docs/GarbageCollection.html

<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
                      "http://www.w3.org/TR/html4/strict.dtd">
<html>
<head>
  <meta http-equiv="Content-Type" Content="text/html; charset=UTF-8" >
  <title>Accurate Garbage Collection with LLVM</title>
  <link rel="stylesheet" href="llvm.css" type="text/css">
  <style type="text/css">
    .rowhead { text-align: left; background: inherit; }
    .indent { padding-left: 1em; }
    .optl { color: #BFBFBF; }
  </style>
</head>
<body>

<div class="doc_title">
  Accurate Garbage Collection with LLVM
</div>

<ol>
  <li><a href="#introduction">Introduction</a>
    <ul>
    <li><a href="#feature">Goals and non-goals</a></li>
    </ul>
  </li>

  <li><a href="#quickstart">Getting started</a>
    <ul>
    <li><a href="#quickstart-compiler">In your compiler</a></li>
    <li><a href="#quickstart-runtime">In your runtime library</a></li>
    <li><a href="#shadow-stack">About the shadow stack</a></li>
    </ul>
  </li>

  <li><a href="#core">Core support</a>
    <ul>
    <li><a href="#gcattr">Specifying GC code generation:
      <tt>gc "..."</tt></a></li>
    <li><a href="#gcroot">Identifying GC roots on the stack:
      <tt>llvm.gcroot</tt></a></li>
    <li><a href="#barriers">Reading and writing references in the heap</a>
      <ul>
      <li><a href="#gcwrite">Write barrier: <tt>llvm.gcwrite</tt></a></li>
      <li><a href="#gcread">Read barrier: <tt>llvm.gcread</tt></a></li>
      </ul>
    </li>
    </ul>
  </li>
  
  <li><a href="#plugin">Compiler plugin interface</a>
    <ul>
    <li><a href="#collector-algos">Overview of available features</a></li>
    <li><a href="#stack-map">Computing stack maps</a></li>
    <li><a href="#init-roots">Initializing roots to null:
      <tt>InitRoots</tt></a></li>
    <li><a href="#custom">Custom lowering of intrinsics: <tt>CustomRoots</tt>, 
      <tt>CustomReadBarriers</tt>, and <tt>CustomWriteBarriers</tt></a></li>
    <li><a href="#safe-points">Generating safe points:
      <tt>NeededSafePoints</tt></a></li>
    <li><a href="#assembly">Emitting assembly code:
      <tt>GCMetadataPrinter</tt></a></li>
    </ul>
  </li>

  <li><a href="#runtime-impl">Implementing a collector runtime</a>
    <ul>
      <li><a href="#gcdescriptors">Tracing GC pointers from heap
      objects</a></li>
    </ul>
  </li>
  
  <li><a href="#references">References</a></li>
  
</ol>

<div class="doc_author">
  <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a> and
     Gordon Henriksen</p>
</div>

<!-- *********************************************************************** -->
<div class="doc_section">
  <a name="introduction">Introduction</a>
</div>
<!-- *********************************************************************** -->

<div class="doc_text">

<p>Garbage collection is a widely used technique that frees the programmer from
having to know the lifetimes of heap objects, making software easier to produce
and maintain. Many programming languages rely on garbage collection for
automatic memory management. There are two primary forms of garbage collection:
conservative and accurate.</p>

<p>Conservative garbage collection often does not require any special support
from either the language or the compiler: it can handle non-type-safe
programming languages (such as C/C++) and does not require any special
information from the compiler. The
<a href="http://www.hpl.hp.com/personal/Hans_Boehm/gc/">Boehm collector</a> is
an example of a state-of-the-art conservative collector.</p>

<p>Accurate garbage collection requires the ability to identify all pointers in
the program at run-time (which requires that the source-language be type-safe in
most cases). Identifying pointers at run-time requires compiler support to
locate all places that hold live pointer variables at run-time, including the
<a href="#gcroot">processor stack and registers</a>.</p>

<p>Conservative garbage collection is attractive because it does not require any
special compiler support, but it does have problems. In particular, because the
conservative garbage collector cannot <i>know</i> that a particular word in the
machine is a pointer, it cannot move live objects in the heap (preventing the
use of compacting and generational GC algorithms) and it can occasionally suffer
from memory leaks due to integer values that happen to point to objects in the
program. In addition, some aggressive compiler transformations can break
conservative garbage collectors (though these seem rare in practice).</p>

<p>Accurate garbage collectors do not suffer from any of these problems, but
they can suffer from degraded scalar optimization of the program. In particular,
because the runtime must be able to identify and update all pointers active in
the program, some optimizations are less effective. In practice, however, the
locality and performance benefits of using aggressive garbage collection
techniques dominates any low-level losses.</p>

<p>This document describes the mechanisms and interfaces provided by LLVM to
support accurate garbage collection.</p>

</div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="feature">Goals and non-goals</a>
</div>

<div class="doc_text">

<p>LLVM's intermediate representation provides <a href="#intrinsics">garbage
collection intrinsics</a> that offer support for a broad class of
collector models. For instance, the intrinsics permit:</p>

<ul>
  <li>semi-space collectors</li>
  <li>mark-sweep collectors</li>
  <li>generational collectors</li>
  <li>reference counting</li>
  <li>incremental collectors</li>
  <li>concurrent collectors</li>
  <li>cooperative collectors</li>
</ul>

<p>We hope that the primitive support built into the LLVM IR is sufficient to
support a broad class of garbage collected languages including Scheme, ML, Java,
C#, Perl, Python, Lua, Ruby, other scripting languages, and more.</p>

