diff --git a/frame/include/blis.h b/frame/include/blis.h
index d87018d009..6292f47452 100644
--- a/frame/include/blis.h
+++ b/frame/include/blis.h
@@ -80,21 +80,6 @@ extern "C" {
 #include "bli_pragma_macro_defs.h"
 
 
-// -- Threading definitions --
-
-#include "bli_thread.h"
-#include "bli_thread_range.h"
-#include "bli_thread_range_slab_rr.h"
-#include "bli_thread_range_tlb.h"
-
-#include "bli_pthread.h"
-
-
-// -- Constant definitions --
-
-#include "bli_extern_defs.h"
-
-
 // -- BLIS architecture/kernel definitions --
 
 #include "bli_pre_ker_params.h"
@@ -116,6 +101,21 @@ extern "C" {
 #include "bli_kernel_macro_defs.h"
 
 
+// -- Threading definitions --
+
+#include "bli_thread.h"
+#include "bli_thread_range.h"
+#include "bli_thread_range_slab_rr.h"
+#include "bli_thread_range_tlb.h"
+
+#include "bli_pthread.h"
+
+
+// -- Constant definitions --
+
+#include "bli_extern_defs.h"
+
+
 // -- Base operation prototypes --
 
 #include "bli_init.h"
diff --git a/frame/thread/bli_thrcomm.c b/frame/thread/bli_thrcomm.c
index e9f9d9dc70..79618f1a83 100644
--- a/frame/thread/bli_thrcomm.c
+++ b/frame/thread/bli_thrcomm.c
@@ -206,6 +206,8 @@ void* bli_thrcomm_bcast
 	return object;
 }
 
+#ifndef BLIS_TREE_BARRIER
+
 // Use __sync_* builtins (assumed available) if __atomic_* ones are not present.
 #ifndef __ATOMIC_RELAXED
 
@@ -214,14 +216,10 @@ void* bli_thrcomm_bcast
 #define __ATOMIC_RELEASE
 #define __ATOMIC_ACQ_REL
 
-#define __atomic_load_n(ptr, constraint) \
-    __sync_fetch_and_add(ptr, 0)
-#define __atomic_add_fetch(ptr, value, constraint) \
-    __sync_add_and_fetch(ptr, value)
-#define __atomic_fetch_add(ptr, value, constraint) \
-    __sync_fetch_and_add(ptr, value)
-#define __atomic_fetch_xor(ptr, value, constraint) \
-    __sync_fetch_and_xor(ptr, value)
+#define __atomic_load_n(    ptr,        constraint ) __sync_fetch_and_add( ptr, 0     )
+#define __atomic_add_fetch( ptr, value, constraint ) __sync_add_and_fetch( ptr, value )
+#define __atomic_fetch_add( ptr, value, constraint ) __sync_fetch_and_add( ptr, value )
+#define __atomic_fetch_xor( ptr, value, constraint ) __sync_fetch_and_xor( ptr, value )
 
 #endif
 
@@ -269,3 +267,5 @@ void bli_thrcomm_barrier_atomic( dim_t t_id, thrcomm_t* comm )
 	}
 }
 
+#endif
+
diff --git a/frame/thread/bli_thrcomm.h b/frame/thread/bli_thrcomm.h
index 04cb23a38a..436b057116 100644
--- a/frame/thread/bli_thrcomm.h
+++ b/frame/thread/bli_thrcomm.h
@@ -51,17 +51,17 @@ struct barrier_s
 	// the fields above and fields below.
 	char   padding1[ BLIS_CACHE_LINE_SIZE ];
 
-	int               count;
+	dim_t             count;
 
 	// We insert a cache line of padding here to eliminate false sharing between
 	// the fields above and fields below.
 	char   padding2[ BLIS_CACHE_LINE_SIZE ];
 
-	volatile int      signal;
+	gint_t            signal;
 
 	// We insert a cache line of padding here to eliminate false sharing between
 	// this struct and the next one.
-	char   padding2[ BLIS_CACHE_LINE_SIZE ];
+	char   padding3[ BLIS_CACHE_LINE_SIZE ];
 };
 typedef struct barrier_s barrier_t;
 #endif
diff --git a/frame/thread/bli_thrcomm_openmp.c b/frame/thread/bli_thrcomm_openmp.c
index a42dabe180..487832cd17 100644
--- a/frame/thread/bli_thrcomm_openmp.c
+++ b/frame/thread/bli_thrcomm_openmp.c
@@ -114,6 +114,10 @@ void bli_thrcomm_cleanup_openmp( thrcomm_t* comm )
 
 void bli_thrcomm_barrier_openmp( dim_t t_id, thrcomm_t* comm )
 {
+	// Return early if the comm is NULL or if there is only one
+	// thread participating.
+	if ( comm == NULL || comm->n_threads == 1 ) return;
+
 	bli_thrcomm_tree_barrier( comm->barriers[t_id] );
 }
 
@@ -176,27 +180,42 @@ void bli_thrcomm_tree_barrier_free( barrier_t* barrier )
 	return;
 }
 
+// Use __sync_* builtins (assumed available) if __atomic_* ones are not present.
+#ifndef __ATOMIC_RELAXED
+
+#define __ATOMIC_RELAXED
+#define __ATOMIC_ACQUIRE
+#define __ATOMIC_RELEASE
+#define __ATOMIC_ACQ_REL
+
+//#define __atomic_add_fetch( ptr, value, constraint ) __sync_add_and_fetch( ptr, value )
+//#define __atomic_fetch_add( ptr, value, constraint ) __sync_fetch_and_add( ptr, value )
+
+#define __atomic_load_n(    ptr,        constraint ) __sync_fetch_and_add( ptr, 0     )
+#define __atomic_sub_fetch( ptr, value, constraint ) __sync_sub_and_fetch( ptr, value )
+#define __atomic_fetch_xor( ptr, value, constraint ) __sync_fetch_and_xor( ptr, value )
+
+#endif
+
 void bli_thrcomm_tree_barrier( barrier_t* barack )
 {
-	int my_signal = barack->signal;
-	int my_count;
+	gint_t my_signal = __atomic_load_n( &barack->signal, __ATOMIC_RELAXED );
 
-	_Pragma( "omp atomic capture" )
-		my_count = barack->count--;
+	dim_t my_count =
+	__atomic_sub_fetch( &barack->count, 1, __ATOMIC_ACQ_REL );
 
-	if ( my_count == 1 )
+	if ( my_count == 0 )
 	{
 		if ( barack->dad != NULL )
 		{
 			bli_thrcomm_tree_barrier( barack->dad );
 		}
 		barack->count = barack->arity;
-		barack->signal = !barack->signal;
+		__atomic_fetch_xor( &barack->signal, 1, __ATOMIC_RELEASE );
 	}
 	else
 	{
-		volatile int* listener = &barack->signal;
-		while ( *listener == my_signal ) {}
+		while ( __atomic_load_n( &barack->signal, __ATOMIC_ACQUIRE ) == my_signal ) {}
 	}
 }