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_points3_general ) 38848b8605SmrgHIDDEN(_mesa_3dnow_transform_points3_general) 39848b8605SmrgGLNAME( _mesa_3dnow_transform_points3_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_2 ) ) 61848b8605Smrg 62848b8605Smrg PREFETCHW ( REGIND(EDX) ) 63848b8605Smrg 64848b8605SmrgALIGNTEXT16 65848b8605SmrgLLBL( G3TPGR_1 ): 66848b8605Smrg 67848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 68848b8605Smrg 69848b8605Smrg MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ 70848b8605Smrg MOVD ( REGOFF(8, EAX), MM2 ) /* | x2 */ 71848b8605Smrg 72848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 73848b8605Smrg PREFETCH ( REGIND(EAX) ) 74848b8605Smrg 75848b8605Smrg MOVQ ( MM0, MM1 ) /* x1 | x0 */ 76848b8605Smrg PUNPCKLDQ ( MM2, MM2 ) /* x2 | x2 */ 77848b8605Smrg 78848b8605Smrg PUNPCKLDQ ( MM0, MM0 ) /* x0 | x0 */ 79848b8605Smrg MOVQ ( MM2, MM5 ) /* x2 | x2 */ 80848b8605Smrg 81848b8605Smrg PUNPCKHDQ ( MM1, MM1 ) /* x1 | x1 */ 82848b8605Smrg PFMUL ( REGOFF(32, ECX), MM2 ) /* x2*m9 | x2*m8 */ 83848b8605Smrg 84848b8605Smrg MOVQ ( MM0, MM3 ) /* x0 | x0 */ 85848b8605Smrg PFMUL ( REGOFF(40, ECX), MM5 ) /* x2*m11 | x2*m10 */ 86848b8605Smrg 87848b8605Smrg MOVQ ( MM1, MM4 ) /* x1 | x1 */ 88848b8605Smrg PFMUL ( REGIND(ECX), MM0 ) /* x0*m1 | x0*m0 */ 89848b8605Smrg 90848b8605Smrg PFADD ( REGOFF(48, ECX), MM2 ) /* x2*m9+m13 | x2*m8+m12 */ 91848b8605Smrg PFMUL ( REGOFF(16, ECX), MM1 ) /* x1*m5 | x1*m4 */ 92848b8605Smrg 93848b8605Smrg PFADD ( REGOFF(56, ECX), MM5 ) /* x2*m11+m15 | x2*m10+m14 */ 94848b8605Smrg PFADD ( MM0, MM1 ) /* x0*m1+x1*m5 | x0*m0+x1*m4 */ 95848b8605Smrg 96848b8605Smrg PFMUL ( REGOFF(8, ECX), MM3 ) /* x0*m3 | x0*m2 */ 97848b8605Smrg PFADD ( MM1, MM2 ) /* r1 | r0 */ 98848b8605Smrg 99848b8605Smrg PFMUL ( REGOFF(24, ECX), MM4 ) /* x1*m7 | x1*m6 */ 100848b8605Smrg ADD_L ( CONST(16), EDX ) /* next output vertex */ 101848b8605Smrg 102848b8605Smrg PFADD ( MM3, MM4 ) /* x0*m3+x1*m7 | x0*m2+x1*m6 */ 103848b8605Smrg MOVQ ( MM2, REGOFF(-16, EDX) ) /* write r0, r1 */ 104848b8605Smrg 105848b8605Smrg PFADD ( MM4, MM5 ) /* r3 | r2 */ 106848b8605Smrg MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */ 107848b8605Smrg 108848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 109848b8605Smrg JNZ ( LLBL( G3TPGR_1 ) ) /* cnt > 0 ? -> process next vertex */ 110848b8605Smrg 111848b8605SmrgLLBL( G3TPGR_2 ): 112848b8605Smrg 113848b8605Smrg FEMMS 114848b8605Smrg POP_L ( EDI ) 115848b8605Smrg POP_L ( ESI ) 116848b8605Smrg RET 117848b8605Smrg 118848b8605Smrg 119848b8605Smrg 120848b8605Smrg 121848b8605SmrgALIGNTEXT16 122848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points3_perspective ) 123848b8605SmrgHIDDEN(_mesa_3dnow_transform_points3_perspective) 124848b8605SmrgGLNAME( _mesa_3dnow_transform_points3_perspective ): 125848b8605Smrg 126848b8605Smrg PUSH_L ( ESI ) 127848b8605Smrg 128848b8605Smrg MOV_L ( ARG_DEST, ECX ) 129848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 130848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 131848b8605Smrg MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) ) 132848b8605Smrg OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) ) 133848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 134848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 135848b8605Smrg 136848b8605Smrg PUSH_L ( EDI ) 137848b8605Smrg 138848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 