diff --git a/lib/Target/X86/README-SSE.txt b/lib/Target/X86/README-SSE.txt
index 1f5f203..58b4384 100644
--- a/lib/Target/X86/README-SSE.txt
+++ b/lib/Target/X86/README-SSE.txt
@@ -505,46 +505,6 @@
 
 //===---------------------------------------------------------------------===//
 
-We should compile this:
-#include <xmmintrin.h>
-typedef union {
-  int i[4];
-  float f[4];
-  __m128 v;
-} vector4_t;
-void swizzle (const void *a, vector4_t * b, vector4_t * c) {
-  b->v = _mm_loadl_pi (b->v, (__m64 *) a);
-  c->v = _mm_loadl_pi (c->v, ((__m64 *) a) + 1);
-}
-
-to:
-
-_swizzle:
-        movl    4(%esp), %eax
-        movl    8(%esp), %edx
-        movl    12(%esp), %ecx
-        movlps  (%eax), %xmm0
-        movlps  %xmm0, (%edx)
-        movlps  8(%eax), %xmm0
-        movlps  %xmm0, (%ecx)
-        ret
-
-not:
-
-swizzle:
-        movl 8(%esp), %eax
-        movaps (%eax), %xmm0
-        movl 4(%esp), %ecx
-        movlps (%ecx), %xmm0
-        movaps %xmm0, (%eax)
-        movl 12(%esp), %eax
-        movaps (%eax), %xmm0
-        movlps 8(%ecx), %xmm0
-        movaps %xmm0, (%eax)
-        ret
-
-//===---------------------------------------------------------------------===//
-
 We should materialize vector constants like "all ones" and "signbit" with 
 code like:
 
diff --git a/lib/Target/X86/X86ISelDAGToDAG.cpp b/lib/Target/X86/X86ISelDAGToDAG.cpp
index c92e769..576661d 100644
--- a/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -363,6 +363,32 @@
                          Store.getOperand(2), Store.getOperand(3));
 }
 
+/// isRMWLoad - Return true if N is a load that's part of RMW sub-DAG.
+/// 
+static bool isRMWLoad(SDOperand N, SDOperand Chain, SDOperand Address,
+                      SDOperand &Load) {
+  if (N.getOpcode() == ISD::BIT_CONVERT)
+    N = N.getOperand(0);
+
+  LoadSDNode *LD = dyn_cast<LoadSDNode>(N);
+  if (!LD || LD->isVolatile())
+    return false;
+  if (LD->getAddressingMode() != ISD::UNINDEXED)
+    return false;
+
+  ISD::LoadExtType ExtType = LD->getExtensionType();
+  if (ExtType != ISD::NON_EXTLOAD && ExtType != ISD::EXTLOAD)
+    return false;
+
+  if (N.hasOneUse() &&
+      N.getOperand(1) == Address &&
+      N.Val->isOperandOf(Chain.Val)) {
+    Load = N;
+    return true;
+  }
+  return false;
+}
+
 /// PreprocessForRMW - Preprocess the DAG to make instruction selection better.
 /// This is only run if not in -fast mode (aka -O0).
 /// This allows the instruction selector to pick more read-modify-write
@@ -414,8 +440,8 @@
 
     SDOperand N1 = I->getOperand(1);
     SDOperand N2 = I->getOperand(2);
-    if (MVT::isFloatingPoint(N1.getValueType()) ||
-        MVT::isVector(N1.getValueType()) ||
+    if ((MVT::isFloatingPoint(N1.getValueType()) &&
+         !MVT::isVector(N1.getValueType())) ||
         !N1.hasOneUse())
       continue;
 
@@ -429,20 +455,13 @@
       case ISD::OR:
       case ISD::XOR:
       case ISD::ADDC:
-      case ISD::ADDE: {
+      case ISD::ADDE:
+      case ISD::VECTOR_SHUFFLE: {
         SDOperand N10 = N1.getOperand(0);
         SDOperand N11 = N1.getOperand(1);
-        if (ISD::isNON_EXTLoad(N10.Val))
-          RModW = true;
-        else if (ISD::isNON_EXTLoad(N11.Val)) {
-          RModW = true;
-          std::swap(N10, N11);
-        }
-        RModW = RModW && N10.Val->isOperandOf(Chain.Val) && N10.hasOneUse() &&
-          (N10.getOperand(1) == N2) &&
-          (N10.Val->getValueType(0) == N1.getValueType());
-        if (RModW)
-          Load = N10;
+        RModW = isRMWLoad(N10, Chain, N2, Load);
+        if (!RModW)
+          RModW = isRMWLoad(N11, Chain, N2, Load);
         break;
       }
       case ISD::SUB:
@@ -456,12 +475,7 @@
       case X86ISD::SHLD:
       case X86ISD::SHRD: {
         SDOperand N10 = N1.getOperand(0);
-        if (ISD::isNON_EXTLoad(N10.Val))
-          RModW = N10.Val->isOperandOf(Chain.Val) && N10.hasOneUse() &&
-            (N10.getOperand(1) == N2) &&
-            (N10.Val->getValueType(0) == N1.getValueType());
-        if (RModW)
-          Load = N10;
+        RModW = isRMWLoad(N10, Chain, N2, Load);
         break;
       }
     }
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index e9c6481..06b7a3a 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -2977,13 +2977,15 @@
                   MOVHP_shuffle_mask)),
           (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
 
