diff --git a/lib/Target/ARM/ARM.td b/lib/Target/ARM/ARM.td
index bf4315f..e690e18 100644
--- a/lib/Target/ARM/ARM.td
+++ b/lib/Target/ARM/ARM.td
@@ -51,6 +51,12 @@
 // to just not use them.
 def FeatureHasSlowFPVMLx : SubtargetFeature<"slowfpvmlx", "SlowFPVMLx", "true",
                                          "Disable VFP / NEON MAC instructions">;
+
+// Cortex-A8 / A9 Advanced SIMD has multiplier accumulator forwarding.
+def FeatureVMLxForwarding : SubtargetFeature<"vmlx-forwarding",
+                                       "HasVMLxForwarding", "true",
+                                       "Has multiplier accumulator forwarding">;
+
 // Some processors benefit from using NEON instructions for scalar
 // single-precision FP operations.
 def FeatureNEONForFP : SubtargetFeature<"neonfp", "UseNEONForSinglePrecisionFP",
@@ -100,11 +106,12 @@
 def ProcA8      : SubtargetFeature<"a8", "ARMProcFamily", "CortexA8",
                                    "Cortex-A8 ARM processors",
                                    [FeatureSlowFPBrcc, FeatureNEONForFP,
-                                    FeatureHasSlowFPVMLx, FeatureT2XtPk]>;
+                                    FeatureHasSlowFPVMLx, FeatureVMLxForwarding,
+                                    FeatureT2XtPk]>;
 def ProcA9      : SubtargetFeature<"a9", "ARMProcFamily", "CortexA9",
                                    "Cortex-A9 ARM processors",
-                                   [FeatureHasSlowFPVMLx, FeatureT2XtPk,
-                                    FeatureFP16]>;
+                                   [FeatureHasSlowFPVMLx, FeatureVMLxForwarding,
+                                    FeatureT2XtPk, FeatureFP16]>;
 
 class ProcNoItin<string Name, list<SubtargetFeature> Features>
  : Processor<Name, GenericItineraries, Features>;
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 16b110f..5838181 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -5224,6 +5224,42 @@
   return SDValue();
 }
 
+/// PerformVMULCombine
+/// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
+/// special multiplier accumulator forwarding.
+///   vmul d3, d0, d2
+///   vmla d3, d1, d2
+/// is faster than
+///   vadd d3, d0, d1
+///   vmul d3, d3, d2
+static SDValue PerformVMULCombine(SDNode *N,
+                                  TargetLowering::DAGCombinerInfo &DCI,
+                                  const ARMSubtarget *Subtarget) {
+  if (!Subtarget->hasVMLxForwarding())
+    return SDValue();
+
+  SelectionDAG &DAG = DCI.DAG;
+  SDValue N0 = N->getOperand(0);
+  SDValue N1 = N->getOperand(1);
+  unsigned Opcode = N0.getOpcode();
+  if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
+      Opcode != ISD::FADD && Opcode != ISD::FSUB) {
+    Opcode = N0.getOpcode();
+    if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
+        Opcode != ISD::FADD && Opcode != ISD::FSUB)
+      return SDValue();
+    std::swap(N0, N1);
+  }
+
+  EVT VT = N->getValueType(0);
+  DebugLoc DL = N->getDebugLoc();
+  SDValue N00 = N0->getOperand(0);
+  SDValue N01 = N0->getOperand(1);
+  return DAG.getNode(Opcode, DL, VT,
+                     DAG.getNode(ISD::MUL, DL, VT, N00, N1),
+                     DAG.getNode(ISD::MUL, DL, VT, N01, N1));
+}
+
 static SDValue PerformMULCombine(SDNode *N,
                                  TargetLowering::DAGCombinerInfo &DCI,
                                  const ARMSubtarget *Subtarget) {
@@ -5236,6 +5272,8 @@
     return SDValue();
 
   EVT VT = N->getValueType(0);
+  if (VT.is64BitVector() || VT.is128BitVector())
+    return PerformVMULCombine(N, DCI, Subtarget);
   if (VT != MVT::i32)
     return SDValue();
 
diff --git a/lib/Target/ARM/ARMSubtarget.h b/lib/Target/ARM/ARMSubtarget.h
index 76c1c3f..e024182 100644
--- a/lib/Target/ARM/ARMSubtarget.h
+++ b/lib/Target/ARM/ARMSubtarget.h
@@ -61,6 +61,10 @@
   /// whether the FP VML[AS] instructions are slow (if so, don't use them).
   bool SlowFPVMLx;
 
+  /// HasVMLxForwarding - If true, NEON has special multiplier accumulator
+  /// forwarding to allow mul + mla being issued back to back.
+  bool HasVMLxForwarding;
+
   /// SlowFPBrcc - True if floating point compare + branch is slow.
   bool SlowFPBrcc;
 
@@ -182,6 +186,7 @@
   bool hasT2ExtractPack() const { return HasT2ExtractPack; }
   bool hasDataBarrier() const { return HasDataBarrier; }
   bool useFPVMLx() const { return !SlowFPVMLx; }
+  bool hasVMLxForwarding() const { return HasVMLxForwarding; }
   bool isFPBrccSlow() const { return SlowFPBrcc; }
   bool isFPOnlySP() const { return FPOnlySP; }
   bool prefers32BitThumb() const { return Pref32BitThumb; }