ORC Design and Implementation

Introduction

This document aims to provide a high-level overview of the design and implementation of the ORC JIT APIs. Except where otherwise stated, all discussion applies to the design of the APIs as of LLVM verison 9 (ORCv2).

Use-cases

ORC provides a modular API for building JIT compilers. There are a range of use cases for such an API. For example:

1. The LLVM tutorials use a simple ORC-based JIT class to execute expressions compiled from a toy languge: Kaleidoscope.

2. The LLVM debugger, LLDB, uses a cross-compiling JIT for expression evaluation. In this use case, cross compilation allows expressions compiled in the debugger process to be executed on the debug target process, which may be on a different device/architecture.

3. In high-performance JITs (e.g. JVMs, Julia) that want to make use of LLVM’s optimizations within an existing JIT infrastructure.

4. In interpreters and REPLs, e.g. Cling (C++) and the Swift interpreter.

By adoping a modular, library-based design we aim to make ORC useful in as many of these contexts as possible.

Features

ORC provides the following features:

• JIT-linking links relocatable object files (COFF, ELF, MachO) [1] into a target process an runtime. The target process may be the same process that contains the JIT session object and jit-linker, or may be another process (even one running on a different machine or architecture) that communicates with the JIT via RPC.
• LLVM IR compilation, which is provided by off the shelf components (IRCompileLayer, SimpleCompiler, ConcurrentIRCompiler) that make it easy to add LLVM IR to a JIT’d process.
• Eager and lazy compilation. By default, ORC will compile symbols as soon as they are looked up in the JIT session object (ExecutionSession). Compiling eagerly by default makes it easy to use ORC as a simple in-memory compiler for an existing JIT. ORC also provides a simple mechanism, lazy-reexports, for deferring compilation until first call.
• Support for custom compilers and program representations. Clients can supply custom compilers for each symbol that they define in their JIT session. ORC will run the user-supplied compiler when the a definition of a symbol is needed. ORC is actually fully language agnostic: LLVM IR is not treated specially, and is supported via the same wrapper mechanism (the MaterializationUnit class) that is used for custom compilers.
• Concurrent JIT’d code and concurrent compilation. JIT’d code may spawn multiple threads, and may re-enter the JIT (e.g. for lazy compilation) concurrently from multiple threads. The ORC APIs also support running multiple compilers concurrently, and provides off-the-shelf infrastructure to track dependencies on running compiles (e.g. to ensure that we never call into code until it is safe to do so, even if that involves waiting on multiple compiles).
• Orthogonality and composability: Each of the features above can be used (or not) independently. It is possible to put ORC components together to make a non-lazy, in-process, single threaded JIT or a lazy, out-of-process, concurrent JIT, or anything in between.

LLJIT and LLLazyJIT

ORC provides two basic JIT classes off-the-shelf. These are useful both as examples of how to assemble ORC components to make a JIT, and as replacements for earlier LLVM JIT APIs (e.g. MCJIT).

The LLJIT class uses an IRCompileLayer and RTDyldObjectLinkingLayer to support compilation of LLVM IR and linking of relocatable object files. All operations are performed eagerly on symbol lookup (i.e. a symbol’s definition is compiled as soon as you attempt to look up its address). LLJIT is a suitable replacement for MCJIT in most cases (note: some more advanced features, e.g. JITEventListeners are not supported yet).

The LLLazyJIT extends LLJIT and adds a CompileOnDemandLayer to enable lazy compilation of LLVM IR. When an LLVM IR module is added via the addLazyIRModule method, function bodies in that module will not be compiled until they are first called. LLLazyJIT aims to provide a replacement of LLVM’s original (pre-MCJIT) JIT API.

