diff --git a/lib/CodeGen/SelectionDAG/FastISel.cpp b/lib/CodeGen/SelectionDAG/FastISel.cpp
index f0fe179e050..cff37c2c860 100644
--- a/lib/CodeGen/SelectionDAG/FastISel.cpp
+++ b/lib/CodeGen/SelectionDAG/FastISel.cpp
@@ -437,6 +437,11 @@ bool FastISel::SelectGetElementPtr(const User *I) {
 
   bool NIsKill = hasTrivialKill(I->getOperand(0));
 
+  // Keep a running tab of the total offset to coalesce multiple N = N + Offset
+  // into a single N = N + TotalOffset.
+  uint64_t TotalOffs = 0;
+  // FIXME: What's a good SWAG number for MaxOffs?
+  uint64_t MaxOffs = 2048;
   Type *Ty = I->getOperand(0)->getType();
   MVT VT = TLI.getPointerTy();
   for (GetElementPtrInst::const_op_iterator OI = I->op_begin()+1,
@@ -446,14 +451,15 @@ bool FastISel::SelectGetElementPtr(const User *I) {
       unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
       if (Field) {
         // N = N + Offset
-        uint64_t Offs = TD.getStructLayout(StTy)->getElementOffset(Field);
-        // FIXME: This can be optimized by combining the add with a
-        // subsequent one.
-        N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, Offs, VT);
-        if (N == 0)
-          // Unhandled operand. Halt "fast" selection and bail.
-          return false;
-        NIsKill = true;
+        TotalOffs += TD.getStructLayout(StTy)->getElementOffset(Field);
+        if (TotalOffs >= MaxOffs) {
+          N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
+          if (N == 0)
+            // Unhandled operand. Halt "fast" selection and bail.
+            return false;
+          NIsKill = true;
+          TotalOffs = 0;
+        }
       }
       Ty = StTy->getElementType(Field);
     } else {
@@ -462,14 +468,26 @@ bool FastISel::SelectGetElementPtr(const User *I) {
       // If this is a constant subscript, handle it quickly.
       if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
         if (CI->isZero()) continue;
-        uint64_t Offs =
+        // N = N + Offset
+        TotalOffs += 
           TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
-        N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, Offs, VT);
+        if (TotalOffs >= MaxOffs) {
+          N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
+          if (N == 0)
+            // Unhandled operand. Halt "fast" selection and bail.
+            return false;
+          NIsKill = true;
+          TotalOffs = 0;
+        }
+        continue;
+      }
+      if (TotalOffs) {
+        N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
         if (N == 0)
           // Unhandled operand. Halt "fast" selection and bail.
           return false;
         NIsKill = true;
-        continue;
+        TotalOffs = 0;
       }
 
       // N = N + Idx * ElementSize;
@@ -494,6 +512,12 @@ bool FastISel::SelectGetElementPtr(const User *I) {
         return false;
     }
   }
+  if (TotalOffs) {
+    N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
+    if (N == 0)
+      // Unhandled operand. Halt "fast" selection and bail.
+      return false;
+  }
 
   // We successfully emitted code for the given LLVM Instruction.
   UpdateValueMap(I, N);
diff --git a/test/CodeGen/ARM/fast-isel-GEP-coalesce.ll b/test/CodeGen/ARM/fast-isel-GEP-coalesce.ll
new file mode 100644
index 00000000000..dbb634df0a1
--- /dev/null
+++ b/test/CodeGen/ARM/fast-isel-GEP-coalesce.ll
@@ -0,0 +1,65 @@
+; RUN: llc < %s -O0 -fast-isel-abort -relocation-model=dynamic-no-pic -mtriple=armv7-apple-darwin | FileCheck %s --check-prefix=ARM
+; RUN: llc < %s -O0 -fast-isel-abort -relocation-model=dynamic-no-pic -mtriple=thumbv7-apple-darwin | FileCheck %s --check-prefix=THUMB
+
+%struct.A = type { i32, [2 x [2 x i32]], i8, [3 x [3 x [3 x i32]]] }
+%struct.B = type { i32, [2 x [2 x [2 x %struct.A]]] }
+
+@arr = common global [2 x [2 x [2 x [2 x [2 x i32]]]]] zeroinitializer, align 4
+@A = common global [3 x [3 x %struct.A]] zeroinitializer, align 4
+@B = common global [2 x [2 x [2 x %struct.B]]] zeroinitializer, align 4
+
+define i32* @t1() nounwind {
+entry:
+; ARM: t1
+; THUMB: t1
+  %addr = alloca i32*, align 4
+  store i32* getelementptr inbounds ([2 x [2 x [2 x [2 x [2 x i32]]]]]* @arr, i32 0, i32 1, i32 1, i32 1, i32 1, i32 1), i32** %addr, align 4
+; ARM: add r0, r0, #124
+; THUMB: adds r0, #124
+  %0 = load i32** %addr, align 4
+  ret i32* %0
+}
+
+define i32* @t2() nounwind {
+entry:
+; ARM: t2
+; THUMB: t2
+  %addr = alloca i32*, align 4
+  store i32* getelementptr inbounds ([3 x [3 x %struct.A]]* @A, i32 0, i32 2, i32 2, i32 3, i32 1, i32 2, i32 2), i32** %addr, align 4
+; ARM: movw r1, #1148
+; ARM: add r0, r0, r1
+; THUMB: addw r0, r0, #1148
+  %0 = load i32** %addr, align 4
+  ret i32* %0
+}
+
+define i32* @t3() nounwind {
+entry:
+; ARM: t3
+; THUMB: t3
+  %addr = alloca i32*, align 4
+  store i32* getelementptr inbounds ([3 x [3 x %struct.A]]* @A, i32 0, i32 0, i32 1, i32 1, i32 0, i32 1), i32** %addr, align 4
+; ARM: add r0, r0, #140
+; THUMB: adds r0, #140
+  %0 = load i32** %addr, align 4
+  ret i32* %0
+}
+
+define i32* @t4() nounwind {
+entry:
+; ARM: t4
+; THUMB: t4
+  %addr = alloca i32*, align 4
+  store i32* getelementptr inbounds ([2 x [2 x [2 x %struct.B]]]* @B, i32 0, i32 0, i32 0, i32 1, i32 1, i32 0, i32 0, i32 1, i32 3, i32 1, i32 2, i32 1), i32** %addr, align 4
+; ARM-NOT: movw r{{[0-9]}}, #1060
+; ARM-NOT: add r{{[0-9]}}, r{{[0-9]}}, #4
+; ARM-NOT: add r{{[0-9]}}, r{{[0-9]}}, #132
+; ARM-NOT: add r{{[0-9]}}, r{{[0-9]}}, #24
+; ARM-NOT: add r{{[0-9]}}, r{{[0-9]}}, #36
+; ARM-NOT: add r{{[0-9]}}, r{{[0-9]}}, #24
+; ARM-NOT: add r{{[0-9]}}, r{{[0-9]}}, #4
+; ARM: movw r{{[0-9]}}, #1284
+; THUMB: addw r{{[0-9]}}, r{{[0-9]}}, #1284
+  %0 = load i32** %addr, align 4
+  ret i32* %0
+}