mygc", "My bespoke garbage collector.");
570 This boilerplate collector does nothing. More specifically:
572 * ``llvm.gcread`` calls are replaced with the corresponding ``load``
575 * ``llvm.gcwrite`` calls are replaced with the corresponding ``store``
578 * No safe points are added to the code.
580 * The stack map is not compiled into the executable.
582 Using the LLVM makefiles, this code
583 can be compiled as a plugin using a simple makefile:
593 include $(LEVEL)/Makefile.common
595 Once the plugin is compiled, code using it may be compiled using ``llc
596 -load=MyGC.so`` (though MyGC.so may have some other platform-specific
602 define void @f() gc "mygc" {
606 $ llvm-as < sample.ll | llc -load=MyGC.so
608 It is also possible to statically link the collector plugin into tools, such as
609 a language-specific compiler front-end.
613 Overview of available features
614 ------------------------------
616 ``GCStrategy`` provides a range of features through which a plugin may do useful
617 work. Some of these are callbacks, some are algorithms that can be enabled,
618 disabled, or customized. This matrix summarizes the supported (and planned)
619 features and correlates them with the collection techniques which typically
622 .. |v| unicode:: 0x2714
625 .. |x| unicode:: 0x2718
628 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
629 | Algorithm | Done | Shadow | refcount | mark- | copying | incremental | threaded | concurrent |
630 | | | stack | | sweep | | | | |
631 +============+======+========+==========+=======+=========+=============+==========+============+
632 | stack map | |v| | | | |x| | |x| | |x| | |x| | |x| |
633 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
634 | initialize | |v| | |x| | |x| | |x| | |x| | |x| | |x| | |x| |
635 | roots | | | | | | | | |
636 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
637 | derived | NO | | | | | | **N**\* | **N**\* |
638 | pointers | | | | | | | | |
639 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
640 | **custom | |v| | | | | | | | |
641 | lowering** | | | | | | | | |
642 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
643 | *gcroot* | |v| | |x| | |x| | | | | | |
644 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
645 | *gcwrite* | |v| | | |x| | | | |x| | | |x| |
646 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
647 | *gcread* | |v| | | | | | | | |x| |
648 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
649 | **safe | | | | | | | | |
650 | points** | | | | | | | | |
651 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
652 | *in | |v| | | | |x| | |x| | |x| | |x| | |x| |
653 | calls* | | | | | | | | |
654 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
655 | *before | |v| | | | | | | |x| | |x| |
656 | calls* | | | | | | | | |
657 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
658 | *for | NO | | | | | | **N** | **N** |
659 | loops* | | | | | | | | |
660 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
661 | *before | |v| | | | | | | |x| | |x| |
662 | escape* | | | | | | | | |
663 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
664 | emit code | NO | | | | | | **N** | **N** |
665 | at safe | | | | | | | | |
666 | points | | | | | | | | |
667 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
668 | **output** | | | | | | | | |
669 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
670 | *assembly* | |v| | | | |x| | |x| | |x| | |x| | |x| |
671 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
672 | *JIT* | NO | | | **?** | **?** | **?** | **?** | **?** |
673 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
674 | *obj* | NO | | | **?** | **?** | **?** | **?** | **?** |
675 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
676 | live | NO | | | **?** | **?** | **?** | **?** | **?** |
677 | analysis | | | | | | | | |
678 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
679 | register | NO | | | **?** | **?** | **?** | **?** | **?** |
680 | map | | | | | | | | |
681 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
682 | \* Derived pointers only pose a hasard to copying collections. |
683 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
684 | **?** denotes a feature which could be utilized if available. |
685 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
687 To be clear, the collection techniques above are defined as:
690 The mutator carefully maintains a linked list of stack roots.
693 The mutator maintains a reference count for each object and frees an object
694 when its count falls to zero.
697 When the heap is exhausted, the collector marks reachable objects starting
698 from the roots, then deallocates unreachable objects in a sweep phase.
701 As reachability analysis proceeds, the collector copies objects from one heap
702 area to another, compacting them in the process. Copying collectors enable
703 highly efficient "bump pointer" allocation and can improve locality of
707 (Including generational collectors.) Incremental collectors generally have all
708 the properties of a copying collector (regardless of whether the mature heap
709 is compacting), but bring the added complexity of requiring write barriers.
712 Denotes a multithreaded mutator; the collector must still stop the mutator
713 ("stop the world") before beginning reachability analysis. Stopping a
714 multithreaded mutator is a complicated problem. It generally requires highly
715 platform-specific code in the runtime, and the production of carefully
716 designed machine code at safe points.
719 In this technique, the mutator and the collector run concurrently, with the
720 goal of eliminating pause times. In a *cooperative* collector, the mutator
721 further aids with collection should a pause occur, allowing collection to take
722 advantage of multiprocessor hosts. The "stop the world" problem of threaded
723 collectors is generally still present to a limited extent. Sophisticated
724 marking algorithms are necessary. Read barriers may be necessary.
