diff --git a/llvm/lib/Target/AIE/AIEPostPipeliner.cpp b/llvm/lib/Target/AIE/AIEPostPipeliner.cpp
index 5cc062a2db73..9875df3be2d3 100644
--- a/llvm/lib/Target/AIE/AIEPostPipeliner.cpp
+++ b/llvm/lib/Target/AIE/AIEPostPipeliner.cpp
@@ -19,7 +19,8 @@
 #include <unordered_set>
 
 #define DEBUG_TYPE "postpipeliner"
-#define LLVM_SUMMARY(X) DEBUG_WITH_TYPE("postpipeliner-summary", X)
+#define DEBUG_SUMMARY(X) DEBUG_WITH_TYPE("postpipeliner-summary", X)
+#define DEBUG_FULL(X) DEBUG_WITH_TYPE("postpipeliner-full", X)
 
 namespace llvm::AIE {
 
@@ -209,6 +210,7 @@ void PostPipeliner::scheduleNode(SUnit &SU, int Cycle) {
     if (NewEarliest > Info[SNum].Earliest) {
       Info[SNum].LastEarliestPusher = SU.NodeNum;
       Info[SNum].Earliest = NewEarliest;
+      Info[SU.NodeNum].NumPushedEarliest++;
       LLVM_DEBUG(dbgs() << SNum << " to " << Info[SNum].Earliest << " -; ");
     }
   }
@@ -223,6 +225,7 @@ void PostPipeliner::scheduleNode(SUnit &SU, int Cycle) {
     if (NewLatest < Info[PNum].Latest) {
       Info[PNum].LastLatestPusher = SU.NodeNum;
       Info[PNum].Latest = NewLatest;
+      Info[SU.NodeNum].NumPushedLatest++;
       LLVM_DEBUG(dbgs() << PNum << " to - " << Info[PNum].Latest << "; ");
     }
   }
@@ -463,12 +466,16 @@ int PostPipeliner::mostUrgent(PostPipelinerStrategy &Strategy) {
   return Best;
 }
 
-void PostPipeliner::resetSchedule() {
+void PostPipeliner::resetSchedule(bool ResetCritical) {
   Scoreboard.clear();
   for (int K = 0; K < NTotalInstrs; K++) {
     auto &N = Info[K];
-    N.LastEarliestPusher = -1;
-    N.LastLatestPusher = -1;
+    if (ResetCritical) {
+      N.NumPushedEarliest = 0;
+      N.NumPushedLatest = 0;
+      N.LastEarliestPusher = -1;
+      N.LastLatestPusher = -1;
+    }
     if (K < NInstr) {
       N.Earliest = N.StaticEarliest;
       N.Latest = N.StaticLatest;
@@ -518,7 +525,7 @@ bool PostPipeliner::scheduleFirstIteration(PostPipelinerStrategy &Strategy) {
 
     scheduleNode(SU, Actual);
     Info[N].Scheduled = true;
-    LLVM_DEBUG(dbgs() << "Scoreboard\n"; Scoreboard.dumpFull(););
+    DEBUG_FULL(dbgs() << "Scoreboard\n"; Scoreboard.dumpFull(););
   }
   LLVM_DEBUG(dbgs() << "==== First iteration scheduled by " << Strategy.name()
                     << "====\n");
@@ -557,78 +564,83 @@ PostPipelinerStrategy::~PostPipelinerStrategy() {}
 
 class ConfigStrategy : public PostPipelinerStrategy {
 public:
-  enum class Modulation { OneWay, Alternate, Split, RandomSwing };
+  enum class Modulation { OneWay, Alternate };
 
 private:
   Modulation Mode = Modulation::OneWay;
   bool FromTop = true;
-
   uint64_t Param = 0;
-  uint64_t RandomState;
-  int Count = 0;
-
   std::set<int> SuccSiblingScheduled;
-
-  // Most trivial seeds lead to boring initial sequences,
-  // so spice them up a bit. The multiplication spreads the bits across the word
-  // and the addition makes 0 interesting. We may hit the one fixpoint value,
-  // but chances are really slim.
-  uint64_t spice(uint64_t Value) {
-    return Value * 0x1555555555555555ull + 0x123456789abcdefull;
-  }
-  // Maximum length 63 bits LFSR
-  int randomBit() {
-    const uint64_t BitIn = (RandomState >> 62) ^ (RandomState >> 61);
-    RandomState = (RandomState << 1) | (BitIn & 1);
-    return RandomState & 1;
-  }
+  std::function<bool(const SUnit &A, const SUnit &B)> Discriminators[8] = {
+      [&](const SUnit &A, const SUnit &B) {
+        auto &IA = Info[A.NodeNum];
+        auto &IB = Info[B.NodeNum];
+        return IA.NumPushedEarliest > IB.NumPushedEarliest;
+      },
+      [&](const SUnit &A, const SUnit &B) {
+        return SuccSiblingScheduled.count(A.NodeNum) >
+               SuccSiblingScheduled.count(B.NodeNum);
+      },
+      [&](const SUnit &A, const SUnit &B) {
+        auto &IA = Info[A.NodeNum];
+        auto &IB = Info[B.NodeNum];
+        return IA.LCDLatest < IB.LCDLatest;
+      },
+      [&](const SUnit &A, const SUnit &B) { return A.NodeNum < B.NodeNum; },
+      [&](const SUnit &A, const SUnit &B) {
+        auto &IA = Info[A.NodeNum];
+        auto &IB = Info[B.NodeNum];
+        return IA.NumPushedLatest > IB.NumPushedLatest;
+      },
+      [&](const SUnit &A, const SUnit &B) {
+        auto &IA = Info[A.NodeNum];
+        auto &IB = Info[B.NodeNum];
+        return IA.Earliest > IB.Earliest;
+      },
+      [&](const SUnit &A, const SUnit &B) { return false; },
+      [&](const SUnit &A, const SUnit &B) { return A.NodeNum > B.NodeNum; },
+  };
+  int Priority[3] = {0, 1, 2};
 
