1=====================================
2Garbage Collection with LLVM
3=====================================
4
5.. contents::
6   :local:
7
8Abstract
9========
10
11This document covers how to integrate LLVM into a compiler for a language which
12supports garbage collection.  **Note that LLVM itself does not provide a
13garbage collector.**  You must provide your own.
14
15Quick Start
16============
17
18First, you should pick a collector strategy.  LLVM includes a number of built
19in ones, but you can also implement a loadable plugin with a custom definition.
20Note that the collector strategy is a description of how LLVM should generate
21code such that it interacts with your collector and runtime, not a description
22of the collector itself.
23
24Next, mark your generated functions as using your chosen collector strategy.
25From c++, you can call:
26
27.. code-block:: c++
28
29  F.setGC(<collector description name>);
30
31
32This will produce IR like the following fragment:
33
34.. code-block:: llvm
35
36  define void @foo() gc "<collector description name>" { ... }
37
38
39When generating LLVM IR for your functions, you will need to:
40
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
44  this step.
45
46* Use the memory allocation routines provided by your garbage collector's
47  runtime library.
48
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.
53
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
57  as a safepoint poll.
58
59You will need to identify roots (i.e. references to heap objects your collector
60needs to know about) in your generated IR, so that LLVM can encode them into
61your final stack maps.  Depending on the collector strategy chosen, this is
62accomplished by using either the ``@llvm.gcroot`` intrinsics or an
63``gc.statepoint`` relocation sequence.
64
65Don't forget to create a root for each intermediate value that is generated when
66evaluating an expression.  In ``h(f(), g())``, the result of ``f()`` could
67easily be collected if evaluating ``g()`` triggers a collection.
68
69Finally, you need to link your runtime library with the generated program
70executable (for a static compiler) or ensure the appropriate symbols are
71available for the runtime linker (for a JIT compiler).
72
73
74Introduction
75============
76
77What is Garbage Collection?
78---------------------------
79
80Garbage collection is a widely used technique that frees the programmer from
81having to know the lifetimes of heap objects, making software easier to produce
82and maintain.  Many programming languages rely on garbage collection for
83automatic memory management.  There are two primary forms of garbage collection:
84conservative and accurate.
85
86Conservative garbage collection often does not require any special support from
87either the language or the compiler: it can handle non-type-safe programming
88languages (such as C/C++) and does not require any special information from the
89compiler.  The `Boehm collector
90<http://www.hpl.hp.com/personal/Hans_Boehm/gc/>`__ is an example of a
91state-of-the-art conservative collector.
92
93Accurate garbage collection requires the ability to identify all pointers in the
94program at run-time (which requires that the source-language be type-safe in
95most cases).  Identifying pointers at run-time requires compiler support to
96locate all places that hold live pointer variables at run-time, including the
97:ref:`processor stack and registers <gcroot>`.
98
99Conservative garbage collection is attractive because it does not require any
100special compiler support, but it does have problems.  In particular, because the
101conservative garbage collector cannot *know* that a particular word in the
102machine is a pointer, it cannot move live objects in the heap (preventing the
103use of compacting and generational GC algorithms) and it can occasionally suffer
104from memory leaks due to integer values that happen to point to objects in the
105program.  In addition, some aggressive compiler transformations can break
106conservative garbage collectors (though these seem rare in practice).
107
108Accurate garbage collectors do not suffer from any of these problems, but they
109can suffer from degraded scalar optimization of the program.  In particular,
110because the runtime must be able to identify and update all pointers active in
111the program, some optimizations are less effective.  In practice, however, the
112locality and performance benefits of using aggressive garbage collection
113techniques dominates any low-level losses.
114
115This document describes the mechanisms and interfaces provided by LLVM to
116support accurate garbage collection.
117
118Goals and non-goals
119-------------------
120
121LLVM's intermediate representation provides :ref:`garbage collection intrinsics
122<gc_intrinsics>` that offer support for a broad class of collector models.  For
123instance, the intrinsics permit:
124
125* semi-space collectors
126
127* mark-sweep collectors
128
129* generational collectors
130
131* incremental collectors
132
133* concurrent collectors
134
135* cooperative collectors
136
137* reference counting
138
139We hope that the support built into the LLVM IR is sufficient to support a
140broad class of garbage collected languages including Scheme, ML, Java, C#,
141Perl, Python, Lua, Ruby, other scripting languages, and more.
142
143Note that LLVM **does not itself provide a garbage collector** --- this should
144be part of your language's runtime library.  LLVM provides a framework for
145describing the garbage collectors requirements to the compiler.  In particular,
146LLVM provides support for generating stack maps at call sites, polling for a
147safepoint, and emitting load and store barriers.  You can also extend LLVM -
148possibly through a loadable :ref:`code generation plugins <plugin>` - to
149generate code and data structures which conforms to the *binary interface*
150specified by the *runtime library*.  This is similar to the relationship between
151LLVM and DWARF debugging info, for example.  The difference primarily lies in
152the lack of an established standard in the domain of garbage collection --- thus
153the need for a flexible extension mechanism.
