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/** TODO:
277117f1b4Smrg  * - insert PREFETCH instructions to avoid cache-misses !
287117f1b4Smrg  * - some more optimizations are possible...
297117f1b4Smrg  * - for 40-50% more performance in the SSE-functions, the
307117f1b4Smrg  *   data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned !
317117f1b4Smrg  */
327117f1b4Smrg
337117f1b4Smrg#ifdef USE_SSE_ASM
34c1f859d4Smrg#include "assyntax.h"
357ec681f3Smrg#define MATH_ASM_PTR_SIZE 4
367ec681f3Smrg#include "math/m_vector_asm.h"
377117f1b4Smrg#include "xform_args.h"
387117f1b4Smrg
397117f1b4Smrg   SEG_TEXT
407117f1b4Smrg
417117f1b4Smrg#define S(i) 	REGOFF(i * 4, ESI)
427117f1b4Smrg#define D(i) 	REGOFF(i * 4, EDI)
437117f1b4Smrg#define M(i) 	REGOFF(i * 4, EDX)
447117f1b4Smrg
457117f1b4Smrg
467117f1b4SmrgALIGNTEXT4
477117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points1_general)
487117f1b4SmrgHIDDEN( _mesa_sse_transform_points1_general )
497117f1b4SmrgGLNAME( _mesa_sse_transform_points1_general ):
507ec681f3Smrg    _CET_ENDBR
517117f1b4Smrg#define FRAME_OFFSET 8
527117f1b4Smrg    PUSH_L    ( ESI )
537117f1b4Smrg    PUSH_L    ( EDI )
547117f1b4Smrg
557117f1b4Smrg    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
567117f1b4Smrg    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
577117f1b4Smrg
587117f1b4Smrg    MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
597117f1b4Smrg    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
607117f1b4Smrg
617117f1b4Smrg    CMP_L( CONST(0), ECX )			/* count == 0 ? */
627117f1b4Smrg    JE( LLBL(K_GTP1GR_finish) )			/* yes -> nothing to do. */
637117f1b4Smrg
647117f1b4Smrg    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
657117f1b4Smrg    OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
667117f1b4Smrg
677117f1b4Smrg    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
687117f1b4Smrg    MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
697117f1b4Smrg
707117f1b4Smrg    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
717117f1b4Smrg    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
727117f1b4Smrg
737117f1b4Smrg    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
747117f1b4Smrg    ADD_L( EDI, ECX ) 				/* count += dest ptr */
757117f1b4Smrg
767117f1b4Smrg
777117f1b4SmrgALIGNTEXT32
787117f1b4Smrg    MOVAPS( M(0), XMM0 )			/* m3  | m2  | m1  | m0  */
797117f1b4Smrg    MOVAPS( M(12), XMM1 )			/* m15 | m14 | m13 | m12 */
807117f1b4Smrg
817117f1b4SmrgALIGNTEXT32
827117f1b4SmrgLLBL(K_GTP1GR_top):
837117f1b4Smrg    MOVSS( S(0), XMM2 )				/* ox */
847117f1b4Smrg    SHUFPS( CONST(0x0), XMM2, XMM2 )		/* ox | ox | ox | ox */
857117f1b4Smrg    MULPS( XMM0, XMM2 )				/* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
867117f1b4Smrg    ADDPS( XMM1, XMM2 )				/* + | + | + | + */
877117f1b4Smrg    MOVUPS( XMM2, D(0) )
887117f1b4Smrg
897117f1b4SmrgLLBL(K_GTP1GR_skip):
907117f1b4Smrg    ADD_L     ( CONST(16), EDI )
917117f1b4Smrg    ADD_L     ( EAX, ESI )
927117f1b4Smrg    CMP_L     ( ECX, EDI )
