Performance Advantage of the Register Stack in Intel® ItaniumTM Processors

The Intel® ItaniumTM architecture provides a virtual register stack of unlimited size for use by software. New virtual registers are allocated on a procedure call and deallocated on return. Itanium processors implement the register stack by means of a large physical register file, a mapping from virtual to physical registers, and a Register Stack Engine (RSE) that saves and restores the contents of the physical registers to memory without explicit program intervention. The combination of these features significantly reduces the number of loads and stores required to save registers across procedure calls compared to a conventional architecture. In this paper, we show that the Itanium register stack reduces load and store traffic to the stack by at least a factor of three across select SpecInt2000 and Oracle database benchmarks. Furthermore, we examine the effects of the register stack on data cache miss rates and program execution time. When compared to a conventional architecture, the Itanium architecture on average achieves 7%-8.3% and 10.2%-12% performance advantage on in-order and out-of-order processor models, respectively, as a result of the register stack. Finally we analyze the vitality of stack loads and show that in general few stack loads are vital in an in-order model. However, a larger percentage of stack loads become vital in the out-of-order model leading to a greater performance benefit from the register stack.