1 =====================================
2 Accurate Garbage Collection with LLVM
3 =====================================
11 Garbage collection is a widely used technique that frees the programmer from
12 having to know the lifetimes of heap objects, making software easier to produce
13 and maintain. Many programming languages rely on garbage collection for
14 automatic memory management. There are two primary forms of garbage collection:
15 conservative and accurate.
17 Conservative garbage collection often does not require any special support from
18 either the language or the compiler: it can handle non-type-safe programming
19 languages (such as C/C++) and does not require any special information from the
20 compiler. The `Boehm collector
21 <http://www.hpl.hp.com/personal/Hans_Boehm/gc/>`__ is an example of a
22 state-of-the-art conservative collector.
24 Accurate garbage collection requires the ability to identify all pointers in the
25 program at run-time (which requires that the source-language be type-safe in
26 most cases). Identifying pointers at run-time requires compiler support to
27 locate all places that hold live pointer variables at run-time, including the
28 :ref:`processor stack and registers <gcroot>`.
30 Conservative garbage collection is attractive because it does not require any
31 special compiler support, but it does have problems. In particular, because the
32 conservative garbage collector cannot *know* that a particular word in the
33 machine is a pointer, it cannot move live objects in the heap (preventing the
34 use of compacting and generational GC algorithms) and it can occasionally suffer
35 from memory leaks due to integer values that happen to point to objects in the
36 program. In addition, some aggressive compiler transformations can break
37 conservative garbage collectors (though these seem rare in practice).
39 Accurate garbage collectors do not suffer from any of these problems, but they
40 can suffer from degraded scalar optimization of the program. In particular,
41 because the runtime must be able to identify and update all pointers active in
42 the program, some optimizations are less effective. In practice, however, the
43 locality and performance benefits of using aggressive garbage collection
44 techniques dominates any low-level losses.
46 This document describes the mechanisms and interfaces provided by LLVM to
47 support accurate garbage collection.
52 LLVM's intermediate representation provides :ref:`garbage collection intrinsics
53 <gc_intrinsics>` that offer support for a broad class of collector models. For
54 instance, the intrinsics permit:
56 * semi-space collectors
58 * mark-sweep collectors
60 * generational collectors
64 * incremental collectors
66 * concurrent collectors
68 * cooperative collectors
70 We hope that the primitive support built into the LLVM IR is sufficient to
71 support a broad class of garbage collected languages including Scheme, ML, Java,
72 C#, Perl, Python, Lua, Ruby, other scripting languages, and more.
74 However, LLVM does not itself provide a garbage collector --- this should be
75 part of your language's runtime library. LLVM provides a framework for compile
76 time :ref:`code generation plugins <plugin>`. The role of these plugins is to
77 generate code and data structures which conforms to the *binary interface*
78 specified by the *runtime library*. This is similar to the relationship between
79 LLVM and DWARF debugging info, for example. The difference primarily lies in
80 the lack of an established standard in the domain of garbage collection --- thus
83 The aspects of the binary interface with which LLVM's GC support is
86 * Creation of GC-safe points within code where collection is allowed to execute
89 * Computation of the stack map. For each safe point in the code, object
90 references within the stack frame must be identified so that the collector may
91 traverse and perhaps update them.
93 * Write barriers when storing object references to the heap. These are commonly
94 used to optimize incremental scans in generational collectors.
96 * Emission of read barriers when loading object references. These are useful
97 for interoperating with concurrent collectors.
99 There are additional areas that LLVM does not directly address:
101 * Registration of global roots with the runtime.
103 * Registration of stack map entries with the runtime.
105 * The functions used by the program to allocate memory, trigger a collection,
108 * Computation or compilation of type maps, or registration of them with the
109 runtime. These are used to crawl the heap for object references.