   bool better(const SUnit &A, const SUnit &B) override {
-    auto &IA = Info[A.NodeNum];
-    auto &IB = Info[B.NodeNum];
-    if (FromTop && SuccSiblingScheduled.count(A.NodeNum) >
-                       SuccSiblingScheduled.count(B.NodeNum)) {
-      return true;
-    }
-    if (FromTop) {
-      return IA.LCDLatest < IB.LCDLatest;
+    int PBias = FromTop ? 0 : 3;
+    for (auto P : Priority) {
+      if (Discriminators[P + PBias](A, B)) {
+        return true;
+      }
     }
-
-    return IA.Earliest > IB.Earliest;
+    return false;
   }
   bool fromTop() override { return FromTop; }
+
   void selected(const SUnit &N) override {
-    switch (Mode) {
-    case Modulation::Alternate:
-      FromTop = !FromTop;
-      break;
-    case Modulation::Split:
-      if (Count) {
-        Count--;
-        if (!Count) {
-          FromTop = !FromTop;
-        }
+    if (FromTop) {
+      // Promote the critical path
+      NodeInfo *Pushed = &Info[N.NodeNum];
+      while (Pushed->LastEarliestPusher >= 0) {
+        Pushed = &Info[Pushed->LastEarliestPusher];
+        Pushed->NumPushedEarliest++;
       }
-      break;
-    case Modulation::RandomSwing:
-      FromTop = randomBit();
-      break;
-    case Modulation::OneWay:
+
       // Promote my siblings
-      if (FromTop) {
-        for (auto &SDep : N.Succs) {
-          if (SDep.getKind() != SDep::Data) {
+      for (auto &SDep : N.Succs) {
+        if (SDep.getKind() != SDep::Data) {
+          continue;
+        }
+        for (auto &PDep : SDep.getSUnit()->Preds) {
+          if (PDep.getKind() != SDep::Data) {
             continue;
           }
-          for (auto &PDep : SDep.getSUnit()->Preds) {
-            if (PDep.getKind() != SDep::Data) {
-              continue;
-            }
-            SuccSiblingScheduled.insert(PDep.getSUnit()->NodeNum);
-          }
+          SuccSiblingScheduled.insert(PDep.getSUnit()->NodeNum);
         }
       }
-      break;
+    }
+    // else {
+    //  TODO : fill in !FromTop
+    // }
+
+    if (Mode == Modulation::Alternate) {
+      FromTop = !FromTop;
     }
   }
 
@@ -641,10 +653,10 @@ class ConfigStrategy : public PostPipelinerStrategy {
   ConfigStrategy(ScheduleDAGInstrs &DAG, std::vector<NodeInfo> &Info,
                  int Length, enum Modulation Mode, bool FromTop, uint64_t Param)
       : PostPipelinerStrategy(DAG, Info, Length), Mode(Mode), FromTop(FromTop),
-        Param(Param), RandomState(spice(Param)) {
-    if (Mode == Modulation::Split) {
-      Count = Length * Param / 1000;
-    }
+        Param(Param) {
+    Priority[0] = (Param >> 0) & 0x3;
+    Priority[1] = (Param >> 2) & 0x3;
+    Priority[2] = (Param >> 4) & 0x3;
   }
 };
 
@@ -665,21 +677,25 @@ bool PostPipeliner::tryHeuristics() {
     SearchVolume *= Node.Latest + MinLength - Node.Earliest + 1;
   }
 
-  LLVM_SUMMARY(dbgs() << "-- MinLength=" << MinLength
-                      << " SearchVolume=" << SearchVolume << "\n");
+  DEBUG_SUMMARY(dbgs() << "-- MinLength=" << MinLength
+                       << " SearchVolume=" << SearchVolume << "\n");
 
-  constexpr auto OneWay = ConfigStrategy::Modulation::OneWay;
-  constexpr auto Alternate = ConfigStrategy::Modulation::Alternate;
-  constexpr auto Split = ConfigStrategy::Modulation::Split;
+  // Critical, sibling, min latest
+  static const int P012 = 0x24;
+  // Original order
+  static const int POrg = 0x3f;
   static const std::tuple<int, ConfigStrategy::Modulation, bool, uint64_t>
       Configs[] = {
-          {0, OneWay, true, 0},     {0, OneWay, false, 0},
-          {1, OneWay, true, 0},     {1, OneWay, false, 0},
-          {0, Split, true, 300},    {0, Split, true, 600},
-          {0, Split, false, 300},   {0, Split, false, 600},
-          {1, Alternate, false, 0}, {1, OneWay, true, 0},
-          {1, OneWay, false, 0},    {1, Alternate, false, 0},
-
+          // Loosely speaking, a lower value of the first parameter targets
+          // a lower stage count, which benefits code size.
+          {0, ConfigStrategy::Modulation::OneWay, true, P012},
+          {0, ConfigStrategy::Modulation::OneWay, false, P012},
+          {0, ConfigStrategy::Modulation::OneWay, true, POrg},
+          {0, ConfigStrategy::Modulation::OneWay, false, POrg},
+          {1, ConfigStrategy::Modulation::OneWay, true, P012},
+          {1, ConfigStrategy::Modulation::OneWay, false, P012},
+          {1, ConfigStrategy::Modulation::OneWay, true, POrg},
+          {1, ConfigStrategy::Modulation::OneWay, false, POrg},
       };
   int H = 0;
   for (auto &[Extra, Mode, FromTop, Param] : Configs) {
@@ -688,16 +704,25 @@ bool PostPipeliner::tryHeuristics() {
       continue;
     }
     ConfigStrategy S(*DAG, Info, MinLength + Extra * II, Mode, FromTop, Param);
-    resetSchedule();
-    LLVM_SUMMARY(dbgs() << "--- Strategy " << S.name());
+    resetSchedule(/*ResetCritical=*/true);
+    DEBUG_SUMMARY(dbgs() << "--- Strategy " << S.name());
+    if (scheduleFirstIteration(S) && scheduleOtherIterations()) {
+      DEBUG_SUMMARY(dbgs() << " found II=" << II << "\n");
+      return true;
+    }
+
+    DEBUG_SUMMARY(dbgs() << " failed\n");
+    resetSchedule(/*ResetCritial=*/false);
+    DEBUG_SUMMARY(dbgs() << "--- Strategy " << S.name()
+                         << " with critical path");
     if (scheduleFirstIteration(S) && scheduleOtherIterations()) {
-      LLVM_SUMMARY(dbgs() << " found II=" << II << "\n");
+      DEBUG_SUMMARY(dbgs() << " found II=" << II << "\n");
       return true;
     }
-    LLVM_SUMMARY(dbgs() << " failed\n");
+    DEBUG_SUMMARY(dbgs() << " failed\n");
     H++;
   }
-  LLVM_SUMMARY(dbgs() << "=== II=" << II << " Failed ===\n");
+  DEBUG_SUMMARY(dbgs() << "=== II=" << II << " Failed ===\n");
   return false;
 }
 
diff --git a/llvm/lib/Target/AIE/AIEPostPipeliner.h b/llvm/lib/Target/AIE/AIEPostPipeliner.h
index 579f709dcffc..64318abe2620 100644
--- a/llvm/lib/Target/AIE/AIEPostPipeliner.h
+++ b/llvm/lib/Target/AIE/AIEPostPipeliner.h
@@ -93,8 +93,12 @@ class NodeInfo {
   int Latest = -1;
 
   // Record critical path components
+  // The Pred/Succ that pushed my Earliest/Latest
   int LastEarliestPusher = -1;
   int LastLatestPusher = -1;
+  // The number of Succs/Preds whose Earliest/Latest I have pushed.
+  int NumPushedEarliest = 0;
+  int NumPushedLatest = 0;
 
