Content area
Abstract
A technique for representing programs abstractly and independently of the eventual target architecture is presented that yields a file representation twice as compact as machine code for a CISC processor. It forms the basis of an implementation, in which the process of code generation is deferred until the time of loading. At that point, native code is created on-the-fly by a code-generating loader.
The process of loading with dynamic code-generation is so fast that it requires little more time than the input of equivalent native code from a disk storage medium. This is predominantly due to the compactness of the abstract program representation, which allows to counterbalance the additional effort of code-generation by shorter input times. Since processor power is currently rising more rapidly than disk-access times and transfer rates are falling, the proposed technique is likely to become even more competitive as hardware technology evolves.
To users of the implemented system, working with modules in the abstract representation is as convenient: as working with native object-files. Both kinds of file-representation coexist in the implemented system; they are completely interchangeable and modules in either representation can import from the other kind. Separate compilation of program modules with type-safe interfaces, and dynamic loading on a per-module basis are both fully supported.
Deferring code-generation until loading time can provide several new capabilities, especially when the intermediate program representation is machine-independent and thereby portable. It is argued that the combination of portability with practicality denotes an important step toward a software-component industry. Further benefits include a potential for reducing the number of recompilations after changes in source text, and a mechanism to decide at load time whether or not run-time integrity checks should be generated for a library module, eliminating the need to distinguish between development and production library configurations. All of these new possibilities have the potential of lowering the cost of software development and maintenance.
In the long run, fast on-the-fly code-generation may even replace binary compatibility for achieving software portability among processors implementing the same architecture. Already today, different models of a processor family are diverging more and more and it is becoming increasingly difficult to serve all of them equally well with just one version of native code. If code is generated only at the time of loading, however, it can always be custom-tailored toward the specific processor that it will eventually run on.