<p>However, LLVM does not itself provide a garbage collector&#151;this should
be part of your language's runtime library. LLVM provides a framework for
compile time <a href="#plugin">code generation plugins</a>. The role of these
plugins is to generate code and data structures which conforms to the <em>binary
interface</em> specified by the <em>runtime library</em>. This is similar to the
relationship between LLVM and DWARF debugging info, for example. The
difference primarily lies in the lack of an established standard in the domain
of garbage collection&#151;thus the plugins.</p>

<p>The aspects of the binary interface with which LLVM's GC support is
concerned are:</p>

<ul>
  <li>Creation of GC-safe points within code where collection is allowed to
      execute safely.</li>
  <li>Computation of the stack map. For each safe point in the code, object
      references within the stack frame must be identified so that the
      collector may traverse and perhaps update them.</li>
  <li>Write barriers when storing object references to the heap. These are
      commonly used to optimize incremental scans in generational
      collectors.</li>
  <li>Emission of read barriers when loading object references. These are
      useful for interoperating with concurrent collectors.</li>
</ul>

<p>There are additional areas that LLVM does not directly address:</p>

<ul>
  <li>Registration of global roots with the runtime.</li>
  <li>Registration of stack map entries with the runtime.</li>
  <li>The functions used by the program to allocate memory, trigger a
      collection, etc.</li>
  <li>Computation or compilation of type maps, or registration of them with
      the runtime. These are used to crawl the heap for object
      references.</li>
</ul>

<p>In general, LLVM's support for GC does not include features which can be
adequately addressed with other features of the IR and does not specify a
particular binary interface. On the plus side, this means that you should be
able to integrate LLVM with an existing runtime. On the other hand, it leaves
a lot of work for the developer of a novel language. However, it's easy to get
started quickly and scale up to a more sophisticated implementation as your
compiler matures.</p>

</div>

<!-- *********************************************************************** -->
<div class="doc_section">
  <a name="quickstart">Getting started</a>
</div>
<!-- *********************************************************************** -->

<div class="doc_text">

<p>Using a GC with LLVM implies many things, for example:</p>

<ul>
  <li>Write a runtime library or find an existing one which implements a GC
      heap.<ol>
    <li>Implement a memory allocator.</li>
    <li>Design a binary interface for the stack map, used to identify
        references within a stack frame on the machine stack.*</li>
    <li>Implement a stack crawler to discover functions on the call stack.*</li>
    <li>Implement a registry for global roots.</li>
    <li>Design a binary interface for type maps, used to identify references
        within heap objects.</li>
    <li>Implement a collection routine bringing together all of the above.</li>
  </ol></li>
  <li>Emit compatible code from your compiler.<ul>
    <li>Initialization in the main function.</li>
    <li>Use the <tt>gc "..."</tt> attribute to enable GC code generation
        (or <tt>F.setGC("...")</tt>).</li>
    <li>Use <tt>@llvm.gcroot</tt> to mark stack roots.</li>
    <li>Use <tt>@llvm.gcread</tt> and/or <tt>@llvm.gcwrite</tt> to
        manipulate GC references, if necessary.</li>
    <li>Allocate memory using the GC allocation routine provided by the
        runtime library.</li>
    <li>Generate type maps according to your runtime's binary interface.</li>
  </ul></li>
  <li>Write a compiler plugin to interface LLVM with the runtime library.*<ul>
    <li>Lower <tt>@llvm.gcread</tt> and <tt>@llvm.gcwrite</tt> to appropriate
        code sequences.*</li>
    <li>Compile LLVM's stack map to the binary form expected by the
        runtime.</li>
  </ul></li>
  <li>Load the plugin into the compiler. Use <tt>llc -load</tt> or link the
      plugin statically with your language's compiler.*</li>
  <li>Link program executables with the runtime.</li>
</ul>

<p>To help with several of these tasks (those indicated with a *), LLVM
includes a highly portable, built-in ShadowStack code generator. It is compiled
into <tt>llc</tt> and works even with the interpreter and C backends.</p>

</div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="quickstart-compiler">In your compiler</a>
</div>

<div class="doc_text">

<p>To turn the shadow stack on for your functions, first call:</p>

<div class="doc_code"><pre
>F.setGC("shadow-stack");</pre></div>

<p>for each function your compiler emits. Since the shadow stack is built into
LLVM, you do not need to load a plugin.</p>

<p>Your compiler must also use <tt>@llvm.gcroot</tt> as documented.
Don't forget to create a root for each intermediate value that is generated
when evaluating an expression. In <tt>h(f(), g())</tt>, the result of
<tt>f()</tt> could easily be collected if evaluating <tt>g()</tt> triggers a
collection.</p>

<p>There's no need to use <tt>@llvm.gcread</tt> and <tt>@llvm.gcwrite</tt> over
plain <tt>load</tt> and <tt>store</tt> for now. You will need them when
switching to a more advanced GC.</p>

</div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="quickstart-runtime">In your runtime</a>
</div>

<div class="doc_text">

<p>The shadow stack doesn't imply a memory allocation algorithm. A semispace
collector or building atop <tt>malloc</tt> are great places to start, and can
be implemented with very little code.</p>

<p>When it comes time to collect, however, your runtime needs to traverse the
stack roots, and for this it needs to integrate with the shadow stack. Luckily,
doing so is very simple. (This code is heavily commented to help you
understand the data structure, but there are only 20 lines of meaningful
code.)</p>

</div>

<div class="doc_code"><pre
>/// @brief The map for a single function's stack frame. One of these is
///        compiled as constant data into the executable for each function.
/// 
/// Storage of metadata values is elided if the %metadata parameter to
/// @llvm.gcroot is null.
struct FrameMap {
  int32_t NumRoots;    //&lt; Number of roots in stack frame.
  int32_t NumMeta;     //&lt; Number of metadata entries. May be &lt; NumRoots.
  const void *Meta[0]; //&lt; Metadata for each root.
};