   // Latest corrected taking Earliest of an LCD successor into account
   int LCDLatest = -1;
@@ -196,8 +200,9 @@ class PostPipeliner {
   void computeForward();
   bool computeBackward();
 
-  /// Forget the previous round of scheduling
-  void resetSchedule();
+  /// Forget the previous round of scheduling. Also forget the critical path
+  /// if ResetCritical is set.
+  void resetSchedule(bool ResetCritical);
 
   /// Try all heuristics, stop at the first that fits the II
   /// If it returns true, a valid schedule is laid down in Info.
diff --git a/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm-1.mir b/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm-1.mir
index 305655f47d9a..2cb5d356493f 100644
--- a/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm-1.mir
+++ b/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm-1.mir
@@ -35,52 +35,50 @@
   ; CHECK-NEXT:    vldb wh9, [p1], m6; nopa ; padds [p0], #128; nopxm ; nopv
   ; CHECK-NEXT:    vldb wl10, [p0], #32; nopx
   ; CHECK-NEXT:    vldb wh10, [p0], #32
+  ; CHECK-NEXT:    vldb wl3, [p0], #32
+  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0
   ; CHECK-NEXT:    vldb wl5, [p1], m5
   ; CHECK-NEXT:    vldb wh5, [p1], m6
-  ; CHECK-NEXT:    vldb wl3, [p0], #32
-  ; CHECK-NEXT:    vldb wl7, [p1], m5
-  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0; vshuffle x11, x9, x9, r2
+  ; CHECK-NEXT:    vldb wl7, [p1], m5; vshuffle x11, x9, x9, r2
   ; CHECK-NEXT:    vldb wh7, [p1], m6
   ; CHECK-NEXT:    vldb wl3, [p1], m5; vshuffle x1, x8, x10, r4
   ; CHECK-NEXT:    vldb.3d wh3, [p1], d1; vshuffle x8, x8, x10, r16
-  ; CHECK-NEXT:    vshuffle x5, x5, x5, r2; vmac.f bmh0, bmh0, x1, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh1, bmh1, x8, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh4, bmh4, x1, x5, r3
-  ; CHECK-NEXT:    vshuffle x0, x6, x3, r4; vmac.f bmh5, bmh5, x8, x5, r3
-  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16
-  ; CHECK-NEXT:    vshuffle x7, x7, x7, r2; vmac.f bmh2, bmh2, x0, x11, r3
-  ; CHECK-NEXT:    vshuffle x3, x3, x3, r2; vmac.f bmh3, bmh3, x6, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh6, bmh6, x0, x5, r3
+  ; CHECK-NEXT:    vshuffle x0, x6, x3, r4; vmac.f bmh0, bmh0, x1, x11, r3
+  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16; vmac.f bmh1, bmh1, x8, x11, r3
+  ; CHECK-NEXT:    vshuffle x5, x5, x5, r2; vmac.f bmh2, bmh2, x0, x11, r3
+  ; CHECK-NEXT:    vmac.f bmh3, bmh3, x6, x11, r3
+  ; CHECK-NEXT:    vshuffle x7, x7, x7, r2; vmac.f bmh4, bmh4, x1, x5, r3
+  ; CHECK-NEXT:    vmac.f bmh5, bmh5, x8, x5, r3
+  ; CHECK-NEXT:    vshuffle x3, x3, x3, r2; vmac.f bmh6, bmh6, x0, x5, r3
   ; CHECK-NEXT:    .p2align 4
   ; CHECK-NEXT:  .LBB0_2: // %for.body
   ; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
-  ; CHECK-NEXT:    nopa ; vldb wl8, [p0, #0]; nopxm ; vmac.f bmh7, bmh7, x6, x5, r3
-  ; CHECK-NEXT:    vldb wh8, [p0, #32]; vmac.f bmh8, bmh8, x1, x7, r3
+  ; CHECK-NEXT:    vldb wl8, [p0, #0]; nopa ; nops ; nopxm ; vmac.f bmh7, bmh7, x6, x5, r3
+  ; CHECK-NEXT:    vldb wh8, [p0, #32]; nopx ; vmac.f bmh8, bmh8, x1, x7, r3
   ; CHECK-NEXT:    vldb wl6, [p0, #64]; vmac.f bml0, bml0, x8, x7, r3
   ; CHECK-NEXT:    vldb wh6, [p0, #96]; vmac.f bml1, bml1, x0, x7, r3
   ; CHECK-NEXT:    padds [p0], m4; vldb wl9, [p1], m5; vmac.f bml2, bml2, x6, x7, r3
   ; CHECK-NEXT:    padds [p0], #128; vldb wh9, [p1], m6; vmac.f bml3, bml3, x1, x3, r3
   ; CHECK-NEXT:    vldb wl10, [p0], #32; vmac.f bml5, bml5, x8, x3, r3
   ; CHECK-NEXT:    vldb wh10, [p0], #32; vmac.f bml6, bml6, x0, x3, r3
-  ; CHECK-NEXT:    vldb wl5, [p1], m5; vmac.f bml4, bml4, x6, x3, r3
+  ; CHECK-NEXT:    vldb wl3, [p0], #32; vmac.f bml4, bml4, x6, x3, r3
+  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0
+  ; CHECK-NEXT:    vldb wl5, [p1], m5
   ; CHECK-NEXT:    vldb wh5, [p1], m6
-  ; CHECK-NEXT:    vldb wl3, [p0], #32
-  ; CHECK-NEXT:    vldb wl7, [p1], m5
-  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0; vshuffle x11, x9, x9, r2
+  ; CHECK-NEXT:    vldb wl7, [p1], m5; vshuffle x11, x9, x9, r2
   ; CHECK-NEXT:    vldb wh7, [p1], m6
   ; CHECK-NEXT:    vldb wl3, [p1], m5; vshuffle x1, x8, x10, r4
   ; CHECK-NEXT:    vldb.3d wh3, [p1], d1; vshuffle x8, x8, x10, r16
-  ; CHECK-NEXT:    vshuffle x5, x5, x5, r2; vmac.f bmh0, bmh0, x1, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh1, bmh1, x8, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh4, bmh4, x1, x5, r3