111 In general, LLVM's support for GC does not include features which can be
112 adequately addressed with other features of the IR and does not specify a
113 particular binary interface. On the plus side, this means that you should be
114 able to integrate LLVM with an existing runtime. On the other hand, it leaves a
115 lot of work for the developer of a novel language. However, it's easy to get
116 started quickly and scale up to a more sophisticated implementation as your
122 Using a GC with LLVM implies many things, for example:
124 * Write a runtime library or find an existing one which implements a GC heap.
126 #. Implement a memory allocator.
128 #. Design a binary interface for the stack map, used to identify references
129 within a stack frame on the machine stack.\*
131 #. Implement a stack crawler to discover functions on the call stack.\*
133 #. Implement a registry for global roots.
135 #. Design a binary interface for type maps, used to identify references
138 #. Implement a collection routine bringing together all of the above.
140 * Emit compatible code from your compiler.
142 * Initialization in the main function.
144 * Use the ``gc "..."`` attribute to enable GC code generation (or
147 * Use ``@llvm.gcroot`` to mark stack roots.
149 * Use ``@llvm.gcread`` and/or ``@llvm.gcwrite`` to manipulate GC references,
152 * Allocate memory using the GC allocation routine provided by the runtime
155 * Generate type maps according to your runtime's binary interface.
157 * Write a compiler plugin to interface LLVM with the runtime library.\*
159 * Lower ``@llvm.gcread`` and ``@llvm.gcwrite`` to appropriate code
162 * Compile LLVM's stack map to the binary form expected by the runtime.
164 * Load the plugin into the compiler. Use ``llc -load`` or link the plugin
165 statically with your language's compiler.\*
167 * Link program executables with the runtime.
169 To help with several of these tasks (those indicated with a \*), LLVM includes a
170 highly portable, built-in ShadowStack code generator. It is compiled into
171 ``llc`` and works even with the interpreter and C backends.
176 To turn the shadow stack on for your functions, first call:
180 F.setGC("shadow-stack");
182 for each function your compiler emits. Since the shadow stack is built into
183 LLVM, you do not need to load a plugin.
185 Your compiler must also use ``@llvm.gcroot`` as documented. Don't forget to
186 create a root for each intermediate value that is generated when evaluating an
187 expression. In ``h(f(), g())``, the result of ``f()`` could easily be collected
188 if evaluating ``g()`` triggers a collection.
190 There's no need to use ``@llvm.gcread`` and ``@llvm.gcwrite`` over plain
191 ``load`` and ``store`` for now. You will need them when switching to a more
197 The shadow stack doesn't imply a memory allocation algorithm. A semispace
198 collector or building atop ``malloc`` are great places to start, and can be
199 implemented with very little code.
201 When it comes time to collect, however, your runtime needs to traverse the stack
202 roots, and for this it needs to integrate with the shadow stack. Luckily, doing
203 so is very simple. (This code is heavily commented to help you understand the
204 data structure, but there are only 20 lines of meaningful code.)
208 /// @brief The map for a single function's stack frame. One of these is
209 /// compiled as constant data into the executable for each function.
211 /// Storage of metadata values is elided if the %metadata parameter to
212 /// @llvm.gcroot is null.
214 int32_t NumRoots; //< Number of roots in stack frame.
215 int32_t NumMeta; //< Number of metadata entries. May be < NumRoots.
216 const void *Meta[0]; //< Metadata for each root.
219 /// @brief A link in the dynamic shadow stack. One of these is embedded in
220 /// the stack frame of each function on the call stack.
222 StackEntry *Next; //< Link to next stack entry (the caller's).
223 const FrameMap *Map; //< Pointer to constant FrameMap.
224 void *Roots[0]; //< Stack roots (in-place array).
227 /// @brief The head of the singly-linked list of StackEntries. Functions push
228 /// and pop onto this in their prologue and epilogue.
230 /// Since there is only a global list, this technique is not threadsafe.
231 StackEntry *llvm_gc_root_chain;
233 /// @brief Calls Visitor(root, meta) for each GC root on the stack.
234 /// root and meta are exactly the values passed to
237 /// Visitor could be a function to recursively mark live objects. Or it
238 /// might copy them to another heap or generation.
