optimize branchless C CABAC decoder

Originally committed as revision 6607 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/cabac.c

@@ -133,11 +133,19 @@ void ff_init_cabac_states(CABACContext *c, uint8_t const (*lps_range)[4],
         c->mps_state[2*i+3]= 2*mps_state[i]+3;
 
         if( i ){
+#ifdef BRANCHLESS_CABAD
+            c->mps_state[-2*i-3]= 2*lps_state[i]+2; //FIXME yes this is not valid C but iam lazy, cleanup welcome
+            c->mps_state[-2*i-4]= 2*lps_state[i]+3;
+        }else{
+            c->mps_state[-2*i-3]= 3;
+            c->mps_state[-2*i-4]= 2;
+#else
             c->lps_state[2*i+2]= 2*lps_state[i]+2;
             c->lps_state[2*i+3]= 2*lps_state[i]+3;
         }else{
             c->lps_state[2*i+2]= 3;
             c->lps_state[2*i+3]= 2;
+#endif
         }
     }
 }

libavcodec/cabac.h

@@ -452,7 +452,7 @@ static int get_cabac(CABACContext *c, uint8_t * const state){
     int bit, lps_mask attribute_unused;
 
     c->range -= RangeLPS;
-#if 1
+#ifndef BRANCHLESS_CABAD
     if(c->low < c->range){
         bit= s&1;
         *state= c->mps_state[s];
@@ -475,8 +475,9 @@ static int get_cabac(CABACContext *c, uint8_t * const state){
     c->low -= c->range & lps_mask;
     c->range += (RangeLPS - c->range) & lps_mask;
 
-    bit= (s^lps_mask)&1;
-    *state= c->mps_state[s - (130&lps_mask)];
+    s^=lps_mask;
+    *state= c->mps_state[s];
+    bit= s&1;
 
     lps_mask= ff_h264_norm_shift[c->range>>(CABAC_BITS+3)];
     c->range<<= lps_mask;