LLVMSwift is a pure Swift interface to the LLVM API and its associated libraries. It provides native, easy-to-use components to make compiler development fun.

The root unit of organization of an LLVM IR program is a Module

let module = Module(name: "main")

LLVM IR construction is handled by IRBuilder objects. An IRBuilder is a cursor pointed inside a context, and as such has ways of extending that context and moving around inside of it.

Defining a function and moving the cursor to a point where we can begin inserting instructions is done like so:

let builder = IRBuilder(module: module)

let main = builder.addFunction("main",
                               type: FunctionType([], IntType.int64))
let entry = main.appendBasicBlock(named: "entry")
builder.positionAtEnd(of: entry)

Inserting instructions creates native IRValue placeholder objects that allow us to structure LLVM IR programs just like Swift programs:

let constant = IntType.int64.constant(21)
let sum = builder.buildAdd(constant, constant)
builder.buildRet(sum)

This simple program generates the following IR:

// module.dump()

define i64 @main() {
entry:
  ret i64 42
}

LLVM IR is a strong, statically typed language. As such, values and functions are tagged with their types, and conversions between them must be explicit (see Conversion Operators). LLVMSwift represents this with values conforming to the IRType protocol and defines the following types:

Control flow is changed through the unconditional and conditional br instruction.

LLVM is also famous for a control-flow specific IR construct called a PHI node. Because all instructions in LLVM IR are in SSA (Single Static Assignment) form, a PHI node is necessary when the value of a variable assignment depends on the path the flow of control takes through the program. For example, let's try to build the following Swift program in IR:

func calculateFibs(_ backward : Bool) -> Double {
  let retVal : Double
  if !backward {
    // the fibonacci series (sort of)
    retVal = 1/89
  } else {
    // the fibonacci series (sort of) backwards
    retVal = 1/109
  }
  return retVal
}

Notice that the value of retVal depends on the path the flow of control takes through this program, so we must emit a PHI node to properly initialize it:

let function = builder.addFunction("calculateFibs", 
                                   type: FunctionType([IntType.int1], 
                                                      FloatType.double))
let entryBB = function.appendBasicBlock(named: "entry")
builder.positionAtEnd(of: entryBB)

// allocate space for a local value		
let local = builder.buildAlloca(type: FloatType.double, name: "local")

// Compare to the condition
let test = builder.buildICmp(function.parameters[0], IntType.int1.zero(), .equal)

// Create basic blocks for "then", "else", and "merge"
let thenBB = function.appendBasicBlock(named: "then")
let elseBB = function.appendBasicBlock(named: "else")
let mergeBB = function.appendBasicBlock(named: "merge")

builder.buildCondBr(condition: test, then: thenBB, else: elseBB)

// MARK: Then Block
builder.positionAtEnd(of: thenBB)
// local = 1/89, the fibonacci series (sort of)
let thenVal = FloatType.double.constant(1/89)
// Branch to the merge block
builder.buildBr(mergeBB)

// MARK: Else Block
builder.positionAtEnd(of: elseBB)
// local = 1/109, the fibonacci series (sort of) backwards
let elseVal = FloatType.double.constant(1/109)
// Branch to the merge block
builder.buildBr(mergeBB)

// MARK: Merge Block
builder.positionAtEnd(of: mergeBB)
let phi = builder.buildPhi(FloatType.double, name: "phi_example")
phi.addIncoming([
  (thenVal, thenBB),
  (elseVal, elseBB),
])
builder.buildStore(phi, to: local)
let ret = builder.buildLoad(local, type: FloatType.double, name: "ret")
builder.buildRet(ret)

This program generates the following IR:

define double @calculateFibs(i1) {
entry:
  %local = alloca double
  %1 = icmp ne i1 %0, false
  br i1 %1, label %then, label %else

then:                                             ; preds = %entry
  br label %merge

else:                                             ; preds = %entry
  br label %merge

merge:                                            ; preds = %else, %then
  %phi_example = phi double [ 0x3F8702E05C0B8170, %then ], [ 0x3F82C9FB4D812CA0, %else ]
  store double %phi_example, double* %local
  %ret = load double, double* %local
  ret double %ret
}

LLVMSwift provides a JIT abstraction to make executing code in LLVM modules quick and easy. Let's execute the PHI node example from before:

// Setup the JIT
let jit = try JIT(machine: TargetMachine())
typealias FnPtr = @convention(c) (Bool) -> Double
_ = try jit.addEagerlyCompiledIR(module) { (name) -> JIT.TargetAddress in
  return JIT.TargetAddress()
}
// Retrieve a handle to the function we're going to invoke
let addr = try jit.address(of: "calculateFibs")
let fn = unsafeBitCast(addr, to: FnPtr.self)
// Call the function!
print(fn(true)) // 0.00917431192660551...
print(fn(false)) // 0.0112359550561798...

There are a couple annoying steps you need to accomplish before building LLVMSwift:

• Install LLVM 11.0+ using your favorite package manager. For example:
  • brew install llvm@11
• Ensure llvm-config is in your PATH
  • That will reside in the /bin folder wherever your package manager installed LLVM.
• Create a pkg-config file for your specific LLVM installation.
  • We have a utility for this: swift utils/make-pkgconfig.swift

Once you do that, you can add LLVMSwift as a dependency for your own Swift compiler projects!

.package(url: "https://github.com/llvm-swift/LLVMSwift.git", from: "0.8.0")

We really recommend using SwiftPM with LLVMSwift, but if your project is structured in such a way that makes using SwiftPM impractical or impossible, use the following instructions:

• Xcode:
  • Add this repository as a git submodule
  • Add the files in Sources/ to your Xcode project.
  • Under Library Search Paths add the output of llvm-config --libdir
  • Under Header Search Paths add the output of llvm-config --includedir
  • Under Link Target with Libraries drag in /path/to/your/llvm/lib/libLLVM.dylib

This project is used by Trill for all its code generation.

• Harlan Haskins (@harlanhaskins)
• Robert Widmann (@CodaFi)

This project is released under the MIT license, a copy of which is available in this repo.