937117f1b4Smrg    JNE       ( LLBL(K_GTP1GR_top) )
947117f1b4Smrg
957117f1b4SmrgLLBL(K_GTP1GR_finish):
967117f1b4Smrg    POP_L     ( EDI )
977117f1b4Smrg    POP_L     ( ESI )
987117f1b4Smrg    RET
997117f1b4Smrg#undef FRAME_OFFSET
1007117f1b4Smrg
1017117f1b4Smrg
1027117f1b4Smrg
1037117f1b4SmrgALIGNTEXT4
1047117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points1_identity)
1057117f1b4SmrgHIDDEN(_mesa_sse_transform_points1_identity)
1067117f1b4SmrgGLNAME( _mesa_sse_transform_points1_identity ):
1077ec681f3Smrg    _CET_ENDBR
1087117f1b4Smrg#define FRAME_OFFSET 8
1097117f1b4Smrg    PUSH_L    ( ESI )
1107117f1b4Smrg    PUSH_L    ( EDI )
1117117f1b4Smrg
1127117f1b4Smrg    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
1137117f1b4Smrg    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
1147117f1b4Smrg
1157117f1b4Smrg    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
1167117f1b4Smrg
1177117f1b4Smrg    TEST_L( ECX, ECX)
1187117f1b4Smrg    JZ( LLBL(K_GTP1IR_finish) ) 		/* count was zero; go to finish */
1197117f1b4Smrg
1207117f1b4Smrg    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
1217117f1b4Smrg    OR_L( CONST(VEC_SIZE_1), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
1227117f1b4Smrg
1237117f1b4Smrg    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
1247117f1b4Smrg    MOV_L( CONST(1), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
1257117f1b4Smrg
1267117f1b4Smrg    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
1277117f1b4Smrg    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
1287117f1b4Smrg
1297117f1b4Smrg    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
1307117f1b4Smrg    ADD_L( EDI, ECX ) 				/* count += dest ptr */
1317117f1b4Smrg
1327117f1b4Smrg    CMP_L( ESI, EDI )
1337117f1b4Smrg    JE( LLBL(K_GTP1IR_finish) )
1347117f1b4Smrg
1357117f1b4Smrg
1367117f1b4SmrgALIGNTEXT32
1377117f1b4SmrgLLBL(K_GTP1IR_top):
1387117f1b4Smrg    MOV_L( S(0), EDX )
1397117f1b4Smrg    MOV_L( EDX, D(0) )
1407117f1b4Smrg
1417117f1b4SmrgLLBL(K_GTP1IR_skip):
1427117f1b4Smrg    ADD_L     ( CONST(16), EDI )
1437117f1b4Smrg    ADD_L     ( EAX, ESI )
1447117f1b4Smrg    CMP_L     ( ECX, EDI )
1457117f1b4Smrg    JNE       ( LLBL(K_GTP1IR_top) )
1467117f1b4Smrg
1477117f1b4SmrgLLBL(K_GTP1IR_finish):
1487117f1b4Smrg    POP_L     ( EDI )
1497117f1b4Smrg    POP_L     ( ESI )
1507117f1b4Smrg    RET
1517117f1b4Smrg#undef FRAME_OFFSET
1527117f1b4Smrg
1537117f1b4Smrg
1547117f1b4Smrg
1557117f1b4SmrgALIGNTEXT4
1567117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points1_3d_no_rot)
1577117f1b4SmrgHIDDEN(_mesa_sse_transform_points1_3d_no_rot)
1587117f1b4SmrgGLNAME(_mesa_sse_transform_points1_3d_no_rot):
1597ec681f3Smrg    _CET_ENDBR
1607117f1b4Smrg#define FRAME_OFFSET 8
1617117f1b4Smrg    PUSH_L( ESI )
1627117f1b4Smrg    PUSH_L( EDI )
1637117f1b4Smrg
1647117f1b4Smrg    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI )	/* ptr to source GLvector4f */
1657117f1b4Smrg    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
