Register Rematerialization In GCC

In most modern processor architectures, difference in data access time for values kept in registers as compared to those in memory is quite high. Thus, compilers should implement efficient register allocation strategies to improve the runtime performance of the code. Register rematerialization is a technique to improve register allocation effectively by improving spill code generation. It is often desirable to compute expressions at a “use” rather than use an earlier spilled value. Normally, “rematerializable” values are derived from registers that are live throughout the function. On register-starved architectures with with addressing modes supporting limited displacement, spilling values, which can be rematerialized, incurs an additional loss in performance due to instructions generated to fetch data from the frame. Hence, rematerialization aids in good usage of registers to give a good gain in execution performance of the code generated. Experimental results indicate a gain of 1-6% in code size and 1-4% improvement in execution performance.