/// @brief A link in the dynamic shadow stack. One of these is embedded in the
///        stack frame of each function on the call stack.
struct StackEntry {
  StackEntry *Next;    //&lt; Link to next stack entry (the caller's).
  const FrameMap *Map; //&lt; Pointer to constant FrameMap.
  void *Roots[0];      //&lt; Stack roots (in-place array).
};

/// @brief The head of the singly-linked list of StackEntries. Functions push
///        and pop onto this in their prologue and epilogue.
/// 
/// Since there is only a global list, this technique is not threadsafe.
StackEntry *llvm_gc_root_chain;

/// @brief Calls Visitor(root, meta) for each GC root on the stack.
///        root and meta are exactly the values passed to
///        <tt>@llvm.gcroot</tt>.
/// 
/// Visitor could be a function to recursively mark live objects. Or it
/// might copy them to another heap or generation.
/// 
/// @param Visitor A function to invoke for every GC root on the stack.
void visitGCRoots(void (*Visitor)(void **Root, const void *Meta)) {
  for (StackEntry *R = llvm_gc_root_chain; R; R = R->Next) {
    unsigned i = 0;
    
    // For roots [0, NumMeta), the metadata pointer is in the FrameMap.
    for (unsigned e = R->Map->NumMeta; i != e; ++i)
      Visitor(&amp;R->Roots[i], R->Map->Meta[i]);
    
    // For roots [NumMeta, NumRoots), the metadata pointer is null.
    for (unsigned e = R->Map->NumRoots; i != e; ++i)
      Visitor(&amp;R->Roots[i], NULL);
  }
}</pre></div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="shadow-stack">About the shadow stack</a>
</div>

<div class="doc_text">

<p>Unlike many GC algorithms which rely on a cooperative code generator to
compile stack maps, this algorithm carefully maintains a linked list of stack
roots [<a href="#henderson02">Henderson2002</a>]. This so-called "shadow stack"
mirrors the machine stack. Maintaining this data structure is slower than using
a stack map compiled into the executable as constant data, but has a significant
portability advantage because it requires no special support from the target
code generator, and does not require tricky platform-specific code to crawl
the machine stack.</p>

<p>The tradeoff for this simplicity and portability is:</p>

<ul>
  <li>High overhead per function call.</li>
  <li>Not thread-safe.</li>
</ul>

<p>Still, it's an easy way to get started. After your compiler and runtime are
up and running, writing a <a href="#plugin">plugin</a> will allow you to take
advantage of <a href="#collector-algos">more advanced GC features</a> of LLVM
in order to improve performance.</p>

</div>

<!-- *********************************************************************** -->
<div class="doc_section">
  <a name="core">IR features</a><a name="intrinsics"></a>
</div>
<!-- *********************************************************************** -->

<div class="doc_text">

<p>This section describes the garbage collection facilities provided by the
<a href="LangRef.html">LLVM intermediate representation</a>. The exact behavior
of these IR features is specified by the binary interface implemented by a
<a href="#plugin">code generation plugin</a>, not by this document.</p>

<p>These facilities are limited to those strictly necessary; they are not
intended to be a complete interface to any garbage collector. A program will
need to interface with the GC library using the facilities provided by that
program.</p>

</div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="gcattr">Specifying GC code generation: <tt>gc "..."</tt></a>
</div>

<div class="doc_code"><tt>
  define <i>ty</i> @<i>name</i>(...) <span style="text-decoration: underline">gc "<i>name</i>"</span> { ...
</tt></div>

<div class="doc_text">

<p>The <tt>gc</tt> function attribute is used to specify the desired GC style
to the compiler. Its programmatic equivalent is the <tt>setGC</tt> method of
<tt>Function</tt>.</p>

<p>Setting <tt>gc "<i>name</i>"</tt> on a function triggers a search for a
matching code generation plugin "<i>name</i>"; it is that plugin which defines
the exact nature of the code generated to support GC. If none is found, the
compiler will raise an error.</p>

<p>Specifying the GC style on a per-function basis allows LLVM to link together
programs that use different garbage collection algorithms (or none at all).</p>

</div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="gcroot">Identifying GC roots on the stack: <tt>llvm.gcroot</tt></a>
</div>

<div class="doc_code"><tt>
  void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
</tt></div>

<div class="doc_text">

<p>The <tt>llvm.gcroot</tt> intrinsic is used to inform LLVM that a stack
variable references an object on the heap and is to be tracked for garbage
collection. The exact impact on generated code is specified by a <a
href="#plugin">compiler plugin</a>.</p>

<p>A compiler which uses mem2reg to raise imperative code using <tt>alloca</tt>
into SSA form need only add a call to <tt>@llvm.gcroot</tt> for those variables
which a pointers into the GC heap.</p>

<p>It is also important to mark intermediate values with <tt>llvm.gcroot</tt>.
For example, consider <tt>h(f(), g())</tt>. Beware leaking the result of
<tt>f()</tt> in the case that <tt>g()</tt> triggers a collection.</p>

<p>The first argument <b>must</b> be a value referring to an alloca instruction
or a bitcast of an alloca. The second contains a pointer to metadata that
should be associated with the pointer, and <b>must</b> be a constant or global
value address. If your target collector uses tags, use a null pointer for
metadata.</p>

<p>The <tt>%metadata</tt> argument can be used to avoid requiring heap objects
to have 'isa' pointers or tag bits. [<a href="#appel89">Appel89</a>, <a
href="#goldberg91">Goldberg91</a>, <a href="#tolmach94">Tolmach94</a>] If
specified, its value will be tracked along with the location of the pointer in
the stack frame.</p>

<p>Consider the following fragment of Java code:</p>

<pre>
       {
         Object X;   // A null-initialized reference to an object
         ...
       }
</pre>

<p>This block (which may be located in the middle of a function or in a loop
nest), could be compiled to this LLVM code:</p>

<pre>
Entry:
   ;; In the entry block for the function, allocate the
   ;; stack space for X, which is an LLVM pointer.
   %X = alloca %Object*
   
   ;; Tell LLVM that the stack space is a stack root.
   ;; Java has type-tags on objects, so we pass null as metadata.
   %tmp = bitcast %Object** %X to i8**
   call void @llvm.gcroot(i8** %X, i8* null)
   ...

