diff --git a/llvm/lib/Target/ARM/ARMGlobalMerge.cpp b/llvm/lib/Target/ARM/ARMGlobalMerge.cpp
index 72cab6462c45de1c938666fc8ec86edb2c6ba505..3f0238387a2ba2d0972b317686fdfaeb5e73b1df 100644
--- a/llvm/lib/Target/ARM/ARMGlobalMerge.cpp
+++ b/llvm/lib/Target/ARM/ARMGlobalMerge.cpp
@@ -65,6 +65,7 @@
 #include "llvm/Pass.h"
 #include "llvm/Target/TargetData.h"
 #include "llvm/Target/TargetLowering.h"
+#include "llvm/Target/TargetLoweringObjectFile.h"
 using namespace llvm;
 
 namespace {
@@ -74,7 +75,7 @@ namespace {
     const TargetLowering *TLI;
 
     bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
-                 Module &M, bool) const;
+                 Module &M, bool isConst) const;
 
   public:
     static char ID;             // Pass identification, replacement for typeid.
@@ -161,7 +162,7 @@ bool ARMGlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
 
 
 bool ARMGlobalMerge::doInitialization(Module &M) {
-  SmallVector<GlobalVariable*, 16> Globals, ConstGlobals;
+  SmallVector<GlobalVariable*, 16> Globals, ConstGlobals, BSSGlobals;
   const TargetData *TD = TLI->getTargetData();
   unsigned MaxOffset = TLI->getMaximalGlobalOffset();
   bool Changed = false;
@@ -183,7 +184,10 @@ bool ARMGlobalMerge::doInitialization(Module &M) {
       continue;
 
     if (TD->getTypeAllocSize(I->getType()->getElementType()) < MaxOffset) {
-      if (I->isConstant())
+      const TargetLoweringObjectFile &TLOF = TLI->getObjFileLowering();
+      if (TLOF.getKindForGlobal(I, TLI->getTargetMachine()).isBSSLocal())
+        BSSGlobals.push_back(I);
+      else if (I->isConstant())
         ConstGlobals.push_back(I);
       else
         Globals.push_back(I);
@@ -192,10 +196,12 @@ bool ARMGlobalMerge::doInitialization(Module &M) {
 
   if (Globals.size() > 1)
     Changed |= doMerge(Globals, M, false);
+  if (BSSGlobals.size() > 1)
+    Changed |= doMerge(BSSGlobals, M, false);
+
   // FIXME: This currently breaks the EH processing due to way how the 
   // typeinfo detection works. We might want to detect the TIs and ignore 
   // them in the future.
-  
   // if (ConstGlobals.size() > 1)
   //  Changed |= doMerge(ConstGlobals, M, true);
 
diff --git a/llvm/test/CodeGen/ARM/global-merge.ll b/llvm/test/CodeGen/ARM/global-merge.ll
index e519dde33a0db02c07166ed36255408391b1114e..28bf2214740a71385aba5235a2e3a66ba849d818 100644
--- a/llvm/test/CodeGen/ARM/global-merge.ll
+++ b/llvm/test/CodeGen/ARM/global-merge.ll
@@ -1,4 +1,4 @@
-; RUN: llc < %s -march=thumb | FileCheck %s
+; RUN: llc < %s -mtriple=thumb-apple-darwin | FileCheck %s
 ; Test the ARMGlobalMerge pass.  Use -march=thumb because it has a small
 ; value for the maximum offset (127).
 
@@ -15,3 +15,9 @@
 ; 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 ]
+
+; Global variables that can be placed in BSS should be kept together in a
+; separate pool of merged globals.
+; CHECK: _MergedGlobals2
+@g4 = internal global i32 0
+@g5 = internal global i32 0