154
155The aspects of the binary interface with which LLVM's GC support is
156concerned are:
157
158* Creation of GC safepoints within code where collection is allowed to execute
159  safely.
160
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.
164
165* Write barriers when storing object references to the heap.  These are commonly
166  used to optimize incremental scans in generational collectors.
167
168* Emission of read barriers when loading object references.  These are useful
169  for interoperating with concurrent collectors.
170
171There are additional areas that LLVM does not directly address:
172
173* Registration of global roots with the runtime.
174
175* Registration of stack map entries with the runtime.
176
177* The functions used by the program to allocate memory, trigger a collection,
178  etc.
179
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.
182
183In general, LLVM's support for GC does not include features which can be
184adequately addressed with other features of the IR and does not specify a
185particular binary interface.  On the plus side, this means that you should be
186able to integrate LLVM with an existing runtime.  On the other hand, it can
187have the effect of leaving a lot of work for the developer of a novel
188language.  We try to mitigate this by providing built in collector strategy
189descriptions that can work with many common collector designs and easy
190extension points.  If you don't already have a specific binary interface
191you need to support, we recommend trying to use one of these built in collector
192strategies.
193
194.. _gc_intrinsics:
195
196LLVM IR Features
197================
198
199This section describes the garbage collection facilities provided by the
200:doc:`LLVM intermediate representation <LangRef>`.  The exact behavior of these
201IR features is specified by the selected :ref:`GC strategy description
202<plugin>`.
203
204Specifying GC code generation: ``gc "..."``
205-------------------------------------------
206
207.. code-block:: llvm
208
209  define <returntype> @name(...) gc "name" { ... }
210
211The ``gc`` function attribute is used to specify the desired GC strategy to the
212compiler.  Its programmatic equivalent is the ``setGC`` method of ``Function``.
213
214Setting ``gc "name"`` on a function triggers a search for a matching subclass
215of GCStrategy.  Some collector strategies are built in.  You can add others
216using either the loadable plugin mechanism, or by patching your copy of LLVM.
217It is the selected GC strategy which defines the exact nature of the code
218generated to support GC.  If none is found, the compiler will raise an error.
219
220Specifying the GC style on a per-function basis allows LLVM to link together
221programs that use different garbage collection algorithms (or none at all).
222
223.. _gcroot:
224
225Identifying GC roots on the stack
226----------------------------------
227
228LLVM currently supports two different mechanisms for describing references in
229compiled code at safepoints.  ``llvm.gcroot`` is the older mechanism;
230``gc.statepoint`` has been added more recently.  At the moment, you can choose
231either implementation (on a per :ref:`GC strategy <plugin>` basis).  Longer
232term, we will probably either migrate away from ``llvm.gcroot`` entirely, or
233substantially merge their implementations. Note that most new development
234work is focused on ``gc.statepoint``.
235
236Using ``gc.statepoint``
237^^^^^^^^^^^^^^^^^^^^^^^^
238:doc:`This page <Statepoints>` contains detailed documentation for
239``gc.statepoint``.
240
241Using ``llvm.gcwrite``
242^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
243
244.. code-block:: llvm
245
246  void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
247
248The ``llvm.gcroot`` intrinsic is used to inform LLVM that a stack variable
249references an object on the heap and is to be tracked for garbage collection.
250The exact impact on generated code is specified by the Function's selected
251:ref:`GC strategy <plugin>`.  All calls to ``llvm.gcroot`` **must** reside
252inside the first basic block.
253
254The first argument **must** be a value referring to an alloca instruction or a
255bitcast of an alloca.  The second contains a pointer to metadata that should be
256associated with the pointer, and **must** be a constant or global value
257address.  If your target collector uses tags, use a null pointer for metadata.
258
259A compiler which performs manual SSA construction **must** ensure that SSA
260values representing GC references are stored in to the alloca passed to the
261respective ``gcroot`` before every call site and reloaded after every call.
262A compiler which uses mem2reg to raise imperative code using ``alloca`` into
263SSA form need only add a call to ``@llvm.gcroot`` for those variables which
264are pointers into the GC heap.
265
266It is also important to mark intermediate values with ``llvm.gcroot``.  For
267example, consider ``h(f(), g())``.  Beware leaking the result of ``f()`` in the
268case that ``g()`` triggers a collection.  Note, that stack variables must be
269initialized and marked with ``llvm.gcroot`` in function's prologue.
270
271The ``%metadata`` argument can be used to avoid requiring heap objects to have
272'isa' pointers or tag bits. [Appel89_, Goldberg91_, Tolmach94_] If specified,
273its value will be tracked along with the location of the pointer in the stack
274frame.
275
276Consider the following fragment of Java code:
277
278.. code-block:: java
279
280   {
281     Object X;   // A null-initialized reference to an object
282     ...
283   }
284
285This block (which may be located in the middle of a function or in a loop nest),
286could be compiled to this LLVM code:
287
288.. code-block:: llvm
289
290  Entry:
291     ;; In the entry block for the function, allocate the
292     ;; stack space for X, which is an LLVM pointer.
293     %X = alloca %Object*
294
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)
299     ...
