mp3lib/decode_i586.c

   1 /*
   2  * Modified for use with MPlayer, for details see the changelog at
   3  * http://svn.mplayerhq.hu/mplayer/trunk/
   4  * $Id$
   5  */
   6
   7 /*
   8 * mpg123_synth_1to1 works the same way as the c version of this
   9 * file.  only two types of changes have been made:
  10 * - reordered floating point instructions to
  11 *   prevent pipline stalls
  12 * - made WRITE_SAMPLE use integer instead of
  13 *   (slower) floating point
  14 * all kinds of x86 processors should benefit from these
  15 * modifications.
  16 *
  17 * useful sources of information on optimizing x86 code include:
  18 *
  19 *     Intel Architecture Optimization Manual
  20 *     http://www.intel.com/design/pentium/manuals/242816.htm
  21 *
  22 *     Cyrix 6x86 Instruction Set Summary
  23 *     ftp://ftp.cyrix.com/6x86/6x-dbch6.pdf
  24 *
  25 *     AMD-K5 Processor Software Development
  26 *     http://www.amd.com/products/cpg/techdocs/appnotes/20007e.pdf
  27 *
  28 * Stefan Bieschewski <stb@acm.org>
  29 *
  30 * $Id$
  31 */
  32 #include "config.h"
  33 #include "mangle.h"
  34 #include "mpg123.h"
  35
  36 static int attribute_used buffs[1088]={0};
  37 static int attribute_used bo=1;
  38 static int attribute_used saved_ebp=0;
  39
  40 int synth_1to1_pent(real *bandPtr, int channel, short *samples)
  41 {
  42   real tmp[3];
  43   register int retval;
  44     __asm__ volatile(
  45 "        movl %%ebp,"MANGLE(saved_ebp)"\n\t"
  46 "        movl %1,%%eax\n\t"/*bandPtr*/
  47 "        movl %3,%%esi\n\t"
  48 "        xorl %%edi,%%edi\n\t"
  49 "        movl "MANGLE(bo)",%%ebp\n\t"
  50 "        cmpl %%edi,%2\n\t"
  51 "        jne .L48\n\t"
  52 "        decl %%ebp\n\t"
  53 "        andl $15,%%ebp\n\t"
  54 "        movl %%ebp,"MANGLE(bo)"\n\t"
  55 "        movl $"MANGLE(buffs)",%%ecx\n\t"
  56 "        jmp .L49\n\t"
  57 ".L48:\n\t"
  58 "        addl $2,%%esi\n\t"
  59 "        movl $"MANGLE(buffs)"+2176,%%ecx\n\t"
  60 ".L49:\n\t"
  61 "        testl $1,%%ebp\n\t"
  62 "        je .L50\n\t"
  63 "        movl %%ecx,%%ebx\n\t"
  64 "        movl %%ebp,%4\n\t"
  65 "        pushl %%eax\n\t"
  66 "        movl 4+%4,%%edx\n\t"
  67 "        leal (%%ebx,%%edx,4),%%eax\n\t"
  68 "        pushl %%eax\n\t"
  69 "        movl 8+%4,%%eax\n\t"
  70 "        incl %%eax\n\t"
  71 "        andl $15,%%eax\n\t"
  72 "        leal 1088(,%%eax,4),%%eax\n\t"
  73 "        addl %%ebx,%%eax\n\t"
  74 "        jmp .L74\n\t"
  75 ".L50:\n\t"
  76 "        leal 1088(%%ecx),%%ebx\n\t"
  77 "        leal 1(%%ebp),%%edx\n\t"
  78 "        movl %%edx,%4\n\t"
  79 "        pushl %%eax\n\t"
  80 "        leal 1092(%%ecx,%%ebp,4),%%eax\n\t"
