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_points2_general)
487117f1b4SmrgHIDDEN (_mesa_sse_transform_points2_general)
497117f1b4SmrgGLNAME( _mesa_sse_transform_points2_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    TEST_L( ECX, ECX )
627117f1b4Smrg    JZ( LLBL(K_GTP2GR_finish) )			/* count was zero; go to finish */
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
767117f1b4SmrgALIGNTEXT32
777117f1b4Smrg    MOVAPS( M(0), XMM0 )			/* m3  | m2  | m1  | m0 */
787117f1b4Smrg    MOVAPS( M(4), XMM1 )			/* m7  | m6  | m5  | m4 */
797117f1b4Smrg    MOVAPS( M(12), XMM2 )			/* m15 | m14 | m13 | m12 */
807117f1b4Smrg
817117f1b4SmrgALIGNTEXT32
827117f1b4SmrgLLBL(K_GTP2GR_top):
837117f1b4Smrg    MOVSS( S(0), XMM3 )				/* ox */
847117f1b4Smrg    SHUFPS( CONST(0x0), XMM3, XMM3 )		/* ox | ox | ox | ox */
857117f1b4Smrg    MULPS( XMM0, XMM3 )				/* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
867117f1b4Smrg    MOVSS( S(1), XMM4 )				/* oy */
877117f1b4Smrg    SHUFPS( CONST(0x0), XMM4, XMM4 )		/* oy | oy | oy | oy */
887117f1b4Smrg    MULPS( XMM1, XMM4 )				/* oy*m7 | oy*m6 | oy*m5 | oy*m4 */
897117f1b4Smrg
907117f1b4Smrg    ADDPS( XMM4, XMM3 )
917117f1b4Smrg    ADDPS( XMM2, XMM3 )
927117f1b4Smrg    MOVAPS( XMM3, D(0) )
937117f1b4Smrg
947117f1b4SmrgLLBL(K_GTP2GR_skip):
957117f1b4Smrg    ADD_L     ( CONST(16), EDI )
967117f1b4Smrg    ADD_L     ( EAX, ESI )
977117f1b4Smrg    CMP_L     ( ECX, EDI )
987117f1b4Smrg    JNE       ( LLBL(K_GTP2GR_top) )
997117f1b4Smrg
1007117f1b4SmrgLLBL(K_GTP2GR_finish):
1017117f1b4Smrg    POP_L     ( EDI )
1027117f1b4Smrg    POP_L     ( ESI )
1037117f1b4Smrg    RET
1047117f1b4Smrg#undef FRAME_OFFSET
1057117f1b4Smrg
1067117f1b4Smrg
1077117f1b4SmrgALIGNTEXT4
1087117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points2_identity)
1097117f1b4SmrgHIDDEN(_mesa_sse_transform_points2_identity)
1107117f1b4SmrgGLNAME( _mesa_sse_transform_points2_identity ):
1117ec681f3Smrg    _CET_ENDBR
1127117f1b4Smrg#define FRAME_OFFSET 8
1137117f1b4Smrg    PUSH_L    ( ESI )
1147117f1b4Smrg    PUSH_L    ( EDI )
1157117f1b4Smrg
1167117f1b4Smrg    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
1177117f1b4Smrg    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
1187117f1b4Smrg
1197117f1b4Smrg    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
1207117f1b4Smrg
1217117f1b4Smrg    TEST_L( ECX, ECX)
1227117f1b4Smrg    JZ( LLBL(K_GTP2IR_finish) )			/* count was zero; go to finish */
1237117f1b4Smrg
1247117f1b4Smrg    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
1257117f1b4Smrg    OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
1267117f1b4Smrg
1277117f1b4Smrg    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
1287117f1b4Smrg    MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
1297117f1b4Smrg
1307117f1b4Smrg    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
1317117f1b4Smrg    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
