Mercurial > audlegacy
annotate Plugins/Input/mpg123/decode_i586.s @ 130:7dc049844b73 trunk
[svn] Fix libmikmod test.
| author | nenolod |
|---|---|
| date | Wed, 09 Nov 2005 19:57:44 -0800 |
| parents | 05d824e30afd |
| children |
| rev | line source |
|---|---|
|
127
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
1 # |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
2 # synth_1to1 works the same way as the c version of this |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
3 # file. only two types of changes have been made: |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
4 # - reordered floating point instructions to |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
5 # prevent pipline stalls |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
6 # - made WRITE_SAMPLE use integer instead of |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
7 # (slower) floating point |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
8 # all kinds of x86 processors should benefit from these |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
9 # modifications. |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
10 # |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
11 # useful sources of information on optimizing x86 code include: |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
12 # |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
13 # Intel Architecture Optimization Manual |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
14 # http:#/www.intel.com/design/pentium/manuals/242816.htm |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
15 # |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
16 # Cyrix 6x86 Instruction Set Summary |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
17 # ftp:#/ftp.cyrix.com/6x86/6x-dbch6.pdf |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
18 # |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
19 # AMD-K5 Processor Software Development |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
20 # http:#/www.amd.com/products/cpg/techdocs/appnotes/20007e.pdf |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
21 # |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
22 # Stefan Bieschewski <stb@acm.org> |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
23 # |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
24 # You can use this part under GPL. |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
25 # |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
26 # $Id: decode_i586.s,v 1.3 2000/10/25 11:05:23 hippm Exp $ |
|
05d824e30afd
[svn] Synchronize mpg123 code with mpg123-0.59s.mc3. This brings us superior accuracy to libMAD, apparently.
nenolod
parents:
61
diff
changeset
|
27 # |
| 61 | 28 .bss |
| 29 .comm buffs,4352,4 | |
| 30 .data | |
| 31 .align 4 | |
| 32 bo: | |
| 33 .long 1 | |
| 34 .section .rodata | |
| 35 .align 8 | |
| 36 .LC0: | |
| 37 .long 0x0,0x40dfffc0 | |
| 38 .align 8 | |
| 39 .LC1: | |
| 40 .long 0x0,0xc0e00000 | |
| 41 .align 8 | |
| 42 .text | |
| 43 .globl mpg123_synth_1to1_pent | |
| 44 mpg123_synth_1to1_pent: | |
| 45 subl $12,%esp | |
| 46 pushl %ebp | |
| 47 pushl %edi | |
| 48 pushl %esi | |
| 49 pushl %ebx | |
| 50 movl 32(%esp),%eax | |