300
301     ;; "CodeBlock" is the block corresponding to the start
302     ;;  of the scope above.
303  CodeBlock:
304     ;; Java null-initializes pointers.
305     store %Object* null, %Object** %X
306
307     ...
308
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
312     ...
313
314Reading and writing references in the heap
315------------------------------------------
316
317Some collectors need to be informed when the mutator (the program that needs
318garbage collection) either reads a pointer from or writes a pointer to a field
319of a heap object.  The code fragments inserted at these points are called *read
320barriers* and *write barriers*, respectively.  The amount of code that needs to
321be executed is usually quite small and not on the critical path of any
322computation, so the overall performance impact of the barrier is tolerable.
323
324Barriers often require access to the *object pointer* rather than the *derived
325pointer* (which is a pointer to the field within the object).  Accordingly,
326these intrinsics take both pointers as separate arguments for completeness.  In
327this snippet, ``%object`` is the object pointer, and ``%derived`` is the derived
328pointer:
329
330.. code-block:: llvm
331
332  ;; An array type.
333  %class.Array = type { %class.Object, i32, [0 x %class.Object*] }
334  ...
335
336  ;; Load the object pointer from a gcroot.
337  %object = load %class.Array** %object_addr
338
339  ;; Compute the derived pointer.
340  %derived = getelementptr %object, i32 0, i32 2, i32 %n
341
342LLVM does not enforce this relationship between the object and derived pointer
343(although a particular :ref:`collector strategy <plugin>` might).  However, it
344would be an unusual collector that violated it.
345
346The use of these intrinsics is naturally optional if the target GC does not
347require the corresponding barrier.  The GC strategy used with such a collector
348should replace the intrinsic calls with the corresponding ``load`` or
349``store`` instruction if they are used.
350
351One known deficiency with the current design is that the barrier intrinsics do
352not include the size or alignment of the underlying operation performed.  It is
353currently assumed that the operation is of pointer size and the alignment is
354assumed to be the target machine's default alignment.
355
356Write barrier: ``llvm.gcwrite``
357^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
358
359.. code-block:: llvm
360
361  void @llvm.gcwrite(i8* %value, i8* %object, i8** %derived)
362
363For write barriers, LLVM provides the ``llvm.gcwrite`` intrinsic function.  It
364has exactly the same semantics as a non-volatile ``store`` to the derived
365pointer (the third argument).  The exact code generated is specified by the
366Function's selected :ref:`GC strategy <plugin>`.
367
368Many important algorithms require write barriers, including generational and
369concurrent collectors.  Additionally, write barriers could be used to implement
370reference counting.
371
372Read barrier: ``llvm.gcread``
373^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
374
375.. code-block:: llvm
376
377  i8* @llvm.gcread(i8* %object, i8** %derived)
378
379For read barriers, LLVM provides the ``llvm.gcread`` intrinsic function.  It has
380exactly 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
382selected :ref:`GC strategy <plugin>`.
383
384Read barriers are needed by fewer algorithms than write barriers, and may have a
385greater performance impact since pointer reads are more frequent than writes.
386
387.. _plugin:
388
389.. _builtin-gc-strategies:
390
391Built In GC Strategies
392======================
393
394LLVM includes built in support for several varieties of garbage collectors.
395
396The Shadow Stack GC
397----------------------
398
399To use this collector strategy, mark your functions with:
400
401.. code-block:: c++
402
403  F.setGC("shadow-stack");
404
405Unlike many GC algorithms which rely on a cooperative code generator to compile
406stack maps, this algorithm carefully maintains a linked list of stack roots
407[:ref:`Henderson2002 <henderson02>`].  This so-called "shadow stack" mirrors the
408machine stack.  Maintaining this data structure is slower than using a stack map
409compiled into the executable as constant data, but has a significant portability
410advantage because it requires no special support from the target code generator,
411and does not require tricky platform-specific code to crawl the machine stack.
412
413The tradeoff for this simplicity and portability is:
414
415* High overhead per function call.
416
417* Not thread-safe.
418
419Still, it's an easy way to get started.  After your compiler and runtime are up
420and running, writing a :ref:`plugin <plugin>` will allow you to take advantage
421of :ref:`more advanced GC features <collector-algos>` of LLVM in order to
422improve performance.
423
424
425The shadow stack doesn't imply a memory allocation algorithm.  A semispace
426collector or building atop ``malloc`` are great places to start, and can be
427implemented with very little code.
428
429When it comes time to collect, however, your runtime needs to traverse the stack
430roots, and for this it needs to integrate with the shadow stack.  Luckily, doing
431so is very simple. (This code is heavily commented to help you understand the
432data structure, but there are only 20 lines of meaningful code.)
433
434.. code-block:: c++
435
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.
438  ///
439  /// Storage of metadata values is elided if the %metadata parameter to
440  /// @llvm.gcroot is null.
441  struct FrameMap {
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.
445  };
446
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.
449  struct StackEntry {
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).
453  };
454
455  /// @brief The head of the singly-linked list of StackEntries.  Functions push
456  ///        and pop onto this in their prologue and epilogue.
