Index: test-suite/trunk/Bitcode/CMakeLists.txt
===================================================================
--- test-suite/trunk/Bitcode/CMakeLists.txt
+++ test-suite/trunk/Bitcode/CMakeLists.txt
@@ -1,6 +1,6 @@
 if(NOT TEST_SUITE_BENCHMARKING_ONLY)
   llvm_add_subdirectories(Regression)
-  if(ARCH STREQUAL "x86")
+  if(ARCH STREQUAL "x86" OR ARCH STREQUAL "AArch64")
     llvm_add_subdirectories(simd_ops)
   endif()
 endif()
Index: test-suite/trunk/Bitcode/simd_ops/CMakeLists.txt
===================================================================
--- test-suite/trunk/Bitcode/simd_ops/CMakeLists.txt
+++ test-suite/trunk/Bitcode/simd_ops/CMakeLists.txt
@@ -11,7 +11,7 @@
   string(REGEX REPLACE ".[cp]+$" "" pathbc ${sourcebc})
   string(REGEX REPLACE ".*/" "" namebc ${pathbc})
   string(REPLACE "." "" namebc ${namebc})
-  set(Source ${CMAKE_CURRENT_SOURCE_DIR}/simd_ops.cpp ${CMAKE_CURRENT_SOURCE_DIR}/simd_op_check_runtime.bc ${CMAKE_CURRENT_SOURCE_DIR}/${ARCH}_tests/${namebc}.bc ${CMAKE_CURRENT_SOURCE_DIR}/${ARCH}_scalar_tests/scalar_${namebc}.bc)
+  set(Source ${CMAKE_CURRENT_SOURCE_DIR}/simd_ops.cpp ${CMAKE_CURRENT_SOURCE_DIR}/${ARCH}_halide_runtime.bc ${CMAKE_CURRENT_SOURCE_DIR}/${ARCH}_tests/${namebc}.bc ${CMAKE_CURRENT_SOURCE_DIR}/${ARCH}_scalar_tests/scalar_${namebc}.bc)
   set(PROG simd_ops_${namebc})
   llvm_multisource()
 endforeach()
Index: test-suite/trunk/Bitcode/simd_ops/simd_ops.cpp
===================================================================
--- test-suite/trunk/Bitcode/simd_ops/simd_ops.cpp
+++ test-suite/trunk/Bitcode/simd_ops/simd_ops.cpp
@@ -16,6 +16,12 @@
 #endif
 }
 
+#if defined(__aarch64__)
+#define FACTOR 5
+#else
+#define FACTOR 1
+#endif
+
 template<typename T>
 T rand_value() {
     return (T)(rand() * 0.125) - 100;
@@ -65,7 +71,7 @@
         seed = time(NULL);
         srand (seed);
     }
-    int W = 256, H = 100;
+    int W = 256*FACTOR, H = 100;
     // Make some input buffers
     buffer_t bufs[] = {
         make_buffer<float>(W, H),
@@ -79,6 +85,7 @@
         make_buffer<int64_t>(W, H),
         make_buffer<uint64_t>(W, H)
     };
+    W/=FACTOR;
 
     int NO = 2;
     buffer_t out[] = {