Efficient global register allocation

TL;DR

This paper introduces a lightweight register allocation algorithm that significantly reduces the need for complex liveness analysis, enabling more flexible and efficient register management in compilers.

Contribution

The novel algorithm uses a 'future-active' set to minimize liveness modeling, improving flexibility and efficiency over traditional methods like graph coloring and linear scan.

Findings

Removes over 90% of liveness analysis for instructions

Reduces memory and runtime costs of register allocation

Improves handling of fixed registers and loop values

Abstract

In a compiler, an essential component is the register allocator. Two main algorithms have dominated implementations, graph coloring and linear scan, differing in how live values are modeled. Graph coloring uses an edge in an `interference graph' to show that two values cannot reside in the same register. Linear scan numbers all values, creates intervals between definition and uses, and then intervals that do not overlap may be allocated to the same register. For both algorithms the liveness models are computed at considerable runtime and memory cost. Furthermore, these algorithms do little to improve code quality, where the target architecture and register coalescing are important concerns. We describe a new register allocation algorithm with lightweight implementation characteristics. The algorithm introduces a `future-active' set for values that will reside in a register later in the allocation. Registers are allocated and freed in the manner of linear scan, although other ordering heuristics could improve code quality or lower runtime cost. An advantageous property of the approach is an ability to make these trade-offs. A key result is the `future-active' set can remove any liveness model for over 90% of instructions and 80% of methods. The major contribution is the allocation algorithm that, for example, solves an inability of the similarly motivated Treescan register allocator to look ahead of the instruction being allocated - allowing an unconstrained allocation order, and an ability to better handle fixed registers and loop carried values. The approach also is not reliant on properties of SSA form, similar to the original linear scan work. An analysis is presented in a production compiler for Java code compiled through SSA form to Android dex files.