In response to bug #20362, change the order of operations in vector move assignment so that if the allocator move assignment throws, we aren't left with two objects pointing at the same memory. This is not a complete fix; I am unconvinced that a complete fix is possible. With this change in place, we will leak the old contents of the vector. LWG issue #2106, when adopted, will make this problem illegal. Thanks to Thomas Koeppe for the report and analysis.

git-svn-id: https://llvm.org/svn/llvm-project/libcxx/trunk@213546 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/include/vector b/include/vector
index 2cc23e5..d370214 100644
--- a/include/vector
+++ b/include/vector
@@ -1331,10 +1331,10 @@
     _NOEXCEPT_(is_nothrow_move_assignable<allocator_type>::value)
 {
     deallocate();
+    __base::__move_assign_alloc(__c); // this can throw
     this->__begin_ = __c.__begin_;
     this->__end_ = __c.__end_;
     this->__end_cap() = __c.__end_cap();
-    __base::__move_assign_alloc(__c);
     __c.__begin_ = __c.__end_ = __c.__end_cap() = nullptr;
 #if _LIBCPP_DEBUG_LEVEL >= 2
     __get_db()->swap(this, &__c);