1667117f1b4Smrg
1677117f1b4Smrg    MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
1687117f1b4Smrg    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
1697117f1b4Smrg
1707117f1b4Smrg    TEST_L( ECX, ECX)
1717117f1b4Smrg    JZ( LLBL(K_GTP13DNRR_finish) ) 		/* count was zero; go to finish */
1727117f1b4Smrg
1737117f1b4Smrg    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
1747117f1b4Smrg    OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
1757117f1b4Smrg
1767117f1b4Smrg    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
1777117f1b4Smrg    MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
1787117f1b4Smrg
1797117f1b4Smrg    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
1807117f1b4Smrg    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
1817117f1b4Smrg
1827117f1b4Smrg    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
1837117f1b4Smrg    ADD_L( EDI, ECX ) 				/* count += dest ptr */
1847117f1b4Smrg
1857117f1b4Smrg
1867117f1b4SmrgALIGNTEXT32
1877117f1b4Smrg    MOVSS( M(0), XMM0 )				/* m0 */
1887117f1b4Smrg    MOVSS( M(12), XMM1 )			/* m12 */
1897117f1b4Smrg    MOVSS( M(13), XMM2 )			/* m13 */
1907117f1b4Smrg    MOVSS( M(14), XMM3 )			/* m14 */
1917117f1b4Smrg
1927117f1b4SmrgALIGNTEXT32
1937117f1b4SmrgLLBL(K_GTP13DNRR_top):
1947117f1b4Smrg    MOVSS( S(0), XMM4 )				/* ox */
1957117f1b4Smrg    MULSS( XMM0, XMM4 )				/* ox*m0 */
1967117f1b4Smrg    ADDSS( XMM1, XMM4 )				/* ox*m0+m12 */
1977117f1b4Smrg    MOVSS( XMM4, D(0) )
1987117f1b4Smrg
1997117f1b4Smrg    MOVSS( XMM2, D(1) )
2007117f1b4Smrg    MOVSS( XMM3, D(2) )
2017117f1b4Smrg
2027117f1b4SmrgLLBL(K_GTP13DNRR_skip):
2037117f1b4Smrg    ADD_L    ( CONST(16), EDI )
2047117f1b4Smrg    ADD_L    ( EAX, ESI )
2057117f1b4Smrg    CMP_L    ( ECX, EDI )
2067117f1b4Smrg    JNE      ( LLBL(K_GTP13DNRR_top) )
2077117f1b4Smrg
2087117f1b4SmrgLLBL(K_GTP13DNRR_finish):
2097117f1b4Smrg    POP_L    ( EDI )
2107117f1b4Smrg    POP_L    ( ESI )
2117117f1b4Smrg    RET
2127117f1b4Smrg#undef FRAME_OFFSET
2137117f1b4Smrg
2147117f1b4Smrg
2157117f1b4Smrg
2167117f1b4SmrgALIGNTEXT4
2177117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points1_perspective)
2187117f1b4SmrgHIDDEN(_mesa_sse_transform_points1_perspective)
2197117f1b4SmrgGLNAME(_mesa_sse_transform_points1_perspective):
2207ec681f3Smrg    _CET_ENDBR
2217117f1b4Smrg#define FRAME_OFFSET 8
2227117f1b4Smrg    PUSH_L   ( ESI )
2237117f1b4Smrg    PUSH_L   ( EDI )
2247117f1b4Smrg
2257117f1b4Smrg    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
2267117f1b4Smrg    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
2277117f1b4Smrg
2287117f1b4Smrg    MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
2297117f1b4Smrg    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
2307117f1b4Smrg
2317117f1b4Smrg    TEST_L( ECX, ECX)
2327117f1b4Smrg    JZ( LLBL(K_GTP13PR_finish) )		/* count was zero; go to finish */
2337117f1b4Smrg
2347117f1b4Smrg    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