457  ///
458  /// Since there is only a global list, this technique is not threadsafe.
459  StackEntry *llvm_gc_root_chain;
460
461  /// @brief Calls Visitor(root, meta) for each GC root on the stack.
462  ///        root and meta are exactly the values passed to
463  ///        @llvm.gcroot.
464  ///
465  /// Visitor could be a function to recursively mark live objects.  Or it
466  /// might copy them to another heap or generation.
467  ///
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) {
471      unsigned i = 0;
472
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]);
476
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);
480    }
481  }
482
483
484The 'Erlang' and 'Ocaml' GCs
485-----------------------------
486
487LLVM ships with two example collectors which leverage the ``gcroot``
488mechanisms.  To our knowledge, these are not actually used by any language
489runtime, but they do provide a reasonable starting point for someone interested
490in writing an ``gcroot`` compatible GC plugin.  In particular, these are the
491only in tree examples of how to produce a custom binary stack map format using
492a ``gcroot`` strategy.
493
494As there names imply, the binary format produced is intended to model that
495used by the Erlang and OCaml compilers respectively.
496
497.. _statepoint_example_gc:
498
499The Statepoint Example GC
500-------------------------
501
502.. code-block:: c++
503
504  F.setGC("statepoint-example");
505
506This GC provides an example of how one might use the infrastructure provided
507by ``gc.statepoint``. This example GC is compatible with the
508:ref:`PlaceSafepoints` and :ref:`RewriteStatepointsForGC` utility passes
509which simplify ``gc.statepoint`` sequence insertion. If you need to build a
510custom GC strategy around the ``gc.statepoints`` mechanisms, it is recommended
511that you use this one as a starting point.
512
513This GC strategy does not support read or write barriers.  As a result, these
514intrinsics are lowered to normal loads and stores.
515
516The stack map format generated by this GC strategy can be found in the
517:ref:`stackmap-section` using a format documented :ref:`here
518<statepoint-stackmap-format>`. This format is intended to be the standard
519format supported by LLVM going forward.
520
521The CoreCLR GC
522-------------------------
523
524.. code-block:: c++
525
526  F.setGC("coreclr");
527
528This GC leverages the ``gc.statepoint`` mechanism to support the
529`CoreCLR <https://github.com/dotnet/coreclr>`__ runtime.
530
531Support for this GC strategy is a work in progress. This strategy will
532differ from
533:ref:`statepoint-example GC<statepoint_example_gc>` strategy in
534certain aspects like:
535
536* Base-pointers of interior pointers are not explicitly
537  tracked and reported.
538
539* A different format is used for encoding stack maps.
540
541* Safe-point polls are only needed before loop-back edges
542  and before tail-calls (not needed at function-entry).
543
544Custom GC Strategies
545====================
546
547If none of the built in GC strategy descriptions met your needs above, you will
548need to define a custom GCStrategy and possibly, a custom LLVM pass to perform
549lowering.  Your best example of where to start defining a custom GCStrategy
550would be to look at one of the built in strategies.
551
552You may be able to structure this additional code as a loadable plugin library.
553Loadable plugins are sufficient if all you need is to enable a different
554combination of built in functionality, but if you need to provide a custom
555lowering pass, you will need to build a patched version of LLVM.  If you think
556you need a patched build, please ask for advice on llvm-dev.  There may be an
557easy way we can extend the support to make it work for your use case without
558requiring a custom build.
559
560Collector Requirements
561----------------------
562
563You should be able to leverage any existing collector library that includes the following elements:
564
565#. A memory allocator which exposes an allocation function your compiled
566   code can call.
567
568#. A binary format for the stack map.  A stack map describes the location
569   of references at a safepoint and is used by precise collectors to identify
570   references within a stack frame on the machine stack. Note that collectors
571   which conservatively scan the stack don't require such a structure.
572
573#. A stack crawler to discover functions on the call stack, and enumerate the
574   references listed in the stack map for each call site.
575
576#. A mechanism for identifying references in global locations (e.g. global
577   variables).
578
579#. If you collector requires them, an LLVM IR implementation of your collectors
580   load and store barriers.  Note that since many collectors don't require
581   barriers at all, LLVM defaults to lowering such barriers to normal loads
582   and stores unless you arrange otherwise.
583
584
585Implementing a collector plugin
586-------------------------------
587
588User code specifies which GC code generation to use with the ``gc`` function
589attribute or, equivalently, with the ``setGC`` method of ``Function``.
590
591To implement a GC plugin, it is necessary to subclass ``llvm::GCStrategy``,
592which can be accomplished in a few lines of boilerplate code.  LLVM's
593infrastructure provides access to several important algorithms.  For an
594uncontroversial collector, all that remains may be to compile LLVM's computed
595stack map to assembly code (using the binary representation expected by the
596runtime library).  This can be accomplished in about 100 lines of code.
597
598This is not the appropriate place to implement a garbage collected heap or a
599garbage collector itself.  That code should exist in the language's runtime
600library.  The compiler plugin is responsible for generating code which conforms
601to the binary interface defined by library, most essentially the :ref:`stack map
602<stack-map>`.
