LLVM Basic Structure

LLVM is a modular compiler framework designed to separate the traditional compiler pipeline into Front-end, Middle-end, and Back-end, allowing each part to be independently replaced or extended.
For example, the C/C++ compiler front-end Clang parses C/C++ source code into LLVM IR (Intermediate Representation), which is then passed to the LLVM Optimizer (middle-end).
The Optimizer applies multiple optimization passes to transform the IR into a new optimized IR, which is then passed to the back-end to be emitted as assembly or binary code for the target machine.
Thanks to this structure, LLVM can be used by many different front-ends (e.g., Swift, Rust, Fortran) to translate into a common IR, apply rich optimizations, and then generate code for multiple architectures (x86, ARM, PowerPC, etc.).

The overall compilation flow of LLVM can be summarized as follows:

Front-end stage: The source code is lexed and parsed into an AST, which is then translated into LLVM IR. (e.g., Clang generates IR from C/C++ source code).
Middle-end (Optimizer) stage: The IR undergoes various optimization passes to improve performance. Since IR is hardware-independent, these optimizations are language- and platform-agnostic.
Back-end stage: The optimized IR is translated into target machine code. At this stage, the virtual registers and abstract instructions of the IR are mapped to the actual registers and instructions of the CPU.

LLVM IR acts as a hub connecting the front-end and back-end. Multiple front-ends can emit LLVM IR, and multiple back-ends can consume the same IR, enabling efficient support for many languages and platforms.
Front-end developers only need to translate source code into IR, while back-end developers only need to work with IR. This makes LLVM a powerful framework that provides both language independence and architecture independence.