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_points2_general ) 707117f1b4SmrgHIDDEN(_mesa_x86_transform_points2_general) 717117f1b4SmrgGLNAME( _mesa_x86_transform_points2_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_p2_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_p2_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 FXCH( ST(3) ) /* F4 F6 F5 F7 */ 1277117f1b4Smrg FADD_S( MAT12 ) 1287117f1b4Smrg FXCH( ST(2) ) /* F5 F6 F4 F7 */ 1297117f1b4Smrg FADD_S( MAT13 ) 1307117f1b4Smrg FXCH( ST(1) ) /* F6 F5 F4 F7 */ 1317117f1b4Smrg FADD_S( MAT14 ) 1327117f1b4Smrg FXCH( ST(3) ) /* F7 F5 F4 F6 */ 1337117f1b4Smrg FADD_S( MAT15 ) 1347117f1b4Smrg 1357117f1b4Smrg FXCH( ST(2) ) /* F4 F5 F7 F6 */ 1367117f1b4Smrg FSTP_S( DST0 ) /* F5 F7 F6 */ 1377117f1b4Smrg FSTP_S( DST1 ) /* F7 F6 */ 1387117f1b4Smrg FXCH( ST(1) ) /* F6 F7 */ 1397117f1b4Smrg FSTP_S( DST2 ) /* F7 */ 1407117f1b4Smrg FSTP_S( DST3 ) /* */ 1417117f1b4Smrg 1427117f1b4SmrgLLBL(x86_p2_gr_skip): 1437117f1b4Smrg 1447117f1b4Smrg ADD_L( CONST(16), EDI ) 1457117f1b4Smrg ADD_L( EAX, ESI ) 1467117f1b4Smrg CMP_L( ECX, EDI ) 1477117f1b4Smrg JNE( LLBL(x86_p2_gr_loop) ) 1487117f1b4Smrg 1497117f1b4SmrgLLBL(x86_p2_gr_done): 1507117f1b4Smrg 1517117f1b4Smrg POP_L( EDI ) 1527117f1b4Smrg POP_L( ESI ) 1537117f1b4Smrg RET 1547117f1b4Smrg#undef FRAME_OFFSET 1557117f1b4Smrg 1567117f1b4Smrg 1577117f1b4Smrg 1587117f1b4Smrg 1597117f1b4SmrgALIGNTEXT16 1607117f1b4SmrgGLOBL GLNAME( _mesa_x86_transform_points2_perspective ) 1617117f1b4SmrgHIDDEN(_mesa_x86_transform_points2_perspective) 1627117f1b4SmrgGLNAME( _mesa_x86_transform_points2_perspective ): 1637117f1b4Smrg 1647117f1b4Smrg#define FRAME_OFFSET 12 1657117f1b4Smrg PUSH_L( ESI ) 1667117f1b4Smrg PUSH_L( EDI ) 1677117f1b4Smrg PUSH_L( EBX ) 1687117f1b4Smrg 1697117f1b4Smrg MOV_L( ARG_SOURCE, ESI ) 1707117f1b4Smrg MOV_L( ARG_DEST, EDI ) 1717117f1b4Smrg 1727117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) 1737117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 1747117f1b4Smrg 1757117f1b4Smrg TEST_L( ECX, ECX ) 1767117f1b4Smrg JZ( LLBL(x86_p2_pr_done) ) 1777117f1b4Smrg 1787117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 1797117f1b4Smrg OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 1807117f1b4Smrg 1817117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 1827117f1b4Smrg MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 1837117f1b4Smrg 1847117f1b4Smrg SHL_L( CONST(4), ECX ) 1857117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) 1867117f1b4Smrg 1877117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) 1887117f1b4Smrg ADD_L( EDI, ECX ) 1897117f1b4Smrg 1907117f1b4Smrg