guard aarch64 assembly with __aarch64__ (#3905)

The symbol HAVE_NEON_AARCH64 is defined by the build system and not
a full compile time check. When building multiarch binaries (apple
universal) then also check for the compiler defined symbol __aarch64__
to ensure we are in the arm64 compiler block

Co-authored-by: Armin Novak <[email protected]>

Author: akallabeth

codec/common/arm64/arm_aarch64_common.h

@@ -35,6 +35,8 @@
 #ifndef CODEC_COMMON_ARM64_ARM_AARCH64_COMMON_H_
 #define CODEC_COMMON_ARM64_ARM_AARCH64_COMMON_H_
 
+#if defined(__aarch64__)
+
 /*
  ; Support macros for
  ;   - Armv8.3-A Pointer Authentication and
@@ -82,4 +84,6 @@
   .popsection;
 #endif
 
+#endif
+
 #endif /* CODEC_COMMON_ARM64_ARM_AARCH64_COMMON_H_ */

codec/common/arm64/copy_mb_aarch64_neon.S

@@ -30,7 +30,7 @@
  *
  */
 
-#ifdef HAVE_NEON_AARCH64
+#if defined(HAVE_NEON_AARCH64) && defined(__aarch64__)
 #include "arm_arch64_common_macro.S"
 
 .macro LOAD_ALIGNED_DATA_WITH_STRIDE arg0, arg1, arg2, arg3, arg4, arg5

codec/common/arm64/deblocking_aarch64_neon.S

@@ -30,7 +30,7 @@
  *
  */
 
-#ifdef HAVE_NEON_AARCH64
+#if defined(HAVE_NEON_AARCH64) && defined(__aarch64__)
 
 #include "arm_arch64_common_macro.S"
 

codec/common/arm64/expand_picture_aarch64_neon.S

@@ -30,7 +30,7 @@
  *
  */
 
-#ifdef HAVE_NEON_AARCH64
+#if defined(HAVE_NEON_AARCH64) && defined(__aarch64__)
 #include "arm_arch64_common_macro.S"
 //void ExpandPictureLuma_AArch64_neon (uint8_t* pDst, const int32_t kiStride, const int32_t kiPicW, const int32_t kiPicH);
 WELS_ASM_AARCH64_FUNC_BEGIN ExpandPictureLuma_AArch64_neon

codec/common/arm64/intra_pred_common_aarch64_neon.S

@@ -30,7 +30,7 @@
  *
  */
 
-#ifdef HAVE_NEON_AARCH64
+#if defined(HAVE_NEON_AARCH64) && defined(__aarch64__)
 #include "arm_arch64_common_macro.S"
 
 //for Luma 16x16

codec/common/arm64/mc_aarch64_neon.S

@@ -30,7 +30,7 @@
  *
  */
 
-#ifdef HAVE_NEON_AARCH64
+#if defined(HAVE_NEON_AARCH64) && defined(__aarch64__)
 #include "arm_arch64_common_macro.S"
 .align 4
 filter_para: .short 0, 1, -5, 20, 0, 0, 0, 0

codec/common/inc/copy_mb.h

@@ -67,7 +67,7 @@ void WelsCopy16x8NotAligned_neon (uint8_t* pDst, int32_t iStrideD, uint8_t* pSrc
 void WelsCopy8x16_neon (uint8_t* pDst, int32_t iStrideD, uint8_t* pSrc, int32_t iStrideS);
 #endif
 
-#if defined (HAVE_NEON_AARCH64)
+#if defined (HAVE_NEON_AARCH64) && defined(__aarch64__)
 void WelsCopy8x8_AArch64_neon (uint8_t* pDst, int32_t iStrideD, uint8_t* pSrc, int32_t iStrideS);
 void WelsCopy16x16_AArch64_neon (uint8_t* pDst, int32_t iStrideD, uint8_t* pSrc, int32_t iStrideS);
 void WelsCopy16x16NotAligned_AArch64_neon (uint8_t* pDst, int32_t iStrideD, uint8_t* pSrc, int32_t iStrideS);

codec/common/inc/deblocking_common.h

@@ -62,7 +62,7 @@ void DeblockChromaEq4H_neon (uint8_t* pPixCb, uint8_t* pPixCr, int32_t iStride,
 void WelsNonZeroCount_neon (int8_t* pNonZeroCount);
 #endif
 
-#if defined(HAVE_NEON_AARCH64)
+#if defined(HAVE_NEON_AARCH64) && defined(__aarch64__)
 void DeblockLumaLt4V_AArch64_neon (uint8_t* pPixY, int32_t iStride, int32_t iAlpha, int32_t iBeta, int8_t* pTc);
 void DeblockLumaEq4V_AArch64_neon (uint8_t* pPixY, int32_t iStride, int32_t iAlpha, int32_t iBeta);
 void DeblockLumaLt4H_AArch64_neon (uint8_t* pPixY, int32_t iStride, int32_t iAlpha, int32_t iBeta, int8_t* pTc);

codec/common/inc/expand_pic.h

@@ -68,7 +68,7 @@ void ExpandPictureChromaUnalign_sse2 (uint8_t* pDst,
 void ExpandPictureLuma_neon (uint8_t* pDst, const int32_t kiStride, const int32_t kiPicW, const int32_t kiPicH);
 void ExpandPictureChroma_neon (uint8_t* pDst, const int32_t kiStride, const int32_t kiPicW, const int32_t kiPicH);
 #endif
-#if defined(HAVE_NEON_AARCH64)
+#if defined(HAVE_NEON_AARCH64) && defined(__aarch64__)
 void ExpandPictureLuma_AArch64_neon (uint8_t* pDst, const int32_t kiStride, const int32_t kiPicW, const int32_t kiPicH);
 void ExpandPictureChroma_AArch64_neon (uint8_t* pDst, const int32_t kiStride, const int32_t kiPicW,
                                        const int32_t kiPicH);

codec/common/inc/intra_pred_common.h

@@ -63,7 +63,7 @@ void WelsI16x16LumaPredV_neon (uint8_t* pPred, uint8_t* pRef, const int32_t kiSt
 void WelsI16x16LumaPredH_neon (uint8_t* pPred, uint8_t* pRef, const int32_t kiStride);
 #endif//HAVE_NEON
 
-#if defined(HAVE_NEON_AARCH64)
+#if defined(HAVE_NEON_AARCH64) && defined(__aarch64__)
 void WelsI16x16LumaPredV_AArch64_neon (uint8_t* pPred, uint8_t* pRef, const int32_t kiStride);
 void WelsI16x16LumaPredH_AArch64_neon (uint8_t* pPred, uint8_t* pRef, const int32_t kiStride);
 #endif//HAVE_NEON_AARCH64