-  ; CHECK-NEXT:    vshuffle x0, x6, x3, r4; vmac.f bmh5, bmh5, x8, x5, r3
-  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16
-  ; CHECK-NEXT:    vshuffle x7, x7, x7, r2; vmac.f bmh2, bmh2, x0, x11, r3
-  ; CHECK-NEXT:    vshuffle x3, x3, x3, r2; vmac.f bmh3, bmh3, x6, x11, r3
+  ; CHECK-NEXT:    vshuffle x0, x6, x3, r4; vmac.f bmh0, bmh0, x1, x11, r3
+  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16; vmac.f bmh1, bmh1, x8, x11, r3
+  ; CHECK-NEXT:    vshuffle x5, x5, x5, r2; vmac.f bmh2, bmh2, x0, x11, r3
+  ; CHECK-NEXT:    vmac.f bmh3, bmh3, x6, x11, r3
+  ; CHECK-NEXT:    vshuffle x7, x7, x7, r2; vmac.f bmh4, bmh4, x1, x5, r3
+  ; CHECK-NEXT:    vmac.f bmh5, bmh5, x8, x5, r3
   ; CHECK-NEXT:  .L_LEnd0:
-  ; CHECK-NEXT:    nopb ; nopa ; nops ; nopxm ; vmac.f bmh6, bmh6, x0, x5, r3
+  ; CHECK-NEXT:    nopb ; nopa ; nops ; nopx ; vshuffle x3, x3, x3, r2; vmac.f bmh6, bmh6, x0, x5, r3
   ; CHECK-NEXT:  // %bb.3: // %for.cond.cleanup
-  ; CHECK-NEXT:    nopb ; nopa ; nops ; nopxm ; vmac.f bmh7, bmh7, x6, x5, r3
+  ; CHECK-NEXT:    vmac.f bmh7, bmh7, x6, x5, r3
   ; CHECK-NEXT:    vmac.f bmh8, bmh8, x1, x7, r3
   ; CHECK-NEXT:    vmac.f bml0, bml0, x8, x7, r3
   ; CHECK-NEXT:    vmac.f bml1, bml1, x0, x7, r3
@@ -89,7 +87,6 @@
   ; CHECK-NEXT:    vmac.f bml5, bml5, x8, x3, r3
   ; CHECK-NEXT:    vmac.f bml6, bml6, x0, x3, r3
   ; CHECK-NEXT:    vmac.f bml4, bml4, x6, x3, r3
-  ; CHECK-NEXT:    nopx
   ; CHECK-NEXT:    nop
   ; CHECK-NEXT:    nop
   ; CHECK-NEXT:    nop
diff --git a/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm-2.mir b/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm-2.mir
index 2e943946aec3..90cf2d235f1c 100644
--- a/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm-2.mir
+++ b/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm-2.mir
@@ -27,60 +27,64 @@
   ; CHECK-NEXT:    add.nc lc, r0, #-1
   ; CHECK-NEXT:    movxm ls, #.LBB0_2
   ; CHECK-NEXT:    movxm le, #.L_LEnd0
+  ; CHECK-NEXT:    vldb wl9, [p1], m5; nopa ; nops ; nopxm ; nopv
+  ; CHECK-NEXT:    vldb wh9, [p1], m6; nopa ; nops ; nopxm ; nopv
+  ; CHECK-NEXT:    vldb wl5, [p1], m5; nopa ; nops ; nopxm ; nopv
+  ; CHECK-NEXT:    vldb wh5, [p1], m6; nopa ; nops ; nopxm ; nopv
+  ; CHECK-NEXT:    vldb wl7, [p1], m5; nopa ; nops ; nopxm ; nopv
+  ; CHECK-NEXT:    vldb wh7, [p1], m6; nopa ; nops ; nopxm ; nopv
   ; CHECK-NEXT:    vldb wl8, [p0, #0]; nopa ; nops ; nopxm ; nopv
-  ; CHECK-NEXT:    vldb wh8, [p0, #32]; nopa ; nops ; nopxm ; nopv
-  ; CHECK-NEXT:    vldb wl6, [p0, #64]; nopa ; nops ; nopxm ; nopv
-  ; CHECK-NEXT:    vldb wh6, [p0, #96]; nopa ; nops ; nopxm ; nopv
-  ; CHECK-NEXT:    vldb wl9, [p1], m5; nopa ; padds [p0], m4; nopxm ; nopv
-  ; CHECK-NEXT:    vldb wh9, [p1], m6; nopa ; padds [p0], #128; nopxm ; nopv
-  ; CHECK-NEXT:    vldb wl10, [p0], #32; nopx
-  ; CHECK-NEXT:    vldb wh10, [p0], #32
-  ; CHECK-NEXT:    vldb wl5, [p1], m5
-  ; CHECK-NEXT:    vldb wh5, [p1], m6
+  ; CHECK-NEXT:    vldb wh8, [p0, #32]; nopx
+  ; CHECK-NEXT:    vldb wl6, [p0, #64]; vshuffle x11, x9, x9, r2
+  ; CHECK-NEXT:    vldb wh6, [p0, #96]; padds [p0], m4
+  ; CHECK-NEXT:    padds [p0], #128; vshuffle x5, x5, x5, r2
+  ; CHECK-NEXT:    vldb wl10, [p0], #32
+  ; CHECK-NEXT:    vldb wh10, [p0], #32; vshuffle x7, x7, x7, r2
   ; CHECK-NEXT:    vldb wl3, [p0], #32
-  ; CHECK-NEXT:    vldb wl7, [p1], m5
-  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0; vshuffle x11, x9, x9, r2
-  ; CHECK-NEXT:    vldb wh7, [p1], m6
-  ; CHECK-NEXT:    vldb wl3, [p1], m5; vshuffle x1, x8, x10, r4
-  ; CHECK-NEXT:    vldb.3d wh3, [p1], d1; vshuffle x0, x8, x10, r16
-  ; CHECK-NEXT:    vshuffle x5, x5, x5, r2; vmac.f bmh0, bmh0, x1, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh1, bmh1, x0, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh4, bmh4, x1, x5, r3
-  ; CHECK-NEXT:    vshuffle x10, x6, x3, r4; vmac.f bmh5, bmh5, x0, x5, r3
-  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16
-  ; CHECK-NEXT:    vshuffle x7, x7, x7, r2; vmac.f bmh2, bmh2, x10, x11, r3
-  ; CHECK-NEXT:    vshuffle x3, x3, x3, r2; vmac.f bmh3, bmh3, x6, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh6, bmh6, x10, x5, r3
+  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0
+  ; CHECK-NEXT:    nop
+  ; CHECK-NEXT:    vldb wl3, [p1], m5
+  ; CHECK-NEXT:    vldb.3d wh3, [p1], d1
