sse_xform3.S revision c1f859d4
1 2/* 3 * Mesa 3-D graphics library 4 * Version: 3.5 5 * 6 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved. 7 * 8 * Permission is hereby granted, free of charge, to any person obtaining a 9 * copy of this software and associated documentation files (the "Software"), 10 * to deal in the Software without restriction, including without limitation 11 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 12 * and/or sell copies of the Software, and to permit persons to whom the 13 * Software is furnished to do so, subject to the following conditions: 14 * 15 * The above copyright notice and this permission notice shall be included 16 * in all copies or substantial portions of the Software. 17 * 18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 21 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN 22 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 23 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 24 */ 25 26/** TODO: 27 * - insert PREFETCH instructions to avoid cache-misses ! 28 * - some more optimizations are possible... 29 * - for 40-50% more performance in the SSE-functions, the 30 * data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned ! 31 */ 32 33#ifdef USE_SSE_ASM 34#include "assyntax.h" 35#include "matypes.h" 36#include "xform_args.h" 37 38 SEG_TEXT 39 40#define S(i) REGOFF(i * 4, ESI) 41#define D(i) REGOFF(i * 4, EDI) 42#define M(i) REGOFF(i * 4, EDX) 43 44 45ALIGNTEXT4 46GLOBL GLNAME(_mesa_sse_transform_points3_general) 47HIDDEN(_mesa_sse_transform_points3_general) 48GLNAME( _mesa_sse_transform_points3_general ): 49 50#define FRAME_OFFSET 8 51 PUSH_L ( ESI ) 52 PUSH_L ( EDI ) 53 54 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 55 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 56 57 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 58 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 59 60 CMP_L ( CONST(0), ECX ) /* count == 0 ? */ 61 JE ( LLBL(K_GTPGR_finish) ) /* yes -> nothing to do. */ 62 63 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 64 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 65 66 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 67 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 68 69 SHL_L( CONST(4), ECX ) /* count *= 16 */ 70 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 71 72 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 73 ADD_L( EDI, ECX ) /* count += dest ptr */ 74 75 76ALIGNTEXT32 77 MOVAPS ( REGOFF(0, EDX), XMM0 ) /* m0 | m1 | m2 | m3 */ 78 MOVAPS ( REGOFF(16, EDX), XMM1 ) /* m4 | m5 | m6 | m7 */ 79 MOVAPS ( REGOFF(32, EDX), XMM2 ) /* m8 | m9 | m10 | m11 */ 80 MOVAPS ( REGOFF(48, EDX), XMM3 ) /* m12 | m13 | m14 | m15 */ 81 82 83ALIGNTEXT32 84LLBL(K_GTPGR_top): 85 MOVSS ( REGOFF(0, ESI), XMM4 ) /* | | | ox */ 86 SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox | ox */ 87 MOVSS ( REGOFF(4, ESI), XMM5 ) /* | | | oy */ 88 SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy | oy */ 89 MOVSS ( REGOFF(8, ESI), XMM6 ) /* | | | oz */ 90 SHUFPS ( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz | oz */ 91 92 MULPS ( XMM0, XMM4 ) /* m3*ox | m2*ox | m1*ox | m0*ox */ 93 MULPS ( XMM1, XMM5 ) /* m7*oy | m6*oy | m5*oy | m4*oy */ 94 MULPS ( XMM2, XMM6 ) /* m11*oz | m10*oz | m9*oz | m8*oz */ 95 96 ADDPS ( XMM5, XMM4 ) 97 ADDPS ( XMM6, XMM4 ) 98 ADDPS ( XMM3, XMM4 ) 99 100 MOVAPS ( XMM4, REGOFF(0, EDI) ) 101 102LLBL(K_GTPGR_skip): 103 ADD_L ( CONST(16), EDI ) 104 ADD_L ( EAX, ESI ) 105 CMP_L ( ECX, EDI ) 106 JNE ( LLBL(K_GTPGR_top) ) 107 108LLBL(K_GTPGR_finish): 109 POP_L ( EDI ) 110 POP_L ( ESI ) 111 RET 112#undef FRAME_OFFSET 113 114 115ALIGNTEXT4 116GLOBL GLNAME(_mesa_sse_transform_points3_identity) 117HIDDEN(_mesa_sse_transform_points3_identity) 118GLNAME( _mesa_sse_transform_points3_identity ): 119 120#define