17117f1b4Smrg 27117f1b4Smrg/* 37117f1b4Smrg * Mesa 3-D graphics library 47117f1b4Smrg * 57117f1b4Smrg * Copyright (C) 1999-2001 Brian Paul All Rights Reserved. 67117f1b4Smrg * 77117f1b4Smrg * Permission is hereby granted, free of charge, to any person obtaining a 87117f1b4Smrg * copy of this software and associated documentation files (the "Software"), 97117f1b4Smrg * to deal in the Software without restriction, including without limitation 107117f1b4Smrg * the rights to use, copy, modify, merge, publish, distribute, sublicense, 117117f1b4Smrg * and/or sell copies of the Software, and to permit persons to whom the 127117f1b4Smrg * Software is furnished to do so, subject to the following conditions: 137117f1b4Smrg * 147117f1b4Smrg * The above copyright notice and this permission notice shall be included 157117f1b4Smrg * in all copies or substantial portions of the Software. 167117f1b4Smrg * 177117f1b4Smrg * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 187117f1b4Smrg * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 197117f1b4Smrg * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 20af69d88dSmrg * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR 21af69d88dSmrg * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 22af69d88dSmrg * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 23af69d88dSmrg * OTHER DEALINGS IN THE SOFTWARE. 247117f1b4Smrg */ 257117f1b4Smrg 267117f1b4Smrg#ifdef USE_3DNOW_ASM 27c1f859d4Smrg#include "assyntax.h" 287ec681f3Smrg#define MATH_ASM_PTR_SIZE 4 297ec681f3Smrg#include "math/m_vector_asm.h" 307117f1b4Smrg#include "xform_args.h" 317117f1b4Smrg 327117f1b4Smrg SEG_TEXT 337117f1b4Smrg 347117f1b4Smrg#define FRAME_OFFSET 4 357117f1b4Smrg 367117f1b4Smrg 377117f1b4SmrgALIGNTEXT16 387117f1b4SmrgGLOBL GLNAME( _mesa_3dnow_transform_points3_general ) 397117f1b4SmrgHIDDEN(_mesa_3dnow_transform_points3_general) 407117f1b4SmrgGLNAME( _mesa_3dnow_transform_points3_general ): 417ec681f3Smrg _CET_ENDBR 427117f1b4Smrg PUSH_L ( ESI ) 437117f1b4Smrg 447117f1b4Smrg MOV_L ( ARG_DEST, ECX ) 457117f1b4Smrg MOV_L ( ARG_MATRIX, ESI ) 467117f1b4Smrg MOV_L ( ARG_SOURCE, EAX ) 477117f1b4Smrg MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) ) 487117f1b4Smrg OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) ) 497117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 507117f1b4Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 517117f1b4Smrg 527117f1b4Smrg PUSH_L ( EDI ) 537117f1b4Smrg 547117f1b4Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 557117f1b4Smrg MOV_L ( ESI, ECX ) 567117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 577117f1b4Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 587117f1b4Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 597117f1b4Smrg 607117f1b4Smrg TEST_L ( ESI, ESI ) 617117f1b4Smrg JZ ( LLBL( G3TPGR_2 ) ) 627117f1b4Smrg 637117f1b4Smrg PREFETCHW ( REGIND(EDX) ) 647117f1b4Smrg 657117f1b4SmrgALIGNTEXT16 667117f1b4SmrgLLBL( G3TPGR_1 ): 677117f1b4Smrg 687117f1b4Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 697117f1b4Smrg 707117f1b4Smrg MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ 717117f1b4Smrg MOVD ( REGOFF(8, EAX), MM2 ) /* | x2 */ 727117f1b4Smrg 737117f1b4Smrg ADD_L ( EDI, EAX ) /* next vertex */ 747117f1b4Smrg PREFETCH ( REGIND(EAX) ) 757117f1b4Smrg 767117f1b4Smrg MOVQ ( MM0, MM1 ) /* x1 | x0 */ 777117f1b4Smrg PUNPCKLDQ ( MM2, MM2 ) /* x2 | x2 */ 787117f1b4Smrg 797117f1b4Smrg PUNPCKLDQ ( MM0, MM0 ) /* x0 | x0 */ 807117f1b4Smrg MOVQ ( MM2, MM5 ) /* x2 | x2 */ 817117f1b4Smrg 827117f1b4Smrg PUNPCKHDQ ( MM1, MM1 ) /* x1 | x1 */ 837117f1b4Smrg PFMUL ( REGOFF(32, ECX), MM2 ) /* x2*m9 | x2*m8 */ 847117f1b4Smrg 857117f1b4Smrg MOVQ ( MM0, MM3 ) /* x0 | x0 */ 867117f1b4Smrg PFMUL ( REGOFF(40, ECX), MM5 ) /* x2*m11 | x2*m10 */ 877117f1b4Smrg 887117f1b4Smrg MOVQ ( MM1, MM4 ) /* x1 | x1 */ 897117f1b4Smrg PFMUL ( REGIND(ECX), MM0 ) /* x0*m1 | x0*m0 */ 907117f1b4Smrg 917117f1b4Smrg PFADD ( REGOFF(48, ECX), MM2 ) /* x2*m9+m13 | x2*m8+m12 */ 927117f1b4Smrg PFMUL ( REGOFF(16, ECX), MM1 ) /* x1*m5 | x1*m4 */ 937117f1b4Smrg 947117f1b4Smrg PFADD ( REGOFF(56, ECX), MM5 ) /* x2*m11+m15 | x2*m10+m14 */ 957117f1b4Smrg PFADD ( MM0, MM1 ) /* x0*m1+x1*m5 | x0*m0+x1*m4 */ 967117f1b4Smrg 977117f1b4Smrg PFMUL ( REGOFF(8, ECX), MM3 ) /* x0*m3 | x0*m2 */ 987117f1b4Smrg PFADD ( MM1, MM2 ) /* r1 | r0 */ 997117f1b4Smrg 1007117f1b4Smrg PFMUL ( REGOFF(24, ECX), MM4 ) /* x1*m7 | x1*m6 */ 1017117f1b4Smrg ADD_L ( CONST(16), EDX ) /* next output vertex */ 1027117f1b4Smrg 1037117f1b4Smrg PFADD ( MM3, MM4 ) /* x0*m3+x1*m7 | x0*m2+x1*m6 */ 1047117f1b4Smrg MOVQ ( MM2, REGOFF(-16, EDX) ) /* write r0, r1 */ 1057117f1b4Smrg 1067117f1b4Smrg PFADD ( MM4, MM5 ) /* r3 | r2 */ 1077117f1b4Smrg MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */ 1087117f1b4Smrg 1097117f1b4Smrg DEC_L ( ESI ) /* decrement vertex counter */ 1107117f1b4Smrg JNZ ( LLBL( G3TPGR_1 ) ) /* cnt > 0 ? -> process next vertex */ 1117117f1b4Smrg 1127117f1b4SmrgLLBL( G3TPGR_2 ): 1137117f1b4Smrg 1147117f1b4Smrg FEMMS 1157117f1b4Smrg POP_L ( EDI ) 1167117f1b4Smrg POP_L ( ESI ) 1177117f1b4Smrg RET 1187117f1b4Smrg 1197117f1b4Smrg 1207117f1b4Smrg 1217117f1b4Smrg 1227117f1b4SmrgALIGNTEXT16 1237117f1b4SmrgGLOBL GLNAME( _mesa_3dnow_transform_points3_perspective ) 1247117f1b4SmrgHIDDEN(_mesa_3dnow_transform_points3_perspective) 1257117f1b4SmrgGLNAME( _mesa_3dnow_transform_points3_perspective ): 1267ec681f3Smrg _CET_ENDBR 1277117f1b4Smrg PUSH_L ( ESI ) 1287117f1b4Smrg 1297117f1b4Smrg MOV_L ( ARG_DEST, ECX ) 1307117f1b4Smrg MOV_L ( ARG_MATRIX, ESI ) 1317117f1b4Smrg MOV_L ( ARG_SOURCE, EAX ) 1327117f1b4Smrg MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) ) 1337117f1b4Smrg OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) ) 