diff --git a/src/tests/cpu-test.c b/src/tests/cpu-test.c
index e38e63296..82d827789 100644
--- a/src/tests/cpu-test.c
+++ b/src/tests/cpu-test.c
@@ -65,48 +65,67 @@
 
 /* Common defines for svolume tests */
 #define CHANNELS 2
-#define SAMPLES 1022
+#define SAMPLES 1028
 #define TIMES 1000
 #define TIMES2 100
 #define PADDING 16
 
-static void run_volume_test(pa_do_volume_func_t func, pa_do_volume_func_t orig_func) {
-    int16_t samples[SAMPLES];
-    int16_t samples_ref[SAMPLES];
-    int16_t samples_orig[SAMPLES];
+static void run_volume_test(pa_do_volume_func_t func, pa_do_volume_func_t orig_func, int align, pa_bool_t correct,
+        pa_bool_t perf) {
+    PA_DECLARE_ALIGNED(8, int16_t, s[SAMPLES]) = { 0 };
+    PA_DECLARE_ALIGNED(8, int16_t, s_ref[SAMPLES]) = { 0 };
+    PA_DECLARE_ALIGNED(8, int16_t, s_orig[SAMPLES]) = { 0 };
     int32_t volumes[CHANNELS + PADDING];
-    int i, padding;
+    int16_t *samples, *samples_ref, *samples_orig;
+    int i, padding, nsamples, size;
 
-    pa_random(samples, sizeof(samples));
-    memcpy(samples_ref, samples, sizeof(samples));
-    memcpy(samples_orig, samples, sizeof(samples));
+    /* Force sample alignment as requested */
+    samples = s + (8 - align);
+    samples_ref = s_ref + (8 - align);
+    samples_orig = s_orig + (8 - align);
+    nsamples = SAMPLES - (8 - align);
+    nsamples += nsamples % CHANNELS;
+    size = nsamples * sizeof(*samples);
+
+    pa_random(samples, size);
+    memcpy(samples_ref, samples, size);
+    memcpy(samples_orig, samples, size);
 
     for (i = 0; i < CHANNELS; i++)
         volumes[i] = PA_CLAMP_VOLUME((pa_volume_t)(rand() >> 15));
     for (padding = 0; padding < PADDING; padding++, i++)
         volumes[i] = volumes[padding];
 
-    orig_func(samples_ref, volumes, CHANNELS, sizeof(samples));
-    func(samples, volumes, CHANNELS, sizeof(samples));
-    for (i = 0; i < SAMPLES; i++) {
-        if (samples[i] != samples_ref[i]) {
-            printf("%d: %04x != %04x (%04x * %08x)\n", i, samples[i], samples_ref[i],
-                  samples_orig[i], volumes[i % CHANNELS]);
-            fail();
+    if (correct) {
+        pa_log_debug("Testing svolume correctness with %d byte alignment", align);
+
+        orig_func(samples_ref, volumes, CHANNELS, size);
+        func(samples, volumes, CHANNELS, size);
+
+        for (i = 0; i < nsamples; i++) {
+            if (samples[i] != samples_ref[i]) {
+                printf("%d: %04x != %04x (%04x * %08x)\n", i, samples[i], samples_ref[i],
+                        samples_orig[i], volumes[i % CHANNELS]);
+                fail();
+            }
         }
     }
 
-    PA_CPU_TEST_RUN_START("func", TIMES, TIMES2) {
-        memcpy(samples, samples_orig, sizeof(samples));
-        func(samples, volumes, CHANNELS, sizeof(samples));
-    } PA_CPU_TEST_RUN_STOP
+    if (perf) {
+        pa_log_debug("Testing svolume performance with %d byte alignment", align);
 
-    PA_CPU_TEST_RUN_START("orig", TIMES, TIMES2) {
-        memcpy(samples_ref, samples_orig, sizeof(samples));
-        orig_func(samples_ref, volumes, CHANNELS, sizeof(samples));
-    } PA_CPU_TEST_RUN_STOP
+        PA_CPU_TEST_RUN_START("func", TIMES, TIMES2) {
+            memcpy(samples, samples_orig, size);
+            func(samples, volumes, CHANNELS, size);
+        } PA_CPU_TEST_RUN_STOP
 
-    fail_unless(memcmp(samples_ref, samples, sizeof(samples)) == 0);
+        PA_CPU_TEST_RUN_START("orig", TIMES, TIMES2) {
+            memcpy(samples_ref, samples_orig, size);
+            orig_func(samples_ref, volumes, CHANNELS, size);
+        } PA_CPU_TEST_RUN_STOP
+
+        fail_unless(memcmp(samples_ref, samples, size) == 0);
+    }
 }
 
