From dc0e2412fdcbb365a2ff8d72d41f0e60f5eb773a Mon Sep 17 00:00:00 2001
From: Yining Karl Li <betajippity@gmail.com>
Date: Thu, 16 Mar 2023 09:50:19 -0700
Subject: [PATCH] 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>
---
 BUILD.bazel                      |  1 +
 cmake/OpenEXRConfigInternal.h.in |  7 +++++++
 cmake/OpenEXRSetup.cmake         | 18 ++++++++++++++++++
 src/lib/OpenEXR/ImfSimd.h        | 14 ++++++++++++++
 4 files changed, 40 insertions(+)

diff --git a/BUILD.bazel b/BUILD.bazel
index f9283d1143..56facf391d 100644
--- a/BUILD.bazel
+++ b/BUILD.bazel
@@ -72,6 +72,7 @@ expand_template(
         "#cmakedefine OPENEXR_IMF_HAVE_GCC_INLINE_ASM_AVX 1": "/* #undef OPENEXR_IMF_HAVE_GCC_INLINE_ASM_AVX */",
         "#cmakedefine OPENEXR_IMF_HAVE_LINUX_PROCFS 1": "/* #undef OPENEXR_IMF_HAVE_LINUX_PROCFS */",
         "#cmakedefine OPENEXR_IMF_HAVE_SYSCONF_NPROCESSORS_ONLN 1": "/* #undef OPENEXR_IMF_HAVE_SYSCONF_NPROCESSORS_ONLN */",
+        "#cmakedefine OPENEXR_MISSING_ARM_VLD1 0": "/* #undef OPENEXR_MISSING_ARM_VLD1 */",
     },
     template = "cmake/OpenEXRConfigInternal.h.in",
 )
diff --git a/cmake/OpenEXRConfigInternal.h.in b/cmake/OpenEXRConfigInternal.h.in
index 30c6f1d586..2e242a6741 100644
--- a/cmake/OpenEXRConfigInternal.h.in
+++ b/cmake/OpenEXRConfigInternal.h.in
@@ -47,6 +47,13 @@
 
 #cmakedefine OPENEXR_IMF_HAVE_GCC_INLINE_ASM_AVX 1
 
+//
+// Define if we need to shim in our own implementation of vld1q_f32_x2 for
+// older compilers that are missing x2 Neon intrinsics on aarch64
+//
+
+#cmakedefine OPENEXR_MISSING_ARM_VLD1 0
+
 // clang-format on
 
 #endif // INCLUDED_OPENEXR_INTERNAL_CONFIG_H
diff --git a/cmake/OpenEXRSetup.cmake b/cmake/OpenEXRSetup.cmake
index 757f233223..458c1b8987 100644
--- a/cmake/OpenEXRSetup.cmake
+++ b/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)
+    set(OPENEXR_MISSING_ARM_VLD1 TRUE)
+  endif()
+endif()
diff --git a/src/lib/OpenEXR/ImfSimd.h b/src/lib/OpenEXR/ImfSimd.h
index 3053a5d4e4..810b1b1bc3 100644
--- a/src/lib/OpenEXR/ImfSimd.h
+++ b/src/lib/OpenEXR/ImfSimd.h
@@ -62,4 +62,18 @@ extern "C" {
 
 }
 
+#include "OpenEXRConfigInternal.h"
+#ifdef OPENEXR_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