1327117f1b4Smrg
1337117f1b4Smrg    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
1347117f1b4Smrg    ADD_L( EDI, ECX ) 				/* count += dest ptr */
1357117f1b4Smrg
1367117f1b4Smrg    CMP_L( ESI, EDI )
1377117f1b4Smrg    JE( LLBL(K_GTP2IR_finish) )
1387117f1b4Smrg
1397117f1b4Smrg
1407117f1b4SmrgALIGNTEXT32
1417117f1b4SmrgLLBL(K_GTP2IR_top):
1427117f1b4Smrg    MOV_L     ( S(0), EDX )
1437117f1b4Smrg    MOV_L     ( EDX, D(0) )
1447117f1b4Smrg    MOV_L     ( S(1), EDX )
1457117f1b4Smrg    MOV_L     ( EDX, D(1) )
1467117f1b4Smrg
1477117f1b4SmrgLLBL(K_GTP2IR_skip):
1487117f1b4Smrg    ADD_L     ( CONST(16), EDI )
1497117f1b4Smrg    ADD_L     ( EAX, ESI )
1507117f1b4Smrg    CMP_L     ( ECX, EDI )
1517117f1b4Smrg    JNE       ( LLBL(K_GTP2IR_top) )
1527117f1b4Smrg
1537117f1b4SmrgLLBL(K_GTP2IR_finish):
1547117f1b4Smrg    POP_L     ( EDI )
1557117f1b4Smrg    POP_L     ( ESI )
1567117f1b4Smrg    RET
1577117f1b4Smrg#undef FRAME_OFFSET
1587117f1b4Smrg
1597117f1b4Smrg
1607117f1b4SmrgALIGNTEXT4
1617117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points2_3d_no_rot)
1627117f1b4SmrgHIDDEN(_mesa_sse_transform_points2_3d_no_rot)
1637117f1b4SmrgGLNAME(_mesa_sse_transform_points2_3d_no_rot):
1647ec681f3Smrg    _CET_ENDBR
1657117f1b4Smrg#define FRAME_OFFSET 8
1667117f1b4Smrg    PUSH_L( ESI )
1677117f1b4Smrg    PUSH_L( EDI )
1687117f1b4Smrg
1697117f1b4Smrg    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
1707117f1b4Smrg    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
1717117f1b4Smrg
1727117f1b4Smrg    MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
1737117f1b4Smrg    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
1747117f1b4Smrg
1757117f1b4Smrg    TEST_L( ECX, ECX)
1767117f1b4Smrg    JZ( LLBL(K_GTP23DNRR_finish) ) 		/* count was zero; go to finish */
1777117f1b4Smrg
1787117f1b4Smrg    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
1797117f1b4Smrg    OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
1807117f1b4Smrg
1817117f1b4Smrg    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
1827117f1b4Smrg    MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
1837117f1b4Smrg
1847117f1b4Smrg    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
1857117f1b4Smrg    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
1867117f1b4Smrg
1877117f1b4Smrg    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
1887117f1b4Smrg    ADD_L( EDI, ECX ) 				/* count += dest ptr */
1897117f1b4Smrg
1904a49301eSmrg    XORPS( XMM0, XMM0 )                         /* clean the working register */
1917117f1b4Smrg
1927117f1b4SmrgALIGNTEXT32
1937117f1b4Smrg    MOVSS    ( M(0), XMM1 )			/* - | - |  -  | m0  */
1947117f1b4Smrg    MOVSS    ( M(5), XMM2 )			/* - | - |  -  | m5  */
1957117f1b4Smrg    UNPCKLPS ( XMM2, XMM1 )			/* - | - | m5  | m0  */
1967117f1b4Smrg    MOVLPS   ( M(12), XMM2 )			/* - | - | m13 | m12 */
