sse_xform3.S revision 462183fe4cb6df6d90632d9e2cee881c8d26b1cb
1462183fe4cb6df6d90632d9e2cee881c8d26b1cbAlan Hourihane/* $Id: sse_xform3.S,v 1.2 2004/04/26 10:10:25 alanh Exp $ */
21b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
31b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes/*
41b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * Mesa 3-D graphics library
51b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * Version:  3.5
61b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes *
71b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * Copyright (C) 1999-2001  Brian Paul   All Rights Reserved.
81b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes *
91b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * Permission is hereby granted, free of charge, to any person obtaining a
101b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * copy of this software and associated documentation files (the "Software"),
111b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * to deal in the Software without restriction, including without limitation
121b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * the rights to use, copy, modify, merge, publish, distribute, sublicense,
131b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * and/or sell copies of the Software, and to permit persons to whom the
141b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * Software is furnished to do so, subject to the following conditions:
151b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes *
161b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * The above copyright notice and this permission notice shall be included
171b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * in all copies or substantial portions of the Software.
181b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes *
191b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
201b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
211b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
221b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
231b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
241b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
251b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes */
261b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
271b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes/** TODO:
281b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes  * - insert PREFETCH instructions to avoid cache-misses !
291b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes  * - some more optimizations are possible...
301b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes  * - for 40-50% more performance in the SSE-functions, the
311b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes  *   data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned !
321b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes  */
331b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
34462183fe4cb6df6d90632d9e2cee881c8d26b1cbAlan Hourihane#ifdef USE_SSE_ASM
351b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#include "matypes.h"
361b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#include "xform_args.h"
371b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
381b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes   SEG_TEXT
391b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
401b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define S(i) 	REGOFF(i * 4, ESI)
411b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define D(i) 	REGOFF(i * 4, EDI)
421b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define M(i) 	REGOFF(i * 4, EDX)
431b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
441b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
451b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT4
461b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLOBL GLNAME(_mesa_sse_transform_points3_general)
471b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME( _mesa_sse_transform_points3_general ):
481b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
491b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8
501b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    PUSH_L    ( ESI )
511b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    PUSH_L    ( EDI )
521b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
531b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
541b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
551b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
561b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
571b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
581b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
591b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    CMP_L     ( CONST(0), ECX )			/* count == 0 ? */
601b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    JE        ( LLBL(K_GTPGR_finish) )		/* yes -> nothing to do. */
611b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
621b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
631b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
641b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
651b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
661b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
671b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
681b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
691b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
701b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
711b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
721b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L( EDI, ECX ) 				/* count += dest ptr */
731b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
741b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
751b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32
761b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVAPS    ( REGOFF(0, EDX), XMM0 )	/* m0  | m1  | m2  | m3 */
771b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVAPS    ( REGOFF(16, EDX), XMM1 )	/* m4  | m5  | m6  | m7 */
781b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVAPS    ( REGOFF(32, EDX), XMM2 )	/* m8  | m9  | m10 | m11 */
791b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVAPS    ( REGOFF(48, EDX), XMM3 )	/* m12 | m13 | m14 | m15 */
801b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
811b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
821b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32
831b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTPGR_top):
841b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS     ( REGOFF(0, ESI), XMM4 )		/*    |    |    | ox */
851b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    SHUFPS    ( CONST(0x0), XMM4, XMM4 )	/* ox | ox | ox | ox */
861b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS     ( REGOFF(4, ESI), XMM5 )		/*    |    |    | oy */
871b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    SHUFPS    ( CONST(0x0), XMM5, XMM5 )	/* oy | oy | oy | oy */
881b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS     ( REGOFF(8, ESI), XMM6 )		/*    |    |    | oz */
891b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    SHUFPS    ( CONST(0x0), XMM6, XMM6 )	/* oz | oz | oz | oz */
901b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
911b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MULPS     ( XMM0, XMM4 )		/* m3*ox  | m2*ox  | m1*ox | m0*ox */
921b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MULPS     ( XMM1, XMM5 )		/* m7*oy  | m6*oy  | m5*oy | m4*oy */