FRAME_OFFSET 8 121 PUSH_L ( ESI ) 122 PUSH_L ( EDI ) 123 124 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 125 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 126 127 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 128 129 TEST_L( ECX, ECX) 130 JZ( LLBL(K_GTPIR_finish) ) /* count was zero; go to finish */ 131 132 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 133 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 134 135 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 136 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 137 138 SHL_L( CONST(4), ECX ) /* count *= 16 */ 139 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 140 141 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 142 ADD_L( EDI, ECX ) /* count += dest ptr */ 143 144 CMP_L( ESI, EDI ) 145 JE( LLBL(K_GTPIR_finish) ) 146 147 148ALIGNTEXT32 149LLBL(K_GTPIR_top): 150 MOVLPS ( S(0), XMM0 ) 151 MOVLPS ( XMM0, D(0) ) 152 MOVSS ( S(2), XMM0 ) 153 MOVSS ( XMM0, D(2) ) 154 155LLBL(K_GTPIR_skip): 156 ADD_L ( CONST(16), EDI ) 157 ADD_L ( EAX, ESI ) 158 CMP_L ( ECX, EDI ) 159 JNE ( LLBL(K_GTPIR_top) ) 160 161LLBL(K_GTPIR_finish): 162 POP_L ( EDI ) 163 POP_L ( ESI ) 164 RET 165#undef FRAME_OFFSET 166 167 168 169 170ALIGNTEXT4 171GLOBL GLNAME(_mesa_sse_transform_points3_3d_no_rot) 172HIDDEN(_mesa_sse_transform_points3_3d_no_rot) 173GLNAME(_mesa_sse_transform_points3_3d_no_rot): 174 175#define FRAME_OFFSET 8 176 PUSH_L( ESI ) 177 PUSH_L( EDI ) 178 179 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 180 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 181 182 183 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 184 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 185 186 TEST_L( ECX, ECX) 187 JZ( LLBL(K_GTP3DNRR_finish) ) /* count was zero; go to finish */ 188 189 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 190 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 191 192 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 193 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 194 195 SHL_L( CONST(4), ECX ) /* count *= 16 */ 196 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 197 198 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 199 ADD_L( EDI, ECX ) /* count += dest ptr */ 200 201 202ALIGNTEXT32 203 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 204 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 205 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 206 MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */ 207 MOVSS ( M(10), XMM3 ) /* - | - | - | m10 */ 208 MOVSS ( M(14), XMM4 ) /* - | - | - | m14 */ 209 210ALIGNTEXT32 211LLBL(K_GTP3DNRR_top): 212 213 MOVLPS ( S(0), XMM0 ) /* - | - | s1 | s0 */ 214 MULPS ( XMM1, XMM0 ) /* - | - | s1*m5 | s0*m0 */ 215 ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */ 216 MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */ 217 218 MOVSS ( S(2), XMM0 ) /* sz */ 219 MULSS ( XMM3, XMM0 ) /* sz*m10 */ 220 ADDSS ( XMM4, XMM0 ) /* +m14 */ 221 MOVSS ( XMM0, D(2) ) /* -> D(2) */ 222 223LLBL(K_GTP3DNRR_skip): 224 ADD_L ( CONST(16), EDI ) 225 ADD_L ( EAX, ESI ) 226 CMP_L ( ECX, EDI ) 227 JNE ( LLBL(K_GTP3DNRR_top) ) 228 229LLBL(K_GTP3DNRR_finish): 230 POP_L ( EDI ) 231 POP_L ( ESI ) 232 RET 233#undef FRAME_OFFSET 234 235 236 237ALIGNTEXT4 238GLOBL GLNAME(_mesa_sse_transform_points3_perspective) 239HIDDEN(_mesa_sse_transform_points3_perspective) 240GLNAME(_mesa_sse_transform_points3_perspective): 241 242#define FRAME_OFFSET 8 243 PUSH_L ( ESI ) 244 PUSH_L ( EDI ) 245 246 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 247 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 248 249 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 250 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 251 252 TEST_L( ECX, ECX) 253 JZ( LLBL(K_GTP3PR_finish) ) /* count