Alias Verification for Fortran Code Optimization
Thi Viet Nga Nguyen (Ecole des Mines de Paris, France)
Francois Irigoin (Ecole des Mines de Paris, France)
Abstract: Alias analysis for Fortran is less complicated than for programming languages with pointers but many real Fortran programs violate the standard: a formal parameter or a common variable that is aliased with another formal parameter is modified. Compilers, assuming standard-conforming programs, consider that an assignment to one variable will not change the value of any other variable, allowing optimizations involving the aliased variables. Higher performance results but anything may happen: the program may appear to run normally, or produce incorrect answers, or behave unpredictably. The results may depend on the compiler and the optimization level.
To guarantee the standard conformance of programs in order to make program analyses exact and program optimizations safe, precise alias information, i.e the determination of overlaps among arrays is studied in this paper. Static analyses and code instrumentation are used to find all violations of the aliasing rules in Fortran code. Alias violation tests are inserted only at places where it cannot be proved statically that they are useless in order to reduce the number of dynamic checks at run-time. A specific memory location naming technique is used to obtain compact representation and to enhance the precision of alias analysis. Modifications on the dependence graph created by aliasing are also studied to show the impact of aliases on the correctness of some program optimizing transformations. Experimental results on SPEC95 CFP benchmark are presented and some related issues are also discussed.
Keywords: alias, dummy aliasing, optimization, verification
Categories: D.2.4, D.2.5, D.3.4