LLJIT and LLLazyJIT instances can be created using their respective builder classes: LLJITBuilder and LLazyJITBuilder. For example, assuming you have a module M loaded on an ThreadSafeContext Ctx:

// Try to detect the host arch and construct an LLJIT instance.
auto JIT = LLJITBuilder().create();

// If we could not construct an instance, return an error.
if (!JIT)
  return JIT.takeError();

// Add the module.
if (auto Err = JIT->addIRModule(TheadSafeModule(std::move(M), Ctx)))
  return Err;

// Look up the JIT'd code entry point.
auto EntrySym = JIT->lookup("entry");
if (!EntrySym)
  return EntrySym.takeError();

auto *Entry = (void(*)())EntrySym.getAddress();

Entry();

The builder clasess provide a number of configuration options that can be specified before the JIT instance is constructed. For example:

// Build an LLLazyJIT instance that uses four worker threads for compilation,
// and jumps to a specific error handler (rather than null) on lazy compile
// failures.

void handleLazyCompileFailure() {
  // JIT'd code will jump here if lazy compilation fails, giving us an
  // opportunity to exit or throw an exception into JIT'd code.
  throw JITFailed();
}

auto JIT = LLLazyJITBuilder()
             .setNumCompileThreads(4)
             .setLazyCompileFailureAddr(
                 toJITTargetAddress(&handleLazyCompileFailure))
             .create();

// ...

For users wanting to get started with LLJIT a minimal example program can be found at llvm/examples/HowToUseLLJIT.

Design Overview

ORC’s JIT’d program model aims to emulate the linking and symbol resolution rules used by the static and dynamic linkers. This allows ORC to JIT arbitrary LLVM IR, including IR produced by an ordinary static compiler (e.g. clang) that uses constructs like symbol linkage and visibility, and weak and common symbol definitions.

To see how this works, imagine a program foo which links against a pair of dynamic libraries: libA and libB. On the command line, building this program might look like:

$ clang++ -shared -o libA.dylib a1.cpp a2.cpp
$ clang++ -shared -o libB.dylib b1.cpp b2.cpp
$ clang++ -o myapp myapp.cpp -L. -lA -lB
$ ./myapp

In ORC, this would translate into API calls on a “CXXCompilingLayer” (with error checking omitted for brevity) as:

ExecutionSession ES;
RTDyldObjectLinkingLayer ObjLinkingLayer(
    ES, []() { return llvm::make_unique<SectionMemoryManager>(); });
CXXCompileLayer CXXLayer(ES, ObjLinkingLayer);

// Create JITDylib "A" and add code to it using the CXX layer.
auto &LibA = ES.createJITDylib("A");
CXXLayer.add(LibA, MemoryBuffer::getFile("a1.cpp"));
CXXLayer.add(LibA, MemoryBuffer::getFile("a2.cpp"));

// Create JITDylib "B" and add code to it using the CXX layer.
auto &LibB = ES.createJITDylib("B");
CXXLayer.add(LibB, MemoryBuffer::getFile("b1.cpp"));
CXXLayer.add(LibB, MemoryBuffer::getFile("b2.cpp"));

// Specify the search order for the main JITDylib. This is equivalent to a
// "links against" relationship in a command-line link.
ES.getMainJITDylib().setSearchOrder({{&LibA, false}, {&LibB, false}});
CXXLayer.add(ES.getMainJITDylib(), MemoryBuffer::getFile("main.cpp"));

// Look up the JIT'd main, cast it to a function pointer, then call it.
auto MainSym = ExitOnErr(ES.lookup({&ES.getMainJITDylib()}, "main"));
auto *Main = (int(*)(int, char*[]))MainSym.getAddress();

v int Result = Main(…);

This example tells us nothing about how or when compilation will happen. That will depend on the implementation of the hypothetical CXXCompilingLayer. The same linker-based symbol resolution rules will apply regardless of that implementation, however. For example, if a1.cpp and a2.cpp both define a function “foo” then ORCv2 will generate a duplicate definition error. On the other hand, if a1.cpp and b1.cpp both define “foo” there is no error (different dynamic libraries may define the same symbol). If main.cpp refers to “foo”, it should bind to the definition in LibA rather than the one in LibB, since main.cpp is part of the “main” dylib, and the main dylib links against LibA before LibB.