-def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
+def : Pat<(v4i32 (vector_shuffle VR128:$src1,
+                                 (bc_v4i32 (memopv2i64 addr:$src2)),
                   MOVLP_shuffle_mask)),
           (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memop addr:$src2),
                   MOVLP_shuffle_mask)),
           (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
-def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
+def : Pat<(v4i32 (vector_shuffle VR128:$src1,
+                                 (bc_v4i32 (memopv2i64 addr:$src2)),
                   MOVHP_shuffle_mask)),
           (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memop addr:$src2),
@@ -2991,6 +2993,37 @@
           (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
 }
 
+// (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
+// (store (vector_shuffle (load addr), v2, <0, 1, 4, 5>), addr) using MOVHPS
+def : Pat<(store (v4f32 (vector_shuffle (memop addr:$src1), VR128:$src2,
+                         MOVLP_shuffle_mask)), addr:$src1),
+          (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
+def : Pat<(store (v2f64 (vector_shuffle (memop addr:$src1), VR128:$src2,
+                         MOVLP_shuffle_mask)), addr:$src1),
+          (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
+def : Pat<(store (v4f32 (vector_shuffle (memop addr:$src1), VR128:$src2,
+                         MOVHP_shuffle_mask)), addr:$src1),
+          (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
+def : Pat<(store (v2f64 (vector_shuffle (memop addr:$src1), VR128:$src2,
+                         MOVHP_shuffle_mask)), addr:$src1),
+          (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
+
+def : Pat<(store (v4i32 (vector_shuffle
+                         (bc_v4i32 (memopv2i64 addr:$src1)), VR128:$src2,
+                         MOVLP_shuffle_mask)), addr:$src1),
+          (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
+def : Pat<(store (v2i64 (vector_shuffle (memop addr:$src1), VR128:$src2,
+                         MOVLP_shuffle_mask)), addr:$src1),
+          (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
+def : Pat<(store (v4i32 (vector_shuffle
+                         (bc_v4i32 (memopv2i64 addr:$src1)), VR128:$src2,
+                         MOVHP_shuffle_mask)), addr:$src1),
+          (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
+def : Pat<(store (v2i64 (vector_shuffle (memop addr:$src1), VR128:$src2,
+                         MOVHP_shuffle_mask)), addr:$src1),
+          (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
+
+
 let AddedComplexity = 15 in {
 // Setting the lowest element in the vector.
 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
diff --git a/test/CodeGen/X86/vec_shuffle-18.ll b/test/CodeGen/X86/vec_shuffle-18.ll
new file mode 100644
index 0000000..5e05690
--- /dev/null
+++ b/test/CodeGen/X86/vec_shuffle-18.ll
@@ -0,0 +1,25 @@
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2 | grep mov | count 7
+
+	%struct.vector4_t = type { <4 x float> }
+
+define void @swizzle(i8* %a, %struct.vector4_t* %b, %struct.vector4_t* %c) nounwind  {
+entry:
+	%tmp9 = getelementptr %struct.vector4_t* %b, i32 0, i32 0		; <<4 x float>*> [#uses=2]
+	%tmp10 = load <4 x float>* %tmp9, align 16		; <<4 x float>> [#uses=1]
+	%tmp14 = bitcast i8* %a to double*		; <double*> [#uses=1]
+	%tmp15 = load double* %tmp14		; <double> [#uses=1]
+	%tmp16 = insertelement <2 x double> undef, double %tmp15, i32 0		; <<2 x double>> [#uses=1]
+	%tmp18 = bitcast <2 x double> %tmp16 to <4 x float>		; <<4 x float>> [#uses=1]
+	%tmp19 = shufflevector <4 x float> %tmp10, <4 x float> %tmp18, <4 x i32> < i32 4, i32 5, i32 2, i32 3 >		; <<4 x float>> [#uses=1]
+	store <4 x float> %tmp19, <4 x float>* %tmp9, align 16
+	%tmp28 = getelementptr %struct.vector4_t* %c, i32 0, i32 0		; <<4 x float>*> [#uses=2]
+	%tmp29 = load <4 x float>* %tmp28, align 16		; <<4 x float>> [#uses=1]
+	%tmp26 = getelementptr i8* %a, i32 8		; <i8*> [#uses=1]
+	%tmp33 = bitcast i8* %tmp26 to double*		; <double*> [#uses=1]
+	%tmp34 = load double* %tmp33		; <double> [#uses=1]
+	%tmp35 = insertelement <2 x double> undef, double %tmp34, i32 0		; <<2 x double>> [#uses=1]
+	%tmp37 = bitcast <2 x double> %tmp35 to <4 x float>		; <<4 x float>> [#uses=1]
+	%tmp38 = shufflevector <4 x float> %tmp29, <4 x float> %tmp37, <4 x i32> < i32 4, i32 5, i32 2, i32 3 >		; <<4 x float>> [#uses=1]
+	store <4 x float> %tmp38, <4 x float>* %tmp28, align 16
+	ret void
+}