931b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MULPS     ( XMM2, XMM6 )		/* m11*oz | m10*oz | m9*oz | m8*oz */
941b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
951b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADDPS     ( XMM5, XMM4 )
961b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADDPS     ( XMM6, XMM4 )
971b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADDPS     ( XMM3, XMM4 )
981b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
991b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVAPS    ( XMM4, REGOFF(0, EDI) )
1001b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1011b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTPGR_skip):
1021b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L     ( CONST(16), EDI )
1031b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L     ( EAX, ESI )
1041b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    CMP_L     ( ECX, EDI )
1051b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    JNE       ( LLBL(K_GTPGR_top) )
1061b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1071b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTPGR_finish):
1081b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    POP_L     ( EDI )
1091b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    POP_L     ( ESI )
1101b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    RET
1111b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET
1121b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1131b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1141b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT4
1151b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLOBL GLNAME(_mesa_sse_transform_points3_identity)
1161b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME( _mesa_sse_transform_points3_identity ):
1171b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1181b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8
1191b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    PUSH_L    ( ESI )
1201b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    PUSH_L    ( EDI )
1211b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1221b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
1231b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
1241b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1251b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
1261b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1271b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    TEST_L( ECX, ECX)
1281b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    JZ( LLBL(K_GTPIR_finish) ) 			/* count was zero; go to finish */
1291b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1301b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
1311b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
1321b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1331b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
1341b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
1351b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1361b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
1371b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
1381b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1391b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
1401b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L( EDI, ECX ) 				/* count += dest ptr */
1411b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1421b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    CMP_L( ESI, EDI )
1431b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    JE( LLBL(K_GTPIR_finish) )
1441b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1451b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1461b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32
1471b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTPIR_top):
1481b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVLPS    ( S(0), XMM0 )
1491b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVLPS    ( XMM0, D(0) )
1501b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS     ( S(2), XMM0 )
1511b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS     ( XMM0, D(2) )
1521b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1531b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTPIR_skip):
1541b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L     ( CONST(16), EDI )
1551b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L     ( EAX, ESI )
1561b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    CMP_L     ( ECX, EDI )
1571b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    JNE       ( LLBL(K_GTPIR_top) )
1581b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1591b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTPIR_finish):
1601b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    POP_L     ( EDI )
1611b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    POP_L     ( ESI )
1621b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    RET
1631b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET
1641b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1651b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1661b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1671b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1681b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT4
1691b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLOBL GLNAME(_mesa_sse_transform_points3_3d_no_rot)
1701b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME(_mesa_sse_transform_points3_3d_no_rot):
1711b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1721b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8
1731b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    PUSH_L( ESI )
1741b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    PUSH_L( EDI )
1751b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1761b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
1771b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
1781b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1791b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1801b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
1811b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
1821b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1831b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    TEST_L( ECX, ECX)
1841b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    JZ( LLBL(K_GTP3DNRR_finish) ) 		/* count was zero; go to finish */
1851b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1861b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
1871b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
1881b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1891b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
1901b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
1911b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1921b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
1931b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
1941b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1951b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
1961b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L( EDI, ECX ) 				/* count += dest ptr */
1971b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1981b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
1991b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32
