← All projects

EECS 583 Advanced Compilers Winter 2024

Compiler Optimization through Loop Perforation

Team course project

Archived course report (PDF)

Loop perforation reduces computation by skipping some loop work, accepting a change in the result. This team project implemented two existing approaches in LLVM: skipping whole iterations and selectively skipping instructions within them. We compared execution time and numerical accuracy.

The evaluation used four small benchmarks: summation, nested loops, matrix multiplication, and convolution.

Workflow diagram from the report showing analysis and transformation steps for selective loop perforation.
Selective loop perforation (report, Figure 1): green marks computations selected for skipping; red marks pointer and loop-control instructions that remain.

Faster execution often came with substantial output error. The report evaluates both together, showing why a speedup alone is insufficient to judge an approximate compiler transformation.

Project team: Yuhang Ning, Chen Huang, Yuanzhe Zheng, Houming Chen, Yucheng Huo.