Accéder directement au contenu Accéder directement à la navigation

Schedule-Sensitive Register Pressure Reduction in Innermost Loops, Basic Blocks and Super-Blocks

Abstract : This report makes a massive experimental study of an efficient heuristic for the SIRA framework \cite{sira04}. The heuristic, called SIRALINA \cite{siralina07}, bounds the register requirement of a data dependence graph before instruction scheduling under resource constraints. Our aim is to guarantee the absence of spilling before any instruction scheduling process, without hurting instruction level parallelism if possible. Our register pressure reduction methods are sensitive for both software pipelining (innermost loops) and acyclic scheduling (basic blocks and super-blocks). The SIRALINA method that we experiment in this report is shown efficient in terms of compilation times, in terms of register requirement reduction and in terms of shorted schedule increase. Our experiments are done on thousands standalone DDG extracted from FFMPEG, MEDIABENCH, SPEC2000 and SPEC2006 benchmarks. We consider processor architectures with multiple register type and we model delayed access times to registers. Our register pressure reduction method is distributed as a C independent library (\texttt{SIRAlib}.
Type de document :
Liste complète des métadonnées
Contributeur : Sid Touati <>
Soumis le : lundi 30 novembre 2009 - 16:19:36
Dernière modification le : vendredi 10 janvier 2020 - 15:42:20
Archivage à long terme le : : mardi 16 octobre 2012 - 14:55:50


  • HAL Id : inria-00436348, version 1




Sébastien Briais, Sid Touati. Schedule-Sensitive Register Pressure Reduction in Innermost Loops, Basic Blocks and Super-Blocks. [Research Report] 2009, pp.53. ⟨inria-00436348⟩



Consultations de la notice


Téléchargements de fichiers