Add SSE4.1 support.

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.
-#<smmintrin.h> SSE4.1
 #<nmmintrin.h> SSE4.2
 #<ammintrin.h> SSE4A
 #<wmmintrin.h> AES
@@ -31,6 +30,7 @@ simdarr = [['-mmmx', 'HAVE_MMX', 'simd_mmx', 'simd_mmx.c'],
  ['-msse2', 'HAVE_SSE2', 'simd_sse2', 'simd_sse2.c'],
  ['-msse3', 'HAVE_SSE3', 'simd_sse3', 'simd_sse3.c'],
  ['-mssse3', 'HAVE_SSSE3', 'simd_ssse3', 'simd_ssse3.c'],
+ ['-msse4.1', 'HAVE_SSE41', 'simd_sse41', 'simd_sse41.c'],
 ]
 
 foreach ia : simdarr

test cases/common/139 simd/simd_sse41.c

@@ -0,0 +1,26 @@
+#include<simdconfig.h>
+#include<simdfuncs.h>
+
+#include<smmintrin.h>
+#include<cpuid.h>
+#include<stdint.h>
+
+int sse41_available() {
+    return __builtin_cpu_supports("sse4.1");
+}
+
+void increment_sse41(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);
+    result = _mm_ceil_pd(result); /* A no-op, only here to use a SSE4.1 intrinsic. */
+    _mm_store_pd(darr, result);
+    result = _mm_add_pd(val2, one);
+    _mm_store_pd(&darr[2], result);
+    arr[0] = (float)darr[1];
+    arr[1] = (float)darr[0];
+    arr[2] = (float)darr[3];
+    arr[3] = (float)darr[2];
+}

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_SSE41
+    if(fptr == NULL && sse41_available()) {
+        fptr = increment_sse41;
+        type = "SSE41";
+    }
+#endif
 #if HAVE_SSSE3
     if(fptr == NULL && ssse3_available()) {
         fptr = increment_ssse3;

test cases/common/139 simd/simdfuncs.h

@@ -34,6 +34,11 @@ int ssse3_available();
 void increment_ssse3(float arr[4]);
 #endif
 
+#if HAVE_SSE41
+int sse41_available();
+void increment_sse41(float arr[4]);
+#endif
+
 #if HAVE_AVX
 int avx_available();
 void increment_avx(float arr[4]);