1848b8605Smrg 2848b8605Smrg/* 3848b8605Smrg * Mesa 3-D graphics library 4848b8605Smrg * 5848b8605Smrg * Copyright (C) 1999-2001 Brian Paul All Rights Reserved. 6848b8605Smrg * 7848b8605Smrg * Permission is hereby granted, free of charge, to any person obtaining a 8848b8605Smrg * copy of this software and associated documentation files (the "Software"), 9848b8605Smrg * to deal in the Software without restriction, including without limitation 10848b8605Smrg * the rights to use, copy, modify, merge, publish, distribute, sublicense, 11848b8605Smrg * and/or sell copies of the Software, and to permit persons to whom the 12848b8605Smrg * Software is furnished to do so, subject to the following conditions: 13848b8605Smrg * 14848b8605Smrg * The above copyright notice and this permission notice shall be included 15848b8605Smrg * in all copies or substantial portions of the Software. 16848b8605Smrg * 17848b8605Smrg * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 18848b8605Smrg * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 19848b8605Smrg * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 20848b8605Smrg * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR 21848b8605Smrg * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 22848b8605Smrg * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 23848b8605Smrg * OTHER DEALINGS IN THE SOFTWARE. 24848b8605Smrg */ 25848b8605Smrg 26848b8605Smrg#ifdef USE_3DNOW_ASM 27848b8605Smrg#include "assyntax.h" 28848b8605Smrg#include "matypes.h" 29848b8605Smrg#include "xform_args.h" 30848b8605Smrg 31848b8605Smrg SEG_TEXT 32848b8605Smrg 33848b8605Smrg#define FRAME_OFFSET 4 34848b8605Smrg 35848b8605Smrg 36848b8605SmrgALIGNTEXT16 37848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points1_general ) 38848b8605SmrgHIDDEN(_mesa_3dnow_transform_points1_general) 39848b8605SmrgGLNAME( _mesa_3dnow_transform_points1_general ): 40848b8605Smrg 41848b8605Smrg PUSH_L ( ESI ) 42848b8605Smrg 43848b8605Smrg MOV_L ( ARG_DEST, ECX ) 44848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 45848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 46848b8605Smrg MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) ) 47848b8605Smrg OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) ) 48848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 49848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 50848b8605Smrg 51848b8605Smrg PUSH_L ( EDI ) 52848b8605Smrg 53848b8605Smrg MOV_L ( REGOFF(4, ECX), EDX ) 54848b8605Smrg MOV_L ( ESI, ECX ) 55848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 56848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 57848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 58848b8605Smrg 59848b8605Smrg TEST_L ( ESI, ESI ) 60848b8605Smrg JZ ( LLBL( G3TPGR_3 ) ) 61848b8605Smrg 62848b8605Smrg MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */ 63848b8605Smrg MOVQ ( REGOFF(8, ECX), MM1 ) /* m03 | m02 */ 64848b8605Smrg 65848b8605Smrg MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */ 66848b8605Smrg MOVQ ( REGOFF(56, ECX), MM3 ) /* m33 | m32 */ 67848b8605Smrg 68848b8605SmrgALIGNTEXT16 69848b8605SmrgLLBL( G3TPGR_2 ): 70848b8605Smrg 71848b8605Smrg MOVD ( REGIND(EAX), MM4 ) /* | x0 */ 72848b8605Smrg PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */ 73848b8605Smrg 74848b8605Smrg MOVQ ( MM4, MM5 ) /* x0 | x0 */ 75848b8605Smrg PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */ 76848b8605Smrg 77848b8605Smrg PFMUL ( MM1, MM5 ) /* x0*m03 | x0*m02 */ 78848b8605Smrg PFADD ( MM2, MM4 ) /* x0*m01+m31 | x0*m00+m30 */ 79848b8605Smrg 80848b8605Smrg PFADD ( MM3, MM5 ) /* x0*m03+m33 | x0*m02+m32 */ 81848b8605Smrg MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */ 82848b8605Smrg 83848b8605Smrg MOVQ ( MM5, REGOFF(8, EDX) ) /* write r3, r2 */ 84848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 85848b8605Smrg 86848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 87848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 88848b8605Smrg 89848b8605Smrg JNZ ( LLBL( G3TPGR_2 ) ) /* cnt > 0 ? -> process next vertex */ 90848b8605Smrg 91848b8605SmrgLLBL( G3TPGR_3 ): 92848b8605Smrg 93848b8605Smrg FEMMS 94848b8605Smrg POP_L ( EDI ) 95848b8605Smrg POP_L ( ESI ) 96848b8605Smrg RET 97848b8605Smrg 98848b8605Smrg 99848b8605Smrg 100848b8605Smrg 101848b8605SmrgALIGNTEXT16 102848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points1_identity ) 103848b8605SmrgHIDDEN(_mesa_3dnow_transform_points1_identity) 