139848b8605Smrg MOV_L ( ESI, ECX ) 140848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 141848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 142848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 143848b8605Smrg 144848b8605Smrg TEST_L ( ESI, ESI ) 145848b8605Smrg JZ ( LLBL( G3TPPR_2 ) ) 146848b8605Smrg 147848b8605Smrg PREFETCH ( REGIND(EAX) ) 148848b8605Smrg PREFETCHW ( REGIND(EDX) ) 149848b8605Smrg 150848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 151848b8605Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */ 152848b8605Smrg 153848b8605Smrg MOVQ ( REGOFF(32, ECX), MM1 ) /* m21 | m20 */ 154848b8605Smrg MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */ 155848b8605Smrg 156848b8605Smrg MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */ 157848b8605Smrg 158848b8605SmrgALIGNTEXT16 159848b8605SmrgLLBL( G3TPPR_1 ): 160848b8605Smrg 161848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 162848b8605Smrg 163848b8605Smrg MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */ 164848b8605Smrg MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ 165848b8605Smrg 166848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 167848b8605Smrg PREFETCH ( REGIND(EAX) ) 168848b8605Smrg 169848b8605Smrg PXOR ( MM7, MM7 ) /* 0 | 0 */ 170848b8605Smrg MOVQ ( MM5, MM6 ) /* | x2 */ 171848b8605Smrg 172848b8605Smrg PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */ 173848b8605Smrg PFSUB ( MM5, MM7 ) /* | -x2 */ 174848b8605Smrg 175848b8605Smrg PFMUL ( MM2, MM6 ) /* | x2*m22 */ 176848b8605Smrg PUNPCKLDQ ( MM5, MM5 ) /* x2 | x2 */ 177848b8605Smrg 178848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 179848b8605Smrg PFMUL ( MM1, MM5 ) /* x2*m21 | x2*m20 */ 180848b8605Smrg 181848b8605Smrg PFADD ( MM3, MM6 ) /* | x2*m22+m32 */ 182848b8605Smrg PFADD ( MM4, MM5 ) /* x1*m11+x2*m21 | x0*m00+x2*m20 */ 183848b8605Smrg 184848b8605Smrg MOVQ ( MM5, REGOFF(-16, EDX) ) /* write r0, r1 */ 185848b8605Smrg MOVD ( MM6, REGOFF(-8, EDX) ) /* write r2 */ 186848b8605Smrg 187848b8605Smrg MOVD ( MM7, REGOFF(-4, EDX) ) /* write r3 */ 188848b8605Smrg 189848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 190848b8605Smrg JNZ ( LLBL( G3TPPR_1 ) ) /* cnt > 0 ? -> process next vertex */ 191848b8605Smrg 192848b8605SmrgLLBL( G3TPPR_2 ): 193848b8605Smrg 194848b8605Smrg FEMMS 195848b8605Smrg POP_L ( EDI ) 196848b8605Smrg POP_L ( ESI ) 197848b8605Smrg RET 198848b8605Smrg 199848b8605Smrg 200848b8605Smrg 201848b8605Smrg 202848b8605SmrgALIGNTEXT16 203848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points3_3d ) 204848b8605SmrgHIDDEN(_mesa_3dnow_transform_points3_3d) 205848b8605SmrgGLNAME( _mesa_3dnow_transform_points3_3d ): 206848b8605Smrg 207848b8605Smrg PUSH_L ( ESI ) 208848b8605Smrg 209848b8605Smrg MOV_L ( ARG_DEST, ECX ) 210848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 211848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 212848b8605Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 213848b8605Smrg OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) ) 214848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 215848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 216848b8605Smrg 217848b8605Smrg PUSH_L ( EDI ) 218848b8605Smrg 219848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 220848b8605Smrg MOV_L ( ESI, ECX ) 221848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 222848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 