[llvm] [SLP] Buildvector for alternate instructions with non-profitable gather operands (PR #84978)

[via llvm-commits]

================
@@ -6056,6 +6056,100 @@ BoUpSLP::TreeEntry::EntryState BoUpSLP::getScalarsVectorizationState(
       LLVM_DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n");
       return TreeEntry::NeedToGather;
     }
+    // Check that the operand node does not generate buildvector sequence. If it
+    // is, then probably not worth it to build alternate shuffle, if number of
+    // buildvector operands + alternate instruction > than the number of
+    // buildvector instructions.
+    SmallVector<ValueList> Operands;
+    for (unsigned I : seq<unsigned>(0, VL0->getNumOperands())) {
+      Operands.emplace_back();
+      // Prepare the operand vector.
+      for (Value *V : VL)
+        Operands.back().push_back(cast<Instruction>(V)->getOperand(I));
+    }
+    if (Operands.size() == 2) {
+      // Try find best operands candidates.
+      for (unsigned I : seq<unsigned>(0, VL.size() - 1)) {
+        SmallVector<std::pair<Value *, Value *>> Candidates(3);
+        Candidates[0] = std::make_pair(Operands[0][I], Operands[0][I + 1]);
+        Candidates[1] = std::make_pair(Operands[0][I], Operands[1][I + 1]);
+        Candidates[2] = std::make_pair(Operands[1][I], Operands[0][I + 1]);
+        std::optional<int> Res = findBestRootPair(Candidates);
+        switch (Res.value_or(0)) {
+        case 0:
+          break;
+        case 1:
+          std::swap(Operands[0][I + 1], Operands[1][I + 1]);
+          break;
+        case 2:
+          std::swap(Operands[0][I], Operands[1][I]);
+          break;
+        default:
+          llvm_unreachable("Unexpected index.");
+        }
+      }
+    }
+    DenseSet<unsigned> UniqueOpcodes;
+    constexpr unsigned NumAltInsts = 3; // main + alt + shuffle.
+    unsigned NonInstCnt = 0;
+    unsigned UndefCnt = 0;
+    unsigned ExtraShuffleInsts = 0;
+    if (Operands.size() == 2) {
+      // Do not count same operands twice.
+      if (Operands.front() == Operands.back()) {
+        Operands.erase(Operands.begin());
+      } else if (!allConstant(Operands.front()) &&
+                 all_of(Operands.front(), [&](Value *V) {
+                   return is_contained(Operands.back(), V);
+                 })) {
+        Operands.erase(Operands.begin());
+        ++ExtraShuffleInsts;
+      }
+    }
+    const Loop *L = LI->getLoopFor(VL0->getParent());
+    if (any_of(Operands,
+               [&](ArrayRef<Value *> Op) {
+                 if (allConstant(Op) ||
+                     (!isSplat(Op) && allSameBlock(Op) && allSameType(Op) &&
+                      getSameOpcode(Op, *TLI).MainOp))
+                   return false;
+                 DenseMap<Value *, unsigned> Uniques;
----------------
goldsteinn wrote:

SmallPtrSet?

https://github.com/llvm/llvm-project/pull/84978