240 /// @param Visitor A function to invoke for every GC root on the stack.
241 void visitGCRoots(void (*Visitor)(void **Root, const void *Meta)) {
242 for (StackEntry *R = llvm_gc_root_chain; R; R = R->Next) {
245 // For roots [0, NumMeta), the metadata pointer is in the FrameMap.
246 for (unsigned e = R->Map->NumMeta; i != e; ++i)
247 Visitor(&R->Roots[i], R->Map->Meta[i]);
249 // For roots [NumMeta, NumRoots), the metadata pointer is null.
250 for (unsigned e = R->Map->NumRoots; i != e; ++i)
251 Visitor(&R->Roots[i], NULL);
255 About the shadow stack
256 ----------------------
258 Unlike many GC algorithms which rely on a cooperative code generator to compile
259 stack maps, this algorithm carefully maintains a linked list of stack roots
260 [:ref:`Henderson2002 <henderson02>`]. This so-called "shadow stack" mirrors the
261 machine stack. Maintaining this data structure is slower than using a stack map
262 compiled into the executable as constant data, but has a significant portability
263 advantage because it requires no special support from the target code generator,
264 and does not require tricky platform-specific code to crawl the machine stack.
266 The tradeoff for this simplicity and portability is:
268 * High overhead per function call.
272 Still, it's an easy way to get started. After your compiler and runtime are up
273 and running, writing a :ref:`plugin <plugin>` will allow you to take advantage
274 of :ref:`more advanced GC features <collector-algos>` of LLVM in order to
282 This section describes the garbage collection facilities provided by the
283 :doc:`LLVM intermediate representation <LangRef>`. The exact behavior of these
284 IR features is specified by the binary interface implemented by a :ref:`code
285 generation plugin <plugin>`, not by this document.
287 These facilities are limited to those strictly necessary; they are not intended
288 to be a complete interface to any garbage collector. A program will need to
289 interface with the GC library using the facilities provided by that program.
291 Specifying GC code generation: ``gc "..."``
292 -------------------------------------------
296 define ty @name(...) gc "name" { ...
298 The ``gc`` function attribute is used to specify the desired GC style to the
299 compiler. Its programmatic equivalent is the ``setGC`` method of ``Function``.
301 Setting ``gc "name"`` on a function triggers a search for a matching code
302 generation plugin "*name*"; it is that plugin which defines the exact nature of
303 the code generated to support GC. If none is found, the compiler will raise an
306 Specifying the GC style on a per-function basis allows LLVM to link together
307 programs that use different garbage collection algorithms (or none at all).
311 Identifying GC roots on the stack: ``llvm.gcroot``
312 --------------------------------------------------
316 void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
318 The ``llvm.gcroot`` intrinsic is used to inform LLVM that a stack variable
319 references an object on the heap and is to be tracked for garbage collection.
320 The exact impact on generated code is specified by a :ref:`compiler plugin
321 <plugin>`. All calls to ``llvm.gcroot`` **must** reside inside the first basic
324 A compiler which uses mem2reg to raise imperative code using ``alloca`` into SSA
325 form need only add a call to ``@llvm.gcroot`` for those variables which a
326 pointers into the GC heap.
328 It is also important to mark intermediate values with ``llvm.gcroot``. For
329 example, consider ``h(f(), g())``. Beware leaking the result of ``f()`` in the
330 case that ``g()`` triggers a collection. Note, that stack variables must be
331 initialized and marked with ``llvm.gcroot`` in function's prologue.
333 The first argument **must** be a value referring to an alloca instruction or a
334 bitcast of an alloca. The second contains a pointer to metadata that should be
335 associated with the pointer, and **must** be a constant or global value
336 address. If your target collector uses tags, use a null pointer for metadata.