  81 "        pushl %%eax\n\t"
  82 "        leal (%%ecx,%%ebp,4),%%eax\n\t"
  83 ".L74:\n\t"
  84 "        pushl %%eax\n\t"
  85 "        call "MANGLE(mp3lib_dct64)"\n\t"
  86 "        addl $12,%%esp\n\t"
  87 "        movl %4,%%edx\n\t"
  88 "        leal 0(,%%edx,4),%%edx\n\t"
  89 "        movl $"MANGLE(mp3lib_decwin)"+64,%%eax\n\t"
  90 "        movl %%eax,%%ecx\n\t"
  91 "        subl %%edx,%%ecx\n\t"
  92 "        movl $16,%%ebp\n\t"
  93 ".L55:\n\t"
  94 "        flds (%%ecx)\n\t"
  95 "        fmuls (%%ebx)\n\t"
  96 "        flds 4(%%ecx)\n\t"
  97 "        fmuls 4(%%ebx)\n\t"
  98 "        fxch %%st(1)\n\t"
  99 "        flds 8(%%ecx)\n\t"
 100 "        fmuls 8(%%ebx)\n\t"
 101 "        fxch %%st(2)\n\t"
 102 "        fsubrp %%st,%%st(1)\n\t"
 103 "        flds 12(%%ecx)\n\t"
 104 "        fmuls 12(%%ebx)\n\t"
 105 "        fxch %%st(2)\n\t"
 106 "        faddp %%st,%%st(1)\n\t"
 107 "        flds 16(%%ecx)\n\t"
 108 "        fmuls 16(%%ebx)\n\t"
 109 "        fxch %%st(2)\n\t"
 110 "        fsubrp %%st,%%st(1)\n\t"
 111 "        flds 20(%%ecx)\n\t"
 112 "        fmuls 20(%%ebx)\n\t"
 113 "        fxch %%st(2)\n\t"
 114 "        faddp %%st,%%st(1)\n\t"
 115 "        flds 24(%%ecx)\n\t"
 116 "        fmuls 24(%%ebx)\n\t"
 117 "        fxch %%st(2)\n\t"
 118 "        fsubrp %%st,%%st(1)\n\t"
 119 "        flds 28(%%ecx)\n\t"
 120 "        fmuls 28(%%ebx)\n\t"
 121 "        fxch %%st(2)\n\t"
 122 "        faddp %%st,%%st(1)\n\t"
 123 "        flds 32(%%ecx)\n\t"
 124 "        fmuls 32(%%ebx)\n\t"
 125 "        fxch %%st(2)\n\t"
 126 "        fsubrp %%st,%%st(1)\n\t"
 127 "        flds 36(%%ecx)\n\t"
 128 "        fmuls 36(%%ebx)\n\t"
 129 "        fxch %%st(2)\n\t"
 130 "        faddp %%st,%%st(1)\n\t"
 131 "        flds 40(%%ecx)\n\t"
 132 "        fmuls 40(%%ebx)\n\t"
 133 "        fxch %%st(2)\n\t"
 134 "        fsubrp %%st,%%st(1)\n\t"
 135 "        flds 44(%%ecx)\n\t"
 136 "        fmuls 44(%%ebx)\n\t"
 137 "        fxch %%st(2)\n\t"
 138 "        faddp %%st,%%st(1)\n\t"
 139 "        flds 48(%%ecx)\n\t"
 140 "        fmuls 48(%%ebx)\n\t"
 141 "        fxch %%st(2)\n\t"
 142 "        fsubrp %%st,%%st(1)\n\t"
 143 "        flds 52(%%ecx)\n\t"
 144 "        fmuls 52(%%ebx)\n\t"
 145 "        fxch %%st(2)\n\t"
 146 "        faddp %%st,%%st(1)\n\t"
 147 "        flds 56(%%ecx)\n\t"
 148 "        fmuls 56(%%ebx)\n\t"
 149 "        fxch %%st(2)\n\t"
 150 "        fsubrp %%st,%%st(1)\n\t"
 151 "        flds 60(%%ecx)\n\t"
 152 "        fmuls 60(%%ebx)\n\t"
 153 "        fxch %%st(2)\n\t"
 154 "        subl $4,%%esp\n\t"
 155 "        faddp %%st,%%st(1)\n\t"
 156 "        fxch %%st(1)\n\t"
 157 "        fsubrp %%st,%%st(1)\n\t"
