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

[Alexey Bataev via llvm-commits]

================
@@ -13555,9 +13619,39 @@ bool SLPVectorizerPass::vectorizeStores(ArrayRef<StoreInst *> Stores,
                           << "MinVF (" << MinVF << ")\n");
       }
 
+      unsigned StartIdx = 0;
+      if (VectorizeNonPowerOf2) {
+        // Try vectorizing with a non-power-of-2 VF. At the moment, only
+        // consider cases where VF + 1 is a power-of-2, i.e. almost all vector
+        // lanes are used.
+        unsigned CandVF = Operands.size() + 1;
+        if (isPowerOf2_32(CandVF) && CandVF <= MaxVF) {
+          assert(
+              all_of(
+                  Operands,
+                  [&](Value *V) {
+                    return cast<StoreInst>(V)->getValueOperand()->getType() ==
+                           cast<StoreInst>(Operands.front())
+                               ->getValueOperand()
+                               ->getType();
+                  }) &&
+              "Expected all operands of same type.");
+          if (!VectorizedStores.count(Operands.front()) &&
+              !VectorizedStores.count(Operands.back()) &&
+              TriedSequences
+                  .insert(std::make_pair(Operands.front(), Operands.back()))
+                  .second &&
+              vectorizeStoreChain(Operands, R, Operands.size(), MinVF)) {
+            // Mark the vectorized stores so that we don't vectorize them again.
+            VectorizedStores.insert(Operands.begin(), Operands.end());
+            Changed = true;
+            StartIdx += Operands.size();
+          }
+        }
+      }
+
----------------
alexey-bataev wrote:

Can it be made the part of the loop? Maybe outline the body as lambda and adjust it properly to avoid code duplication

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