338 The ``%metadata`` argument can be used to avoid requiring heap objects to have
339 'isa' pointers or tag bits. [Appel89_, Goldberg91_, Tolmach94_] If specified,
340 its value will be tracked along with the location of the pointer in the stack
343 Consider the following fragment of Java code:
348 Object X; // A null-initialized reference to an object
352 This block (which may be located in the middle of a function or in a loop nest),
353 could be compiled to this LLVM code:
358 ;; In the entry block for the function, allocate the
359 ;; stack space for X, which is an LLVM pointer.
362 ;; Tell LLVM that the stack space is a stack root.
363 ;; Java has type-tags on objects, so we pass null as metadata.
364 %tmp = bitcast %Object** %X to i8**
365 call void @llvm.gcroot(i8** %tmp, i8* null)
368 ;; "CodeBlock" is the block corresponding to the start
369 ;; of the scope above.
371 ;; Java null-initializes pointers.
372 store %Object* null, %Object** %X
376 ;; As the pointer goes out of scope, store a null value into
377 ;; it, to indicate that the value is no longer live.
378 store %Object* null, %Object** %X
381 Reading and writing references in the heap
382 ------------------------------------------
384 Some collectors need to be informed when the mutator (the program that needs
385 garbage collection) either reads a pointer from or writes a pointer to a field
386 of a heap object. The code fragments inserted at these points are called *read
387 barriers* and *write barriers*, respectively. The amount of code that needs to
388 be executed is usually quite small and not on the critical path of any
389 computation, so the overall performance impact of the barrier is tolerable.
391 Barriers often require access to the *object pointer* rather than the *derived
392 pointer* (which is a pointer to the field within the object). Accordingly,
393 these intrinsics take both pointers as separate arguments for completeness. In
394 this snippet, ``%object`` is the object pointer, and ``%derived`` is the derived
400 %class.Array = type { %class.Object, i32, [0 x %class.Object*] }
403 ;; Load the object pointer from a gcroot.
404 %object = load %class.Array** %object_addr
406 ;; Compute the derived pointer.
407 %derived = getelementptr %object, i32 0, i32 2, i32 %n
409 LLVM does not enforce this relationship between the object and derived pointer
410 (although a :ref:`plugin <plugin>` might). However, it would be an unusual
411 collector that violated it.
413 The use of these intrinsics is naturally optional if the target GC does require
414 the corresponding barrier. Such a GC plugin will replace the intrinsic calls
415 with the corresponding ``load`` or ``store`` instruction if they are used.
417 Write barrier: ``llvm.gcwrite``
418 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
422 void @llvm.gcwrite(i8* %value, i8* %object, i8** %derived)
424 For write barriers, LLVM provides the ``llvm.gcwrite`` intrinsic function. It
425 has exactly the same semantics as a non-volatile ``store`` to the derived
426 pointer (the third argument). The exact code generated is specified by a
427 compiler :ref:`plugin <plugin>`.
429 Many important algorithms require write barriers, including generational and
430 concurrent collectors. Additionally, write barriers could be used to implement
433 Read barrier: ``llvm.gcread``
434 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
438 i8* @llvm.gcread(i8* %object, i8** %derived)
440 For read barriers, LLVM provides the ``llvm.gcread`` intrinsic function. It has
441 exactly the same semantics as a non-volatile ``load`` from the derived pointer
442 (the second argument). The exact code generated is specified by a
443 :ref:`compiler plugin <plugin>`.
445 Read barriers are needed by fewer algorithms than write barriers, and may have a
446 greater performance impact since pointer reads are more frequent than writes.
450 Implementing a collector plugin
451 ===============================
453 User code specifies which GC code generation to use with the ``gc`` function
454 attribute or, equivalently, with the ``setGC`` method of ``Function``.
456 To implement a GC plugin, it is necessary to subclass ``llvm::GCStrategy``,
457 which can be accomplished in a few lines of boilerplate code. LLVM's
458 infrastructure provides access to several important algorithms. For an
459 uncontroversial collector, all that remains may be to compile LLVM's computed
460 stack map to assembly code (using the binary representation expected by the
461 runtime library). This can be accomplished in about 100 lines of code.
