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_points2_general ) 38848b8605SmrgHIDDEN(_mesa_3dnow_transform_points2_general) 39848b8605SmrgGLNAME( _mesa_3dnow_transform_points2_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(V4F_START, 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 MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 63848b8605Smrg PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */ 64848b8605Smrg 65848b8605Smrg MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */ 66848b8605Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */ 67848b8605Smrg 68848b8605Smrg MOVD ( REGOFF(8, ECX), MM2 ) /* | m02 */ 69848b8605Smrg PUNPCKLDQ ( REGOFF(24, ECX), MM2 ) /* m12 | m02 */ 70848b8605Smrg 71848b8605Smrg MOVD ( REGOFF(12, ECX), MM3 ) /* | m03 */ 72848b8605Smrg PUNPCKLDQ ( REGOFF(28, ECX), MM3 ) /* m13 | m03 */ 73848b8605Smrg 74848b8605Smrg MOVQ ( REGOFF(48, ECX), MM4 ) /* m31 | m30 */ 75848b8605Smrg MOVQ ( REGOFF(56, ECX), MM5 ) /* m33 | m32 */ 76848b8605Smrg 77848b8605SmrgALIGNTEXT16 78848b8605SmrgLLBL( G3TPGR_2 ): 79848b8605Smrg 80848b8605Smrg MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */ 81848b8605Smrg MOVQ ( MM6, MM7 ) /* x1 | x0 */ 82848b8605Smrg 83848b8605Smrg PFMUL ( MM0, MM6 ) /* x1*m10 | x0*m00 */ 84848b8605Smrg PFMUL ( MM1, MM7 ) /* x1*m11 | x0*m01 */ 85848b8605Smrg 86848b8605Smrg PFACC ( MM7, MM6 ) /* x0*m01+x1*m11 | x0*x00+x1*m10 */ 87848b8605Smrg PFADD ( MM4, MM6 ) /* x0*...*m11+m31 | x0*...*m10+m30 */ 88848b8605Smrg 89848b8605Smrg MOVQ ( MM6, REGIND(EDX) ) /* write r1, r0 */ 90848b8605Smrg MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */ 91848b8605Smrg 92848b8605Smrg MOVQ ( MM6, MM7 ) /* x1 | x0 */ 93848b8605Smrg PFMUL ( MM2, MM6 ) /* x1*m12 | x0*m02 */ 94848b8605Smrg 95848b8605Smrg PFMUL ( MM3, MM7 ) /* x1*m13 | x0*m03 */ 96848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 97848b8605Smrg 98848b8605Smrg PFACC ( MM7, MM6 ) /* x0*m03+x1*m13 | x0*x02+x1*m12 */ 99848b8605Smrg PFADD ( MM5, MM6 ) /* x0*...*m13+m33 | x0*...*m12+m32 */ 100848b8605Smrg 101848b8605Smrg MOVQ ( MM6, REGOFF(8, EDX) ) /* write r3, r2 */ 102848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 103848b8605Smrg 104848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 105848b8605Smrg JNZ ( LLBL( G3TPGR_2 ) ) /* cnt > 0 ? -> process next vertex */ 106848b8605Smrg 107848b8605SmrgLLBL( G3TPGR_3 ): 108848b8605Smrg 109848b8605Smrg FEMMS 110848b8605Smrg POP_L ( EDI ) 111848b8605Smrg POP_L ( ESI ) 112848b8605Smrg RET 113848b8605Smrg 114848b8605Smrg 115848b8605Smrg 116848b8605Smrg 117848b8605SmrgALIGNTEXT16 118848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points2_perspective ) 119848b8605SmrgHIDDEN(_mesa_3dnow_transform_points2_perspective) 120848b8605SmrgGLNAME( _mesa_3dnow_transform_points2_perspective ): 121848b8605Smrg 122848b8605Smrg PUSH_L ( ESI ) 123848b8605Smrg 124848b8605Smrg MOV_L ( ARG_DEST, ECX ) 125848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 126848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 127848b8605Smrg MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) ) 128848b8605Smrg OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) ) 129848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 130848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 131848b8605Smrg 132848b8605Smrg PUSH_L ( EDI ) 133848b8605Smrg 134848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 135848b8605Smrg MOV_L ( ESI, ECX ) 136848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 137848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 138848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 139848b8605Smrg 140848b8605Smrg TEST_L ( ESI, ESI ) 141848b8605Smrg JZ ( LLBL( G3TPPR_3 ) ) 142848b8605Smrg 143848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 144848b8605Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */ 145848b8605Smrg 146848b8605Smrg MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */ 147848b8605Smrg 148848b8605SmrgALIGNTEXT16 149848b8605SmrgLLBL( G3TPPR_2 ): 150848b8605Smrg 151848b8605Smrg MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ 152848b8605Smrg PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */ 153848b8605Smrg 154848b8605Smrg MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */ 155848b8605Smrg MOVQ ( MM3, REGOFF(8, EDX) ) /* write r2 (=m32), r3 (=0) */ 156848b8605Smrg 