[llvm] [VectorCombine] Refine cost model and decision logic in foldSelectShuffle (PR #146694)

[David Green via llvm-commits]

================
@@ -3174,6 +3174,54 @@ bool VectorCombine::foldCastFromReductions(Instruction &I) {
   return true;
 }
 
+/// Returns true if this ShuffleVectorInst eventually feeds into a
+/// vector reduction intrinsic (e.g., vector_reduce_add) by only following
+/// chains of shuffles and binary operators (in any combination/order).
+/// The search does not go deeper than the given Depth.
+static bool feedsIntoVectorReduction(ShuffleVectorInst *SVI, unsigned Depth) {
+  SmallPtrSet<Instruction *, 8> Visited;
+  SmallVector<std::pair<Instruction *, unsigned>, 4> WorkList;
+  bool FoundReduction = false;
+
+  WorkList.push_back({SVI, 0});
+  while (!WorkList.empty()) {
+    auto [I, CurDepth] = WorkList.pop_back_val();
+    if (CurDepth > Depth)
+      return false;
+
+    for (User *U : I->users()) {
+      auto *UI = dyn_cast<Instruction>(U);
+      if (!UI || !Visited.insert(UI).second)
----------------
davemgreen wrote:

I believe a User will always be an Instruction.

https://github.com/llvm/llvm-project/pull/146694