Book Image

Learn LLVM 12

By : Kai Nacke
Book Image

Learn LLVM 12

By: Kai Nacke

Overview of this book

LLVM was built to bridge the gap between compiler textbooks and actual compiler development. It provides a modular codebase and advanced tools which help developers to build compilers easily. This book provides a practical introduction to LLVM, gradually helping you navigate through complex scenarios with ease when it comes to building and working with compilers. You’ll start by configuring, building, and installing LLVM libraries, tools, and external projects. Next, the book will introduce you to LLVM design and how it works in practice during each LLVM compiler stage: frontend, optimizer, and backend. Using a subset of a real programming language as an example, you will then learn how to develop a frontend and generate LLVM IR, hand it over to the optimization pipeline, and generate machine code from it. Later chapters will show you how to extend LLVM with a new pass and how instruction selection in LLVM works. You’ll also focus on Just-in-Time compilation issues and the current state of JIT-compilation support that LLVM provides, before finally going on to understand how to develop a new backend for LLVM. By the end of this LLVM book, you will have gained real-world experience in working with the LLVM compiler development framework with the help of hands-on examples and source code snippets.
Table of Contents (17 chapters)
Section 1 – The Basics of Compiler Construction with LLVM
Section 2 – From Source to Machine Code Generation
Section 3 –Taking LLVM to the Next Level

Understanding the LLVM target backend structure

After the LLVM IR is optimized, the selected LLVM target is used to generate the machine code from it. Among others, the following tasks are performed in the target backend:

  1. The directed acyclic graph (DAG) used for instruction selection, usually referred to as the SelectionDAG, is constructed.
  2. Machine instructions corresponding to the IR code are selected.
  3. The selected machine instructions are ordered in an optimal sequence.
  4. Virtual registers are replaced with machine registers.
  5. Prologue and epilogue code is added to functions.
  6. Basic blocks are ordered in an optimal sequence.
  7. Target-specific passes are run.
  8. Object code or assembly is emitted.

All these steps are implemented as machine function passes, derived from the MachineFunctionPass class. This is a subclass of the FunctionPass class, one of the base classes used by the old pass manager. As of LLVM 12, the conversion of machine function...