diff --git a/lib/Target/ARM/ARM.h b/lib/Target/ARM/ARM.h
index 21445adff97..ff1980d19df 100644
--- a/lib/Target/ARM/ARM.h
+++ b/lib/Target/ARM/ARM.h
@@ -109,6 +109,7 @@ FunctionPass *createNEONPreAllocPass();
 FunctionPass *createNEONMoveFixPass();
 FunctionPass *createThumb2ITBlockPass();
 FunctionPass *createThumb2SizeReductionPass();
+FunctionPass *createARMMaxStackAlignmentCalculatorPass();
 
 extern Target TheARMTarget, TheThumbTarget;
 
diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index 19ffcf72eed..828041c64ec 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -1132,6 +1132,7 @@ bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
       break;
     }
     case ARMII::AddrMode4:
+    case ARMII::AddrMode6:
       // Can't fold any offset even if it's zero.
       return false;
     case ARMII::AddrMode5: {
diff --git a/lib/Target/ARM/ARMBaseRegisterInfo.cpp b/lib/Target/ARM/ARMBaseRegisterInfo.cpp
index 9b4368f47dc..c3b1000eb2a 100644
--- a/lib/Target/ARM/ARMBaseRegisterInfo.cpp
+++ b/lib/Target/ARM/ARMBaseRegisterInfo.cpp
@@ -1170,7 +1170,8 @@ ARMBaseRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
   // as much as possible above, handle the rest, providing a register that is
   // SP+LargeImm.
   assert((Offset ||
-          (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode4) &&
+          (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode4 ||
+          (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode6) &&
          "This code isn't needed if offset already handled!");
 
   unsigned ScratchReg = 0;
@@ -1179,7 +1180,7 @@ ARMBaseRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
     ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImm();
   unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg();
   if (Offset == 0)
-    // Must be addrmode4.
+    // Must be addrmode4/6.
     MI.getOperand(i).ChangeToRegister(FrameReg, false, false, false);
   else {
     ScratchReg = MF.getRegInfo().createVirtualRegister(ARM::GPRRegisterClass);
@@ -1462,4 +1463,46 @@ emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const {
     emitSPUpdate(isARM, MBB, MBBI, dl, TII, VARegSaveSize);
 }
 
+namespace {
+  struct MSAC : public MachineFunctionPass {
+    static char ID;
+    MSAC() : MachineFunctionPass(&ID) {}
+
+    virtual bool runOnMachineFunction(MachineFunction &MF) {
+      MachineFrameInfo *FFI = MF.getFrameInfo();
+      MachineRegisterInfo &RI = MF.getRegInfo();
+
+      // Calculate max stack alignment of all already allocated stack objects.
+      unsigned MaxAlign = calculateMaxStackAlignment(FFI);
+
+      // Be over-conservative: scan over all vreg defs and find, whether vector
+      // registers are used. If yes - there is probability, that vector register
+      // will be spilled and thus stack needs to be aligned properly.
+      for (unsigned RegNum = TargetRegisterInfo::FirstVirtualRegister;
+           RegNum < RI.getLastVirtReg(); ++RegNum)
+        MaxAlign = std::max(MaxAlign, RI.getRegClass(RegNum)->getAlignment());
+
+      if (FFI->getMaxAlignment() == MaxAlign)
+        return false;
+
+      FFI->setMaxAlignment(MaxAlign);
+      return true;
+    }
+
+    virtual const char *getPassName() const {
+      return "ARM Maximal Stack Alignment Calculator";
+    }
+
+    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+      AU.setPreservesCFG();
+      MachineFunctionPass::getAnalysisUsage(AU);
+    }
+  };
+
+  char MSAC::ID = 0;
+}
+
+FunctionPass*
+llvm::createARMMaxStackAlignmentCalculatorPass() { return new MSAC(); }
+
 #include "ARMGenRegisterInfo.inc"
diff --git a/lib/Target/ARM/ARMTargetMachine.cpp b/lib/Target/ARM/ARMTargetMachine.cpp
index 4d20a5c1a03..2564ed92547 100644
--- a/lib/Target/ARM/ARMTargetMachine.cpp
+++ b/lib/Target/ARM/ARMTargetMachine.cpp
@@ -93,6 +93,10 @@ bool ARMBaseTargetMachine::addPreRegAlloc(PassManagerBase &PM,
   if (Subtarget.hasNEON())
     PM.add(createNEONPreAllocPass());
 
+  // Calculate and set max stack object alignment early, so we can decide
+  // whether we will need stack realignment (and thus FP).
+  PM.add(createARMMaxStackAlignmentCalculatorPass());
+
   // FIXME: temporarily disabling load / store optimization pass for Thumb1.
   if (OptLevel != CodeGenOpt::None && !Subtarget.isThumb1Only())
     PM.add(createARMLoadStoreOptimizationPass(true));
diff --git a/test/CodeGen/ARM/spill-q.ll b/test/CodeGen/ARM/spill-q.ll
index f4b27a7603e..6b44210af74 100644
--- a/test/CodeGen/ARM/spill-q.ll
+++ b/test/CodeGen/ARM/spill-q.ll
@@ -11,8 +11,9 @@ declare <4 x float> @llvm.arm.neon.vld1.v4f32(i8*) nounwind readonly
 
 define arm_apcscc void @aaa(%quuz* %this, i8* %block) {
 ; CHECK: aaa:
-; CHECK: vstmia sp
-; CHECK: vldmia sp
+; CHECK: bic sp, sp, #15
+; CHECK: vst1.64 {{.*}}sp @128
+; CHECK: vld1.64 {{.*}}sp @128
 entry:
   %0 = call <4 x float> @llvm.arm.neon.vld1.v4f32(i8* undef) nounwind ; <<4 x float>> [#uses=1]
   store float 6.300000e+01, float* undef, align 4
diff --git a/test/CodeGen/Thumb2/thumb2-spill-q.ll b/test/CodeGen/Thumb2/thumb2-spill-q.ll
index 0a7221c6174..6c6a7e19af0 100644
--- a/test/CodeGen/Thumb2/thumb2-spill-q.ll
+++ b/test/CodeGen/Thumb2/thumb2-spill-q.ll
@@ -11,8 +11,9 @@ declare <4 x float> @llvm.arm.neon.vld1.v4f32(i8*) nounwind readonly
 
 define arm_apcscc void @aaa(%quuz* %this, i8* %block) {
 ; CHECK: aaa:
-; CHECK: vstmia sp
-; CHECK: vldmia sp
+; CHECK: bic sp, sp, #15
+; CHECK: vst1.64 {{.*}}sp @128
+; CHECK: vld1.64 {{.*}}sp @128
 entry:
   %0 = call <4 x float> @llvm.arm.neon.vld1.v4f32(i8* undef) nounwind ; <<4 x float>> [#uses=1]
   store float 6.300000e+01, float* undef, align 4