Benoit Perrot. Task graph and data dependencies. CSI Seminar The task system used in the Tiger Compiler is actually based on handwritten tasks's dependencies: for instance the developper, who knows that the "parse" task must be done before the "type-check" one, explicitely wrote a dependency between "parse" and "type-check".

A clever way to express thoose dependencies would be to inform the tasks of what they need (in term of instanciated/allocated/realized objects) and what they produce (idem). This concept is well illustrated in the "make" utility, where for instance, a binary program depends on its objects, and thoose objects depend on their sources.

Building such a graph would allow dynamical data analysis, particularly detection of "dead objects", i.e. objects that could be safely destroyed after use (avoiding memory dummy consumption or leak).

The challenge is here to express tasks and dependencies in terms of "resources" and "products" of "process", considering class hierarchies as manipulated entities.

Indeed, a simple abstract syntax tree is a resource of the "ast-display" task and also of the "type-check" one, but not of "hir-compute". But the product of "type-check", which is a "decorated" (or a derivated ?) abstract syntax tree is always a valid resource of "ast-display" and became a resource of "hir-compute"!

• perrot0503_tasks.pdf
• perrot0503_tasks_report.pdf

to top