diff --git a/include/boost/coroutine2/detail/create_control_block.ipp b/include/boost/coroutine2/detail/create_control_block.ipp
index 9eaea2e..127ea84 100644
--- a/include/boost/coroutine2/detail/create_control_block.ipp
+++ b/include/boost/coroutine2/detail/create_control_block.ipp
@@ -28,7 +28,7 @@ namespace coroutines2 {
 namespace detail {
 
 template< typename ControlBlock, typename StackAllocator, typename Fn >
-ControlBlock * create_control_block( StackAllocator salloc, Fn && fn) {
+ControlBlock * create_control_block( StackAllocator && salloc, Fn && fn) {
     auto sctx = salloc.allocate();
     // reserve space for control structure
 #if defined(BOOST_NO_CXX11_CONSTEXPR) || defined(BOOST_NO_CXX11_STD_ALIGN)
@@ -48,7 +48,7 @@ ControlBlock * create_control_block( StackAllocator salloc, Fn && fn) {
 #endif
     // placment new for control structure on coroutine stack
     return new ( sp) ControlBlock{ context::preallocated( sp, size, sctx),
-                                   salloc, std::forward< Fn >( fn) };
+                                   std::forward< StackAllocator >( salloc), std::forward< Fn >( fn) };
 }
 
 }}}
diff --git a/include/boost/coroutine2/detail/pull_control_block_cc.hpp b/include/boost/coroutine2/detail/pull_control_block_cc.hpp
index 5353fb8..b9ac614 100644
--- a/include/boost/coroutine2/detail/pull_control_block_cc.hpp
+++ b/include/boost/coroutine2/detail/pull_control_block_cc.hpp
@@ -35,7 +35,7 @@ struct pull_coroutine< T >::control_block {
     static void destroy( control_block * cb) noexcept;
 
     template< typename StackAllocator, typename Fn >
-    control_block( context::preallocated, StackAllocator, Fn &&);
+    control_block( context::preallocated, StackAllocator &&, Fn &&);
 
     control_block( typename push_coroutine< T >::control_block *, boost::context::continuation &) noexcept;
 
@@ -76,7 +76,7 @@ struct pull_coroutine< T & >::control_block {
     static void destroy( control_block * cb) noexcept;
 
     template< typename StackAllocator, typename Fn >
-    control_block( context::preallocated, StackAllocator, Fn &&);
+    control_block( context::preallocated, StackAllocator &&, Fn &&);
 
     control_block( typename push_coroutine< T & >::control_block *, boost::context::continuation &) noexcept;
 
@@ -103,7 +103,7 @@ struct pull_coroutine< void >::control_block {
     static void destroy( control_block * cb) noexcept;
 
     template< typename StackAllocator, typename Fn >
-    control_block( context::preallocated, StackAllocator, Fn &&);
+    control_block( context::preallocated, StackAllocator &&, Fn &&);
 
     control_block( push_coroutine< void >::control_block *, boost::context::continuation &) noexcept;
 
diff --git a/include/boost/coroutine2/detail/pull_control_block_cc.ipp b/include/boost/coroutine2/detail/pull_control_block_cc.ipp
index a7dd68e..43fdb76 100644
--- a/include/boost/coroutine2/detail/pull_control_block_cc.ipp
+++ b/include/boost/coroutine2/detail/pull_control_block_cc.ipp
@@ -43,7 +43,7 @@ pull_coroutine< T >::control_block::destroy( control_block * cb) noexcept {
 
 template< typename T >
 template< typename StackAllocator, typename Fn >
-pull_coroutine< T >::control_block::control_block( context::preallocated palloc, StackAllocator salloc,
+pull_coroutine< T >::control_block::control_block( context::preallocated palloc, StackAllocator && salloc,
                                                    Fn && fn) :
     c{},
     other{ nullptr },
@@ -53,7 +53,7 @@ pull_coroutine< T >::control_block::control_block( context::preallocated palloc,
     storage{} {
 #if defined(BOOST_NO_CXX14_GENERIC_LAMBDAS)
     c = boost::context::callcc(
-            std::allocator_arg, palloc, salloc,
+            std::allocator_arg, palloc, std::forward< StackAllocator >( salloc),
             wrap( [this](typename std::decay< Fn >::type & fn_,boost::context::continuation && c) mutable {
                     // create synthesized push_coroutine< T >
                     typename push_coroutine< T >::control_block synthesized_cb{ this, c };
@@ -79,7 +79,7 @@ pull_coroutine< T >::control_block::control_block( context::preallocated palloc,
                  std::forward< Fn >( fn) ) );
 #else
     c = boost::context::callcc(
-            std::allocator_arg, palloc, salloc,
+            std::allocator_arg, palloc, std::forward< StackAllocator >( salloc),
             [this,fn_=std::forward< Fn >( fn)](boost::context::continuation && c) mutable {
                // create synthesized push_coroutine< T >
                typename push_coroutine< T >::control_block synthesized_cb{ this, c };
@@ -193,7 +193,7 @@ pull_coroutine< T & >::control_block::destroy( control_block * cb) noexcept {
 