104848b8605SmrgGLNAME( _mesa_3dnow_transform_points1_identity ): 105848b8605Smrg 106848b8605Smrg PUSH_L ( ESI ) 107848b8605Smrg 108848b8605Smrg MOV_L ( ARG_DEST, ECX ) 109848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 110848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 111848b8605Smrg MOV_L ( CONST(1), REGOFF(V4F_SIZE, ECX) ) 112848b8605Smrg OR_B ( CONST(VEC_SIZE_1), REGOFF(V4F_FLAGS, ECX) ) 113848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 114848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 115848b8605Smrg 116848b8605Smrg PUSH_L ( EDI ) 117848b8605Smrg 118848b8605Smrg MOV_L ( REGOFF(4, ECX), EDX ) 119848b8605Smrg MOV_L ( ESI, ECX ) 120848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 121848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 122848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 123848b8605Smrg 124848b8605Smrg TEST_L ( ESI, ESI ) 125848b8605Smrg JZ ( LLBL( G3TPIR_4) ) 126848b8605Smrg 127848b8605SmrgALIGNTEXT16 128848b8605SmrgLLBL( G3TPIR_3 ): 129848b8605Smrg 130848b8605Smrg MOVD ( REGIND(EAX), MM0 ) /* | x0 */ 131848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 132848b8605Smrg 133848b8605Smrg MOVD ( MM0, REGIND(EDX) ) /* | r0 */ 134848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 135848b8605Smrg 136848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 137848b8605Smrg JNZ ( LLBL( G3TPIR_3 ) ) /* cnt > 0 ? -> process next vertex */ 138848b8605Smrg 139848b8605SmrgLLBL( G3TPIR_4 ): 140848b8605Smrg 141848b8605Smrg FEMMS 142848b8605Smrg POP_L ( EDI ) 143848b8605Smrg POP_L ( ESI ) 144848b8605Smrg RET 145848b8605Smrg 146848b8605Smrg 147848b8605Smrg 148848b8605Smrg 149848b8605SmrgALIGNTEXT16 150848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points1_3d_no_rot ) 151848b8605SmrgHIDDEN(_mesa_3dnow_transform_points1_3d_no_rot) 152848b8605SmrgGLNAME( _mesa_3dnow_transform_points1_3d_no_rot ): 153848b8605Smrg 154848b8605Smrg PUSH_L ( ESI ) 155848b8605Smrg 156848b8605Smrg MOV_L ( ARG_DEST, ECX ) 157848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 158848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 159848b8605Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 160848b8605Smrg OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) ) 161848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 162848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 163848b8605Smrg 164848b8605Smrg PUSH_L ( EDI ) 165848b8605Smrg 166848b8605Smrg MOV_L ( REGOFF(4, ECX), EDX ) 167848b8605Smrg MOV_L ( ESI, ECX ) 168848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 169848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 170848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 171848b8605Smrg 172848b8605Smrg TEST_L ( ESI, ESI ) 173848b8605Smrg JZ ( LLBL( G3TP3NRR_3 ) ) 174848b8605Smrg 175848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 176848b8605Smrg MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */ 177848b8605Smrg 178848b8605Smrg MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */ 179848b8605Smrg 180848b8605SmrgALIGNTEXT16 181848b8605SmrgLLBL( G3TP3NRR_2 ): 182848b8605Smrg 183848b8605Smrg MOVD ( REGIND(EAX), MM4 ) /* | x0 */ 184848b8605Smrg PFMUL ( MM0, MM4 ) /* | x0*m00 */ 185848b8605Smrg 186848b8605Smrg PFADD ( MM2, MM4 ) /* m31 | x0*m00+m30 */ 187848b8605Smrg MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */ 188848b8605Smrg 189848b8605Smrg MOVD ( MM3, REGOFF(8, EDX) ) /* write r2 */ 190848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 191848b8605Smrg 192848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 193848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 194848b8605Smrg 195848b8605Smrg JNZ ( LLBL( G3TP3NRR_2 ) ) /* cnt > 0 ? -> process next vertex */ 196848b8605Smrg 197848b8605SmrgLLBL( G3TP3NRR_3 ): 198848b8605Smrg 199848b8605Smrg FEMMS 200848b8605Smrg POP_L ( EDI ) 201848b8605Smrg POP_L ( ESI ) 202848b8605Smrg RET 203848b8605Smrg 204848b8605Smrg 205848b8605Smrg 206848b8605Smrg 207848b8605SmrgALIGNTEXT16 208848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points1_perspective ) 209848b8605SmrgHIDDEN(_mesa_3dnow_transform_points1_perspective) 210848b8605SmrgGLNAME( _mesa_3dnow_transform_points1_perspective ): 211848b8605Smrg 212848b8605Smrg PUSH_L ( ESI ) 213848b8605Smrg 214848b8605Smrg MOV_L ( ARG_DEST, ECX ) 215848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 216848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 