From 8dbf790f62aabded881b2c2bcf9e0a7c6558f514 Mon Sep 17 00:00:00 2001
From: nmeyer-ur <nils.meyer@ur.de>
Date: Fri, 12 Jun 2020 17:12:34 +0200
Subject: [PATCH] correct tbl2 for sp

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

diff --git a/Grid/simd/Grid_a64fx-2.h b/Grid/simd/Grid_a64fx-2.h
index 51a6bfd0..221562de 100644
--- a/Grid/simd/Grid_a64fx-2.h
+++ b/Grid/simd/Grid_a64fx-2.h
@@ -82,7 +82,7 @@ NAMESPACE_BEGIN(Optimization);
     };
   };
 
-  #else // not defines ARMCLANGCOMPAT
+  #else //  no ARMCLANGCOMPAT
   #define vec_imm vec
   // SIMD vector types
   template <typename T>
@@ -181,7 +181,7 @@ struct acle<float>{
   }
   static inline vec<uint32_t> tbl2(){
     //const vec<uint32_t> t = {2, 3, 0, 1, 6, 7, 4, 5, 10, 11, 8, 9, 14, 15, 12, 13};
-    const vec_imm<uint32_t> t = {4, 5, 6, 7, 0, 1, 2, 3, 12, 13, 14, 15, 8, 9, 10, 11};
+    const vec_imm<uint32_t> t = {2, 3, 0, 1, 6, 7, 4, 5, 10, 11, 8, 9, 14, 15, 12, 13};
     return t;
   }
   static inline vec<uint32_t> tbl_exch1a(){ // Exchange1
@@ -889,6 +889,7 @@ inline Integer Reduce<Integer, veci>::operator()(veci in){
 }
 
 #undef svred
+#undef vec_imm
 
 NAMESPACE_END(Optimization)
 
diff --git a/Grid/simd/Grid_vector_types.h b/Grid/simd/Grid_vector_types.h
index 527fde18..634cf470 100644
--- a/Grid/simd/Grid_vector_types.h
+++ b/Grid/simd/Grid_vector_types.h
@@ -122,7 +122,7 @@ accelerator_inline Grid_half sfw_float_to_half(float ff) {
     #if defined(A64FX) // VLA
       #pragma message("building for A64FX / SVE ACLE VLA")
       #if defined(ARMCLANGCOMPAT)
-        #pragma message("applying armclang patch")
+        #pragma message("applying data types patch")
       #endif
       #include "Grid_a64fx-2.h"
     #endif
@@ -327,8 +327,9 @@ public:
 
   // FIXME -- alias this to an accelerator_inline MAC struct.
 
-  // FIXME VLA build error
-  #if defined(A64FX) || defined(A64FXFIXEDSIZE)
+  // safety exclude fxmac from VLA (causing wrong results?)
+  //#if defined(A64FX) || defined(A64FXFIXEDSIZE)
+  #if defined(A64FXFIXEDSIZE)
   friend accelerator_inline void mac(Grid_simd *__restrict__ y,
 				     const Grid_simd *__restrict__ a,
 				     const Grid_simd *__restrict__ x) {