1977117f1b4Smrg    MOVSS    ( M(14), XMM3 )			/* - | - |  -  | m14 */
1987117f1b4Smrg
1997117f1b4SmrgALIGNTEXT32
2007117f1b4SmrgLLBL(K_GTP23DNRR_top):
2017117f1b4Smrg    MOVLPS   ( S(0), XMM0 )			/* - | - |  oy   | ox */
2027117f1b4Smrg    MULPS    ( XMM1, XMM0 )			/* - | - | oy*m5 | ox*m0 */
2037117f1b4Smrg    ADDPS    ( XMM2, XMM0 )			/* - | - | +m13  | +m12 */
2047117f1b4Smrg    MOVLPS   ( XMM0, D(0) )			/* -> D(1) | -> D(0) */
2057117f1b4Smrg
2067117f1b4Smrg    MOVSS    ( XMM3, D(2) )			/* -> D(2) */
2077117f1b4Smrg
2087117f1b4SmrgLLBL(K_GTP23DNRR_skip):
2097117f1b4Smrg    ADD_L    ( CONST(16), EDI )
2107117f1b4Smrg    ADD_L    ( EAX, ESI )
2117117f1b4Smrg    CMP_L    ( ECX, EDI )
2127117f1b4Smrg    JNE      ( LLBL(K_GTP23DNRR_top) )
2137117f1b4Smrg
2147117f1b4SmrgLLBL(K_GTP23DNRR_finish):
2157117f1b4Smrg    POP_L    ( EDI )
2167117f1b4Smrg    POP_L    ( ESI )
2177117f1b4Smrg    RET
2187117f1b4Smrg#undef FRAME_OFFSET
2197117f1b4Smrg
2207117f1b4Smrg
2217117f1b4SmrgALIGNTEXT4
2227117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points2_perspective)
2237117f1b4SmrgHIDDEN(_mesa_sse_transform_points2_perspective)
2247117f1b4SmrgGLNAME(_mesa_sse_transform_points2_perspective):
2257ec681f3Smrg    _CET_ENDBR
2267117f1b4Smrg#define FRAME_OFFSET 8
2277117f1b4Smrg    PUSH_L   ( ESI )
2287117f1b4Smrg    PUSH_L   ( EDI )
2297117f1b4Smrg
2307117f1b4Smrg    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI )	/* ptr to source GLvector4f */
2317117f1b4Smrg    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
2327117f1b4Smrg
2337117f1b4Smrg    MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
2347117f1b4Smrg    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
2357117f1b4Smrg
2367117f1b4Smrg    TEST_L( ECX, ECX)
2377117f1b4Smrg    JZ( LLBL(K_GTP23PR_finish) )		/* count was zero; go to finish */
2387117f1b4Smrg
2397117f1b4Smrg    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
2407117f1b4Smrg    OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
2417117f1b4Smrg
2427117f1b4Smrg    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
2437117f1b4Smrg    MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
2447117f1b4Smrg
2457117f1b4Smrg    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
2467117f1b4Smrg    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
2477117f1b4Smrg
2487117f1b4Smrg    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
2497117f1b4Smrg    ADD_L( EDI, ECX ) 				/* count += dest ptr */
2507117f1b4Smrg
2517117f1b4SmrgALIGNTEXT32
2527117f1b4Smrg    MOVSS    ( M(0), XMM1 )			/* -  | -  |  -  | m0  */
2537117f1b4Smrg    MOVSS    ( M(5), XMM2 )			/* -  | -  |  -  | m5  */
2547117f1b4Smrg    UNPCKLPS ( XMM2, XMM1 )			/* -  | -  | m5  | m0  */
2557117f1b4Smrg    MOVSS    ( M(14), XMM3 )			/* m14 */
2567117f1b4Smrg    XORPS    ( XMM0, XMM0 )			/* 0 | 0 | 0 | 0 */
2577117f1b4Smrg
2587117f1b4SmrgALIGNTEXT32
2597117f1b4SmrgLLBL(K_GTP23PR_top):
