[SLP]Introduce transformNodes() and transform loads + reverse to strided loads.

Introduced transformNodes() function to perform transformation of the
nodes (cost-based, instruction count based, etc.).
Implemented transformation of consecutive loads + reverse order to
strided loads with stride -1, if profitable.

Reviewers: RKSimon, preames, topperc

Reviewed By: RKSimon

Pull Request: #88530

Author: alexey-bataev

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -1126,6 +1126,9 @@ class BoUpSLP {
   void
   buildExternalUses(const ExtraValueToDebugLocsMap &ExternallyUsedValues = {});
 
+  /// Transforms graph nodes to target specific representations, if profitable.
+  void transformNodes();
+
   /// Clear the internal data structures that are created by 'buildTree'.
   void deleteTree() {
     VectorizableTree.clear();
@@ -7813,6 +7816,43 @@ getGEPCosts(const TargetTransformInfo &TTI, ArrayRef<Value *> Ptrs,
   return std::make_pair(ScalarCost, VecCost);
 }
 
+void BoUpSLP::transformNodes() {
+  constexpr TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
+  for (std::unique_ptr<TreeEntry> &TE : VectorizableTree) {
+    TreeEntry &E = *TE.get();
+    switch (E.getOpcode()) {
+    case Instruction::Load: {
+      Type *ScalarTy = E.getMainOp()->getType();
+      auto *VecTy = FixedVectorType::get(ScalarTy, E.Scalars.size());
+      Align CommonAlignment = computeCommonAlignment<LoadInst>(E.Scalars);
+      // Check if profitable to represent consecutive load + reverse as strided
+      // load with stride -1.
+      if (isReverseOrder(E.ReorderIndices) &&
+          TTI->isLegalStridedLoadStore(VecTy, CommonAlignment)) {
+        SmallVector<int> Mask;
+        inversePermutation(E.ReorderIndices, Mask);
+        auto *BaseLI = cast<LoadInst>(E.Scalars.back());
+        InstructionCost OriginalVecCost =
+            TTI->getMemoryOpCost(Instruction::Load, VecTy, BaseLI->getAlign(),
+                                 BaseLI->getPointerAddressSpace(), CostKind,
+                                 TTI::OperandValueInfo()) +
+            ::getShuffleCost(*TTI, TTI::SK_Reverse, VecTy, Mask, CostKind);
+        InstructionCost StridedCost = TTI->getStridedMemoryOpCost(
+            Instruction::Load, VecTy, BaseLI->getPointerOperand(),
+            /*VariableMask=*/false, CommonAlignment, CostKind, BaseLI);
+        if (StridedCost < OriginalVecCost)
+          // Strided load is more profitable than consecutive load + reverse -
+          // transform the node to strided load.
+          E.State = TreeEntry::StridedVectorize;
+      }
+      break;
+    }
+    default:
+      break;
+    }
+  }
+}
+
 /// Merges shuffle masks and emits final shuffle instruction, if required. It
 /// supports shuffling of 2 input vectors. It implements lazy shuffles emission,
 /// when the actual shuffle instruction is generated only if this is actually
@@ -15189,6 +15229,7 @@ bool SLPVectorizerPass::vectorizeStoreChain(ArrayRef<Value *> Chain, BoUpSLP &R,
   R.buildExternalUses();
 
   R.computeMinimumValueSizes();
+  R.transformNodes();
 
   InstructionCost Cost = R.getTreeCost();
 
@@ -15567,6 +15608,7 @@ bool SLPVectorizerPass::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
       R.buildExternalUses();
 
       R.computeMinimumValueSizes();
+      R.transformNodes();
       InstructionCost Cost = R.getTreeCost();
       CandidateFound = true;
       MinCost = std::min(MinCost, Cost);
@@ -16563,6 +16605,7 @@ class HorizontalReduction {
         V.buildExternalUses(LocalExternallyUsedValues);
 
         V.computeMinimumValueSizes();
+        V.transformNodes();
 
         // Estimate cost.
         InstructionCost TreeCost = V.getTreeCost(VL);

llvm/test/Transforms/SLPVectorizer/RISCV/strided-loads-vectorized.ll

@@ -240,11 +240,10 @@ define void @test3(ptr %p, ptr noalias %s) {
 ; CHECK-LABEL: @test3(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [48 x float], ptr [[P:%.*]], i64 0, i64 0
+; CHECK-NEXT:    [[ARRAYIDX1:%.*]] = getelementptr inbounds [48 x float], ptr [[P]], i64 0, i64 30
 ; CHECK-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds float, ptr [[S:%.*]], i64 0
-; CHECK-NEXT:    [[ARRAYIDX48:%.*]] = getelementptr inbounds [48 x float], ptr [[P]], i64 0, i64 23
 ; CHECK-NEXT:    [[TMP0:%.*]] = call <8 x float> @llvm.experimental.vp.strided.load.v8f32.p0.i64(ptr align 4 [[ARRAYIDX]], i64 16, <8 x i1> <i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true>, i32 8)
-; CHECK-NEXT:    [[TMP1:%.*]] = load <8 x float>, ptr [[ARRAYIDX48]], align 4
-; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <8 x float> [[TMP1]], <8 x float> poison, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
+; CHECK-NEXT:    [[TMP2:%.*]] = call <8 x float> @llvm.experimental.vp.strided.load.v8f32.p0.i64(ptr align 4 [[ARRAYIDX1]], i64 -4, <8 x i1> <i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true>, i32 8)
 ; CHECK-NEXT:    [[TMP3:%.*]] = fsub fast <8 x float> [[TMP2]], [[TMP0]]
 ; CHECK-NEXT:    store <8 x float> [[TMP3]], ptr [[ARRAYIDX2]], align 4
 ; CHECK-NEXT:    ret void