157848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 158848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 159848b8605Smrg 160848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 161848b8605Smrg JNZ ( LLBL( G3TPPR_2 ) ) /* cnt > 0 ? -> process next vertex */ 162848b8605Smrg 163848b8605SmrgLLBL( G3TPPR_3 ): 164848b8605Smrg 165848b8605Smrg FEMMS 166848b8605Smrg POP_L ( EDI ) 167848b8605Smrg POP_L ( ESI ) 168848b8605Smrg RET 169848b8605Smrg 170848b8605Smrg 171848b8605Smrg 172848b8605Smrg 173848b8605SmrgALIGNTEXT16 174848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points2_3d ) 175848b8605SmrgHIDDEN(_mesa_3dnow_transform_points2_3d) 176848b8605SmrgGLNAME( _mesa_3dnow_transform_points2_3d ): 177848b8605Smrg 178848b8605Smrg PUSH_L ( ESI ) 179848b8605Smrg 180848b8605Smrg MOV_L ( ARG_DEST, ECX ) 181848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 182848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 183848b8605Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 184848b8605Smrg OR_B ( CONST(VEC_SIZE_3 ), REGOFF(V4F_FLAGS, ECX) ) 185848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 186848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 187848b8605Smrg 188848b8605Smrg PUSH_L ( EDI ) 189848b8605Smrg 190848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 191848b8605Smrg MOV_L ( ESI, ECX ) 192848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 193848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 194848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 195848b8605Smrg 196848b8605Smrg TEST_L ( ESI, ESI ) 197848b8605Smrg JZ ( LLBL( G3TP3R_3 ) ) 198848b8605Smrg 199848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 200848b8605Smrg PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */ 201848b8605Smrg 202848b8605Smrg MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */ 203848b8605Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */ 204848b8605Smrg 205848b8605Smrg MOVD ( REGOFF(8, ECX), MM2 ) /* | m02 */ 206848b8605Smrg PUNPCKLDQ ( REGOFF(24, ECX), MM2 ) /* m12 | m02 */ 207848b8605Smrg 208848b8605Smrg MOVQ ( REGOFF(48, ECX), MM4 ) /* m31 | m30 */ 209848b8605Smrg MOVD ( REGOFF(56, ECX), MM5 ) /* | m32 */ 210848b8605Smrg 211848b8605SmrgALIGNTEXT16 212848b8605SmrgLLBL( G3TP3R_2 ): 213848b8605Smrg 214848b8605Smrg MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */ 215848b8605Smrg MOVQ ( MM6, MM7 ) /* x1 | x0 */ 216848b8605Smrg 217848b8605Smrg PFMUL ( MM0, MM6 ) /* x1*m10 | x0*m00 */ 218848b8605Smrg PFMUL ( MM1, MM7 ) /* x1*m11 | x0*m01 */ 219848b8605Smrg 220848b8605Smrg PFACC ( MM7, MM6 ) /* x0*m01+x1*m11 | x0*x00+x1*m10 */ 221848b8605Smrg PFADD ( MM4, MM6 ) /* x0*...*m11+m31 | x0*...*m10+m30 */ 222848b8605Smrg 223848b8605Smrg MOVQ ( MM6, REGIND(EDX) ) /* write r1, r0 */ 224848b8605Smrg MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */ 225848b8605Smrg 226848b8605Smrg MOVQ ( MM6, MM7 ) /* x1 | x0 */ 227848b8605Smrg PFMUL ( MM2, MM6 ) /* x1*m12 | x0*m02 */ 228848b8605Smrg 229848b8605Smrg PFACC ( MM7, MM6 ) /* ***trash*** | x0*x02+x1*m12 */ 230848b8605Smrg PFADD ( MM5, MM6 ) /* ***trash*** | x0*...*m12+m32 */ 231848b8605Smrg 232848b8605Smrg MOVD ( MM6, REGOFF(8, EDX) ) /* write r2 */ 233848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 234848b8605Smrg 235848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 236848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 237848b8605Smrg 238848b8605Smrg JNZ ( LLBL( G3TP3R_2 ) ) /* cnt > 0 ? -> process next vertex */ 239848b8605Smrg 240848b8605SmrgLLBL( G3TP3R_3 ): 241848b8605Smrg 242848b8605Smrg FEMMS 243848b8605Smrg POP_L ( EDI ) 244848b8605Smrg POP_L ( ESI ) 245848b8605Smrg RET 246848b8605Smrg 247848b8605Smrg 248848b8605Smrg 249848b8605Smrg 250848b8605SmrgALIGNTEXT16 251848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points2_3d_no_rot ) 252848b8605SmrgHIDDEN(_mesa_3dnow_transform_points2_3d_no_rot) 253848b8605SmrgGLNAME( _mesa_3dnow_transform_points2_3d_no_rot ): 254848b8605Smrg 255848b8605Smrg PUSH_L ( ESI ) 256848b8605Smrg 257848b8605Smrg MOV_L ( ARG_DEST, ECX ) 258848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 259848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 260848b8605Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 261848b8605Smrg OR_B ( CONST(VEC_SIZE_3 ), REGOFF(V4F_FLAGS, ECX) ) 262848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 