Many JIT clients will have no need for this strict adherence to the usual ahead-of-time linking rules, and should be able to get by just fine by putting all of their code in a single JITDylib. However, clients who want to JIT code for languages/projects that traditionally rely on ahead-of-time linking (e.g. C++) will find that this feature makes life much easier.

Symbol lookup in ORC serves two other important functions, beyond providing addresses for symbols: (1) It triggers compilation of the symbol(s) searched for (if they have not been compiled already), and (2) it provides the synchronization mechanism for concurrent compilation. The pseudo-code for the lookup process is:

construct a query object from a query set and query handler
lock the session
lodge query against requested symbols, collect required materializers (if any)
unlock the session
dispatch materializers (if any)

In this context a materializer is something that provides a working definition of a symbol upon request. Usually materializers are just wrappers for compilers, but they may also wrap a jit-linker directly (if the program representation backing the definitions is an object file), or may even be a class that writes bits directly into memory (for example, if the definitions are stubs). Materialization is the blanket term for any actions (compiling, linking, splatting bits, registering with runtimes, etc.) that are requried to generate a symbol definition that is safe to call or access.

As each materializer completes its work it notifies the JITDylib, which in turn notifies any query objects that are waiting on the newly materialized definitions. Each query object maintains a count of the number of symbols that it is still waiting on, and once this count reaches zero the query object calls the query handler with a SymbolMap (a map of symbol names to addresses) describing the result. If any symbol fails to materialize the query immediately calls the query handler with an error.

The collected materialization units are sent to the ExecutionSession to be dispatched, and the dispatch behavior can be set by the client. By default each materializer is run on the calling thread. Clients are free to create new threads to run materializers, or to send the work to a work queue for a thread pool (this is what LLJIT/LLLazyJIT do).

Top Level APIs

Many of ORC’s top-level APIs are visible in the example above:

• ExecutionSession represents the JIT’d program and provides context for the JIT: It contains the JITDylibs, error reporting mechanisms, and dispatches the materializers.
• JITDylibs provide the symbol tables.
• Layers (ObjLinkingLayer and CXXLayer) are wrappers around compilers and allow clients to add uncompiled program representations supported by those compilers to JITDylibs.

Several other important APIs are used explicitly. JIT clients need not be aware of them, but Layer authors will use them:

• MaterializationUnit - When XXXLayer::add is invoked it wraps the given program representation (in this example, C++ source) in a MaterializationUnit, which is then stored in the JITDylib. MaterializationUnits are responsible for describing the definitions they provide, and for unwrapping the program representation and passing it back to the layer when compilation is required (this ownership shuffle makes writing thread-safe layers easier, since the ownership of the program representation will be passed back on the stack, rather than having to be fished out of a Layer member, which would require synchronization).
• MaterializationResponsibility - When a MaterializationUnit hands a program representation back to the layer it comes with an associated MaterializationResponsibility object. This object tracks the definitions that must be materialized and provides a way to notify the JITDylib once they are either successfully materialized or a failure occurs.

Handy utilities

TBD: absolute symbols, aliases, off-the-shelf layers.

Laziness

Laziness in ORC is provided by a utility called “lazy-reexports”. The aim of this utility is to re-use the synchronization provided by the symbol lookup mechanism to make it safe to lazily compile functions, even if calls to the stub occur simultaneously on multiple threads of JIT’d code. It does this by reducing lazy compilation to symbol lookup: The lazy stub performs a lookup of its underlying definition on first call, updating the function body pointer once the definition is available. If additional calls arrive on other threads while compilation is ongoing they will be safely blocked by the normal lookup synchronization guarantee (no result until the result is safe) and can also proceed as soon as compilation completes.

TBD: Usage example.

Supporting Custom Compilers

TBD.