463 This is not the appropriate place to implement a garbage collected heap or a
464 garbage collector itself. That code should exist in the language's runtime
465 library. The compiler plugin is responsible for generating code which conforms
466 to the binary interface defined by library, most essentially the :ref:`stack map
469 To subclass ``llvm::GCStrategy`` and register it with the compiler:
473 // lib/MyGC/MyGC.cpp - Example LLVM GC plugin
475 #include "llvm/CodeGen/GCStrategy.h"
476 #include "llvm/CodeGen/GCMetadata.h"
477 #include "llvm/Support/Compiler.h"
479 using namespace llvm;
482 class LLVM_LIBRARY_VISIBILITY MyGC : public GCStrategy {
487 GCRegistry::Add<MyGC>
488 X("mygc", "My bespoke garbage collector.");
491 This boilerplate collector does nothing. More specifically:
493 * ``llvm.gcread`` calls are replaced with the corresponding ``load``
496 * ``llvm.gcwrite`` calls are replaced with the corresponding ``store``
499 * No safe points are added to the code.
501 * The stack map is not compiled into the executable.
503 Using the LLVM makefiles, this code
504 can be compiled as a plugin using a simple makefile:
514 include $(LEVEL)/Makefile.common
516 Once the plugin is compiled, code using it may be compiled using ``llc
517 -load=MyGC.so`` (though MyGC.so may have some other platform-specific
523 define void @f() gc "mygc" {
527 $ llvm-as < sample.ll | llc -load=MyGC.so
529 It is also possible to statically link the collector plugin into tools, such as
530 a language-specific compiler front-end.
534 Overview of available features
535 ------------------------------
537 ``GCStrategy`` provides a range of features through which a plugin may do useful
538 work. Some of these are callbacks, some are algorithms that can be enabled,
539 disabled, or customized. This matrix summarizes the supported (and planned)
540 features and correlates them with the collection techniques which typically
543 .. |v| unicode:: 0x2714
546 .. |x| unicode:: 0x2718
549 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
550 | Algorithm | Done | Shadow | refcount | mark- | copying | incremental | threaded | concurrent |
551 | | | stack | | sweep | | | | |
552 +============+======+========+==========+=======+=========+=============+==========+============+
553 | stack map | |v| | | | |x| | |x| | |x| | |x| | |x| |
554 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
555 | initialize | |v| | |x| | |x| | |x| | |x| | |x| | |x| | |x| |
556 | roots | | | | | | | | |
557 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
558 | derived | NO | | | | | | **N**\* | **N**\* |
559 | pointers | | | | | | | | |
560 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
561 | **custom | |v| | | | | | | | |
562 | lowering** | | | | | | | | |
563 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
564 | *gcroot* | |v| | |x| | |x| | | | | | |
565 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
566 | *gcwrite* | |v| | | |x| | | | |x| | | |x| |
567 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
568 | *gcread* | |v| | | | | | | | |x| |
569 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
570 | **safe | | | | | | | | |
571 | points** | | | | | | | | |
572 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
573 | *in | |v| | | | |x| | |x| | |x| | |x| | |x| |
574 | calls* | | | | | | | | |
575 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
576 | *before | |v| | | | | | | |x| | |x| |
577 | calls* | | | | | | | | |
578 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
579 | *for | NO | | | | | | **N** | **N** |
580 | loops* | | | | | | | | |
581 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
582 | *before | |v| | | | | | | |x| | |x| |
583 | escape* | | | | | | | | |
584 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
585 | emit code | NO | | | | | | **N** | **N** |
586 | at safe | | | | | | | | |
587 | points | | | | | | | | |
588 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
589 | **output** | | | | | | | | |
590 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
591 | *assembly* | |v| | | | |x| | |x| | |x| | |x| | |x| |
592 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
593 | *JIT* | NO | | | **?** | **?** | **?** | **?** | **?** |
594 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
595 | *obj* | NO | | | **?** | **?** | **?** | **?** | **?** |
596 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
597 | live | NO | | | **?** | **?** | **?** | **?** | **?** |
598 | analysis | | | | | | | | |
599 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
600 | register | NO | | | **?** | **?** | **?** | **?** | **?** |
601 | map | | | | | | | | |
602 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
603 | \* Derived pointers only pose a hasard to copying collections. |
604 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
605 | **?** denotes a feature which could be utilized if available. |
606 +------------+------+--------+----------+-------+---------+-------------+----------+------------+