726 As the matrix indicates, LLVM's garbage collection infrastructure is already
727 suitable for a wide variety of collectors, but does not currently extend to
728 multithreaded programs. This will be added in the future as there is
736 LLVM automatically computes a stack map. One of the most important features
737 of a ``GCStrategy`` is to compile this information into the executable in
738 the binary representation expected by the runtime library.
740 The stack map consists of the location and identity of each GC root in the
741 each function in the module. For each root:
743 * ``RootNum``: The index of the root.
745 * ``StackOffset``: The offset of the object relative to the frame pointer.
747 * ``RootMetadata``: The value passed as the ``%metadata`` parameter to the
748 ``@llvm.gcroot`` intrinsic.
750 Also, for the function as a whole:
752 * ``getFrameSize()``: The overall size of the function's initial stack frame,
753 not accounting for any dynamic allocation.
755 * ``roots_size()``: The count of roots in the function.
757 To access the stack map, use ``GCFunctionMetadata::roots_begin()`` and
758 -``end()`` from the :ref:`GCMetadataPrinter <assembly>`:
762 for (iterator I = begin(), E = end(); I != E; ++I) {
763 GCFunctionInfo *FI = *I;
764 unsigned FrameSize = FI->getFrameSize();
765 size_t RootCount = FI->roots_size();
767 for (GCFunctionInfo::roots_iterator RI = FI->roots_begin(),
768 RE = FI->roots_end();
770 int RootNum = RI->Num;
771 int RootStackOffset = RI->StackOffset;
772 Constant *RootMetadata = RI->Metadata;
776 If the ``llvm.gcroot`` intrinsic is eliminated before code generation by a
777 custom lowering pass, LLVM will compute an empty stack map. This may be useful
778 for collector plugins which implement reference counting or a shadow stack.
782 Initializing roots to null: ``InitRoots``
783 -----------------------------------------
791 When set, LLVM will automatically initialize each root to ``null`` upon entry to
792 the function. This prevents the GC's sweep phase from visiting uninitialized
793 pointers, which will almost certainly cause it to crash. This initialization
794 occurs before custom lowering, so the two may be used together.
796 Since LLVM does not yet compute liveness information, there is no means of
797 distinguishing an uninitialized stack root from an initialized one. Therefore,
798 this feature should be used by all GC plugins. It is enabled by default.
800 Custom lowering of intrinsics: ``CustomRoots``, ``CustomReadBarriers``, and ``CustomWriteBarriers``
801 ---------------------------------------------------------------------------------------------------
803 For GCs which use barriers or unusual treatment of stack roots, these
804 flags allow the collector to perform arbitrary transformations of the
809 class MyGC : public GCStrategy {
813 CustomReadBarriers = true;
814 CustomWriteBarriers = true;
818 If any of these flags are set, LLVM suppresses its default lowering for
819 the corresponding intrinsics. Instead, you must provide a custom Pass
820 which lowers the intrinsics as desired. If you have opted in to custom
821 lowering of a particular intrinsic your pass **must** eliminate all
822 instances of the corresponding intrinsic in functions which opt in to
823 your GC. The best example of such a pass is the ShadowStackGC and it's
824 ShadowStackGCLowering pass.
826 There is currently no way to register such a custom lowering pass
827 without building a custom copy of LLVM.
831 Generating safe points: ``NeededSafePoints``
832 --------------------------------------------
834 LLVM can compute four kinds of safe points:
839 /// PointKind - The type of a collector-safe point.
842 Loop, //< Instr is a loop (backwards branch).
843 Return, //< Instr is a return instruction.
844 PreCall, //< Instr is a call instruction.
845 PostCall //< Instr is the return address of a call.
849 A collector can request any combination of the four by setting the
850 ``NeededSafePoints`` mask:
855 NeededSafePoints = 1 << GC::Loop
861 It can then use the following routines to access safe points.
865 for (iterator I = begin(), E = end(); I != E; ++I) {
866 GCFunctionInfo *MD = *I;
867 size_t PointCount = MD->size();
869 for (GCFunctionInfo::iterator PI = MD->begin(),
870 PE = MD->end(); PI != PE; ++PI) {
871 GC::PointKind PointKind = PI->Kind;
872 unsigned PointNum = PI->Num;
876 Almost every collector requires ``PostCall`` safe points, since these correspond
877 to the moments when the function is suspended during a call to a subroutine.
879 Threaded programs generally require ``Loop`` safe points to guarantee that the
880 application will reach a safe point within a bounded amount of time, even if it
881 is executing a long-running loop which contains no function calls.
883 Threaded collectors may also require ``Return`` and ``PreCall`` safe points to
884 implement "stop the world" techniques using self-modifying code, where it is
885 important that the program not exit the function without reaching a safe point
886 (because only the topmost function has been patched).