MOV_L( MAT14, EBX ) 1917117f1b4Smrg 1927117f1b4SmrgALIGNTEXT16 1937117f1b4SmrgLLBL(x86_p2_pr_loop): 1947117f1b4Smrg 1957117f1b4Smrg FLD_S( SRC0 ) /* F4 */ 1967117f1b4Smrg FMUL_S( MAT0 ) 1977117f1b4Smrg 1987117f1b4Smrg FLD_S( SRC1 ) /* F1 F4 */ 1997117f1b4Smrg FMUL_S( MAT5 ) 2007117f1b4Smrg 2017117f1b4Smrg FXCH( ST(1) ) /* F4 F1 */ 2027117f1b4Smrg FSTP_S( DST0 ) /* F1 */ 2037117f1b4Smrg FSTP_S( DST1 ) /* */ 2047117f1b4Smrg MOV_L( EBX, DST2 ) 2057117f1b4Smrg MOV_L( CONST(FP_ZERO), DST3 ) 2067117f1b4Smrg 2077117f1b4SmrgLLBL(x86_p2_pr_skip): 2087117f1b4Smrg 2097117f1b4Smrg ADD_L( CONST(16), EDI ) 2107117f1b4Smrg ADD_L( EAX, ESI ) 2117117f1b4Smrg CMP_L( ECX, EDI ) 2127117f1b4Smrg JNE( LLBL(x86_p2_pr_loop) ) 2137117f1b4Smrg 2147117f1b4SmrgLLBL(x86_p2_pr_done): 2157117f1b4Smrg 2167117f1b4Smrg POP_L( EBX ) 2177117f1b4Smrg POP_L( EDI ) 2187117f1b4Smrg POP_L( ESI ) 2197117f1b4Smrg RET 2207117f1b4Smrg#undef FRAME_OFFSET 2217117f1b4Smrg 2227117f1b4Smrg 2237117f1b4Smrg 2247117f1b4Smrg 2257117f1b4SmrgALIGNTEXT16 2267117f1b4SmrgGLOBL GLNAME( _mesa_x86_transform_points2_3d ) 2277117f1b4SmrgHIDDEN(_mesa_x86_transform_points2_3d) 2287117f1b4SmrgGLNAME( _mesa_x86_transform_points2_3d ): 2297117f1b4Smrg 2307117f1b4Smrg#define FRAME_OFFSET 8 2317117f1b4Smrg PUSH_L( ESI ) 2327117f1b4Smrg PUSH_L( EDI ) 2337117f1b4Smrg 2347117f1b4Smrg MOV_L( ARG_SOURCE, ESI ) 2357117f1b4Smrg MOV_L( ARG_DEST, EDI ) 2367117f1b4Smrg 2377117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) 2387117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 2397117f1b4Smrg 2407117f1b4Smrg TEST_L( ECX, ECX ) 2417117f1b4Smrg JZ( LLBL(x86_p2_3dr_done) ) 2427117f1b4Smrg 2437117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 2447117f1b4Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 2457117f1b4Smrg 2467117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 2477117f1b4Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 2487117f1b4Smrg 2497117f1b4Smrg SHL_L( CONST(4), ECX ) 2507117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) 2517117f1b4Smrg 2527117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) 2537117f1b4Smrg ADD_L( EDI, ECX ) 2547117f1b4Smrg 2557117f1b4SmrgALIGNTEXT16 2567117f1b4SmrgLLBL(x86_p2_3dr_loop): 2577117f1b4Smrg 2587117f1b4Smrg FLD_S( SRC0 ) /* F4 */ 2597117f1b4Smrg FMUL_S( MAT0 ) 2607117f1b4Smrg FLD_S( SRC0 ) /* F5 F4 */ 2617117f1b4Smrg FMUL_S( MAT1 ) 2627117f1b4Smrg FLD_S( SRC0 ) /* F6 F5 F4 */ 2637117f1b4Smrg FMUL_S( MAT2 ) 2647117f1b4Smrg 2657117f1b4Smrg FLD_S( SRC1 ) /* F0 F6 F5 F4 */ 2667117f1b4Smrg FMUL_S( MAT4 ) 2677117f1b4Smrg FLD_S( SRC1 ) /* F1 F0 F6 F5 F4 */ 2687117f1b4Smrg FMUL_S( MAT5 ) 2697117f1b4Smrg