1347117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 1357117f1b4Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 1367117f1b4Smrg 1377117f1b4Smrg PUSH_L ( EDI ) 1387117f1b4Smrg 1397117f1b4Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 1407117f1b4Smrg MOV_L ( ESI, ECX ) 1417117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 1427117f1b4Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 1437117f1b4Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 1447117f1b4Smrg 1457117f1b4Smrg TEST_L ( ESI, ESI ) 1467117f1b4Smrg JZ ( LLBL( G3TPPR_2 ) ) 1477117f1b4Smrg 1487117f1b4Smrg PREFETCH ( REGIND(EAX) ) 1497117f1b4Smrg PREFETCHW ( REGIND(EDX) ) 1507117f1b4Smrg 1517117f1b4Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 1527117f1b4Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */ 1537117f1b4Smrg 1547117f1b4Smrg MOVQ ( REGOFF(32, ECX), MM1 ) /* m21 | m20 */ 1557117f1b4Smrg MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */ 1567117f1b4Smrg 1577117f1b4Smrg MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */ 1587117f1b4Smrg 1597117f1b4SmrgALIGNTEXT16 1607117f1b4SmrgLLBL( G3TPPR_1 ): 1617117f1b4Smrg 1627117f1b4Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 1637117f1b4Smrg 1647117f1b4Smrg MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */ 1657117f1b4Smrg MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ 1667117f1b4Smrg 1677117f1b4Smrg ADD_L ( EDI, EAX ) /* next vertex */ 1687117f1b4Smrg PREFETCH ( REGIND(EAX) ) 1697117f1b4Smrg 1707117f1b4Smrg PXOR ( MM7, MM7 ) /* 0 | 0 */ 1717117f1b4Smrg MOVQ ( MM5, MM6 ) /* | x2 */ 1727117f1b4Smrg 1737117f1b4Smrg PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */ 1747117f1b4Smrg PFSUB ( MM5, MM7 ) /* | -x2 */ 1757117f1b4Smrg 1767117f1b4Smrg PFMUL ( MM2, MM6 ) /* | x2*m22 */ 1777117f1b4Smrg PUNPCKLDQ ( MM5, MM5 ) /* x2 | x2 */ 1787117f1b4Smrg 1797117f1b4Smrg ADD_L ( CONST(16), EDX ) /* next r */ 1807117f1b4Smrg PFMUL ( MM1, MM5 ) /* x2*m21 | x2*m20 */ 1817117f1b4Smrg 1827117f1b4Smrg PFADD ( MM3, MM6 ) /* | x2*m22+m32 */ 1837117f1b4Smrg PFADD ( MM4, MM5 ) /* x1*m11+x2*m21 | x0*m00+x2*m20 */ 1847117f1b4Smrg 1857117f1b4Smrg MOVQ ( MM5, REGOFF(-16, EDX) ) /* write r0, r1 */ 1867117f1b4Smrg MOVD ( MM6, REGOFF(-8, EDX) ) /* write r2 */ 1877117f1b4Smrg 1887117f1b4Smrg MOVD ( MM7, REGOFF(-4, EDX) ) /* write r3 */ 1897117f1b4Smrg 1907117f1b4Smrg DEC_L ( ESI ) /* decrement vertex counter */ 1917117f1b4Smrg JNZ ( LLBL( G3TPPR_1 ) ) /* cnt > 0 ? -> process next vertex */ 1927117f1b4Smrg 1937117f1b4SmrgLLBL( G3TPPR_2 ): 1947117f1b4Smrg 1957117f1b4Smrg FEMMS 1967117f1b4Smrg POP_L ( EDI ) 1977117f1b4Smrg POP_L ( ESI ) 1987117f1b4Smrg RET 1997117f1b4Smrg 2007117f1b4Smrg 2017117f1b4Smrg 2027117f1b4Smrg 2037117f1b4SmrgALIGNTEXT16 2047117f1b4SmrgGLOBL GLNAME( _mesa_3dnow_transform_points3_3d ) 2057117f1b4SmrgHIDDEN(_mesa_3dnow_transform_points3_3d) 