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/* 277117f1b4Smrg * NOTE: Avoid using spaces in between '(' ')' and arguments, especially 287117f1b4Smrg * with macros like CONST, LLBL that expand to CONCAT(...). Putting spaces 297117f1b4Smrg * in there will break the build on some platforms. 307117f1b4Smrg */ 317117f1b4Smrg 32c1f859d4Smrg#include "assyntax.h" 337ec681f3Smrg#define MATH_ASM_PTR_SIZE 4 347ec681f3Smrg#include "math/m_vector_asm.h" 357117f1b4Smrg#include "xform_args.h" 367117f1b4Smrg 377117f1b4Smrg SEG_TEXT 387117f1b4Smrg 397117f1b4Smrg#define FP_ONE 1065353216 407117f1b4Smrg#define FP_ZERO 0 417117f1b4Smrg 427117f1b4Smrg#define SRC0 REGOFF(0, ESI) 437117f1b4Smrg#define SRC1 REGOFF(4, ESI) 447117f1b4Smrg#define SRC2 REGOFF(8, ESI) 457117f1b4Smrg#define SRC3 REGOFF(12, ESI) 467117f1b4Smrg#define DST0 REGOFF(0, EDI) 477117f1b4Smrg#define DST1 REGOFF(4, EDI) 487117f1b4Smrg#define DST2 REGOFF(8, EDI) 497117f1b4Smrg#define DST3 REGOFF(12, EDI) 507117f1b4Smrg#define MAT0 REGOFF(0, EDX) 517117f1b4Smrg#define MAT1 REGOFF(4, EDX) 527117f1b4Smrg#define MAT2 REGOFF(8, EDX) 537117f1b4Smrg#define MAT3 REGOFF(12, EDX) 547117f1b4Smrg#define MAT4 REGOFF(16, EDX) 557117f1b4Smrg#define MAT5 REGOFF(20, EDX) 567117f1b4Smrg#define MAT6 REGOFF(24, EDX) 577117f1b4Smrg#define MAT7 REGOFF(28, EDX) 587117f1b4Smrg#define MAT8 REGOFF(32, EDX) 597117f1b4Smrg#define MAT9 REGOFF(36, EDX) 607117f1b4Smrg#define MAT10 REGOFF(40, EDX) 617117f1b4Smrg#define MAT11 REGOFF(44, EDX) 627117f1b4Smrg#define MAT12 REGOFF(48, EDX) 637117f1b4Smrg#define MAT13 REGOFF(52, EDX) 647117f1b4Smrg#define MAT14 REGOFF(56, EDX) 657117f1b4Smrg#define MAT15 REGOFF(60, EDX) 667117f1b4Smrg 677117f1b4Smrg 687117f1b4SmrgALIGNTEXT16 697117f1b4SmrgGLOBL GLNAME( _mesa_x86_transform_points4_general ) 707117f1b4SmrgHIDDEN(_mesa_x86_transform_points4_general) 717117f1b4SmrgGLNAME( _mesa_x86_transform_points4_general ): 727117f1b4Smrg 737117f1b4Smrg#define FRAME_OFFSET 8 747117f1b4Smrg PUSH_L( ESI ) 757117f1b4Smrg PUSH_L( EDI ) 767117f1b4Smrg 777117f1b4Smrg MOV_L( ARG_SOURCE, ESI ) 787117f1b4Smrg MOV_L( ARG_DEST, EDI ) 797117f1b4Smrg 807117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) 817117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 827117f1b4Smrg 837117f1b4Smrg TEST_L( ECX, ECX ) 847117f1b4Smrg JZ( LLBL(x86_p4_gr_done) ) 857117f1b4Smrg 867117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 877117f1b4Smrg OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 887117f1b4Smrg 897117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 907117f1b4Smrg MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 917117f1b4Smrg 927117f1b4Smrg SHL_L( CONST(4), ECX ) 937117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) 947117f1b4Smrg 957117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) 967117f1b4Smrg ADD_L( EDI, ECX ) 977117f1b4Smrg 987117f1b4SmrgALIGNTEXT16 997117f1b4SmrgLLBL(x86_p4_gr_loop): 1007117f1b4Smrg 1017117f1b4Smrg FLD_S( SRC0 ) /* F4 */ 1027117f1b4Smrg FMUL_S( MAT0 ) 1037117f1b4Smrg FLD_S( SRC0 ) /* F5 F4 */ 1047117f1b4Smrg FMUL_S( MAT1 ) 1057117f1b4Smrg FLD_S( SRC0 ) /* F6 F5 F4 */ 1067117f1b4Smrg FMUL_S( MAT2 ) 1077117f1b4Smrg FLD_S( SRC0 ) /* F7 F6 F5 F4 */ 1087117f1b4Smrg FMUL_S( MAT3 ) 1097117f1b4Smrg 1107117f1b4Smrg FLD_S( SRC1 ) /* F0 F7 F6 F5 F4 */ 1117117f1b4Smrg FMUL_S( MAT4 ) 1127117f1b4Smrg FLD_S( SRC1 ) /* F1 F0 F7 F6 F5 F4 */ 1137117f1b4Smrg FMUL_S( MAT5 ) 1147117f1b4Smrg FLD_S( SRC1 ) /* F2 F1 F0 F7 F6 F5 F4 */ 1157117f1b4Smrg FMUL_S( MAT6 ) 1167117f1b4Smrg FLD_S( SRC1 ) /* F3 F2 F1 F0 F7 F6 F5 F4 */ 1177117f1b4Smrg FMUL_S( MAT7 ) 1187117f1b4Smrg 1197117f1b4Smrg FXCH( ST(3) ) /* F0 F2 F1 F3 F7 F6 F5 F4 */ 1207117f1b4Smrg FADDP( ST0, ST(7) ) /* F2 F1 F3 F7 F6 F5 F4 */ 1217117f1b4Smrg FXCH( ST(1) ) /* F1 F2 F3 F7 F6 F5 F4 */ 1227117f1b4Smrg FADDP( ST0, ST(5) ) /* F2 F3 F7 F6 F5 F4 */ 1237117f1b4Smrg FADDP( ST0, ST(3) ) /* F3 F7 F6 F5 F4 */ 1247117f1b4Smrg FADDP( ST0, ST(1) ) /* F7 F6 F5 F4 */ 1257117f1b4Smrg 1267117f1b4Smrg FLD_S( SRC2 ) /* F0 F7 F6 F5 F4 */ 1277117f1b4Smrg FMUL_S( MAT8 ) 1287117f1b4Smrg FLD_S( SRC2 ) /* F1 F0 F7 F6 F5 F4 */ 1297117f1b4Smrg FMUL_S( MAT9 ) 1307117f1b4Smrg FLD_S( SRC2 ) /* F2 F1 F0 F7 F6 F5 F4 */ 1317117f1b4Smrg FMUL_S( MAT10 ) 1327117f1b4Smrg FLD_S( SRC2 ) /* F3 F2 F1 F0 F7 F6 F5 F4 */ 1337117f1b4Smrg FMUL_S( MAT11 ) 1347117f1b4Smrg 1357117f1b4Smrg FXCH( ST(3) ) /* F0 F2 F1 F3 F7 F6 F5 F4 */ 1367117f1b4Smrg FADDP( ST0, ST(7) ) /* F2 F1 F3 F7 F6 F5 F4 */ 1377117f1b4Smrg FXCH( ST(1) ) /* F1 F2 F3 F7 F6 F5 F4 */ 1387117f1b4Smrg FADDP( ST0, ST(5) ) /* F2 F3 F7 F6 F5 F4 */ 1397117f1b4Smrg FADDP( ST0, ST(3) ) /* F3 F7 F6 F5 F4 */ 1407117f1b4Smrg FADDP( ST0, ST(1) ) /* F7 F6 F5 F4 */ 1417117f1b4Smrg 1427117f1b4Smrg FLD_S( SRC3 ) /* F0 F7 F6 F5 F4 */ 1437117f1b4Smrg FMUL_S( MAT12 ) 1447117f1b4Smrg FLD_S( SRC3 ) /* F1 F0 F7 F6 F5 F4 */ 1457117f1b4Smrg FMUL_S( MAT13 ) 1467117f1b4Smrg FLD_S( SRC3 ) /* F2 F1 F0 F7 F6 F5 F4 */ 1477117f1b4Smrg FMUL_S( MAT14 ) 1487117f1b4Smrg FLD_S( SRC3 ) /* F3 F2 F1 F0 F7 F6 F5 F4 */ 1497117f1b4Smrg FMUL_S( MAT15 ) 1507117f1b4Smrg 1517117f1b4Smrg FXCH( ST(3) ) /* F0 F2 F1 F3 F7 F6 F5 F4 */ 1527117f1b4Smrg FADDP( ST0, ST(7) ) /* F2 F1 F3 F7 F6 F5 F4 */ 1537117f1b4Smrg FXCH( ST(1) ) /* F1 F2 F3 F7 F6 F5 F4 */ 1547117f1b4Smrg FADDP( ST0, ST(5) ) /* F2 F3 F7 F6 F5 F4 */ 1557117f1b4Smrg FADDP( ST0, ST(3) ) /* F3 F7 F6 F5 F4 */ 1567117f1b4Smrg FADDP( ST0, ST(1) ) /* F7 F6 F5 F4 */ 1577117f1b4Smrg 1587117f1b4Smrg FXCH( ST(3) ) /* F4 F6 F5 F7 */ 1597117f1b4Smrg FSTP_S( DST0 ) /* F6 F5 F7 */ 1607117f1b4Smrg FXCH( ST(1) ) /* F5 F6 F7 */ 1617117f1b4Smrg FSTP_S( DST1 ) /* F6 F7 */ 1627117f1b4Smrg FSTP_S( DST2 ) /* F7 */ 1637117f1b4Smrg FSTP_S( DST3 ) /* */ 1647117f1b4Smrg 1657117f1b4SmrgLLBL(x86_p4_gr_skip): 1667117f1b4Smrg 1677117f1b4Smrg ADD_L( CONST(16), EDI ) 1687117f1b4Smrg ADD_L( EAX, ESI ) 1697117f1b4Smrg CMP_L( ECX, EDI ) 1707117f1b4Smrg JNE( LLBL(x86_p4_gr_loop) ) 1717117f1b4Smrg 