+  ; CHECK-NEXT:    nop
+  ; CHECK-NEXT:    vshuffle x1, x8, x10, r4
+  ; CHECK-NEXT:    vshuffle x0, x8, x10, r16
+  ; CHECK-NEXT:    vshuffle x10, x6, x3, r4; vmac.f bmh0, bmh0, x1, x11, r3
+  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16; vmac.f bmh1, bmh1, x0, x11, r3
   ; CHECK-NEXT:    .p2align 4
   ; CHECK-NEXT:  .LBB0_2: // %for.body
   ; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
-  ; CHECK-NEXT:    nopa ; vldb wl8, [p0, #0]; nopxm ; vmac.f bmh7, bmh7, x6, x5, r3
+  ; CHECK-NEXT:    nopa ; vldb wl9, [p1], m5; nopxm ; vmac.f bmh2, bmh2, x10, x11, r3
+  ; CHECK-NEXT:    vldb wh9, [p1], m6; vshuffle x3, x3, x3, r2; vmac.f bmh3, bmh3, x6, x11, r3
+  ; CHECK-NEXT:    vldb wl5, [p1], m5; vmac.f bmh4, bmh4, x1, x5, r3
+  ; CHECK-NEXT:    vldb wh5, [p1], m6; vmac.f bml4, bml4, x6, x3, r3
+  ; CHECK-NEXT:    vldb wl7, [p1], m5; vmac.f bmh5, bmh5, x0, x5, r3
+  ; CHECK-NEXT:    vldb wh7, [p1], m6; vmac.f bmh6, bmh6, x10, x5, r3
+  ; CHECK-NEXT:    vldb wl8, [p0, #0]; vmac.f bmh7, bmh7, x6, x5, r3
   ; CHECK-NEXT:    vldb wh8, [p0, #32]; vmac.f bmh8, bmh8, x1, x7, r3
-  ; CHECK-NEXT:    vldb wl6, [p0, #64]; vmac.f bml0, bml0, x0, x7, r3
-  ; CHECK-NEXT:    vldb wh6, [p0, #96]; vmac.f bml1, bml1, x10, x7, r3
-  ; CHECK-NEXT:    padds [p0], m4; vldb wl9, [p1], m5; vmac.f bml2, bml2, x6, x7, r3
-  ; CHECK-NEXT:    padds [p0], #128; vldb wh9, [p1], m6; vmac.f bml3, bml3, x1, x3, r3
-  ; CHECK-NEXT:    vldb wl10, [p0], #32; vmac.f bml5, bml5, x0, x3, r3
-  ; CHECK-NEXT:    vldb wh10, [p0], #32; vmac.f bml6, bml6, x10, x3, r3
-  ; CHECK-NEXT:    vldb wl5, [p1], m5; vmac.f bml4, bml4, x6, x3, r3
-  ; CHECK-NEXT:    vldb wh5, [p1], m6
-  ; CHECK-NEXT:    vldb wl3, [p0], #32
-  ; CHECK-NEXT:    vldb wl7, [p1], m5
-  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0; vshuffle x11, x9, x9, r2
-  ; CHECK-NEXT:    vldb wh7, [p1], m6
-  ; CHECK-NEXT:    vldb wl3, [p1], m5; vshuffle x1, x8, x10, r4
-  ; CHECK-NEXT:    vldb.3d wh3, [p1], d1; vshuffle x0, x8, x10, r16
-  ; CHECK-NEXT:    vshuffle x5, x5, x5, r2; vmac.f bmh0, bmh0, x1, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh1, bmh1, x0, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh4, bmh4, x1, x5, r3
-  ; CHECK-NEXT:    vshuffle x10, x6, x3, r4; vmac.f bmh5, bmh5, x0, x5, r3
-  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16
-  ; CHECK-NEXT:    vshuffle x7, x7, x7, r2; vmac.f bmh2, bmh2, x10, x11, r3
-  ; CHECK-NEXT:    vshuffle x3, x3, x3, r2; vmac.f bmh3, bmh3, x6, x11, r3
+  ; CHECK-NEXT:    vldb wl6, [p0, #64]; vshuffle x11, x9, x9, r2; vmac.f bml0, bml0, x0, x7, r3
+  ; CHECK-NEXT:    padds [p0], m4; vldb wh6, [p0, #96]; vmac.f bml1, bml1, x10, x7, r3
+  ; CHECK-NEXT:    padds [p0], #128; vshuffle x5, x5, x5, r2; vmac.f bml2, bml2, x6, x7, r3
+  ; CHECK-NEXT:    vldb wl10, [p0], #32; vmac.f bml3, bml3, x1, x3, r3
+  ; CHECK-NEXT:    vldb wh10, [p0], #32; vshuffle x7, x7, x7, r2; vmac.f bml5, bml5, x0, x3, r3
+  ; CHECK-NEXT:    vldb wl3, [p0], #32; vmac.f bml6, bml6, x10, x3, r3
+  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0
+  ; CHECK-NEXT:    nop
+  ; CHECK-NEXT:    vldb wl3, [p1], m5
+  ; CHECK-NEXT:    vldb.3d wh3, [p1], d1
+  ; CHECK-NEXT:    nop
+  ; CHECK-NEXT:    vshuffle x1, x8, x10, r4
+  ; CHECK-NEXT:    vshuffle x0, x8, x10, r16
+  ; CHECK-NEXT:    vshuffle x10, x6, x3, r4; vmac.f bmh0, bmh0, x1, x11, r3
   ; CHECK-NEXT:  .L_LEnd0:
-  ; CHECK-NEXT:    nopb ; nopa ; nops ; nopxm ; vmac.f bmh6, bmh6, x10, x5, r3
+  ; CHECK-NEXT:    nopb ; nopa ; nops ; nopx ; vshuffle x6, x6, x3, r16; vmac.f bmh1, bmh1, x0, x11, r3
   ; CHECK-NEXT:  // %bb.3: // %for.cond.cleanup
-  ; CHECK-NEXT:    nopb ; nopa ; nops ; nopxm ; vmac.f bmh7, bmh7, x6, x5, r3
+  ; CHECK-NEXT:    vmac.f bmh2, bmh2, x10, x11, r3
+  ; CHECK-NEXT:    nopx ; vshuffle x3, x3, x3, r2; vmac.f bmh3, bmh3, x6, x11, r3
+  ; CHECK-NEXT:    vmac.f bmh4, bmh4, x1, x5, r3
+  ; CHECK-NEXT:    vmac.f bml4, bml4, x6, x3, r3
+  ; CHECK-NEXT:    vmac.f bmh5, bmh5, x0, x5, r3
+  ; CHECK-NEXT:    vmac.f bmh6, bmh6, x10, x5, r3
+  ; CHECK-NEXT:    vmac.f bmh7, bmh7, x6, x5, r3
   ; CHECK-NEXT:    vmac.f bmh8, bmh8, x1, x7, r3
   ; CHECK-NEXT:    vmac.f bml0, bml0, x0, x7, r3
   ; CHECK-NEXT:    vmac.f bml1, bml1, x10, x7, r3
@@ -88,13 +92,6 @@
   ; CHECK-NEXT:    vmac.f bml3, bml3, x1, x3, r3
   ; CHECK-NEXT:    vmac.f bml5, bml5, x0, x3, r3
