[pulseaudio-discuss] [PATCH] echo-cancel: Handle alignment requirement manually

[Arun Raghavan]

PA_ALIGNED can't always guarantee that the alignment we want (the GCC
man page suggests that the linker might not be able to meet the
alignment requirements we desire). Instead, we now allocate some extra
memory and guaratee that the alignment we require is met.
---
 src/modules/echo-cancel/adrian-aec.c |   12 +++++++++---
 src/modules/echo-cancel/adrian-aec.h |    3 ++-
 2 files changed, 11 insertions(+), 4 deletions(-)

diff --git a/src/modules/echo-cancel/adrian-aec.c b/src/modules/echo-cancel/adrian-aec.c
index 04b31e9..dfe8ada 100644
--- a/src/modules/echo-cancel/adrian-aec.c
+++ b/src/modules/echo-cancel/adrian-aec.c
@@ -12,6 +12,7 @@
 
 #include <math.h>
 #include <string.h>
+#include <stdint.h>
 
 #include <pulse/xmalloc.h>
 
@@ -70,7 +71,7 @@ AEC* AEC_init(int RATE, int have_vector)
   a->hangover = 0;
   memset(a->x, 0, sizeof(a->x));
   memset(a->xf, 0, sizeof(a->xf));
-  memset(a->w, 0, sizeof(a->w));
+  memset(a->w_arr, 0, sizeof(a->w_arr));
   a->j = NLMS_EXT;
   a->delta = 0.0f;
   AEC_setambient(a, NoiseFloor);
@@ -89,10 +90,15 @@ AEC* AEC_init(int RATE, int have_vector)
   a->dumpcnt = 0;
   memset(a->ws, 0, sizeof(a->ws));
 
-  if (have_vector)
+  if (have_vector) {
+      /* Get a 16-byte aligned location */
+      a->w = (REAL *) (((uintptr_t) a->w_arr) + (((uintptr_t) a->w_arr) % 16));
       a->dotp = dotp_sse;
-  else
+  } else {
+      /* We don't care about alignment, just use the array as-is */
+      a->w = a->w_arr;
       a->dotp = dotp;
+  }
 
   return a;
 }
diff --git a/src/modules/echo-cancel/adrian-aec.h b/src/modules/echo-cancel/adrian-aec.h
index 9c722b9..efb9e27 100644
--- a/src/modules/echo-cancel/adrian-aec.h
+++ b/src/modules/echo-cancel/adrian-aec.h
@@ -306,7 +306,8 @@ struct AEC {
   // NLMS-pw
   REAL x[NLMS_LEN + NLMS_EXT];  // tap delayed loudspeaker signal
   REAL xf[NLMS_LEN + NLMS_EXT]; // pre-whitening tap delayed signal
-  PA_DECLARE_ALIGNED(16, REAL, w[NLMS_LEN]);             // tap weights
+  REAL w_arr[NLMS_LEN+16];      // tap weights
+  REAL *w;                      // this will be a 16-byte aligned pointer into w_arr
   int j;                        // optimize: less memory copies
   double dotp_xf_xf;            // double to avoid loss of precision
   float delta;                  // noise floor to stabilize NLMS
-- 
1.7.5.rc1