Garbage Collection with LLVM

Abstract

This document covers how to integrate LLVM into a compiler for a language whichsupports garbage collection. Note that LLVM itself does not provide agarbage collector. You must provide your own.

Quick Start

First, you should pick a collector strategy. LLVM includes a number of builtin ones, but you can also implement a loadable plugin with a custom definition.Note that the collector strategy is a description of how LLVM should generatecode such that it interacts with your collector and runtime, not a descriptionof the collector itself.

Next, mark your generated functions as using your chosen collector strategy.From c++, you can call:

F.setGC(<collector description name>);

This will produce IR like the following fragment:

define void @foo() gc "<collector description name>" { ... }

When generating LLVM IR for your functions, you will need to:

• Use @llvm.gcread and/or @llvm.gcwrite in place of standard load andstore instructions. These intrinsics are used to represent load and storebarriers. If you collector does not require such barriers, you can skipthis step.
• Use the memory allocation routines provided by your garbage collector’sruntime library.
• If your collector requires them, generate type maps according to yourruntime’s binary interface. LLVM is not involved in the process. Inparticular, the LLVM type system is not suitable for conveying suchinformation though the compiler.
• Insert any coordination code required for interacting with your collector.Many collectors require running application code to periodically check aflag and conditionally call a runtime function. This is often referred toas a safepoint poll.

You will need to identify roots (i.e. references to heap objects your collectorneeds to know about) in your generated IR, so that LLVM can encode them intoyour final stack maps. Depending on the collector strategy chosen, this isaccomplished by using either the @llvm.gcroot intrinsics or angc.statepoint relocation sequence.

Don’t forget to create a root for each intermediate value that is generated whenevaluating an expression. In h(f(), g()), the result of f() couldeasily be collected if evaluating g() triggers a collection.

Finally, you need to link your runtime library with the generated programexecutable (for a static compiler) or ensure the appropriate symbols areavailable for the runtime linker (for a JIT compiler).

Introduction

What is Garbage Collection?

Garbage collection is a widely used technique that frees the programmer fromhaving to know the lifetimes of heap objects, making software easier to produceand maintain. Many programming languages rely on garbage collection forautomatic memory management. There are two primary forms of garbage collection:conservative and accurate.

Conservative garbage collection often does not require any special support fromeither the language or the compiler: it can handle non-type-safe programminglanguages (such as C/C++) and does not require any special information from thecompiler. The Boehm collector is an example of astate-of-the-art conservative collector.

Accurate garbage collection requires the ability to identify all pointers in theprogram at run-time (which requires that the source-language be type-safe inmost cases). Identifying pointers at run-time requires compiler support tolocate all places that hold live pointer variables at run-time, including theprocessor stack and registers.

Conservative garbage collection is attractive because it does not require anyspecial compiler support, but it does have problems. In particular, because theconservative garbage collector cannot know that a particular word in themachine is a pointer, it cannot move live objects in the heap (preventing theuse of compacting and generational GC algorithms) and it can occasionally sufferfrom memory leaks due to integer values that happen to point to objects in theprogram. In addition, some aggressive compiler transformations can breakconservative garbage collectors (though these seem rare in practice).

Accurate garbage collectors do not suffer from any of these problems, but theycan suffer from degraded scalar optimization of the program. In particular,because the runtime must be able to identify and update all pointers active inthe program, some optimizations are less effective. In practice, however, thelocality and performance benefits of using aggressive garbage collectiontechniques dominates any low-level losses.

This document describes the mechanisms and interfaces provided by LLVM tosupport accurate garbage collection.

Goals and non-goals

LLVM’s intermediate representation provides garbage collection intrinsics that offer support for a broad class of collector models. Forinstance, the intrinsics permit:

• semi-space collectors
• mark-sweep collectors
• generational collectors
• incremental collectors
• concurrent collectors
• cooperative collectors
• reference counting

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

Note that LLVM does not itself provide a garbage collector — this shouldbe part of your language’s runtime library. LLVM provides a framework fordescribing the garbage collectors requirements to the compiler. In particular,LLVM provides support for generating stack maps at call sites, polling for asafepoint, and emitting load and store barriers. You can also extend LLVM -possibly through a loadable code generation plugins - togenerate code and data structures which conforms to the binary interface_specified by the _runtime library. This is similar to the relationship betweenLLVM and DWARF debugging info, for example. The difference primarily lies inthe lack of an established standard in the domain of garbage collection — thusthe need for a flexible extension mechanism.