   ; CHECK-NEXT:    vmac.f bml6, bml6, x10, x3, r3
-  ; CHECK-NEXT:    vmac.f bml4, bml4, x6, x3, r3
-  ; CHECK-NEXT:    nopx
-  ; CHECK-NEXT:    nop
-  ; CHECK-NEXT:    nop
-  ; CHECK-NEXT:    nop
-  ; CHECK-NEXT:    nop
-  ; CHECK-NEXT:    nop
   ; CHECK-NEXT:    nop
   ; CHECK-NEXT:    nop
   ; CHECK-NEXT:    nop
diff --git a/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm-3.mir b/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm-3.mir
index af9e1589e96a..8dec880a39b1 100644
--- a/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm-3.mir
+++ b/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm-3.mir
@@ -27,60 +27,64 @@
   ; CHECK-NEXT:    add.nc lc, r0, #-1
   ; CHECK-NEXT:    movxm ls, #.LBB0_2
   ; CHECK-NEXT:    movxm le, #.L_LEnd0
+  ; CHECK-NEXT:    vldb wl9, [p1], m5; nopa ; nops ; nopxm ; nopv
+  ; CHECK-NEXT:    vldb wh9, [p1], m6; nopa ; nops ; nopxm ; nopv
+  ; CHECK-NEXT:    vldb wl5, [p1], m5; nopa ; nops ; nopxm ; nopv
+  ; CHECK-NEXT:    vldb wh5, [p1], m6; nopa ; nops ; nopxm ; nopv
+  ; CHECK-NEXT:    vldb wl7, [p1], m5; nopa ; nops ; nopxm ; nopv
+  ; CHECK-NEXT:    vldb wh7, [p1], m6; nopa ; nops ; nopxm ; nopv
   ; CHECK-NEXT:    vldb wl8, [p0, #0]; nopa ; nops ; nopxm ; nopv
-  ; CHECK-NEXT:    vldb wh8, [p0, #32]; nopa ; nops ; nopxm ; nopv
-  ; CHECK-NEXT:    vldb wl6, [p0, #64]; nopa ; nops ; nopxm ; nopv
-  ; CHECK-NEXT:    vldb wh6, [p0, #96]; nopa ; nops ; nopxm ; nopv
-  ; CHECK-NEXT:    vldb wl9, [p1], m5; nopa ; padds [p0], m4; nopxm ; nopv
-  ; CHECK-NEXT:    vldb wh9, [p1], m6; nopa ; padds [p0], #128; nopxm ; nopv
-  ; CHECK-NEXT:    vldb wl10, [p0], #32; nopx
-  ; CHECK-NEXT:    vldb wh10, [p0], #32
-  ; CHECK-NEXT:    vldb wl5, [p1], m5
-  ; CHECK-NEXT:    vldb wh5, [p1], m6
+  ; CHECK-NEXT:    vldb wh8, [p0, #32]; nopx
+  ; CHECK-NEXT:    vldb wl6, [p0, #64]; vshuffle x11, x9, x9, r2
+  ; CHECK-NEXT:    vldb wh6, [p0, #96]; padds [p0], m4
+  ; CHECK-NEXT:    padds [p0], #128; vshuffle x5, x5, x5, r2
+  ; CHECK-NEXT:    vldb wl10, [p0], #32
+  ; CHECK-NEXT:    vldb wh10, [p0], #32; vshuffle x7, x7, x7, r2
   ; CHECK-NEXT:    vldb wl3, [p0], #32
-  ; CHECK-NEXT:    vldb wl7, [p1], m5
-  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0; vshuffle x11, x9, x9, r2
-  ; CHECK-NEXT:    vldb wh7, [p1], m6
-  ; CHECK-NEXT:    vldb wl3, [p1], m5; vshuffle x1, x8, x10, r4
-  ; CHECK-NEXT:    vldb.3d wh3, [p1], d1; vshuffle x2, x2, x10, r16
-  ; CHECK-NEXT:    vshuffle x5, x5, x5, r2; vmac.f bmh0, bmh0, x1, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh1, bmh1, x2, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh4, bmh4, x1, x5, r3
-  ; CHECK-NEXT:    vshuffle x0, x6, x3, r4; vmac.f bmh5, bmh5, x2, x5, r3
-  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16
-  ; CHECK-NEXT:    vshuffle x7, x7, x7, r2; vmac.f bmh2, bmh2, x0, x11, r3
-  ; CHECK-NEXT:    vshuffle x3, x3, x3, r2; vmac.f bmh3, bmh3, x6, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh6, bmh6, x0, x5, r3
+  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0
+  ; CHECK-NEXT:    nop
+  ; CHECK-NEXT:    vldb wl3, [p1], m5
+  ; CHECK-NEXT:    vldb.3d wh3, [p1], d1
+  ; CHECK-NEXT:    nop
+  ; CHECK-NEXT:    vshuffle x2, x2, x10, r16
+  ; CHECK-NEXT:    vshuffle x1, x8, x10, r4
+  ; CHECK-NEXT:    vshuffle x0, x6, x3, r4; vmac.f bmh1, bmh1, x2, x11, r3
+  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16; vmac.f bmh0, bmh0, x1, x11, r3
   ; CHECK-NEXT:    .p2align 4
   ; CHECK-NEXT:  .LBB0_2: // %for.body
   ; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
-  ; CHECK-NEXT:    nopa ; vldb wl8, [p0, #0]; nopxm ; vmac.f bmh7, bmh7, x6, x5, r3
+  ; CHECK-NEXT:    nopa ; vldb wl9, [p1], m5; nopxm ; vmac.f bmh2, bmh2, x0, x11, r3
+  ; CHECK-NEXT:    vldb wh9, [p1], m6; vshuffle x3, x3, x3, r2; vmac.f bmh3, bmh3, x6, x11, r3
+  ; CHECK-NEXT:    vldb wl5, [p1], m5; vmac.f bmh4, bmh4, x1, x5, r3
+  ; CHECK-NEXT:    vldb wh5, [p1], m6; vmac.f bml4, bml4, x6, x3, r3
+  ; CHECK-NEXT:    vldb wl7, [p1], m5; vmac.f bmh5, bmh5, x2, x5, r3
+  ; CHECK-NEXT:    vldb wh7, [p1], m6; vmac.f bmh6, bmh6, x0, x5, r3
+  ; CHECK-NEXT:    vldb wl8, [p0, #0]; vmac.f bmh7, bmh7, x6, x5, r3
   ; CHECK-NEXT:    vldb wh8, [p0, #32]; vmac.f bmh8, bmh8, x1, x7, r3
-  ; CHECK-NEXT:    vldb wl6, [p0, #64]; vmac.f bml0, bml0, x2, x7, r3
-  ; CHECK-NEXT:    vldb wh6, [p0, #96]; vmac.f bml1, bml1, x0, x7, r3
