Swift implemented in Swift

Libswift is the part of the Swift compiler, which is implemented in Swift.

With libswift it is possible to add SIL optimization passes written in Swift. It allows to gradually migrate the SIL optimizer from C++ to Swift.

Building

Libswift is a static library and it is built as part of the swift compiler using build-script and CMake.

To enable libswift, add the build-script option --libswift.

Currently the swift-frontend and sil-opt tools use libswift. Tools, which don't use any optimization passes from libswift don't need to link libswift. For example, all tools, which compile Swift source code, but don't optimize it, like SourceKit or lldb, don't need to link libswift. As long as initializeLibSwift() is not called there is no dependency on libswift.

Build modes

There are four build modes for libswift, which can be selected with the --libswift=<mode> build-script option:

• off (= default when no --libswift open is given): the compiler is not built with libswift. In this early stage of development this mode is the default.
• hosttools (= default for a --libswift option without argument): libswift is built with a pre-installed swift toolchain, using a swiftc which is expected to be in the command search path. This mode is the preferred way to build for local development, because it is the fastest way to build. It requires a 5.5 (or newer) swift toolchain to be installed on the host system.
• bootstrapping: The compiler and libswift are built with a two-stage bootstrapping process. This is the preferred mode if no swift toolchain is available on the system or if the build should not depend on any pre-installed toolchain.
• bootstrapping-with-hostlibs: This mode is only available on macOS. It's similar to bootstrapping, but links the compiler against the host system swift libraries instead of the built libraries. The build is faster than with bootstrapping because only the swiftmodule files of the bootstrapping libraries have to be built.

The libswift build mode is cached in the CMakeCache.txt file in the Swift build directory. When building with a different mode, the CMakeCache.txt has to be deleted.

Bootstrapping

The bootstrapping process is completely implemented with CMake dependencies. The build-script is not required to build the whole bootstrapping chain. For example, a ninja swift-frontend invocation builds all the required bootstrapping steps required for the final swift-frontend.

Bootstrapping involves the following steps:

1. The level-0 compiler

The build outputs of level-0 are stored in the bootstrapping0 directory under the main build directory.

In this first step swift-frontend is built, but without libswift. When more optimizations are migrated from C++ to libswift, this compiler will produce worse code than the final compiler.

2. The level-0 library

With the compiler from step 1, a minimal subset of the standard library is built in bootstrapping0/lib/swift. The subset contains the swift core library libswiftCore and, in case of a debug build, the libswiftOnoneSupport library. In the future it will also contain the concurrency library.

This library will be less optimized than the final library build.

This step is skipped when the build mode is bootstrapping-with-hostlibs.

3. The level-1 libswift

The build outputs of level-1 are stored in the bootstrapping1 directory under the main build directory.

The libswift library is built using the level-0 compiler and standard library from step 2 - or the OS libraries in case of bootstrapping-with-hostlibs.

4. The level-1 compiler

In this step, the level-1 swift-frontend is built which includes the libswift from step 3. This compiler already produces the exact same code as the final compiler, but might run slower, because its libswift is not optimized as good as in the final build.

Unless the build mode is bootstrapping-with-hostlibs, the level-1 compiler dynamically links against the level-1 library. This is specified with the RPATH setting in the executable.

5. The level-1 library

Like in step 2, a minimal subset of the standard library is built, using the level-1 compiler from step 4.

In this step, the build system redirects the compiler's dynamic library path to the level-0 library (by setting DY/LD_LIBRARY_PATH in SwiftSource.cmake:_compile_swift_files). This is needed because the level-1 libraries are not built, yet.

This step is skipped when the build mode is bootstrapping-with-hostlibs.

6. The final libswift

The final libswift is built with the level-1 compiler and standard library from step 5 - or the OS libraries in case of bootstrapping-with-hostlibs.

7. The final compiler

Now the final swift-frontend can be built, linking the final libswift from step 6.

8. The final standard library

With the final compiler from step 7, the final and full standard library is built. This library should be binary equivalent to the level-1 library.

Again, unless the build mode is bootstrapping-with-hostlibs, for building the standard library, the build system redirects the compiler's dynamic library path to the level-1 library.

Using and not using libswift

To include libswift in a tool, initializeLibSwift() must be called at the start of the tool. This must be done before any SIL objects are created. So ideally, at the start of the tool, e.g. at the place where INITIALIZE_LLVM() is called.

Currently the swift-frontend and sil-opt tools use libswift.

Tools, which don't use any optimization passes from libswift don't need to link libswift. For example, all tools, which compile Swift source code, but don't optimize it, like SourceKit or lldb, don't need to link libswift. As long as initializeLibSwift() is not called there is no dependency on libswift.

This also means that currently it is not possible to implement mandatory passes in libswift, because this would break tools which compile Swift code but don't use libswift. When we want to implement mandatory passes in libswift in the future, we'll need to link libswift to all those tools.