603
604To subclass ``llvm::GCStrategy`` and register it with the compiler:
605
606.. code-block:: c++
607
608  // lib/MyGC/MyGC.cpp - Example LLVM GC plugin
609
610  #include "llvm/CodeGen/GCStrategy.h"
611  #include "llvm/CodeGen/GCMetadata.h"
612  #include "llvm/Support/Compiler.h"
613
614  using namespace llvm;
615
616  namespace {
617    class LLVM_LIBRARY_VISIBILITY MyGC : public GCStrategy {
618    public:
619      MyGC() {}
620    };
621
622    GCRegistry::Add<MyGC>
623    X("mygc", "My bespoke garbage collector.");
624  }
625
626This boilerplate collector does nothing.  More specifically:
627
628* ``llvm.gcread`` calls are replaced with the corresponding ``load``
629  instruction.
630
631* ``llvm.gcwrite`` calls are replaced with the corresponding ``store``
632  instruction.
633
634* No safe points are added to the code.
635
636* The stack map is not compiled into the executable.
637
638Using the LLVM makefiles, this code
639can be compiled as a plugin using a simple makefile:
640
641.. code-block:: make
642
643  # lib/MyGC/Makefile
644
645  LEVEL := ../..
646  LIBRARYNAME = MyGC
647  LOADABLE_MODULE = 1
648
649  include $(LEVEL)/Makefile.common
650
651Once the plugin is compiled, code using it may be compiled using ``llc
652-load=MyGC.so`` (though MyGC.so may have some other platform-specific
653extension):
654
655::
656
657  $ cat sample.ll
658  define void @f() gc "mygc" {
659  entry:
660    ret void
661  }
662  $ llvm-as < sample.ll | llc -load=MyGC.so
663
664It is also possible to statically link the collector plugin into tools, such as
665a language-specific compiler front-end.
666
667.. _collector-algos:
668
669Overview of available features
670------------------------------
671
672``GCStrategy`` provides a range of features through which a plugin may do useful
673work.  Some of these are callbacks, some are algorithms that can be enabled,
674disabled, or customized.  This matrix summarizes the supported (and planned)
675features and correlates them with the collection techniques which typically
676require them.
677
678.. |v| unicode:: 0x2714
679   :trim:
680
681.. |x| unicode:: 0x2718
682   :trim:
683
684+------------+------+--------+----------+-------+---------+-------------+----------+------------+
685| Algorithm  | Done | Shadow | refcount | mark- | copying | incremental | threaded | concurrent |
686|            |      | stack  |          | sweep |         |             |          |            |
687+============+======+========+==========+=======+=========+=============+==========+============+
688| stack map  | |v|  |        |          | |x|   | |x|     | |x|         | |x|      | |x|        |
689+------------+------+--------+----------+-------+---------+-------------+----------+------------+
690| initialize | |v|  | |x|    | |x|      | |x|   | |x|     | |x|         | |x|      | |x|        |
691| roots      |      |        |          |       |         |             |          |            |
692+------------+------+--------+----------+-------+---------+-------------+----------+------------+
693| derived    | NO   |        |          |       |         |             | **N**\*  | **N**\*    |
694| pointers   |      |        |          |       |         |             |          |            |
695+------------+------+--------+----------+-------+---------+-------------+----------+------------+
696| **custom   | |v|  |        |          |       |         |             |          |            |
697| lowering** |      |        |          |       |         |             |          |            |
698+------------+------+--------+----------+-------+---------+-------------+----------+------------+
699| *gcroot*   | |v|  | |x|    | |x|      |       |         |             |          |            |
700+------------+------+--------+----------+-------+---------+-------------+----------+------------+
701| *gcwrite*  | |v|  |        | |x|      |       |         | |x|         |          | |x|        |
702+------------+------+--------+----------+-------+---------+-------------+----------+------------+
703| *gcread*   | |v|  |        |          |       |         |             |          | |x|        |
704+------------+------+--------+----------+-------+---------+-------------+----------+------------+
705| **safe     |      |        |          |       |         |             |          |            |
706| points**   |      |        |          |       |         |             |          |            |
707+------------+------+--------+----------+-------+---------+-------------+----------+------------+
708| *in        | |v|  |        |          | |x|   | |x|     | |x|         | |x|      | |x|        |
709| calls*     |      |        |          |       |         |             |          |            |
710+------------+------+--------+----------+-------+---------+-------------+----------+------------+
711| *before    | |v|  |        |          |       |         |             | |x|      | |x|        |
712| calls*     |      |        |          |       |         |             |          |            |
713+------------+------+--------+----------+-------+---------+-------------+----------+------------+
714| *for       | NO   |        |          |       |         |             | **N**    | **N**      |
715| loops*     |      |        |          |       |         |             |          |            |
716+------------+------+--------+----------+-------+---------+-------------+----------+------------+
717| *before    | |v|  |        |          |       |         |             | |x|      | |x|        |
718| escape*    |      |        |          |       |         |             |          |            |
719+------------+------+--------+----------+-------+---------+-------------+----------+------------+
720| emit code  | NO   |        |          |       |         |             | **N**    | **N**      |
721| at safe    |      |        |          |       |         |             |          |            |
