Performance Analysis Tools
- PDF / 6,090,891 Bytes
- 245 Pages / 565.087 x 755.008 pts Page_size
- 13 Downloads / 213 Views
Synonyms Parallelization
Discussion Parafrase was a successful project allowing experimentation in source-to-source translation of FORTRAN programs. Parafrase was the work of David J. Kuck and his students at the University of Illinois at Urbana-Champaign. It was supported largely by NSF. The name of system is due to Stott Parker. A wide variety of optimization strategies were developed by researchers, and the performance achieved was measured using a broad cross section of applications, across a variety of theoretical and actual machines. Parafrase was successful because it reduced the effort required of the researcher to implement an optimization strategy, and to perform experiments. Parafrase was structured much like a multi-pass compiler of the same era. The FORTRAN program being compiled was processed one routine at a time. A scanner consumed the source code for a FORTRAN routine and built an equivalent representation in the Internal Language (IL). Then a series of passes that transformed the IL in various ways was run. The researcher would select the passes based on the goals of the research. Then lastly, a FORTRAN code regenerator pass was run. This entire process was repeated for each of the routines in the input program. The FORTRAN program output by Parafrase could be compiled by a FORTRAN compiler and run on any machine. Because the ultimate goal of Parafrase was to regenerate optimized FORTRAN code, the IL used by Parafrase was considerably different than the IL found in a compiler. The IL used by Parafrase was a relatively
straight forward representation of FORTRAN source, with some additional look-aside tables to hold summary information about variables and loop structures. Each pass was required to accept and generate this common IL. Having an IL that was very close to FORTRAN enabled Parafrase to provide an easily understandable IL dump by simply pretty printing the FORTRAN program that was represented in the data structures. The pretty printer could be invoked by Parafrase at the end of a pass, or called by the pass itself to display intermediate results. This made it easy for the researcher to see what an optimization was doing because the researcher only had to understand FORTRAN, not the details of some unusual IL. Parafrase also included an IL verification phase that could be enabled to run after any pass to ensure that that pass was following the rules. This was especially useful when identifying which pass was incorrectly processing a FORTRAN routine. Because all of the passes were required to accept and generate the common IL, the passes could be run in any order. This flexibility allowed the researcher to construct a specialized ordering of the passes to achieve specific goals. Very early in Parafrase’s life, the ability to specify the order of the passes in a text file was added. This enabled the researcher to adjust the order of the passes without having to rebuild the executable image of Parafrase. Parafrase was implemented in PL/I and run on an IBM System/ before virtual memory. Wi
Data Loading...