2607117f1b4Smrg    MOVLPS( S(0), XMM4 )			/* oy | ox */
2617117f1b4Smrg    MULPS( XMM1, XMM4 )				/* oy*m5 | ox*m0 */
2627117f1b4Smrg    MOVLPS( XMM4, D(0) )			/* ->D(1) | ->D(0) */
2637117f1b4Smrg    MOVSS( XMM3, D(2) )				/* ->D(2) */
2647117f1b4Smrg    MOVSS( XMM0, D(3) )				/* ->D(3) */
2657117f1b4Smrg
2667117f1b4SmrgLLBL(K_GTP23PR_skip):
2677117f1b4Smrg    ADD_L( CONST(16), EDI )
2687117f1b4Smrg    ADD_L( EAX, ESI )
2697117f1b4Smrg    CMP_L( ECX, EDI )
2707117f1b4Smrg    JNE( LLBL(K_GTP23PR_top) )
2717117f1b4Smrg
2727117f1b4SmrgLLBL(K_GTP23PR_finish):
2737117f1b4Smrg    POP_L    ( EDI )
2747117f1b4Smrg    POP_L    ( ESI )
2757117f1b4Smrg    RET
2767117f1b4Smrg#undef FRAME_OFFSET
2777117f1b4Smrg
2787117f1b4Smrg
2797117f1b4Smrg
2807117f1b4SmrgALIGNTEXT4
2817117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points2_2d)
2827117f1b4SmrgHIDDEN(_mesa_sse_transform_points2_2d)
2837117f1b4SmrgGLNAME(_mesa_sse_transform_points2_2d):
2847ec681f3Smrg    _CET_ENDBR
2857117f1b4Smrg#define FRAME_OFFSET 8
2867117f1b4Smrg    PUSH_L( ESI )
2877117f1b4Smrg    PUSH_L( EDI )
2887117f1b4Smrg
2897117f1b4Smrg    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
2907117f1b4Smrg    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
2917117f1b4Smrg
2927117f1b4Smrg    MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
2937117f1b4Smrg    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
2947117f1b4Smrg
2957117f1b4Smrg    TEST_L( ECX, ECX)
2967117f1b4Smrg    JZ( LLBL(K_GTP23P2DR_finish) ) 		/* count was zero; go to finish */
2977117f1b4Smrg
2987117f1b4Smrg    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
2997117f1b4Smrg    OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
3007117f1b4Smrg
3017117f1b4Smrg    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
3027117f1b4Smrg    MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
3037117f1b4Smrg
3047117f1b4Smrg    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
3057117f1b4Smrg    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
3067117f1b4Smrg
3077117f1b4Smrg    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
3087117f1b4Smrg    ADD_L( EDI, ECX ) 				/* count += dest ptr */
3097117f1b4Smrg
3107117f1b4SmrgALIGNTEXT32
3117117f1b4Smrg    MOVLPS( M(0), XMM0 )			/* m1  | m0 */
3127117f1b4Smrg    MOVLPS( M(4), XMM1 )			/* m5  | m4 */
3137117f1b4Smrg    MOVLPS( M(12), XMM2 )			/* m13 | m12 */
3147117f1b4Smrg
3157117f1b4SmrgALIGNTEXT32
3167117f1b4SmrgLLBL(K_GTP23P2DR_top):
3177117f1b4Smrg    MOVSS( S(0), XMM3 )				/* ox */
3187117f1b4Smrg    SHUFPS( CONST(0x0), XMM3, XMM3 )		/* ox | ox */
3197117f1b4Smrg    MULPS( XMM0, XMM3 )				/* ox*m1 | ox*m0 */
3207117f1b4Smrg
3217117f1b4Smrg    MOVSS( S(1), XMM4 )				/* oy */
3227117f1b4Smrg    SHUFPS( CONST(0x0), XMM4, XMM4 )		/* oy | oy */
3237117f1b4Smrg    MULPS( XMM1, XMM4 )				/* oy*m5 | oy*m4 */
3247117f1b4Smrg
3257117f1b4Smrg    ADDPS( XMM4, XMM3 )