-  ; CHECK-NEXT:    padds [p0], m4; vldb wl9, [p1], m5; vmac.f bml2, bml2, x6, x7, r3
-  ; CHECK-NEXT:    padds [p0], #128; vldb wh9, [p1], m6; vmac.f bml3, bml3, x1, x3, r3
-  ; CHECK-NEXT:    vldb wl10, [p0], #32; vmac.f bml5, bml5, x2, x3, r3
-  ; CHECK-NEXT:    vldb wh10, [p0], #32; vmac.f bml6, bml6, x0, x3, r3
-  ; CHECK-NEXT:    vldb wl5, [p1], m5; vmac.f bml4, bml4, x6, x3, r3
-  ; CHECK-NEXT:    vldb wh5, [p1], m6
-  ; CHECK-NEXT:    vldb wl3, [p0], #32
-  ; CHECK-NEXT:    vldb wl7, [p1], m5
-  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0; vshuffle x11, x9, x9, r2
-  ; CHECK-NEXT:    vldb wh7, [p1], m6
-  ; CHECK-NEXT:    vldb wl3, [p1], m5; vshuffle x1, x8, x10, r4
-  ; CHECK-NEXT:    vldb.3d wh3, [p1], d1; vshuffle x2, x2, x10, r16
-  ; CHECK-NEXT:    vshuffle x5, x5, x5, r2; vmac.f bmh0, bmh0, x1, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh1, bmh1, x2, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh4, bmh4, x1, x5, r3
-  ; CHECK-NEXT:    vshuffle x0, x6, x3, r4; vmac.f bmh5, bmh5, x2, x5, r3
-  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16
-  ; CHECK-NEXT:    vshuffle x7, x7, x7, r2; vmac.f bmh2, bmh2, x0, x11, r3
-  ; CHECK-NEXT:    vshuffle x3, x3, x3, r2; vmac.f bmh3, bmh3, x6, x11, r3
+  ; CHECK-NEXT:    vldb wl6, [p0, #64]; vshuffle x11, x9, x9, r2; vmac.f bml0, bml0, x2, x7, r3
+  ; CHECK-NEXT:    padds [p0], m4; vldb wh6, [p0, #96]; vmac.f bml1, bml1, x0, x7, r3
+  ; CHECK-NEXT:    padds [p0], #128; vshuffle x5, x5, x5, r2; vmac.f bml2, bml2, x6, x7, r3
+  ; CHECK-NEXT:    vldb wl10, [p0], #32; vmac.f bml3, bml3, x1, x3, r3
+  ; CHECK-NEXT:    vldb wh10, [p0], #32; vshuffle x7, x7, x7, r2; vmac.f bml5, bml5, x2, x3, r3
+  ; CHECK-NEXT:    vldb wl3, [p0], #32; vmac.f bml6, bml6, x0, x3, r3
+  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0
+  ; CHECK-NEXT:    nop
+  ; CHECK-NEXT:    vldb wl3, [p1], m5
+  ; CHECK-NEXT:    vldb.3d wh3, [p1], d1
+  ; CHECK-NEXT:    nop
+  ; CHECK-NEXT:    vshuffle x2, x2, x10, r16
+  ; CHECK-NEXT:    vshuffle x1, x8, x10, r4
+  ; CHECK-NEXT:    vshuffle x0, x6, x3, r4; vmac.f bmh1, bmh1, x2, x11, r3
   ; CHECK-NEXT:  .L_LEnd0:
-  ; CHECK-NEXT:    nopb ; nopa ; nops ; nopxm ; vmac.f bmh6, bmh6, x0, x5, r3
+  ; CHECK-NEXT:    nopb ; nopa ; nops ; nopx ; vshuffle x6, x6, x3, r16; vmac.f bmh0, bmh0, x1, x11, r3
   ; CHECK-NEXT:  // %bb.3: // %for.cond.cleanup
-  ; CHECK-NEXT:    nopb ; nopa ; nops ; nopxm ; vmac.f bmh7, bmh7, x6, x5, r3
+  ; CHECK-NEXT:    vmac.f bmh2, bmh2, x0, x11, r3
+  ; CHECK-NEXT:    nopx ; vshuffle x3, x3, x3, r2; vmac.f bmh3, bmh3, x6, x11, r3
+  ; CHECK-NEXT:    vmac.f bmh4, bmh4, x1, x5, r3
+  ; CHECK-NEXT:    vmac.f bml4, bml4, x6, x3, r3
+  ; CHECK-NEXT:    vmac.f bmh5, bmh5, x2, x5, r3
+  ; CHECK-NEXT:    vmac.f bmh6, bmh6, x0, x5, r3
+  ; CHECK-NEXT:    vmac.f bmh7, bmh7, x6, x5, r3
   ; CHECK-NEXT:    vmac.f bmh8, bmh8, x1, x7, r3
   ; CHECK-NEXT:    vmac.f bml0, bml0, x2, x7, r3
   ; CHECK-NEXT:    vmac.f bml1, bml1, x0, x7, r3
@@ -88,13 +92,6 @@
   ; CHECK-NEXT:    vmac.f bml3, bml3, x1, x3, r3
   ; CHECK-NEXT:    vmac.f bml5, bml5, x2, x3, r3
   ; CHECK-NEXT:    vmac.f bml6, bml6, x0, x3, r3
-  ; CHECK-NEXT:    vmac.f bml4, bml4, x6, x3, r3
-  ; CHECK-NEXT:    nopx
-  ; CHECK-NEXT:    nop
-  ; CHECK-NEXT:    nop
-  ; CHECK-NEXT:    nop
-  ; CHECK-NEXT:    nop
-  ; CHECK-NEXT:    nop
   ; CHECK-NEXT:    nop
   ; CHECK-NEXT:    nop
   ; CHECK-NEXT:    nop
diff --git a/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm.mir b/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm.mir
index fbd31a30aea4..72c73af64e31 100644
--- a/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm.mir
+++ b/llvm/test/CodeGen/AIE/aie2/schedule/postpipeliner/gemm.mir
@@ -35,52 +35,50 @@
   ; CHECK-NEXT:    vldb wh9, [p1], m6; nopa ; padds [p0], #128; nopxm ; nopv
   ; CHECK-NEXT:    vldb wl10, [p0], #32; nopx
   ; CHECK-NEXT:    vldb wh10, [p0], #32
+  ; CHECK-NEXT:    vldb wl3, [p0], #32
+  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0
   ; CHECK-NEXT:    vldb wl5, [p1], m5
   ; CHECK-NEXT:    vldb wh5, [p1], m6
-  ; CHECK-NEXT:    vldb wl3, [p0], #32
-  ; CHECK-NEXT:    vldb wl7, [p1], m5
-  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0; vshuffle x11, x9, x9, r2
+  ; CHECK-NEXT:    vldb wl7, [p1], m5; vshuffle x11, x9, x9, r2
   ; CHECK-NEXT:    vldb wh7, [p1], m6
   ; CHECK-NEXT:    vldb wl3, [p1], m5; vshuffle x1, x8, x10, r4