2067117f1b4SmrgGLNAME( _mesa_3dnow_transform_points3_3d ): 2077ec681f3Smrg _CET_ENDBR 2087117f1b4Smrg PUSH_L ( ESI ) 2097117f1b4Smrg 2107117f1b4Smrg MOV_L ( ARG_DEST, ECX ) 2117117f1b4Smrg MOV_L ( ARG_MATRIX, ESI ) 2127117f1b4Smrg MOV_L ( ARG_SOURCE, EAX ) 2137117f1b4Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 2147117f1b4Smrg OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) ) 2157117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 2167117f1b4Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 2177117f1b4Smrg 2187117f1b4Smrg PUSH_L ( EDI ) 2197117f1b4Smrg 2207117f1b4Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 2217117f1b4Smrg MOV_L ( ESI, ECX ) 2227117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 2237117f1b4Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 2247117f1b4Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 2257117f1b4Smrg 2267117f1b4Smrg TEST_L ( ESI, ESI ) 2277117f1b4Smrg JZ ( LLBL( G3TP3R_2 ) ) 2287117f1b4Smrg 2297117f1b4Smrg PREFETCH ( REGIND(EAX) ) 2307117f1b4Smrg PREFETCH ( REGIND(EDX) ) 2317117f1b4Smrg 2327117f1b4Smrg MOVD ( REGOFF(8, ECX), MM7 ) /* | m2 */ 2337117f1b4Smrg PUNPCKLDQ ( REGOFF(24, ECX), MM7 ) /* m6 | m2 */ 2347117f1b4Smrg 2357117f1b4Smrg 2367117f1b4SmrgALIGNTEXT16 2377117f1b4SmrgLLBL( G3TP3R_1 ): 2387117f1b4Smrg 2397117f1b4Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 2407117f1b4Smrg 2417117f1b4Smrg MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ 2427117f1b4Smrg MOVD ( REGOFF(8, EAX), MM1 ) /* | x2 */ 2437117f1b4Smrg 2447117f1b4Smrg ADD_L ( EDI, EAX ) /* next vertex */ 2457117f1b4Smrg PREFETCH ( REGIND(EAX) ) 2467117f1b4Smrg 2477117f1b4Smrg MOVQ ( MM0, MM2 ) /* x1 | x0 */ 2487117f1b4Smrg ADD_L ( CONST(16), EDX ) /* next r */ 2497117f1b4Smrg 2507117f1b4Smrg PUNPCKLDQ ( MM2, MM2 ) /* x0 | x0 */ 2517117f1b4Smrg MOVQ ( MM0, MM3 ) /* x1 | x0 */ 2527117f1b4Smrg 2537117f1b4Smrg PFMUL ( REGIND(ECX), MM2 ) /* x0*m1 | x0*m0 */ 2547117f1b4Smrg PUNPCKHDQ ( MM3, MM3 ) /* x1 | x1 */ 2557117f1b4Smrg 2567117f1b4Smrg MOVQ ( MM1, MM4 ) /* | x2 */ 2577117f1b4Smrg PFMUL ( REGOFF(16, ECX), MM3 ) /* x1*m5 | x1*m4 */ 2587117f1b4Smrg 2597117f1b4Smrg PUNPCKLDQ ( MM4, MM4 ) /* x2 | x2 */ 2607117f1b4Smrg PFADD ( MM2, MM3 ) /* x0*m1+x1*m5 | x0*m0+x1*m4 */ 2617117f1b4Smrg 2627117f1b4Smrg PFMUL ( REGOFF(32, ECX), MM4 ) /* x2*m9 | x2*m8 */ 2637117f1b4Smrg PFADD ( REGOFF(48, ECX), MM3 ) /* x0*m1+...+m11 | x0*m0+x1*m4+m12 */ 2647117f1b4Smrg 2657117f1b4Smrg PFMUL ( MM7, MM0 ) /* x1*m6 | x0*m2 */ 2667117f1b4Smrg PFADD ( MM4, MM3 ) /* r1 | r0 */ 2677117f1b4Smrg 2687117f1b4Smrg PFMUL ( REGOFF(40, ECX), MM1 ) /* | x2*m10 */ 2697117f1b4Smrg PUNPCKLDQ ( REGOFF(56, ECX), MM1 ) /* m14 | x2*m10 */ 2707117f1b4Smrg 2717117f1b4Smrg PFACC ( MM0, MM1 ) 2727117f1b4Smrg 2737117f1b4Smrg MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */ 2747117f1b4Smrg PFACC ( MM1, MM1 ) /* | r2 */ 