   ;; "CodeBlock" is the block corresponding to the start
   ;;  of the scope above.
CodeBlock:
   ;; Java null-initializes pointers.
   store %Object* null, %Object** %X

   ...

   ;; As the pointer goes out of scope, store a null value into
   ;; it, to indicate that the value is no longer live.
   store %Object* null, %Object** %X
   ...
</pre>

</div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="barriers">Reading and writing references in the heap</a>
</div>

<div class="doc_text">

<p>Some collectors need to be informed when the mutator (the program that needs
garbage collection) either reads a pointer from or writes a pointer to a field
of a heap object. The code fragments inserted at these points are called
<em>read barriers</em> and <em>write barriers</em>, respectively. The amount of
code that needs to be executed is usually quite small and not on the critical
path of any computation, so the overall performance impact of the barrier is
tolerable.</p>

<p>Barriers often require access to the <em>object pointer</em> rather than the
<em>derived pointer</em> (which is a pointer to the field within the
object). Accordingly, these intrinsics take both pointers as separate arguments
for completeness. In this snippet, <tt>%object</tt> is the object pointer, and 
<tt>%derived</tt> is the derived pointer:</p>

<blockquote><pre>
    ;; An array type.
    %class.Array = type { %class.Object, i32, [0 x %class.Object*] }
    ...

    ;; Load the object pointer from a gcroot.
    %object = load %class.Array** %object_addr

    ;; Compute the derived pointer.
    %derived = getelementptr %object, i32 0, i32 2, i32 %n</pre></blockquote>

<p>LLVM does not enforce this relationship between the object and derived
pointer (although a <a href="#plugin">plugin</a> might). However, it would be
an unusual collector that violated it.</p>

<p>The use of these intrinsics is naturally optional if the target GC does
require the corresponding barrier. Such a GC plugin will replace the intrinsic
calls with the corresponding <tt>load</tt> or <tt>store</tt> instruction if they
are used.</p>

</div>

<!-- ======================================================================= -->
<div class="doc_subsubsection">
  <a name="gcwrite">Write barrier: <tt>llvm.gcwrite</tt></a>
</div>

<div class="doc_code"><tt>
void @llvm.gcwrite(i8* %value, i8* %object, i8** %derived)
</tt></div>

<div class="doc_text">

<p>For write barriers, LLVM provides the <tt>llvm.gcwrite</tt> intrinsic
function. It has exactly the same semantics as a non-volatile <tt>store</tt> to
the derived pointer (the third argument). The exact code generated is specified
by a <a href="#plugin">compiler plugin</a>.</p>

<p>Many important algorithms require write barriers, including generational
and concurrent collectors. Additionally, write barriers could be used to
implement reference counting.</p>

</div>

<!-- ======================================================================= -->
<div class="doc_subsubsection">
  <a name="gcread">Read barrier: <tt>llvm.gcread</tt></a>
</div>

<div class="doc_code"><tt>
i8* @llvm.gcread(i8* %object, i8** %derived)<br>
</tt></div>

<div class="doc_text">

<p>For read barriers, LLVM provides the <tt>llvm.gcread</tt> intrinsic function.
It has exactly the same semantics as a non-volatile <tt>load</tt> from the
derived pointer (the second argument). The exact code generated is specified by
a <a href="#plugin">compiler plugin</a>.</p>

<p>Read barriers are needed by fewer algorithms than write barriers, and may
have a greater performance impact since pointer reads are more frequent than
writes.</p>

</div>

<!-- *********************************************************************** -->
<div class="doc_section">
  <a name="plugin">Implementing a collector plugin</a>
</div>
<!-- *********************************************************************** -->

<div class="doc_text">

<p>User code specifies which GC code generation to use with the <tt>gc</tt>
function attribute or, equivalently, with the <tt>setGC</tt> method of
<tt>Function</tt>.</p>

<p>To implement a GC plugin, it is necessary to subclass
<tt>llvm::GCStrategy</tt>, which can be accomplished in a few lines of
boilerplate code. LLVM's infrastructure provides access to several important
algorithms. For an uncontroversial collector, all that remains may be to
compile LLVM's computed stack map to assembly code (using the binary
representation expected by the runtime library). This can be accomplished in
about 100 lines of code.</p>

<p>This is not the appropriate place to implement a garbage collected heap or a
garbage collector itself. That code should exist in the language's runtime
library. The compiler plugin is responsible for generating code which
conforms to the binary interface defined by library, most essentially the
<a href="#stack-map">stack map</a>.</p>

<p>To subclass <tt>llvm::GCStrategy</tt> and register it with the compiler:</p>

<blockquote><pre>// lib/MyGC/MyGC.cpp - Example LLVM GC plugin

#include "llvm/CodeGen/GCStrategy.h"
#include "llvm/CodeGen/GCMetadata.h"
#include "llvm/Support/Compiler.h"

using namespace llvm;

namespace {
  class VISIBILITY_HIDDEN MyGC : public GCStrategy {
  public:
    MyGC() {}
  };
  
  GCRegistry::Add&lt;MyGC&gt;
  X("mygc", "My bespoke garbage collector.");
}</pre></blockquote>

