Manipulating MAXLIVE for Spill-Free Register Allocation

Many embedded systems use single-chip microcontrollers which have no on-chip RAM. In such a system, the processor registers must hold all live data values. Nanocontrollers further reduce the controller circuit complexity so that a nanocontroller can be embedded with each of thousands to millions of sensors, actuators, or other devices on a single chip. This reduction in circuit complexity is accomplished by using a bit-serial multiplexorbased SIMD architecture with just tens of one-bit local registers. These registers not only must hold all declared and temporary values, but also are used to hold program state information in support of MIMD programmability. Implementing word-level operations using bit-serial multiplexor operations often yields huge basic blocks with very complex DAGs, apparently requiring even more registers. Spilling is not possible, so code that needs too many registers simply cannot be run. This paper explores new compilation methods, including Genetic Algorithms (GAs) and a new adaptation of SethiUllman numbering, to aggressively restructure the code and allocate registers so that the number of nanocontroller registers used does not exceed the number available. The approach also is shown to be adaptable to solve the less demanding problem of avoiding register spills for microcontrollers or general-purpose processors.