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

[Florian Hahn 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(),
----------------
fhahn wrote:

I might be missing something, but at least for NEON we seem to already consider VF = 16 for loops that perform dot-products without this change? https://godbolt.org/z/nqqM7Gxvv

It looks like its just that the lowering doesn't make use of dot products for those cases yet

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