diff --git a/lib/CodeGen/CodePlacementOpt.cpp b/lib/CodeGen/CodePlacementOpt.cpp
index 2e1d12d2346..61a8b12860a 100644
--- a/lib/CodeGen/CodePlacementOpt.cpp
+++ b/lib/CodeGen/CodePlacementOpt.cpp
@@ -62,6 +62,8 @@ namespace {
 
   private:
     bool OptimizeIntraLoopEdges();
+    bool HeaderShouldBeAligned(MachineBasicBlock *MBB, MachineLoop *L,
+                               SmallPtrSet<MachineBasicBlock*, 4> &DoNotAlign);
     bool AlignLoops(MachineFunction &MF);
   };
 
@@ -244,14 +246,37 @@ bool CodePlacementOpt::OptimizeIntraLoopEdges() {
 /// should be aligned. For now, we will not align it if all the predcessors
 /// (i.e. loop back edges) are laid out above the header. FIXME: Do not
 /// align small loops.
-static bool HeaderShouldBeAligned(MachineBasicBlock *MBB) {
+bool
+CodePlacementOpt::HeaderShouldBeAligned(MachineBasicBlock *MBB, MachineLoop *L,
+                               SmallPtrSet<MachineBasicBlock*, 4> &DoNotAlign) {
+  if (DoNotAlign.count(MBB))
+    return false;
+
+  bool BackEdgeBelow = false;
   for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
          PE = MBB->pred_end(); PI != PE; ++PI) {
     MachineBasicBlock *PredMBB = *PI;
-    if (PredMBB == MBB || PredMBB->getNumber() > MBB->getNumber())
-      return true;
+    if (PredMBB == MBB || PredMBB->getNumber() > MBB->getNumber()) {
+      BackEdgeBelow = true;
+      break;
+    }
   }
-  return false;
+
+  if (!BackEdgeBelow)
+    return false;
+
+  // Ok, we are going to align this loop header. If it's an inner loop,
+  // do not align its outer loop.
+  MachineBasicBlock *PreHeader = L->getLoopPreheader();
+  if (PreHeader) {
+    MachineLoop *L = MLI->getLoopFor(PreHeader);
+    if (L) {
+      MachineBasicBlock *HeaderBlock = L->getHeader();
+      HeaderBlock->setAlignment(0);
+      DoNotAlign.insert(HeaderBlock);
+    }
+  }
+  return true;
 }
 
 /// AlignLoops - Align loop headers to target preferred alignments.
@@ -269,14 +294,16 @@ bool CodePlacementOpt::AlignLoops(MachineFunction &MF) {
   MF.RenumberBlocks();
 
   bool Changed = false;
+  SmallPtrSet<MachineBasicBlock*, 4> DoNotAlign;
   for (unsigned i = 0, e = LoopHeaders.size(); i != e; ++i) {
     MachineBasicBlock *HeaderMBB = LoopHeaders[i];
     MachineBasicBlock *PredMBB = prior(MachineFunction::iterator(HeaderMBB));
-    if (MLI->getLoopFor(HeaderMBB) == MLI->getLoopFor(PredMBB))
+    MachineLoop *L = MLI->getLoopFor(HeaderMBB);
+    if (L == MLI->getLoopFor(PredMBB))
       // If previously BB is in the same loop, don't align this BB. We want
       // to prevent adding noop's inside a loop.
       continue;
-    if (HeaderShouldBeAligned(HeaderMBB)) {
+    if (HeaderShouldBeAligned(HeaderMBB, L, DoNotAlign)) {
       HeaderMBB->setAlignment(Align);
       Changed = true;
       ++NumHeaderAligned;
diff --git a/test/CodeGen/X86/avoid-loop-align-2.ll b/test/CodeGen/X86/avoid-loop-align-2.ll
new file mode 100644
index 00000000000..9f0aeb32c41
--- /dev/null
+++ b/test/CodeGen/X86/avoid-loop-align-2.ll
@@ -0,0 +1,45 @@
+; RUN: llvm-as < %s | llc -march=x86 | grep align | count 3
+
+@x = external global i32*		; <i32**> [#uses=1]
+
+define i32 @t(i32 %a, i32 %b) nounwind readonly ssp {
+entry:
+	%0 = icmp eq i32 %a, 0		; <i1> [#uses=1]
+	br i1 %0, label %bb5, label %bb.nph12
+
+bb.nph12:		; preds = %entry
+	%1 = icmp eq i32 %b, 0		; <i1> [#uses=1]
+	%2 = load i32** @x, align 8		; <i32*> [#uses=1]
+	br i1 %1, label %bb2.preheader, label %bb2.preheader.us
+
+bb2.preheader.us:		; preds = %bb2.bb3_crit_edge.us, %bb.nph12
+	%indvar18 = phi i32 [ 0, %bb.nph12 ], [ %indvar.next19, %bb2.bb3_crit_edge.us ]		; <i32> [#uses=2]
+	%sum.111.us = phi i32 [ 0, %bb.nph12 ], [ %4, %bb2.bb3_crit_edge.us ]		; <i32> [#uses=0]
+	%tmp16 = mul i32 %indvar18, %a		; <i32> [#uses=1]
+	br label %bb1.us
+
+bb1.us:		; preds = %bb1.us, %bb2.preheader.us
+	%indvar = phi i32 [ 0, %bb2.preheader.us ], [ %indvar.next, %bb1.us ]		; <i32> [#uses=2]
+	%tmp17 = add i32 %indvar, %tmp16		; <i32> [#uses=1]
+	%tmp. = zext i32 %tmp17 to i64		; <i64> [#uses=1]
+	%3 = getelementptr i32* %2, i64 %tmp.		; <i32*> [#uses=1]
+	%4 = load i32* %3, align 4		; <i32> [#uses=2]
+	%indvar.next = add i32 %indvar, 1		; <i32> [#uses=2]
+	%exitcond = icmp eq i32 %indvar.next, %b		; <i1> [#uses=1]
+	br i1 %exitcond, label %bb2.bb3_crit_edge.us, label %bb1.us
+
+bb2.bb3_crit_edge.us:		; preds = %bb1.us
+	%indvar.next19 = add i32 %indvar18, 1		; <i32> [#uses=2]
+	%exitcond22 = icmp eq i32 %indvar.next19, %a		; <i1> [#uses=1]
+	br i1 %exitcond22, label %bb5, label %bb2.preheader.us
+
+bb2.preheader:		; preds = %bb2.preheader, %bb.nph12
+	%indvar24 = phi i32 [ %indvar.next25, %bb2.preheader ], [ 0, %bb.nph12 ]		; <i32> [#uses=1]
+	%indvar.next25 = add i32 %indvar24, 1		; <i32> [#uses=2]
+	%exitcond28 = icmp eq i32 %indvar.next25, %a		; <i1> [#uses=1]
+	br i1 %exitcond28, label %bb5, label %bb2.preheader
+
+bb5:		; preds = %bb2.preheader, %bb2.bb3_crit_edge.us, %entry
+	%sum.1.lcssa = phi i32 [ 0, %entry ], [ 0, %bb2.preheader ], [ %4, %bb2.bb3_crit_edge.us ]		; <i32> [#uses=1]
+	ret i32 %sum.1.lcssa
+}