sse_xform2.S revision fcdc6a7d2488defd66bc7e8398c6d8c9a6190a1a
1fcdc6a7d2488defd66bc7e8398c6d8c9a6190a1aKristian Høgsberg/* $Id: sse_xform2.S,v 1.4 2006/04/17 18:58:24 krh 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_points2_general) 47932dee87e3002be87dc3bcc49efd8ac9ac3e1fa4Brian PaulHIDDEN (_mesa_sse_transform_points2_general) 481b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME( _mesa_sse_transform_points2_general ): 491b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 501b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8 511b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L ( ESI ) 521b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L ( EDI ) 531b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 541b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 551b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 561b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 571b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 581b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 591b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 601b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes TEST_L( ECX, ECX ) 611b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JZ( LLBL(K_GTP2GR_finish) ) /* count was zero; go to finish */ 621b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 631b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 641b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 651b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 661b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 671b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 681b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 691b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes SHL_L( CONST(4), ECX ) /* count *= 16 */ 701b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 711b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 721b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 731b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L( EDI, ECX ) /* count += dest ptr */ 741b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 751b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32 761b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */ 771b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVAPS( M(4), XMM1 ) /* m7 | m6 | m5 | m4 */ 781b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVAPS( M(12), XMM2 ) /* m15 | m14 | m13 | m12 */ 791b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 801b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32 811b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP2GR_top): 821b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS( S(0), XMM3 ) /* ox */ 831b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox | ox */ 841b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MULPS( XMM0, XMM3 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */ 851b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS( S(1), XMM4 ) /* oy */ 861b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy | oy */ 871b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MULPS( XMM1, XMM4 ) /* oy*m7 | oy*m6 | oy*m5 | oy*m4 */ 881b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 891b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADDPS( XMM4, XMM3 ) 901b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADDPS( XMM2, XMM3 ) 911b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVAPS( XMM3, D(0) ) 921b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 931b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP2GR_skip): 941b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L ( CONST(16), EDI ) 951b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L ( EAX, ESI ) 961b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes CMP_L ( ECX, EDI ) 971b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JNE ( LLBL(K_GTP2GR_top) ) 981b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 991b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP2GR_finish): 1001b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L ( EDI ) 1011b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L ( ESI ) 1021b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes RET 1031b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET 1041b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1051b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1061b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT4 1071b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLOBL GLNAME(_mesa_sse_transform_points2_identity) 108932dee87e3002be87dc3bcc49efd8ac9ac3e1fa4Brian PaulHIDDEN(_mesa_sse_transform_points2_identity) 1091b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME( _mesa_sse_transform_points2_identity ): 1101b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1111b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8 1121b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L ( ESI ) 1131b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L ( EDI ) 1141b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1151b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 1161b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 1171b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1181b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 1191b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1201b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes TEST_L( ECX, ECX) 1211b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JZ( LLBL(K_GTP2IR_finish) ) /* count was zero; go to finish */ 1221b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1231b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 1241b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 1251b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1261b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 1271b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 1281b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1291b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes SHL_L( CONST(4), ECX ) /* count *= 16 */ 1301b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 1311b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1321b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 1331b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L( EDI, ECX ) /* count += dest ptr */ 1341b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1351b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes CMP_L( ESI, EDI ) 1361b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JE( LLBL(K_GTP2IR_finish) ) 1371b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1381b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1391b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32 1401b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP2IR_top): 1411b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L ( S(0), EDX ) 1421b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L ( EDX, D(0) ) 1431b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L ( S(1), EDX ) 1441b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L ( EDX, D(1) ) 1451b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1461b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP2IR_skip): 1471b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L ( CONST(16), EDI ) 1481b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L ( EAX, ESI ) 1491b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes CMP_L ( ECX, EDI ) 1501b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JNE ( LLBL(K_GTP2IR_top) ) 1511b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1521b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP2IR_finish): 1531b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L ( EDI ) 1541b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L ( ESI ) 1551b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes RET 1561b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET 1571b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1581b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1591b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT4 1601b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLOBL GLNAME(_mesa_sse_transform_points2_3d_no_rot) 161932dee87e3002be87dc3bcc49efd8ac9ac3e1fa4Brian PaulHIDDEN(_mesa_sse_transform_points2_3d_no_rot) 1621b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME(_mesa_sse_transform_points2_3d_no_rot): 1631b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1641b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8 1651b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L( ESI ) 1661b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L( EDI ) 1671b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1681b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 1691b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 1701b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1711b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 1721b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 1731b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1741b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes TEST_L( ECX, ECX) 1751b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JZ( LLBL(K_GTP23DNRR_finish) ) /* count was zero; go to finish */ 1761b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1771b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 1781b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 1791b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1801b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 1811b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 1821b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1831b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes SHL_L( CONST(4), ECX ) /* count *= 16 */ 1841b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 