diff --git a/lib/Target/ARM/ARMGlobalMerge.cpp b/lib/Target/ARM/ARMGlobalMerge.cpp
index 44d523569eb..72cab6462c4 100644
--- a/lib/Target/ARM/ARMGlobalMerge.cpp
+++ b/lib/Target/ARM/ARMGlobalMerge.cpp
@@ -129,11 +129,14 @@ bool ARMGlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
     uint64_t MergedSize = 0;
     std::vector<const Type*> Tys;
     std::vector<Constant*> Inits;
-    for (j = i; MergedSize < MaxOffset && j != e; ++j) {
+    for (j = i; j != e; ++j) {
       const Type *Ty = Globals[j]->getType()->getElementType();
+      MergedSize += TD->getTypeAllocSize(Ty);
+      if (MergedSize > MaxOffset) {
+        break;
+      }
       Tys.push_back(Ty);
       Inits.push_back(Globals[j]->getInitializer());
-      MergedSize += TD->getTypeAllocSize(Ty);
     }
 
     StructType *MergedTy = StructType::get(M.getContext(), Tys);
diff --git a/test/CodeGen/ARM/global-merge.ll b/test/CodeGen/ARM/global-merge.ll
index dcc16c07ef4..e519dde33a0 100644
--- a/test/CodeGen/ARM/global-merge.ll
+++ b/test/CodeGen/ARM/global-merge.ll
@@ -9,3 +9,9 @@
 ; CHECK: _MergedGlobals:
 @g1 = internal global i32 1
 @g2 = internal global i32 2
+
+; Make sure that the complete variable fits within the range of the maximum
+; offset.  Having the starting offset in range is not sufficient.
+; When this works properly, @g3 is placed in a separate chunk of merged globals.
+; CHECK: _MergedGlobals1:
+@g3 = internal global [30 x i32] [ i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10 ]