1727117f1b4SmrgLLBL(x86_p4_gr_done): 1737117f1b4Smrg 1747117f1b4Smrg POP_L( EDI ) 1757117f1b4Smrg POP_L( ESI ) 1767117f1b4Smrg RET 1777117f1b4Smrg#undef FRAME_OFFSET 1787117f1b4Smrg 1797117f1b4Smrg 1807117f1b4Smrg 1817117f1b4Smrg 1827117f1b4SmrgALIGNTEXT16 1837117f1b4SmrgGLOBL GLNAME( _mesa_x86_transform_points4_perspective ) 1847117f1b4SmrgHIDDEN(_mesa_x86_transform_points4_perspective) 1857117f1b4SmrgGLNAME( _mesa_x86_transform_points4_perspective ): 1867117f1b4Smrg 1877117f1b4Smrg#define FRAME_OFFSET 12 1887117f1b4Smrg PUSH_L( ESI ) 1897117f1b4Smrg PUSH_L( EDI ) 1907117f1b4Smrg PUSH_L( EBX ) 1917117f1b4Smrg 1927117f1b4Smrg MOV_L( ARG_SOURCE, ESI ) 1937117f1b4Smrg MOV_L( ARG_DEST, EDI ) 1947117f1b4Smrg 1957117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) 1967117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 1977117f1b4Smrg 1987117f1b4Smrg TEST_L( ECX, ECX ) 1997117f1b4Smrg JZ( LLBL(x86_p4_pr_done) ) 2007117f1b4Smrg 2017117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 2027117f1b4Smrg OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 2037117f1b4Smrg 2047117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 2057117f1b4Smrg MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 2067117f1b4Smrg 2077117f1b4Smrg SHL_L( CONST(4), ECX ) 2087117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) 2097117f1b4Smrg 2107117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) 2117117f1b4Smrg ADD_L( EDI, ECX ) 2127117f1b4Smrg 2137117f1b4SmrgALIGNTEXT16 2147117f1b4SmrgLLBL(x86_p4_pr_loop): 2157117f1b4Smrg 2167117f1b4Smrg FLD_S( SRC0 ) /* F4 */ 2177117f1b4Smrg FMUL_S( MAT0 ) 2187117f1b4Smrg 2197117f1b4Smrg FLD_S( SRC1 ) /* F5 F4 */ 2207117f1b4Smrg FMUL_S( MAT5 ) 2217117f1b4Smrg 2227117f1b4Smrg FLD_S( SRC2 ) /* F0 F5 F4 */ 2237117f1b4Smrg FMUL_S( MAT8 ) 2247117f1b4Smrg FLD_S( SRC2 ) /* F1 F0 F5 F4 */ 2257117f1b4Smrg FMUL_S( MAT9 ) 2267117f1b4Smrg FLD_S( SRC2 ) /* F6 F1 F0 F5 F4 */ 2277117f1b4Smrg FMUL_S( MAT10 ) 2287117f1b4Smrg 2297117f1b4Smrg FXCH( ST(2) ) /* F0 F1 F6 F5 F4 */ 2307117f1b4Smrg FADDP( ST0, ST(4) ) /* F1 F6 F5 F4 */ 2317117f1b4Smrg FADDP( ST0, ST(2) ) /* F6 F5 F4 */ 2327117f1b4Smrg 2337117f1b4Smrg FLD_S( SRC3 ) /* F2 F6 F5 F4 */ 2347117f1b4Smrg FMUL_S( MAT14 ) 2357117f1b4Smrg 2367117f1b4Smrg FADDP( ST0, ST(1) ) /* F6 F5 F4 */ 2377117f1b4Smrg 2387117f1b4Smrg MOV_L( SRC2, EBX ) 2397117f1b4Smrg XOR_L( CONST(-2147483648), EBX )/* change sign */ 2407117f1b4Smrg 2417117f1b4Smrg FXCH( ST(2) ) /* F4 F5 F6 */ 2427117f1b4Smrg FSTP_S( DST0 ) /* F5 F6 */ 2437117f1b4Smrg FSTP_S( DST1 ) /* F6 */ 2447117f1b4Smrg FSTP_S( DST2 ) /* */ 2457117f1b4Smrg MOV_L( EBX, DST3 ) 2467117f1b4Smrg 2477117f1b4SmrgLLBL(x86_p4_pr_skip): 2487117f1b4Smrg 2497117f1b4Smrg ADD_L( CONST(16), EDI ) 2507117f1b4Smrg ADD_L( EAX, ESI ) 2517117f1b4Smrg CMP_L( ECX, EDI ) 2527117f1b4Smrg JNE( LLBL(x86_p4_pr_loop) ) 2537117f1b4Smrg 2547117f1b4SmrgLLBL(x86_p4_pr_done): 2557117f1b4Smrg 2567117f1b4Smrg POP_L( EBX ) 2577117f1b4Smrg POP_L( EDI ) 