x86: cabac: don't load/store context values in asm
[ffmpeg.git] / libavcodec / x86 / h264_i386.h
index b303347..eb77002 100644 (file)
@@ -32,6 +32,7 @@
 #include <stddef.h>
 
 #include "libavcodec/cabac.h"
+#include "cabac.h"
 
 //FIXME use some macros to avoid duplicating get_cabac (cannot be done yet
 //as that would make optimization work hard)
@@ -40,124 +41,107 @@ static int decode_significance_x86(CABACContext *c, int max_coeff,
                                    uint8_t *significant_coeff_ctx_base,
                                    int *index, x86_reg last_off){
     void *end= significant_coeff_ctx_base + max_coeff - 1;
-    int minusstart= -(int)significant_coeff_ctx_base;
-    int minusindex= 4-(int)index;
+    int minusstart= -(intptr_t)significant_coeff_ctx_base;
+    int minusindex= 4-(intptr_t)index;
     int bit;
     x86_reg coeff_count;
-    int low;
     __asm__ volatile(
-        "movl %a10(%5), %%esi                   \n\t"
-        "movl %a11(%5), %3                      \n\t"
-
         "2:                                     \n\t"
 
-        BRANCHLESS_GET_CABAC("%4", "%5", "(%1)", "%3",
-                             "%w3", "%%esi", "%k0", "%b0", "%a12")
+        BRANCHLESS_GET_CABAC("%4", "%6", "(%1)", "%3",
+                             "%w3", "%5", "%k0", "%b0", "%a11")
 
         "test $1, %4                            \n\t"
         " jz 3f                                 \n\t"
-        "add  %9, %1                            \n\t"
+        "add  %10, %1                           \n\t"
 
-        BRANCHLESS_GET_CABAC("%4", "%5", "(%1)", "%3",
-                             "%w3", "%%esi", "%k0", "%b0", "%a12")
+        BRANCHLESS_GET_CABAC("%4", "%6", "(%1)", "%3",
+                             "%w3", "%5", "%k0", "%b0", "%a11")
 
-        "sub  %9, %1                            \n\t"
+        "sub  %10, %1                           \n\t"
         "mov  %2, %0                            \n\t"
-        "movl %6, %%ecx                         \n\t"
+        "movl %7, %%ecx                         \n\t"
         "add  %1, %%"REG_c"                     \n\t"
         "movl %%ecx, (%0)                       \n\t"
 
         "test $1, %4                            \n\t"
         " jnz 4f                                \n\t"
 
-        "add  $4, %0                            \n\t"
-        "mov  %0, %2                            \n\t"
+        "add  $4, %2                            \n\t"
 
         "3:                                     \n\t"
         "add  $1, %1                            \n\t"
-        "cmp  %7, %1                            \n\t"
+        "cmp  %8, %1                            \n\t"
         " jb 2b                                 \n\t"
         "mov  %2, %0                            \n\t"
-        "movl %6, %%ecx                         \n\t"
+        "movl %7, %%ecx                         \n\t"
         "add  %1, %%"REG_c"                     \n\t"
         "movl %%ecx, (%0)                       \n\t"
         "4:                                     \n\t"
-        "add  %8, %k0                           \n\t"
+        "add  %9, %k0                           \n\t"
         "shr $2, %k0                            \n\t"
-
-        "movl %%esi, %a10(%5)                   \n\t"
-        "movl %3, %a11(%5)                      \n\t"
-        :"=&r"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index),
-         "=&r"(low), "=&r"(bit)
+        :"=&q"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index),
+         "+&r"(c->low), "=&r"(bit), "+&r"(c->range)
         :"r"(c), "m"(minusstart), "m"(end), "m"(minusindex), "m"(last_off),
-         "i"(offsetof(CABACContext, range)), "i"(offsetof(CABACContext, low)),
          "i"(offsetof(CABACContext, bytestream))
-        : "%"REG_c, "%esi", "memory"
+        : "%"REG_c, "memory"
     );
     return coeff_count;
 }
 
