[llvm] [SLP]Initial support for non-power-of-2 (but still whole register) number of elements in operands. (PR #107273)

[Philip Reames via llvm-commits]

preames wrote:

I think I understand, let me state this back to you to make sure.

You're extending the non-power-of-two vectorization support with a different class of non-power-of-two vectors which may be profitable.  (The previous class was VL=2^N-1.)  The class being proposed is VL=VLMAX*N where VLMAX is the number of elements in a single vector register.  You're planning to both allow this class as the initial VL, and also adjust the search order to try VLMAX boundaries instead of reducing by powers of two.

This sounds reasonable.  Let me go look at the code with this in mind.



https://github.com/llvm/llvm-project/pull/107273