1851b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1861b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 1871b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L( EDI, ECX ) /* count += dest ptr */ 1881b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1891b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1901b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32 1911b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 1921b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 1931b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 1941b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */ 1951b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS ( M(14), XMM3 ) /* - | - | - | m14 */ 1961b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 1971b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32 1981b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23DNRR_top): 1991b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVLPS ( S(0), XMM0 ) /* - | - | oy | ox */ 2001b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MULPS ( XMM1, XMM0 ) /* - | - | oy*m5 | ox*m0 */ 2011b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */ 2021b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */ 2031b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2041b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS ( XMM3, D(2) ) /* -> D(2) */ 2051b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2061b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23DNRR_skip): 2071b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L ( CONST(16), EDI ) 2081b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L ( EAX, ESI ) 2091b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes CMP_L ( ECX, EDI ) 2101b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JNE ( LLBL(K_GTP23DNRR_top) ) 2111b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2121b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23DNRR_finish): 2131b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L ( EDI ) 2141b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L ( ESI ) 2151b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes RET 2161b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET 2171b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2181b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2191b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT4 2201b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLOBL GLNAME(_mesa_sse_transform_points2_perspective) 221932dee87e3002be87dc3bcc49efd8ac9ac3e1fa4Brian PaulHIDDEN(_mesa_sse_transform_points2_perspective) 2221b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME(_mesa_sse_transform_points2_perspective): 2231b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2241b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8 2251b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L ( ESI ) 2261b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L ( EDI ) 2271b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2281b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 2291b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 2301b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2311b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 2321b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 2331b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2341b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes TEST_L( ECX, ECX) 2351b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JZ( LLBL(K_GTP23PR_finish) ) /* count was zero; go to finish */ 2361b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2371b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 2381b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 2391b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2401b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 2411b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 2421b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2431b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes SHL_L( CONST(4), ECX ) /* count *= 16 */ 2441b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 2451b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2461b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 2471b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L( EDI, ECX ) /* count += dest ptr */ 2481b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2491b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32 2501b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 2511b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 2521b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 2531b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS ( M(14), XMM3 ) /* m14 */ 2541b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes XORPS ( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */ 2551b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2561b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32 2571b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23PR_top): 2581b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVLPS( S(0), XMM4 ) /* oy | ox */ 2591b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MULPS( XMM1, XMM4 ) /* oy*m5 | ox*m0 */ 2601b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVLPS( XMM4, D(0) ) /* ->D(1) | ->D(0) */ 2611b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS( XMM3, D(2) ) /* ->D(2) */ 2621b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS( XMM0, D(3) ) /* ->D(3) */ 2631b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2641b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23PR_skip): 2651b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L( CONST(16), EDI ) 2661b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L( EAX, ESI ) 2671b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes CMP_L( ECX, EDI ) 2681b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JNE( LLBL(K_GTP23PR_top) ) 2691b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2701b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23PR_finish): 2711b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L ( EDI ) 2721b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L ( ESI ) 2731b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes RET 2741b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET 2751b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2761b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2771b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2781b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT4 2791b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLOBL GLNAME(_mesa_sse_transform_points2_2d) 280932dee87e3002be87dc3bcc49efd8ac9ac3e1fa4Brian PaulHIDDEN(_mesa_sse_transform_points2_2d) 2811b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME(_mesa_sse_transform_points2_2d): 2821b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2831b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8 2841b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L( ESI ) 2851b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L( EDI ) 2861b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2871b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 2881b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 2891b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2901b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 2911b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 2921b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2931b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes TEST_L( ECX, ECX) 2941b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JZ( LLBL(K_GTP23P2DR_finish) ) /* count was zero; go to finish */ 2951b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2961b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 2971b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 2981b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 2991b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 3001b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 3011b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3021b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes SHL_L( CONST(4), ECX ) /* count *= 16 */ 3031b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 3041b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3051b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 3061b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L( EDI, ECX ) /* count += dest ptr */ 3071b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3081b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32 3091b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVLPS( M(0), XMM0 ) /* m1 | m0 */ 3101b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVLPS( M(4), XMM1 ) /* m5 | m4 */ 3111b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVLPS( M(12), XMM2 ) /* m13 | m12 */ 3121b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3131b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32 3141b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23P2DR_top): 3151b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS( S(0), XMM3 ) /* ox */ 3161b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox */ 3171b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MULPS( XMM0, XMM3 ) /* ox*m1 | ox*m0 */ 3181b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3191b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS( S(1), XMM4 ) /* oy */ 3201b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy */ 3211b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MULPS( XMM1, XMM4 ) /* oy*m5 | oy*m4 */ 3221b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3231b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADDPS( XMM4, XMM3 ) 3241b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADDPS( XMM2, XMM3 ) 3251b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */ 3261b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3271b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23P2DR_skip): 3281b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L ( CONST(16), EDI ) 3291b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L ( EAX, ESI ) 3301b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes CMP_L ( ECX, EDI ) 3311b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JNE ( LLBL(K_GTP23P2DR_top) ) 3321b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3331b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23P2DR_finish): 3341b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L ( EDI ) 3351b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L ( ESI ) 3361b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes RET 3371b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET 3381b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3391b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3401b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3411b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT4 3421b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLOBL GLNAME(_mesa_sse_transform_points2_2d_no_rot) 343932dee87e3002be87dc3bcc49efd8ac9ac3e1fa4Brian PaulHIDDEN(_mesa_sse_transform_points2_2d_no_rot) 3441b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME(_mesa_sse_transform_points2_2d_no_rot): 3451b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3461b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8 3471b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L( ESI ) 3481b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L( EDI ) 3491b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3501b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 3511b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 3521b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3531b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 3541b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 3551b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3561b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes TEST_L( ECX, ECX) 3571b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JZ( LLBL(K_GTP23P2DNRR_finish) ) /* count was zero; go to finish */ 3581b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3591b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 3601b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 3611b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3621b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 3631b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 3641b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3651b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes SHL_L( CONST(4), ECX ) /* count *= 16 */ 3661b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 3671b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3681b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 3691b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L( EDI, ECX ) /* count += dest ptr */ 3701b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3711b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32 3721b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS ( M(0), XMM1 ) /* m0 */ 3731b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS ( M(5), XMM2 ) /* m5 */ 3741b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */ 3751b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVLPS ( M(12), XMM2 ) /* m13 | m12 */ 3761b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3771b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32 3781b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23P2DNRR_top): 3791b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVLPS( S(0), XMM0 ) /* oy | ox */ 3801b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */ 3811b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */ 3821b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */ 3831b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3841b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23P2DNRR_skip): 3851b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L( CONST(16), EDI ) 3861b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L( EAX, ESI ) 3871b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes CMP_L( ECX, EDI ) 3881b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JNE( LLBL(K_GTP23P2DNRR_top) ) 3891b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3901b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23P2DNRR_finish): 3911b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L( EDI ) 3921b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L( ESI ) 3931b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes RET 3941b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET 3951b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3961b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3971b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 3981b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT4 3991b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLOBL GLNAME(_mesa_sse_transform_points2_3d) 400932dee87e3002be87dc3bcc49efd8ac9ac3e1fa4Brian PaulHIDDEN(_mesa_sse_transform_points2_3d) 4011b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesGLNAME(_mesa_sse_transform_points2_3d): 4021b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4031b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#define FRAME_OFFSET 8 4041b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L( ESI ) 4051b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes PUSH_L( EDI ) 4061b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4071b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 4081b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 4091b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4101b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 4111b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 4121b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4131b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes TEST_L( ECX, ECX) 4141b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JZ( LLBL(K_GTP23P3DR_finish) ) /* count was zero; go to finish */ 4151b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4161b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 4171b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 4181b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4191b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 4201b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 4211b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4221b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes SHL_L( CONST(4), ECX ) /* count *= 16 */ 4231b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 4241b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4251b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 4261b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L( EDI, ECX ) /* count += dest ptr */ 4271b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4281b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32 4291b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */ 4301b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */ 4311b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVAPS( M(12), XMM2 ) /* m14 | m13 | m12 */ 4321b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4331b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesALIGNTEXT32 4341b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23P3DR_top): 4351b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS( S(0), XMM3 ) /* ox */ 4361b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox */ 4371b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MULPS( XMM0, XMM3 ) /* ox*m2 | ox*m1 | ox*m0 */ 4381b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4391b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS( S(1), XMM4 ) /* oy */ 4401b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy */ 4411b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MULPS( XMM1, XMM4 ) /* oy*m6 | oy*m5 | oy*m4 */ 4421b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4431b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADDPS( XMM4, XMM3 ) 4441b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADDPS( XMM2, XMM3 ) 4451b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4461b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */ 4471b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes UNPCKHPS( XMM3, XMM3 ) 4481b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes MOVSS( XMM3, D(2) ) /* ->D(2) */ 4491b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4501b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23P3DR_skip): 4511b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L( CONST(16), EDI ) 4521b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes ADD_L( EAX, ESI ) 4531b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes CMP_L( ECX, EDI ) 4541b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes JNE( LLBL(K_GTP23P3DR_top) ) 4551b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes 4561b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth HughesLLBL(K_GTP23P3DR_finish): 4571b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L( EDI ) 4581b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes POP_L( ESI ) 4591b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes RET 4601b2fef5c28a40cd001598071e25b876ad4fccdd1Gareth Hughes#undef FRAME_OFFSET 461462183fe4cb6df6d90632d9e2cee881c8d26b1cbAlan Hourihane#endif 462fcdc6a7d2488defd66bc7e8398c6d8c9a6190a1aKristian Høgsberg 463fcdc6a7d2488defd66bc7e8398c6d8c9a6190a1aKristian Høgsberg#if defined (__ELF__) && defined (__linux__) 464fcdc6a7d2488defd66bc7e8398c6d8c9a6190a1aKristian Høgsberg .section .note.GNU-stack,"",%progbits 465fcdc6a7d2488defd66bc7e8398c6d8c9a6190a1aKristian Høgsberg#endif 466