2587117f1b4Smrg POP_L( ESI ) 2597117f1b4Smrg RET 2607117f1b4Smrg#undef FRAME_OFFSET 2617117f1b4Smrg 2627117f1b4Smrg 2637117f1b4Smrg 2647117f1b4Smrg 2657117f1b4SmrgALIGNTEXT16 2667117f1b4SmrgGLOBL GLNAME( _mesa_x86_transform_points4_3d ) 2677117f1b4SmrgHIDDEN(_mesa_x86_transform_points4_3d) 2687117f1b4SmrgGLNAME( _mesa_x86_transform_points4_3d ): 2697117f1b4Smrg 2707117f1b4Smrg#define FRAME_OFFSET 12 2717117f1b4Smrg PUSH_L( ESI ) 2727117f1b4Smrg PUSH_L( EDI ) 2737117f1b4Smrg PUSH_L( EBX ) 2747117f1b4Smrg 2757117f1b4Smrg MOV_L( ARG_SOURCE, ESI ) 2767117f1b4Smrg MOV_L( ARG_DEST, EDI ) 2777117f1b4Smrg 2787117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) 2797117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 2807117f1b4Smrg 2817117f1b4Smrg TEST_L( ECX, ECX ) 2827117f1b4Smrg JZ( LLBL(x86_p4_3dr_done) ) 2837117f1b4Smrg 2847117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 2857117f1b4Smrg OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 2867117f1b4Smrg 2877117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 2887117f1b4Smrg MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 2897117f1b4Smrg 2907117f1b4Smrg SHL_L( CONST(4), ECX ) 2917117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) 2927117f1b4Smrg 2937117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) 2947117f1b4Smrg ADD_L( EDI, ECX ) 2957117f1b4Smrg 2967117f1b4SmrgALIGNTEXT16 2977117f1b4SmrgLLBL(x86_p4_3dr_loop): 2987117f1b4Smrg 2997117f1b4Smrg FLD_S( SRC0 ) /* F4 */ 3007117f1b4Smrg FMUL_S( MAT0 ) 3017117f1b4Smrg FLD_S( SRC0 ) /* F5 F4 */ 3027117f1b4Smrg FMUL_S( MAT1 ) 3037117f1b4Smrg FLD_S( SRC0 ) /* F6 F5 F4 */ 3047117f1b4Smrg FMUL_S( MAT2 ) 3057117f1b4Smrg 3067117f1b4Smrg FLD_S( SRC1 ) /* F0 F6 F5 F4 */ 3077117f1b4Smrg FMUL_S( MAT4 ) 3087117f1b4Smrg FLD_S( SRC1 ) /* F1 F0 F6 F5 F4 */ 3097117f1b4Smrg FMUL_S( MAT5 ) 3107117f1b4Smrg FLD_S( SRC1 ) /* F2 F1 F0 F6 F5 F4 */ 3117117f1b4Smrg FMUL_S( MAT6 ) 3127117f1b4Smrg 3137117f1b4Smrg FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */ 3147117f1b4Smrg FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */ 3157117f1b4Smrg FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */ 3167117f1b4Smrg FADDP( ST0, ST(1) ) /* F6 F5 F4 */ 3177117f1b4Smrg 3187117f1b4Smrg FLD_S( SRC2 ) /* F0 F6 F5 F4 */ 3197117f1b4Smrg FMUL_S( MAT8 ) 3207117f1b4Smrg FLD_S( SRC2 ) /* F1 F0 F6 F5 F4 */ 3217117f1b4Smrg FMUL_S( MAT9 ) 3227117f1b4Smrg FLD_S( SRC2 ) /* F2 F1 F0 F6 F5 F4 */ 3237117f1b4Smrg FMUL_S( MAT10 ) 3247117f1b4Smrg 3257117f1b4Smrg FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */ 3267117f1b4Smrg FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */ 3277117f1b4Smrg FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */ 3287117f1b4Smrg FADDP( ST0, ST(1) ) /* F6 F5 F4 */ 3297117f1b4Smrg 3307117f1b4Smrg FLD_S( SRC3 ) /* F0 F6 F5 F4 */ 3317117f1b4Smrg FMUL_S( MAT12 ) 3327117f1b4Smrg FLD_S( SRC3 ) /* F1 F0 F6 F5 F4 */ 3337117f1b4Smrg FMUL_S( MAT13 ) 3347117f1b4Smrg FLD_S( SRC3 ) /* F2 F1 F0 