[llvm] [SLP] Initial vectorization of non-power-of-2 ops. (PR #77790)

[Florian Hahn via llvm-commits]

================
@@ -14036,10 +14084,22 @@ bool SLPVectorizerPass::vectorizeStores(ArrayRef<StoreInst *> Stores,
 
       unsigned Sz = 1 + Log2_32(MaxVF) - Log2_32(MinVF);
       SmallVector<unsigned> CandidateVFs(Sz);
+      auto VFsToFill = make_range(CandidateVFs.begin(), CandidateVFs.end());
----------------
fhahn wrote:

We need to add the extra non-power-of-2 VF at the front so it gets tried first, so the for-each should only process 1..end elements. Could also insert at index 0, after for-each, but that would be more expensive.

The non-power-of-2 VF cannot be handled via the division, as it would compute incorrect VFs.

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