217848b8605Smrg MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) ) 218848b8605Smrg OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) ) 219848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 220848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 221848b8605Smrg 222848b8605Smrg PUSH_L ( EDI ) 223848b8605Smrg 224848b8605Smrg MOV_L ( REGOFF(4, ECX), EDX ) 225848b8605Smrg MOV_L ( ESI, ECX ) 226848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 227848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 228848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 229848b8605Smrg 230848b8605Smrg TEST_L ( ESI, ESI ) 231848b8605Smrg JZ ( LLBL( G3TPPR_3 ) ) 232848b8605Smrg 233848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 234848b8605Smrg MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */ 235848b8605Smrg 236848b8605SmrgALIGNTEXT16 237848b8605SmrgLLBL( G3TPPR_2 ): 238848b8605Smrg 239848b8605Smrg MOVD ( REGIND(EAX), MM4 ) /* 0 | x0 */ 240848b8605Smrg PFMUL ( MM0, MM4 ) /* 0 | x0*m00 */ 241848b8605Smrg 242848b8605Smrg MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */ 243848b8605Smrg MOVQ ( MM3, REGOFF(8, EDX) ) /* write r2 (=m32), r3 (=0) */ 244848b8605Smrg 245848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 246848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 247848b8605Smrg 248848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 249848b8605Smrg JNZ ( LLBL( G3TPPR_2 ) ) /* cnt > 0 ? -> process next vertex */ 250848b8605Smrg 251848b8605SmrgLLBL( G3TPPR_3 ): 252848b8605Smrg 253848b8605Smrg FEMMS 254848b8605Smrg POP_L ( EDI ) 255848b8605Smrg POP_L ( ESI ) 256848b8605Smrg RET 257848b8605Smrg 258848b8605Smrg 259848b8605Smrg 260848b8605Smrg 261848b8605SmrgALIGNTEXT16 262848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points1_2d ) 263848b8605SmrgHIDDEN(_mesa_3dnow_transform_points1_2d) 264848b8605SmrgGLNAME( _mesa_3dnow_transform_points1_2d ): 265848b8605Smrg 266848b8605Smrg PUSH_L ( ESI ) 267848b8605Smrg 268848b8605Smrg MOV_L ( ARG_DEST, ECX ) 269848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 270848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 271848b8605Smrg MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) ) 272848b8605Smrg OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) ) 273848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 274848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 275848b8605Smrg 276848b8605Smrg PUSH_L ( EDI ) 277848b8605Smrg 278848b8605Smrg MOV_L ( REGOFF(4, ECX), EDX ) 279848b8605Smrg MOV_L ( ESI, ECX ) 280848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 281848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 282848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 283848b8605Smrg 284848b8605Smrg TEST_L ( ESI, ESI ) 285848b8605Smrg JZ ( LLBL( G3TP2R_3 ) ) 286848b8605Smrg 287848b8605Smrg MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */ 288848b8605Smrg MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */ 289848b8605Smrg 290848b8605SmrgALIGNTEXT16 291848b8605SmrgLLBL( G3TP2R_2 ): 292848b8605Smrg 293848b8605Smrg MOVD ( REGIND(EAX), MM4 ) /* | x0 */ 294848b8605Smrg PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */ 295848b8605Smrg 296848b8605Smrg PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */ 297848b8605Smrg PFADD ( MM2, MM4 ) /* x0*m01+m31 | x0*m00+m30 */ 298848b8605Smrg 299848b8605Smrg MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */ 300848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 301848b8605Smrg 302848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 303848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 304848b8605Smrg 305848b8605Smrg JNZ ( LLBL( G3TP2R_2 ) ) /* cnt > 0 ? -> process next vertex */ 306848b8605Smrg 307848b8605SmrgLLBL( G3TP2R_3 ): 308848b8605Smrg 309848b8605Smrg FEMMS 310848b8605Smrg POP_L ( EDI ) 311848b8605Smrg POP_L ( ESI ) 312848b8605Smrg RET 313848b8605Smrg 314848b8605Smrg 315848b8605Smrg 316848b8605Smrg 317848b8605SmrgALIGNTEXT16 318848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points1_2d_no_rot ) 319848b8605SmrgHIDDEN(_mesa_3dnow_transform_points1_2d_no_rot) 320848b8605SmrgGLNAME( _mesa_3dnow_transform_points1_2d_no_rot ): 321848b8605Smrg 322848b8605Smrg PUSH_L ( ESI ) 323848b8605Smrg 324848b8605Smrg MOV_L ( ARG_DEST, ECX ) 325848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 326848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 327848b8605Smrg MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) ) 328848b8605Smrg OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) ) 329848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 330848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 331848b8605Smrg 332848b8605Smrg PUSH_L ( EDI ) 333848b8605Smrg 334848b8605Smrg MOV_L ( REGOFF(4, ECX), EDX ) 335848b8605Smrg MOV_L ( ESI, ECX ) 336848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 337848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 338848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 339848b8605Smrg 340848b8605Smrg TEST_L ( ESI, ESI ) 341848b8605Smrg JZ ( LLBL( G3TP2NRR_3 ) ) 342848b8605Smrg 343848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 344848b8605Smrg MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */ 345848b8605Smrg 346848b8605SmrgALIGNTEXT16 347848b8605SmrgLLBL( G3TP2NRR_2 ): 348848b8605Smrg 349848b8605Smrg MOVD ( REGIND(EAX), MM4 ) /* | x0 */ 350848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 351848b8605Smrg 352848b8605Smrg PFMUL ( MM0, MM4 ) /* | x0*m00 */ 353848b8605Smrg PFADD ( MM2, MM4 ) /* m31 | x0*m00+m30 */ 354848b8605Smrg 355848b8605Smrg MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */ 356848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 357848b8605Smrg 358848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 359848b8605Smrg JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */ 360848b8605Smrg 361848b8605SmrgLLBL( G3TP2NRR_3 ): 362848b8605Smrg 363848b8605Smrg FEMMS 364848b8605Smrg POP_L ( EDI ) 365848b8605Smrg POP_L ( ESI ) 366848b8605Smrg RET 367848b8605Smrg 368848b8605Smrg 369848b8605Smrg 370848b8605Smrg 371848b8605SmrgALIGNTEXT16 372848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points1_3d ) 373848b8605SmrgHIDDEN(_mesa_3dnow_transform_points1_3d) 374848b8605SmrgGLNAME( _mesa_3dnow_transform_points1_3d ): 375848b8605Smrg 376848b8605Smrg PUSH_L ( ESI ) 377848b8605Smrg 378848b8605Smrg MOV_L ( ARG_DEST, ECX ) 379848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 380848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 381848b8605Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 382848b8605Smrg OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) ) 383848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 384848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 385848b8605Smrg 386848b8605Smrg PUSH_L ( EDI ) 387848b8605Smrg 388848b8605Smrg MOV_L ( REGOFF(4, ECX), EDX ) 389848b8605Smrg MOV_L ( ESI, ECX ) 390848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 391848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 392848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 393848b8605Smrg 394848b8605Smrg TEST_L ( ESI, ESI ) 395848b8605Smrg JZ ( LLBL( G3TP3R_3 ) ) 396848b8605Smrg 397848b8605Smrg MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */ 398848b8605Smrg MOVD ( REGOFF(8, ECX), MM1 ) /* | m02 */ 399848b8605Smrg 400848b8605Smrg MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */ 401848b8605Smrg MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */ 402848b8605Smrg 403848b8605SmrgALIGNTEXT16 404848b8605SmrgLLBL( G3TP3R_2 ): 405848b8605Smrg 406848b8605Smrg MOVD ( REGIND(EAX), MM4 ) /* | x0 */ 407848b8605Smrg PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */ 408848b8605Smrg 409848b8605Smrg MOVQ ( MM4, MM5 ) /* | x0 */ 410848b8605Smrg PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */ 411848b8605Smrg 412848b8605Smrg PFMUL ( MM1, MM5 ) /* | x0*m02 */ 413848b8605Smrg PFADD ( MM2, MM4 ) /* x0*m01+m31 | x0*m00+m30 */ 414848b8605Smrg 415848b8605Smrg PFADD ( MM3, MM5 ) /* | x0*m02+m32 */ 416848b8605Smrg MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */ 417848b8605Smrg 418848b8605Smrg MOVD ( MM5, REGOFF(8, EDX) ) /* write r2 */ 419848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 420848b8605Smrg 421848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 422848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 423848b8605Smrg 424848b8605Smrg JNZ ( LLBL( G3TP3R_2 ) ) /* cnt > 0 ? -> process next vertex */ 425848b8605Smrg 426848b8605SmrgLLBL( G3TP3R_3 ): 427848b8605Smrg 428848b8605Smrg FEMMS 429848b8605Smrg POP_L ( EDI ) 430848b8605Smrg POP_L ( ESI ) 431848b8605Smrg RET 432848b8605Smrg 433848b8605Smrg#endif 434848b8605Smrg 435848b8605Smrg#if defined (__ELF__) && defined (__linux__) 436848b8605Smrg .section .note.GNU-stack,"",%progbits 437848b8605Smrg#endif 438