 template< typename T >
 template< typename StackAllocator, typename Fn >
-pull_coroutine< T & >::control_block::control_block( context::preallocated palloc, StackAllocator salloc,
+pull_coroutine< T & >::control_block::control_block( context::preallocated palloc, StackAllocator && salloc,
                                                      Fn && fn) :
     c{},
     other{ nullptr },
@@ -203,7 +203,7 @@ pull_coroutine< T & >::control_block::control_block( context::preallocated pallo
     storage{} {
 #if defined(BOOST_NO_CXX14_GENERIC_LAMBDAS)
     c = boost::context::callcc(
-            std::allocator_arg, palloc, salloc,
+            std::allocator_arg, palloc, std::forward< StackAllocator >( salloc),
             wrap( [this](typename std::decay< Fn >::type & fn_,boost::context::continuation && c) mutable {
                     // create synthesized push_coroutine< T & >
                     typename push_coroutine< T & >::control_block synthesized_cb{ this, c };
@@ -229,7 +229,7 @@ pull_coroutine< T & >::control_block::control_block( context::preallocated pallo
                  std::forward< Fn >( fn) ) );
 #else
     c = boost::context::callcc(
-            std::allocator_arg, palloc, salloc,
+            std::allocator_arg, palloc, std::forward< StackAllocator >( salloc),
             [this,fn_=std::forward< Fn >( fn)](boost::context::continuation && c) mutable {
                // create synthesized push_coroutine< T & >
                typename push_coroutine< T & >::control_block synthesized_cb{ this, c };
@@ -319,7 +319,7 @@ pull_coroutine< void >::control_block::destroy( control_block * cb) noexcept {
 }
 
 template< typename StackAllocator, typename Fn >
-pull_coroutine< void >::control_block::control_block( context::preallocated palloc, StackAllocator salloc,
+pull_coroutine< void >::control_block::control_block( context::preallocated palloc, StackAllocator && salloc,
                                                       Fn && fn) :
     c{},
     other{ nullptr },
@@ -327,7 +327,7 @@ pull_coroutine< void >::control_block::control_block( context::preallocated pall
     except{} {
 #if defined(BOOST_NO_CXX14_GENERIC_LAMBDAS)
     c = boost::context::callcc(
-            std::allocator_arg, palloc, salloc,
+            std::allocator_arg, palloc, std::forward< StackAllocator >( salloc),
             wrap( [this](typename std::decay< Fn >::type & fn_,boost::context::continuation && c) mutable {
                     // create synthesized push_coroutine< void >
                     typename push_coroutine< void >::control_block synthesized_cb{ this, c };
@@ -353,7 +353,7 @@ pull_coroutine< void >::control_block::control_block( context::preallocated pall
                  std::forward< Fn >( fn) ) );
 #else
     c = boost::context::callcc(
-            std::allocator_arg, palloc, salloc,
+            std::allocator_arg, palloc, std::forward< StackAllocator >( salloc),
             [this,fn_=std::forward< Fn >( fn)]( boost::context::continuation && c) mutable {
                // create synthesized push_coroutine< void >
                typename push_coroutine< void >::control_block synthesized_cb{ this, c };
diff --git a/include/boost/coroutine2/detail/pull_coroutine.hpp b/include/boost/coroutine2/detail/pull_coroutine.hpp
index f4c4073..32cd68e 100644
--- a/include/boost/coroutine2/detail/pull_coroutine.hpp
+++ b/include/boost/coroutine2/detail/pull_coroutine.hpp
@@ -45,7 +45,7 @@ public:
     explicit pull_coroutine( Fn &&);
 
     template< typename StackAllocator, typename Fn >
-    pull_coroutine( StackAllocator, Fn &&);
+    pull_coroutine( StackAllocator &&, Fn &&);
 
     ~pull_coroutine();
 
