Citation Context

... that predication may be hindering application of DSWP in other benchmarks too. 5.3 Case Study: 179.art loop The source code of the179.art loop is shown in Figure 11. Performing accumulator expansion =-=[18]-=- on the summing variable (Y[yj].y) allows the compiler to obtain 14 SCCs, 7 up from the original, giving a 60% speedup with an IPC of 0.17 and 0.12 for the two cores. Accumulator expansion also improv...

Citation Context

...lation [10].Finally, the compiler should leverage modern transformations and analyses to avoid over-estimating dependences.These techniques range from traditional techniques like reduction expansions =-=[19, 21]-=- to more modern techniques,such as aggressive alias analysis [5], speculative pointer analysis [28], or variable value analysis [22]. Proving twomemory operations do not conflict or proving that a var...

Citation Context

...e the load instruction may not be speculated above the store in an attempt to improve the schedule. Using the same superblocks formed in Figure 1b) and applying superblock loop invariant code removal =-=[11]-=-, the load may be removed from the superblock loop BCE and placed in the loop preheader A, see Figure 2b). In this example, the superblock generated (BCE) with profile information avoids the hazard an...

Citation Context

... variable elimination, and global variable migration. Next, several profile-based code transformations are performed that increase the available instruction-level parallelism of the intermediate code =-=[17]-=- [20]. These optimizations increase the size of superblocks and eliminate data dependences between instructions. They are applied only to the most frequently executed superblocks to control code expan...

Citation Context

...es for optimizations which reduce the resource requirements and dependence height. The benefits of unrolling and these unrolling-based optimizations have been quantified for acyclicly scheduled loops =-=[10]-=-, but have not been measured for modulo scheduled loops. This paper describes the benefits of unrolling, and a set of optimizations for unrolled loops which have been implemented in the IMPACT compile...

Citation Context

...es for increasing the available ILP in programs have become increasingly important. One such code improvement technique is loop unrolling (LU) [9, 18] which in conjunction with register renaming (RR) =-=[3, 14]-=- can increase ILP. LU replicates the original loop body multiple times, adjusts the loop termination code and eliminates redundant branch instructions. The resulting larger basic block increases the p...