[SLP]Use proper order when calculating costs for geps/extracts to correctly identify profitability

Need to reorder properly the scalars, when evaluating the costs for the
external uses/geps to prevent differences in the calculating of the
profitability costs, used to choose between gather/compressed loads.

Fixes #132099 (comment)

Author: alexey-bataev

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -5491,12 +5491,16 @@ static bool isMaskedLoadCompress(
   const unsigned Sz = VL.size();
   auto *VecTy = getWidenedType(ScalarTy, Sz);
   constexpr TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
+  SmallVector<int> Mask;
+  if (!Order.empty())
+    inversePermutation(Order, Mask);
   // Check external uses.
   for (const auto [I, V] : enumerate(VL)) {
     if (AreAllUsersVectorized(V))
       continue;
     InstructionCost ExtractCost =
-        TTI.getVectorInstrCost(Instruction::ExtractElement, VecTy, CostKind, I);
+        TTI.getVectorInstrCost(Instruction::ExtractElement, VecTy, CostKind,
+                               Mask.empty() ? I : Mask[I]);
     InstructionCost ScalarCost =
         TTI.getInstructionCost(cast<Instruction>(V), CostKind);
     if (ExtractCost <= ScalarCost)
@@ -5536,8 +5540,11 @@ static bool isMaskedLoadCompress(
   bool IsStrided =
       buildCompressMask(PointerOps, Order, ScalarTy, DL, SE, CompressMask);
   assert(CompressMask.size() >= 2 && "At least two elements are required");
+  SmallVector<Value *> OrderedPointerOps(PointerOps);
+  if (!Order.empty())
+    reorderScalars(OrderedPointerOps, Mask);
   auto [ScalarGEPCost, VectorGEPCost] =
-      getGEPCosts(TTI, PointerOps, PointerOps.front(),
+      getGEPCosts(TTI, OrderedPointerOps, OrderedPointerOps.front(),
                   Instruction::GetElementPtr, CostKind, ScalarTy, LoadVecTy);
   // The cost of scalar loads.
   InstructionCost ScalarLoadsCost =
@@ -5564,17 +5571,16 @@ static bool isMaskedLoadCompress(
         TTI.getMemoryOpCost(Instruction::Load, LoadVecTy, CommonAlignment,
                             LI->getPointerAddressSpace(), CostKind);
   }
-  SmallVector<int> Mask;
-  if (!Order.empty())
-    inversePermutation(Order, Mask);
   if (IsStrided) {
     // Check for potential segmented(interleaved) loads.
     if (TTI.isLegalInterleavedAccessType(LoadVecTy, CompressMask[1],
                                          CommonAlignment,
                                          LI->getPointerAddressSpace())) {
-      InstructionCost InterleavedCost = TTI.getInterleavedMemoryOpCost(
-          Instruction::Load, LoadVecTy, CompressMask[1], std::nullopt,
-          CommonAlignment, LI->getPointerAddressSpace(), CostKind, IsMasked);
+      InstructionCost InterleavedCost =
+          VectorGEPCost + TTI.getInterleavedMemoryOpCost(
+                              Instruction::Load, LoadVecTy, CompressMask[1],
+                              std::nullopt, CommonAlignment,
+                              LI->getPointerAddressSpace(), CostKind, IsMasked);
       if (!Mask.empty())
         InterleavedCost += ::getShuffleCost(TTI, TTI::SK_PermuteSingleSrc,
                                             VecTy, Mask, CostKind);

llvm/test/Transforms/SLPVectorizer/AArch64/reordered-loads.ll

@@ -0,0 +1,29 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S --passes=slp-vectorizer -mtriple=aarch64-unknown-linux-gnu < %s -mcpu=neoverse-512tvb | FileCheck %s
+
+define i32 @test(ptr %0, i64 %1) vscale_range(2,2) {
+; CHECK-LABEL: define i32 @test(
+; CHECK-SAME: ptr [[TMP0:%.*]], i64 [[TMP1:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    br label %[[FOR_BODY48:.*]]
+; CHECK:       [[FOR_BODY48]]:
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr float, ptr [[TMP0]], i64 [[TMP1]]
+; CHECK-NEXT:    [[ARRAYIDX52:%.*]] = getelementptr i8, ptr [[TMP2]], i64 8
+; CHECK-NEXT:    [[TMP3:%.*]] = load float, ptr [[ARRAYIDX52]], align 4
+; CHECK-NEXT:    [[TMP4:%.*]] = load float, ptr [[TMP2]], align 4
+; CHECK-NEXT:    [[ADD56:%.*]] = fadd float [[TMP3]], [[TMP4]]
+; CHECK-NEXT:    store float [[ADD56]], ptr [[TMP0]], align 4
+; CHECK-NEXT:    br label %[[FOR_BODY48]]
+;
+entry:
+  br label %for.body48
+
+for.body48:
+  %2 = getelementptr float, ptr %0, i64 %1
+  %arrayidx52 = getelementptr i8, ptr %2, i64 8
+  %3 = load float, ptr %arrayidx52, align 4
+  %4 = load float, ptr %2, align 4
+  %add56 = fadd float %3, %4
+  store float %add56, ptr %0, align 4
+  br label %for.body48
+}