F6 F5 F4 */ 3357117f1b4Smrg FMUL_S( MAT14 ) 3367117f1b4Smrg 3377117f1b4Smrg FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */ 3387117f1b4Smrg FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */ 3397117f1b4Smrg FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */ 3407117f1b4Smrg FADDP( ST0, ST(1) ) /* F6 F5 F4 */ 3417117f1b4Smrg 3427117f1b4Smrg MOV_L( SRC3, EBX ) 3437117f1b4Smrg 3447117f1b4Smrg FXCH( ST(2) ) /* F4 F5 F6 */ 3457117f1b4Smrg FSTP_S( DST0 ) /* F5 F6 */ 3467117f1b4Smrg FSTP_S( DST1 ) /* F6 */ 3477117f1b4Smrg FSTP_S( DST2 ) /* */ 3487117f1b4Smrg MOV_L( EBX, DST3 ) 3497117f1b4Smrg 3507117f1b4SmrgLLBL(x86_p4_3dr_skip): 3517117f1b4Smrg 3527117f1b4Smrg ADD_L( CONST(16), EDI ) 3537117f1b4Smrg ADD_L( EAX, ESI ) 3547117f1b4Smrg CMP_L( ECX, EDI ) 3557117f1b4Smrg JNE( LLBL(x86_p4_3dr_loop) ) 3567117f1b4Smrg 3577117f1b4SmrgLLBL(x86_p4_3dr_done): 3587117f1b4Smrg 3597117f1b4Smrg POP_L( EBX ) 3607117f1b4Smrg POP_L( EDI ) 3617117f1b4Smrg POP_L( ESI ) 3627117f1b4Smrg RET 3637117f1b4Smrg#undef FRAME_OFFSET 3647117f1b4Smrg 3657117f1b4Smrg 3667117f1b4Smrg 3677117f1b4Smrg 3687117f1b4SmrgALIGNTEXT16 3697117f1b4SmrgGLOBL GLNAME(_mesa_x86_transform_points4_3d_no_rot) 3707117f1b4SmrgHIDDEN(_mesa_x86_transform_points4_3d_no_rot) 3717117f1b4SmrgGLNAME(_mesa_x86_transform_points4_3d_no_rot): 3727117f1b4Smrg 3737117f1b4Smrg#define FRAME_OFFSET 12 3747117f1b4Smrg PUSH_L( ESI ) 3757117f1b4Smrg PUSH_L( EDI ) 3767117f1b4Smrg PUSH_L( EBX ) 3777117f1b4Smrg 3787117f1b4Smrg MOV_L( ARG_SOURCE, ESI ) 3797117f1b4Smrg MOV_L( ARG_DEST, EDI ) 3807117f1b4Smrg 3817117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) 3827117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 3837117f1b4Smrg 3847117f1b4Smrg TEST_L( ECX, ECX ) 3857117f1b4Smrg JZ( LLBL(x86_p4_3dnrr_done) ) 3867117f1b4Smrg 3877117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 3887117f1b4Smrg OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 3897117f1b4Smrg 3907117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 3917117f1b4Smrg MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 3927117f1b4Smrg 3937117f1b4Smrg SHL_L( CONST(4), ECX ) 3947117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) 3957117f1b4Smrg 3967117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) 3977117f1b4Smrg ADD_L( EDI, ECX ) 3987117f1b4Smrg 3997117f1b4SmrgALIGNTEXT16 4007117f1b4SmrgLLBL(x86_p4_3dnrr_loop): 4017117f1b4Smrg 4027117f1b4Smrg FLD_S( SRC0 ) /* F4 */ 4037117f1b4Smrg FMUL_S( MAT0 ) 4047117f1b4Smrg 4057117f1b4Smrg FLD_S( SRC1 ) /* F5 F4 */ 4067117f1b4Smrg FMUL_S( MAT5 ) 4077117f1b4Smrg 4087117f1b4Smrg FLD_S( SRC2 ) /* F6 F5 F4 */ 4097117f1b4Smrg FMUL_S( MAT10 ) 4107117f1b4Smrg 4117117f1b4Smrg FLD_S( SRC3 ) /* F0 F6 F5 F4 */ 4127117f1b4Smrg FMUL_S( MAT12 ) 4137117f1b4Smrg FLD_S( SRC3 ) /* F1 F0 F6 F5 F4 */ 4147117f1b4Smrg FMUL_S( MAT13 ) 4157117f1b4Smrg FLD_S( SRC3 ) /* F2 F1 F0 F6 F5 F4 */ 4167117f1b4Smrg FMUL_S( MAT14 ) 4177117f1b4Smrg 4187117f1b4Smrg FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */ 4197117f1b4Smrg FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */ 4207117f1b4Smrg FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */ 4217117f1b4Smrg FADDP( ST0, ST(1) ) /* F6 F5 F4 */ 4227117f1b4Smrg 4237117f1b4Smrg MOV_L( SRC3, EBX ) 4247117f1b4Smrg 4257117f1b4Smrg FXCH( ST(2) ) /* F4 F5 F6 */ 4267117f1b4Smrg FSTP_S( DST0 ) /* F5 F6 */ 4277117f1b4Smrg FSTP_S( DST1 ) /* F6 */ 4287117f1b4Smrg FSTP_S( DST2 ) /* */ 4297117f1b4Smrg MOV_L( EBX, DST3 ) 4307117f1b4Smrg 4317117f1b4SmrgLLBL(x86_p4_3dnrr_skip): 4327117f1b4Smrg 4337117f1b4Smrg ADD_L( CONST(16), EDI ) 4347117f1b4Smrg ADD_L( EAX, ESI ) 4357117f1b4Smrg CMP_L( ECX, EDI ) 4367117f1b4Smrg JNE( LLBL(x86_p4_3dnrr_loop) ) 4377117f1b4Smrg 4387117f1b4SmrgLLBL(x86_p4_3dnrr_done): 4397117f1b4Smrg 4407117f1b4Smrg POP_L( EBX ) 4417117f1b4Smrg POP_L( EDI ) 4427117f1b4Smrg POP_L( ESI ) 4437117f1b4Smrg RET 4447117f1b4Smrg#undef FRAME_OFFSET 4457117f1b4Smrg 4467117f1b4Smrg 4477117f1b4Smrg 4487117f1b4Smrg 4497117f1b4SmrgALIGNTEXT16 4507117f1b4SmrgGLOBL GLNAME( _mesa_x86_transform_points4_2d ) 4517117f1b4SmrgHIDDEN(_mesa_x86_transform_points4_2d) 4527117f1b4SmrgGLNAME( _mesa_x86_transform_points4_2d ): 4537117f1b4Smrg 4547117f1b4Smrg#define FRAME_OFFSET 16 4557117f1b4Smrg PUSH_L( ESI ) 4567117f1b4Smrg PUSH_L( EDI ) 4577117f1b4Smrg PUSH_L( EBX ) 4587117f1b4Smrg PUSH_L( EBP ) 4597117f1b4Smrg 4607117f1b4Smrg MOV_L( ARG_SOURCE, ESI ) 4617117f1b4Smrg MOV_L( ARG_DEST, EDI ) 4627117f1b4Smrg 4637117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) 4647117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 4657117f1b4Smrg 4667117f1b4Smrg TEST_L( ECX, ECX ) 4677117f1b4Smrg JZ( LLBL(x86_p4_2dr_done) ) 4687117f1b4Smrg 4697117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 4707117f1b4Smrg OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 4717117f1b4Smrg 4727117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 4737117f1b4Smrg MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 4747117f1b4Smrg 4757117f1b4Smrg SHL_L( CONST(4), ECX ) 4767117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) 4777117f1b4Smrg 4787117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) 4797117f1b4Smrg ADD_L( EDI, ECX ) 4807117f1b4Smrg 4817117f1b4SmrgALIGNTEXT16 4827117f1b4SmrgLLBL(x86_p4_2dr_loop): 4837117f1b4Smrg 4847117f1b4Smrg FLD_S( SRC0 ) /* F4 */ 4857117f1b4Smrg FMUL_S( MAT0 ) 4867117f1b4Smrg FLD_S( SRC0 ) /* F5 F4 */ 4877117f1b4Smrg FMUL_S( MAT1 ) 4887117f1b4Smrg 4897117f1b4Smrg FLD_S( SRC1 ) /* F0 F5 F4 */ 4907117f1b4Smrg FMUL_S( MAT4 ) 4917117f1b4Smrg FLD_S( SRC1 ) /* F1 F0 F5 F4 */ 4927117f1b4Smrg FMUL_S( MAT5 ) 4937117f1b4Smrg 4947117f1b4Smrg FXCH( ST(1) ) /* F0 F1 F5 F4 */ 4957117f1b4Smrg FADDP( ST0, ST(3) ) /* F1 F5 F4 */ 4967117f1b4Smrg FADDP( ST0, ST(1) ) /* F5 F4 */ 4977117f1b4Smrg 4987117f1b4Smrg FLD_S( SRC3 ) /* F0 F5 F4 */ 4997117f1b4Smrg FMUL_S( MAT12 ) 5007117f1b4Smrg FLD_S( SRC3 ) /* F1 F0 F5 F4 */ 5017117f1b4Smrg