W. G. Griswold, D. Notkin,
"Architectural Tradeoffs for a Meaning-Preserving Program Restructuring Tool",
IEEE Transactions of Software Engineering, April, 1995.
See also an
on-line talk
on this subject.
Abstract
Maintaining the consistency of multiple program representations--such as
abstract syntax trees and program dependence graphs--in a program
manipulation tool is difficult. This paper describes a hybrid software
architecture for a meaning-preserving program restructuring tool. Layering
is the primary architectural paradigm, which successively provides
increasingly integrated and unified abstract machines to implement the
tool. However, layering does not provide adequate control
over extensibility or the independence of components. Consequently, we also
adopted the paradigm of keeping the key program abstractions separate
throughout the layering, providing independent ``columns'' of
abstract data types. A pair of columns is integrated by a
mapping column that translates elements in one column's data type
into related elements in the other column's data type. Thus
integration of function and separation of representation can be achieved
simultaneously in this complex domain.
This hybrid architecture was crucial in overcoming severe performance
problems--classic in traditional layered systems--that became apparent
once the basic tool was completed. By taking advantage
of the independence of the columns and the special characteristics of
meaning-preserving restructuring, it was possible to extend one representation
column of the architecture to the uppermost layer to provide the
required access for efficient update without compromising independence.
The cost of the extended architecture is that the upper layers are
no longer as simple because they expose operations
that only guarantee consistency under careful usage.
However, the structural constraints of the hybrid architecture and
provided models for building the more complicated layers minimizes the
negative impact of this tradeoff.