What's remaining open problem in polly?

[李拯先]

Hello everyone:

  I have little compiler development experimence and now learned some loop optimization, I'm interested in developping more improvement for polly. And I saw some open projects in polly.

  I'm interested in BLAS and Integrate Polly with the LLVM vectorizers.

  Where should i start from?

Yours,lzx

[Michael Kruse]

Better integration of Polly with LLVM's LoopVectorize pass sounds like a good idea. At moment, Polly has two vectorization "modes". With --polly-vectorizer=polly, Polly emits vector instructions itself. With -polly-vectorizer=stripmine, it only splits loops into vector-sized chunks.

What you could do:

1. Emit Loop metadata ('"llvm.loop.vectorize.enable") to the innermost loops generated by -polly-vectorizer=stripmine. [1] and [2,3]

2. Emit that loop metadata even if not stripmining for loops that might be vectorizable.

3. Polly's vectorization heuristic is kind of dumb, it always assumes a vector width of 4 by default. An improvement would be to ask the backend for the native vector width(s). Furthermore, it could use LoopVectorizer's heuristic as well to determine whether vectorization is profitable at all.

Michael

[1] https://github.com/llvm/llvm-project/blob/3cea2505fd8d99a9ba0cb625aecfe28a47c4e3f8/polly/lib/CodeGen/IslNodeBuilder.cpp#L425

[2] https://llvm.org/docs/LangRef.html#llvm-loop-vectorize-and-llvm-loop-interleave

[3] https://llvm.org/docs/TransformMetadata.html#loop-vectorization-and-interleaving