Improve optimization barriers

3 files changed, 29 insertions, 29 deletions

diff --git a/floatprops.h b/floatprops.h
index c7a3b7f..e284c5f 100644
--- a/floatprops.h
+++ b/floatprops.h
@@ -283,6 +283,30 @@ V &set_elt(V &v, const int n, const E e) {
   return v;
 }
 
+template <typename real_t> real_t barrier(real_t x) {
+#if defined __GNUC__ && !defined __clang__ && !defined __ICC
+// GCC crashes when +X is used as constraint
+#if defined __SSE2__
+  __asm__("" : "+x"(x));
+#elif defined __PPC64__ // maybe also __PPC__
+  __asm__("" : "+f"(x));
+#elif defined __arm__
+  __asm__("" : "+w"(x));
+#else
+#error "Floating point barrier undefined on this architecture"
+#endif
+#elif defined __clang__
+  __asm__("" : "+x"(x));
+#elif defined __ICC
+  __asm__("" : "+x"(x));
+#elif defined __IBMCPP__
+  __asm__("" : "+f"(x));
+#else
+#error "Floating point barrier undefined on this architecture"
+#endif
+  return x;
+}
+
 } // namespace vecmathlib
 
 #endif // #ifndef FLOATPROPS_H
diff --git a/vec_builtin.h b/vec_builtin.h
index 2f1ff90..e810ddd 100644
--- a/vec_builtin.h
+++ b/vec_builtin.h
@@ -296,7 +296,10 @@ template <typename T, int N> struct realbuiltinvec : floatprops<T> {
     return name_.c_str();
   }
 #endif
-  void barrier() { volatile vector_t x __attribute__((__unused__)) = v; }
+  void barrier() {
+    volatile vector_t x = v;
+    v = x;
+  }
 
   typedef boolbuiltinvec<real_t, size> boolvec_t;
   typedef intbuiltinvec<real_t, size> intvec_t;
diff --git a/vec_pseudo.h b/vec_pseudo.h
index c4cbbc1..fa2895c 100644
--- a/vec_pseudo.h
+++ b/vec_pseudo.h
@@ -472,35 +472,8 @@ template <typename T, int N> struct realpseudovec : floatprops<T> {
   }
 #endif
   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
+      v[d] = vecmathlib::barrier(v[d]);
   }
 
   typedef boolpseudovec<real_t, size> boolvec_t;