We present MIR-14, a minimal yet practical intermediate representation that evolves from an initial 27-instruction set through aggressive reduction to 13 instructions, then pragmatically adds back one essential operation (UnaryOp) for a final count of 14 core instructions. This evolution demonstrates the tension between theoretical minimalism and practical efficiency in compiler design.
Our key contributions are: (1) systematic instruction set evolution from Core-27 → Core-13 → Core-14 through empirical validation; (2) the BoxCall unification architecture that elegantly handles all data access patterns; (3) recognition that certain primitive operations (UnaryOp for negation and NOT) significantly improve both compilation efficiency and runtime performance; (4) optimization strategies including inline caching (33x speedup), AOT compilation, and typed array specialization that complement the minimal instruction set; (5) proof that the "Everything is Box" philosophy can be effectively realized at the IR level while maintaining practical performance.
Implementation results show that despite halving the instruction count, our benchmarks maintain performance within ±5% of the baseline while reducing MIR code size by 20-50%. The system successfully compiles complex applications including GUI programs, web servers, and distributed systems. This work demonstrates that IR minimalism, when coupled with strategic optimization placement, can achieve both extreme simplicity and production-level performance.
Our approach challenges the trend toward increasingly complex intermediate representations (e.g., LLVM's 60+ opcodes), showing that careful design can achieve more with less. We believe MIR-14 opens new possibilities for compiler construction, optimization research, and language implementation education.