608 To be clear, the collection techniques above are defined as:
611 The mutator carefully maintains a linked list of stack roots.
614 The mutator maintains a reference count for each object and frees an object
615 when its count falls to zero.
618 When the heap is exhausted, the collector marks reachable objects starting
619 from the roots, then deallocates unreachable objects in a sweep phase.
622 As reachability analysis proceeds, the collector copies objects from one heap
623 area to another, compacting them in the process. Copying collectors enable
624 highly efficient "bump pointer" allocation and can improve locality of
628 (Including generational collectors.) Incremental collectors generally have all
629 the properties of a copying collector (regardless of whether the mature heap
630 is compacting), but bring the added complexity of requiring write barriers.
633 Denotes a multithreaded mutator; the collector must still stop the mutator
634 ("stop the world") before beginning reachability analysis. Stopping a
635 multithreaded mutator is a complicated problem. It generally requires highly
636 platform-specific code in the runtime, and the production of carefully
637 designed machine code at safe points.
640 In this technique, the mutator and the collector run concurrently, with the
641 goal of eliminating pause times. In a *cooperative* collector, the mutator
642 further aids with collection should a pause occur, allowing collection to take
643 advantage of multiprocessor hosts. The "stop the world" problem of threaded
644 collectors is generally still present to a limited extent. Sophisticated
645 marking algorithms are necessary. Read barriers may be necessary.
647 As the matrix indicates, LLVM's garbage collection infrastructure is already
648 suitable for a wide variety of collectors, but does not currently extend to
649 multithreaded programs. This will be added in the future as there is
657 LLVM automatically computes a stack map. One of the most important features
658 of a ``GCStrategy`` is to compile this information into the executable in
659 the binary representation expected by the runtime library.
661 The stack map consists of the location and identity of each GC root in the
662 each function in the module. For each root:
664 * ``RootNum``: The index of the root.
666 * ``StackOffset``: The offset of the object relative to the frame pointer.
668 * ``RootMetadata``: The value passed as the ``%metadata`` parameter to the
669 ``@llvm.gcroot`` intrinsic.
671 Also, for the function as a whole:
673 * ``getFrameSize()``: The overall size of the function's initial stack frame,
674 not accounting for any dynamic allocation.
676 * ``roots_size()``: The count of roots in the function.
678 To access the stack map, use ``GCFunctionMetadata::roots_begin()`` and
679 -``end()`` from the :ref:`GCMetadataPrinter <assembly>`:
683 for (iterator I = begin(), E = end(); I != E; ++I) {
684 GCFunctionInfo *FI = *I;
685 unsigned FrameSize = FI->getFrameSize();
686 size_t RootCount = FI->roots_size();
688 for (GCFunctionInfo::roots_iterator RI = FI->roots_begin(),
689 RE = FI->roots_end();
691 int RootNum = RI->Num;
692 int RootStackOffset = RI->StackOffset;
693 Constant *RootMetadata = RI->Metadata;
697 If the ``llvm.gcroot`` intrinsic is eliminated before code generation by a
698 custom lowering pass, LLVM will compute an empty stack map. This may be useful
699 for collector plugins which implement reference counting or a shadow stack.
703 Initializing roots to null: ``InitRoots``
704 -----------------------------------------
712 When set, LLVM will automatically initialize each root to ``null`` upon entry to
713 the function. This prevents the GC's sweep phase from visiting uninitialized
714 pointers, which will almost certainly cause it to crash. This initialization
715 occurs before custom lowering, so the two may be used together.
