From c80c540be32e7b5872f2ef1af9cabd64af1c15c0 Mon Sep 17 00:00:00 2001
From: Isaac Aronson <i@pingas.org>
Date: Wed, 6 Sep 2023 17:17:09 -0500
Subject: [PATCH] Support older GCC and clang that lack some intrinsics

---
 src/video/SDL_blit_A_avx2.c   |  2 ++
 src/video/SDL_blit_A_sse4_1.c |  3 +++
 src/video/SDL_blit_A_sse4_1.h | 32 ++++++++++++++++++++++++++++++++
 3 files changed, 37 insertions(+)

diff --git a/src/video/SDL_blit_A_avx2.c b/src/video/SDL_blit_A_avx2.c
index d2a241874d..4f8d01b313 100644
--- a/src/video/SDL_blit_A_avx2.c
+++ b/src/video/SDL_blit_A_avx2.c
@@ -4,6 +4,8 @@
 
 #ifdef SDL_AVX2_INTRINSICS
 
+#define SDL_blit_A_avx2_c
+
 #include "SDL_blit.h"
 #include "SDL_blit_A_sse4_1.h"
 
diff --git a/src/video/SDL_blit_A_sse4_1.c b/src/video/SDL_blit_A_sse4_1.c
index 2135c87001..b8bf6f4902 100644
--- a/src/video/SDL_blit_A_sse4_1.c
+++ b/src/video/SDL_blit_A_sse4_1.c
@@ -4,7 +4,10 @@
 
 #ifdef SDL_SSE4_1_INTRINSICS
 
+#define SDL_blit_A_sse4_1_c
+
 #include "SDL_blit.h"
+#include "SDL_blit_A_sse4_1.h"
 
 /**
  * Using the SSE4.1 instruction set, blit four pixels with alpha blending
diff --git a/src/video/SDL_blit_A_sse4_1.h b/src/video/SDL_blit_A_sse4_1.h
index f26196efc6..2850290fd5 100644
--- a/src/video/SDL_blit_A_sse4_1.h
+++ b/src/video/SDL_blit_A_sse4_1.h
@@ -12,4 +12,36 @@ void SDL_TARGETING("sse4.1") BlitNtoNPixelAlpha_SSE4_1(SDL_BlitInfo *info);
 
 #endif
 
+/* for compatibility with older compilers: */
+#if defined(SDL_blit_A_sse4_1_c) || defined(SDL_blit_A_avx2_c)
+/* _mm_loadu_si64 : missing in clang < 3.9, missing in gcc < 9
+ * _mm_storeu_si64: missing in clang < 8.0, missing in gcc < 9
+ * __m128i_u type (to be used to define the missing two above):
+ *		      missing in gcc < 7, missing in clang < 9
+ */
+#if defined(__clang__)
+#if (__clang_major__ < 9)
+#define MISSING__m128i_u
+#endif
+#if (__clang_major__ < 8)
+#define MISSING__mm_storeu_si64
+#endif
+#elif defined(__GNUC__)
+#if (__GNUC__ < 7)
+#define MISSING__m128i_u
+#endif
+#if (__GNUC__ < 9)
+#define MISSING__mm_storeu_si64
+#endif
+#endif
+
+#ifdef MISSING__m128i_u
+typedef long long __m128i_u __attribute__((__vector_size__(16), __may_alias__, __aligned__(1)));
+#endif
+#ifdef MISSING__mm_storeu_si64
+#define _mm_loadu_si64(_x) _mm_loadl_epi64((__m128i_u*)(_x))
+#define _mm_storeu_si64(_x,_y)  _mm_storel_epi64((__m128i_u*)(_x),(_y))
+#endif
+#endif /**/
+
 #endif //SDL_SDL_BLIT_A_SSE4_1_H