FMUL_S( MAT13 ) 5027117f1b4Smrg 5037117f1b4Smrg FXCH( ST(1) ) /* F0 F1 F5 F4 */ 5047117f1b4Smrg FADDP( ST0, ST(3) ) /* F1 F5 F4 */ 5057117f1b4Smrg FADDP( ST0, ST(1) ) /* F5 F4 */ 5067117f1b4Smrg 5077117f1b4Smrg MOV_L( SRC2, EBX ) 5087117f1b4Smrg MOV_L( SRC3, EBP ) 5097117f1b4Smrg 5107117f1b4Smrg FXCH( ST(1) ) /* F4 F5 */ 5117117f1b4Smrg FSTP_S( DST0 ) /* F5 */ 5127117f1b4Smrg FSTP_S( DST1 ) /* */ 5137117f1b4Smrg MOV_L( EBX, DST2 ) 5147117f1b4Smrg MOV_L( EBP, DST3 ) 5157117f1b4Smrg 5167117f1b4SmrgLLBL(x86_p4_2dr_skip): 5177117f1b4Smrg 5187117f1b4Smrg ADD_L( CONST(16), EDI ) 5197117f1b4Smrg ADD_L( EAX, ESI ) 5207117f1b4Smrg CMP_L( ECX, EDI ) 5217117f1b4Smrg JNE( LLBL(x86_p4_2dr_loop) ) 5227117f1b4Smrg 5237117f1b4SmrgLLBL(x86_p4_2dr_done): 5247117f1b4Smrg 5257117f1b4Smrg POP_L( EBP ) 5267117f1b4Smrg POP_L( EBX ) 5277117f1b4Smrg POP_L( EDI ) 5287117f1b4Smrg POP_L( ESI ) 5297117f1b4Smrg RET 5307117f1b4Smrg#undef FRAME_OFFSET 5317117f1b4Smrg 5327117f1b4Smrg 5337117f1b4Smrg 5347117f1b4Smrg 5357117f1b4SmrgALIGNTEXT16 5367117f1b4SmrgGLOBL GLNAME( _mesa_x86_transform_points4_2d_no_rot ) 5377117f1b4SmrgHIDDEN(_mesa_x86_transform_points4_2d_no_rot) 5387117f1b4SmrgGLNAME( _mesa_x86_transform_points4_2d_no_rot ): 5397117f1b4Smrg 5407117f1b4Smrg#define FRAME_OFFSET 16 5417117f1b4Smrg PUSH_L( ESI ) 5427117f1b4Smrg PUSH_L( EDI ) 5437117f1b4Smrg PUSH_L( EBX ) 5447117f1b4Smrg PUSH_L( EBP ) 5457117f1b4Smrg 5467117f1b4Smrg MOV_L( ARG_SOURCE, ESI ) 5477117f1b4Smrg MOV_L( ARG_DEST, EDI ) 5487117f1b4Smrg 5497117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) 5507117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 5517117f1b4Smrg 5527117f1b4Smrg TEST_L( ECX, ECX ) 5537117f1b4Smrg JZ( LLBL(x86_p4_2dnrr_done) ) 5547117f1b4Smrg 5557117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 5567117f1b4Smrg OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 5577117f1b4Smrg 5587117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 5597117f1b4Smrg MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 5607117f1b4Smrg 5617117f1b4Smrg SHL_L( CONST(4), ECX ) 5627117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) 5637117f1b4Smrg 5647117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) 5657117f1b4Smrg ADD_L( EDI, ECX ) 5667117f1b4Smrg 5677117f1b4SmrgALIGNTEXT16 5687117f1b4SmrgLLBL(x86_p4_2dnrr_loop): 5697117f1b4Smrg 5707117f1b4Smrg FLD_S( SRC0 ) /* F4 */ 5717117f1b4Smrg FMUL_S( MAT0 ) 5727117f1b4Smrg 5737117f1b4Smrg FLD_S( SRC1 ) /* F5 F4 */ 5747117f1b4Smrg FMUL_S( MAT5 ) 5757117f1b4Smrg 5767117f1b4Smrg FLD_S( SRC3 ) /* F0 F5 F4 */ 5777117f1b4Smrg FMUL_S( MAT12 ) 5787117f1b4Smrg FLD_S( SRC3 ) /* F1 F0 F5 F4 */ 5797117f1b4Smrg FMUL_S( MAT13 ) 5807117f1b4Smrg 5817117f1b4Smrg FXCH( ST(1) ) /* F0 F1 F5 F4 */ 5827117f1b4Smrg FADDP( ST0, ST(3) ) /* F1 F5 F4 */ 5837117f1b4Smrg FADDP( ST0, ST(1) ) /* F5 F4 */ 5847117f1b4Smrg 5857117f1b4Smrg MOV_L( SRC2, EBX ) 5867117f1b4Smrg MOV_L( SRC3, EBP ) 5877117f1b4Smrg 5887117f1b4Smrg FXCH( ST(1) ) /* F4 