Index: lib/Demangle/ItaniumDemangle.cpp
===================================================================
--- lib/Demangle/ItaniumDemangle.cpp
+++ lib/Demangle/ItaniumDemangle.cpp
@@ -1945,40 +1945,53 @@
   struct BlockMeta {
     BlockMeta* Next;
     size_t Current;
+
+    BlockMeta(BlockMeta *N) : Next(N),
+      Current(Alignment -
+              (reinterpret_cast<size_t>(this + 1) & (Alignment - 1))) { }
+    void *getTail() { return reinterpret_cast<char*>(this + 1) + Current; }
   };
 
+  static constexpr size_t Alignment = 16;
   static constexpr size_t AllocSize = 4096;
-  static constexpr size_t UsableAllocSize = AllocSize - sizeof(BlockMeta);
+  static constexpr size_t UsableAllocSize = AllocSize - sizeof(BlockMeta) -
+                                            Alignment;
 
-  alignas(16) char InitialBuffer[AllocSize];
+  char InitialBuffer[AllocSize];
   BlockMeta* BlockList = nullptr;
 
   void grow() {
     char* NewMeta = new char[AllocSize];
-    BlockList = new (NewMeta) BlockMeta{BlockList, 0};
+    BlockList = new (NewMeta) BlockMeta(BlockList);
   }
 
   void* allocateMassive(size_t NBytes) {
-    NBytes += sizeof(BlockMeta);
-    BlockMeta* NewMeta = reinterpret_cast<BlockMeta*>(new char[NBytes]);
-    BlockList->Next = new (NewMeta) BlockMeta{BlockList->Next, 0};
-    return static_cast<void*>(NewMeta + 1);
+    size_t AllocSz = NBytes + sizeof(BlockMeta) + Alignment;
+    char *Mem = new char[AllocSz];
+    BlockList->Next = new (Mem) BlockMeta(BlockList->Next);
+    void *Result = BlockList->getTail();
+    BlockList->Current += NBytes;
+    return Result;
+  }
+
+  static constexpr size_t getAlignedSize(size_t Sz) {
+    return (Sz + (Alignment - 1)) & ~(Alignment - 1);
   }
 
 public:
   BumpPointerAllocator()
-      : BlockList(new (InitialBuffer) BlockMeta{nullptr, 0}) {}
+      : BlockList(new (InitialBuffer) BlockMeta(nullptr)) {}
 
   void* allocate(size_t N) {
-    N = (N + 15u) & ~15u;
+    N = getAlignedSize(N);
     if (N + BlockList->Current >= UsableAllocSize) {
       if (N > UsableAllocSize)
         return allocateMassive(N);
       grow();
     }
+    void *Result = BlockList->getTail();
     BlockList->Current += N;
-    return static_cast<void*>(reinterpret_cast<char*>(BlockList + 1) +
-                              BlockList->Current - N);
+    return Result;
   }
 
   void reset() {
@@ -1988,7 +2001,7 @@
       if (reinterpret_cast<char*>(Tmp) != InitialBuffer)
         delete[] reinterpret_cast<char*>(Tmp);
     }
-    BlockList = new (InitialBuffer) BlockMeta{nullptr, 0};
+    BlockList = new (InitialBuffer) BlockMeta(nullptr);
   }
 
   ~BumpPointerAllocator() { reset(); }