Bug fix for crashes when SSE is used on unaligned arrays.
authorBenjamin Larsson <banan@ludd.ltu.se>
Thu, 14 Dec 2006 17:50:23 +0000 (17:50 +0000)
committerBenjamin Larsson <banan@ludd.ltu.se>
Thu, 14 Dec 2006 17:50:23 +0000 (17:50 +0000)
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.

Originally committed as revision 7318 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/wmadec.c

index 684aea2c8f65f91c403120201e045b4fa832e1a2..bbf4970cec8243c7430fffc5f2ed180c8f2ac4f5 100644 (file)
@@ -115,6 +115,8 @@ typedef struct WMADecodeContext {
     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 @@ static int wma_decode_block(WMADecodeContext *s)
 {
     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 @@ static int wma_decode_block(WMADecodeContext *s)
         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 @@ static int wma_decode_block(WMADecodeContext *s)
         }
 
         /* 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 @@ static int wma_decode_block(WMADecodeContext *s)
 
     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 @@ static int wma_decode_block(WMADecodeContext *s)
             /* 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);
             }
         }
     }