FLD_S( SRC1 ) /* F2 F1 F0 F6 F5 F4 */ 2707117f1b4Smrg FMUL_S( MAT6 ) 2717117f1b4Smrg 2727117f1b4Smrg FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */ 2737117f1b4Smrg FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */ 2747117f1b4Smrg FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */ 2757117f1b4Smrg FADDP( ST0, ST(1) ) /* F6 F5 F4 */ 2767117f1b4Smrg 2777117f1b4Smrg FXCH( ST(2) ) /* F4 F5 F6 */ 2787117f1b4Smrg FADD_S( MAT12 ) 2797117f1b4Smrg FXCH( ST(1) ) /* F5 F4 F6 */ 2807117f1b4Smrg FADD_S( MAT13 ) 2817117f1b4Smrg FXCH( ST(2) ) /* F6 F4 F5 */ 2827117f1b4Smrg FADD_S( MAT14 ) 2837117f1b4Smrg 2847117f1b4Smrg FXCH( ST(1) ) /* F4 F6 F5 */ 2857117f1b4Smrg FSTP_S( DST0 ) /* F6 F5 */ 2867117f1b4Smrg FXCH( ST(1) ) /* F5 F6 */ 2877117f1b4Smrg FSTP_S( DST1 ) /* F6 */ 2887117f1b4Smrg FSTP_S( DST2 ) /* */ 2897117f1b4Smrg 2907117f1b4SmrgLLBL(x86_p2_3dr_skip): 2917117f1b4Smrg 2927117f1b4Smrg ADD_L( CONST(16), EDI ) 2937117f1b4Smrg ADD_L( EAX, ESI ) 2947117f1b4Smrg CMP_L( ECX, EDI ) 2957117f1b4Smrg JNE( LLBL(x86_p2_3dr_loop) ) 2967117f1b4Smrg 2977117f1b4SmrgLLBL(x86_p2_3dr_done): 2987117f1b4Smrg 2997117f1b4Smrg POP_L( EDI ) 3007117f1b4Smrg POP_L( ESI ) 3017117f1b4Smrg RET 3027117f1b4Smrg#undef FRAME_OFFSET 3037117f1b4Smrg 3047117f1b4Smrg 3057117f1b4Smrg 3067117f1b4Smrg 3077117f1b4SmrgALIGNTEXT16 3087117f1b4SmrgGLOBL GLNAME( _mesa_x86_transform_points2_3d_no_rot ) 3097117f1b4SmrgHIDDEN(_mesa_x86_transform_points2_3d_no_rot) 3107117f1b4SmrgGLNAME( _mesa_x86_transform_points2_3d_no_rot ): 3117117f1b4Smrg 3127117f1b4Smrg#define FRAME_OFFSET 12 3137117f1b4Smrg PUSH_L( ESI ) 3147117f1b4Smrg PUSH_L( EDI ) 3157117f1b4Smrg PUSH_L( EBX ) 3167117f1b4Smrg 3177117f1b4Smrg MOV_L( ARG_SOURCE, ESI ) 3187117f1b4Smrg MOV_L( ARG_DEST, EDI ) 3197117f1b4Smrg 3207117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) 3217117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 3227117f1b4Smrg 3237117f1b4Smrg TEST_L( ECX, ECX ) 3247117f1b4Smrg JZ( LLBL(x86_p2_3dnrr_done) ) 3257117f1b4Smrg 3267117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 3277117f1b4Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 3287117f1b4Smrg 3297117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 3307117f1b4Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 3317117f1b4Smrg 3327117f1b4Smrg SHL_L( CONST(4), ECX ) 3337117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) 3347117f1b4Smrg 3357117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) 3367117f1b4Smrg ADD_L( EDI, ECX ) 3377117f1b4Smrg 3387117f1b4Smrg MOV_L( MAT14, EBX ) 3397117f1b4Smrg 3407117f1b4SmrgALIGNTEXT16 3417117f1b4SmrgLLBL(x86_p2_3dnrr_loop): 