722| points     |      |        |          |       |         |             |          |            |
723+------------+------+--------+----------+-------+---------+-------------+----------+------------+
724| **output** |      |        |          |       |         |             |          |            |
725+------------+------+--------+----------+-------+---------+-------------+----------+------------+
726| *assembly* | |v|  |        |          | |x|   | |x|     | |x|         | |x|      | |x|        |
727+------------+------+--------+----------+-------+---------+-------------+----------+------------+
728| *JIT*      | NO   |        |          | **?** | **?**   | **?**       | **?**    | **?**      |
729+------------+------+--------+----------+-------+---------+-------------+----------+------------+
730| *obj*      | NO   |        |          | **?** | **?**   | **?**       | **?**    | **?**      |
731+------------+------+--------+----------+-------+---------+-------------+----------+------------+
732| live       | NO   |        |          | **?** | **?**   | **?**       | **?**    | **?**      |
733| analysis   |      |        |          |       |         |             |          |            |
734+------------+------+--------+----------+-------+---------+-------------+----------+------------+
735| register   | NO   |        |          | **?** | **?**   | **?**       | **?**    | **?**      |
736| map        |      |        |          |       |         |             |          |            |
737+------------+------+--------+----------+-------+---------+-------------+----------+------------+
738| \* Derived pointers only pose a hasard to copying collections.                                |
739+------------+------+--------+----------+-------+---------+-------------+----------+------------+
740| **?** denotes a feature which could be utilized if available.                                 |
741+------------+------+--------+----------+-------+---------+-------------+----------+------------+
742
743To be clear, the collection techniques above are defined as:
744
745Shadow Stack
746  The mutator carefully maintains a linked list of stack roots.
747
748Reference Counting
749  The mutator maintains a reference count for each object and frees an object
750  when its count falls to zero.
751
752Mark-Sweep
753  When the heap is exhausted, the collector marks reachable objects starting
754  from the roots, then deallocates unreachable objects in a sweep phase.
755
756Copying
757  As reachability analysis proceeds, the collector copies objects from one heap
758  area to another, compacting them in the process.  Copying collectors enable
759  highly efficient "bump pointer" allocation and can improve locality of
760  reference.
761
762Incremental
763  (Including generational collectors.) Incremental collectors generally have all
764  the properties of a copying collector (regardless of whether the mature heap
765  is compacting), but bring the added complexity of requiring write barriers.
766
767Threaded
768  Denotes a multithreaded mutator; the collector must still stop the mutator
769  ("stop the world") before beginning reachability analysis.  Stopping a
770  multithreaded mutator is a complicated problem.  It generally requires highly
771  platform-specific code in the runtime, and the production of carefully
772  designed machine code at safe points.
773
774Concurrent
775  In this technique, the mutator and the collector run concurrently, with the
776  goal of eliminating pause times.  In a *cooperative* collector, the mutator
777  further aids with collection should a pause occur, allowing collection to take
778  advantage of multiprocessor hosts.  The "stop the world" problem of threaded
779  collectors is generally still present to a limited extent.  Sophisticated
780  marking algorithms are necessary.  Read barriers may be necessary.
781
782As the matrix indicates, LLVM's garbage collection infrastructure is already
783suitable for a wide variety of collectors, but does not currently extend to
784multithreaded programs.  This will be added in the future as there is
785interest.
786
787.. _stack-map:
788
789Computing stack maps
790--------------------
791
792LLVM automatically computes a stack map.  One of the most important features
793of a ``GCStrategy`` is to compile this information into the executable in
794the binary representation expected by the runtime library.
795
796The stack map consists of the location and identity of each GC root in the
797each function in the module.  For each root:
798
799* ``RootNum``: The index of the root.
800
801* ``StackOffset``: The offset of the object relative to the frame pointer.
802
803* ``RootMetadata``: The value passed as the ``%metadata`` parameter to the
804  ``@llvm.gcroot`` intrinsic.
805
806Also, for the function as a whole:
807
808* ``getFrameSize()``: The overall size of the function's initial stack frame,
809   not accounting for any dynamic allocation.
810
811* ``roots_size()``: The count of roots in the function.
812
813To access the stack map, use ``GCFunctionMetadata::roots_begin()`` and
814-``end()`` from the :ref:`GCMetadataPrinter <assembly>`:
815
816.. code-block:: c++
817
818  for (iterator I = begin(), E = end(); I != E; ++I) {
819    GCFunctionInfo *FI = *I;
820    unsigned FrameSize = FI->getFrameSize();
821    size_t RootCount = FI->roots_size();
822
823    for (GCFunctionInfo::roots_iterator RI = FI->roots_begin(),
824                                        RE = FI->roots_end();
825                                        RI != RE; ++RI) {
826      int RootNum = RI->Num;
827      int RootStackOffset = RI->StackOffset;
828      Constant *RootMetadata = RI->Metadata;
829    }
830  }
831
832If the ``llvm.gcroot`` intrinsic is eliminated before code generation by a
833custom lowering pass, LLVM will compute an empty stack map.  This may be useful
834for collector plugins which implement reference counting or a shadow stack.