   ; CHECK-NEXT:    vldb.3d wh3, [p1], d1; vshuffle x8, x8, x10, r16
-  ; CHECK-NEXT:    vshuffle x5, x5, x5, r2; vmac.f bmh0, bmh0, x1, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh1, bmh1, x8, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh4, bmh4, x1, x5, r3
-  ; CHECK-NEXT:    vshuffle x10, x6, x3, r4; vmac.f bmh5, bmh5, x8, x5, r3
-  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16
-  ; CHECK-NEXT:    vshuffle x7, x7, x7, r2; vmac.f bmh2, bmh2, x10, x11, r3
-  ; CHECK-NEXT:    vshuffle x3, x3, x3, r2; vmac.f bmh3, bmh3, x6, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh6, bmh6, x10, x5, r3
+  ; CHECK-NEXT:    vshuffle x10, x6, x3, r4; vmac.f bmh0, bmh0, x1, x11, r3
+  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16; vmac.f bmh1, bmh1, x8, x11, r3
+  ; CHECK-NEXT:    vshuffle x5, x5, x5, r2; vmac.f bmh2, bmh2, x10, x11, r3
+  ; CHECK-NEXT:    vmac.f bmh3, bmh3, x6, x11, r3
+  ; CHECK-NEXT:    vshuffle x7, x7, x7, r2; vmac.f bmh4, bmh4, x1, x5, r3
+  ; CHECK-NEXT:    vmac.f bmh5, bmh5, x8, x5, r3
+  ; CHECK-NEXT:    vshuffle x3, x3, x3, r2; vmac.f bmh6, bmh6, x10, x5, r3
   ; CHECK-NEXT:    .p2align 4
   ; CHECK-NEXT:  .LBB0_2: // %for.body
   ; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
-  ; CHECK-NEXT:    nopa ; vldb wl8, [p0, #0]; nopxm ; vmac.f bmh7, bmh7, x6, x5, r3
-  ; CHECK-NEXT:    vldb wh8, [p0, #32]; vmac.f bmh8, bmh8, x1, x7, r3
+  ; CHECK-NEXT:    vldb wl8, [p0, #0]; nopa ; nops ; nopxm ; vmac.f bmh7, bmh7, x6, x5, r3
+  ; CHECK-NEXT:    vldb wh8, [p0, #32]; nopx ; vmac.f bmh8, bmh8, x1, x7, r3
   ; CHECK-NEXT:    vldb wl6, [p0, #64]; vmac.f bml0, bml0, x8, x7, r3
   ; CHECK-NEXT:    vldb wh6, [p0, #96]; vmac.f bml1, bml1, x10, x7, r3
   ; CHECK-NEXT:    padds [p0], m4; vldb wl9, [p1], m5; vmac.f bml2, bml2, x6, x7, r3
   ; CHECK-NEXT:    padds [p0], #128; vldb wh9, [p1], m6; vmac.f bml3, bml3, x1, x3, r3
   ; CHECK-NEXT:    vldb wl10, [p0], #32; vmac.f bml5, bml5, x8, x3, r3
   ; CHECK-NEXT:    vldb wh10, [p0], #32; vmac.f bml6, bml6, x10, x3, r3
-  ; CHECK-NEXT:    vldb wl5, [p1], m5; vmac.f bml4, bml4, x6, x3, r3
+  ; CHECK-NEXT:    vldb wl3, [p0], #32; vmac.f bml4, bml4, x6, x3, r3
+  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0
+  ; CHECK-NEXT:    vldb wl5, [p1], m5
   ; CHECK-NEXT:    vldb wh5, [p1], m6
-  ; CHECK-NEXT:    vldb wl3, [p0], #32
-  ; CHECK-NEXT:    vldb wl7, [p1], m5
-  ; CHECK-NEXT:    vldb.3d wh3, [p0], d0; vshuffle x11, x9, x9, r2
+  ; CHECK-NEXT:    vldb wl7, [p1], m5; vshuffle x11, x9, x9, r2
   ; CHECK-NEXT:    vldb wh7, [p1], m6
   ; CHECK-NEXT:    vldb wl3, [p1], m5; vshuffle x1, x8, x10, r4
   ; CHECK-NEXT:    vldb.3d wh3, [p1], d1; vshuffle x8, x8, x10, r16
-  ; CHECK-NEXT:    vshuffle x5, x5, x5, r2; vmac.f bmh0, bmh0, x1, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh1, bmh1, x8, x11, r3
-  ; CHECK-NEXT:    vmac.f bmh4, bmh4, x1, x5, r3
-  ; CHECK-NEXT:    vshuffle x10, x6, x3, r4; vmac.f bmh5, bmh5, x8, x5, r3
-  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16
-  ; CHECK-NEXT:    vshuffle x7, x7, x7, r2; vmac.f bmh2, bmh2, x10, x11, r3
-  ; CHECK-NEXT:    vshuffle x3, x3, x3, r2; vmac.f bmh3, bmh3, x6, x11, r3
+  ; CHECK-NEXT:    vshuffle x10, x6, x3, r4; vmac.f bmh0, bmh0, x1, x11, r3
+  ; CHECK-NEXT:    vshuffle x6, x6, x3, r16; vmac.f bmh1, bmh1, x8, x11, r3
+  ; CHECK-NEXT:    vshuffle x5, x5, x5, r2; vmac.f bmh2, bmh2, x10, x11, r3
+  ; CHECK-NEXT:    vmac.f bmh3, bmh3, x6, x11, r3
+  ; CHECK-NEXT:    vshuffle x7, x7, x7, r2; vmac.f bmh4, bmh4, x1, x5, r3
+  ; CHECK-NEXT:    vmac.f bmh5, bmh5, x8, x5, r3
   ; CHECK-NEXT:  .L_LEnd0:
-  ; CHECK-NEXT:    nopb ; nopa ; nops ; nopxm ; vmac.f bmh6, bmh6, x10, x5, r3
+  ; CHECK-NEXT:    nopb ; nopa ; nops ; nopx ; vshuffle x3, x3, x3, r2; vmac.f bmh6, bmh6, x10, x5, r3
   ; CHECK-NEXT:  // %bb.3: // %for.cond.cleanup
-  ; CHECK-NEXT:    nopb ; nopa ; nops ; nopxm ; vmac.f bmh7, bmh7, x6, x5, r3
+  ; CHECK-NEXT:    vmac.f bmh7, bmh7, x6, x5, r3
   ; CHECK-NEXT:    vmac.f bmh8, bmh8, x1, x7, r3
   ; CHECK-NEXT:    vmac.f bml0, bml0, x8, x7, r3
   ; CHECK-NEXT:    vmac.f bml1, bml1, x10, x7, r3
@@ -89,7 +87,6 @@
   ; CHECK-NEXT:    vmac.f bml5, bml5, x8, x3, r3
   ; CHECK-NEXT:    vmac.f bml6, bml6, x10, x3, r3
   ; CHECK-NEXT:    vmac.f bml4, bml4, x6, x3, r3
-  ; CHECK-NEXT:    nopx
   ; CHECK-NEXT:    nop
   ; CHECK-NEXT:    nop
   ; CHECK-NEXT:    nop