2001b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( M(0), XMM1 )			/* - | - |  -  | m0  */
2011b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( M(5), XMM2 )			/* - | - |  -  | m5  */
2021b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    UNPCKLPS ( XMM2, XMM1 )			/* - | - | m5  | m0  */
2031b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVLPS   ( M(12), XMM2 )			/* - | - | m13 | m12 */
2041b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( M(10), XMM3 )			/* - | - |  -  | m10 */
2051b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( M(14), XMM4 )			/* - | - |  -  | m14 */
2061b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2071b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32
2081b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3DNRR_top):
2091b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2101b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVLPS   ( S(0), XMM0 )			/* - | - |  s1   | s0 */
2111b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MULPS    ( XMM1, XMM0 )			/* - | - | s1*m5 | s0*m0 */
2121b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADDPS    ( XMM2, XMM0 )			/* - | - | +m13  | +m12 */
2131b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVLPS   ( XMM0, D(0) )			/* -> D(1) | -> D(0) */
2141b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2151b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( S(2), XMM0 )			/* sz */
2161b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MULSS    ( XMM3, XMM0 )			/* sz*m10 */
2171b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADDSS    ( XMM4, XMM0 )			/* +m14 */
2181b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( XMM0, D(2) )			/* -> D(2) */
2191b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2201b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3DNRR_skip):
2211b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L    ( CONST(16), EDI )
2221b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L    ( EAX, ESI )
2231b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    CMP_L    ( ECX, EDI )
2241b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    JNE      ( LLBL(K_GTP3DNRR_top) )
2251b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2261b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3DNRR_finish):
2271b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    POP_L    ( EDI )
2281b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    POP_L    ( ESI )
2291b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    RET
2301b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET
2311b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2321b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2331b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2341b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT4
2351b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLOBL GLNAME(_mesa_sse_transform_points3_perspective)
2361b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME(_mesa_sse_transform_points3_perspective):
2371b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2381b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8
2391b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    PUSH_L   ( ESI )
2401b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    PUSH_L   ( EDI )
2411b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2421b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
2431b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
2441b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2451b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
2461b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
2471b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2481b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    TEST_L( ECX, ECX)
2491b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    JZ( LLBL(K_GTP3PR_finish) )			/* count was zero; go to finish */
2501b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2511b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
2521b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
2531b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2541b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
2551b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
2561b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2571b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
2581b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
2591b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2601b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
2611b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L( EDI, ECX ) 				/* count += dest ptr */
2621b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2631b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32
2641b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( M(0), XMM1 )			/* -  | -  |  -  | m0  */
2651b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( M(5), XMM2 )			/* -  | -  |  -  | m5  */
2661b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    UNPCKLPS ( XMM2, XMM1 )			/* -  | -  | m5  | m0  */
2671b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVLPS   ( M(8), XMM2 )			/* -  | -  | m9  | m8  */
2681b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( M(10), XMM3 )			/* m10 */
2691b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( M(14), XMM4 )			/* m14 */
2701b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    XORPS    ( XMM6, XMM6 )			/* 0 */
2711b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2721b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32
2731b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3PR_top):
2741b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVLPS   ( S(0), XMM0 )			/* oy | ox */
2751b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MULPS    ( XMM1, XMM0 )			/* oy*m5 | ox*m0 */
2761b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( S(2), XMM5 )			/* oz */
2771b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    SHUFPS   ( CONST(0x0), XMM5, XMM5 )		/* oz | oz */
2781b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MULPS    ( XMM2, XMM5 )			/* oz*m9 | oz*m8 */
2791b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADDPS    ( XMM5, XMM0 )			/* +oy*m5 | +ox*m0 */
2801b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVLPS   ( XMM0, D(0) )			/* ->D(1) | ->D(0) */
2811b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2821b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( S(2), XMM0 )			/* oz */
2831b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MULSS    ( XMM3, XMM0 )			/* oz*m10 */
2841b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADDSS    ( XMM4, XMM0 )			/* +m14 */
2851b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( XMM0, D(2) )			/* ->D(2) */
2861b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2871b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( S(2), XMM0 )			/* oz */
2881b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( XMM6, XMM5 )			/* 0 */
2891b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    SUBPS    ( XMM0, XMM5 )			/* -oz */
2901b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( XMM5, D(3) )			/* ->D(3) */
2911b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2921b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3PR_skip):
2931b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L( CONST(16), EDI )
2941b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L( EAX, ESI )
2951b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    CMP_L( ECX, EDI )