@@ -157,7 +157,7 @@ public:
     explicit pull_coroutine( Fn &&);
 
     template< typename StackAllocator, typename Fn >
-    pull_coroutine( StackAllocator, Fn &&);
+    pull_coroutine( StackAllocator &&, Fn &&);
 
     ~pull_coroutine();
 
@@ -267,7 +267,7 @@ public:
     explicit pull_coroutine( Fn &&);
 
     template< typename StackAllocator, typename Fn >
-    pull_coroutine( StackAllocator, Fn &&);
+    pull_coroutine( StackAllocator &&, Fn &&);
 
     ~pull_coroutine();
 
diff --git a/include/boost/coroutine2/detail/pull_coroutine.ipp b/include/boost/coroutine2/detail/pull_coroutine.ipp
index 9a81893..699d5ab 100644
--- a/include/boost/coroutine2/detail/pull_coroutine.ipp
+++ b/include/boost/coroutine2/detail/pull_coroutine.ipp
@@ -50,8 +50,8 @@ pull_coroutine< T >::pull_coroutine( Fn && fn) :
 
 template< typename T >
 template< typename StackAllocator, typename Fn >
-pull_coroutine< T >::pull_coroutine( StackAllocator salloc, Fn && fn) :
-    cb_{ create_control_block< control_block >( salloc, std::forward< Fn >( fn) ) } {
+pull_coroutine< T >::pull_coroutine( StackAllocator && salloc, Fn && fn) :
+    cb_{ create_control_block< control_block >( std::forward< StackAllocator >( salloc), std::forward< Fn >( fn) ) } {
     if ( ! cb_->valid() ) {
         cb_->deallocate();
         cb_ = nullptr;
@@ -119,8 +119,8 @@ pull_coroutine< T & >::pull_coroutine( Fn && fn) :
 
 template< typename T >
 template< typename StackAllocator, typename Fn >
-pull_coroutine< T & >::pull_coroutine( StackAllocator salloc, Fn && fn) :
-    cb_{ create_control_block< control_block >( salloc, std::forward< Fn >( fn) ) } {
+pull_coroutine< T & >::pull_coroutine( StackAllocator && salloc, Fn && fn) :
+    cb_{ create_control_block< control_block >( std::forward< StackAllocator >( salloc), std::forward< Fn >( fn) ) } {
     if ( ! cb_->valid() ) {
         cb_->deallocate();
         cb_ = nullptr;
@@ -180,8 +180,8 @@ pull_coroutine< void >::pull_coroutine( Fn && fn) :
 }
 
 template< typename StackAllocator, typename Fn >
-pull_coroutine< void >::pull_coroutine( StackAllocator salloc, Fn && fn) :
-    cb_{ create_control_block< control_block >( salloc, std::forward< Fn >( fn) ) } {
+pull_coroutine< void >::pull_coroutine( StackAllocator && salloc, Fn && fn) :
+    cb_{ create_control_block< control_block >( std::forward< StackAllocator >( salloc), std::forward< Fn >( fn) ) } {
     if ( ! cb_->valid() ) {
         cb_->deallocate();
         cb_ = nullptr;
diff --git a/include/boost/coroutine2/detail/push_control_block_cc.hpp b/include/boost/coroutine2/detail/push_control_block_cc.hpp
index d01c98f..8dda44f 100644
--- a/include/boost/coroutine2/detail/push_control_block_cc.hpp
+++ b/include/boost/coroutine2/detail/push_control_block_cc.hpp
@@ -32,7 +32,7 @@ struct push_coroutine< T >::control_block {
     static void destroy( control_block * cb) noexcept;
 
     template< typename StackAllocator, typename Fn >
-    control_block( context::preallocated, StackAllocator, Fn &&);
+    control_block( context::preallocated, StackAllocator &&, Fn &&);
 
     control_block( typename pull_coroutine< T >::control_block *, boost::context::continuation &) noexcept;
 
@@ -58,7 +58,7 @@ struct push_coroutine< T & >::control_block {
     static void destroy( control_block * cb) noexcept;
 
     template< typename StackAllocator, typename Fn >
-    control_block( context::preallocated, StackAllocator, Fn &&);
+    control_block( context::preallocated, StackAllocator &&, Fn &&);
 
     control_block( typename pull_coroutine< T & >::control_block *, boost::context::continuation &) noexcept;
 