Transitioning from ORCv1 to ORCv2

Since LLVM 7.0, new ORC development work has focused on adding support for concurrent JIT compilation. The new APIs (including new layer interfaces and implementations, and new utilities) that support concurrency are collectively referred to as ORCv2, and the original, non-concurrent layers and utilities are now referred to as ORCv1.

The majority of the ORCv1 layers and utilities were renamed with a ‘Legacy’ prefix in LLVM 8.0, and have deprecation warnings attached in LLVM 9.0. In LLVM 10.0 ORCv1 will be removed entirely.

Transitioning from ORCv1 to ORCv2 should be easy for most clients. Most of the ORCv1 layers and utilities have ORCv2 counterparts[2]_ that can be directly substituted. However there are some design differences between ORCv1 and ORCv2 to be aware of:

1. ORCv2 fully adopts the JIT-as-linker model that began with MCJIT. Modules (and other program representations, e.g. Object Files) are no longer added directly to JIT classes or layers. Instead, they are added to JITDylib instances by layers. The JITDylib determines where the definitions reside, the layers determine how the definitions will be compiled. Linkage relationships between JITDylibs determine how inter-module references are resolved, and symbol resolvers are no longer used. See the section Design Overview for more details.

   Unless multiple JITDylibs are needed to model linkage relationsips, ORCv1 clients should place all code in the main JITDylib (returned by ExecutionSession::getMainJITDylib()). MCJIT clients should use LLJIT (see LLJIT and LLLazyJIT).

2. All JIT stacks now need an ExecutionSession instance. ExecutionSession manages the string pool, error reporting, synchronization, and symbol lookup.

3. ORCv2 uses uniqued strings (SymbolStringPtr instances) rather than string values in order to reduce memory overhead and improve lookup performance. See the subsection How to manage symbol strings.

4. IR layers require ThreadSafeModule instances, rather than std::unique_ptr<Module>s. ThreadSafeModule is a wrapper that ensures that Modules that use the same LLVMContext are not accessed concurrently. See How to use ThreadSafeModule and ThreadSafeContext.

5. Symbol lookup is no longer handled by layers. Instead, there is a lookup method on JITDylib that takes a list of JITDylibs to scan.

   ExecutionSession ES;
   JITDylib &JD1 = ...;
   JITDylib &JD2 = ...;
   
   auto Sym = ES.lookup({&JD1, &JD2}, ES.intern("_main"));
   

6. Module removal is not yet supported. There is no equivalent of the layer concept removeModule/removeObject methods. Work on resource tracking and removal in ORCv2 is ongoing.

For code examples and suggestions of how to use the ORCv2 APIs, please see the section How-tos.

How-tos

How to manage symbol strings

Symbol strings in ORC are uniqued to improve lookup performance, reduce memory overhead, and allow symbol names to function as efficient keys. To get the unique SymbolStringPtr for a string value, call the ExecutionSession::intern method:

ExecutionSession ES;
/// ...
auto MainSymbolName = ES.intern("main");

If you wish to perform lookup using the C/IR name of a symbol you will also need to apply the platform linker-mangling before interning the string. On Linux this mangling is a no-op, but on other platforms it usually involves adding a prefix to the string (e.g. ‘_’ on Darwin). The mangling scheme is based on the DataLayout for the target. Given a DataLayout and an ExecutionSession, you can create a MangleAndInterner function object that will perform both jobs for you:

ExecutionSession ES;
const DataLayout &DL = ...;
MangleAndInterner Mangle(ES, DL);

// ...

// Portable IR-symbol-name lookup:
auto Sym = ES.lookup({&ES.getMainJITDylib()}, Mangle("main"));

How to create JITDylibs and set up linkage relationships

In ORC, all symbol definitions reside in JITDylibs. JITDylibs are created by calling the ExecutionSession::createJITDylib method with a unique name:

ExecutionSession ES;
auto &JD = ES.createJITDylib("libFoo.dylib");

The JITDylib is owned by the ExecutionEngine instance and will be freed when it is destroyed.

