[llvm] [SLP]Improve/fix subvectors in gather/buildvector nodes handling (PR #104144)

[Alexey Bataev via llvm-commits]

================
@@ -8361,8 +8367,49 @@ getGEPCosts(const TargetTransformInfo &TTI, ArrayRef<Value *> Ptrs,
 
 void BoUpSLP::transformNodes() {
   constexpr TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
-  for (std::unique_ptr<TreeEntry> &TE : VectorizableTree) {
-    TreeEntry &E = *TE;
+  // The tree may grow here, so iterate over nodes, built before.
+  for (unsigned Idx : seq<unsigned>(VectorizableTree.size())) {
+    TreeEntry &E = *VectorizableTree[Idx];
+    if (E.isGather()) {
+      ArrayRef<Value *> VL = E.Scalars;
+      const unsigned Sz = getVectorElementSize(VL.front());
+      unsigned MinVF = getMinVF(2 * Sz);
+      if (VL.size() <= 2 ||
+          (E.getOpcode() &&
+           (E.isAltShuffle() || E.getOpcode() != Instruction::Load)))
+        continue;
+      // Try to find vectorizable sequences and transform them into a serie of
+      // insertvector instructions.
+      unsigned StartIdx = 0;
+      unsigned End = VL.size();
+      for (unsigned VF = VL.size() / 2; VF >= MinVF; VF /= 2) {
+        for (unsigned Cnt = StartIdx; Cnt + VF <= End; Cnt += VF) {
+          ArrayRef<Value *> Slice = VL.slice(Cnt, VF);
+          InstructionsState S = getSameOpcode(Slice, *TLI);
+          if (!S.getOpcode() || S.isAltShuffle() ||
+              (S.getOpcode() != Instruction::Load &&
+               any_of(Slice, [&](Value *V) {
+                 return !areAllUsersVectorized(cast<Instruction>(V),
+                                               UserIgnoreList);
+               })))
+            continue;
+          if (!getTreeEntry(Slice.front()) && !getTreeEntry(Slice.back())) {
+            unsigned PrevSize = VectorizableTree.size();
+            buildTree_rec(Slice, 0, EdgeInfo(&E, UINT_MAX));
+            if (PrevSize + 1 == VectorizableTree.size() &&
+                VectorizableTree[PrevSize]->isGather()) {
+              VectorizableTree.pop_back();
----------------
alexey-bataev wrote:

Yes. The vector of nodes may only grow here, we need to iterate over the nodes that existed before this loop, so it is safe.

https://github.com/llvm/llvm-project/pull/104144