3267117f1b4Smrg    ADDPS( XMM2, XMM3 )
3277117f1b4Smrg    MOVLPS( XMM3, D(0) )			/* ->D(1) | ->D(0) */
3287117f1b4Smrg
3297117f1b4SmrgLLBL(K_GTP23P2DR_skip):
3307117f1b4Smrg    ADD_L    ( CONST(16), EDI )
3317117f1b4Smrg    ADD_L    ( EAX, ESI )
3327117f1b4Smrg    CMP_L    ( ECX, EDI )
3337117f1b4Smrg    JNE      ( LLBL(K_GTP23P2DR_top) )
3347117f1b4Smrg
3357117f1b4SmrgLLBL(K_GTP23P2DR_finish):
3367117f1b4Smrg    POP_L    ( EDI )
3377117f1b4Smrg    POP_L    ( ESI )
3387117f1b4Smrg    RET
3397117f1b4Smrg#undef FRAME_OFFSET
3407117f1b4Smrg
3417117f1b4Smrg
3427117f1b4Smrg
3437117f1b4SmrgALIGNTEXT4
3447117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points2_2d_no_rot)
3457117f1b4SmrgHIDDEN(_mesa_sse_transform_points2_2d_no_rot)
3467117f1b4SmrgGLNAME(_mesa_sse_transform_points2_2d_no_rot):
3477ec681f3Smrg	_CET_ENDBR
3487117f1b4Smrg#define FRAME_OFFSET 8
3497117f1b4Smrg	PUSH_L( ESI )
3507117f1b4Smrg	PUSH_L( EDI )
3517117f1b4Smrg
3527117f1b4Smrg	MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
3537117f1b4Smrg	MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
3547117f1b4Smrg
3557117f1b4Smrg	MOV_L( ARG_MATRIX, EDX ) 		/* ptr to matrix */
3567117f1b4Smrg	MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
3577117f1b4Smrg
3587117f1b4Smrg	TEST_L( ECX, ECX)
3597117f1b4Smrg	JZ( LLBL(K_GTP23P2DNRR_finish) ) 	/* count was zero; go to finish */
3607117f1b4Smrg
3617117f1b4Smrg	MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
3627117f1b4Smrg	OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
3637117f1b4Smrg
3647117f1b4Smrg	MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
3657117f1b4Smrg	MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
3667117f1b4Smrg
3677117f1b4Smrg	SHL_L( CONST(4), ECX ) 			/* count *= 16 */
3687117f1b4Smrg	MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
3697117f1b4Smrg
3707117f1b4Smrg	MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
3717117f1b4Smrg	ADD_L( EDI, ECX ) 			/* count += dest ptr */
3727117f1b4Smrg
3737117f1b4SmrgALIGNTEXT32
3747117f1b4Smrg	MOVSS    ( M(0), XMM1 )			/* m0 */
3757117f1b4Smrg	MOVSS    ( M(5), XMM2 )			/* m5 */
3767117f1b4Smrg	UNPCKLPS ( XMM2, XMM1 )			/* m5 | m0 */
3777117f1b4Smrg	MOVLPS   ( M(12), XMM2 )		/* m13 | m12 */
3787117f1b4Smrg
3797117f1b4SmrgALIGNTEXT32
3807117f1b4SmrgLLBL(K_GTP23P2DNRR_top):
3817117f1b4Smrg	MOVLPS( S(0), XMM0 )			/* oy | ox */
3827117f1b4Smrg	MULPS( XMM1, XMM0 )			/* oy*m5 | ox*m0 */
3837117f1b4Smrg	ADDPS( XMM2, XMM0 )			/* +m13 | +m12 */
3847117f1b4Smrg	MOVLPS( XMM0, D(0) )			/* ->D(1) | ->D(0) */
3857117f1b4Smrg
3867117f1b4SmrgLLBL(K_GTP23P2DNRR_skip):
3877117f1b4Smrg	ADD_L( CONST(16), EDI )
3887117f1b4Smrg	ADD_L( EAX, ESI )
3897117f1b4Smrg	CMP_L( ECX, EDI )
3907117f1b4Smrg	JNE( LLBL(K_GTP23P2DNRR_top) )
3917117f1b4Smrg
3927117f1b4SmrgLLBL(K_GTP23P2DNRR_finish):
3937117f1b4Smrg	POP_L( EDI )
3947117f1b4Smrg	POP_L( ESI )
3957117f1b4Smrg	RET
3967117f1b4Smrg#undef FRAME_OFFSET
