Embedded Software for SoC - Grupo de Mecatrônica EESC/USP

Control Flow Driven Splitting of Loop Nests 217
flow in the hot-spots of an application. Furthermore, accesses to memory are
reduced significantly since a large number of branching, arithmetic and logical
instructions and index variable accesses is removed.
The techniques presented in this article are fully automated. Since a loop
analysis is crucial for our techniques, a high-level intermediate representation
preserving loop information is required. Since this is not the case for
state-of-the-art optimizing compilers (e.g. Sun-IR [18], LANCE [12]), our
techniques are not realized as a compiler optimization but as a processor
independent source code transformation. This independence enables a very
detailed benchmarking using 10 different programmable processors.
A survey of work related to loop optimizations and source code transformations
is provided in section 2. Section 3 presents the analytical models
and algorithms for loop nest splitting. Section 4 shows the effects of our optimization
on pipelines and caches, runtimes, energy dissipation and code sizes.
Section 5 summarizes and concludes this article.
2. RELATED WORK
Loop transformations have been described in literature on compiler design for
many years (see e.g. [1, 19]) and are often integrated into today’s optimizing
compilers. Classical loop splitting (or loop distribution/fission) creates several
loops out of an original one and distributes the statements of the original
loop body among all loops. The main goal of this optimization is to enable