The aspects of the binary interface with which LLVM’s GC support isconcerned are:

• Creation of GC safepoints within code where collection is allowed to executesafely.
• Computation of the stack map. For each safe point in the code, objectreferences within the stack frame must be identified so that the collector maytraverse and perhaps update them.
• Write barriers when storing object references to the heap. These are commonlyused to optimize incremental scans in generational collectors.
• Emission of read barriers when loading object references. These are usefulfor interoperating with concurrent collectors.

There are additional areas that LLVM does not directly address:

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

In general, LLVM’s support for GC does not include features which can beadequately addressed with other features of the IR and does not specify aparticular binary interface. On the plus side, this means that you should beable to integrate LLVM with an existing runtime. On the other hand, it canhave the effect of leaving a lot of work for the developer of a novellanguage. We try to mitigate this by providing built in collector strategydescriptions that can work with many common collector designs and easyextension points. If you don’t already have a specific binary interfaceyou need to support, we recommend trying to use one of these built in collectorstrategies.

LLVM IR Features

This section describes the garbage collection facilities provided by theLLVM intermediate representation. The exact behavior of theseIR features is specified by the selected GC strategy description.

Specifying GC code generation: gc "…"

define <returntype> @name(...) gc "name" { ... }

The gc function attribute is used to specify the desired GC strategy to thecompiler. Its programmatic equivalent is the setGC method of Function.

Setting gc "name" on a function triggers a search for a matching subclassof GCStrategy. Some collector strategies are built in. You can add othersusing either the loadable plugin mechanism, or by patching your copy of LLVM.It is the selected GC strategy which defines the exact nature of the codegenerated to support GC. If none is found, the compiler will raise an error.

Specifying the GC style on a per-function basis allows LLVM to link togetherprograms that use different garbage collection algorithms (or none at all).

Identifying GC roots on the stack

LLVM currently supports two different mechanisms for describing references incompiled code at safepoints. llvm.gcroot is the older mechanism;gc.statepoint has been added more recently. At the moment, you can chooseeither implementation (on a per GC strategy basis). Longerterm, we will probably either migrate away from llvm.gcroot entirely, orsubstantially merge their implementations. Note that most new developmentwork is focused on gc.statepoint.

Using gc.statepoint

This page contains detailed documentation forgc.statepoint.

Using llvm.gcwrite

void @llvm.gcroot(i8** %ptrloc, i8* %metadata)

The llvm.gcroot intrinsic is used to inform LLVM that a stack variablereferences an object on the heap and is to be tracked for garbage collection.The exact impact on generated code is specified by the Function’s selectedGC strategy. All calls to llvm.gcroot must resideinside the first basic block.

The first argument must be a value referring to an alloca instruction or abitcast of an alloca. The second contains a pointer to metadata that should beassociated with the pointer, and must be a constant or global valueaddress. If your target collector uses tags, use a null pointer for metadata.

A compiler which performs manual SSA construction must ensure that SSAvalues representing GC references are stored in to the alloca passed to therespective gcroot before every call site and reloaded after every call.A compiler which uses mem2reg to raise imperative code using alloca intoSSA form need only add a call to @llvm.gcroot for those variables whichare pointers into the GC heap.

It is also important to mark intermediate values with llvm.gcroot. Forexample, consider h(f(), g()). Beware leaking the result of f() in thecase that g() triggers a collection. Note, that stack variables must beinitialized and marked with llvm.gcroot in function’s prologue.

The %metadata argument can be used to avoid requiring heap objects to have‘isa’ pointers or tag bits. [Appel89, Goldberg91, Tolmach94] If specified,its value will be tracked along with the location of the pointer in the stackframe.

Consider the following fragment of Java code:

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

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

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** %tmp, 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
   ...

Reading and writing references in the heap

Some collectors need to be informed when the mutator (the program that needsgarbage collection) either reads a pointer from or writes a pointer to a fieldof a heap object. The code fragments inserted at these points are called readbarriers and write barriers, respectively. The amount of code that needs tobe executed is usually quite small and not on the critical path of anycomputation, so the overall performance impact of the barrier is tolerable.

