"movq      %%mm4, (%%edi)  \n\t"                  "addl      $8, %%esi       \n\t" // inc by 8 bytes processed                  "addl      $8, %%edi       \n\t"                  "subl      $8, %%ecx       \n\t" // dec by 8 pixels processed                  "ja        mainloop8       \n\t"                "mainloop8end:               \n\t"// preload        "movl      diff, %%ecx     \n\t" // (diff is in eax)                  "movl      %%eax, %%ecx    \n\t"                  "cmpl      $0, %%ecx       \n\t"                  "jz        end8            \n\t"// preload        "movl      mask, %%edx     \n\t"                  "sall      $24, %%edx      \n\t" // make low byte, high byte                "secondloop8:                \n\t"                  "sall      %%edx           \n\t" // move high bit to CF                  "jnc       skip8           \n\t" // if CF = 0                  "movb      (%%esi), %%al   \n\t"                  "movb      %%al, (%%edi)   \n\t"                "skip8:                      \n\t"                  "incl      %%esi           \n\t"                  "incl      %%edi           \n\t"                  "decl      %%ecx           \n\t"                  "jnz       secondloop8     \n\t"                "end8:                       \n\t"                  "EMMS                      \n\t"  // DONE                  : "=a" (dummy_value_a),           // output regs (dummy)                    "=d" (dummy_value_d),                    "=c" (dummy_value_c),                    "=S" (dummy_value_S),                    "=D" (dummy_value_D)                  : "3" (srcptr),      // esi       // input regs                    "4" (dstptr),      // edi                    "0" (diff),        // eax// was (unmask)     "b"    RESERVED    // ebx       // Global Offset Table idx                    "2" (len),         // ecx                    "1" (mask)         // edx#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */                  : "%mm0", "%mm4", "%mm6", "%mm7"  // clobber list#endif               );            }            else /* mmx _not supported - Use modified C routine */#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */            {               register png_uint_32 i;               png_uint_32 initial_val = png_pass_start[png_ptr->pass];                 /* png.c:  png_pass_start[] = {0, 4, 0, 2, 0, 1, 0}; */               register int stride = png_pass_inc[png_ptr->pass];                 /* png.c:  png_pass_inc[] = {8, 8, 4, 4, 2, 2, 1}; */               register int rep_bytes = png_pass_width[png_ptr->pass];                 /* png.c:  png_pass_width[] = {8, 4, 4, 2, 2, 1, 1}; */               png_uint_32 len = png_ptr->width &~7;  /* reduce to mult. of 8 */               int diff = (int) (png_ptr->width & 7); /* amount lost */               register png_uint_32 final_val = len;  /* GRR bugfix */               srcptr = png_ptr->row_buf + 1 + initial_val;               dstptr = row + initial_val;               for (i = initial_val; i < final_val; i += stride)               {                  png_memcpy(dstptr, srcptr, rep_bytes);                  srcptr += stride;                  dstptr += stride;               }               if (diff)  /* number of leftover pixels:  3 for pngtest */               {                  final_val+=diff /* *BPP1 */ ;                  for (; i < final_val; i += stride)                  {                     if (rep_bytes > (int)(final_val-i))                        rep_bytes = (int)(final_val-i);                     png_memcpy(dstptr, srcptr, rep_bytes);                     srcptr += stride;                     dstptr += stride;                  }               }            } /* end of else (_mmx_supported) */            break;         }       /* end 8 bpp */         case 16:       /* png_ptr->row_info.pixel_depth */         {            png_bytep srcptr;            png_bytep dstptr;#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_THREAD_UNSAFE_OK)#if !defined(PNG_1_0_X)            if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_COMBINE_ROW)                /* && _mmx_supported */ )#else            if (_mmx_supported)#endif            {               png_uint_32 len;               int diff;               int dummy_value_a;   // fix 'forbidden register spilled' error               int dummy_value_d;               int dummy_value_c;               int dummy_value_S;               int dummy_value_D;               _unmask = ~mask;            // global variable for -fPIC version               srcptr = png_ptr->row_buf + 1;               dstptr = row;               len  = png_ptr->width &~7;  // reduce to multiple of 8               diff = (int) (png_ptr->width & 7); // amount lost //               __asm__ __volatile__ (                  "movd      _unmask, %%mm7   \n\t" // load bit pattern                  "psubb     %%mm6, %%mm6     \n\t" // zero mm6                  "punpcklbw %%mm7, %%mm7     \n\t"                  "punpcklwd %%mm7, %%mm7     \n\t"                  "punpckldq %%mm7, %%mm7     \n\t" // fill reg with 8 masks                  "movq      _mask16_0, %%mm0 \n\t"                  "movq      _mask16_1, %%mm1 \n\t"                  "pand      %%mm7, %%mm0     \n\t"                  "pand      %%mm7, %%mm1     \n\t"                  "pcmpeqb   %%mm6, %%mm0     \n\t"                  "pcmpeqb   %%mm6, %%mm1     \n\t"// preload        "movl      len, %%ecx       \n\t" // load length of line// preload        "movl      srcptr, %%esi    \n\t" // load source// preload        "movl      dstptr, %%edi    \n\t" // load dest                  "cmpl      $0, %%ecx        \n\t"                  "jz        mainloop16end    \n\t"                "mainloop16:                  \n\t"                  "movq      (%%esi), %%mm4   \n\t"                  "pand      %%mm0, %%mm4     \n\t"                  "movq      %%mm0, %%mm6     \n\t"                  "movq      (%%edi), %%mm7   \n\t"                  "pandn     %%mm7, %%mm6     \n\t"                  "por       %%mm6, %%mm4     \n\t"                  "movq      %%mm4, (%%edi)   \n\t"                  "movq      8(%%esi), %%mm5  \n\t"                  "pand      %%mm1, %%mm5     \n\t"                  "movq      %%mm1, %%mm7     \n\t"                  "movq      8(%%edi), %%mm6  \n\t"                  "pandn     %%mm6, %%mm7     \n\t"                  "por       %%mm7, %%mm5     \n\t"                  "movq      %%mm5, 8(%%edi)  \n\t"                  "addl      $16, %%esi       \n\t" // inc by 16 bytes processed                  "addl      $16, %%edi       \n\t"                  "subl      $8, %%ecx        \n\t" // dec by 8 pixels processed                  "ja        mainloop16       \n\t"                "mainloop16end:               \n\t"// preload        "movl      diff, %%ecx      \n\t" // (diff is in eax)                  "movl      %%eax, %%ecx     \n\t"                  "cmpl      $0, %%ecx        \n\t"                  "jz        end16            \n\t"// preload        "movl      mask, %%edx      \n\t"                  "sall      $24, %%edx       \n\t" // make low byte, high byte                "secondloop16:                \n\t"                  "sall      %%edx            \n\t" // move high bit to CF                  "jnc       skip16           \n\t" // if CF = 0                  "movw      (%%esi), %%ax    \n\t"                  "movw      %%ax, (%%edi)    \n\t"                "skip16:                      \n\t"                  "addl      $2, %%esi        \n\t"                  "addl      $2, %%edi        \n\t"                  "decl      %%ecx            \n\t"                  "jnz       secondloop16     \n\t"                "end16:                       \n\t"                  "EMMS                       \n\t" // DONE                  : "=a" (dummy_value_a),           // output regs (dummy)                    "=c" (dummy_value_c),                    "=d" (dummy_value_d),                    "=S" (dummy_value_S),                    "=D" (dummy_value_D)                  : "0" (diff),        // eax       // input regs// was (unmask)     " "    RESERVED    // ebx       // Global Offset Table idx                    "1" (len),         // ecx                    "2" (mask),        // edx                    "3" (srcptr),      // esi                    "4" (dstptr)       // edi#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */                  : "%mm0", "%mm1", "%mm4"          // clobber list                  , "%mm5", "%mm6", "%mm7"#endif               );            }            else /* mmx _not supported - Use modified C routine */#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */            {               register png_uint_32 i;               png_uint_32 initial_val = BPP2 * png_pass_start[png_ptr->pass];                 /* png.c:  png_pass_start[] = {0, 4, 0, 2, 0, 1, 0}; */               register int stride = BPP2 * png_pass_inc[png_ptr->pass];                 /* png.c:  png_pass_inc[] = {8, 8, 4, 4, 2, 2, 1}; */               register int rep_bytes = BPP2 * png_pass_width[png_ptr->pass];                 /* png.c:  png_pass_width[] = {8, 4, 4, 2, 2, 1, 1}; */               png_uint_32 len = png_ptr->width &~7;  /* reduce to mult. of 8 */               int diff = (int) (png_ptr->width & 7); /* amount lost */               register png_uint_32 final_val = BPP2 * len;   /* GRR bugfix */               srcptr = png_ptr->row_buf + 1 + initial_val;               dstptr = row + initial_val;               for (i = initial_val; i < final_val; i += stride)               {                  png_memcpy(dstptr, srcptr, rep_bytes);                  srcptr += stride;                  dstptr += stride;               }               if (diff)  /* number of leftover pixels:  3 for pngtest */               {                  final_val+=diff*BPP2;                  for (; i < final_val; i += stride)                  {                     if (rep_bytes > (int)(final_val-i))                        rep_bytes = (int)(final_val-i);                     png_memcpy(dstptr, srcptr, rep_bytes);                     srcptr += stride;                     dstptr += stride;                  }               }            } /* end of else (_mmx_supported) */            break;         }       /* end 16 bpp */         case 24:       /* png_ptr->row_info.pixel_depth */         {            png_bytep srcptr;            png_bytep dstptr;#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_THREAD_UNSAFE_OK)#if !defined(PNG_1_0_X)            if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_COMBINE_ROW)                /* && _mmx_supported */ )#else            if (_mmx_supported)#endif            {               png_uint_32 len;               int diff;               int dummy_value_a;   // fix 'forbidden register spilled' error               int dummy_value_d;               int dummy_value_c;               int dummy_value_S;               int dummy_value_D;               _unmask = ~mask;            // global variable for -fPIC version               srcptr = png_ptr->row_buf + 1;               dstptr = row;               len  = png_ptr->width &~7;  // reduce to multiple of 8               diff = (int) (png_ptr->width & 7); // amount lost //               __asm__ __volatile__ (                  "movd      _unmask, %%mm7   \n\t" // load bit pattern                  "psubb     %%mm6, %%mm6     \n\t" // zero mm6                  "punpcklbw %%mm7, %%mm7     \n\t"                  "punpcklwd %%mm7, %%mm7     \n\t"                  "punpckldq %%mm7, %%mm7     \n\t" // fill reg with 8 masks                  "movq      _mask24_0, %%mm0 \n\t"                  "movq      _mask24_1, %%mm1 \n\t"                  "movq      _mask24_2, %%mm2 \n\t"                  "pand      %%mm7, %%mm0     \n\t"                  "pand      %%mm7, %%mm1     \n\t"                  "pand      %%mm7, %%mm2     \n\t"                  "pcmpeqb   %%mm6, %%mm0     \n\t"                  "pcmpeqb   %%mm6, %%mm1     \n\t"                  "pcmpeqb   %%mm6, %%mm2     \n\t"// preload        "movl      len, %%ecx       \n\t" // load length of line// preload        "movl      srcptr, %%esi    \n\t" // load source// preload        "movl      dstptr, %%edi    \n\t" // load dest                  "cmpl      $0, %%ecx        \n\t"                  "jz        mainloop24end    \n\t"                "mainloop24:                  \n\t"                  "movq      (%%esi), %%mm4   \n\t"                  "pand      %%mm0, %%mm4     \n\t"                  "movq      %%mm0, %%mm6     \n\t"                  "movq      (%%edi), %%mm7   \n\t"                  "pandn     %%mm7, %%mm6     \n\t"                  "por       %%mm6, %%mm4     \n\t"                  "movq      %%mm4, (%%edi)   \n\t"                  "movq      8(%%esi), %%mm5  \n\t"                  "pand      %%mm1, %%mm5     \n\t"                  "movq      %%mm1, %%mm7     \n\t"                  "movq      8(%%edi), %%mm6  \n\t"                  "pandn     %%mm6, %%mm7     \n\t"                  "por       %%mm7, %%mm5     \n\t"                  "movq      %%mm5, 8(%%edi)  \n\t"                  "movq      16(%%esi), %%mm6 \n\t"                  "pand      %%mm2, %%mm6     \n\t"                  "movq      %%mm2, %%mm4     \n\t"                  "movq      16(%%edi), %%mm7 \n\t"                  "pandn     %%mm7, %%mm4     \n\t"                  "por       %%mm4, %%mm6     \n\t"                  "movq      %%mm6, 16(%%edi) \n\t"                  "addl      $24, %%esi       \n\t" // inc by 24 bytes processed                  "addl      $24, %%edi       \n\t"                  "subl      $8, %%ecx        \n\t" // dec by 8 pixels processed                  "ja        mainloop24       \n\t"                "mainloop24end:               \n\t"// preload        "movl      diff, %%ecx      \n\t" // (diff is in eax)