LLVM short comings regarding polyhedral and vectorization optimizations

Dear All;

Is there any work that discusses ​LLVM framework’s short comings regarding auto vectorization and polyhedral optimizations?



did you look at the Polly project:

It performs polyhedral optimizations on LLVM-IR and allows the LLVM
vectorizer to vectorize more loops by:
  - Creating alias free region with runtime-alias checks
  - Annotating parallel loops
  - Moving parallel dimensions to the innermost loop level

I added the Polly discussion list to the CC in case you want to know
more or start a discussion.

Best regards,