835
836.. _init-roots:
837
838Initializing roots to null: ``InitRoots``
839-----------------------------------------
840
841.. code-block:: c++
842
843  MyGC::MyGC() {
844    InitRoots = true;
845  }
846
847When set, LLVM will automatically initialize each root to ``null`` upon entry to
848the function.  This prevents the GC's sweep phase from visiting uninitialized
849pointers, which will almost certainly cause it to crash.  This initialization
850occurs before custom lowering, so the two may be used together.
851
852Since LLVM does not yet compute liveness information, there is no means of
853distinguishing an uninitialized stack root from an initialized one.  Therefore,
854this feature should be used by all GC plugins.  It is enabled by default.
855
856Custom lowering of intrinsics: ``CustomRoots``, ``CustomReadBarriers``, and ``CustomWriteBarriers``
857---------------------------------------------------------------------------------------------------
858
859For GCs which use barriers or unusual treatment of stack roots, these
860flags allow the collector to perform arbitrary transformations of the
861LLVM IR:
862
863.. code-block:: c++
864
865  class MyGC : public GCStrategy {
866  public:
867    MyGC() {
868      CustomRoots = true;
869      CustomReadBarriers = true;
870      CustomWriteBarriers = true;
871    }
872  };
873
874If any of these flags are set, LLVM suppresses its default lowering for
875the corresponding intrinsics.  Instead, you must provide a custom Pass
876which lowers the intrinsics as desired.  If you have opted in to custom
877lowering of a particular intrinsic your pass **must** eliminate all
878instances of the corresponding intrinsic in functions which opt in to
879your GC.  The best example of such a pass is the ShadowStackGC and it's
880ShadowStackGCLowering pass.
881
882There is currently no way to register such a custom lowering pass
883without building a custom copy of LLVM.
884
885.. _safe-points:
886
887Generating safe points: ``NeededSafePoints``
888--------------------------------------------
889
890LLVM can compute four kinds of safe points:
891
892.. code-block:: c++
893
894  namespace GC {
895    /// PointKind - The type of a collector-safe point.
896    ///
897    enum PointKind {
898      Loop,    //< Instr is a loop (backwards branch).
899      Return,  //< Instr is a return instruction.
900      PreCall, //< Instr is a call instruction.
901      PostCall //< Instr is the return address of a call.
902    };
903  }
904
905A collector can request any combination of the four by setting the
906``NeededSafePoints`` mask:
907
908.. code-block:: c++
909
910  MyGC::MyGC()  {
911    NeededSafePoints = 1 << GC::Loop
912                     | 1 << GC::Return
913                     | 1 << GC::PreCall
914                     | 1 << GC::PostCall;
915  }
916
917It can then use the following routines to access safe points.
918
919.. code-block:: c++
920
921  for (iterator I = begin(), E = end(); I != E; ++I) {
922    GCFunctionInfo *MD = *I;
923    size_t PointCount = MD->size();
924
925    for (GCFunctionInfo::iterator PI = MD->begin(),
926                                  PE = MD->end(); PI != PE; ++PI) {
927      GC::PointKind PointKind = PI->Kind;
928      unsigned PointNum = PI->Num;
929    }
930  }
931
932Almost every collector requires ``PostCall`` safe points, since these correspond
933to the moments when the function is suspended during a call to a subroutine.
934
935Threaded programs generally require ``Loop`` safe points to guarantee that the
936application will reach a safe point within a bounded amount of time, even if it
937is executing a long-running loop which contains no function calls.
938
939Threaded collectors may also require ``Return`` and ``PreCall`` safe points to
940implement "stop the world" techniques using self-modifying code, where it is
941important that the program not exit the function without reaching a safe point
942(because only the topmost function has been patched).
943
944.. _assembly:
945
946Emitting assembly code: ``GCMetadataPrinter``
947---------------------------------------------
948
949LLVM allows a plugin to print arbitrary assembly code before and after the rest
950of a module's assembly code.  At the end of the module, the GC can compile the
951LLVM stack map into assembly code. (At the beginning, this information is not
952yet computed.)
953
954Since AsmWriter and CodeGen are separate components of LLVM, a separate abstract
955base class and registry is provided for printing assembly code, the
956``GCMetadaPrinter`` and ``GCMetadataPrinterRegistry``.  The AsmWriter will look
957for such a subclass if the ``GCStrategy`` sets ``UsesMetadata``:
958
959.. code-block:: c++
960
961  MyGC::MyGC() {
962    UsesMetadata = true;
963  }
964
965This separation allows JIT-only clients to be smaller.
966
967Note that LLVM does not currently have analogous APIs to support code generation
968in the JIT, nor using the object writers.
