From 2f0318a2afef17f1991d1b61848b83ed3e0c9aff Mon Sep 17 00:00:00 2001
From: Erik Schnetter <schnetter@gmail.com>
Date: Tue, 18 Jun 2013 18:47:48 -0400
Subject: Implement barrier for vec_pseudo

---
 vec_pseudo.h | 27 +++++++++++++++++++++++++++
 1 file changed, 27 insertions(+)

diff --git a/vec_pseudo.h b/vec_pseudo.h
index abc1b30..0764d94 100644
--- a/vec_pseudo.h
+++ b/vec_pseudo.h
@@ -424,6 +424,33 @@ namespace vecmathlib {
       }
       return name_.c_str();
     }
+    inline void barrier()
+    {
+#if defined __GNUC__ && !defined __clang__ && !defined __ICC
+      // GCC crashes when +X is used as constraint
+#  if defined __SSE2__
+      for (int d=0; d<size; ++d) __asm__("": "+x" (v[d]));
+#  elif defined __PPC64__       // maybe also __PPC__
+      for (int d=0; d<size; ++d) __asm__("": "+f" (v[d]));
+#  elif defined __arm__
+      for (int d=0; d<size; ++d) __asm__("": "+w" (v[d]));
+#  else
+#    error "Floating point barrier undefined on this architecture"
+#  endif
+#elif defined __clang__
+      for (int d=0; d<size; ++d) __asm__("": "+X" (v[d]));
+#elif defined __ICC
+      for (int d=0; d<size; ++d) {
+        real_t tmp = v[d];
+        __asm__("": "+X" (tmp));
+        v[d] = tmp;
+      }
+#elif defined __IBMCPP__
+      for (int d=0; d<size; ++d) __asm__("": "+f" (v[d]));
+#else
+#  error "Floating point barrier undefined on this architecture"
+#endif
+    }
     
     typedef boolpseudovec<real_t, size> boolvec_t;
     typedef intpseudovec<real_t, size> intvec_t;
-- 
cgit v1.1