[VectorCombine] Enable transform 'scalarizeLoadExtract' for scalable vector types

[benshi001]

Summary

The transform 'scalarizeLoadExtract' can be applied to scalable
vector types if the index is less than the minimum number of elements.

[benshi001]

llvm/test/Transforms/VectorCombine/AArch64/load-extractelement-scalarization.ll

97

For scalable vector types, only the index is guaranteed to be less than the minimum length, then this transform is allowed.

[benshi001]

llvm/test/Transforms/VectorCombine/AArch64/load-extractelement-scalarization.ll

268

This is not optimized due to there is no guaranteed that index is less than the minimum length of a scalable vector.

[benshi001]

I have created a GitHub PR as https://github.com/llvm/llvm-project/pull/65443

[nikic]

Abandon this and the parent revision, as they were migrated to GitHub?

llvm/lib/Transforms/Vectorize/VectorCombine.cpp

   if (!match(&I, m_Load(m_Value(Ptr))))
     return false;
 
-  auto *FixedVT = cast<FixedVectorType>(I.getType());
+   auto *VecTy = cast<VectorType>(I.getType());
   auto *LI = cast<LoadInst>(&I);
   const DataLayout &DL = I.getModule()->getDataLayout();
-  if (LI->isVolatile() || !DL.typeSizeEqualsStoreSize(FixedVT))
+  if (LI->isVolatile() || !DL.typeSizeEqualsStoreSize(VecTy))
     return false;
 
   InstructionCost OriginalCost =
-      TTI.getMemoryOpCost(Instruction::Load, FixedVT, LI->getAlign(),
+      TTI.getMemoryOpCost(Instruction::Load, VecTy, LI->getAlign(),
                           LI->getPointerAddressSpace());
   InstructionCost ScalarizedCost = 0;
 

       LastCheckedInst = UI;
     }
 
-    auto ScalarIdx = canScalarizeAccess(FixedVT, UI->getOperand(1), &I, AC, DT);
+    auto ScalarIdx = canScalarizeAccess(VecTy, UI->getOperand(1), &I, AC, DT);
     if (!ScalarIdx.isSafe()) {
       // TODO: Freeze index if it is safe to do so.
       ScalarIdx.discard();

     auto *Index = dyn_cast<ConstantInt>(UI->getOperand(1));
     TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
     OriginalCost +=
-        TTI.getVectorInstrCost(Instruction::ExtractElement, FixedVT, CostKind,
+        TTI.getVectorInstrCost(Instruction::ExtractElement, VecTy, CostKind,
                                Index ? Index->getZExtValue() : -1);
     ScalarizedCost +=
-        TTI.getMemoryOpCost(Instruction::Load, FixedVT->getElementType(),
+        TTI.getMemoryOpCost(Instruction::Load, VecTy->getElementType(),
                             Align(1), LI->getPointerAddressSpace());
-    ScalarizedCost += TTI.getAddressComputationCost(FixedVT->getElementType());
+    ScalarizedCost += TTI.getAddressComputationCost(VecTy->getElementType());
   }
 
   if (ScalarizedCost >= OriginalCost)

 
     Value *Idx = EI->getOperand(1);
     Value *GEP =
-        Builder.CreateInBoundsGEP(FixedVT, Ptr, {Builder.getInt32(0), Idx});
+        Builder.CreateInBoundsGEP(VecTy, Ptr, {Builder.getInt32(0), Idx});
     auto *NewLoad = cast<LoadInst>(Builder.CreateLoad(
-        FixedVT->getElementType(), GEP, EI->getName() + ".scalar"));
+        VecTy->getElementType(), GEP, EI->getName() + ".scalar"));
 
     Align ScalarOpAlignment = computeAlignmentAfterScalarization(
-        LI->getAlign(), FixedVT->getElementType(), Idx, DL);
+        LI->getAlign(), VecTy->getElementType(), Idx, DL);
     NewLoad->setAlignment(ScalarOpAlignment);
 
     replaceValue(*EI, *NewLoad);

       case Instruction::ShuffleVector:
         MadeChange |= widenSubvectorLoad(I);
         break;
-      case Instruction::Load:
-        MadeChange |= scalarizeLoadExtract(I);
-        break;
       default:
         break;
       }

     if (Opcode == Instruction::Store)
       MadeChange |= foldSingleElementStore(I);
 
+    if (isa<VectorType>(I.getType()) && Opcode == Instruction::Load)
+      MadeChange |= scalarizeLoadExtract(I);
 
     // If this is an early pipeline invocation of this pass, we are done.
     if (TryEarlyFoldsOnly)

llvm/test/Transforms/VectorCombine/AArch64/load-extractelement-scalarization.ll

 
 define i32 @load_extract_idx_0_vscale(ptr %x) {
 ; CHECK-LABEL: @load_extract_idx_0_vscale(
-; CHECK-NEXT:    [[LV:%.*]] = load <vscale x 4 x i32>, ptr [[X:%.*]], align 16
-; CHECK-NEXT:    [[R:%.*]] = extractelement <vscale x 4 x i32> [[LV]], i32 3
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds <vscale x 4 x i32>, ptr [[X:%.*]], i32 0, i32 3
+; CHECK-NEXT:    [[R:%.*]] = load i32, ptr [[TMP1]], align 4
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %lv = load <vscale x 4 x i32>, ptr %x

 
 define i32 @load_extract_idx_2_vscale(ptr %x) {
 ; CHECK-LABEL: @load_extract_idx_2_vscale(
-; CHECK-NEXT:    [[LV:%.*]] = load <vscale x 4 x i32>, ptr [[X:%.*]], align 16
-; CHECK-NEXT:    [[R:%.*]] = extractelement <vscale x 4 x i32> [[LV]], i32 2
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds <vscale x 4 x i32>, ptr [[X:%.*]], i32 0, i32 2
+; CHECK-NEXT:    [[R:%.*]] = load i32, ptr [[TMP1]], align 8
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %lv = load <vscale x 4 x i32>, ptr %x