I’m currently in the process of choosing a project for a grad-level compiler course to be done over the next 8 weeks. I’d like to do something related to optimization since I haven’t worked much in that area before, but anything in the field is fair game.
What was the most interesting compiler-related project you’ve done? What did you learn the most from?
Edit: Thank you all for your great suggestions. I apologize for not updating this for so long.
The project I ended up doing was a simple autovectorization optimization on LLVM. LLVM has vector types, but there didn’t seem to be any way to take advantage of them without support for the front-end. This optimization converted normal scalar code into vector code.
Since auto-vectorization is a fairly difficult optimization to implement, we limited our scope as much as we could. First, in order to expose instruction level parallelism in the code, we looked for one-block loops that matched our criteria, then unrolled them a specific number of times so they would be conveniently vectorizable. We then implemented the packing algorithm laid out in Exploiting Superword Level Parallelism with Multimedia Instruction Sets by Larsen and Amarasinghe.
Even a simplified version of this optimization is pretty complicated. There are a lot of constraints; for instance, you don’t want to vectorize a variable that lives out of the loop, since the rest of the program expects it to be scalar. We put in a lot of hours in the last few weeks. The project was a lot of fun though, and we learned a lot.
If you’re interested in optimization, Vectorization of loops using SSE and MMX instruction sets could be interesting.