[llvm] [LV][AArch64] Provide option to use partial reductions by default (PR #216001)

[Benjamin Maxwell via llvm-commits]

================
@@ -5230,13 +5240,56 @@ void VPlanTransforms::createPartialReductions(VPlan &Plan,
   MapVector<VPReductionPHIRecipe *, SmallVector<VPPartialReductionChain>>
       ChainsByPhi;
   VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
+  SmallVector<VPReductionPHIRecipe *, 4> UnorderedReductions;
   for (VPRecipeBase &R : HeaderVPBB->phis()) {
     auto *RedPhiR = dyn_cast<VPReductionPHIRecipe>(&R);
     if (!RedPhiR)
       continue;
 
     if (auto Chains = getScaledReductions(RedPhiR))
       ChainsByPhi.try_emplace(RedPhiR, std::move(*Chains));
+    else if (UsePartialReduceByDefault &&
+             (RedPhiR->getRecurrenceKind() == RecurKind::Add ||
+              (RedPhiR->getRecurrenceKind() == RecurKind::FAdd &&
+               !RedPhiR->isOrdered() && !RedPhiR->isInLoop())))
+      UnorderedReductions.push_back(RedPhiR);
+  }
+
+  // For general unordered reductions which aren't part of a candidate
+  // chain for a scaled partial reduction, we can potentially still use
+  // the intrinsic to allow for more optimization later on.
+  for (auto *Rdx : UnorderedReductions) {
+    auto *Backedge = dyn_cast<VPWidenRecipe>(Rdx->getBackedgeValue());
+    VPValue *OtherOp;
+    if (!Backedge ||
+        !match(Backedge,
+               m_CombineOr(m_c_FAdd(m_Specific(Rdx), m_VPValue(OtherOp)),
+                           m_c_Add(m_Specific(Rdx), m_VPValue(OtherOp)))))
+      continue;
+
+    // If the target indicates that the intrinsic is as cheap as (or cheaper
+    // than) the add, then prefer the intrinsic.
+    if (!LoopVectorizationPlanner::getDecisionAndClampRange(
+            [&CostCtx, Rdx, Backedge](ElementCount VF) {
+              InstructionCost CurrentCost = Backedge->computeCost(VF, CostCtx);
+              Type *ScalarTy = Backedge->getScalarType();
+              InstructionCost PRCost = CostCtx.TTI.getPartialReductionCost(
+                  Backedge->getOpcode(), ScalarTy, ScalarTy, ScalarTy, VF,
----------------
MacDue wrote:

Based on the existing use of `getPartialReductionCost()`  (and docs), we only pass the `InputTypeB` when we have a binary operation. So I think this should just pass `ScalarTy` for the accumulator and `InputTypeA` (then nullptr for `InputTypeB`). 

https://github.com/llvm/llvm-project/pull/216001