[llvm] [LoopVectorizer] Add support for partial reductions (PR #92418)

[Graham Hunter via llvm-commits]

================
@@ -6871,6 +6974,18 @@ void LoopVectorizationCostModel::collectValuesToIgnore() {
     const SmallVectorImpl<Instruction *> &Casts = IndDes.getCastInsts();
     VecValuesToIgnore.insert(Casts.begin(), Casts.end());
   }
+
+  // Ignore any values that we know will be flattened
+  for (auto It : getPartialReductionChains()) {
+    PartialReductionChain Chain = It.second;
+    SmallVector<Value *> PartialReductionValues{Chain.Reduction, Chain.BinOp,
+                                                Chain.ExtendA, Chain.ExtendB,
+                                                Chain.Accumulator};
+    ValuesToIgnore.insert(PartialReductionValues.begin(),
+                          PartialReductionValues.end());
+    VecValuesToIgnore.insert(PartialReductionValues.begin(),
----------------
huntergr-arm wrote:

I dislike adding these to ignore, since it means we discard the values for both determining max VF (which we want to do, since we're hoping for a higher VF) and cost modelling (which we'd prefer to keep at least some of, so we can more accurately compare the scalar vs. vector plans). I don't have a nice solution to this at present, so maybe it just requires the vplan-based cost model to be fully implemented so we can do modelling properly in a future patch.

https://github.com/llvm/llvm-project/pull/92418