Index: lib/Target/X86/X86ISelDAGToDAG.cpp
===================================================================
--- lib/Target/X86/X86ISelDAGToDAG.cpp
+++ lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -2223,24 +2223,32 @@
 
     if (foldedLoad) {
       SDValue Chain;
+      MachineSDNode *CNode = nullptr;
       SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
                         InFlag };
       if (MOpc == X86::MULX32rm || MOpc == X86::MULX64rm) {
         SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Other, MVT::Glue);
-        SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
+        CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
         ResHi = SDValue(CNode, 0);
         ResLo = SDValue(CNode, 1);
         Chain = SDValue(CNode, 2);
         InFlag = SDValue(CNode, 3);
       } else {
         SDVTList VTs = CurDAG->getVTList(MVT::Other, MVT::Glue);
-        SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
+        CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
         Chain = SDValue(CNode, 0);
         InFlag = SDValue(CNode, 1);
       }
 
       // Update the chain.
       ReplaceUses(N1.getValue(1), Chain);
+      // Record the mem-refs
+      LoadSDNode *LoadNode = cast<LoadSDNode>(N1);
+      if (LoadNode) {
+        MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+        MemOp[0] = LoadNode->getMemOperand();
+        CNode->setMemRefs(MemOp, MemOp + 1);
+      }
     } else {
       SDValue Ops[] = { N1, InFlag };
       if (Opc == X86::MULX32rr || Opc == X86::MULX64rr) {
Index: test/CodeGen/X86/hoist-invariant-load.ll
===================================================================
--- test/CodeGen/X86/hoist-invariant-load.ll
+++ test/CodeGen/X86/hoist-invariant-load.ll
@@ -1,7 +1,10 @@
 ; REQUIRES: asserts
-; RUN: llc < %s -stats -O2 2>&1 | grep "2 machine-licm"
+; RUN: llc -mcpu=haswell < %s -stats -O2 2>&1 | grep "4 machine-licm.*hoisted"
+; For test:
 ; 2 invariant loads, 1 for OBJC_SELECTOR_REFERENCES_
 ; and 1 for objc_msgSend from the GOT
+; For test_multi_def:
+; 2 invariant load (full multiply, both loads should be hoisted.)
 
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
 target triple = "x86_64-apple-macosx10.7.2"
@@ -29,4 +32,32 @@
 
 declare i8* @objc_msgSend(i8*, i8*, ...) nonlazybind
 
+define void @test_multi_def(i64* dereferenceable(8) %x1,
+                            i64* dereferenceable(8) %x2,
+                            i128* %y, i64 %count) nounwind {
+entry:
+  br label %for.body
+
+for.check:
+  %inc = add nsw i64 %i, 1
+  %done = icmp sge i64 %inc, %count
+  br i1 %done, label %exit, label %for.body
+
+for.body:
+  %i = phi i64 [ 0, %entry ], [ %inc, %for.check ]
+  %x1_load = load i64, i64* %x1, align 8, !invariant.load !0
+  %x1_zext = zext i64 %x1_load to i128
+  %x2_load = load i64, i64* %x2, align 8, !invariant.load !0
+  %x2_zext = zext i64 %x2_load to i128
+  %x_prod = mul i128 %x1_zext, %x2_zext
+  %y_elem = getelementptr inbounds i128, i128* %y, i64 %i
+  %y_load = load i128, i128* %y_elem, align 8
+  %y_plus = add i128 %x_prod, %y_load
+  store i128 %y_plus, i128* %y_elem, align 8
+  br label %for.check
+
+exit:
+  ret void
+}
+
 !0 = !{}