Correctly align malloc’d vector arrays

Fixes segfault with -march=native

2 files changed, 17 insertions, 4 deletions

diff --git a/flam3.c b/flam3.c
index 4564b3f..c8f43ce 100644
--- a/flam3.c
+++ b/flam3.c
@@ -35,6 +35,7 @@
 #include <unistd.h>
 #endif
 #include <errno.h>
+#include <assert.h>
 
 #include <pthread.h>
 
@@ -3259,7 +3260,8 @@ int flam3_estimate_bounding_box(flam3_genome *cp, double eps, int nsamples,
 
    if (nsamples <= 0) nsamples = 10000;
 
-   points = (double4 *) malloc(sizeof(double4) * nsamples);
+   int ret = posix_memalign ((void **) &points, sizeof (*points), sizeof(*points) * nsamples);
+   assert (ret == 0 && points != NULL);
    const double4 start = (double4) { rand_d11(rc), rand_d11(rc), 0.0, 0.0 };
 
    if (prepare_precalc_flags(cp))
diff --git a/rect.c b/rect.c
index 7e45e16..e566647 100644
--- a/rect.c
+++ b/rect.c
@@ -17,6 +17,7 @@
 */
 
 #include <assert.h>
+#include <stdlib.h>
 
 #include "private.h"
 #include "filters.h"
@@ -380,14 +381,24 @@ int render_rectangle(flam3_frame *spec, void *out,
 
    nbuckets = (long)fic.width * (long)fic.height;
 
-   double4 * const buckets = malloc (nbuckets * sizeof (*buckets));
+   double4 *buckets;
+   int ret = posix_memalign ((void **) &buckets, sizeof (*buckets),
+                             nbuckets * sizeof (*buckets));
+   assert (ret == 0);
    assert (buckets != NULL);
-   double4 * const accumulate = malloc (nbuckets * sizeof (*accumulate));
+   double4 *accumulate;
+   ret = posix_memalign ((void **) &accumulate, sizeof (*accumulate),
+                         nbuckets * sizeof (*accumulate));
+   assert (ret == 0);
    assert (accumulate != NULL);
    double4 ** const iter_storage = malloc (spec->nthreads * sizeof (*iter_storage));
    assert (iter_storage != NULL);
    for (size_t i = 0; i < spec->nthreads; i++) {
-      iter_storage[i] = malloc (spec->sub_batch_size * sizeof (*iter_storage[i]));
+      ret = posix_memalign ((void **) &iter_storage[i],
+	                        sizeof (*iter_storage[i]),
+							spec->sub_batch_size * sizeof (*iter_storage[i]));
+	  assert (ret == 0);
+	  assert (iter_storage[i] != NULL);
    }
 
    if (verbose) {