LLVM 101 | ECOSYSTEM Research

LLVM is not a single compiler, but a collection of modular and reusable compiler and toolchain technologies. For a compiler engineer, its value lies in its strict separation of concerns, allowing you to build a language frontend that compiles to a universal Intermediate Representation (IR), letting LLVM handle the heavy lifting of target-independent optimization and target-specific machine code generation. 1. The Three-Phase Architecture LLVM enforces a strict three-phase design: 1. Frontend: Parses source code, performs semantic analysis, builds an Abstract Syntax Tree (AST), and generates LLVM IR. (e.g., Clang for C/C++, Rustc for Rust, Zig's compiler). 2. Optimizer (Middle-End): Takes LLVM IR, applies a series of target-independent optimizations (passes) to improve runtime performance or reduce code size, and outputs optimized LLVM IR. 3. Backend (Code Generator): Takes the optimized LLVM IR and lowers it to target-specific assembly code or machine code (e.g., x86-64, ARM, RISC-V, WebAssembly). 2. LLVM Intermediate Representation (IR) The heart of LLVM is its IR. It is a strictly typed, Static Single Assignment (SSA) based representation. Every variable is assigned exactly once, making data-flow analysis trivial for optimization passes. LLVM IR exists in three isomorphic forms: In-memory compiler IR: The C++ object model manipulated by the API (llvm::Value, llvm::Instruction). On-disk bitcode (.bc): A dense binary format for fast loading and Link-Time Optimization (LTO). Human-readable assembly (.ll): Used for debugging and testing. Example: C to LLVM IR Compiled to LLVM IR (clang -S -emit-llvm): Key IR Concepts: Basic Blocks: Sequences of instructions with a single entry and a single exit point (terminator instructions like ret or br). Values and Uses: Everything that computes a result inherits from llvm::Value. Instructions that consume these values are llvm::User. Phi Nodes (phi): Required in SSA to merge control flow paths, choosing a value based on which basic block executed previously. 3. The Pass Manager and Optimizations The middle-end operates as a pipeline of passes. Modern LLVM uses the New Pass Manager, which separates passes into two categories: Analysis Passes: Compute information about the IR without modifying it (e.g., Alias Analysis, Dominator Tree, Scalar Evolution). Their results are cached and invalidated only when the IR changes significantly. Transform Passes: Mutate the IR to optimize it (e.g., Dead Code Elimination, Loop Unrolling, Inlining, Vectorization). When designing a frontend, you typically map your language's semantics to IR and rely on standard pipelines (like O2 or O3) to optimize it. You can interact with the pass pipeline using the opt command-line tool to test specific optimizations in isolation. 4. The LLVM Backend (Code Generation) The backend translates LLVM IR into machine code. This is a multi-step lowering process: 1. Instruction Selection (ISel): Translates LLVM IR into a target-specific Directed Acyclic Graph (SelectionDAG) or uses the newer GlobalISel framework. It maps generic IR operations to specific machine instructions. 2. Machine IR (MIR): After ISel, the code is represented as Machine IR, which is much closer to actual assembly but still uses infinite virtual registers. 3. Register Allocation: Maps the infinite virtual registers to finite physical registers available on the target architecture (x86-64, ARM, etc.). Spills values to the stack if necessary. 4. Instruction Scheduling: Reorders machine instructions to maximize instruction-level parallelism and minimize pipeline stalls. 5. Code Emission: Outputs the final object file (ELF, Mach-O, COFF) or raw assembly. 5. Building a Frontend: IRBuilder For compiler engineers building a new language, the llvm::IRBuilder class is the primary interface. It provides helper methods to construct IR instructions and automatically inserts them into the correct basic block. 6. Essential LLVM Tools llc: The LLVM static compiler. Takes LLVM IR (.ll or .bc) and compiles it to assembly or object files for a specific target. opt: The LLVM optimizer. Runs arbitrary optimization passes on IR. llvm-dis / llvm-as: Disassembles .bc to .ll / Assembles .ll to .bc. TableGen (llvm-tblgen): A domain-specific language used heavily in the LLVM backend to describe target architectures (registers, instruction formats, calling conventions) without writing thousands of lines of C++ boilerplate. ORC JIT: LLVM's Just-In-Time compilation API, useful if you are building an interactive REPL or an engine requiring runtime code generation.