# HG changeset patch
# User banan
# Date 1166118623 0
# Node ID b43bd0c56eaaaf025098487247c86d6ca1899d14
# Parent  95044f59417099b48f46006d0ccbdb04e088e614
Bug fix for crashes when SSE is used on unaligned arrays.
No measureable change in speed. This gave random crashes on Win32
and BeOS. The cause for this bug is that gcc doesn't align the
stackframe. Linux and glibc always ensure this to be true thus
this never affected Linux.

diff -r 95044f594170 -r b43bd0c56eaa wmadec.c
--- a/wmadec.c	Wed Dec 13 11:10:41 2006 +0000
+++ b/wmadec.c	Thu Dec 14 17:50:23 2006 +0000
@@ -115,6 +115,8 @@
     float max_exponent[MAX_CHANNELS];
     int16_t coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE];
     DECLARE_ALIGNED_16(float, coefs[MAX_CHANNELS][BLOCK_MAX_SIZE]);
+    DECLARE_ALIGNED_16(FFTSample, output[BLOCK_MAX_SIZE * 2]);
+    DECLARE_ALIGNED_16(float, window[BLOCK_MAX_SIZE * 2]);
     MDCTContext mdct_ctx[BLOCK_NB_SIZES];
     float *windows[BLOCK_NB_SIZES];
     DECLARE_ALIGNED_16(FFTSample, mdct_tmp[BLOCK_MAX_SIZE]); /* temporary storage for imdct */
@@ -717,7 +719,6 @@
 {
     int n, v, a, ch, code, bsize;
     int coef_nb_bits, total_gain, parse_exponents;
-    DECLARE_ALIGNED_16(float, window[BLOCK_MAX_SIZE * 2]);
     int nb_coefs[MAX_CHANNELS];
     float mdct_norm;
 
@@ -1072,7 +1073,7 @@
         next_block_len = 1 << s->next_block_len_bits;
 
         /* right part */
-        wptr = window + block_len;
+        wptr = s->window + block_len;
         if (block_len <= next_block_len) {
             for(i=0;i<block_len;i++)
                 *wptr++ = s->windows[bsize][i];
@@ -1088,7 +1089,7 @@
         }
 
         /* left part */
-        wptr = window + block_len;
+        wptr = s->window + block_len;
         if (block_len <= prev_block_len) {
             for(i=0;i<block_len;i++)
                 *--wptr = s->windows[bsize][i];
@@ -1107,14 +1108,13 @@
 
     for(ch = 0; ch < s->nb_channels; ch++) {
         if (s->channel_coded[ch]) {
-            DECLARE_ALIGNED_16(FFTSample, output[BLOCK_MAX_SIZE * 2]);
             float *ptr;
             int n4, index, n;
 
             n = s->block_len;
             n4 = s->block_len / 2;
             s->mdct_ctx[bsize].fft.imdct_calc(&s->mdct_ctx[bsize],
-                          output, s->coefs[ch], s->mdct_tmp);
+                          s->output, s->coefs[ch], s->mdct_tmp);
 
             /* XXX: optimize all that by build the window and
                multipying/adding at the same time */
@@ -1122,13 +1122,13 @@
             /* multiply by the window and add in the frame */
             index = (s->frame_len / 2) + s->block_pos - n4;
             ptr = &s->frame_out[ch][index];
-            s->dsp.vector_fmul_add_add(ptr,window,output,ptr,0,2*n,1);
+            s->dsp.vector_fmul_add_add(ptr,s->window,s->output,ptr,0,2*n,1);
 
             /* specific fast case for ms-stereo : add to second
                channel if it is not coded */
             if (s->ms_stereo && !s->channel_coded[1]) {
                 ptr = &s->frame_out[1][index];
-                s->dsp.vector_fmul_add_add(ptr,window,output,ptr,0,2*n,1);
+                s->dsp.vector_fmul_add_add(ptr,s->window,s->output,ptr,0,2*n,1);
             }
         }
     }