223848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 224848b8605Smrg 225848b8605Smrg TEST_L ( ESI, ESI ) 226848b8605Smrg JZ ( LLBL( G3TP3R_2 ) ) 227848b8605Smrg 228848b8605Smrg PREFETCH ( REGIND(EAX) ) 229848b8605Smrg PREFETCH ( REGIND(EDX) ) 230848b8605Smrg 231848b8605Smrg MOVD ( REGOFF(8, ECX), MM7 ) /* | m2 */ 232848b8605Smrg PUNPCKLDQ ( REGOFF(24, ECX), MM7 ) /* m6 | m2 */ 233848b8605Smrg 234848b8605Smrg 235848b8605SmrgALIGNTEXT16 236848b8605SmrgLLBL( G3TP3R_1 ): 237848b8605Smrg 238848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 239848b8605Smrg 240848b8605Smrg MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ 241848b8605Smrg MOVD ( REGOFF(8, EAX), MM1 ) /* | x2 */ 242848b8605Smrg 243848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 244848b8605Smrg PREFETCH ( REGIND(EAX) ) 245848b8605Smrg 246848b8605Smrg MOVQ ( MM0, MM2 ) /* x1 | x0 */ 247848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 248848b8605Smrg 249848b8605Smrg PUNPCKLDQ ( MM2, MM2 ) /* x0 | x0 */ 250848b8605Smrg MOVQ ( MM0, MM3 ) /* x1 | x0 */ 251848b8605Smrg 252848b8605Smrg PFMUL ( REGIND(ECX), MM2 ) /* x0*m1 | x0*m0 */ 253848b8605Smrg PUNPCKHDQ ( MM3, MM3 ) /* x1 | x1 */ 254848b8605Smrg 255848b8605Smrg MOVQ ( MM1, MM4 ) /* | x2 */ 256848b8605Smrg PFMUL ( REGOFF(16, ECX), MM3 ) /* x1*m5 | x1*m4 */ 257848b8605Smrg 258848b8605Smrg PUNPCKLDQ ( MM4, MM4 ) /* x2 | x2 */ 259848b8605Smrg PFADD ( MM2, MM3 ) /* x0*m1+x1*m5 | x0*m0+x1*m4 */ 260848b8605Smrg 261848b8605Smrg PFMUL ( REGOFF(32, ECX), MM4 ) /* x2*m9 | x2*m8 */ 262848b8605Smrg PFADD ( REGOFF(48, ECX), MM3 ) /* x0*m1+...+m11 | x0*m0+x1*m4+m12 */ 263848b8605Smrg 264848b8605Smrg PFMUL ( MM7, MM0 ) /* x1*m6 | x0*m2 */ 265848b8605Smrg PFADD ( MM4, MM3 ) /* r1 | r0 */ 266848b8605Smrg 267848b8605Smrg PFMUL ( REGOFF(40, ECX), MM1 ) /* | x2*m10 */ 268848b8605Smrg PUNPCKLDQ ( REGOFF(56, ECX), MM1 ) /* m14 | x2*m10 */ 269848b8605Smrg 270848b8605Smrg PFACC ( MM0, MM1 ) 271848b8605Smrg 272848b8605Smrg MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */ 273848b8605Smrg PFACC ( MM1, MM1 ) /* | r2 */ 274848b8605Smrg 275848b8605Smrg MOVD ( MM1, REGOFF(-8, EDX) ) /* write r2 */ 276848b8605Smrg 277848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 278848b8605Smrg JNZ ( LLBL( G3TP3R_1 ) ) /* cnt > 0 ? -> process next vertex */ 279848b8605Smrg 280848b8605SmrgLLBL( G3TP3R_2 ): 281848b8605Smrg 282848b8605Smrg FEMMS 283848b8605Smrg POP_L ( EDI ) 284848b8605Smrg POP_L ( ESI ) 285848b8605Smrg RET 286848b8605Smrg 287848b8605Smrg 288848b8605Smrg 289848b8605Smrg 290848b8605SmrgALIGNTEXT16 291848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points3_3d_no_rot ) 292848b8605SmrgHIDDEN(_mesa_3dnow_transform_points3_3d_no_rot) 293848b8605SmrgGLNAME( _mesa_3dnow_transform_points3_3d_no_rot ): 294848b8605Smrg 295848b8605Smrg PUSH_L ( ESI ) 296848b8605Smrg 297848b8605Smrg MOV_L ( ARG_DEST, ECX ) 298848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 299848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 300848b8605Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 301848b8605Smrg OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) ) 302848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 303848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 304848b8605Smrg 305848b8605Smrg PUSH_L ( EDI ) 306848b8605Smrg 307848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 308848b8605Smrg