[llvm] [SLP] Emit reduction instead of 2 extracts + scalar op, when vectorizing operands (PR #147583)

[Alexey Bataev via llvm-commits]

================
@@ -23733,15 +23752,60 @@ bool SLPVectorizerPass::tryToVectorize(Instruction *I, BoUpSLP &R) {
       Candidates.emplace_back(A1, B);
   }
 
+  auto TryToReduce = [this, &R, &TTI = *TTI](Instruction *Inst,
+                                             ArrayRef<Value *> Ops) {
+    if (!isReductionCandidate(Inst))
+      return false;
+    Type *Ty = Inst->getType();
+    if (!isValidElementType(Ty) || Ty->isPointerTy())
+      return false;
+    HorizontalReduction HorRdx(Inst, Ops);
+    if (!HorRdx.matchReductionForOperands())
+      return false;
+    // Check the cost of operations.
+    VectorType *VecTy = getWidenedType(Ty, Ops.size());
+    constexpr TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
+    InstructionCost ScalarCost =
+        TTI.getScalarizationOverhead(
+            VecTy, APInt::getAllOnes(getNumElements(VecTy)), /*Insert=*/false,
+            /*Extract=*/true, CostKind) +
+        TTI.getInstructionCost(Inst, CostKind);
+    InstructionCost RedCost;
+    switch (::getRdxKind(Inst)) {
+    case RecurKind::Add:
+    case RecurKind::Mul:
+    case RecurKind::Or:
+    case RecurKind::And:
+    case RecurKind::Xor:
+    case RecurKind::FAdd:
+    case RecurKind::FMul: {
----------------
alexey-bataev wrote:

We do not vectorize it here for now, just binaryoperations and compares

https://github.com/llvm/llvm-project/pull/147583