3977117f1b4Smrg
3987117f1b4Smrg
3997117f1b4Smrg
4007117f1b4SmrgALIGNTEXT4
4017117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points2_3d)
4027117f1b4SmrgHIDDEN(_mesa_sse_transform_points2_3d)
4037117f1b4SmrgGLNAME(_mesa_sse_transform_points2_3d):
4047ec681f3Smrg	_CET_ENDBR
4057117f1b4Smrg#define FRAME_OFFSET 8
4067117f1b4Smrg	PUSH_L( ESI )
4077117f1b4Smrg	PUSH_L( EDI )
4087117f1b4Smrg
4097117f1b4Smrg	MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
4107117f1b4Smrg	MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
4117117f1b4Smrg
4127117f1b4Smrg	MOV_L( ARG_MATRIX, EDX ) 		/* ptr to matrix */
4137117f1b4Smrg	MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
4147117f1b4Smrg
4157117f1b4Smrg	TEST_L( ECX, ECX)
4167117f1b4Smrg	JZ( LLBL(K_GTP23P3DR_finish) ) 	/* count was zero; go to finish */
4177117f1b4Smrg
4187117f1b4Smrg	MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
4197117f1b4Smrg	OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
4207117f1b4Smrg
4217117f1b4Smrg	MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
4227117f1b4Smrg	MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
4237117f1b4Smrg
4247117f1b4Smrg	SHL_L( CONST(4), ECX ) 			/* count *= 16 */
4257117f1b4Smrg	MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
4267117f1b4Smrg
4277117f1b4Smrg	MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
4287117f1b4Smrg	ADD_L( EDI, ECX ) 			/* count += dest ptr */
4297117f1b4Smrg
4307117f1b4SmrgALIGNTEXT32
4317117f1b4Smrg	MOVAPS( M(0), XMM0 )			/* m2  | m1  | m0 */
4327117f1b4Smrg	MOVAPS( M(4), XMM1 )			/* m6  | m5  | m4 */
4337117f1b4Smrg	MOVAPS( M(12), XMM2 )			/* m14 | m13 | m12 */
4347117f1b4Smrg
4357117f1b4SmrgALIGNTEXT32
4367117f1b4SmrgLLBL(K_GTP23P3DR_top):
4377117f1b4Smrg	MOVSS( S(0), XMM3 )			/* ox */
4387117f1b4Smrg	SHUFPS( CONST(0x0), XMM3, XMM3 )	/* ox | ox | ox */
4397117f1b4Smrg	MULPS( XMM0, XMM3 )			/* ox*m2 | ox*m1 | ox*m0 */
4407117f1b4Smrg
4417117f1b4Smrg	MOVSS( S(1), XMM4 )			/* oy */
4427117f1b4Smrg	SHUFPS( CONST(0x0), XMM4, XMM4 )	/* oy | oy | oy */
4437117f1b4Smrg	MULPS( XMM1, XMM4 )			/* oy*m6 | oy*m5 | oy*m4 */
4447117f1b4Smrg
4457117f1b4Smrg	ADDPS( XMM4, XMM3 )
4467117f1b4Smrg	ADDPS( XMM2, XMM3 )
4477117f1b4Smrg
4487117f1b4Smrg	MOVLPS( XMM3, D(0) )			/* ->D(1) | ->D(0) */
4497117f1b4Smrg	UNPCKHPS( XMM3, XMM3 )
4507117f1b4Smrg	MOVSS( XMM3, D(2) )			/* ->D(2) */
4517117f1b4Smrg
4527117f1b4SmrgLLBL(K_GTP23P3DR_skip):
4537117f1b4Smrg	ADD_L( CONST(16), EDI )
4547117f1b4Smrg	ADD_L( EAX, ESI )
4557117f1b4Smrg	CMP_L( ECX, EDI )
4567117f1b4Smrg	JNE( LLBL(K_GTP23P3DR_top) )
4577117f1b4Smrg
4587117f1b4SmrgLLBL(K_GTP23P3DR_finish):
4597117f1b4Smrg	POP_L( EDI )
4607117f1b4Smrg	POP_L( ESI )
4617117f1b4Smrg	RET
4627117f1b4Smrg#undef FRAME_OFFSET
4637117f1b4Smrg#endif
4647117f1b4Smrg
4657117f1b4Smrg#if defined (__ELF__) && defined (__linux__)
4667117f1b4Smrg	.section .note.GNU-stack,"",%progbits
4677117f1b4Smrg#endif
468