| 51 movl 40(%esp),%esi | |
| 52 xorl %edi,%edi | |
| 53 movl bo,%ebp | |
| 54 cmpl %edi,36(%esp) | |
| 55 jne .L48 | |
| 56 decl %ebp | |
| 57 andl $15,%ebp | |
| 58 movl %ebp,bo | |
| 59 movl $buffs,%ecx | |
| 60 jmp .L49 | |
| 61 .L48: | |
| 62 addl $2,%esi | |
| 63 movl $buffs+2176,%ecx | |
| 64 .L49: | |
| 65 testl $1,%ebp | |
| 66 je .L50 | |
| 67 movl %ecx,%ebx | |
| 68 movl %ebp,16(%esp) | |
| 69 pushl %eax | |
| 70 movl 20(%esp),%edx | |
| 71 leal (%ebx,%edx,4),%eax | |
| 72 pushl %eax | |
| 73 movl 24(%esp),%eax | |
| 74 incl %eax | |
| 75 andl $15,%eax | |
| 76 leal 1088(,%eax,4),%eax | |
| 77 addl %ebx,%eax | |
| 78 jmp .L74 | |
| 79 .L50: | |
| 80 leal 1088(%ecx),%ebx | |
| 81 leal 1(%ebp),%edx | |
| 82 movl %edx,16(%esp) | |
| 83 pushl %eax | |
| 84 leal 1092(%ecx,%ebp,4),%eax | |
| 85 pushl %eax | |
| 86 leal (%ecx,%ebp,4),%eax | |
| 87 .L74: | |
| 88 pushl %eax | |
| 89 call mpg123_dct64 | |
| 90 addl $12,%esp | |
| 91 movl 16(%esp),%edx | |
| 92 leal 0(,%edx,4),%edx | |
| 93 movl $mpg123_decwin+64,%eax | |
| 94 movl %eax,%ecx | |
| 95 subl %edx,%ecx | |
| 96 movl $16,%ebp | |
| 97 .L55: | |
| 98 flds (%ecx) | |
| 99 fmuls (%ebx) | |
| 100 flds 4(%ecx) | |
| 101 fmuls 4(%ebx) | |
| 102 fxch %st(1) | |
| 103 flds 8(%ecx) | |
| 104 fmuls 8(%ebx) | |
| 105 fxch %st(2) | |
| 106 fsubrp %st,%st(1) | |
| 107 flds 12(%ecx) | |
| 108 fmuls 12(%ebx) | |
| 109 fxch %st(2) | |
| 110 faddp %st,%st(1) | |
| 111 flds 16(%ecx) | |
| 112 fmuls 16(%ebx) | |
| 113 fxch %st(2) | |
| 114 fsubrp %st,%st(1) | |
| 115 flds 20(%ecx) | |
| 116 fmuls 20(%ebx) | |
| 117 fxch %st(2) | |
| 118 faddp %st,%st(1) | |
| 119 flds 24(%ecx) | |
| 120 fmuls 24(%ebx) | |
| 121 fxch %st(2) | |
| 122 fsubrp %st,%st(1) | |
| 123 flds 28(%ecx) | |
| 124 fmuls 28(%ebx) | |
| 125 fxch %st(2) | |
| 126 faddp %st,%st(1) | |
| 127 flds 32(%ecx) | |
| 128 fmuls 32(%ebx) | |
| 129 fxch %st(2) | |
| 130 fsubrp %st,%st(1) | |
| 131 flds 36(%ecx) | |
| 132 fmuls 36(%ebx) | |
| 133 fxch %st(2) | |
| 134 faddp %st,%st(1) | |
| 135 flds 40(%ecx) | |
| 136 fmuls 40(%ebx) | |
| 137 fxch %st(2) | |
| 138 fsubrp %st,%st(1) | |
| 139 flds 44(%ecx) | |
| 140 fmuls 44(%ebx) | |
| 141 fxch %st(2) | |
| 142 faddp %st,%st(1) | |
| 143 flds 48(%ecx) | |
| 144 fmuls 48(%ebx) | |
| 145 fxch %st(2) | |
| 146 fsubrp %st,%st(1) | |
| 147 flds 52(%ecx) | |
| 148 fmuls 52(%ebx) | |
| 149 fxch %st(2) | |
| 150 faddp %st,%st(1) | |
| 151 flds 56(%ecx) | |
| 152 fmuls 56(%ebx) | |
| 153 fxch %st(2) | |
| 154 fsubrp %st,%st(1) | |
| 155 flds 60(%ecx) | |
| 156 fmuls 60(%ebx) | |
| 157 fxch %st(2) | |
| 158 subl $4,%esp | |
| 159 faddp %st,%st(1) | |
| 160 fxch %st(1) | |
| 161 fsubrp %st,%st(1) | |
| 162 fistpl (%esp) | |
| 163 popl %eax | |
| 164 cmpl $32767,%eax | |
| 165 jg 1f | |
| 166 cmpl $-32768,%eax | |
| 167 jl 2f | |
| 168 movw %ax,(%esi) | |
| 169 jmp 4f | |
| 170 1: movw $32767,(%esi) | |
| 171 jmp 3f | |
| 172 2: movw $-32768,(%esi) | |
| 173 3: incl %edi | |
| 174 4: | |
| 175 .L54: | |
| 176 addl $64,%ebx | |
| 177 subl $-128,%ecx | |
| 178 addl $4,%esi | |
| 179 decl %ebp | |
| 180 jnz .L55 | |
| 181 flds (%ecx) | |
| 182 fmuls (%ebx) | |
| 183 flds 8(%ecx) | |
| 184 fmuls 8(%ebx) | |
| 185 flds 16(%ecx) | |
| 186 fmuls 16(%ebx) | |