<p>This boilerplate collector does nothing. More specifically:</p>

<ul>
  <li><tt>llvm.gcread</tt> calls are replaced with the corresponding
      <tt>load</tt> instruction.</li>
  <li><tt>llvm.gcwrite</tt> calls are replaced with the corresponding
      <tt>store</tt> instruction.</li>
  <li>No safe points are added to the code.</li>
  <li>The stack map is not compiled into the executable.</li>
</ul>

<p>Using the LLVM makefiles (like the <a
href="http://llvm.org/viewvc/llvm-project/llvm/trunk/projects/sample/">sample
project</a>), this code can be compiled as a plugin using a simple
makefile:</p>

<blockquote><pre
># lib/MyGC/Makefile

LEVEL := ../..
LIBRARYNAME = <var>MyGC</var>
LOADABLE_MODULE = 1

include $(LEVEL)/Makefile.common</pre></blockquote>

<p>Once the plugin is compiled, code using it may be compiled using <tt>llc
-load=<var>MyGC.so</var></tt> (though <var>MyGC.so</var> may have some other
platform-specific extension):</p>

<blockquote><pre
>$ cat sample.ll
define void @f() gc "mygc" {
entry:
        ret void
}
$ llvm-as &lt; sample.ll | llc -load=MyGC.so</pre></blockquote>

<p>It is also possible to statically link the collector plugin into tools, such
as a language-specific compiler front-end.</p>

</div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="collector-algos">Overview of available features</a>
</div>

<div class="doc_text">

<p><tt>GCStrategy</tt> provides a range of features through which a plugin
may do useful work. Some of these are callbacks, some are algorithms that can
be enabled, disabled, or customized. This matrix summarizes the supported (and
planned) features and correlates them with the collection techniques which
typically require them.</p>

<table>
  <tr>
    <th>Algorithm</th>
    <th>Done</th>
    <th>shadow stack</th>
    <th>refcount</th>
    <th>mark-sweep</th>
    <th>copying</th>
    <th>incremental</th>
    <th>threaded</th>
    <th>concurrent</th>
  </tr>
  <tr>
    <th class="rowhead"><a href="#stack-map">stack map</a></th>
    <td>&#10004;</td>
    <td></td>
    <td></td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
  </tr>
  <tr>
    <th class="rowhead"><a href="#init-roots">initialize roots</a></th>
    <td>&#10004;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
  </tr>
  <tr class="doc_warning">
    <th class="rowhead">derived pointers</th>
    <td>NO</td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td>&#10008;*</td>
    <td>&#10008;*</td>
  </tr>
  <tr>
    <th class="rowhead"><em><a href="#custom">custom lowering</a></em></th>
    <td>&#10004;</td>
    <th></th>
    <th></th>
    <th></th>
    <th></th>
    <th></th>
    <th></th>
    <th></th>
  </tr>
  <tr>
    <th class="rowhead indent">gcroot</th>
    <td>&#10004;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
  </tr>
  <tr>
    <th class="rowhead indent">gcwrite</th>
    <td>&#10004;</td>
    <td></td>
    <td>&#10008;</td>
    <td></td>
    <td></td>
    <td>&#10008;</td>
    <td></td>
    <td>&#10008;</td>
  </tr>
  <tr>
    <th class="rowhead indent">gcread</th>
    <td>&#10004;</td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td>&#10008;</td>
  </tr>
  <tr>
    <th class="rowhead"><em><a href="#safe-points">safe points</a></em></th>
    <td></td>
    <th></th>
    <th></th>
    <th></th>
    <th></th>
    <th></th>
    <th></th>
    <th></th>
  </tr>
  <tr>
    <th class="rowhead indent">in calls</th>
    <td>&#10004;</td>
    <td></td>
    <td></td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
  </tr>
  <tr>
    <th class="rowhead indent">before calls</th>
    <td>&#10004;</td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td>&#10008;</td>
    <td>&#10008;</td>
  </tr>
  <tr class="doc_warning">
    <th class="rowhead indent">for loops</th>
    <td>NO</td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td>&#10008;</td>
    <td>&#10008;</td>
  </tr>
  <tr>
    <th class="rowhead indent">before escape</th>
    <td>&#10004;</td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td>&#10008;</td>
    <td>&#10008;</td>
  </tr>
  <tr class="doc_warning">
    <th class="rowhead">emit code at safe points</th>
    <td>NO</td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td></td>
    <td>&#10008;</td>
    <td>&#10008;</td>
  </tr>
  <tr>
    <th class="rowhead"><em>output</em></th>
    <td></td>
    <th></th>
    <th></th>
    <th></th>
    <th></th>
    <th></th>
    <th></th>
    <th></th>
  </tr>
  <tr>
    <th class="rowhead indent"><a href="#assembly">assembly</a></th>
    <td>&#10004;</td>
    <td></td>
    <td></td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
    <td>&#10008;</td>
  </tr>
  <tr class="doc_warning">
    <th class="rowhead indent">JIT</th>
    <td>NO</td>
    <td></td>
    <td></td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
  </tr>
  <tr class="doc_warning">
    <th class="rowhead indent">obj</th>
    <td>NO</td>
    <td></td>
    <td></td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
  </tr>
  <tr class="doc_warning">
    <th class="rowhead">live analysis</th>
    <td>NO</td>
    <td></td>
    <td></td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
  </tr>
  <tr class="doc_warning">
    <th class="rowhead">register map</th>
    <td>NO</td>
    <td></td>
    <td></td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
    <td class="optl">&#10008;</td>
  </tr>
  <tr>
    <td colspan="10">
      <div><span class="doc_warning">*</span> Derived pointers only pose a
           hazard to copying collectors.</div>
      <div><span class="optl">&#10008;</span> in gray denotes a feature which
           could be utilized if available.</div>
    </td>
  </tr>
</table>