2961b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    JNE( LLBL(K_GTP3PR_top) )
2971b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
2981b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3PR_finish):
2991b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    POP_L    ( EDI )
3001b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    POP_L    ( ESI )
3011b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    RET
3021b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET
3031b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3041b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3051b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3061b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT4
3071b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLOBL GLNAME(_mesa_sse_transform_points3_2d)
3081b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME(_mesa_sse_transform_points3_2d):
3091b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3101b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8
3111b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    PUSH_L( ESI )
3121b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    PUSH_L( EDI )
3131b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3141b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
3151b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
3161b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3171b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
3181b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
3191b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3201b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    TEST_L( ECX, ECX)
3211b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    JZ( LLBL(K_GTP3P2DR_finish) ) 		/* count was zero; go to finish */
3221b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3231b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
3241b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
3251b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3261b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
3271b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
3281b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3291b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
3301b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
3311b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3321b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
3331b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L( EDI, ECX ) 				/* count += dest ptr */
3341b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3351b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32
3361b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVLPS( M(0), XMM0 )			/* m1  | m0 */
3371b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVLPS( M(4), XMM1 )			/* m5  | m4 */
3381b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVLPS( M(12), XMM2 )			/* m13 | m12 */
3391b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3401b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32
3411b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3P2DR_top):
3421b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( S(0), XMM3 )			/* ox */
3431b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    SHUFPS   ( CONST(0x0), XMM3, XMM3 )		/* ox | ox */
3441b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MULPS    ( XMM0, XMM3 )			/* ox*m1 | ox*m0 */
3451b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( S(1), XMM4 )			/* oy */
3461b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    SHUFPS   ( CONST(0x0), XMM4, XMM4 )		/* oy | oy */
3471b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MULPS    ( XMM1, XMM4 )			/* oy*m5 | oy*m4 */
3481b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3491b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADDPS    ( XMM4, XMM3 )
3501b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADDPS    ( XMM2, XMM3 )
3511b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVLPS   ( XMM3, D(0) )
3521b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3531b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( S(2), XMM3 )
3541b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    MOVSS    ( XMM3, D(2) )
3551b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3561b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3P2DR_skip):
3571b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L    ( CONST(16), EDI )
3581b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    ADD_L    ( EAX, ESI )
3591b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    CMP_L    ( ECX, EDI )
3601b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    JNE      ( LLBL(K_GTP3P2DR_top) )
3611b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3621b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3P2DR_finish):
3631b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    POP_L    ( EDI )
3641b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    POP_L    ( ESI )
3651b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes    RET
3661b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET
3671b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3681b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3691b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3701b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT4
3711b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLOBL GLNAME(_mesa_sse_transform_points3_2d_no_rot)
3721b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME(_mesa_sse_transform_points3_2d_no_rot):
3731b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3741b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8
3751b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	PUSH_L( ESI )
3761b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	PUSH_L( EDI )
3771b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3781b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
3791b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
3801b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3811b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( ARG_MATRIX, EDX ) 		/* ptr to matrix */
3821b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
3831b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3841b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	TEST_L( ECX, ECX)
3851b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	JZ( LLBL(K_GTP3P2DNRR_finish) ) 	/* count was zero; go to finish */
3861b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3871b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
3881b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
3891b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3901b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
3911b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
3921b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3931b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	SHL_L( CONST(4), ECX ) 			/* count *= 16 */
3941b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
3951b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3961b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
3971b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	ADD_L( EDI, ECX ) 			/* count += dest ptr */
3981b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
3991b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32
4001b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVSS    ( M(0), XMM1 )			/* m0 */