3427117f1b4Smrg 3437117f1b4Smrg FLD_S( SRC0 ) /* F4 */ 3447117f1b4Smrg FMUL_S( MAT0 ) 3457117f1b4Smrg 3467117f1b4Smrg FLD_S( SRC1 ) /* F1 F4 */ 3477117f1b4Smrg FMUL_S( MAT5 ) 3487117f1b4Smrg 3497117f1b4Smrg FXCH( ST(1) ) /* F4 F1 */ 3507117f1b4Smrg FADD_S( MAT12 ) 3517117f1b4Smrg FLD_S( MAT13 ) /* F5 F4 F1 */ 3527117f1b4Smrg FXCH( ST(2) ) /* F1 F4 F5 */ 3537117f1b4Smrg FADDP( ST0, ST(2) ) /* F4 F5 */ 3547117f1b4Smrg 3557117f1b4Smrg FSTP_S( DST0 ) /* F5 */ 3567117f1b4Smrg FSTP_S( DST1 ) /* */ 3577117f1b4Smrg MOV_L( EBX, DST2 ) 3587117f1b4Smrg 3597117f1b4SmrgLLBL(x86_p2_3dnrr_skip): 3607117f1b4Smrg 3617117f1b4Smrg ADD_L( CONST(16), EDI ) 3627117f1b4Smrg ADD_L( EAX, ESI ) 3637117f1b4Smrg CMP_L( ECX, EDI ) 3647117f1b4Smrg JNE( LLBL(x86_p2_3dnrr_loop) ) 3657117f1b4Smrg 3667117f1b4SmrgLLBL(x86_p2_3dnrr_done): 3677117f1b4Smrg 3687117f1b4Smrg POP_L( EBX ) 3697117f1b4Smrg POP_L( EDI ) 3707117f1b4Smrg POP_L( ESI ) 3717117f1b4Smrg RET 3727117f1b4Smrg#undef FRAME_OFFSET 3737117f1b4Smrg 3747117f1b4Smrg 3757117f1b4Smrg 3767117f1b4Smrg 3777117f1b4SmrgALIGNTEXT16 3787117f1b4SmrgGLOBL GLNAME( _mesa_x86_transform_points2_2d ) 3797117f1b4SmrgHIDDEN(_mesa_x86_transform_points2_2d) 3807117f1b4SmrgGLNAME( _mesa_x86_transform_points2_2d ): 3817117f1b4Smrg 3827117f1b4Smrg#define FRAME_OFFSET 8 3837117f1b4Smrg PUSH_L( ESI ) 3847117f1b4Smrg PUSH_L( EDI ) 3857117f1b4Smrg 3867117f1b4Smrg MOV_L( ARG_SOURCE, ESI ) 3877117f1b4Smrg MOV_L( ARG_DEST, EDI ) 3887117f1b4Smrg 3897117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) 3907117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 3917117f1b4Smrg 3927117f1b4Smrg TEST_L( ECX, ECX ) 3937117f1b4Smrg JZ( LLBL(x86_p2_2dr_done) ) 3947117f1b4Smrg 3957117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 3967117f1b4Smrg OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) 3977117f1b4Smrg 3987117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 3997117f1b4Smrg MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) 4007117f1b4Smrg 4017117f1b4Smrg SHL_L( CONST(4), ECX ) 4027117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) 4037117f1b4Smrg 4047117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) 4057117f1b4Smrg ADD_L( EDI, ECX ) 4067117f1b4Smrg 4077117f1b4SmrgALIGNTEXT16 4087117f1b4SmrgLLBL(x86_p2_2dr_loop): 4097117f1b4Smrg 4107117f1b4Smrg FLD_S( SRC0 ) /* F4 */ 4117117f1b4Smrg FMUL_S( MAT0 ) 4127117f1b4Smrg FLD_S( SRC0 ) /* F5 F4 */ 4137117f1b4Smrg FMUL_S( MAT1 ) 4147117f1b4Smrg 4157117f1b4Smrg FLD_S( SRC1 ) /* F0 F5 F4 */ 4167117f1b4Smrg FMUL_S( MAT4 ) 4177117f1b4Smrg FLD_S( SRC1 ) /* F1 F0 F5 F4 */ 4187117f1b4Smrg FMUL_S( MAT5 ) 4197117f1b4Smrg 4207117f1b4Smrg