<p>To be clear, the collection techniques above are defined as:</p>

<dl>
  <dt>Shadow Stack</dt>
  <dd>The mutator carefully maintains a linked list of stack roots.</dd>
  <dt>Reference Counting</dt>
  <dd>The mutator maintains a reference count for each object and frees an
      object when its count falls to zero.</dd>
  <dt>Mark-Sweep</dt>
  <dd>When the heap is exhausted, the collector marks reachable objects starting
      from the roots, then deallocates unreachable objects in a sweep
      phase.</dd>
  <dt>Copying</dt>
  <dd>As reachability analysis proceeds, the collector copies objects from one
      heap area to another, compacting them in the process. Copying collectors
      enable highly efficient "bump pointer" allocation and can improve locality
      of reference.</dd>
  <dt>Incremental</dt>
  <dd>(Including generational collectors.) Incremental collectors generally have
      all the properties of a copying collector (regardless of whether the
      mature heap is compacting), but bring the added complexity of requiring
      write barriers.</dd>
  <dt>Threaded</dt>
  <dd>Denotes a multithreaded mutator; the collector must still stop the mutator
      ("stop the world") before beginning reachability analysis. Stopping a
      multithreaded mutator is a complicated problem. It generally requires
      highly platform specific code in the runtime, and the production of
      carefully designed machine code at safe points.</dd>
  <dt>Concurrent</dt>
  <dd>In this technique, the mutator and the collector run concurrently, with
      the goal of eliminating pause times. In a <em>cooperative</em> collector,
      the mutator further aids with collection should a pause occur, allowing
      collection to take advantage of multiprocessor hosts. The "stop the world"
      problem of threaded collectors is generally still present to a limited
      extent. Sophisticated marking algorithms are necessary. Read barriers may
      be necessary.</dd>
</dl>

<p>As the matrix indicates, LLVM's garbage collection infrastructure is already
suitable for a wide variety of collectors, but does not currently extend to
multithreaded programs. This will be added in the future as there is
interest.</p>

</div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="stack-map">Computing stack maps</a>
</div>

<div class="doc_text">

<p>LLVM automatically computes a stack map. One of the most important features
of a <tt>GCStrategy</tt> is to compile this information into the executable in
the binary representation expected by the runtime library.</p>

<p>The stack map consists of the location and identity of each GC root in the
each function in the module. For each root:</p>

<ul>
  <li><tt>RootNum</tt>: The index of the root.</li>
  <li><tt>StackOffset</tt>: The offset of the object relative to the frame
      pointer.</li>
  <li><tt>RootMetadata</tt>: The value passed as the <tt>%metadata</tt>
      parameter to the <a href="#gcroot"><tt>@llvm.gcroot</tt></a> intrinsic.</li>
</ul>

<p>Also, for the function as a whole:</p>

<ul>
  <li><tt>getFrameSize()</tt>: The overall size of the function's initial
      stack frame, not accounting for any dynamic allocation.</li>
  <li><tt>roots_size()</tt>: The count of roots in the function.</li>
</ul>

<p>To access the stack map, use <tt>GCFunctionMetadata::roots_begin()</tt> and
-<tt>end()</tt> from the <tt><a
href="#assembly">GCMetadataPrinter</a></tt>:</p>

<blockquote><pre
>for (iterator I = begin(), E = end(); I != E; ++I) {
  GCFunctionInfo *FI = *I;
  unsigned FrameSize = FI-&gt;getFrameSize();
  size_t RootCount = FI-&gt;roots_size();

  for (GCFunctionInfo::roots_iterator RI = FI-&gt;roots_begin(),
                                      RE = FI-&gt;roots_end();
                                      RI != RE; ++RI) {
    int RootNum = RI->Num;
    int RootStackOffset = RI->StackOffset;
    Constant *RootMetadata = RI->Metadata;
  }
}</pre></blockquote>

<p>If the <tt>llvm.gcroot</tt> intrinsic is eliminated before code generation by
a custom lowering pass, LLVM will compute an empty stack map. This may be useful
for collector plugins which implement reference counting or a shadow stack.</p>

</div>


<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="init-roots">Initializing roots to null: <tt>InitRoots</tt></a>
</div>

<div class="doc_text">

<blockquote><pre
>MyGC::MyGC() {
  InitRoots = true;
}</pre></blockquote>

<p>When set, LLVM will automatically initialize each root to <tt>null</tt> upon
entry to the function. This prevents the GC's sweep phase from visiting
uninitialized pointers, which will almost certainly cause it to crash. This
initialization occurs before custom lowering, so the two may be used
together.</p>

<p>Since LLVM does not yet compute liveness information, there is no means of
distinguishing an uninitialized stack root from an initialized one. Therefore,
this feature should be used by all GC plugins. It is enabled by default.</p>

</div>


<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="custom">Custom lowering of intrinsics: <tt>CustomRoots</tt>, 
    <tt>CustomReadBarriers</tt>, and <tt>CustomWriteBarriers</tt></a>
</div>

<div class="doc_text">

<p>For GCs which use barriers or unusual treatment of stack roots, these
flags allow the collector to perform arbitrary transformations of the LLVM
IR:</p>

<blockquote><pre
>class MyGC : public GCStrategy {
public:
  MyGC() {
    CustomRoots = true;
    CustomReadBarriers = true;
    CustomWriteBarriers = true;
  }
  
  virtual bool initializeCustomLowering(Module &amp;M);
  virtual bool performCustomLowering(Function &amp;F);
};</pre></blockquote>

