From 0062595e13ab6c7a14150903696d94428530d608 Mon Sep 17 00:00:00 2001
From: Jussi Pakkanen <jpakkane@gmail.com>
Date: Mon, 13 Feb 2017 22:14:23 +0200
Subject: [PATCH] Added SSE3 support.

---
 test cases/common/139 simd/meson.build   |  4 ++--
 test cases/common/139 simd/simd_sse3.c   | 26 ++++++++++++++++++++++++
 test cases/common/139 simd/simdchecker.c |  6 ++++++
 test cases/common/139 simd/simdfuncs.h   |  5 +++++
 4 files changed, 39 insertions(+), 2 deletions(-)
 create mode 100644 test cases/common/139 simd/simd_sse3.c

diff --git a/test cases/common/139 simd/meson.build b/test cases/common/139 simd/meson.build
index d717dba7e..96852576d 100644
--- a/test cases/common/139 simd/meson.build	
+++ b/test cases/common/139 simd/meson.build	
@@ -17,7 +17,6 @@ cdata = configuration_data()
 # and then have a target that uses the result in links_with.
 
 # The following headers need to be added. Also Thumb and Altivec.
-#<pmmintrin.h> SSE3
 #<tmmintrin.h> SSSE3
 #<smmintrin.h> SSE4.1
 #<nmmintrin.h> SSE4.2
@@ -28,9 +27,10 @@ cdata = configuration_data()
 
 simdlibs = []
 
-simdarr = [['-mmx', 'HAVE_MMX', 'simd_mmx', 'simd_mmx.c'],
+simdarr = [['-mmmx', 'HAVE_MMX', 'simd_mmx', 'simd_mmx.c'],
  ['-msse', 'HAVE_SSE', 'simd_sse', 'simd_sse.c'],
  ['-msse2', 'HAVE_SSE2', 'simd_sse2', 'simd_sse2.c'],
+ ['-msse3', 'HAVE_SSE3', 'simd_sse3', 'simd_sse3.c'],
 ]
 
 foreach ia : simdarr
diff --git a/test cases/common/139 simd/simd_sse3.c b/test cases/common/139 simd/simd_sse3.c
new file mode 100644
index 000000000..db0eef62d
--- /dev/null
+++ b/test cases/common/139 simd/simd_sse3.c	
@@ -0,0 +1,26 @@
+#include<simdconfig.h>
+#include<simdfuncs.h>
+
+#include<pmmintrin.h>
+#include<cpuid.h>
+#include<stdint.h>
+
+int sse3_available() {
+    return __builtin_cpu_supports("sse3");
+}
+
+void increment_sse3(float arr[4]) {
+    double darr[4];
+    __m128d val1 = _mm_set_pd(arr[0], arr[1]);
+    __m128d val2 = _mm_set_pd(arr[2], arr[3]);
+    __m128d one = _mm_set_pd1(1.0);
+    __m128d result = _mm_add_pd(val1, one);
+    _mm_store_pd(darr, result);
+    result = _mm_add_pd(val2, one);
+    _mm_store_pd(&darr[2], result);
+    result = _mm_hadd_pd(val1, val2); /* This does nothing. Only here so we use an SSE3 instruction. */
+    arr[0] = (float)darr[1];
+    arr[1] = (float)darr[0];
+    arr[2] = (float)darr[3];
+    arr[3] = (float)darr[2];
+}
diff --git a/test cases/common/139 simd/simdchecker.c b/test cases/common/139 simd/simdchecker.c
index ce28f054d..5e5fe4305 100644
--- a/test cases/common/139 simd/simdchecker.c	
+++ b/test cases/common/139 simd/simdchecker.c	
@@ -17,6 +17,12 @@ int main(int argc, char **argv) {
 /* Add here. The first matched one is used so put "better" instruction
  * sets at the top.
  */
+#if HAVE_SSE3
+    if(fptr == NULL && sse3_available()) {
+        fptr = increment_sse3;
+        type = "SSE3";
+    }
+#endif
 #if HAVE_SSE2
     if(fptr == NULL && sse2_available()) {
         fptr = increment_sse2;
diff --git a/test cases/common/139 simd/simdfuncs.h b/test cases/common/139 simd/simdfuncs.h
index 62c6bc7f7..dc6da147f 100644
--- a/test cases/common/139 simd/simdfuncs.h	
+++ b/test cases/common/139 simd/simdfuncs.h	
@@ -24,6 +24,11 @@ int sse2_available();
 void increment_sse2(float arr[4]);
 #endif
 
+#if HAVE_SSE3
+int sse3_available();
+void increment_sse3(float arr[4]);
+#endif
+
 #if HAVE_AVX
 int avx_available();
 void increment_avx(float arr[4]);