 #if defined (__i386__) || defined (__amd64__)
@@ -126,7 +145,14 @@ START_TEST (svolume_mmx_test) {
     mmx_func = pa_get_volume_func(PA_SAMPLE_S16NE);
 
     pa_log_debug("Checking MMX svolume");
-    run_volume_test(mmx_func, orig_func);
+    run_volume_test(mmx_func, orig_func, 0, TRUE, FALSE);
+    run_volume_test(mmx_func, orig_func, 1, TRUE, FALSE);
+    run_volume_test(mmx_func, orig_func, 2, TRUE, FALSE);
+    run_volume_test(mmx_func, orig_func, 3, TRUE, FALSE);
+    run_volume_test(mmx_func, orig_func, 4, TRUE, FALSE);
+    run_volume_test(mmx_func, orig_func, 5, TRUE, FALSE);
+    run_volume_test(mmx_func, orig_func, 6, TRUE, FALSE);
+    run_volume_test(mmx_func, orig_func, 7, TRUE, TRUE);
 }
 END_TEST
 
@@ -146,7 +172,14 @@ START_TEST (svolume_sse_test) {
     sse_func = pa_get_volume_func(PA_SAMPLE_S16NE);
 
     pa_log_debug("Checking SSE2 svolume");
-    run_volume_test(sse_func, orig_func);
+    run_volume_test(sse_func, orig_func, 0, TRUE, FALSE);
+    run_volume_test(sse_func, orig_func, 1, TRUE, FALSE);
+    run_volume_test(sse_func, orig_func, 2, TRUE, FALSE);
+    run_volume_test(sse_func, orig_func, 3, TRUE, FALSE);
+    run_volume_test(sse_func, orig_func, 4, TRUE, FALSE);
+    run_volume_test(sse_func, orig_func, 5, TRUE, FALSE);
+    run_volume_test(sse_func, orig_func, 6, TRUE, FALSE);
+    run_volume_test(sse_func, orig_func, 7, TRUE, TRUE);
 }
 END_TEST
 #endif /* defined (__i386__) || defined (__amd64__) */
@@ -168,7 +201,14 @@ START_TEST (svolume_arm_test) {
     arm_func = pa_get_volume_func(PA_SAMPLE_S16NE);
 
     pa_log_debug("Checking ARM svolume");
-    run_volume_test(arm_func, orig_func);
+    run_volume_test(arm_func, orig_func, 0, TRUE, TRUE);
+    run_volume_test(arm_func, orig_func, 1, TRUE, TRUE);
+    run_volume_test(arm_func, orig_func, 2, TRUE, TRUE);
+    run_volume_test(arm_func, orig_func, 3, TRUE, TRUE);
+    run_volume_test(arm_func, orig_func, 4, TRUE, TRUE);
+    run_volume_test(arm_func, orig_func, 5, TRUE, TRUE);
+    run_volume_test(arm_func, orig_func, 6, TRUE, TRUE);
+    run_volume_test(arm_func, orig_func, 7, TRUE, TRUE);
 }
 END_TEST
 #endif /* defined (__arm__) && defined (__linux__) */
@@ -193,7 +233,14 @@ START_TEST (svolume_orc_test) {
     orc_func = pa_get_volume_func(PA_SAMPLE_S16NE);
 
     pa_log_debug("Checking Orc svolume");
-    run_volume_test(orc_func, orig_func);
+    run_volume_test(orc_func, orig_func, 0, TRUE, FALSE);
+    run_volume_test(orc_func, orig_func, 1, TRUE, FALSE);
+    run_volume_test(orc_func, orig_func, 2, TRUE, FALSE);
+    run_volume_test(orc_func, orig_func, 3, TRUE, FALSE);
+    run_volume_test(orc_func, orig_func, 4, TRUE, FALSE);
+    run_volume_test(orc_func, orig_func, 5, TRUE, FALSE);
+    run_volume_test(orc_func, orig_func, 6, TRUE, FALSE);
+    run_volume_test(orc_func, orig_func, 7, TRUE, TRUE);
 }
 END_TEST
 
@@ -402,6 +449,7 @@ int main(int argc, char *argv[]) {
     tcase_add_test(tc, svolume_arm_test);
 #endif
     tcase_add_test(tc, svolume_orc_test);
+    tcase_set_timeout(tc, 120);
     suite_add_tcase(s, tc);
 
     /* Converstion tests */