2757117f1b4Smrg 2767117f1b4Smrg MOVD ( MM1, REGOFF(-8, EDX) ) /* write r2 */ 2777117f1b4Smrg 2787117f1b4Smrg DEC_L ( ESI ) /* decrement vertex counter */ 2797117f1b4Smrg JNZ ( LLBL( G3TP3R_1 ) ) /* cnt > 0 ? -> process next vertex */ 2807117f1b4Smrg 2817117f1b4SmrgLLBL( G3TP3R_2 ): 2827117f1b4Smrg 2837117f1b4Smrg FEMMS 2847117f1b4Smrg POP_L ( EDI ) 2857117f1b4Smrg POP_L ( ESI ) 2867117f1b4Smrg RET 2877117f1b4Smrg 2887117f1b4Smrg 2897117f1b4Smrg 2907117f1b4Smrg 2917117f1b4SmrgALIGNTEXT16 2927117f1b4SmrgGLOBL GLNAME( _mesa_3dnow_transform_points3_3d_no_rot ) 2937117f1b4SmrgHIDDEN(_mesa_3dnow_transform_points3_3d_no_rot) 2947117f1b4SmrgGLNAME( _mesa_3dnow_transform_points3_3d_no_rot ): 2957ec681f3Smrg _CET_ENDBR 2967117f1b4Smrg PUSH_L ( ESI ) 2977117f1b4Smrg 2987117f1b4Smrg MOV_L ( ARG_DEST, ECX ) 2997117f1b4Smrg MOV_L ( ARG_MATRIX, ESI ) 3007117f1b4Smrg MOV_L ( ARG_SOURCE, EAX ) 3017117f1b4Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 3027117f1b4Smrg OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) ) 3037117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 3047117f1b4Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 3057117f1b4Smrg 3067117f1b4Smrg PUSH_L ( EDI ) 3077117f1b4Smrg 3087117f1b4Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 3097117f1b4Smrg MOV_L ( ESI, ECX ) 3107117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 3117117f1b4Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 3127117f1b4Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 3137117f1b4Smrg 3147117f1b4Smrg TEST_L ( ESI, ESI ) 3157117f1b4Smrg JZ ( LLBL( G3TP3NRR_2 ) ) 3167117f1b4Smrg 3177117f1b4Smrg PREFETCH ( REGIND(EAX) ) 3187117f1b4Smrg PREFETCHW ( REGIND(EDX) ) 3197117f1b4Smrg 3207117f1b4Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 3217117f1b4Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */ 3227117f1b4Smrg 3237117f1b4Smrg MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */ 3247117f1b4Smrg PUNPCKLDQ ( MM2, MM2 ) /* m22 | m22 */ 3257117f1b4Smrg 3267117f1b4Smrg MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */ 3277117f1b4Smrg MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */ 3287117f1b4Smrg 3297117f1b4Smrg PUNPCKLDQ ( MM3, MM3 ) /* m32 | m32 */ 3307117f1b4Smrg 3317117f1b4Smrg 3327117f1b4SmrgALIGNTEXT16 3337117f1b4SmrgLLBL( G3TP3NRR_1 ): 3347117f1b4Smrg 3357117f1b4Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 3367117f1b4Smrg 3377117f1b4Smrg MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ 3387117f1b4Smrg MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */ 3397117f1b4Smrg 3407117f1b4Smrg ADD_L ( EDI, EAX ) /* next vertex */ 3417117f1b4Smrg PREFETCHW ( REGIND(EAX) ) 3427117f1b4Smrg 3437117f1b4Smrg PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */ 3447117f1b4Smrg 3457117f1b4Smrg PFADD ( MM1, MM4 ) /* x1*m11+m31 | x0*m00+m30 */ 3467117f1b4Smrg PFMUL ( MM2, MM5 ) /* | x2*m22 */ 3477117f1b4Smrg 3487117f1b4Smrg PFADD ( MM3, MM5 ) /* | x2*m22+m32 */ 3497117f1b4Smrg MOVQ ( MM4, REGIND(EDX) ) /* write r0, r1 */ 3507117f1b4Smrg 3517117f1b4Smrg ADD_L ( CONST(16), EDX ) /* next r */ 3527117f1b4Smrg DEC_L ( ESI ) /* decrement vertex counter */ 3537117f1b4Smrg 3547117f1b4Smrg MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 */ 3557117f1b4Smrg JNZ ( LLBL( G3TP3NRR_1 ) ) /* cnt > 0 ? -> process next vertex */ 3567117f1b4Smrg 3577117f1b4SmrgLLBL( G3TP3NRR_2 ): 3587117f1b4Smrg 3597117f1b4Smrg FEMMS 3607117f1b4Smrg POP_L ( EDI ) 3617117f1b4Smrg POP_L ( ESI ) 3627117f1b4Smrg RET 3637117f1b4Smrg 3647117f1b4Smrg 3657117f1b4Smrg 3667117f1b4Smrg 3677117f1b4SmrgALIGNTEXT16 3687117f1b4SmrgGLOBL GLNAME( _mesa_3dnow_transform_points3_2d ) 3697117f1b4SmrgHIDDEN(_mesa_3dnow_transform_points3_2d) 3707117f1b4SmrgGLNAME( _mesa_3dnow_transform_points3_2d ): 3717ec681f3Smrg _CET_ENDBR 3727117f1b4Smrg PUSH_L ( ESI ) 3737117f1b4Smrg 3747117f1b4Smrg MOV_L ( ARG_DEST, ECX ) 3757117f1b4Smrg MOV_L ( ARG_MATRIX, ESI ) 3767117f1b4Smrg MOV_L ( ARG_SOURCE, EAX ) 3777117f1b4Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 3787117f1b4Smrg OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) ) 3797117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 3807117f1b4Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 3817117f1b4Smrg 3827117f1b4Smrg PUSH_L ( EDI ) 3837117f1b4Smrg 3847117f1b4Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 3857117f1b4Smrg MOV_L ( ESI, ECX ) 3867117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 3877117f1b4Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 3887117f1b4Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 3897117f1b4Smrg 3907117f1b4Smrg TEST_L ( ESI, ESI ) 3917117f1b4Smrg JZ ( LLBL( G3TP2R_3) ) 3927117f1b4Smrg 3937117f1b4Smrg PREFETCH ( REGIND(EAX) ) 3947117f1b4Smrg PREFETCHW ( REGIND(EDX) ) 3957117f1b4Smrg 3967117f1b4Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 3977117f1b4Smrg PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */ 3987117f1b4Smrg 3997117f1b4Smrg MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */ 4007117f1b4Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */ 4017117f1b4Smrg 4027117f1b4Smrg MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */ 4037117f1b4Smrg 4047117f1b4SmrgALIGNTEXT16 4057117f1b4SmrgLLBL( G3TP2R_2 ): 4067117f1b4Smrg 4077117f1b4Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 4087117f1b4Smrg 4097117f1b4Smrg MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */ 4107117f1b4Smrg MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */ 4117117f1b4Smrg 4127117f1b4Smrg ADD_L ( EDI, EAX ) /* next vertex */ 4137117f1b4Smrg PREFETCH ( REGIND(EAX) ) 4147117f1b4Smrg 4157117f1b4Smrg MOVQ ( MM3, MM4 ) /* x1 | x0 */ 4167117f1b4Smrg PFMUL ( MM0, MM3 ) /* x1*m10 | x0*m00 */ 4177117f1b4Smrg 4187117f1b4Smrg ADD_L ( CONST(16), EDX ) /* next r */ 4197117f1b4Smrg PFMUL ( MM1, MM4 ) /* x1*m11 | x0*m01 */ 4207117f1b4Smrg 4217117f1b4Smrg PFACC ( MM4, MM3 ) /* x0*m00+x1*m10 | x0*m01+x1*m11 */ 4227117f1b4Smrg MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 (=x2) */ 4237117f1b4Smrg 4247117f1b4Smrg PFADD ( MM2, MM3 ) /* x0*...