Intermediate representation

Intermediate representation (IR) is the data structure or format used internally by a compiler to represent source code between the front-end parsing phase and the back-end code generation phase. It is the lingua franca of compiler construction — a deliberately simplified language that strips away syntactic sugar and surface-level idiosyncrasies to expose the program's operational semantics in a form that optimization passes can manipulate. Every major compiler uses some form of IR: GCC uses GIMPLE and RTL, LLVM uses LLVM IR, and the Java compiler uses bytecode.

The design of an IR involves trade-offs between expressiveness and analyzability. A high-level IR preserves more of the source program's structure, making it easier to perform source-level optimizations and produce good error messages. A low-level IR exposes more machine details, enabling register allocation, instruction scheduling, and target-specific optimizations. Modern compilers typically use multiple IRs at different levels, transforming the program through a series of progressively lower-level representations. The SSA form is a common property enforced on IRs to simplify dataflow analysis.

The choice of IR is one of the most consequential design decisions in compiler architecture, yet it receives far less attention than parsing or optimization algorithms. A badly designed IR constrains every optimization that operates on it; a well-designed IR enables optimizations that the compiler writers did not anticipate. The IR is the hidden grammar of compilation — the language the compiler speaks to itself.