was zero; go to finish */ 254 255 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 256 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 257 258 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 259 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 260 261 SHL_L( CONST(4), ECX ) /* count *= 16 */ 262 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 263 264 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 265 ADD_L( EDI, ECX ) /* count += dest ptr */ 266 267ALIGNTEXT32 268 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 269 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 270 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 271 MOVLPS ( M(8), XMM2 ) /* - | - | m9 | m8 */ 272 MOVSS ( M(10), XMM3 ) /* m10 */ 273 MOVSS ( M(14), XMM4 ) /* m14 */ 274 XORPS ( XMM6, XMM6 ) /* 0 */ 275 276ALIGNTEXT32 277LLBL(K_GTP3PR_top): 278 MOVLPS ( S(0), XMM0 ) /* oy | ox */ 279 MULPS ( XMM1, XMM0 ) /* oy*m5 | ox*m0 */ 280 MOVSS ( S(2), XMM5 ) /* oz */ 281 SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oz | oz */ 282 MULPS ( XMM2, XMM5 ) /* oz*m9 | oz*m8 */ 283 ADDPS ( XMM5, XMM0 ) /* +oy*m5 | +ox*m0 */ 284 MOVLPS ( XMM0, D(0) ) /* ->D(1) | ->D(0) */ 285 286 MOVSS ( S(2), XMM0 ) /* oz */ 287 MULSS ( XMM3, XMM0 ) /* oz*m10 */ 288 ADDSS ( XMM4, XMM0 ) /* +m14 */ 289 MOVSS ( XMM0, D(2) ) /* ->D(2) */ 290 291 MOVSS ( S(2), XMM0 ) /* oz */ 292 MOVSS ( XMM6, XMM5 ) /* 0 */ 293 SUBPS ( XMM0, XMM5 ) /* -oz */ 294 MOVSS ( XMM5, D(3) ) /* ->D(3) */ 295 296LLBL(K_GTP3PR_skip): 297 ADD_L( CONST(16), EDI ) 298 ADD_L( EAX, ESI ) 299 CMP_L( ECX, EDI ) 300 JNE( LLBL(K_GTP3PR_top) ) 301 302LLBL(K_GTP3PR_finish): 303 POP_L ( EDI ) 304 POP_L ( ESI ) 305 RET 306#undef FRAME_OFFSET 307 308 309 310ALIGNTEXT4 311GLOBL GLNAME(_mesa_sse_transform_points3_2d) 312HIDDEN(_mesa_sse_transform_points3_2d) 313GLNAME(_mesa_sse_transform_points3_2d): 314 315#define FRAME_OFFSET 8 316 PUSH_L( ESI ) 317 PUSH_L( EDI ) 318 319 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 320 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 321 322 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 323 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 324 325 TEST_L( ECX, ECX) 326 JZ( LLBL(K_GTP3P2DR_finish) ) /* count was zero; go to finish */ 327 328 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 329 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 330 331 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 332 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 333 334 SHL_L( CONST(4), ECX ) /* count *= 16 */ 335 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 336 337 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 338 ADD_L( EDI, ECX ) /* count += dest ptr */ 339 340ALIGNTEXT32 341 MOVLPS( M(0), XMM0 ) /* m1 | m0 */ 342 MOVLPS( M(4), XMM1 ) /* m5 | m4 */ 343 MOVLPS( M(12), XMM2 ) /* m13 | m12 */ 344 345ALIGNTEXT32 346LLBL(K_GTP3P2DR_top): 347 MOVSS ( S(0), XMM3 ) /* ox */ 348 SHUFPS ( CONST(0x0), XMM3, XMM3 ) /* ox | ox */ 349 MULPS ( XMM0, XMM3 ) /* ox*m1 | ox*m0 */ 350 MOVSS ( S(1), XMM4 ) /* oy */ 351 SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* oy | oy */ 352 MULPS ( XMM1, XMM4 ) /* oy*m5 | oy*m4 */ 353 354 ADDPS ( XMM4, XMM3 ) 355 ADDPS ( XMM2, XMM3 ) 356 MOVLPS ( XMM3, D(0) ) 357 358 MOVSS ( S(2), XMM3 ) 359 MOVSS ( XMM3, D(2) ) 360 361LLBL(K_GTP3P2DR_skip): 362 ADD_L ( CONST(16), EDI ) 363 ADD_L ( EAX, ESI ) 364 CMP_L ( ECX, EDI ) 365 JNE ( LLBL(K_GTP3P2DR_top) ) 366 367LLBL(K_GTP3P2DR_finish): 368 POP_L ( EDI ) 369 POP_L ( ESI ) 370 RET 371#undef FRAME_OFFSET 372 373 374 375ALIGNTEXT4 376GLOBL GLNAME(_mesa_sse_transform_points3_2d_no_rot) 377HIDDEN(_mesa_sse_transform_points3_2d_no_rot) 378GLNAME(_mesa_sse_transform_points3_2d_no_rot): 379 380#define FRAME_OFFSET 8 381 PUSH_L( ESI ) 382 PUSH_L( EDI ) 383 384 