Index: lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- lib/Target/X86/X86ISelLowering.cpp
+++ lib/Target/X86/X86ISelLowering.cpp
@@ -22927,6 +22927,29 @@
                                 MinAlign(St->getAlignment(), 4));
     return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
   }
+
+  // This is similar to the above case, but here we handle a scalar 64-bit
+  // integer store that is extracted from a vector on a 32-bit target.
+  // If we have SSE, then we can treat it like a floating-point double
+  // to get past legalization. The execution dependencies fixup pass will
+  // choose the optimal machine instruction for the store if this really is
+  // an integer or v2f32 rather than an f64.
+  if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
+      St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
+    SDValue OldExtract = St->getOperand(1);
+    SDValue ExtOp0 = OldExtract.getOperand(0);
+    unsigned VecSize = ExtOp0.getValueSizeInBits();
+    MVT VecVT = MVT::getVectorVT(MVT::f64, VecSize / 64);
+    SDValue BitCast = DAG.getNode(ISD::BITCAST, dl, VecVT, ExtOp0);
+    SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
+                                     BitCast, OldExtract.getOperand(1));
+    return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
+                        St->getPointerInfo(),
+                        St->isVolatile(),
+                        St->isNonTemporal(),
+                        St->getAlignment());
+  }
+
   return SDValue();
 }
 
Index: test/CodeGen/X86/i64-mem-copy.ll
===================================================================
--- test/CodeGen/X86/i64-mem-copy.ll
+++ test/CodeGen/X86/i64-mem-copy.ll
@@ -1,5 +1,6 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=sse2 | FileCheck %s --check-prefix=X64
 ; RUN: llc < %s -mtriple=i386-unknown-unknown -mattr=sse2 | FileCheck %s --check-prefix=X32
+; RUN: llc < %s -mtriple=i386-unknown-unknown -mattr=avx2 | FileCheck %s --check-prefix=X32AVX
 
 ; Use movq or movsd to load / store i64 values if sse2 is available.
 ; rdar://6659858
@@ -18,9 +19,46 @@
 ; X32-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
 ; X32-NEXT:    movsd %xmm0, (%eax)
 ; X32-NEXT:    retl
-
   %tmp1 = load i64, i64* %y, align 8
   store i64 %tmp1, i64* %x, align 8
   ret void
 }
 
+; Verify that a 64-bit chunk extracted from a vector is stored with a movq
+; regardless of whether the system is 64-bit.
+
+define void @store_i64_from_vector(<8 x i16> %x, <8 x i16> %y, i64* %i) {
+; X64-LABEL: store_i64_from_vector:
+; X64:       # BB#0:
+; X64-NEXT:    paddw %xmm1, %xmm0
+; X64-NEXT:    movq %xmm0, (%rdi)
+; X64-NEXT:    retq
+;
+; X32-LABEL: store_i64_from_vector:
+; X32:       # BB#0:
+; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NEXT:    paddw %xmm1, %xmm0
+; X32-NEXT:    movq %xmm0, (%eax)
+; X32-NEXT:    retl
+  %z = add <8 x i16> %x, %y                          ; force execution domain
+  %bc = bitcast <8 x i16> %z to <2 x i64>
+  %vecext = extractelement <2 x i64> %bc, i32 0
+  store i64 %vecext, i64* %i, align 8
+  ret void
+}
+
+define void @store_i64_from_vector256(<16 x i16> %x, <16 x i16> %y, i64* %i) {
+; X32AVX-LABEL: store_i64_from_vector256:
+; X32AVX:       # BB#0:
+; X32AVX-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32AVX-NEXT:    vpaddw %ymm1, %ymm0, %ymm0
+; X32AVX-NEXT:    vmovq %xmm0, (%eax)
+; X32AVX-NEXT:    vzeroupper
+; X32AVX-NEXT:    retl
+  %z = add <16 x i16> %x, %y                          ; force execution domain
+  %bc = bitcast <16 x i16> %z to <4 x i64>
+  %vecext = extractelement <4 x i64> %bc, i32 0
+  store i64 %vecext, i64* %i, align 8
+  ret void
+}
+