A graph-based IR is an intermediate representation in which a program is modeled as a graph of operations and dependencies rather than as a linear sequence of instructions or a hierarchy of structured blocks. In this representation, nodes are operations, edges are data dependencies or control constraints, and the entire program exists as a single connected structure that the compiler traverses and rewrites. The Graal Compiler is the most prominent industrial example, but the idea traces back to the value dependence graph and sea-of-nodes representations developed in compiler research.
The graph-based approach eliminates the artificial distinction between high-level and low-level program representations. The same graph structure can represent a virtual method call, an arithmetic operation, or a machine instruction, with only the node type differing. This uniformity makes the compiler more extensible — new languages and new backends can be added by defining new node types — but it also makes the compiler more complex, since optimizations must reason about global graph properties rather than local instruction sequences.