[llvm] [SLP] Vectorize non-power-of-2 ops with padding. (PR #77790)

[Florian Hahn via llvm-commits]

fhahn wrote:

> This new node will be a burden to support, later the genric Vectorization will be extended to support non-power-of-2 nodes directly.

Agreed, it turns out the new node is not really needed and non-power-of-2 vector ops are handled quite well directly by backends (checked AArch64). Adjusted the patch to drop the special TreeEntry, tracking of NumPadding and generate non-power-of-2 vector ops directly for all nodes in the tree, WDYT as a starting point? https://github.com/llvm/llvm-project/commit/0ddbdd3052e694eedef6bccce3b17f92dff7add7

https://github.com/llvm/llvm-project/pull/77790