Detect missing vst1q_f32_x2 and provide replacement if necessary

Older versions of GCC (< 9) do not provide the vst1q_f32_x2 intrinsic on
aarch64, so we must detect when vst1q_f32_x2 is not available and provide
our own implementation instead.

Signed-off-by: Yining Karl Li <betajippity@gmail.com>

cmake/OpenEXRSetup.cmake

@@ -312,3 +312,21 @@ else()
     message(STATUS "Imath interface dirs ${IMATH_HEADER_ONLY_INCLUDE_DIRS}")
   endif()
 endif()
+
+###########################################
+# Check if we need to emulate vld1q_f32_x2
+###########################################
+
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64")
+  include(CheckCSourceCompiles)
+  check_c_source_compiles("#include <arm_neon.h>
+int main() {
+  float a[] = {1.0, 1.0};
+  vld1q_f32_x2(a);
+  return 0;
+}" HAS_VLD1)
+
+  if(NOT HAS_VLD1)
+    string(APPEND CMAKE_CXX_FLAGS " -DMISSING_ARM_VLD1")
+  endif()
+endif()

src/lib/OpenEXR/ImfSimd.h

@@ -62,4 +62,17 @@ extern "C" {
 
 }
 
+#if defined (MISSING_ARM_VLD1)
+/* Workaround for missing vld1q_f32_x2 in older gcc versions.  */
+
+__extension__ extern __inline float32x4x2_t
+    __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
+    vld1q_f32_x2 (const float32_t* __a)
+{
+    float32x4x2_t ret;
+    asm ("ld1 {%S0.4s - %T0.4s}, [%1]" : "=w"(ret) : "r"(__a) :);
+    return ret;
+}
+#endif
+
 #endif