Liens transverses ENS de Lyon

Agenda de l'ENS de Lyon

Compiling Trees : Combining Data Layouts and the Polyhedral Model

lun 02 mai 2022



Amphi B


Soutenance de thèse de M. IANNETTA paul sous la Direction de Mme GONNORD Laure

Langue(s) des interventions

Description générale

The polyhedral model is a algebraic-based framework which enables efficient code optimization for computer-intensive programs which has been a prolific area of research since its inception. However, its strong assumptions about the shape of the control flow and memory accesses makes its application quite narrow in practice, even if many efforts have been done to extend it, mostly in the direction of more complex control-flow.
In this thesis, we propose to explore two research directions in order to deal with arbitrary control flow and memory pattern accesses, and to target other data structures than arrays.
Our first contribution is a semantic-based rephrasing of the framework that partially answers the first question and could sustain more in-depth research on the extension of the scope of the model to encompass "almost polyhedral programs". This contribution highlights the static program properties used by the polyhedral algorithms.
Our second contribution deals with algebraic data types, such as trees, which are, if not as common as arrays in scientific programs, ubiquitous in many algorithms. We propose a compilation scheme for algebraic data types laid out in memory according to a fixed layout. Memory movements are characterized "in the polyhedral model style" which enables optimized C code generation.
All in all, this thesis contributes to the on-going research of the polyhedral model on two points: by giving another point of view of polyhedral programs and by exploring how algebraic data types could reuse ideas from the framework.

Mots clés