Barriers often require access to the object pointer rather than the derivedpointer (which is a pointer to the field within the object). Accordingly,these intrinsics take both pointers as separate arguments for completeness. Inthis snippet, %object is the object pointer, and %derived is the derivedpointer:

;; 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

LLVM does not enforce this relationship between the object and derived pointer(although a particular collector strategy might). However, itwould be an unusual collector that violated it.

The use of these intrinsics is naturally optional if the target GC does notrequire the corresponding barrier. The GC strategy used with such a collectorshould replace the intrinsic calls with the corresponding load orstore instruction if they are used.

One known deficiency with the current design is that the barrier intrinsics donot include the size or alignment of the underlying operation performed. It iscurrently assumed that the operation is of pointer size and the alignment isassumed to be the target machine’s default alignment.

Write barrier: llvm.gcwrite

void @llvm.gcwrite(i8* %value, i8* %object, i8** %derived)

For write barriers, LLVM provides the llvm.gcwrite intrinsic function. Ithas exactly the same semantics as a non-volatile store to the derivedpointer (the third argument). The exact code generated is specified by theFunction’s selected GC strategy.

Many important algorithms require write barriers, including generational andconcurrent collectors. Additionally, write barriers could be used to implementreference counting.

Read barrier: llvm.gcread

i8* @llvm.gcread(i8* %object, i8** %derived)

For read barriers, LLVM provides the llvm.gcread intrinsic function. It hasexactly the same semantics as a non-volatile load from the derived pointer(the second argument). The exact code generated is specified by the Function’sselected GC strategy.

Read barriers are needed by fewer algorithms than write barriers, and may have agreater performance impact since pointer reads are more frequent than writes.

Built In GC Strategies

LLVM includes built in support for several varieties of garbage collectors.

The Shadow Stack GC

To use this collector strategy, mark your functions with:

F.setGC("shadow-stack");

Unlike many GC algorithms which rely on a cooperative code generator to compilestack maps, this algorithm carefully maintains a linked list of stack roots[Henderson2002]. This so-called “shadow stack” mirrors themachine stack. Maintaining this data structure is slower than using a stack mapcompiled into the executable as constant data, but has a significant portabilityadvantage because it requires no special support from the target code generator,and does not require tricky platform-specific code to crawl the machine stack.

The tradeoff for this simplicity and portability is:

• High overhead per function call.
• Not thread-safe.

Still, it’s an easy way to get started. After your compiler and runtime are upand running, writing a plugin will allow you to take advantageof more advanced GC features of LLVM in order toimprove performance.

The shadow stack doesn’t imply a memory allocation algorithm. A semispacecollector or building atop malloc are great places to start, and can beimplemented with very little code.

When it comes time to collect, however, your runtime needs to traverse the stackroots, and for this it needs to integrate with the shadow stack. Luckily, doingso is very simple. (This code is heavily commented to help you understand thedata structure, but there are only 20 lines of meaningful code.)

/// 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;    //< Number of roots in stack frame.
  int32_t NumMeta;     //< Number of metadata entries.  May be < NumRoots.
  const void *Meta[0]; //< Metadata for each root.
};
 
/// 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;    //< Link to next stack entry (the caller's).
  const FrameMap *Map; //< Pointer to constant FrameMap.
  void *Roots[0];      //< Stack roots (in-place array).
};
 
/// 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;
 
/// Calls Visitor(root, meta) for each GC root on the stack.
///        root and meta are exactly the values passed to
///        @llvm.gcroot.
///
/// 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(&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(&R->Roots[i], NULL);
  }
}

The ‘Erlang’ and ‘Ocaml’ GCs

LLVM ships with two example collectors which leverage the gcrootmechanisms. To our knowledge, these are not actually used by any languageruntime, but they do provide a reasonable starting point for someone interestedin writing an gcroot compatible GC plugin. In particular, these are theonly in tree examples of how to produce a custom binary stack map format usinga gcroot strategy.

As there names imply, the binary format produced is intended to model thatused by the Erlang and OCaml compilers respectively.

The Statepoint Example GC

F.setGC("statepoint-example");

This GC provides an example of how one might use the infrastructure providedby gc.statepoint. This example GC is compatible with thePlaceSafepoints and RewriteStatepointsForGC utility passeswhich simplify gc.statepoint sequence insertion. If you need to build acustom GC strategy around the gc.statepoints mechanisms, it is recommendedthat you use this one as a starting point.

This GC strategy does not support read or write barriers. As a result, theseintrinsics are lowered to normal loads and stores.