 158 "        fistpl (%%esp)\n\t"
 159 "        popl %%eax\n\t"
 160 "        cmpl $32767,%%eax\n\t"
 161 "        jg 1f\n\t"
 162 "        cmpl $-32768,%%eax\n\t"
 163 "        jl 2f\n\t"
 164 "        movw %%ax,(%%esi)\n\t"
 165 "        jmp 4f\n\t"
 166 "1:      movw $32767,(%%esi)\n\t"
 167 "        jmp 3f\n\t"
 168 "2:      movw $-32768,(%%esi)\n\t"
 169 "3:      incl %%edi\n\t"
 170 "4:\n\t"
 171 ".L54:\n\t"
 172 "        addl $64,%%ebx\n\t"
 173 "        subl $-128,%%ecx\n\t"
 174 "        addl $4,%%esi\n\t"
 175 "        decl %%ebp\n\t"
 176 "        jnz .L55\n\t"
 177 "        flds (%%ecx)\n\t"
 178 "        fmuls (%%ebx)\n\t"
 179 "        flds 8(%%ecx)\n\t"
 180 "        fmuls 8(%%ebx)\n\t"
 181 "        flds 16(%%ecx)\n\t"
 182 "        fmuls 16(%%ebx)\n\t"
 183 "        fxch %%st(2)\n\t"
 184 "        faddp %%st,%%st(1)\n\t"
 185 "        flds 24(%%ecx)\n\t"
 186 "        fmuls 24(%%ebx)\n\t"
 187 "        fxch %%st(2)\n\t"
 188 "        faddp %%st,%%st(1)\n\t"
 189 "        flds 32(%%ecx)\n\t"
 190 "        fmuls 32(%%ebx)\n\t"
 191 "        fxch %%st(2)\n\t"
 192 "        faddp %%st,%%st(1)\n\t"
 193 "        flds 40(%%ecx)\n\t"
 194 "        fmuls 40(%%ebx)\n\t"
 195 "        fxch %%st(2)\n\t"
 196 "        faddp %%st,%%st(1)\n\t"
 197 "        flds 48(%%ecx)\n\t"
 198 "        fmuls 48(%%ebx)\n\t"
 199 "        fxch %%st(2)\n\t"
 200 "        faddp %%st,%%st(1)\n\t"
 201 "        flds 56(%%ecx)\n\t"
 202 "        fmuls 56(%%ebx)\n\t"
 203 "        fxch %%st(2)\n\t"
 204 "        subl $4,%%esp\n\t"
 205 "        faddp %%st,%%st(1)\n\t"
 206 "        fxch %%st(1)\n\t"
 207 "        faddp %%st,%%st(1)\n\t"
 208 "        fistpl (%%esp)\n\t"
 209 "        popl %%eax\n\t"
 210 "        cmpl $32767,%%eax\n\t"
 211 "        jg 1f\n\t"
 212 "        cmpl $-32768,%%eax\n\t"
 213 "        jl 2f\n\t"
 214 "        movw %%ax,(%%esi)\n\t"
 215 "        jmp 4f\n\t"
 216 "1:      movw $32767,(%%esi)\n\t"
 217 "        jmp 3f\n\t"
 218 "2:      movw $-32768,(%%esi)\n\t"
 219 "3:      incl %%edi\n\t"
 220 "4:\n\t"
 221 ".L62:\n\t"
 222 "        addl $-64,%%ebx\n\t"
 223 "        addl $4,%%esi\n\t"
 224 "        movl %4,%%edx\n\t"
 225 "        leal -128(%%ecx,%%edx,8),%%ecx\n\t"
 226 "        movl $15,%%ebp\n\t"
 227 ".L68:\n\t"
 228 "        flds -4(%%ecx)\n\t"
 229 "        fchs\n\t"
 230 "        fmuls (%%ebx)\n\t"
 231 "        flds -8(%%ecx)\n\t"
 232 "        fmuls 4(%%ebx)\n\t"
 233 "        fxch %%st(1)\n\t"
 234 "        flds -12(%%ecx)\n\t"
 235 "        fmuls 8(%%ebx)\n\t"
 236 "        fxch %%st(2)\n\t"
 237 "        fsubrp %%st,%%st(1)\n\t"
 238 "        flds -16(%%ecx)\n\t"
 239 "        fmuls 12(%%ebx)\n\t"
 240 "        fxch %%st(2)\n\t"
