Bytewise Register Allocation

Abstract

Traditionally, variables have been considered as atoms by register allocation: Each variable was to be placed in one register, or spilt (placed in main memory) or rematerialized (recalculated as needed). Some flexibility arose from what would be considered a register: Register aliasing allowed to treat a register meant to hold a 16-bit variable as two registers that could hold an 8-bit variable each. We allow for far more lexibility in register allocation: We decide on the storage of variables bytewise, i. e. we decide for each individual byte in a variable whether to store it in memory or a register, and consider any byte of any register as a possible storage location. We implemented a backend for the STM8 architecture (STMicroelectronics' current 8-bit architecture) in the C-compiler sdcc, and experimentally evaluate the beneits of bytewise register allocation. The results show that bytewise register allocation can result in substantial improvements in the generated code. Optimizing for code size we obtained 27.2%, 13.2% and 9.2% reductions in code size in the Whetstone, Dhrystone and Coremark benchmarks, respectively, when using bytewise allocation and spilling compared to conventional allocation.