SIL

The design of SIL in libswift matches very closely the design on the C++ side. For example, there are functions, basic blocks, instructions, SIL values, etc.

Though, there are some small deviations from the C++ SIL design. Either due to the nature of the Swift language (e.g. the SIL Value is a protocol, not a class), or improvements, which could be done in C++ as well.

Bridging SIL between C++ and Swift is toll-free, i.e. does not involve any "conversion" between C++ and Swift SIL.

The bridging layer

The bridging layer is a small interface layer which enables calling into the SIL C++ API from the Swift side. Currently the bridging layer is implemented in C using C interop. In future it can be replaced by a C++ implementation by using C++ interop, which will further simplify the bridging layer or make it completely obsolete. But this is an implementation detail and does not affect the API of SIL in libswift.

The bridging layer consists of the C header file SILBridging.h and its implementation file SILBridging.cpp. The header file contains all the bridging functions and some C data structures like BridgedStringRef (once we use C++ interop, those C data structures are not required anymore and can be removed).

SIL C++ objects in Swift

The core SIL C++ classes have corresponding classes in libswift, for example Function, BasicBlock or all the instruction classes.

This makes libswift easy to use and it allows to program in a "Swifty" style. For example one can write

  for inst in block.instructions {
    if let cfi = inst as? CondFailInst {
      // ...
    }
  }

Bridging SIL classes is implemented by including a two word Swift object header (SwiftObjectHeader) in the C++ definition of a class, like in SILFunction, SILBasicBlock or SILNode. This enables to use SIL objects on both, the C++ and the Swift, side.

The Swift class metatypes are "registered" by registerClass(), called from initializeLibSwift(). On the C++ side, they are stored in static global variables (see registerBridgedClass()) and then used to initialize the object headers in the class constructors.

The reference counts in the object header are initialized to "immortal", which let's all ARC operations on SIL objects be no-ops.

The Swift classes don't define any stored properties, because those would overlap data fields in the C++ classes. Instead, data fields are accessed via computed properties, which call bridging functions to retrieve the actual data values.

Lazy implementation

In the current state the SIL functionality and API is not completely implemented, yet. For example, not all instruction classes have a corresponding class in libswift. Whenever a new libswift optimization needs a specific SIL feature, like an instruction, a Builder-function or an accessor to a data field, it's easy to add the missing parts.

For example, to add a new instruction class:

• replace the macro which defines the instruction in SILNodes.def with a BRIDGED-* macro
• add the instruction class in Instruction.swift
• register the class in registerSILClasses()
• if needed, add bridging functions to access the instruction's data fields.

No yet implemented instruction classes are mapped to a "placeholder" instruction, e.g UnimplementedInstruction. This ensures that optimizations can process any kind of SIL, even if some instructions don't have a representation in libswift yet.

The Optimizer

Similar to SIL, the optimizer also uses a small bridging layer (OptimizerBridging.h). Passes are registered in registerSwiftPasses(), called from initializeLibSwift(). The C++ PassManager can then call a libwift pass like any other SILFunctionTransform pass.

To add a new function pass:

• add a SWIFT_FUNCTION_PASS entry in Passes.def
• create a new Swift file in libswift/Sources/Optimizer/FunctionPasses
• add a FunctionPass global
• register the pass in registerSwiftPasses()

All SIL modifications, which a pass can do, are going through the FunctionPassContext - the second parameter of its run-function. In other words, the context is the central place to make modifications. This enables automatic change notifications to the pass manager. Also, it makes it easier to build a concurrent pass manager in future.

Instruction Passes

In addition to function passes, libswift provides the infrastructure for instruction passes. Instruction passes are invoked from SILCombine (in the C++ SILOptimizer) and correspond to a visit-function in SILCombine.

With instruction passes it's possible to implement small peephole optimizations for certain instruction classes.

To add a new instruction pass:

• add a SWIFT_INSTRUCTION_PASS entry in Passes.def
• create a new Swift file in libswift/Sources/Optimizer/InstructionPasses
• add an InstructionPass global
• register the pass in registerSwiftPasses()
• if this passes replaces an existing SILCombiner visit function, remove the old visit function

Performance

The performance of libswift is very important, because compile time is critical. Some performance considerations:

• Memory is managed on the C++ side. On the Swift side, SIL objects are treated as "immortal" objects, which avoids (most of) ARC overhead. ARC runtime functions are still being called, but no atomic reference counting operations are done. In future we could add a compiler feature to mark classes as immortal to avoid the runtime calls at all.

• Minimizing memory allocations: libswift provides data structures which are malloc-free. For example StackList can be used in optimizations to implement work lists without any memory allocations. (Not yet done: BasicBlockSet, BasicBlockData)

But most importantly, if there are performance issues with the current compiler, the design of libswift should make it possible to fix performance deficiencies with future compiler improvements.