The stack map format generated by this GC strategy can be found in theStack Map Section using a format documented here. This format is intended to be the standardformat supported by LLVM going forward.

The CoreCLR GC

F.setGC("coreclr");

This GC leverages the gc.statepoint mechanism to support theCoreCLR runtime.

Support for this GC strategy is a work in progress. This strategy willdiffer fromstatepoint-example GC strategy incertain aspects like:

• Base-pointers of interior pointers are not explicitlytracked and reported.
• A different format is used for encoding stack maps.
• Safe-point polls are only needed before loop-back edgesand before tail-calls (not needed at function-entry).

Custom GC Strategies

If none of the built in GC strategy descriptions met your needs above, you willneed to define a custom GCStrategy and possibly, a custom LLVM pass to performlowering. Your best example of where to start defining a custom GCStrategywould be to look at one of the built in strategies.

You may be able to structure this additional code as a loadable plugin library.Loadable plugins are sufficient if all you need is to enable a differentcombination of built in functionality, but if you need to provide a customlowering pass, you will need to build a patched version of LLVM. If you thinkyou need a patched build, please ask for advice on llvm-dev. There may be aneasy way we can extend the support to make it work for your use case withoutrequiring a custom build.

Collector Requirements

You should be able to leverage any existing collector library that includes the following elements:

• A memory allocator which exposes an allocation function your compiledcode can call.
• A binary format for the stack map. A stack map describes the locationof references at a safepoint and is used by precise collectors to identifyreferences within a stack frame on the machine stack. Note that collectorswhich conservatively scan the stack don’t require such a structure.
• A stack crawler to discover functions on the call stack, and enumerate thereferences listed in the stack map for each call site.
• A mechanism for identifying references in global locations (e.g. globalvariables).
• If you collector requires them, an LLVM IR implementation of your collectorsload and store barriers. Note that since many collectors don’t requirebarriers at all, LLVM defaults to lowering such barriers to normal loadsand stores unless you arrange otherwise.

Implementing a collector plugin

User code specifies which GC code generation to use with the gc functionattribute or, equivalently, with the setGC method of Function.

To implement a GC plugin, it is necessary to subclass llvm::GCStrategy,which can be accomplished in a few lines of boilerplate code. LLVM’sinfrastructure provides access to several important algorithms. For anuncontroversial collector, all that remains may be to compile LLVM’s computedstack map to assembly code (using the binary representation expected by theruntime library). This can be accomplished in about 100 lines of code.

This is not the appropriate place to implement a garbage collected heap or agarbage collector itself. That code should exist in the language’s runtimelibrary. The compiler plugin is responsible for generating code which conformsto the binary interface defined by library, most essentially the stack map.

To subclass llvm::GCStrategy and register it with the compiler:

// 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 LLVM_LIBRARY_VISIBILITY MyGC : public GCStrategy {
  public:
    MyGC() {}
  };
 
  GCRegistry::Add<MyGC>
  X("mygc", "My bespoke garbage collector.");
}

This boilerplate collector does nothing. More specifically:

• llvm.gcread calls are replaced with the corresponding loadinstruction.
• llvm.gcwrite calls are replaced with the corresponding storeinstruction.
• No safe points are added to the code.
• The stack map is not compiled into the executable.

Using the LLVM makefiles, this codecan be compiled as a plugin using a simple makefile:

# lib/MyGC/Makefile
 
LEVEL := ../..
LIBRARYNAME = MyGC
LOADABLE_MODULE = 1
 
include $(LEVEL)/Makefile.common

Once the plugin is compiled, code using it may be compiled using llc-load=MyGC.so (though MyGC.so may have some other platform-specificextension):

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

It is also possible to statically link the collector plugin into tools, such asa language-specific compiler front-end.

Overview of available features

GCStrategy provides a range of features through which a plugin may do usefulwork. 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 typicallyrequire them.

