* src/bitstream.h:

Fix some more bugs in the optimized path and reenable it again. 


git-svn-id: https://core.fluendo.com/gstreamer/svn/trunk/gst-fluendo-mp3@2273 2b0047a9-a6d8-0310-accf-f7200b2a168c

1 files changed, 47 insertions, 26 deletions

diff --git a/src/bitstream.h b/src/bitstream.h
index a90c5c7..f732f9e 100644
--- a/src/bitstream.h
+++ b/src/bitstream.h
@@ -16,7 +16,7 @@
 #include <gst/gst.h>
 
 /* Accumulator optimization on bitstream management */
-#define ENABLE_OPT_BS 0
+#define ENABLE_OPT_BS 1
 
 /* Bit stream reader definitions */
 #define         MAX_LENGTH      32      /* Maximum length of word written or
@@ -160,7 +160,11 @@ static inline void h_byte_align (huffdec_bitbuf * bb);
 static inline guint h_bytes_avail (huffdec_bitbuf * bb);
 
 /* Return the current bit stream position (in bits) */
+#if ENABLE_OPT_BS
+#define h_sstell(bb) ((bb->buf_byte_idx * 8) - bb->buf_bit_idx)
+#else
 #define h_sstell(bb) ((bb->buf_byte_idx * 8) + (BS_ACUM_SIZE - bb->buf_bit_idx))
+#endif
 
 static inline guint32
 bs_get1bit (Bit_stream_struc * bs)
@@ -380,34 +384,51 @@ h_get1bit (huffdec_bitbuf * bb)
 static inline void
 h_flushbits (huffdec_bitbuf * bb, guint N)
 {
-  guint bits = N % 8;
-  guint bytes = N / 8;
-  guint acc_bytes;
-
-  if (bb->buf_bit_idx >= bits)
-    bb->buf_bit_idx -= bits;
-
-  acc_bytes = bb->buf_bit_idx / 8;
-
-  if (acc_bytes >= bytes)
-    bb->buf_bit_idx -= bytes * 8;
-  else {
-    bb->buf_bit_idx -= acc_bytes * 8;
-    bytes -= acc_bytes;
-    if (bytes <= bb->remaining) {
-      bytes--;
-      bits = 8 - bb->buf_bit_idx;
-      if (bytes > 0) {
-        bb->buf += bytes;
-        bb->remaining -= bytes;
+  guint bits;
+  guint bytes;
+
+  if (N < 32) {
+    bits = N;
+  } else {
+    N -= bb->buf_bit_idx;
+    bytes = N >> 3;
+    bits = N & 0x7;
+    bb->buf_byte_idx += bytes;
+    bb->remaining -= bytes;
+    bb->buf_bit_idx = 0;
+
+    if (bb->remaining >= 4) {
+      /* reload the accumulator */
+      bb->buf_bit_idx = 32;       /* subtract 1 bit */
+      bb->remaining -= 4;
+
+      /* we need reverse the byte order */
+  #if defined(HAVE_CPU_I386) && !defined (_MSC_VER)
+      register guint32 tmp = *((guint32 *) (bb->buf + bb->buf_byte_idx));
+    __asm__ ("bswap %0 \n\t": "=r" (tmp):"0" (tmp));
+      bb->accumulator = tmp;
+  #else
+      bb->accumulator = (guint32) bb->buf[bb->buf_byte_idx + 3];
+      bb->accumulator |= (guint32) (bb->buf[bb->buf_byte_idx + 2]) << 8;
+      bb->accumulator |= (guint32) (bb->buf[bb->buf_byte_idx + 1]) << 16;
+      bb->accumulator |= (guint32) (bb->buf[bb->buf_byte_idx + 0]) << 24;
+  #endif
+      bb->buf_byte_idx += 4;
+    } else if (bb->remaining > 0) {
+      /* reload the accumulator */
+      bb->buf_bit_idx = (bb->remaining * 8) - 1;  /* subtract 1 bit  */
+
+      bb->accumulator = 0;
+      /* load remaining bytes into the accumulator in the right order */
+      for (; bb->remaining > 0; bb->remaining--) {
+        bb->accumulator <<= 8;
+        bb->accumulator |= (guint32) bb->buf[bb->buf_byte_idx++];
       }
-      bb->buf_bit_idx = 0;
-      h_getbits (bb, bits);
-    } else {
-      bb->remaining = 0;
-      bb->buf_bit_idx = 0;
     }
   }
+
+  if (bits)
+    h_getbits (bb, bits);
 }
 
 #else