 241 "        fsubrp %%st,%%st(1)\n\t"
 242 "        flds -20(%%ecx)\n\t"
 243 "        fmuls 16(%%ebx)\n\t"
 244 "        fxch %%st(2)\n\t"
 245 "        fsubrp %%st,%%st(1)\n\t"
 246 "        flds -24(%%ecx)\n\t"
 247 "        fmuls 20(%%ebx)\n\t"
 248 "        fxch %%st(2)\n\t"
 249 "        fsubrp %%st,%%st(1)\n\t"
 250 "        flds -28(%%ecx)\n\t"
 251 "        fmuls 24(%%ebx)\n\t"
 252 "        fxch %%st(2)\n\t"
 253 "        fsubrp %%st,%%st(1)\n\t"
 254 "        flds -32(%%ecx)\n\t"
 255 "        fmuls 28(%%ebx)\n\t"
 256 "        fxch %%st(2)\n\t"
 257 "        fsubrp %%st,%%st(1)\n\t"
 258 "        flds -36(%%ecx)\n\t"
 259 "        fmuls 32(%%ebx)\n\t"
 260 "        fxch %%st(2)\n\t"
 261 "        fsubrp %%st,%%st(1)\n\t"
 262 "        flds -40(%%ecx)\n\t"
 263 "        fmuls 36(%%ebx)\n\t"
 264 "        fxch %%st(2)\n\t"
 265 "        fsubrp %%st,%%st(1)\n\t"
 266 "        flds -44(%%ecx)\n\t"
 267 "        fmuls 40(%%ebx)\n\t"
 268 "        fxch %%st(2)\n\t"
 269 "        fsubrp %%st,%%st(1)\n\t"
 270 "        flds -48(%%ecx)\n\t"
 271 "        fmuls 44(%%ebx)\n\t"
 272 "        fxch %%st(2)\n\t"
 273 "        fsubrp %%st,%%st(1)\n\t"
 274 "        flds -52(%%ecx)\n\t"
 275 "        fmuls 48(%%ebx)\n\t"
 276 "        fxch %%st(2)\n\t"
 277 "        fsubrp %%st,%%st(1)\n\t"
 278 "        flds -56(%%ecx)\n\t"
 279 "        fmuls 52(%%ebx)\n\t"
 280 "        fxch %%st(2)\n\t"
 281 "        fsubrp %%st,%%st(1)\n\t"
 282 "        flds -60(%%ecx)\n\t"
 283 "        fmuls 56(%%ebx)\n\t"
 284 "        fxch %%st(2)\n\t"
 285 "        fsubrp %%st,%%st(1)\n\t"
 286 "        flds (%%ecx)\n\t"
 287 "        fmuls 60(%%ebx)\n\t"
 288 "        fxch %%st(2)\n\t"
 289 "        subl $4,%%esp\n\t"
 290 "        fsubrp %%st,%%st(1)\n\t"
 291 "        fxch %%st(1)\n\t"
 292 "        fsubrp %%st,%%st(1)\n\t"
 293 "        fistpl (%%esp)\n\t"
 294 "        popl %%eax\n\t"
 295 "        cmpl $32767,%%eax\n\t"
 296 "        jg 1f\n\t"
 297 "        cmpl $-32768,%%eax\n\t"
 298 "        jl 2f\n\t"
 299 "        movw %%ax,(%%esi)\n\t"
 300 "        jmp 4f\n\t"
 301 "1:      movw $32767,(%%esi)\n\t"
 302 "        jmp 3f\n\t"
 303 "2:      movw $-32768,(%%esi)\n\t"
 304 "3:      incl %%edi\n\t"
 305 "4:\n\t"
 306 ".L67:\n\t"
 307 "        addl $-64,%%ebx\n\t"
 308 "        addl $-128,%%ecx\n\t"
 309 "        addl $4,%%esi\n\t"
 310 "        decl %%ebp\n\t"
 311 "        jnz .L68\n\t"
 312 "        movl %%edi,%%eax\n\t"
 313 "        movl "MANGLE(saved_ebp)",%%ebp\n\t"
 314         :"=a"(retval)
 315         :"m"(bandPtr),"m"(channel),"m"(samples),"m"(tmp[0])
 316         :"memory","%edi","%esi","%ebx","%ecx","%edx");
 317   return retval;
 318 }