*m10+m30 | x0*...*m11+m31 */ 4257117f1b4Smrg MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */ 4267117f1b4Smrg 4277117f1b4Smrg DEC_L ( ESI ) /* decrement vertex counter */ 4287117f1b4Smrg JNZ ( LLBL( G3TP2R_2 ) ) /* cnt > 0 ? -> process next vertex */ 4297117f1b4Smrg 4307117f1b4SmrgLLBL( G3TP2R_3 ): 4317117f1b4Smrg 4327117f1b4Smrg FEMMS 4337117f1b4Smrg POP_L ( EDI ) 4347117f1b4Smrg POP_L ( ESI ) 4357117f1b4Smrg RET 4367117f1b4Smrg 4377117f1b4Smrg 4387117f1b4Smrg 4397117f1b4Smrg 4407117f1b4SmrgALIGNTEXT16 4417117f1b4SmrgGLOBL GLNAME( _mesa_3dnow_transform_points3_2d_no_rot ) 4427117f1b4SmrgHIDDEN(_mesa_3dnow_transform_points3_2d_no_rot) 4437117f1b4SmrgGLNAME( _mesa_3dnow_transform_points3_2d_no_rot ): 4447ec681f3Smrg _CET_ENDBR 4457117f1b4Smrg PUSH_L ( ESI ) 4467117f1b4Smrg 4477117f1b4Smrg MOV_L ( ARG_DEST, ECX ) 4487117f1b4Smrg MOV_L ( ARG_MATRIX, ESI ) 4497117f1b4Smrg MOV_L ( ARG_SOURCE, EAX ) 4507117f1b4Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 4517117f1b4Smrg OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) ) 4527117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 4537117f1b4Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 4547117f1b4Smrg 4557117f1b4Smrg PUSH_L ( EDI ) 4567117f1b4Smrg 4577117f1b4Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 4587117f1b4Smrg MOV_L ( ESI, ECX ) 4597117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 4607117f1b4Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 4617117f1b4Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 4627117f1b4Smrg 4637117f1b4Smrg TEST_L ( ESI, ESI ) 4647117f1b4Smrg JZ ( LLBL( G3TP2NRR_2 ) ) 4657117f1b4Smrg 4667117f1b4Smrg PREFETCH ( REGIND(EAX) ) 4677117f1b4Smrg PREFETCHW ( REGIND(EDX) ) 4687117f1b4Smrg 4697117f1b4Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 4707117f1b4Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */ 4717117f1b4Smrg 4727117f1b4Smrg MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */ 4737117f1b4Smrg 4747117f1b4Smrg 4757117f1b4SmrgALIGNTEXT16 4767117f1b4SmrgLLBL( G3TP2NRR_1 ): 4777117f1b4Smrg 4787117f1b4Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 4797117f1b4Smrg 4807117f1b4Smrg MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ 4817117f1b4Smrg MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */ 4827117f1b4Smrg 4837117f1b4Smrg ADD_L ( EDI, EAX ) /* next vertex */ 4847117f1b4Smrg PREFETCH ( REGIND(EAX) ) 4857117f1b4Smrg 4867117f1b4Smrg PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */ 4877117f1b4Smrg ADD_L ( CONST(16), EDX ) /* next r */ 4887117f1b4Smrg 