263848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 264848b8605Smrg 265848b8605Smrg PUSH_L ( EDI ) 266848b8605Smrg 267848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 268848b8605Smrg MOV_L ( ESI, ECX ) 269848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 270848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 271848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 272848b8605Smrg 273848b8605Smrg TEST_L ( ESI, ESI ) 274848b8605Smrg JZ ( LLBL( G3TP3NRR_3 ) ) 275848b8605Smrg 276848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 277848b8605Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */ 278848b8605Smrg 279848b8605Smrg MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */ 280848b8605Smrg MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */ 281848b8605Smrg 282848b8605SmrgALIGNTEXT16 283848b8605SmrgLLBL( G3TP3NRR_2 ): 284848b8605Smrg 285848b8605Smrg MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ 286848b8605Smrg PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */ 287848b8605Smrg 288848b8605Smrg PFADD ( MM2, MM4 ) /* x1*m11+m31 | x0*m00+m30 */ 289848b8605Smrg MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */ 290848b8605Smrg 291848b8605Smrg MOVD ( MM3, REGOFF(8, EDX) ) /* write r2 */ 292848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 293848b8605Smrg 294848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 295848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 296848b8605Smrg 297848b8605Smrg JNZ ( LLBL( G3TP3NRR_2 ) ) /* cnt > 0 ? -> process next vertex */ 298848b8605Smrg 299848b8605SmrgLLBL( G3TP3NRR_3 ): 300848b8605Smrg 301848b8605Smrg FEMMS 302848b8605Smrg POP_L ( EDI ) 303848b8605Smrg POP_L ( ESI ) 304848b8605Smrg RET 305848b8605Smrg 306848b8605Smrg 307848b8605Smrg 308848b8605Smrg 309848b8605SmrgALIGNTEXT16 310848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points2_2d ) 311848b8605SmrgHIDDEN(_mesa_3dnow_transform_points2_2d) 312848b8605SmrgGLNAME( _mesa_3dnow_transform_points2_2d ): 313848b8605Smrg 314848b8605Smrg PUSH_L ( ESI ) 315848b8605Smrg 316848b8605Smrg MOV_L ( ARG_DEST, ECX ) 317848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 318848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 319848b8605Smrg MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) ) 320848b8605Smrg OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) ) 321848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 322848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 323848b8605Smrg 324848b8605Smrg PUSH_L ( EDI ) 325848b8605Smrg 326848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 327848b8605Smrg MOV_L ( ESI, ECX ) 328848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 329848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 330848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 331848b8605Smrg 332848b8605Smrg TEST_L ( ESI, ESI ) 333848b8605Smrg JZ ( LLBL( G3TP2R_3 ) ) 334848b8605Smrg 335848b8605Smrg MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */ 336848b8605Smrg MOVQ ( REGOFF(16, ECX), MM1 ) /* m11 | m10 */ 337848b8605Smrg 338848b8605Smrg MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */ 339848b8605Smrg 340848b8605SmrgALIGNTEXT16 341848b8605SmrgLLBL( G3TP2R_2 ): 342848b8605Smrg 343848b8605Smrg MOVD ( REGIND(EAX), MM4 ) /* | x0 */ 344848b8605Smrg MOVD ( REGOFF(4, EAX), MM5 ) /* | x1 */ 345848b8605Smrg 346848b8605Smrg PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */ 347848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 348848b8605Smrg 349848b8605Smrg PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */ 350848b8605Smrg PUNPCKLDQ ( MM5, MM5 ) /* x1 | x1 */ 351848b8605Smrg 352848b8605Smrg PFMUL ( MM1, MM5 ) /* x1*m11 | x1*m10 */ 353848b8605Smrg PFADD ( MM2, MM4 ) /* x...x1*m11+31 | x0*..