 static int decode_significance_8x8_x86(CABACContext *c,
                                        uint8_t *significant_coeff_ctx_base,
-                                       int *index, x86_reg last_off, const uint8_t *sig_off){
-    int minusindex= 4-(int)index;
+                                       int *index, uint8_t *last_coeff_ctx_base, const uint8_t *sig_off){
+    int minusindex= 4-(intptr_t)index;
     int bit;
     x86_reg coeff_count;
-    int low;
     x86_reg last=0;
+    x86_reg state;
     __asm__ volatile(
-        "movl %a10(%5), %%esi                   \n\t"
-        "movl %a11(%5), %3                      \n\t"
-
-        "mov %1, %%"REG_D"                      \n\t"
+        "mov %1, %6                             \n\t"
         "2:                                     \n\t"
 
-        "mov %8, %0                             \n\t"
-        "movzbl (%0, %%"REG_D"), %%edi          \n\t"
-        "add %7, %%"REG_D"                      \n\t"
+        "mov %10, %0                            \n\t"
+        "movzbl (%0, %6), %k6                   \n\t"
+        "add %9, %6                             \n\t"
 
-        BRANCHLESS_GET_CABAC("%4", "%5", "(%%"REG_D")", "%3",
-                             "%w3", "%%esi", "%k0", "%b0", "%a12")
+        BRANCHLESS_GET_CABAC("%4", "%7", "(%6)", "%3",
+                             "%w3", "%5", "%k0", "%b0", "%a12")
 
-        "mov %1, %%edi                          \n\t"
+        "mov %1, %k6                            \n\t"
         "test $1, %4                            \n\t"
         " jz 3f                                 \n\t"
 
-        "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%%edi), %%edi\n\t"
-        "add %7, %%"REG_D"                      \n\t"
-        "add %9, %%"REG_D"                      \n\t"
+        "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%k6), %k6\n\t"
+        "add %11, %6                            \n\t"
 
-        BRANCHLESS_GET_CABAC("%4", "%5", "(%%"REG_D")", "%3",
-                             "%w3", "%%esi", "%k0", "%b0", "%a12")
+        BRANCHLESS_GET_CABAC("%4", "%7", "(%6)", "%3",
+                             "%w3", "%5", "%k0", "%b0", "%a12")
 
         "mov %2, %0                             \n\t"
-        "mov %1, %%edi                          \n\t"
-        "movl %%edi, (%0)                       \n\t"
+        "mov %1, %k6                            \n\t"
+        "movl %k6, (%0)                         \n\t"
 
         "test $1, %4                            \n\t"
         " jnz 4f                                \n\t"
 
-        "add $4, %0                             \n\t"
-        "mov %0, %2                             \n\t"
+        "add  $4, %2                            \n\t"
 
         "3:                                     \n\t"
-        "addl $1, %%edi                         \n\t"
-        "mov %%edi, %1                          \n\t"
-        "cmpl $63, %%edi                        \n\t"
+        "addl $1, %k6                           \n\t"
+        "mov %k6, %1                            \n\t"
+        "cmpl $63, %k6                          \n\t"
         " jb 2b                                 \n\t"
         "mov %2, %0                             \n\t"
-        "movl %%edi, (%0)                       \n\t"
+        "movl %k6, (%0)                         \n\t"
         "4:                                     \n\t"
-        "addl %6, %k0                           \n\t"
+        "addl %8, %k0                           \n\t"
         "shr $2, %k0                            \n\t"
-
-        "movl %%esi, %a10(%5)                   \n\t"
-        "movl %3, %a11(%5)                      \n\t"
-        :"=&r"(coeff_count),"+m"(last), "+m"(index), "=&r"(low), "=&r"(bit)
-        :"r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base), "m"(sig_off), "m"(last_off),
-         "i"(offsetof(CABACContext, range)), "i"(offsetof(CABACContext, low)),
+        :"=&q"(coeff_count),"+m"(last), "+m"(index), "+&r"(c->low), "=&r"(bit),
+         "+&r"(c->range), "=&r"(state)
+        :"r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base), "m"(sig_off), "m"(last_coeff_ctx_base),
          "i"(offsetof(CABACContext, bytestream))
-        : "%"REG_c, "%esi", "%"REG_D, "memory"
+        : "%"REG_c, "memory"
     );
     return coeff_count;
 }