h264: use one table instead of several for cabac functions
[ffmpeg.git] / libavcodec / x86 / cabac.h
index 3c3652d..32ce2b2 100644 (file)
 #include "libavcodec/cabac.h"
 #include "libavutil/attributes.h"
 #include "libavutil/x86_cpu.h"
+#include "libavutil/internal.h"
 #include "config.h"
 
 #if HAVE_FAST_CMOV
-#define BRANCHLESS_GET_CABAC_UPDATE(ret, statep, low, lowword, range, tmp)\
+#define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\
         "mov    "tmp"       , %%ecx     \n\t"\
         "shl    $17         , "tmp"     \n\t"\
         "cmp    "low"       , "tmp"     \n\t"\
@@ -37,7 +38,7 @@
         "xor    %%ecx       , "ret"     \n\t"\
         "sub    "tmp"       , "low"     \n\t"
 #else /* HAVE_FAST_CMOV */
-#define BRANCHLESS_GET_CABAC_UPDATE(ret, statep, low, lowword, range, tmp)\
+#define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\
         "mov    "tmp"       , %%ecx     \n\t"\
         "shl    $17         , "tmp"     \n\t"\
         "sub    "low"       , "tmp"     \n\t"\
         "xor    "tmp"       , "ret"     \n\t"
 #endif /* HAVE_FAST_CMOV */
 
-#define BRANCHLESS_GET_CABAC(ret, statep, low, lowword, range, tmp, tmpbyte, byte) \
+#define BRANCHLESS_GET_CABAC(ret, statep, low, lowword, range, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off) \
         "movzbl "statep"    , "ret"                                     \n\t"\
         "mov    "range"     , "tmp"                                     \n\t"\
         "and    $0xC0       , "range"                                   \n\t"\
-        "movzbl "MANGLE(ff_h264_lps_range)"("ret", "range", 2), "range" \n\t"\
+        "movzbl "MANGLE(ff_h264_cabac_tables)"+"lps_off"("ret", "range", 2), "range" \n\t"\
         "sub    "range"     , "tmp"                                     \n\t"\
-        BRANCHLESS_GET_CABAC_UPDATE(ret, statep, low, lowword, range, tmp)   \
-        "movzbl " MANGLE(ff_h264_norm_shift) "("range"), %%ecx          \n\t"\
+        BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)                    \
+        "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"("range"), %%ecx    \n\t"\
         "shl    %%cl        , "range"                                   \n\t"\
-        "movzbl "MANGLE(ff_h264_mlps_state)"+128("ret"), "tmp"          \n\t"\
+        "movzbl "MANGLE(ff_h264_cabac_tables)"+"mlps_off"+128("ret"), "tmp"  \n\t"\
         "shl    %%cl        , "low"                                     \n\t"\
         "mov    "tmpbyte"   , "statep"                                  \n\t"\
         "test   "lowword"   , "lowword"                                 \n\t"\
-        " jnz   1f                                                      \n\t"\
+        " jnz   2f                                                      \n\t"\
         "mov    "byte"      , %%"REG_c"                                 \n\t"\
+        "cmp    "end"       , %%"REG_c"                                 \n\t"\
+        "jge    1f                                                      \n\t"\
         "add"OPSIZE" $2     , "byte"                                    \n\t"\
+        "1:                                                             \n\t"\
         "movzwl (%%"REG_c")     , "tmp"                                 \n\t"\
         "lea    -1("low")   , %%ecx                                     \n\t"\
         "xor    "low"       , %%ecx                                     \n\t"\
         "shr    $15         , %%ecx                                     \n\t"\
         "bswap  "tmp"                                                   \n\t"\
         "shr    $15         , "tmp"                                     \n\t"\
-        "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx            \n\t"\
+        "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"(%%ecx), %%ecx \n\t"\
         "sub    $0xFFFF     , "tmp"                                     \n\t"\
         "neg    %%ecx                                                   \n\t"\
         "add    $7          , %%ecx                                     \n\t"\
         "shl    %%cl        , "tmp"                                     \n\t"\
         "add    "tmp"       , "low"                                     \n\t"\
-        "1:                                                             \n\t"
+        "2:                                                             \n\t"
 