890 Emitting assembly code: ``GCMetadataPrinter``
891 ---------------------------------------------
893 LLVM allows a plugin to print arbitrary assembly code before and after the rest
894 of a module's assembly code. At the end of the module, the GC can compile the
895 LLVM stack map into assembly code. (At the beginning, this information is not
898 Since AsmWriter and CodeGen are separate components of LLVM, a separate abstract
899 base class and registry is provided for printing assembly code, the
900 ``GCMetadaPrinter`` and ``GCMetadataPrinterRegistry``. The AsmWriter will look
901 for such a subclass if the ``GCStrategy`` sets ``UsesMetadata``:
909 This separation allows JIT-only clients to be smaller.
911 Note that LLVM does not currently have analogous APIs to support code generation
912 in the JIT, nor using the object writers.
916 // lib/MyGC/MyGCPrinter.cpp - Example LLVM GC printer
918 #include "llvm/CodeGen/GCMetadataPrinter.h"
919 #include "llvm/Support/Compiler.h"
921 using namespace llvm;
924 class LLVM_LIBRARY_VISIBILITY MyGCPrinter : public GCMetadataPrinter {
926 virtual void beginAssembly(AsmPrinter &AP);
928 virtual void finishAssembly(AsmPrinter &AP);
931 GCMetadataPrinterRegistry::Add<MyGCPrinter>
932 X("mygc", "My bespoke garbage collector.");
935 The collector should use ``AsmPrinter`` to print portable assembly code. The
936 collector itself contains the stack map for the entire module, and may access
937 the ``GCFunctionInfo`` using its own ``begin()`` and ``end()`` methods. Here's
942 #include "llvm/CodeGen/AsmPrinter.h"
943 #include "llvm/IR/Function.h"
944 #include "llvm/IR/DataLayout.h"
945 #include "llvm/Target/TargetAsmInfo.h"
946 #include "llvm/Target/TargetMachine.h"
948 void MyGCPrinter::beginAssembly(AsmPrinter &AP) {
952 void MyGCPrinter::finishAssembly(AsmPrinter &AP) {
953 MCStreamer &OS = AP.OutStreamer;
954 unsigned IntPtrSize = AP.TM.getSubtargetImpl()->getDataLayout()->getPointerSize();
956 // Put this in the data section.
957 OS.SwitchSection(AP.getObjFileLowering().getDataSection());
959 // For each function...
960 for (iterator FI = begin(), FE = end(); FI != FE; ++FI) {
961 GCFunctionInfo &MD = **FI;
963 // A compact GC layout. Emit this data structure:
966 // int32_t PointCount;
967 // void *SafePointAddress[PointCount];
968 // int32_t StackFrameSize; // in words
969 // int32_t StackArity;
970 // int32_t LiveCount;
971 // int32_t LiveOffsets[LiveCount];
972 // } __gcmap_<FUNCTIONNAME>;
974 // Align to address width.
975 AP.EmitAlignment(IntPtrSize == 4 ? 2 : 3);
978 OS.AddComment("safe point count");
979 AP.EmitInt32(MD.size());
981 // And each safe point...
982 for (GCFunctionInfo::iterator PI = MD.begin(),
983 PE = MD.end(); PI != PE; ++PI) {
984 // Emit the address of the safe point.
985 OS.AddComment("safe point address");
986 MCSymbol *Label = PI->Label;
987 AP.EmitLabelPlusOffset(Label/*Hi*/, 0/*Offset*/, 4/*Size*/);
990 // Stack information never change in safe points! Only print info from the
992 GCFunctionInfo::iterator PI = MD.begin();
994 // Emit the stack frame size.
995 OS.AddComment("stack frame size (in words)");
996 AP.EmitInt32(MD.getFrameSize() / IntPtrSize);
998 // Emit stack arity, i.e. the number of stacked arguments.
999 unsigned RegisteredArgs = IntPtrSize == 4 ? 5 : 6;
1000 unsigned StackArity = MD.getFunction().arg_size() > RegisteredArgs ?
1001 MD.getFunction().arg_size() - RegisteredArgs : 0;
1002 OS.AddComment("stack arity");
1003 AP.EmitInt32(StackArity);
1005 // Emit the number of live roots in the function.
1006 OS.AddComment("live root count");
1007 AP.EmitInt32(MD.live_size(PI));
1009 // And for each live root...
1010 for (GCFunctionInfo::live_iterator LI = MD.live_begin(PI),
1011 LE = MD.live_end(PI);
1013 // Emit live root's offset within the stack frame.
1014 OS.AddComment("stack index (offset / wordsize)");
1015 AP.EmitInt32(LI->StackOffset);
1025 [Appel89] Runtime Tags Aren't Necessary. Andrew W. Appel. Lisp and Symbolic
1026 Computation 19(7):703-705, July 1989.
1030 [Goldberg91] Tag-free garbage collection for strongly typed programming
1031 languages. Benjamin Goldberg. ACM SIGPLAN PLDI'91.
1035 [Tolmach94] Tag-free garbage collection using explicit type parameters. Andrew
1036 Tolmach. Proceedings of the 1994 ACM conference on LISP and functional
1041 [Henderson2002] `Accurate Garbage Collection in an Uncooperative Environment
1042 <http://citeseer.ist.psu.edu/henderson02accurate.html>`__