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 385 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 386 387 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 388 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 389 390 TEST_L( ECX, ECX) 391 JZ( LLBL(K_GTP3P2DNRR_finish) ) /* count was zero; go to finish */ 392 393 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 394 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 395 396 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 397 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 398 399 SHL_L( CONST(4), ECX ) /* count *= 16 */ 400 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 401 402 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 403 ADD_L( EDI, ECX ) /* count += dest ptr */ 404 405ALIGNTEXT32 406 MOVSS ( M(0), XMM1 ) /* m0 */ 407 MOVSS ( M(5), XMM2 ) /* m5 */ 408 UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */ 409 MOVLPS ( M(12), XMM2 ) /* m13 | m12 */ 410 411ALIGNTEXT32 412LLBL(K_GTP3P2DNRR_top): 413 MOVLPS( S(0), XMM0 ) /* oy | ox */ 414 MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */ 415 ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */ 416 MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */ 417 418 MOVSS( S(2), XMM0 ) 419 MOVSS( XMM0, D(2) ) 420 421LLBL(K_GTP3P2DNRR_skip): 422 ADD_L( CONST(16), EDI ) 423 ADD_L( EAX, ESI ) 424 CMP_L( ECX, EDI ) 425 JNE( LLBL(K_GTP3P2DNRR_top) ) 426 427LLBL(K_GTP3P2DNRR_finish): 428 POP_L( EDI ) 429 POP_L( ESI ) 430 RET 431#undef FRAME_OFFSET 432 433 434 435 436ALIGNTEXT4 437GLOBL GLNAME(_mesa_sse_transform_points3_3d) 438HIDDEN(_mesa_sse_transform_points3_3d) 439GLNAME(_mesa_sse_transform_points3_3d): 440 441#define FRAME_OFFSET 8 442 PUSH_L( ESI ) 443 PUSH_L( EDI ) 444 445 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 446 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 447 448 449 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 450 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 451 452 TEST_L( ECX, ECX) 453 JZ( LLBL(K_GTP3P3DR_finish) ) /* count was zero; go to finish */ 454 455 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 456 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 457 458 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 459 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 460 461 SHL_L( CONST(4), ECX ) /* count *= 16 */ 462 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 463 464 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 465 ADD_L( EDI, ECX ) /* count += dest ptr */ 466 467 468ALIGNTEXT32 469 MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */ 470 MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */ 471 MOVAPS( M(8), XMM2 ) /* m10 | m9 | m8 */ 472 MOVAPS( M(12), XMM3 ) /* m14 | m13 | m12 */ 473 474ALIGNTEXT32 475LLBL(K_GTP3P3DR_top): 476 MOVSS( S(0), XMM4 ) 477 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox */ 478 MULPS( XMM0, XMM4 ) /* ox*m2 | ox*m1 | ox*m0 */ 479 480 MOVSS( S(1), XMM5 ) 481 SHUFPS( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy */ 482 MULPS( XMM1, XMM5 ) /* oy*m6 | oy*m5 | oy*m4 */ 483 484 MOVSS( S(2), XMM6 ) 485 SHUFPS( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz */ 486 MULPS( XMM2, XMM6 ) /* oz*m10 | oz*m9 | oz*m8 */ 487 488 ADDPS( XMM5, XMM4 ) /* + | + | + */ 489 ADDPS( XMM6, XMM4 ) /* + | + | + */ 490 ADDPS( XMM3, XMM4 ) /* + | + | + */ 491 492 MOVLPS( XMM4, D(0) ) /* => D(1) | => D(0) */ 493 UNPCKHPS( XMM4, XMM4 ) 494 MOVSS( XMM4, D(2) ) 495 496LLBL(K_GTP3P3DR_skip): 497 ADD_L( CONST(16), EDI ) 498 ADD_L( EAX, ESI ) 499 CMP_L( ECX, EDI ) 500 JNE( LLBL(K_GTP3P3DR_top) ) 501 502LLBL(K_GTP3P3DR_finish): 503 POP_L( EDI ) 504 POP_L( ESI ) 505 RET 506#undef FRAME_OFFSET 507#endif 508 509#if defined (__ELF__) && defined (__linux__) 510 .section .note.GNU-stack,"",%progbits 511#endif 512