@@ -81,7 +81,7 @@ struct push_coroutine< void >::control_block {
     static void destroy( control_block * cb) noexcept;
 
     template< typename StackAllocator, typename Fn >
-    control_block( context::preallocated, StackAllocator, Fn &&);
+    control_block( context::preallocated, StackAllocator &&, Fn &&);
 
     control_block( pull_coroutine< void >::control_block *, boost::context::continuation &) noexcept;
 
diff --git a/include/boost/coroutine2/detail/push_control_block_cc.ipp b/include/boost/coroutine2/detail/push_control_block_cc.ipp
index 98b24f2..ae7dee1 100644
--- a/include/boost/coroutine2/detail/push_control_block_cc.ipp
+++ b/include/boost/coroutine2/detail/push_control_block_cc.ipp
@@ -42,7 +42,7 @@ push_coroutine< T >::control_block::destroy( control_block * cb) noexcept {
 
 template< typename T >
 template< typename StackAllocator, typename Fn >
-push_coroutine< T >::control_block::control_block( context::preallocated palloc, StackAllocator salloc,
+push_coroutine< T >::control_block::control_block( context::preallocated palloc, StackAllocator && salloc,
                                                    Fn && fn) :
     c{},
     other{ nullptr },
@@ -50,7 +50,7 @@ push_coroutine< T >::control_block::control_block( context::preallocated palloc,
     except{} {
 #if defined(BOOST_NO_CXX14_GENERIC_LAMBDAS)
     c = boost::context::callcc(
-            std::allocator_arg, palloc, salloc,
+            std::allocator_arg, palloc, std::forward< StackAllocator >( salloc),
             wrap( [this](typename std::decay< Fn >::type & fn_,boost::context::continuation && c) mutable {
                     // create synthesized pull_coroutine< T >
                     typename pull_coroutine< T >::control_block synthesized_cb{ this, c };
@@ -77,7 +77,7 @@ push_coroutine< T >::control_block::control_block( context::preallocated palloc,
                  std::forward< Fn >( fn) ) );
 #else
     c = boost::context::callcc(
-            std::allocator_arg, palloc, salloc,
+            std::allocator_arg, palloc, std::forward< StackAllocator >( salloc),
             [this,fn_=std::forward< Fn >( fn)](boost::context::continuation && c) mutable {
                // create synthesized pull_coroutine< T >
                typename pull_coroutine< T >::control_block synthesized_cb{ this, c };
@@ -166,7 +166,7 @@ push_coroutine< T & >::control_block::destroy( control_block * cb) noexcept {
 
 template< typename T >
 template< typename StackAllocator, typename Fn >
-push_coroutine< T & >::control_block::control_block( context::preallocated palloc, StackAllocator salloc,
+push_coroutine< T & >::control_block::control_block( context::preallocated palloc, StackAllocator && salloc,
                                                      Fn && fn) :
     c{},
     other{ nullptr },
@@ -174,7 +174,7 @@ push_coroutine< T & >::control_block::control_block( context::preallocated pallo
     except{} {
 #if defined(BOOST_NO_CXX14_GENERIC_LAMBDAS)
     c = boost::context::callcc(
-            std::allocator_arg, palloc, salloc,
+            std::allocator_arg, palloc, std::forward< StackAllocator >( salloc),
             wrap( [this](typename std::decay< Fn >::type & fn_,boost::context::continuation && c) mutable {
                     // create synthesized pull_coroutine< T & >
                     typename pull_coroutine< T & >::control_block synthesized_cb{ this, c };
@@ -201,7 +201,7 @@ push_coroutine< T & >::control_block::control_block( context::preallocated pallo
                  std::forward< Fn >( fn) ) );
 #else
     c = boost::context::callcc(
-            std::allocator_arg, palloc, salloc,
+            std::allocator_arg, palloc, std::forward< StackAllocator >( salloc),
             [this,fn_=std::forward< Fn >( fn)](boost::context::continuation && c) mutable {
                // create synthesized pull_coroutine< T & >
                typename pull_coroutine< T & >::control_block synthesized_cb{ this, c };
@@ -277,14 +277,14 @@ push_coroutine< void >::control_block::destroy( control_block * cb) noexcept {
 }
 