 #if HAVE_7REGS && !defined(BROKEN_RELOCATIONS)
 #define get_cabac_inline get_cabac_inline_x86
@@ -89,11 +93,16 @@ static av_always_inline int get_cabac_inline_x86(CABACContext *c,
     int bit, tmp;
 
     __asm__ volatile(
-        BRANCHLESS_GET_CABAC("%0", "(%5)", "%1", "%w1", "%2",
-                             "%3", "%b3", "%4")
-        :"=&r"(bit), "+&r"(c->low), "+&r"(c->range), "=&q"(tmp),
-         "+m"(c->bytestream)
-        :"r"(state)
+        BRANCHLESS_GET_CABAC("%0", "(%4)", "%1", "%w1",
+                             "%2", "%3", "%b3",
+                             "%a6(%5)", "%a7(%5)", "%a8", "%a9", "%a10")
+        : "=&r"(bit), "+&r"(c->low), "+&r"(c->range), "=&q"(tmp)
+        : "r"(state), "r"(c),
+          "i"(offsetof(CABACContext, bytestream)),
+          "i"(offsetof(CABACContext, bytestream_end)),
+          "i"(H264_NORM_SHIFT_OFFSET),
+          "i"(H264_LPS_RANGE_OFFSET),
+          "i"(H264_MLPS_STATE_OFFSET)
         : "%"REG_c, "memory"
     );
     return bit & 1;
@@ -105,32 +114,37 @@ static av_always_inline int get_cabac_bypass_sign_x86(CABACContext *c, int val)
 {
     x86_reg tmp;
     __asm__ volatile(
-        "movl %4, %k1                           \n\t"
-        "movl %2, %%eax                         \n\t"
-        "shl $17, %k1                           \n\t"
-        "add %%eax, %%eax                       \n\t"
-        "sub %k1, %%eax                         \n\t"
-        "cltd                                   \n\t"
-        "and %%edx, %k1                         \n\t"
-        "add %k1, %%eax                         \n\t"
-        "xor %%edx, %%ecx                       \n\t"
-        "sub %%edx, %%ecx                       \n\t"
-        "test %%ax, %%ax                        \n\t"
-        " jnz 1f                                \n\t"
-        "mov  %3, %1                            \n\t"
-        "subl $0xFFFF, %%eax                    \n\t"
-        "movzwl (%1), %%edx                     \n\t"
-        "bswap %%edx                            \n\t"
-        "shrl $15, %%edx                        \n\t"
-        "add  $2, %1                            \n\t"
-        "addl %%edx, %%eax                      \n\t"
-        "mov  %1, %3                            \n\t"
-        "1:                                     \n\t"
-        "movl %%eax, %2                         \n\t"
+        "movl        %a6(%2), %k1       \n\t"
+        "movl        %a3(%2), %%eax     \n\t"
+        "shl             $17, %k1       \n\t"
+        "add           %%eax, %%eax     \n\t"
+        "sub             %k1, %%eax     \n\t"
+        "cltd                           \n\t"
+        "and           %%edx, %k1       \n\t"
+        "add             %k1, %%eax     \n\t"
+        "xor           %%edx, %%ecx     \n\t"
+        "sub           %%edx, %%ecx     \n\t"
+        "test           %%ax, %%ax      \n\t"
+        "jnz              1f            \n\t"
+        "mov         %a4(%2), %1        \n\t"
+        "subl        $0xFFFF, %%eax     \n\t"
+        "movzwl         (%1), %%edx     \n\t"
+        "bswap         %%edx            \n\t"
+        "shrl            $15, %%edx     \n\t"
+        "addl          %%edx, %%eax     \n\t"
+        "cmp         %a5(%2), %1        \n\t"
+        "jge              1f            \n\t"
+        "add"OPSIZE"      $2, %a4(%2)   \n\t"
+        "1:                             \n\t"
+        "movl          %%eax, %a3(%2)   \n\t"
 
-        :"+c"(val), "=&r"(tmp), "+m"(c->low), "+m"(c->bytestream)
-        :"m"(c->range)
-        : "%eax", "%edx"
+        : "+c"(val), "=&r"(tmp)
+        : "r"(c),
+          "i"(offsetof(CABACContext, low)),
+          "i"(offsetof(CABACContext, bytestream)),
+          "i"(offsetof(CABACContext, bytestream_end)),
+          "i"(offsetof(CABACContext, range))
+        : "%eax", "%edx", "memory"
     );
     return val;
 }