Fix simd instruction alignment on 32bit kernels.

Resolves segfaults on i686 kernel. Also execute all available simd
instructions, not just the best one.

Signed-off-by: Dimitri John Ledkov <xnox@ubuntu.com>
Signed-off-by: Michael Hudson <michael.hudson@ubuntu.com>

Bug-Debian: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=894774

test cases/common/152 simd/simd_sse2.c

@@ -1,6 +1,7 @@
 #include<simdconfig.h>
 #include<simdfuncs.h>
 #include<emmintrin.h>
+#include<stdalign.h>
 
 #ifdef _MSC_VER
 int sse2_available() {
@@ -21,7 +22,7 @@ int sse2_available() {
 #endif
 
 void increment_sse2(float arr[4]) {
-    double darr[4];
+    alignas(16) 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_pd(1.0, 1.0);

test cases/common/152 simd/simd_sse3.c

@@ -1,5 +1,6 @@
 #include<simdconfig.h>
 #include<simdfuncs.h>
+#include<stdalign.h>
 
 #ifdef _MSC_VER
 #include<intrin.h>
@@ -22,7 +23,7 @@ int sse3_available() {
 #endif
 
 void increment_sse3(float arr[4]) {
-    double darr[4];
+    alignas(16) 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_pd(1.0, 1.0);

test cases/common/152 simd/simd_sse41.c

@@ -1,6 +1,7 @@
 #include<simdconfig.h>
 #include<simdfuncs.h>
 
+#include<stdalign.h>
 #include<stdint.h>
 
 #ifdef _MSC_VER
@@ -24,7 +25,7 @@ int sse41_available() {
 #endif
 
 void increment_sse41(float arr[4]) {
-    double darr[4];
+    alignas(16) 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_pd(1.0, 1.0);

test cases/common/152 simd/simd_sse42.c

@@ -1,6 +1,7 @@
 #include<simdconfig.h>
 #include<simdfuncs.h>
 #include<stdint.h>
+#include<stdalign.h>
 
 #ifdef _MSC_VER
 #include<intrin.h>
@@ -27,7 +28,7 @@ int sse42_available() {
 #endif
 
 void increment_sse42(float arr[4]) {
-    double darr[4];
+    alignas(16) 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_pd(1.0, 1.0);

test cases/common/152 simd/simd_ssse3.c

@@ -1,5 +1,6 @@
 #include<simdconfig.h>
 #include<simdfuncs.h>
+#include<stdalign.h>
 
 #include<emmintrin.h>
 #include<tmmintrin.h>
@@ -30,7 +31,7 @@ int ssse3_available() {
 #endif
 
 void increment_ssse3(float arr[4]) {
-    double darr[4];
+    alignas(16) 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_pd(1.0, 1.0);

test cases/common/152 simd/simdchecker.c

@@ -1,5 +1,7 @@
 #include<simdfuncs.h>
+#include<stdalign.h>
 #include<stdio.h>
+#include<string.h>
 
 /*
  * A function that checks at runtime which simd accelerations are
@@ -8,86 +10,89 @@
  */
 
 int main(int argc, char **argv) {
-    float four[4] = {2.0, 3.0, 4.0, 5.0};
+    static const float four_initial[4] = {2.0, 3.0, 4.0, 5.0};
+    alignas(16) float four[4];
     const float expected[4] = {3.0, 4.0, 5.0, 6.0};
     void (*fptr)(float[4]) = NULL;
     const char *type;
-    int i;
+    int i, r=0;
 
 /* Add here. The first matched one is used so put "better" instruction
  * sets at the top.
  */
 #if HAVE_NEON
-    if(fptr == NULL && neon_available()) {
+    if(neon_available()) {
         fptr = increment_neon;
         type = "NEON";
+    #include<simdtest.h>
     }
 #endif
 #if HAVE_AVX2
-    if(fptr == NULL && avx2_available()) {
+    if(avx2_available()) {
         fptr = increment_avx2;
         type = "AVX2";
+    #include<simdtest.h>
     }
 #endif
 #if HAVE_AVX
-    if(fptr == NULL && avx_available()) {
+    if(avx_available()) {
         fptr = increment_avx;
         type = "AVX";
+    #include<simdtest.h>
     }
 #endif
 #if HAVE_SSE42
-    if(fptr == NULL && sse42_available()) {
+    if(sse42_available()) {
         fptr = increment_sse42;
         type = "SSE42";
+    #include<simdtest.h>
     }
 #endif
 #if HAVE_SSE41
-    if(fptr == NULL && sse41_available()) {
+    if(sse41_available()) {
         fptr = increment_sse41;
         type = "SSE41";
+    #include<simdtest.h>
     }
 #endif
 #if HAVE_SSSE3
-    if(fptr == NULL && ssse3_available()) {
+    if(ssse3_available()) {
         fptr = increment_ssse3;
         type = "SSSE3";
+    #include<simdtest.h>
     }
 #endif
 #if HAVE_SSE3
-    if(fptr == NULL && sse3_available()) {
+    if(sse3_available()) {
         fptr = increment_sse3;
         type = "SSE3";
+    #include<simdtest.h>
     }
 #endif
 #if HAVE_SSE2
-    if(fptr == NULL && sse2_available()) {
+    if(sse2_available()) {
         fptr = increment_sse2;
         type = "SSE2";
+    #include<simdtest.h>
     }
 #endif
 #if HAVE_SSE
-    if(fptr == NULL && sse_available()) {
+    if(sse_available()) {
         fptr = increment_sse;
         type = "SSE";
+    #include<simdtest.h>
     }
 #endif
 #if HAVE_MMX
-    if(fptr == NULL && mmx_available()) {
+    if(mmx_available()) {
         fptr = increment_mmx;
         type = "MMX";
+    #include<simdtest.h>
     }
 #endif
-    if(fptr == NULL) {
-        fptr = increment_fallback;
-        type = "fallback";
-    }
-    printf("Using %s.\n", type);
-    fptr(four);
-    for(i=0; i<4; i++) {
-        if(four[i] != expected[i]) {
-            printf("Increment function failed, got %f expected %f.\n", four[i], expected[i]);
-            return 1;
-        }
-    }
-    return 0;
+    fptr = increment_fallback;
+    type = "fallback";
+    #include<simdtest.h>
+
+    return r;
 }

test cases/common/152 simd/simdtest.h

@@ -0,0 +1,9 @@
+memcpy(four, four_initial, sizeof(four_initial));
+printf("Using %s.\n", type);
+fptr(four);
+for(i=0; i<4; i++) {
+    if(four[i] != expected[i]) {
+        printf("Increment function failed, got %f expected %f.\n", four[i], expected[i]);
+        r=1;
+    }
+}