COMPILER TECHNIQUES FOR MATLAB PROGRAMS ... - CiteSeerX

A forward/backward traversal scheme for type inference is described in [ASU85]. In this
scheme, type inference is performed with data-ow analysis on a ow graph of the program.
The in and out set of variables for each block of the program are mapped onto sets of
possible types. The scheme uses an iterative process that propagates information forward
and backward, reducing the set of types associated with each variable until a xed point
is reached. One assumption of this scheme is that variables do not change types during
the execution of the program. In our case, this is not a valid assumption since variables
in MATLAB can change types during run-time thereby not always allowing a backward
step. Hence, our inference mechanism concentrates on a forward data-ow analysis and, as
discussed later, performs a backward step only when backward inference is possible.
SETL is a set-theoretically oriented language of very high-level. A SETL program maybe
considered to represent an algorithm before it is codied into a language of lower-level [Sch75].
It treats types in a fully dynamic way, withnotype declarations. However, the types of the
objects that appear in SETL programs can be deduced with the use of an appropriate global
analysis [Sch75]. For this type inference, a type algebra that operates on the structural type
of SETL objects is used. This algebra is implemented using tables whose entries describe
the action, on the symbolic entities of the algebra, of each of the primitives of the language
to be analyzed [Sch75].
2.2 Compilation of APL
APL is similar to MATLAB in that it can be executed interactively, is usually interpreted,
and operates on aggregate data structures. A few compilers for APL have beendeveloped
in the past. These compilers are also based on forward/backward dataow analysis.
Budd [Bud88] and Ching [Chi86] have independently developed compilers for APL.
Budd's compiler translates APL programs into C, while Ching's compiler produces IBM
System/370 assembly code directly. The motivation for their work was to investigate the
issues raised by the development of a compiler for a very high-level language, and to exploit
9