Align various libwma buffers. Saves about 1 MHz on the Gigabeat S.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@28640 a1c6a512-1295-4272-9138-f99709370657

2 files changed, 14 insertions, 12 deletions

diff --git a/apps/codecs/libwma/wmadec.h b/apps/codecs/libwma/wmadec.h
index d84485ca0f..d7fa79d268 100644
--- a/apps/codecs/libwma/wmadec.h
+++ b/apps/codecs/libwma/wmadec.h
@@ -130,7 +130,7 @@ typedef struct WMADecodeContext
     uint8_t ms_stereo; /* true if mid/side stereo mode */
     uint8_t channel_coded[MAX_CHANNELS]; /* true if channel is coded */
     int exponents_bsize[MAX_CHANNELS];      // log2 ratio frame/exp. length
-    fixed32 exponents[MAX_CHANNELS][BLOCK_MAX_SIZE];
+    fixed32 exponents[MAX_CHANNELS][BLOCK_MAX_SIZE] MEM_ALIGN_ATTR;
     fixed32 max_exponent[MAX_CHANNELS];
     int16_t coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE];
     fixed32 (*coefs)[MAX_CHANNELS][BLOCK_MAX_SIZE];
@@ -139,14 +139,14 @@ typedef struct WMADecodeContext
     fixed32 (*frame_out)[MAX_CHANNELS][BLOCK_MAX_SIZE*2];
 
     /* last frame info */
-    uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4]; /* padding added */
+    uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4] MEM_ALIGN_ATTR; /* padding added */
     int last_bitoffset;
     int last_superframe_len;
     fixed32 *noise_table;
     int noise_index;
     fixed32 noise_mult; /* XXX: suppress that and integrate it in the noise array */
     /* lsp_to_curve tables */
-    fixed32 lsp_cos_table[BLOCK_MAX_SIZE];
+    fixed32 lsp_cos_table[BLOCK_MAX_SIZE] MEM_ALIGN_ATTR;
     void *lsp_pow_m_table1;
     void *lsp_pow_m_table2;
 
diff --git a/apps/codecs/libwma/wmadeci.c b/apps/codecs/libwma/wmadeci.c
index 9333f19ff1..98c10e486d 100644
--- a/apps/codecs/libwma/wmadeci.c
+++ b/apps/codecs/libwma/wmadeci.c
@@ -33,17 +33,19 @@ static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len);
 
 /*declarations of statically allocated variables used to remove malloc calls*/
 
-fixed32 coefsarray[MAX_CHANNELS][BLOCK_MAX_SIZE] IBSS_ATTR;
+fixed32 coefsarray[MAX_CHANNELS][BLOCK_MAX_SIZE] IBSS_ATTR MEM_ALIGN_ATTR;
 /*decode and window into IRAM on targets with at least 80KB of codec IRAM*/
-fixed32 frame_out_buf[MAX_CHANNELS][BLOCK_MAX_SIZE * 2] IBSS_ATTR_WMA_LARGE_IRAM;
+fixed32 frame_out_buf[MAX_CHANNELS][BLOCK_MAX_SIZE * 2] IBSS_ATTR_WMA_LARGE_IRAM MEM_ALIGN_ATTR;
 
 /*MDCT reconstruction windows*/
-fixed32 stat0[2048], stat1[1024], stat2[512], stat3[256], stat4[128];    
+fixed32 stat0[2048] MEM_ALIGN_ATTR, stat1[1024] MEM_ALIGN_ATTR, 
+    stat2[512] MEM_ALIGN_ATTR, stat3[256] MEM_ALIGN_ATTR, stat4[128] MEM_ALIGN_ATTR;    
 
 /*VLC lookup tables*/
 uint16_t *runtabarray[2], *levtabarray[2];                                        
 
-uint16_t runtab_big[1336], runtab_small[1072], levtab_big[1336], levtab_small[1072];
+uint16_t runtab_big[1336] MEM_ALIGN_ATTR, runtab_small[1072] MEM_ALIGN_ATTR,
+    levtab_big[1336] MEM_ALIGN_ATTR, levtab_small[1072] MEM_ALIGN_ATTR;
 
 #define VLCBUF1SIZE 4598
 #define VLCBUF2SIZE 3574
@@ -52,11 +54,11 @@ uint16_t runtab_big[1336], runtab_small[1072], levtab_big[1336], levtab_small[10
 
 /*putting these in IRAM actually makes PP slower*/
 
-VLC_TYPE vlcbuf1[VLCBUF1SIZE][2];
-VLC_TYPE vlcbuf2[VLCBUF2SIZE][2];
+VLC_TYPE vlcbuf1[VLCBUF1SIZE][2] MEM_ALIGN_ATTR;
+VLC_TYPE vlcbuf2[VLCBUF2SIZE][2] MEM_ALIGN_ATTR;
 /* This buffer gets reused for lsp tables */
-VLC_TYPE vlcbuf3[VLCBUF3SIZE][2] __attribute__((aligned (sizeof(fixed32))));
-VLC_TYPE vlcbuf4[VLCBUF4SIZE][2];
+VLC_TYPE vlcbuf3[VLCBUF3SIZE][2] MEM_ALIGN_ATTR;
+VLC_TYPE vlcbuf4[VLCBUF4SIZE][2] MEM_ALIGN_ATTR;
 
 
 
@@ -1232,7 +1234,7 @@ static int wma_decode_block(WMADecodeContext *s)
     for(ch = 0; ch < s->nb_channels; ++ch)
     { 
         /* BLOCK_MAX_SIZE is 2048 (samples) and MAX_CHANNELS is 2. */
-        static uint32_t scratch_buf[BLOCK_MAX_SIZE * MAX_CHANNELS] IBSS_ATTR;
+        static uint32_t scratch_buf[BLOCK_MAX_SIZE * MAX_CHANNELS] IBSS_ATTR MEM_ALIGN_ATTR;
         if (s->channel_coded[ch])
         {
             int n4, index;