4897117f1b4Smrg PFADD ( MM1, MM4 ) /* x1*m11+m31 | x0*m00+m30 */ 4907117f1b4Smrg 4917117f1b4Smrg MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */ 4927117f1b4Smrg MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 (=x2) */ 4937117f1b4Smrg 4947117f1b4Smrg DEC_L ( ESI ) /* decrement vertex counter */ 4957117f1b4Smrg JNZ ( LLBL( G3TP2NRR_1 ) ) /* cnt > 0 ? -> process next vertex */ 4967117f1b4Smrg 4977117f1b4SmrgLLBL( G3TP2NRR_2 ): 4987117f1b4Smrg 4997117f1b4Smrg FEMMS 5007117f1b4Smrg POP_L ( EDI ) 5017117f1b4Smrg POP_L ( ESI ) 5027117f1b4Smrg RET 5037117f1b4Smrg 5047117f1b4Smrg 5057117f1b4Smrg 5067117f1b4Smrg 5077117f1b4SmrgALIGNTEXT16 5087117f1b4SmrgGLOBL GLNAME( _mesa_3dnow_transform_points3_identity ) 5097117f1b4SmrgHIDDEN(_mesa_3dnow_transform_points3_identity) 5107117f1b4SmrgGLNAME( _mesa_3dnow_transform_points3_identity ): 5117ec681f3Smrg _CET_ENDBR 5127117f1b4Smrg PUSH_L ( ESI ) 5137117f1b4Smrg 5147117f1b4Smrg MOV_L ( ARG_DEST, ECX ) 5157117f1b4Smrg MOV_L ( ARG_MATRIX, ESI ) 5167117f1b4Smrg MOV_L ( ARG_SOURCE, EAX ) 5177117f1b4Smrg MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) ) 5187117f1b4Smrg OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) ) 5197117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 5207117f1b4Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 5217117f1b4Smrg 5227117f1b4Smrg PUSH_L ( EDI ) 5237117f1b4Smrg 5247117f1b4Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 5257117f1b4Smrg MOV_L ( ESI, ECX ) 5267117f1b4Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 5277117f1b4Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 5287117f1b4Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 5297117f1b4Smrg 5307117f1b4Smrg TEST_L ( ESI, ESI ) 5317117f1b4Smrg JZ ( LLBL( G3TPIR_2 ) ) 5327117f1b4Smrg 5337117f1b4Smrg PREFETCHW ( REGIND(EDX) ) 5347117f1b4Smrg 5357117f1b4SmrgALIGNTEXT16 5367117f1b4SmrgLLBL( G3TPIR_1 ): 5377117f1b4Smrg 5387117f1b4Smrg PREFETCHW ( REGOFF(32, EDX) ) 5397117f1b4Smrg 5407117f1b4Smrg MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ 5417117f1b4Smrg MOVD ( REGOFF(8, EAX), MM1 ) /* | x2 */ 5427117f1b4Smrg 5437117f1b4Smrg ADD_L ( EDI, EAX ) /* next vertex */ 5447117f1b4Smrg ADD_L ( CONST(16), EDX ) /* next r */ 5457117f1b4Smrg 5467117f1b4Smrg DEC_L ( ESI ) /* decrement vertex counter */ 5477117f1b4Smrg MOVQ ( MM0, REGOFF(-16, EDX) ) /* r1 | r0 */ 5487117f1b4Smrg 5497117f1b4Smrg MOVD ( MM1, REGOFF(-8, EDX) ) /* | r2 */ 5507117f1b4Smrg JNZ ( LLBL( G3TPIR_1 ) ) /* cnt > 0 ? -> process next vertex */ 5517117f1b4Smrg 5527117f1b4SmrgLLBL( G3TPIR_2 ): 5537117f1b4Smrg 5547117f1b4Smrg FEMMS 5557117f1b4Smrg POP_L ( EDI ) 5567117f1b4Smrg POP_L ( ESI ) 5577117f1b4Smrg RET 5587117f1b4Smrg#endif 5597117f1b4Smrg 5607117f1b4Smrg#if defined (__ELF__) && defined (__linux__) 5617117f1b4Smrg .section .note.GNU-stack,"",%progbits 5627117f1b4Smrg#endif 563