A JITDylib representing the JIT main program is created by ExecutionEngine by default. A reference to it can be obtained by calling ExecutionSession::getMainJITDylib():

ExecutionSession ES;
auto &MainJD = ES.getMainJITDylib();

How to use ThreadSafeModule and ThreadSafeContext

ThreadSafeModule and ThreadSafeContext are wrappers around Modules and LLVMContexts respectively. A ThreadSafeModule is a pair of a std::unique_ptr<Module> and a (possibly shared) ThreadSafeContext value. A ThreadSafeContext is a pair of a std::unique_ptr<LLVMContext> and a lock. This design serves two purposes: providing both a locking scheme and lifetime management for LLVMContexts. The ThreadSafeContext may be locked to prevent accidental concurrent access by two Modules that use the same LLVMContext. The underlying LLVMContext is freed once all ThreadSafeContext values pointing to it are destroyed, allowing the context memory to be reclaimed as soon as the Modules referring to it are destroyed.

ThreadSafeContexts can be explicitly constructed from a std::unique_ptr<LLVMContext>:

ThreadSafeContext TSCtx(llvm::make_unique<LLVMContext>());

ThreadSafeModules can be constructed from a pair of a std::unique_ptr<Module> and a ThreadSafeContext value. ThreadSafeContext values may be shared between multiple ThreadSafeModules:

ThreadSafeModule TSM1(
  llvm::make_unique<Module>("M1", *TSCtx.getContext()), TSCtx);

ThreadSafeModule TSM2(
  llvm::make_unique<Module>("M2", *TSCtx.getContext()), TSCtx);

Before using a ThreadSafeContext, clients should ensure that either the context is only accessible on the current thread, or that the context is locked. In the example above (where the context is never locked) we rely on the fact that both TSM1 and TSM2, and TSCtx are all created on one thread. If a context is going to be shared between threads then it must be locked before the context, or any Modules attached to it, are accessed. When code is added to in-tree IR layers this locking is is done automatically by the BasicIRLayerMaterializationUnit::materialize method. In all other situations, for example when writing a custom IR materialization unit, or constructing a new ThreadSafeModule from higher-level program representations, locking must be done explicitly:

void HighLevelRepresentationLayer::emit(MaterializationResponsibility R,
                                        HighLevelProgramRepresentation H) {
  // Get or create a context value that may be shared between threads.
  ThreadSafeContext TSCtx = getContext();

  // Lock the context to prevent concurrent access.
  auto Lock = TSCtx.getLock();

  // IRGen a module onto the locked Context.
  ThreadSafeModule TSM(IRGen(H, *TSCtx.getContext()), TSCtx);

  // Emit the module to the base layer with the context still locked.
  BaseIRLayer.emit(std::move(R), std::move(TSM));
}

Clients wishing to maximize possibilities for concurrent compilation will want to create every new ThreadSafeModule on a new ThreadSafeContext. For this reason a convenience constructor for ThreadSafeModule is provided that implicitly constructs a new ThreadSafeContext value from a std::unique_ptr<LLVMContext>:

// Maximize concurrency opportunities by loading every module on a
// separate context.
for (const auto &IRPath : IRPaths) {
  auto Ctx = llvm::make_unique<LLVMContext>();
  auto M = llvm::make_unique<LLVMContext>("M", *Ctx);
  CompileLayer.add(ES.getMainJITDylib(),
                   ThreadSafeModule(std::move(M), std::move(Ctx)));
}

Clients who plan to run single-threaded may choose to save memory by loading all modules on the same context:

// Save memory by using one context for all Modules:
ThreadSafeContext TSCtx(llvm::make_unique<LLVMContext>());
for (const auto &IRPath : IRPaths) {
  ThreadSafeModule TSM(parsePath(IRPath, *TSCtx.getContext()), TSCtx);
  CompileLayer.add(ES.getMainJITDylib(), ThreadSafeModule(std::move(TSM));
}