717 Since LLVM does not yet compute liveness information, there is no means of
718 distinguishing an uninitialized stack root from an initialized one. Therefore,
719 this feature should be used by all GC plugins. It is enabled by default.
721 Custom lowering of intrinsics: ``CustomRoots``, ``CustomReadBarriers``, and ``CustomWriteBarriers``
722 ---------------------------------------------------------------------------------------------------
724 For GCs which use barriers or unusual treatment of stack roots, these
725 flags allow the collector to perform arbitrary transformations of the
730 class MyGC : public GCStrategy {
734 CustomReadBarriers = true;
735 CustomWriteBarriers = true;
739 If any of these flags are set, LLVM suppresses its default lowering for
740 the corresponding intrinsics. Instead, you must provide a custom Pass
741 which lowers the intrinsics as desired. If you have opted in to custom
742 lowering of a particular intrinsic your pass **must** eliminate all
743 instances of the corresponding intrinsic in functions which opt in to
744 your GC. The best example of such a pass is the ShadowStackGC and it's
745 ShadowStackGCLowering pass.
747 There is currently no way to register such a custom lowering pass
748 without building a custom copy of LLVM.
752 Generating safe points: ``NeededSafePoints``
753 --------------------------------------------
755 LLVM can compute four kinds of safe points:
760 /// PointKind - The type of a collector-safe point.
763 Loop, //< Instr is a loop (backwards branch).
764 Return, //< Instr is a return instruction.
765 PreCall, //< Instr is a call instruction.
766 PostCall //< Instr is the return address of a call.
770 A collector can request any combination of the four by setting the
771 ``NeededSafePoints`` mask:
776 NeededSafePoints = 1 << GC::Loop
782 It can then use the following routines to access safe points.
786 for (iterator I = begin(), E = end(); I != E; ++I) {
787 GCFunctionInfo *MD = *I;
788 size_t PointCount = MD->size();
790 for (GCFunctionInfo::iterator PI = MD->begin(),
791 PE = MD->end(); PI != PE; ++PI) {
792 GC::PointKind PointKind = PI->Kind;
793 unsigned PointNum = PI->Num;
797 Almost every collector requires ``PostCall`` safe points, since these correspond
798 to the moments when the function is suspended during a call to a subroutine.
800 Threaded programs generally require ``Loop`` safe points to guarantee that the
801 application will reach a safe point within a bounded amount of time, even if it
802 is executing a long-running loop which contains no function calls.
804 Threaded collectors may also require ``Return`` and ``PreCall`` safe points to
805 implement "stop the world" techniques using self-modifying code, where it is
806 important that the program not exit the function without reaching a safe point
807 (because only the topmost function has been patched).
811 Emitting assembly code: ``GCMetadataPrinter``
812 ---------------------------------------------
814 LLVM allows a plugin to print arbitrary assembly code before and after the rest
815 of a module's assembly code. At the end of the module, the GC can compile the
816 LLVM stack map into assembly code. (At the beginning, this information is not
819 Since AsmWriter and CodeGen are separate components of LLVM, a separate abstract
820 base class and registry is provided for printing assembly code, the
821 ``GCMetadaPrinter`` and ``GCMetadataPrinterRegistry``. The AsmWriter will look
822 for such a subclass if the ``GCStrategy`` sets ``UsesMetadata``:
830 This separation allows JIT-only clients to be smaller.
832 Note that LLVM does not currently have analogous APIs to support code generation
833 in the JIT, nor using the object writers.
