Section: Software and Platforms

LAO Developments in Aggressive Compilation

Participants : Benoit Boissinot, Florent Bouchez, Florian Brandner, Quentin Colombet, Alain Darte, Benoît Dupont de Dinechin [Kalray] , Christophe Guillon [STMicroelectronics] , Sebastian Hack [Former post-doc in Compsys] , Fabrice Rastello, Cédric Vincent [Former student in Compsys] .

Our past aggressive optimization techniques are all implemented in stand-alone experimental tools (as for example for register coalescing algorithms) or within LAO, the back-end compiler of STMicroelectronics, or both. They concern SSA construction and destruction, instruction-cache optimizations, register allocation. Here, we report only our activities related to register allocation.

Our developments on register allocation within the STMicroelectronics compiler started when Cédric Vincent (bachelor degree, under Alain Darte supervision) developed a complete register allocator in LAO, the assembly-code optimizer of STMicroelectronics. This was the first time a complete implementation was done with success, outside the MCDT (now CEC) team, in their optimizer. This continued with developments made during the master internships and PhD theses of Florent Bouchez, Benoit Boissinot, and Quentin Colombet, and post-doctoral works of Sebastian Hack and Florian Brandner. In 2009, Quentin Colombet started to develop and integrate into the main trunk of LAO a full implementation of a two-phases register allocation. This implementation now includes two different decoupled spilling phases, the first one as described in Sebastian Hack's PhD thesis and a second ILP-based solution. It also includes an up-to-date graph-based register coalescing. Finally, since all these optimizations take place under SSA form, it includes also a mechanism for going out of colored-SSA (register-allocated SSA) form that can handle critical edges and does further optimizations. See details in the “new results” presented in previous Compsys activity reports.