2357117f1b4Smrg    OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
2367117f1b4Smrg
2377117f1b4Smrg    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
2387117f1b4Smrg    MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
2397117f1b4Smrg
2407117f1b4Smrg    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
2417117f1b4Smrg    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
2427117f1b4Smrg
2437117f1b4Smrg    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
2447117f1b4Smrg    ADD_L( EDI, ECX ) 				/* count += dest ptr */
2457117f1b4Smrg
2467117f1b4Smrg
2477117f1b4SmrgALIGNTEXT32
2487117f1b4Smrg    XORPS( XMM0, XMM0 )				/* 0 | 0 | 0 | 0 */
2497117f1b4Smrg    MOVSS( M(0), XMM1 )				/* m0 */
2507117f1b4Smrg    MOVSS( M(14), XMM2 )			/* m14 */
2517117f1b4Smrg
2527117f1b4SmrgALIGNTEXT32
2537117f1b4SmrgLLBL(K_GTP13PR_top):
2547117f1b4Smrg    MOVSS( S(0), XMM3 )				/* ox */
2557117f1b4Smrg    MULSS( XMM1, XMM3 )				/* ox*m0 */
2567117f1b4Smrg    MOVSS( XMM3, D(0) )				/* ox*m0->D(0) */
2577117f1b4Smrg    MOVSS( XMM2, D(2) )				/* m14->D(2) */
2587117f1b4Smrg
2597117f1b4Smrg    MOVSS( XMM0, D(1) )
2607117f1b4Smrg    MOVSS( XMM0, D(3) )
2617117f1b4Smrg
2627117f1b4SmrgLLBL(K_GTP13PR_skip):
2637117f1b4Smrg    ADD_L( CONST(16), EDI )
2647117f1b4Smrg    ADD_L( EAX, ESI )
2657117f1b4Smrg    CMP_L( ECX, EDI )
2667117f1b4Smrg    JNE( LLBL(K_GTP13PR_top) )
2677117f1b4Smrg
2687117f1b4SmrgLLBL(K_GTP13PR_finish):
2697117f1b4Smrg    POP_L    ( EDI )
2707117f1b4Smrg    POP_L    ( ESI )
2717117f1b4Smrg    RET
2727117f1b4Smrg#undef FRAME_OFFSET
2737117f1b4Smrg
2747117f1b4Smrg
2757117f1b4SmrgALIGNTEXT4
2767117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points1_2d)
2777117f1b4SmrgHIDDEN(_mesa_sse_transform_points1_2d)
2787117f1b4SmrgGLNAME(_mesa_sse_transform_points1_2d):
2797ec681f3Smrg    _CET_ENDBR
2807117f1b4Smrg#define FRAME_OFFSET 8
2817117f1b4Smrg    PUSH_L( ESI )
2827117f1b4Smrg    PUSH_L( EDI )
2837117f1b4Smrg
2847117f1b4Smrg    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
2857117f1b4Smrg    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
2867117f1b4Smrg
2877117f1b4Smrg    MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
2887117f1b4Smrg    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
2897117f1b4Smrg
2907117f1b4Smrg    TEST_L( ECX, ECX)
2917117f1b4Smrg    JZ( LLBL(K_GTP13P2DR_finish) ) 		/* count was zero; go to finish */
2927117f1b4Smrg
2937117f1b4Smrg    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
2947117f1b4Smrg    OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
2957117f1b4Smrg
2967117f1b4Smrg    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
2977117f1b4Smrg    MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
2987117f1b4Smrg
2997117f1b4Smrg    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
3007117f1b4Smrg    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
3017117f1b4Smrg
3027117f1b4Smrg    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
3037117f1b4Smrg    ADD_L( EDI, ECX ) 				/* count += dest ptr */
