From 6ddcef1bca97423ac0a77f096554518ef71fa34c Mon Sep 17 00:00:00 2001
From: nmeyer-ur <nils.meyer@ur.de>
Date: Thu, 21 May 2020 21:21:03 +0200
Subject: [PATCH] fix build error enabling fcmla/mac in vector types for VLA

---
 Grid/simd/Grid_a64fx-2.h      |  5 +++--
 Grid/simd/Grid_vector_types.h | 10 ++++------
 2 files changed, 7 insertions(+), 8 deletions(-)

diff --git a/Grid/simd/Grid_a64fx-2.h b/Grid/simd/Grid_a64fx-2.h
index 0333299f..a0463a10 100644
--- a/Grid/simd/Grid_a64fx-2.h
+++ b/Grid/simd/Grid_a64fx-2.h
@@ -396,7 +396,7 @@ struct MultComplex{
 struct MultAddComplex{
   // Complex a*b+c
   template <typename T>
-  inline void mac(const vec<T> &a, const vec<T> b, const vec<T> c){
+  inline vec<T> operator()(vec<T> a, vec<T> b, vec<T> c){
     vec<T> out;
     svbool_t pg1 = acle<T>::pg1();
     typename acle<T>::vt a_v = svld1(pg1, a.v);
@@ -406,8 +406,9 @@ struct MultAddComplex{
     // using FCMLA
     typename acle<T>::vt r_v = svcmla_x(pg1, c_v, a_v, b_v, 0);
     r_v = svcmla_x(pg1, r_v, a_v, b_v, 90);
+    svst1(pg1, out.v, r_v);
 
-    svst1(pg1, a.v, r_v);
+    return out;
   }
 };
 
diff --git a/Grid/simd/Grid_vector_types.h b/Grid/simd/Grid_vector_types.h
index f8de3d30..e1eb330d 100644
--- a/Grid/simd/Grid_vector_types.h
+++ b/Grid/simd/Grid_vector_types.h
@@ -299,8 +299,7 @@ public:
   // FIXME -- alias this to an accelerator_inline MAC struct.
 
   // FIXME VLA build error
-  //#if defined(A64FX) || defined(A64FXFIXEDSIZE)  // VLA only: build error
-  #if defined(A64FXFIXEDSIZE)
+  #if defined(A64FX) || defined(A64FXFIXEDSIZE)
   friend accelerator_inline void mac(Grid_simd *__restrict__ y,
 				     const Grid_simd *__restrict__ a,
 				     const Grid_simd *__restrict__ x) {
@@ -791,10 +790,9 @@ accelerator_inline Grid_simd<S, V> operator*(Grid_simd<S, V> a, Grid_simd<S, V>
   return ret;
 };
 
-// ----------------A64FX MAC ---------------------
+// ---------------- A64FX MAC -------------------
 // Distinguish between complex types and others
-//#if defined(A64FX) || defined(A64FXFIXEDSIZE)  // VLA only: build error
-#if defined(A64FXFIXEDSIZE)
+#if defined(A64FX) || defined(A64FXFIXEDSIZE)
 template <class S, class V, IfComplex<S> = 0>
 accelerator_inline Grid_simd<S, V> fxmac(Grid_simd<S, V> a, Grid_simd<S, V> b, Grid_simd<S, V> c) {
   Grid_simd<S, V> ret;
@@ -810,7 +808,7 @@ accelerator_inline Grid_simd<S, V> fxmac(Grid_simd<S, V> a, Grid_simd<S, V> b, G
   return ret;
 };
 #endif
-// -------------------------------------
+// ----------------------------------------------
 
 
 // Distinguish between complex types and others