How to Add Process and Library Symbols to the JITDylibs

JIT’d code typically needs access to symbols in the host program or in supporting libraries. References to process symbols can be “baked in” to code as it is compiled by turning external references into pre-resolved integer constants, however this ties the JIT’d code to the current process’s virtual memory layout (meaning that it can not be cached between runs) and makes debugging lower level program representations difficult (as all external references are opaque integer values). A bettor solution is to maintain symbolic external references and let the jit-linker bind them for you at runtime. To allow the JIT linker to find these external definitions their addresses must be added to a JITDylib that the JIT’d definitions link against.

Adding definitions for external symbols could be done using the absoluteSymbols function:

const DataLayout &DL = getDataLayout();
MangleAndInterner Mangle(ES, DL);

auto &JD = ES.getMainJITDylib();

JD.define(
  absoluteSymbols({
    { Mangle("puts"), pointerToJITTargetAddress(&puts)},
    { Mangle("gets"), pointerToJITTargetAddress(&getS)}
  }));

Manually adding absolute symbols for a large or changing interface is cumbersome however, so ORC provides an alternative to generate new definitions on demand: definition generators. If a definition generator is attached to a JITDylib, then any unsuccessful lookup on that JITDylib will fall back to calling the definition generator, and the definition generator may choose to generate a new definition for the missing symbols. Of particular use here is the DynamicLibrarySearchGenerator utility. This can be used to reflect the whole exported symbol set of the process or a specific dynamic library, or a subset of either of these determined by a predicate.

For example, to load the whole interface of a runtime library:

const DataLayout &DL = getDataLayout();
auto &JD = ES.getMainJITDylib();

JD.setGenerator(DynamicLibrarySearchGenerator::Load("/path/to/lib"
                                                    DL.getGlobalPrefix()));

// IR added to JD can now link against all symbols exported by the library
// at '/path/to/lib'.
CompileLayer.add(JD, loadModule(...));

Or, to expose a whitelisted set of symbols from the main process:

const DataLayout &DL = getDataLayout();
MangleAndInterner Mangle(ES, DL);

auto &JD = ES.getMainJITDylib();

DenseSet<SymbolStringPtr> Whitelist({
    Mangle("puts"),
    Mangle("gets")
  });

// Use GetForCurrentProcess with a predicate function that checks the
// whitelist.
JD.setGenerator(
  DynamicLibrarySearchGenerator::GetForCurrentProcess(
    DL.getGlobalPrefix(),
    [&](const SymbolStringPtr &S) { return Whitelist.count(S); }));

// IR added to JD can now link against any symbols exported by the process
// and contained in the whitelist.
CompileLayer.add(JD, loadModule(...));

Future Features

TBD: Speculative compilation. Object Caches.

[1]	Formats/architectures vary in terms of supported features. MachO and ELF tend to have better support than COFF. Patches very welcome!

[2]

The LazyEmittingLayer, RemoteObjectClientLayer and RemoteObjectServerLayer do not have counterparts in the new system. In the case of LazyEmittingLayer it was simply no longer needed: in ORCv2, deferring compilation until symbols are looked up is the default. The removal of RemoteObjectClientLayer and RemoteObjectServerLayer means that JIT stacks can no longer be split across processes, however this functionality appears not to have been used.

[3]

Sharing ThreadSafeModules in a concurrent compilation can be dangerous: if interdependent modules are loaded on the same context, but compiled on different threads a deadlock may occur (with each compile waiting for the other(s) to complete, and the other(s) unable to proceed because the context is locked).

[4]	Mostly. Weak definitions are handled correctly within dylibs, but if multiple dylibs provide a weak definition of a symbol each will end up with its own definition (similar to how weak symbols in Windows DLLs behave). This will be fixed in the future.