| 187 fxch %st(2) | |
| 188 faddp %st,%st(1) | |
| 189 flds 24(%ecx) | |
| 190 fmuls 24(%ebx) | |
| 191 fxch %st(2) | |
| 192 faddp %st,%st(1) | |
| 193 flds 32(%ecx) | |
| 194 fmuls 32(%ebx) | |
| 195 fxch %st(2) | |
| 196 faddp %st,%st(1) | |
| 197 flds 40(%ecx) | |
| 198 fmuls 40(%ebx) | |
| 199 fxch %st(2) | |
| 200 faddp %st,%st(1) | |
| 201 flds 48(%ecx) | |
| 202 fmuls 48(%ebx) | |
| 203 fxch %st(2) | |
| 204 faddp %st,%st(1) | |
| 205 flds 56(%ecx) | |
| 206 fmuls 56(%ebx) | |
| 207 fxch %st(2) | |
| 208 subl $4,%esp | |
| 209 faddp %st,%st(1) | |
| 210 fxch %st(1) | |
| 211 faddp %st,%st(1) | |
| 212 fistpl (%esp) | |
| 213 popl %eax | |
| 214 cmpl $32767,%eax | |
| 215 jg 1f | |
| 216 cmpl $-32768,%eax | |
| 217 jl 2f | |
| 218 movw %ax,(%esi) | |
| 219 jmp 4f | |
| 220 1: movw $32767,(%esi) | |
| 221 jmp 3f | |
| 222 2: movw $-32768,(%esi) | |
| 223 3: incl %edi | |
| 224 4: | |
| 225 .L62: | |
| 226 addl $-64,%ebx | |
| 227 addl $4,%esi | |
| 228 movl 16(%esp),%edx | |
| 229 leal -128(%ecx,%edx,8),%ecx | |
| 230 movl $15,%ebp | |
| 231 .L68: | |
| 232 flds -4(%ecx) | |
| 233 fchs | |
| 234 fmuls (%ebx) | |
| 235 flds -8(%ecx) | |
| 236 fmuls 4(%ebx) | |
| 237 fxch %st(1) | |
| 238 flds -12(%ecx) | |
| 239 fmuls 8(%ebx) | |
| 240 fxch %st(2) | |
| 241 fsubrp %st,%st(1) | |
| 242 flds -16(%ecx) | |
| 243 fmuls 12(%ebx) | |
| 244 fxch %st(2) | |
| 245 fsubrp %st,%st(1) | |
| 246 flds -20(%ecx) | |
| 247 fmuls 16(%ebx) | |
| 248 fxch %st(2) | |
| 249 fsubrp %st,%st(1) | |
| 250 flds -24(%ecx) | |
| 251 fmuls 20(%ebx) | |
| 252 fxch %st(2) | |
| 253 fsubrp %st,%st(1) | |
| 254 flds -28(%ecx) | |
| 255 fmuls 24(%ebx) | |
| 256 fxch %st(2) | |
| 257 fsubrp %st,%st(1) | |
| 258 flds -32(%ecx) | |
| 259 fmuls 28(%ebx) | |
| 260 fxch %st(2) | |
| 261 fsubrp %st,%st(1) | |
| 262 flds -36(%ecx) | |
| 263 fmuls 32(%ebx) | |
| 264 fxch %st(2) | |
| 265 fsubrp %st,%st(1) | |
| 266 flds -40(%ecx) | |
| 267 fmuls 36(%ebx) | |
| 268 fxch %st(2) | |
| 269 fsubrp %st,%st(1) | |
| 270 flds -44(%ecx) | |
| 271 fmuls 40(%ebx) | |
| 272 fxch %st(2) | |
| 273 fsubrp %st,%st(1) | |
| 274 flds -48(%ecx) | |
| 275 fmuls 44(%ebx) | |
| 276 fxch %st(2) | |
| 277 fsubrp %st,%st(1) | |
| 278 flds -52(%ecx) | |
| 279 fmuls 48(%ebx) | |
| 280 fxch %st(2) | |
| 281 fsubrp %st,%st(1) | |
| 282 flds -56(%ecx) | |
| 283 fmuls 52(%ebx) | |
| 284 fxch %st(2) | |
| 285 fsubrp %st,%st(1) | |
| 286 flds -60(%ecx) | |
| 287 fmuls 56(%ebx) | |
| 288 fxch %st(2) | |
| 289 fsubrp %st,%st(1) | |
| 290 flds (%ecx) | |
| 291 fmuls 60(%ebx) | |
| 292 fxch %st(2) | |
| 293 subl $4,%esp | |
| 294 fsubrp %st,%st(1) | |
| 295 fxch %st(1) | |
| 296 fsubrp %st,%st(1) | |
| 297 fistpl (%esp) | |
| 298 popl %eax | |
| 299 cmpl $32767,%eax | |
| 300 jg 1f | |
| 301 cmpl $-32768,%eax | |
| 302 jl 2f | |
| 303 movw %ax,(%esi) | |
| 304 jmp 4f | |
| 305 1: movw $32767,(%esi) | |
| 306 jmp 3f | |
| 307 2: movw $-32768,(%esi) | |
| 308 3: incl %edi | |
| 309 4: | |
| 310 .L67: | |
| 311 addl $-64,%ebx | |
| 312 addl $-128,%ecx | |
| 313 addl $4,%esi | |
| 314 decl %ebp | |
| 315 jnz .L68 | |
| 316 movl %edi,%eax | |
| 317 popl %ebx | |
| 318 popl %esi | |
| 319 popl %edi | |
| 320 popl %ebp | |
| 321 addl $12,%esp | |
| 322 ret | |
| 323 |