3047117f1b4Smrg
3057117f1b4SmrgALIGNTEXT32
3067117f1b4Smrg    MOVLPS( M(0), XMM0 )			/* m1  | m0  */
3077117f1b4Smrg    MOVLPS( M(12), XMM1 )			/* m13 | m12 */
3087117f1b4Smrg
3097117f1b4SmrgALIGNTEXT32
3107117f1b4SmrgLLBL(K_GTP13P2DR_top):
3117117f1b4Smrg    MOVSS( S(0), XMM2 )				/* ox */
3127117f1b4Smrg    SHUFPS( CONST(0x0), XMM2, XMM2 )		/* ox | ox | ox | ox */
3137117f1b4Smrg    MULPS( XMM0, XMM2 )				/* - | - | ox*m1 | ox*m0 */
3147117f1b4Smrg    ADDPS( XMM1, XMM2 )				/* - | - | ox*m1+m13 | ox*m0+m12 */
3157117f1b4Smrg    MOVLPS( XMM2, D(0) )
3167117f1b4Smrg
3177117f1b4SmrgLLBL(K_GTP13P2DR_skip):
3187117f1b4Smrg    ADD_L    ( CONST(16), EDI )
3197117f1b4Smrg    ADD_L    ( EAX, ESI )
3207117f1b4Smrg    CMP_L    ( ECX, EDI )
3217117f1b4Smrg    JNE      ( LLBL(K_GTP13P2DR_top) )
3227117f1b4Smrg
3237117f1b4SmrgLLBL(K_GTP13P2DR_finish):
3247117f1b4Smrg    POP_L    ( EDI )
3257117f1b4Smrg    POP_L    ( ESI )
3267117f1b4Smrg    RET
3277117f1b4Smrg#undef FRAME_OFFSET
3287117f1b4Smrg
3297117f1b4Smrg
3307117f1b4SmrgALIGNTEXT4
3317117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points1_2d_no_rot)
3327117f1b4SmrgHIDDEN(_mesa_sse_transform_points1_2d_no_rot)
3337117f1b4SmrgGLNAME(_mesa_sse_transform_points1_2d_no_rot):
3347ec681f3Smrg	_CET_ENDBR
3357117f1b4Smrg#define FRAME_OFFSET 8
3367117f1b4Smrg	PUSH_L( ESI )
3377117f1b4Smrg	PUSH_L( EDI )
3387117f1b4Smrg
3397117f1b4Smrg	MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
3407117f1b4Smrg	MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
3417117f1b4Smrg
3427117f1b4Smrg	MOV_L( ARG_MATRIX, EDX ) 		/* ptr to matrix */
3437117f1b4Smrg	MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
3447117f1b4Smrg
3457117f1b4Smrg	TEST_L( ECX, ECX)
3467117f1b4Smrg	JZ( LLBL(K_GTP13P2DNRR_finish) ) 	/* count was zero; go to finish */
3477117f1b4Smrg
3487117f1b4Smrg	MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
3497117f1b4Smrg	OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
3507117f1b4Smrg
3517117f1b4Smrg	MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
3527117f1b4Smrg	MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
3537117f1b4Smrg
3547117f1b4Smrg	SHL_L( CONST(4), ECX ) 			/* count *= 16 */
3557117f1b4Smrg	MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
3567117f1b4Smrg
3577117f1b4Smrg	MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
3587117f1b4Smrg	ADD_L( EDI, ECX ) 			/* count += dest ptr */
3597117f1b4Smrg
3607117f1b4SmrgALIGNTEXT32
3617117f1b4Smrg	MOVSS( M(0), XMM0 )			/* m0 */
3627117f1b4Smrg	MOVSS( M(12), XMM1 )			/* m12 */
3637117f1b4Smrg	MOVSS( M(13), XMM2 )			/* m13 */
3647117f1b4Smrg
3657117f1b4SmrgALIGNTEXT32
3667117f1b4SmrgLLBL(K_GTP13P2DNRR_top):
3677117f1b4Smrg	MOVSS( S(0), XMM3 )			/* ox */
3687117f1b4Smrg	MULSS( XMM0, XMM3 )			/* ox*m0 */
3697117f1b4Smrg	ADDSS( XMM1, XMM3 )			/* ox*m0+m12 */
3707117f1b4Smrg	MOVSS( XMM3, D(0) )
3717117f1b4Smrg	MOVSS( XMM2, D(1) )
3727117f1b4Smrg
3737117f1b4SmrgLLBL(K_GTP13P2DNRR_skip):
3747117f1b4Smrg	ADD_L( CONST(16), EDI )