To be clear, the collection techniques above are defined as:

• Shadow Stack
• The mutator carefully maintains a linked list of stack roots.
• Reference Counting
• The mutator maintains a reference count for each object and frees an objectwhen its count falls to zero.
• Mark-Sweep
• When the heap is exhausted, the collector marks reachable objects startingfrom the roots, then deallocates unreachable objects in a sweep phase.
• Copying
• As reachability analysis proceeds, the collector copies objects from one heaparea to another, compacting them in the process. Copying collectors enablehighly efficient “bump pointer” allocation and can improve locality ofreference.
• Incremental
• (Including generational collectors.) Incremental collectors generally have allthe properties of a copying collector (regardless of whether the mature heapis compacting), but bring the added complexity of requiring write barriers.
• Threaded
• Denotes a multithreaded mutator; the collector must still stop the mutator(“stop the world”) before beginning reachability analysis. Stopping amultithreaded mutator is a complicated problem. It generally requires highlyplatform-specific code in the runtime, and the production of carefullydesigned machine code at safe points.
• Concurrent
• In this technique, the mutator and the collector run concurrently, with thegoal of eliminating pause times. In a cooperative collector, the mutatorfurther aids with collection should a pause occur, allowing collection to takeadvantage of multiprocessor hosts. The “stop the world” problem of threadedcollectors is generally still present to a limited extent. Sophisticatedmarking algorithms are necessary. Read barriers may be necessary.

As the matrix indicates, LLVM’s garbage collection infrastructure is alreadysuitable for a wide variety of collectors, but does not currently extend tomultithreaded programs. This will be added in the future as there isinterest.

Computing stack maps

LLVM automatically computes a stack map. One of the most important featuresof a GCStrategy is to compile this information into the executable inthe binary representation expected by the runtime library.

The stack map consists of the location and identity of each GC root in theeach function in the module. For each root:

• RootNum: The index of the root.
• StackOffset: The offset of the object relative to the frame pointer.
• RootMetadata: The value passed as the %metadata parameter to the@llvm.gcroot intrinsic.

Also, for the function as a whole:

• • getFrameSize(): The overall size of the function’s initial stack frame,
  • not accounting for any dynamic allocation.
• roots_size(): The count of roots in the function.

To access the stack map, use GCFunctionMetadata::roots_begin() and-end() from the GCMetadataPrinter:

for (iterator I = begin(), E = end(); I != E; ++I) {
  GCFunctionInfo *FI = *I;
  unsigned FrameSize = FI->getFrameSize();
  size_t RootCount = FI->roots_size();
 
  for (GCFunctionInfo::roots_iterator RI = FI->roots_begin(),
                                      RE = FI->roots_end();
                                      RI != RE; ++RI) {
    int RootNum = RI->Num;
    int RootStackOffset = RI->StackOffset;
    Constant *RootMetadata = RI->Metadata;
  }
}

If the llvm.gcroot intrinsic is eliminated before code generation by acustom lowering pass, LLVM will compute an empty stack map. This may be usefulfor collector plugins which implement reference counting or a shadow stack.

Initializing roots to null

It is recommended that frontends initialize roots explicitly to avoidpotentially confusing the optimizer. This prevents the GC from visitinguninitialized pointers, which will almost certainly cause it to crash.

As a fallback, LLVM will automatically initialize each root to nullupon entry to the function. Support for this mode in code generation islargely a legacy detail to keep old collector implementations working.

Custom lowering of intrinsics

For GCs which use barriers or unusual treatment of stack roots, theimplementor is responsibly for providing a custom pass to lower theintrinsics with the desired semantics. If you have opted in to customlowering of a particular intrinsic your pass must eliminate allinstances of the corresponding intrinsic in functions which opt in toyour GC. The best example of such a pass is the ShadowStackGC and it’sShadowStackGCLowering pass.

There is currently no way to register such a custom lowering passwithout building a custom copy of LLVM.

Generating safe points

LLVM provides support for associating stackmaps with the return address ofa call. Any loop or return safepoints required by a given collector designcan be modeled via calls to runtime routines, or potentially patchable callsequences. Using gcroot, all call instructions are inferred to be possiblesafepoints and will thus have an associated stackmap.

Emitting assembly code: GCMetadataPrinter

LLVM allows a plugin to print arbitrary assembly code before and after the restof a module’s assembly code. At the end of the module, the GC can compile theLLVM stack map into assembly code. (At the beginning, this information is notyet computed.)

Since AsmWriter and CodeGen are separate components of LLVM, a separate abstractbase class and registry is provided for printing assembly code, theGCMetadaPrinter and GCMetadataPrinterRegistry. The AsmWriter will lookfor such a subclass if the GCStrategy sets UsesMetadata:

MyGC::MyGC() {
  UsesMetadata = true;
}

This separation allows JIT-only clients to be smaller.