 template< typename StackAllocator, typename Fn >
-push_coroutine< void >::control_block::control_block( context::preallocated palloc, StackAllocator salloc, Fn && fn) :
+push_coroutine< void >::control_block::control_block( context::preallocated palloc, StackAllocator && salloc, Fn && fn) :
     c{},
     other{ nullptr },
     state{ state_t::unwind },
     except{} {
 #if defined(BOOST_NO_CXX14_GENERIC_LAMBDAS)
     c = boost::context::callcc(
-            std::allocator_arg, palloc, salloc,
+            std::allocator_arg, palloc, std::forward< StackAllocator >( salloc),
             wrap( [this](typename std::decay< Fn >::type & fn_,boost::context::continuation && c) mutable {
                     // create synthesized pull_coroutine< void >
                     typename pull_coroutine< void >::control_block synthesized_cb{ this, c };
@@ -311,7 +311,7 @@ push_coroutine< void >::control_block::control_block( context::preallocated pall
                  std::forward< Fn >( fn) ) );
 #else
     c = boost::context::callcc(
-            std::allocator_arg, palloc, salloc,
+            std::allocator_arg, palloc, std::forward< StackAllocator >( salloc),
             [this,fn_=std::forward< Fn >( fn)](boost::context::continuation && c) mutable {
                // create synthesized pull_coroutine< void >
                typename pull_coroutine< void >::control_block synthesized_cb{ this, c};
diff --git a/include/boost/coroutine2/detail/push_coroutine.hpp b/include/boost/coroutine2/detail/push_coroutine.hpp
index 533d4f9..d152be2 100644
--- a/include/boost/coroutine2/detail/push_coroutine.hpp
+++ b/include/boost/coroutine2/detail/push_coroutine.hpp
@@ -43,7 +43,7 @@ public:
     explicit push_coroutine( Fn &&);
 
     template< typename StackAllocator, typename Fn >
-    push_coroutine( StackAllocator, Fn &&);
+    push_coroutine( StackAllocator &&, Fn &&);
 
     ~push_coroutine();
 
@@ -123,7 +123,7 @@ public:
     explicit push_coroutine( Fn &&);
 
     template< typename StackAllocator, typename Fn >
-    push_coroutine( StackAllocator, Fn &&);
+    push_coroutine( StackAllocator &&, Fn &&);
 
     ~push_coroutine();
 
@@ -201,7 +201,7 @@ public:
     explicit push_coroutine( Fn &&);
 
     template< typename StackAllocator, typename Fn >
-    push_coroutine( StackAllocator, Fn &&);
+    push_coroutine( StackAllocator &&, Fn &&);
 
     ~push_coroutine();
 
diff --git a/include/boost/coroutine2/detail/push_coroutine.ipp b/include/boost/coroutine2/detail/push_coroutine.ipp
index 303961b..dd51403 100644
--- a/include/boost/coroutine2/detail/push_coroutine.ipp
+++ b/include/boost/coroutine2/detail/push_coroutine.ipp
@@ -43,8 +43,8 @@ push_coroutine< T >::push_coroutine( Fn && fn) :
 
 template< typename T >
 template< typename StackAllocator, typename Fn >
-push_coroutine< T >::push_coroutine( StackAllocator salloc, Fn && fn) :
-    cb_{ create_control_block< control_block >( salloc, std::forward< Fn >( fn) ) } {
+push_coroutine< T >::push_coroutine( StackAllocator && salloc, Fn && fn) :
+    cb_{ create_control_block< control_block >( std::forward< StackAllocator >( salloc), std::forward< Fn >( fn) ) } {
 }
 
 template< typename T >
@@ -103,8 +103,8 @@ push_coroutine< T & >::push_coroutine( Fn && fn) :
 
 template< typename T >
 template< typename StackAllocator, typename Fn >
-push_coroutine< T & >::push_coroutine( StackAllocator salloc, Fn && fn) :
-    cb_{ create_control_block< control_block >( salloc, std::forward< Fn >( fn) ) } {
+push_coroutine< T & >::push_coroutine( StackAllocator && salloc, Fn && fn) :
+    cb_{ create_control_block< control_block >( std::forward< StackAllocator >( salloc), std::forward< Fn >( fn) ) } {
 }
 
 template< typename T >
@@ -154,8 +154,8 @@ push_coroutine< void >::push_coroutine( Fn && fn) :
 }
 
 template< typename StackAllocator, typename Fn >
-push_coroutine< void >::push_coroutine( StackAllocator salloc, Fn && fn) :
-    cb_{ create_control_block< control_block >( salloc, std::forward< Fn >( fn) ) } {
+push_coroutine< void >::push_coroutine( StackAllocator && salloc, Fn && fn) :
+    cb_{ create_control_block< control_block >( std::forward< StackAllocator >( salloc), std::forward< Fn >( fn) ) } {
 }
 
 inline