[llvm] [SLP] More accurately cost RISCV scalar splats (PR #213104)

[Alexey Bataev via llvm-commits]

================
@@ -22858,10 +22684,30 @@ ResTy BoUpSLP::processBuildVector(const TreeEntry *E, Type *ScalarTy,
     // Gather unique scalars and all constants.
     SmallVector<int> ReuseMask(GatheredScalars.size(), PoisonMaskElem);
     TryPackScalars(GatheredScalars, ReuseMask, /*IsRootPoison=*/true);
-    Value *BV = ShuffleBuilder.gather(GatheredScalars, ReuseMask.size());
-    ShuffleBuilder.add(BV, ReuseMask);
+    auto GatherUserOps = [&](SmallVectorImpl<TTI::BuildVectorUseOp> &UserOps) {
+      bool HasMatches = false;
+      for (const auto &TE : VectorizableTree) {
+        if (DeletedNodes.contains(TE.get()))
+          continue;
+        if (!(TE->isGather() && E->isSame(TE->Scalars)))
+          continue;
+        HasMatches = true;
+        auto *UserTE = TE->UserTreeIndex.UserTE;
+        if (!UserTE || !UserTE->hasState() || UserTE->isAltShuffle())
+          return false;
+        UserOps.push_back(
+            {UserTE->getOpcode(), static_cast<int>(TE->UserTreeIndex.EdgeIdx)});
----------------
alexey-bataev wrote:

```suggestion
        UserOps.emplace_back(
            UserTE->getOpcode(), TE->UserTreeIndex.EdgeIdx);
```

https://github.com/llvm/llvm-project/pull/213104