FXCH( ST(1) ) /* F0 F1 F5 F4 */ 4217117f1b4Smrg FADDP( ST0, ST(3) ) /* F1 F5 F4 */ 4227117f1b4Smrg FADDP( ST0, ST(1) ) /* F5 F4 */ 4237117f1b4Smrg 4247117f1b4Smrg FXCH( ST(1) ) /* F4 F5 */ 4257117f1b4Smrg FADD_S( MAT12 ) 4267117f1b4Smrg FXCH( ST(1) ) /* F5 F4 */ 4277117f1b4Smrg FADD_S( MAT13 ) 4287117f1b4Smrg 4297117f1b4Smrg FXCH( ST(1) ) /* F4 F5 */ 4307117f1b4Smrg FSTP_S( DST0 ) /* F5 */ 4317117f1b4Smrg FSTP_S( DST1 ) /* */ 4327117f1b4Smrg 4337117f1b4SmrgLLBL(x86_p2_2dr_skip): 4347117f1b4Smrg 4357117f1b4Smrg ADD_L( CONST(16), EDI ) 4367117f1b4Smrg ADD_L( EAX, ESI ) 4377117f1b4Smrg CMP_L( ECX, EDI ) 4387117f1b4Smrg JNE( LLBL(x86_p2_2dr_loop) ) 4397117f1b4Smrg 4407117f1b4SmrgLLBL(x86_p2_2dr_done): 4417117f1b4Smrg 4427117f1b4Smrg POP_L( EDI ) 4437117f1b4Smrg POP_L( ESI ) 4447117f1b4Smrg RET 4457117f1b4Smrg#undef FRAME_OFFSET 4467117f1b4Smrg 4477117f1b4Smrg 4487117f1b4Smrg 4497117f1b4Smrg 4507117f1b4SmrgALIGNTEXT4 4517117f1b4SmrgGLOBL GLNAME( _mesa_x86_transform_points2_2d_no_rot ) 4527117f1b4SmrgHIDDEN(_mesa_x86_transform_points2_2d_no_rot) 4537117f1b4SmrgGLNAME( _mesa_x86_transform_points2_2d_no_rot ): 4547117f1b4Smrg 4557117f1b4Smrg#define FRAME_OFFSET 8 4567117f1b4Smrg PUSH_L( ESI ) 4577117f1b4Smrg PUSH_L( EDI ) 4587117f1b4Smrg 4597117f1b4Smrg MOV_L( ARG_SOURCE, ESI ) 4607117f1b4Smrg MOV_L( ARG_DEST, EDI ) 4617117f1b4Smrg 4627117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) 4637117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 4647117f1b4Smrg 4657117f1b4Smrg TEST_L( ECX, ECX ) 4667117f1b4Smrg JZ( LLBL(x86_p2_2dnrr_done) ) 4677117f1b4Smrg 4687117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 4697117f1b4Smrg OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) 4707117f1b4Smrg 4717117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 4727117f1b4Smrg MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) 4737117f1b4Smrg 4747117f1b4Smrg SHL_L( CONST(4), ECX ) 4757117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) 4767117f1b4Smrg 4777117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) 4787117f1b4Smrg ADD_L( EDI, ECX ) 4797117f1b4Smrg 4807117f1b4SmrgALIGNTEXT16 4817117f1b4SmrgLLBL(x86_p2_2dnrr_loop): 4827117f1b4Smrg 4837117f1b4Smrg FLD_S( SRC0 ) /* F4 */ 4847117f1b4Smrg FMUL_S( MAT0 ) 4857117f1b4Smrg 4867117f1b4Smrg FLD_S( SRC1 ) /* F1 F4 */ 4877117f1b4Smrg FMUL_S( MAT5 ) 4887117f1b4Smrg 4897117f1b4Smrg FXCH( ST(1) ) /* F4 F1 */ 4907117f1b4Smrg FADD_S( MAT12 ) 4917117f1b4Smrg FLD_S( MAT13 ) /* F5 F4 F1 */ 4927117f1b4Smrg FXCH( ST(2) ) /* F1 F4 F5 */ 4937117f1b4Smrg FADDP( ST0, ST(2) ) /* F4 F5 */ 4947117f1b4Smrg 4957117f1b4Smrg FSTP_S( DST0 ) /* F5 */ 4967117f1b4Smrg