<p>If any of these flags are set, then LLVM suppresses its default lowering for
the corresponding intrinsics and instead calls
<tt>performCustomLowering</tt>.</p>

<p>LLVM's default action for each intrinsic is as follows:</p>

<ul>
  <li><tt>llvm.gcroot</tt>: Leave it alone. The code generator must see it
                            or the stack map will not be computed.</li>
  <li><tt>llvm.gcread</tt>: Substitute a <tt>load</tt> instruction.</li>
  <li><tt>llvm.gcwrite</tt>: Substitute a <tt>store</tt> instruction.</li>
</ul>

<p>If <tt>CustomReadBarriers</tt> or <tt>CustomWriteBarriers</tt> are specified,
then <tt>performCustomLowering</tt> <strong>must</strong> eliminate the
corresponding barriers.</p>

<p><tt>performCustomLowering</tt> must comply with the same restrictions as <a
href="WritingAnLLVMPass.html#runOnFunction"><tt
>FunctionPass::runOnFunction</tt></a>.
Likewise, <tt>initializeCustomLowering</tt> has the same semantics as <a
href="WritingAnLLVMPass.html#doInitialization_mod"><tt
>Pass::doInitialization(Module&amp;)</tt></a>.</p>

<p>The following can be used as a template:</p>

<blockquote><pre
>#include "llvm/Module.h"
#include "llvm/IntrinsicInst.h"

bool MyGC::initializeCustomLowering(Module &amp;M) {
  return false;
}

bool MyGC::performCustomLowering(Function &amp;F) {
  bool MadeChange = false;
  
  for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
    for (BasicBlock::iterator II = BB-&gt;begin(), E = BB-&gt;end(); II != E; )
      if (IntrinsicInst *CI = dyn_cast&lt;IntrinsicInst&gt;(II++))
        if (Function *F = CI-&gt;getCalledFunction())
          switch (F-&gt;getIntrinsicID()) {
          case Intrinsic::gcwrite:
            // Handle llvm.gcwrite.
            CI-&gt;eraseFromParent();
            MadeChange = true;
            break;
          case Intrinsic::gcread:
            // Handle llvm.gcread.
            CI-&gt;eraseFromParent();
            MadeChange = true;
            break;
          case Intrinsic::gcroot:
            // Handle llvm.gcroot.
            CI-&gt;eraseFromParent();
            MadeChange = true;
            break;
          }
  
  return MadeChange;
}</pre></blockquote>

</div>


<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="safe-points">Generating safe points: <tt>NeededSafePoints</tt></a>
</div>

<div class="doc_text">

<p>LLVM can compute four kinds of safe points:</p>

<blockquote><pre
>namespace GC {
  /// PointKind - The type of a collector-safe point.
  /// 
  enum PointKind {
    Loop,    //&lt; Instr is a loop (backwards branch).
    Return,  //&lt; Instr is a return instruction.
    PreCall, //&lt; Instr is a call instruction.
    PostCall //&lt; Instr is the return address of a call.
  };
}</pre></blockquote>

<p>A collector can request any combination of the four by setting the 
<tt>NeededSafePoints</tt> mask:</p>

<blockquote><pre
>MyGC::MyGC() {
  NeededSafePoints = 1 &lt;&lt; GC::Loop
                   | 1 &lt;&lt; GC::Return
                   | 1 &lt;&lt; GC::PreCall
                   | 1 &lt;&lt; GC::PostCall;
}</pre></blockquote>

<p>It can then use the following routines to access safe points.</p>

<blockquote><pre
>for (iterator I = begin(), E = end(); I != E; ++I) {
  GCFunctionInfo *MD = *I;
  size_t PointCount = MD-&gt;size();

  for (GCFunctionInfo::iterator PI = MD-&gt;begin(),
                                PE = MD-&gt;end(); PI != PE; ++PI) {
    GC::PointKind PointKind = PI-&gt;Kind;
    unsigned PointNum = PI-&gt;Num;
  }
}
</pre></blockquote>

<p>Almost every collector requires <tt>PostCall</tt> safe points, since these
correspond to the moments when the function is suspended during a call to a
subroutine.</p>

<p>Threaded programs generally require <tt>Loop</tt> safe points to guarantee
that the application will reach a safe point within a bounded amount of time,
even if it is executing a long-running loop which contains no function
calls.</p>

<p>Threaded collectors may also require <tt>Return</tt> and <tt>PreCall</tt>
safe points to implement "stop the world" techniques using self-modifying code,
where it is important that the program not exit the function without reaching a
safe point (because only the topmost function has been patched).</p>

</div>


<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="assembly">Emitting assembly code: <tt>GCMetadataPrinter</tt></a>
</div>

<div class="doc_text">

<p>LLVM allows a plugin to print arbitrary assembly code before and after the
rest of a module's assembly code. At the end of the module, the GC can compile
the LLVM stack map into assembly code. (At the beginning, this information is not
yet computed.)</p>

<p>Since AsmWriter and CodeGen are separate components of LLVM, a separate
abstract base class and registry is provided for printing assembly code, the
<tt>GCMetadaPrinter</tt> and <tt>GCMetadataPrinterRegistry</tt>. The AsmWriter
will look for such a subclass if the <tt>GCStrategy</tt> sets
<tt>UsesMetadata</tt>:</p>

<blockquote><pre
>MyGC::MyGC() {
  UsesMetadata = true;
}</pre></blockquote>

<p>This separation allows JIT-only clients to be smaller.</p>

<p>Note that LLVM does not currently have analogous APIs to support code
generation in the JIT, nor using the object writers.</p>