969
970.. code-block:: c++
971
972  // lib/MyGC/MyGCPrinter.cpp - Example LLVM GC printer
973
974  #include "llvm/CodeGen/GCMetadataPrinter.h"
975  #include "llvm/Support/Compiler.h"
976
977  using namespace llvm;
978
979  namespace {
980    class LLVM_LIBRARY_VISIBILITY MyGCPrinter : public GCMetadataPrinter {
981    public:
982      virtual void beginAssembly(AsmPrinter &AP);
983
984      virtual void finishAssembly(AsmPrinter &AP);
985    };
986
987    GCMetadataPrinterRegistry::Add<MyGCPrinter>
988    X("mygc", "My bespoke garbage collector.");
989  }
990
991The collector should use ``AsmPrinter`` to print portable assembly code.  The
992collector itself contains the stack map for the entire module, and may access
993the ``GCFunctionInfo`` using its own ``begin()`` and ``end()`` methods.  Here's
994a realistic example:
995
996.. code-block:: c++
997
998  #include "llvm/CodeGen/AsmPrinter.h"
999  #include "llvm/IR/Function.h"
1000  #include "llvm/IR/DataLayout.h"
1001  #include "llvm/Target/TargetAsmInfo.h"
1002  #include "llvm/Target/TargetMachine.h"
1003
1004  void MyGCPrinter::beginAssembly(AsmPrinter &AP) {
1005    // Nothing to do.
1006  }
1007
1008  void MyGCPrinter::finishAssembly(AsmPrinter &AP) {
1009    MCStreamer &OS = AP.OutStreamer;
1010    unsigned IntPtrSize = AP.TM.getSubtargetImpl()->getDataLayout()->getPointerSize();
1011
1012    // Put this in the data section.
1013    OS.SwitchSection(AP.getObjFileLowering().getDataSection());
1014
1015    // For each function...
1016    for (iterator FI = begin(), FE = end(); FI != FE; ++FI) {
1017      GCFunctionInfo &MD = **FI;
1018
1019      // A compact GC layout. Emit this data structure:
1020      //
1021      // struct {
1022      //   int32_t PointCount;
1023      //   void *SafePointAddress[PointCount];
1024      //   int32_t StackFrameSize; // in words
1025      //   int32_t StackArity;
1026      //   int32_t LiveCount;
1027      //   int32_t LiveOffsets[LiveCount];
1028      // } __gcmap_<FUNCTIONNAME>;
1029
1030      // Align to address width.
1031      AP.EmitAlignment(IntPtrSize == 4 ? 2 : 3);
1032
1033      // Emit PointCount.
1034      OS.AddComment("safe point count");
1035      AP.EmitInt32(MD.size());
1036
1037      // And each safe point...
1038      for (GCFunctionInfo::iterator PI = MD.begin(),
1039                                    PE = MD.end(); PI != PE; ++PI) {
1040        // Emit the address of the safe point.
1041        OS.AddComment("safe point address");
1042        MCSymbol *Label = PI->Label;
1043        AP.EmitLabelPlusOffset(Label/*Hi*/, 0/*Offset*/, 4/*Size*/);
1044      }
1045
1046      // Stack information never change in safe points! Only print info from the
1047      // first call-site.
1048      GCFunctionInfo::iterator PI = MD.begin();
1049
1050      // Emit the stack frame size.
1051      OS.AddComment("stack frame size (in words)");
1052      AP.EmitInt32(MD.getFrameSize() / IntPtrSize);
1053
1054      // Emit stack arity, i.e. the number of stacked arguments.
1055      unsigned RegisteredArgs = IntPtrSize == 4 ? 5 : 6;
1056      unsigned StackArity = MD.getFunction().arg_size() > RegisteredArgs ?
1057                            MD.getFunction().arg_size() - RegisteredArgs : 0;
1058      OS.AddComment("stack arity");
1059      AP.EmitInt32(StackArity);
1060
1061      // Emit the number of live roots in the function.
1062      OS.AddComment("live root count");
1063      AP.EmitInt32(MD.live_size(PI));
1064
1065      // And for each live root...
1066      for (GCFunctionInfo::live_iterator LI = MD.live_begin(PI),
1067                                         LE = MD.live_end(PI);
1068                                         LI != LE; ++LI) {
1069        // Emit live root's offset within the stack frame.
1070        OS.AddComment("stack index (offset / wordsize)");
1071        AP.EmitInt32(LI->StackOffset);
1072      }
1073    }
1074  }
1075
1076References
1077==========
1078
1079.. _appel89:
1080
1081[Appel89] Runtime Tags Aren't Necessary. Andrew W. Appel. Lisp and Symbolic
1082Computation 19(7):703-705, July 1989.
1083
1084.. _goldberg91:
1085
1086[Goldberg91] Tag-free garbage collection for strongly typed programming
1087languages. Benjamin Goldberg. ACM SIGPLAN PLDI'91.
1088
1089.. _tolmach94:
1090
1091[Tolmach94] Tag-free garbage collection using explicit type parameters. Andrew
1092Tolmach. Proceedings of the 1994 ACM conference on LISP and functional
1093programming.
1094
1095.. _henderson02:
1096
1097[Henderson2002] `Accurate Garbage Collection in an Uncooperative Environment
1098<http://citeseer.ist.psu.edu/henderson02accurate.html>`__
1099