F5 */ 5897117f1b4Smrg FSTP_S( DST0 ) /* F5 */ 5907117f1b4Smrg FSTP_S( DST1 ) /* */ 5917117f1b4Smrg MOV_L( EBX, DST2 ) 5927117f1b4Smrg MOV_L( EBP, DST3 ) 5937117f1b4Smrg 5947117f1b4SmrgLLBL(x86_p4_2dnrr_skip): 5957117f1b4Smrg 5967117f1b4Smrg ADD_L( CONST(16), EDI ) 5977117f1b4Smrg ADD_L( EAX, ESI ) 5987117f1b4Smrg CMP_L( ECX, EDI ) 5997117f1b4Smrg JNE( LLBL(x86_p4_2dnrr_loop) ) 6007117f1b4Smrg 6017117f1b4SmrgLLBL(x86_p4_2dnrr_done): 6027117f1b4Smrg 6037117f1b4Smrg POP_L( EBP ) 6047117f1b4Smrg POP_L( EBX ) 6057117f1b4Smrg POP_L( EDI ) 6067117f1b4Smrg POP_L( ESI ) 6077117f1b4Smrg RET 6087117f1b4Smrg#undef FRAME_OFFSET 6097117f1b4Smrg 6107117f1b4Smrg 6117117f1b4Smrg 6127117f1b4Smrg 6137117f1b4SmrgALIGNTEXT16 6147117f1b4SmrgGLOBL GLNAME( _mesa_x86_transform_points4_identity ) 6157117f1b4SmrgHIDDEN(_mesa_x86_transform_points4_identity) 6167117f1b4SmrgGLNAME( _mesa_x86_transform_points4_identity ): 6177117f1b4Smrg 6187117f1b4Smrg#define FRAME_OFFSET 12 6197117f1b4Smrg PUSH_L( ESI ) 6207117f1b4Smrg PUSH_L( EDI ) 6217117f1b4Smrg PUSH_L( EBX ) 6227117f1b4Smrg 6237117f1b4Smrg MOV_L( ARG_SOURCE, ESI ) 6247117f1b4Smrg MOV_L( ARG_DEST, EDI ) 6257117f1b4Smrg 6267117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) 6277117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 6287117f1b4Smrg 6297117f1b4Smrg TEST_L( ECX, ECX ) 6307117f1b4Smrg JZ( LLBL(x86_p4_ir_done) ) 6317117f1b4Smrg 6327117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 6337117f1b4Smrg OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 6347117f1b4Smrg 6357117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 6367117f1b4Smrg MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 6377117f1b4Smrg 6387117f1b4Smrg SHL_L( CONST(4), ECX ) 6397117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) 6407117f1b4Smrg 6417117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) 6427117f1b4Smrg ADD_L( EDI, ECX ) 6437117f1b4Smrg 6447117f1b4Smrg CMP_L( ESI, EDI ) 6457117f1b4Smrg JE( LLBL(x86_p4_ir_done) ) 6467117f1b4Smrg 6477117f1b4SmrgALIGNTEXT16 6487117f1b4SmrgLLBL(x86_p4_ir_loop): 6497117f1b4Smrg 6507117f1b4Smrg MOV_L( SRC0, EBX ) 6517117f1b4Smrg MOV_L( SRC1, EDX ) 6527117f1b4Smrg 6537117f1b4Smrg MOV_L( EBX, DST0 ) 6547117f1b4Smrg MOV_L( EDX, DST1 ) 6557117f1b4Smrg 6567117f1b4Smrg MOV_L( SRC2, EBX ) 6577117f1b4Smrg MOV_L( SRC3, EDX ) 6587117f1b4Smrg 6597117f1b4Smrg MOV_L( EBX, DST2 ) 6607117f1b4Smrg MOV_L( EDX, DST3 ) 6617117f1b4Smrg 6627117f1b4SmrgLLBL(x86_p4_ir_skip): 6637117f1b4Smrg 6647117f1b4Smrg ADD_L( CONST(16), EDI ) 6657117f1b4Smrg ADD_L( EAX, ESI ) 6667117f1b4Smrg CMP_L( ECX, EDI ) 6677117f1b4Smrg JNE( LLBL(x86_p4_ir_loop) ) 6687117f1b4Smrg 6697117f1b4SmrgLLBL(x86_p4_ir_done): 6707117f1b4Smrg 6717117f1b4Smrg POP_L( EBX ) 6727117f1b4Smrg POP_L( EDI ) 6737117f1b4Smrg POP_L( ESI ) 6747117f1b4Smrg RET 6757117f1b4Smrg 6767117f1b4Smrg#if defined (__ELF__) && defined (__linux__) 6777117f1b4Smrg .section .note.GNU-stack,"",%progbits 6787117f1b4Smrg#endif 679