Scalar replacement transformation has been mainly used for improvement of loop balance as being combined with unroll-and-jam transformation[4, 5]. These researches considered the effect of register usage to improve loop balance, but did not consider t...
Scalar replacement transformation has been mainly used for improvement of loop balance as being combined with unroll-and-jam transformation[4, 5]. These researches considered the effect of register usage to improve loop balance, but did not consider the effect of data cache usage. This paper shows the effect of data cache usage by scalarization transformation, the simple form of scalar replacement transformation. We implemented scalarization transformation as the separate module in SUIF compiler system, and experimented the data cache simulation with SPEC CFP95. In the direct-mapped cache simulation, the data cache misses are reduced by 10.8% on average. In the reduction of these cache misses, the reduction of conflict misses is 19.7%, speciallay the reduction of tomcatv benchmark is 51.8%. The result shows the importance of the effect of data cache performance by scalarization or scalar replacement. Therefore, in researches using scalarization or scalar replacement tranformations, the researchers should consider the effect of data cache usage by these transformations.