3757117f1b4Smrg	ADD_L( EAX, ESI )
3767117f1b4Smrg	CMP_L( ECX, EDI )
3777117f1b4Smrg	JNE( LLBL(K_GTP13P2DNRR_top) )
3787117f1b4Smrg
3797117f1b4SmrgLLBL(K_GTP13P2DNRR_finish):
3807117f1b4Smrg	POP_L( EDI )
3817117f1b4Smrg	POP_L( ESI )
3827117f1b4Smrg	RET
3837117f1b4Smrg#undef FRAME_OFFSET
3847117f1b4Smrg
3857117f1b4Smrg
3867117f1b4Smrg
3877117f1b4SmrgALIGNTEXT4
3887117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points1_3d)
3897117f1b4SmrgHIDDEN(_mesa_sse_transform_points1_3d)
3907117f1b4SmrgGLNAME(_mesa_sse_transform_points1_3d):
3917ec681f3Smrg	_CET_ENDBR
3927117f1b4Smrg#define FRAME_OFFSET 8
3937117f1b4Smrg	PUSH_L( ESI )
3947117f1b4Smrg	PUSH_L( EDI )
3957117f1b4Smrg
3967117f1b4Smrg	MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
3977117f1b4Smrg	MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
3987117f1b4Smrg
3997117f1b4Smrg	MOV_L( ARG_MATRIX, EDX ) 		/* ptr to matrix */
4007117f1b4Smrg	MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
4017117f1b4Smrg
4027117f1b4Smrg	TEST_L( ECX, ECX)
4037117f1b4Smrg	JZ( LLBL(K_GTP13P3DR_finish) ) 	/* count was zero; go to finish */
4047117f1b4Smrg
4057117f1b4Smrg	MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
4067117f1b4Smrg	OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
4077117f1b4Smrg
4087117f1b4Smrg	MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
4097117f1b4Smrg	MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
4107117f1b4Smrg
4117117f1b4Smrg	SHL_L( CONST(4), ECX ) 			/* count *= 16 */
4127117f1b4Smrg	MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
4137117f1b4Smrg
4147117f1b4Smrg	MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
4157117f1b4Smrg	ADD_L( EDI, ECX ) 			/* count += dest ptr */
4167117f1b4Smrg
4177117f1b4Smrg
4187117f1b4SmrgALIGNTEXT32
4197117f1b4Smrg	MOVAPS( M(0), XMM0 )			/* m3  | m2  | m1  |  m0 */
4207117f1b4Smrg	MOVAPS( M(12), XMM1 )			/* m15 | m14 | m13 | m12 */
4217117f1b4Smrg
4227117f1b4SmrgALIGNTEXT32
4237117f1b4SmrgLLBL(K_GTP13P3DR_top):
4247117f1b4Smrg	MOVSS( S(0), XMM2 )			/* ox */
4257117f1b4Smrg	SHUFPS( CONST(0x0), XMM2, XMM2 )	/* ox | ox | ox | ox */
4267117f1b4Smrg	MULPS( XMM0, XMM2 )			/* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
4277117f1b4Smrg	ADDPS( XMM1, XMM2 )			/* +m15  | +m14  | +m13  | +m12  */
4287117f1b4Smrg	MOVLPS( XMM2, D(0) )			/*   -   |   -   | ->D(1)| ->D(0)*/
4297117f1b4Smrg	UNPCKHPS( XMM2, XMM2 )			/* ox*m3+m15 | ox*m3+m15 | ox*m2+m14 | ox*m2+m14 */
4307117f1b4Smrg	MOVSS( XMM2, D(2) )
4317117f1b4Smrg
4327117f1b4SmrgLLBL(K_GTP13P3DR_skip):
4337117f1b4Smrg	ADD_L( CONST(16), EDI )
4347117f1b4Smrg	ADD_L( EAX, ESI )
4357117f1b4Smrg	CMP_L( ECX, EDI )
4367117f1b4Smrg	JNE( LLBL(K_GTP13P3DR_top) )
4377117f1b4Smrg
4387117f1b4SmrgLLBL(K_GTP13P3DR_finish):
4397117f1b4Smrg	POP_L( EDI )
4407117f1b4Smrg	POP_L( ESI )
4417117f1b4Smrg	RET
4427117f1b4Smrg#undef FRAME_OFFSET
4437117f1b4Smrg#endif
4447117f1b4Smrg
4457117f1b4Smrg#if defined (__ELF__) && defined (__linux__)
4467117f1b4Smrg	.section .note.GNU-stack,"",%progbits
4477117f1b4Smrg#endif
448