<blockquote><pre
>// lib/MyGC/MyGCPrinter.cpp - Example LLVM GC printer

#include "llvm/CodeGen/GCMetadataPrinter.h"
#include "llvm/Support/Compiler.h"

using namespace llvm;

namespace {
  class VISIBILITY_HIDDEN MyGCPrinter : public GCMetadataPrinter {
  public:
    virtual void beginAssembly(std::ostream &amp;OS, AsmPrinter &amp;AP,
                               const TargetAsmInfo &amp;TAI);
  
    virtual void finishAssembly(std::ostream &amp;OS, AsmPrinter &amp;AP,
                                const TargetAsmInfo &amp;TAI);
  };
  
  GCMetadataPrinterRegistry::Add&lt;MyGCPrinter&gt;
  X("mygc", "My bespoke garbage collector.");
}</pre></blockquote>

<p>The collector should use <tt>AsmPrinter</tt> and <tt>TargetAsmInfo</tt> to
print portable assembly code to the <tt>std::ostream</tt>. The collector itself
contains the stack map for the entire module, and may access the
<tt>GCFunctionInfo</tt> using its own <tt>begin()</tt> and <tt>end()</tt>
methods. Here's a realistic example:</p>

<blockquote><pre
>#include "llvm/CodeGen/AsmPrinter.h"
#include "llvm/Function.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetAsmInfo.h"

void MyGCPrinter::beginAssembly(std::ostream &amp;OS, AsmPrinter &amp;AP,
                                const TargetAsmInfo &amp;TAI) {
  // Nothing to do.
}

void MyGCPrinter::finishAssembly(std::ostream &amp;OS, AsmPrinter &amp;AP,
                                 const TargetAsmInfo &amp;TAI) {
  // Set up for emitting addresses.
  const char *AddressDirective;
  int AddressAlignLog;
  if (AP.TM.getTargetData()->getPointerSize() == sizeof(int32_t)) {
    AddressDirective = TAI.getData32bitsDirective();
    AddressAlignLog = 2;
  } else {
    AddressDirective = TAI.getData64bitsDirective();
    AddressAlignLog = 3;
  }
  
  // Put this in the data section.
  AP.SwitchToDataSection(TAI.getDataSection());
  
  // For each function...
  for (iterator FI = begin(), FE = end(); FI != FE; ++FI) {
    GCFunctionInfo &amp;MD = **FI;
    
    // Emit this data structure:
    // 
    // struct {
    //   int32_t PointCount;
    //   struct {
    //     void *SafePointAddress;
    //     int32_t LiveCount;
    //     int32_t LiveOffsets[LiveCount];
    //   } Points[PointCount];
    // } __gcmap_&lt;FUNCTIONNAME&gt;;
    
    // Align to address width.
    AP.EmitAlignment(AddressAlignLog);
    
    // Emit the symbol by which the stack map entry can be found.
    std::string Symbol;
    Symbol += TAI.getGlobalPrefix();
    Symbol += "__gcmap_";
    Symbol += MD.getFunction().getName();
    if (const char *GlobalDirective = TAI.getGlobalDirective())
      OS &lt;&lt; GlobalDirective &lt;&lt; Symbol &lt;&lt; "\n";
    OS &lt;&lt; TAI.getGlobalPrefix() &lt;&lt; Symbol &lt;&lt; ":\n";
    
    // Emit PointCount.
    AP.EmitInt32(MD.size());
    AP.EOL("safe point count");
    
    // And each safe point...
    for (GCFunctionInfo::iterator PI = MD.begin(),
                                     PE = MD.end(); PI != PE; ++PI) {
      // Align to address width.
      AP.EmitAlignment(AddressAlignLog);
      
      // Emit the address of the safe point.
      OS &lt;&lt; AddressDirective
         &lt;&lt; TAI.getPrivateGlobalPrefix() &lt;&lt; "label" &lt;&lt; PI-&gt;Num;
      AP.EOL("safe point address");
      
      // Emit the stack frame size.
      AP.EmitInt32(MD.getFrameSize());
      AP.EOL("stack frame size");
      
      // Emit the number of live roots in the function.
      AP.EmitInt32(MD.live_size(PI));
      AP.EOL("live root count");
      
      // And for each live root...
      for (GCFunctionInfo::live_iterator LI = MD.live_begin(PI),
                                            LE = MD.live_end(PI);
                                            LI != LE; ++LI) {
        // Print its offset within the stack frame.
        AP.EmitInt32(LI-&gt;StackOffset);
        AP.EOL("stack offset");
      }
    }
  }
}
</pre></blockquote>

</div>


<!-- *********************************************************************** -->
<div class="doc_section">
  <a name="references">References</a>
</div>
<!-- *********************************************************************** -->

<div class="doc_text">

<p><a name="appel89">[Appel89]</a> Runtime Tags Aren't Necessary. Andrew
W. Appel. Lisp and Symbolic Computation 19(7):703-705, July 1989.</p>

<p><a name="goldberg91">[Goldberg91]</a> Tag-free garbage collection for
strongly typed programming languages. Benjamin Goldberg. ACM SIGPLAN
PLDI'91.</p>

<p><a name="tolmach94">[Tolmach94]</a> Tag-free garbage collection using
explicit type parameters. Andrew Tolmach. Proceedings of the 1994 ACM
conference on LISP and functional programming.</p>

<p><a name="henderson02">[Henderson2002]</a> <a
href="http://citeseer.ist.psu.edu/henderson02accurate.html">
Accurate Garbage Collection in an Uncooperative Environment</a>.
Fergus Henderson. International Symposium on Memory Management 2002.</p>

</div>


<!-- *********************************************************************** -->

<hr>
<address>
  <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
  src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
  <a href="http://validator.w3.org/check/referer"><img
  src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>

  <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
  <a href="http://llvm.org">LLVM Compiler Infrastructure</a><br>
  Last modified: $Date$
</address>

</body>
</html>