FSTP_S( DST1 ) /* */ 4977117f1b4Smrg 4987117f1b4SmrgLLBL(x86_p2_2dnrr_skip): 4997117f1b4Smrg 5007117f1b4Smrg ADD_L( CONST(16), EDI ) 5017117f1b4Smrg ADD_L( EAX, ESI ) 5027117f1b4Smrg CMP_L( ECX, EDI ) 5037117f1b4Smrg JNE( LLBL(x86_p2_2dnrr_loop) ) 5047117f1b4Smrg 5057117f1b4SmrgLLBL(x86_p2_2dnrr_done): 5067117f1b4Smrg 5077117f1b4Smrg POP_L( EDI ) 5087117f1b4Smrg POP_L( ESI ) 5097117f1b4Smrg RET 5107117f1b4Smrg#undef FRAME_OFFSET 5117117f1b4Smrg 5127117f1b4Smrg 5137117f1b4Smrg 5147117f1b4Smrg 5157117f1b4SmrgALIGNTEXT16 5167117f1b4SmrgGLOBL GLNAME( _mesa_x86_transform_points2_identity ) 5177117f1b4SmrgHIDDEN(_mesa_x86_transform_points2_identity) 5187117f1b4SmrgGLNAME( _mesa_x86_transform_points2_identity ): 5197117f1b4Smrg 5207117f1b4Smrg#define FRAME_OFFSET 12 5217117f1b4Smrg PUSH_L( ESI ) 5227117f1b4Smrg PUSH_L( EDI ) 5237117f1b4Smrg PUSH_L( EBX ) 5247117f1b4Smrg 5257117f1b4Smrg MOV_L( ARG_SOURCE, ESI ) 5267117f1b4Smrg MOV_L( ARG_DEST, EDI ) 5277117f1b4Smrg 5287117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) 5297117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 5307117f1b4Smrg 5317117f1b4Smrg TEST_L( ECX, ECX ) 5327117f1b4Smrg JZ( LLBL(x86_p2_ir_done) ) 5337117f1b4Smrg 5347117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 5357117f1b4Smrg OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) 5367117f1b4Smrg 5377117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 5387117f1b4Smrg MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) 5397117f1b4Smrg 5407117f1b4Smrg SHL_L( CONST(4), ECX ) 5417117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) 5427117f1b4Smrg 5437117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) 5447117f1b4Smrg ADD_L( EDI, ECX ) 5457117f1b4Smrg 5467117f1b4Smrg CMP_L( ESI, EDI ) 5477117f1b4Smrg JE( LLBL(x86_p2_ir_done) ) 5487117f1b4Smrg 5497117f1b4SmrgALIGNTEXT16 5507117f1b4SmrgLLBL(x86_p2_ir_loop): 5517117f1b4Smrg 5527117f1b4Smrg MOV_L( SRC0, EBX ) 5537117f1b4Smrg MOV_L( SRC1, EDX ) 5547117f1b4Smrg 5557117f1b4Smrg MOV_L( EBX, DST0 ) 5567117f1b4Smrg MOV_L( EDX, DST1 ) 5577117f1b4Smrg 5587117f1b4SmrgLLBL(x86_p2_ir_skip): 5597117f1b4Smrg 5607117f1b4Smrg ADD_L( CONST(16), EDI ) 5617117f1b4Smrg ADD_L( EAX, ESI ) 5627117f1b4Smrg CMP_L( ECX, EDI ) 5637117f1b4Smrg JNE( LLBL(x86_p2_ir_loop) ) 5647117f1b4Smrg 5657117f1b4SmrgLLBL(x86_p2_ir_done): 5667117f1b4Smrg 5677117f1b4Smrg POP_L( EBX ) 5687117f1b4Smrg POP_L( EDI ) 5697117f1b4Smrg POP_L( ESI ) 5707117f1b4Smrg RET 5717117f1b4Smrg#undef FRAME_OFFSET 5727117f1b4Smrg 5737117f1b4Smrg#if defined (__ELF__) && defined (__linux__) 5747117f1b4Smrg .section .note.GNU-stack,"",%progbits 5757117f1b4Smrg#endif 576