837 // lib/MyGC/MyGCPrinter.cpp - Example LLVM GC printer
839 #include "llvm/CodeGen/GCMetadataPrinter.h"
840 #include "llvm/Support/Compiler.h"
842 using namespace llvm;
845 class LLVM_LIBRARY_VISIBILITY MyGCPrinter : public GCMetadataPrinter {
847 virtual void beginAssembly(AsmPrinter &AP);
849 virtual void finishAssembly(AsmPrinter &AP);
852 GCMetadataPrinterRegistry::Add<MyGCPrinter>
853 X("mygc", "My bespoke garbage collector.");
856 The collector should use ``AsmPrinter`` to print portable assembly code. The
857 collector itself contains the stack map for the entire module, and may access
858 the ``GCFunctionInfo`` using its own ``begin()`` and ``end()`` methods. Here's
863 #include "llvm/CodeGen/AsmPrinter.h"
864 #include "llvm/IR/Function.h"
865 #include "llvm/IR/DataLayout.h"
866 #include "llvm/Target/TargetAsmInfo.h"
867 #include "llvm/Target/TargetMachine.h"
869 void MyGCPrinter::beginAssembly(AsmPrinter &AP) {
873 void MyGCPrinter::finishAssembly(AsmPrinter &AP) {
874 MCStreamer &OS = AP.OutStreamer;
875 unsigned IntPtrSize = AP.TM.getSubtargetImpl()->getDataLayout()->getPointerSize();
877 // Put this in the data section.
878 OS.SwitchSection(AP.getObjFileLowering().getDataSection());
880 // For each function...
881 for (iterator FI = begin(), FE = end(); FI != FE; ++FI) {
882 GCFunctionInfo &MD = **FI;
884 // A compact GC layout. Emit this data structure:
887 // int32_t PointCount;
888 // void *SafePointAddress[PointCount];
889 // int32_t StackFrameSize; // in words
890 // int32_t StackArity;
891 // int32_t LiveCount;
892 // int32_t LiveOffsets[LiveCount];
893 // } __gcmap_<FUNCTIONNAME>;
895 // Align to address width.
896 AP.EmitAlignment(IntPtrSize == 4 ? 2 : 3);
899 OS.AddComment("safe point count");
900 AP.EmitInt32(MD.size());
902 // And each safe point...
903 for (GCFunctionInfo::iterator PI = MD.begin(),
904 PE = MD.end(); PI != PE; ++PI) {
905 // Emit the address of the safe point.
906 OS.AddComment("safe point address");
907 MCSymbol *Label = PI->Label;
908 AP.EmitLabelPlusOffset(Label/*Hi*/, 0/*Offset*/, 4/*Size*/);
911 // Stack information never change in safe points! Only print info from the
913 GCFunctionInfo::iterator PI = MD.begin();
915 // Emit the stack frame size.
916 OS.AddComment("stack frame size (in words)");
917 AP.EmitInt32(MD.getFrameSize() / IntPtrSize);
919 // Emit stack arity, i.e. the number of stacked arguments.
920 unsigned RegisteredArgs = IntPtrSize == 4 ? 5 : 6;
921 unsigned StackArity = MD.getFunction().arg_size() > RegisteredArgs ?
922 MD.getFunction().arg_size() - RegisteredArgs : 0;
923 OS.AddComment("stack arity");
924 AP.EmitInt32(StackArity);
926 // Emit the number of live roots in the function.
927 OS.AddComment("live root count");
928 AP.EmitInt32(MD.live_size(PI));
930 // And for each live root...
931 for (GCFunctionInfo::live_iterator LI = MD.live_begin(PI),
932 LE = MD.live_end(PI);
934 // Emit live root's offset within the stack frame.
935 OS.AddComment("stack index (offset / wordsize)");
936 AP.EmitInt32(LI->StackOffset);
946 [Appel89] Runtime Tags Aren't Necessary. Andrew W. Appel. Lisp and Symbolic
947 Computation 19(7):703-705, July 1989.
951 [Goldberg91] Tag-free garbage collection for strongly typed programming
952 languages. Benjamin Goldberg. ACM SIGPLAN PLDI'91.
956 [Tolmach94] Tag-free garbage collection using explicit type parameters. Andrew
957 Tolmach. Proceedings of the 1994 ACM conference on LISP and functional
962 [Henderson2002] `Accurate Garbage Collection in an Uncooperative Environment
963 <http://citeseer.ist.psu.edu/henderson02accurate.html>`__