*m10+m30 */ 354848b8605Smrg 355848b8605Smrg PFADD ( MM5, MM4 ) /* x0*m01+x1*m11 | x0*m00+x1*m10 */ 356848b8605Smrg MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */ 357848b8605Smrg 358848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 359848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 360848b8605Smrg 361848b8605Smrg JNZ ( LLBL( G3TP2R_2 ) ) /* cnt > 0 ? -> process next vertex */ 362848b8605Smrg 363848b8605SmrgLLBL( G3TP2R_3 ): 364848b8605Smrg 365848b8605Smrg FEMMS 366848b8605Smrg POP_L ( EDI ) 367848b8605Smrg POP_L ( ESI ) 368848b8605Smrg RET 369848b8605Smrg 370848b8605Smrg 371848b8605Smrg 372848b8605Smrg 373848b8605SmrgALIGNTEXT16 374848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points2_2d_no_rot ) 375848b8605SmrgHIDDEN(_mesa_3dnow_transform_points2_2d_no_rot) 376848b8605SmrgGLNAME( _mesa_3dnow_transform_points2_2d_no_rot ): 377848b8605Smrg 378848b8605Smrg PUSH_L ( ESI ) 379848b8605Smrg 380848b8605Smrg MOV_L ( ARG_DEST, ECX ) 381848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 382848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 383848b8605Smrg MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) ) 384848b8605Smrg OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) ) 385848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 386848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 387848b8605Smrg 388848b8605Smrg PUSH_L ( EDI ) 389848b8605Smrg 390848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 391848b8605Smrg MOV_L ( ESI, ECX ) 392848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 393848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 394848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 395848b8605Smrg 396848b8605Smrg TEST_L ( ESI, ESI ) 397848b8605Smrg JZ ( LLBL( G3TP2NRR_3 ) ) 398848b8605Smrg 399848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 400848b8605Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */ 401848b8605Smrg 402848b8605Smrg MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */ 403848b8605Smrg 404848b8605SmrgALIGNTEXT16 405848b8605SmrgLLBL( G3TP2NRR_2 ): 406848b8605Smrg 407848b8605Smrg MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ 408848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 409848b8605Smrg 410848b8605Smrg PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */ 411848b8605Smrg PFADD ( MM2, MM4 ) /* m31 | x0*m00+m30 */ 412848b8605Smrg 413848b8605Smrg MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */ 414848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 415848b8605Smrg 416848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 417848b8605Smrg JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */ 418848b8605Smrg 419848b8605SmrgLLBL( G3TP2NRR_3 ): 420848b8605Smrg 421848b8605Smrg FEMMS 422848b8605Smrg POP_L ( EDI ) 423848b8605Smrg POP_L ( ESI ) 424848b8605Smrg RET 425848b8605Smrg 426848b8605Smrg 427848b8605Smrg 428848b8605Smrg 429848b8605SmrgALIGNTEXT16 430848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points2_identity ) 431848b8605SmrgHIDDEN(_mesa_3dnow_transform_points2_identity) 432848b8605SmrgGLNAME( _mesa_3dnow_transform_points2_identity ): 433848b8605Smrg 434848b8605Smrg PUSH_L ( ESI ) 435848b8605Smrg 436848b8605Smrg MOV_L ( ARG_DEST, ECX ) 437848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 438848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 439848b8605Smrg MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) ) 440848b8605Smrg OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) ) 441848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 442848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 443848b8605Smrg 444848b8605Smrg PUSH_L ( EDI ) 445848b8605Smrg 446848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 447848b8605Smrg MOV_L ( ESI, ECX ) 448848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 449848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 450848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 451848b8605Smrg 452848b8605Smrg TEST_L ( ESI, ESI ) 453848b8605Smrg JZ ( LLBL( G3TPIR_3 ) ) 454848b8605Smrg 455848b8605SmrgALIGNTEXT16 456848b8605SmrgLLBL( G3TPIR_3 ): 457848b8605Smrg 458848b8605Smrg MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ 459848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 460848b8605Smrg 461848b8605Smrg MOVQ ( MM0, REGIND(EDX) ) /* r1 | r0 */ 462848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 463848b8605Smrg 464848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 465848b8605Smrg JNZ ( LLBL( G3TPIR_3 ) ) /* cnt > 0 ? -> process next vertex */ 466848b8605Smrg 467848b8605SmrgLLBL( G3TPIR_4 ): 468848b8605Smrg 469848b8605Smrg FEMMS 470848b8605Smrg POP_L ( EDI ) 471848b8605Smrg POP_L ( ESI ) 472848b8605Smrg RET 473848b8605Smrg#endif 474848b8605Smrg 475848b8605Smrg#if defined (__ELF__) && defined (__linux__) 476848b8605Smrg .section .note.GNU-stack,"",%progbits 477848b8605Smrg#endif 478