MOV_L ( ESI, ECX ) 309848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 310848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 311848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 312848b8605Smrg 313848b8605Smrg TEST_L ( ESI, ESI ) 314848b8605Smrg JZ ( LLBL( G3TP3NRR_2 ) ) 315848b8605Smrg 316848b8605Smrg PREFETCH ( REGIND(EAX) ) 317848b8605Smrg PREFETCHW ( REGIND(EDX) ) 318848b8605Smrg 319848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 320848b8605Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */ 321848b8605Smrg 322848b8605Smrg MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */ 323848b8605Smrg PUNPCKLDQ ( MM2, MM2 ) /* m22 | m22 */ 324848b8605Smrg 325848b8605Smrg MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */ 326848b8605Smrg MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */ 327848b8605Smrg 328848b8605Smrg PUNPCKLDQ ( MM3, MM3 ) /* m32 | m32 */ 329848b8605Smrg 330848b8605Smrg 331848b8605SmrgALIGNTEXT16 332848b8605SmrgLLBL( G3TP3NRR_1 ): 333848b8605Smrg 334848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 335848b8605Smrg 336848b8605Smrg MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ 337848b8605Smrg MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */ 338848b8605Smrg 339848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 340848b8605Smrg PREFETCHW ( REGIND(EAX) ) 341848b8605Smrg 342848b8605Smrg PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */ 343848b8605Smrg 344848b8605Smrg PFADD ( MM1, MM4 ) /* x1*m11+m31 | x0*m00+m30 */ 345848b8605Smrg PFMUL ( MM2, MM5 ) /* | x2*m22 */ 346848b8605Smrg 347848b8605Smrg PFADD ( MM3, MM5 ) /* | x2*m22+m32 */ 348848b8605Smrg MOVQ ( MM4, REGIND(EDX) ) /* write r0, r1 */ 349848b8605Smrg 350848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 351848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 352848b8605Smrg 353848b8605Smrg MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 */ 354848b8605Smrg JNZ ( LLBL( G3TP3NRR_1 ) ) /* cnt > 0 ? -> process next vertex */ 355848b8605Smrg 356848b8605SmrgLLBL( G3TP3NRR_2 ): 357848b8605Smrg 358848b8605Smrg FEMMS 359848b8605Smrg POP_L ( EDI ) 360848b8605Smrg POP_L ( ESI ) 361848b8605Smrg RET 362848b8605Smrg 363848b8605Smrg 364848b8605Smrg 365848b8605Smrg 366848b8605SmrgALIGNTEXT16 367848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points3_2d ) 368848b8605SmrgHIDDEN(_mesa_3dnow_transform_points3_2d) 369848b8605SmrgGLNAME( _mesa_3dnow_transform_points3_2d ): 370848b8605Smrg 371848b8605Smrg PUSH_L ( ESI ) 372848b8605Smrg 373848b8605Smrg MOV_L ( ARG_DEST, ECX ) 374848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 375848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 376848b8605Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 377848b8605Smrg OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) ) 378848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 379848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 380848b8605Smrg 381848b8605Smrg PUSH_L ( EDI ) 382848b8605Smrg 383848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 384848b8605Smrg MOV_L ( ESI, ECX ) 385848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 386848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 387848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 388848b8605Smrg 389848b8605Smrg TEST_L ( ESI, ESI ) 390848b8605Smrg JZ ( LLBL( G3TP2R_3) ) 391848b8605Smrg 392848b8605Smrg PREFETCH ( REGIND(EAX) ) 393848b8605Smrg PREFETCHW ( REGIND(EDX) ) 394848b8605Smrg 395848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 396848b8605Smrg PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */ 397848b8605Smrg 398848b8605Smrg MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */ 399848b8605Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */ 400848b8605Smrg 401848b8605Smrg MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */ 402848b8605Smrg 403848b8605SmrgALIGNTEXT16 404848b8605SmrgLLBL( G3TP2R_2 ): 405848b8605Smrg 406848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 407848b8605Smrg 408848b8605Smrg MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */ 409848b8605Smrg MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */ 410848b8605Smrg 411848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 412848b8605Smrg PREFETCH ( REGIND(EAX) ) 413848b8605Smrg 414848b8605Smrg MOVQ ( MM3, MM4 ) /* x1 | x0 */ 415848b8605Smrg PFMUL ( MM0, MM3 ) /* x1*m10 | x0*m00 */ 416848b8605Smrg 417848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 418848b8605Smrg PFMUL ( MM1, MM4 ) /* x1*m11 | x0*m01 */ 419848b8605Smrg 420848b8605Smrg PFACC ( MM4, MM3 ) /* x0*m00+x1*m10 | x0*m01+x1*m11 */ 421848b8605Smrg MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 (=x2) */ 422848b8605Smrg 423848b8605Smrg PFADD ( MM2, MM3 ) /* x0*...*m10+m30 | x0*...*m11+m31 */ 424848b8605Smrg MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */ 425848b8605Smrg 426848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 427848b8605Smrg JNZ ( LLBL( G3TP2R_2 ) ) /* cnt > 0 ? -> process next vertex */ 428848b8605Smrg 429848b8605SmrgLLBL( G3TP2R_3 ): 430848b8605Smrg 431848b8605Smrg FEMMS 432848b8605Smrg POP_L ( EDI ) 433848b8605Smrg POP_L ( ESI ) 434848b8605Smrg RET 435848b8605Smrg 436848b8605Smrg 437848b8605Smrg 438848b8605Smrg 439848b8605SmrgALIGNTEXT16 440848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points3_2d_no_rot ) 441848b8605SmrgHIDDEN(_mesa_3dnow_transform_points3_2d_no_rot) 442848b8605SmrgGLNAME( _mesa_3dnow_transform_points3_2d_no_rot ): 443848b8605Smrg 444848b8605Smrg PUSH_L ( ESI ) 445848b8605Smrg 446848b8605Smrg MOV_L ( ARG_DEST, ECX ) 447848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 448848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 449848b8605Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 450848b8605Smrg OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) ) 451848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 452848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 453848b8605Smrg 454848b8605Smrg PUSH_L ( EDI ) 455848b8605Smrg 456848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 457848b8605Smrg MOV_L ( ESI, ECX ) 458848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 459848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 460848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 461848b8605Smrg 462848b8605Smrg TEST_L ( ESI, ESI ) 463848b8605Smrg JZ ( LLBL( G3TP2NRR_2 ) ) 464848b8605Smrg 465848b8605Smrg PREFETCH ( REGIND(EAX) ) 466848b8605Smrg PREFETCHW ( REGIND(EDX) ) 467848b8605Smrg 468848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 469848b8605Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */ 470848b8605Smrg 471848b8605Smrg MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */ 472848b8605Smrg 473848b8605Smrg 474848b8605SmrgALIGNTEXT16 475848b8605SmrgLLBL( G3TP2NRR_1 ): 476848b8605Smrg 477848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 478848b8605Smrg 479848b8605Smrg MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ 480848b8605Smrg MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */ 481848b8605Smrg 482848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 483848b8605Smrg PREFETCH ( REGIND(EAX) ) 484848b8605Smrg 485848b8605Smrg PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */ 486848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 487848b8605Smrg 488848b8605Smrg PFADD ( MM1, MM4 ) /* x1*m11+m31 | x0*m00+m30 */ 489848b8605Smrg 490848b8605Smrg MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */ 491848b8605Smrg MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 (=x2) */ 492848b8605Smrg 493848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 494848b8605Smrg JNZ ( LLBL( G3TP2NRR_1 ) ) /* cnt > 0 ? -> process next vertex */ 495848b8605Smrg 496848b8605SmrgLLBL( G3TP2NRR_2 ): 497848b8605Smrg 498848b8605Smrg FEMMS 499848b8605Smrg POP_L ( EDI ) 500848b8605Smrg POP_L ( ESI ) 501848b8605Smrg RET 502848b8605Smrg 503848b8605Smrg 504848b8605Smrg 505848b8605Smrg 506848b8605SmrgALIGNTEXT16 507848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points3_identity ) 508848b8605SmrgHIDDEN(_mesa_3dnow_transform_points3_identity) 509848b8605SmrgGLNAME( _mesa_3dnow_transform_points3_identity ): 510848b8605Smrg 511848b8605Smrg PUSH_L ( ESI ) 512848b8605Smrg 513848b8605Smrg MOV_L ( ARG_DEST, ECX ) 514848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 515848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 516848b8605Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 517848b8605Smrg OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) ) 518848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 519848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 520848b8605Smrg 521848b8605Smrg PUSH_L ( EDI ) 522848b8605Smrg 523848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 524848b8605Smrg MOV_L ( ESI, ECX ) 525848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 526848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 527848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 528848b8605Smrg 529848b8605Smrg TEST_L ( ESI, ESI ) 530848b8605Smrg JZ ( LLBL( G3TPIR_2 ) ) 531848b8605Smrg 532848b8605Smrg PREFETCHW ( REGIND(EDX) ) 533848b8605Smrg 534848b8605SmrgALIGNTEXT16 535848b8605SmrgLLBL( G3TPIR_1 ): 536848b8605Smrg 537848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) 538848b8605Smrg 539848b8605Smrg MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ 540848b8605Smrg MOVD ( REGOFF(8, EAX), MM1 ) /* | x2 */ 541848b8605Smrg 542848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 543848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 544848b8605Smrg 545848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 546848b8605Smrg MOVQ ( MM0, REGOFF(-16, EDX) ) /* r1 | r0 */ 547848b8605Smrg 548848b8605Smrg MOVD ( MM1, REGOFF(-8, EDX) ) /* | r2 */ 549848b8605Smrg JNZ ( LLBL( G3TPIR_1 ) ) /* cnt > 0 ? -> process next vertex */ 550848b8605Smrg 551848b8605SmrgLLBL( G3TPIR_2 ): 552848b8605Smrg 553848b8605Smrg FEMMS 554848b8605Smrg POP_L ( EDI ) 555848b8605Smrg POP_L ( ESI ) 556848b8605Smrg RET 557848b8605Smrg#endif 558848b8605Smrg 559848b8605Smrg#if defined (__ELF__) && defined (__linux__) 560848b8605Smrg .section .note.GNU-stack,"",%progbits 561848b8605Smrg#endif 562