4011b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVSS    ( M(5), XMM2 )			/* m5 */
4021b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	UNPCKLPS ( XMM2, XMM1 )			/* m5 | m0 */
4031b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVLPS   ( M(12), XMM2 )		/* m13 | m12 */
4041b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4051b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32
4061b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3P2DNRR_top):
4071b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVLPS( S(0), XMM0 )			/* oy | ox */
4081b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MULPS( XMM1, XMM0 )			/* oy*m5 | ox*m0 */
4091b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	ADDPS( XMM2, XMM0 )			/* +m13 | +m12 */
4101b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVLPS( XMM0, D(0) )			/* ->D(1) | ->D(0) */
4111b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4121b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVSS( S(2), XMM0 )
4131b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVSS( XMM0, D(2) )
4141b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4151b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3P2DNRR_skip):
4161b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	ADD_L( CONST(16), EDI )
4171b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	ADD_L( EAX, ESI )
4181b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	CMP_L( ECX, EDI )
4191b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	JNE( LLBL(K_GTP3P2DNRR_top) )
4201b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4211b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3P2DNRR_finish):
4221b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	POP_L( EDI )
4231b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	POP_L( ESI )
4241b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	RET
4251b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET
4261b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4271b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4281b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4291b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4301b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT4
4311b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLOBL GLNAME(_mesa_sse_transform_points3_3d)
4321b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME(_mesa_sse_transform_points3_3d):
4331b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4341b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8
4351b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	PUSH_L( ESI )
4361b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	PUSH_L( EDI )
4371b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4381b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
4391b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
4401b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4411b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4421b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( ARG_MATRIX, EDX ) 		/* ptr to matrix */
4431b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
4441b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4451b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	TEST_L( ECX, ECX)
4461b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	JZ( LLBL(K_GTP3P3DR_finish) ) 	/* count was zero; go to finish */
4471b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4481b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
4491b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
4501b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4511b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
4521b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
4531b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4541b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	SHL_L( CONST(4), ECX ) 			/* count *= 16 */
4551b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
4561b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4571b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
4581b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	ADD_L( EDI, ECX ) 			/* count += dest ptr */
4591b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4601b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4611b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32
4621b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVAPS( M(0), XMM0 )			/* m2  | m1  | m0 */
4631b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVAPS( M(4), XMM1 )			/* m6  | m5  | m4 */
4641b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVAPS( M(8), XMM2 )			/* m10 | m9  | m8 */
4651b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVAPS( M(12), XMM3 )			/* m14 | m13 | m12 */
4661b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4671b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32
4681b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3P3DR_top):
4691b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVSS( S(0), XMM4 )
4701b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	SHUFPS( CONST(0x0), XMM4, XMM4 )	/* ox | ox | ox */
4711b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MULPS( XMM0, XMM4 )			/* ox*m2 | ox*m1 | ox*m0 */
4721b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4731b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVSS( S(1), XMM5 )
4741b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	SHUFPS( CONST(0x0), XMM5, XMM5 )	/* oy | oy | oy */
4751b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MULPS( XMM1, XMM5 )			/* oy*m6 | oy*m5 | oy*m4 */
4761b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4771b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVSS( S(2), XMM6 )
4781b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	SHUFPS( CONST(0x0), XMM6, XMM6 )	/* oz | oz | oz */
4791b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MULPS( XMM2, XMM6 )			/* oz*m10 | oz*m9 | oz*m8 */
4801b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4811b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	ADDPS( XMM5, XMM4 )			/* + | + | + */
4821b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	ADDPS( XMM6, XMM4 )			/* + | + | + */
4831b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	ADDPS( XMM3, XMM4 )			/* + | + | + */
4841b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4851b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVLPS( XMM4, D(0) )			/* => D(1) | => D(0) */
4861b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	UNPCKHPS( XMM4, XMM4 )
4871b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	MOVSS( XMM4, D(2) )
4881b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4891b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3P3DR_skip):
4901b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	ADD_L( CONST(16), EDI )
4911b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	ADD_L( EAX, ESI )
4921b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	CMP_L( ECX, EDI )
4931b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	JNE( LLBL(K_GTP3P3DR_top) )
4941b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes
4951b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP3P3DR_finish):
4961b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	POP_L( EDI )
4971b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	POP_L( ESI )
4981b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes	RET
4991b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET
500462183fe4cb6df6d90632d9e2cee881c8d26b1cbAlan Hourihane#endif
501