diff --git a/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index f2be01b73be..908038b35e4 100644
--- a/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -479,57 +479,57 @@ Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS,
   // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
   // this.
   bool Swapped = false;
-  GetElementPtrInst *GEP = 0;
-  ConstantExpr *CstGEP = 0;
-  
-  // TODO: Could also optimize &A[i] - &A[j] -> "i-j", and "&A.foo[i] - &A.foo".
+  GEPOperator *GEP1 = 0, *GEP2 = 0;
+
   // For now we require one side to be the base pointer "A" or a constant
-  // expression derived from it.
-  if (GetElementPtrInst *LHSGEP = dyn_cast<GetElementPtrInst>(LHS)) {
+  // GEP derived from it.
+  if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
     // (gep X, ...) - X
     if (LHSGEP->getOperand(0) == RHS) {
-      GEP = LHSGEP;
+      GEP1 = LHSGEP;
       Swapped = false;
-    } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(RHS)) {
-      // (gep X, ...) - (ce_gep X, ...)
-      if (CE->getOpcode() == Instruction::GetElementPtr &&
-          LHSGEP->getOperand(0) == CE->getOperand(0)) {
-        CstGEP = CE;
-        GEP = LHSGEP;
+    } else if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
+      // (gep X, ...) - (gep X, ...)
+      if (LHSGEP->getOperand(0)->stripPointerCasts() ==
+            RHSGEP->getOperand(0)->stripPointerCasts()) {
+        GEP2 = RHSGEP;
+        GEP1 = LHSGEP;
         Swapped = false;
       }
     }
   }
   
-  if (GetElementPtrInst *RHSGEP = dyn_cast<GetElementPtrInst>(RHS)) {
+  if (GEPOperator *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
     // X - (gep X, ...)
     if (RHSGEP->getOperand(0) == LHS) {
-      GEP = RHSGEP;
+      GEP1 = RHSGEP;
       Swapped = true;
-    } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(LHS)) {
-      // (ce_gep X, ...) - (gep X, ...)
-      if (CE->getOpcode() == Instruction::GetElementPtr &&
-          RHSGEP->getOperand(0) == CE->getOperand(0)) {
-        CstGEP = CE;
-        GEP = RHSGEP;
+    } else if (GEPOperator *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
+      // (gep X, ...) - (gep X, ...)
+      if (RHSGEP->getOperand(0)->stripPointerCasts() ==
+            LHSGEP->getOperand(0)->stripPointerCasts()) {
+        GEP2 = LHSGEP;
+        GEP1 = RHSGEP;
         Swapped = true;
       }
     }
   }
   
-  if (GEP == 0)
+  // Avoid duplicating the arithmetic if GEP2 has non-constant indices and
+  // multiple users.
+  if (GEP1 == 0 ||
+      (GEP2 != 0 && !GEP2->hasAllConstantIndices() && !GEP2->hasOneUse()))
     return 0;
   
   // Emit the offset of the GEP and an intptr_t.
-  Value *Result = EmitGEPOffset(GEP);
+  Value *Result = EmitGEPOffset(GEP1);
   
   // If we had a constant expression GEP on the other side offsetting the
   // pointer, subtract it from the offset we have.
-  if (CstGEP) {
-    Value *CstOffset = EmitGEPOffset(CstGEP);
-    Result = Builder->CreateSub(Result, CstOffset);
+  if (GEP2) {
+    Value *Offset = EmitGEPOffset(GEP2);
+    Result = Builder->CreateSub(Result, Offset);
   }
-  
 
   // If we have p - gep(p, ...)  then we have to negate the result.
   if (Swapped)
diff --git a/test/Transforms/InstCombine/sub.ll b/test/Transforms/InstCombine/sub.ll
index 37de3281358..b71ec8c98f8 100644
--- a/test/Transforms/InstCombine/sub.ll
+++ b/test/Transforms/InstCombine/sub.ll
@@ -301,3 +301,29 @@ define i32 @test28(i32 %x, i32 %y, i32 %z) {
 ; CHECK-NEXT: add i32
 ; CHECK-NEXT: ret i32
 }
+
+define i64 @test29(i8* %foo, i64 %i, i64 %j) {
+  %gep1 = getelementptr inbounds i8* %foo, i64 %i
+  %gep2 = getelementptr inbounds i8* %foo, i64 %j
+  %cast1 = ptrtoint i8* %gep1 to i64
+  %cast2 = ptrtoint i8* %gep2 to i64
+  %sub = sub i64 %cast1, %cast2
+  ret i64 %sub
+; CHECK: @test29
+; CHECK-NEXT: sub i64 %i, %j
+; CHECK-NEXT: ret i64
+}
+
+define i64 @test30(i8* %foo, i64 %i, i64 %j) {
+  %bit = bitcast i8* %foo to i32*
+  %gep1 = getelementptr inbounds i32* %bit, i64 %i
+  %gep2 = getelementptr inbounds i8* %foo, i64 %j
+  %cast1 = ptrtoint i32* %gep1 to i64
+  %cast2 = ptrtoint i8* %gep2 to i64
+  %sub = sub i64 %cast1, %cast2
+  ret i64 %sub
+; CHECK: @test30
+; CHECK-NEXT: %gep1.idx = shl nuw i64 %i, 2
+; CHECK-NEXT: sub i64 %gep1.idx, %j
+; CHECK-NEXT: ret i64
+}