Note that LLVM does not currently have analogous APIs to support code generationin the JIT, nor using the object writers.

// lib/MyGC/MyGCPrinter.cpp - Example LLVM GC printer
 
#include "llvm/CodeGen/GCMetadataPrinter.h"
#include "llvm/Support/Compiler.h"
 
using namespace llvm;
 
namespace {
  class LLVM_LIBRARY_VISIBILITY MyGCPrinter : public GCMetadataPrinter {
  public:
    virtual void beginAssembly(AsmPrinter &AP);
 
    virtual void finishAssembly(AsmPrinter &AP);
  };
 
  GCMetadataPrinterRegistry::Add<MyGCPrinter>
  X("mygc", "My bespoke garbage collector.");
}

The collector should use AsmPrinter to print portable assembly code. Thecollector itself contains the stack map for the entire module, and may accessthe GCFunctionInfo using its own begin() and end() methods. Here’sa realistic example:

#include "llvm/CodeGen/AsmPrinter.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/Target/TargetAsmInfo.h"
#include "llvm/Target/TargetMachine.h"
 
void MyGCPrinter::beginAssembly(AsmPrinter &AP) {
  // Nothing to do.
}
 
void MyGCPrinter::finishAssembly(AsmPrinter &AP) {
  MCStreamer &OS = AP.OutStreamer;
  unsigned IntPtrSize = AP.getPointerSize();
 
  // Put this in the data section.
  OS.SwitchSection(AP.getObjFileLowering().getDataSection());
 
  // For each function...
  for (iterator FI = begin(), FE = end(); FI != FE; ++FI) {
    GCFunctionInfo &MD = **FI;
 
    // A compact GC layout. Emit this data structure:
    //
    // struct {
    //   int32_t PointCount;
    //   void *SafePointAddress[PointCount];
    //   int32_t StackFrameSize; // in words
    //   int32_t StackArity;
    //   int32_t LiveCount;
    //   int32_t LiveOffsets[LiveCount];
    // } __gcmap_<FUNCTIONNAME>;
 
    // Align to address width.
    AP.emitAlignment(IntPtrSize == 4 ? 2 : 3);
 
    // Emit PointCount.
    OS.AddComment("safe point count");
    AP.emitInt32(MD.size());
 
    // And each safe point...
    for (GCFunctionInfo::iterator PI = MD.begin(),
                                  PE = MD.end(); PI != PE; ++PI) {
      // Emit the address of the safe point.
      OS.AddComment("safe point address");
      MCSymbol *Label = PI->Label;
      AP.emitLabelPlusOffset(Label/*Hi*/, 0/*Offset*/, 4/*Size*/);
    }
 
    // Stack information never change in safe points! Only print info from the
    // first call-site.
    GCFunctionInfo::iterator PI = MD.begin();
 
    // Emit the stack frame size.
    OS.AddComment("stack frame size (in words)");
    AP.emitInt32(MD.getFrameSize() / IntPtrSize);
 
    // Emit stack arity, i.e. the number of stacked arguments.
    unsigned RegisteredArgs = IntPtrSize == 4 ? 5 : 6;
    unsigned StackArity = MD.getFunction().arg_size() > RegisteredArgs ?
                          MD.getFunction().arg_size() - RegisteredArgs : 0;
    OS.AddComment("stack arity");
    AP.emitInt32(StackArity);
 
    // Emit the number of live roots in the function.
    OS.AddComment("live root count");
    AP.emitInt32(MD.live_size(PI));
 
    // And for each live root...
    for (GCFunctionInfo::live_iterator LI = MD.live_begin(PI),
                                       LE = MD.live_end(PI);
                                       LI != LE; ++LI) {
      // Emit live root's offset within the stack frame.
      OS.AddComment("stack index (offset / wordsize)");
      AP.emitInt32(LI->StackOffset);
    }
  }
}

References

[Appel89] Runtime Tags Aren’t Necessary. Andrew W. Appel. Lisp and SymbolicComputation 19(7):703-705, July 1989.

[Goldberg91] Tag-free garbage collection for strongly typed programminglanguages. Benjamin Goldberg. ACM SIGPLAN PLDI‘91.

[Tolmach94] Tag-free garbage collection using explicit type parameters. AndrewTolmach. Proceedings of the 1994 ACM conference on LISP and functionalprogramming.

[Henderson2002] Accurate Garbage Collection in an Uncooperative Environment