1848b8605Smrg 2848b8605Smrg/* 3848b8605Smrg * Mesa 3-D graphics library 4848b8605Smrg * 5848b8605Smrg * Copyright (C) 1999-2001 Brian Paul All Rights Reserved. 6848b8605Smrg * 7848b8605Smrg * Permission is hereby granted, free of charge, to any person obtaining a 8848b8605Smrg * copy of this software and associated documentation files (the "Software"), 9848b8605Smrg * to deal in the Software without restriction, including without limitation 10848b8605Smrg * the rights to use, copy, modify, merge, publish, distribute, sublicense, 11848b8605Smrg * and/or sell copies of the Software, and to permit persons to whom the 12848b8605Smrg * Software is furnished to do so, subject to the following conditions: 13848b8605Smrg * 14848b8605Smrg * The above copyright notice and this permission notice shall be included 15848b8605Smrg * in all copies or substantial portions of the Software. 16848b8605Smrg * 17848b8605Smrg * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 18848b8605Smrg * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 19848b8605Smrg * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 20848b8605Smrg * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR 21848b8605Smrg * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 22848b8605Smrg * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 23848b8605Smrg * OTHER DEALINGS IN THE SOFTWARE. 24848b8605Smrg */ 25848b8605Smrg 26848b8605Smrg#ifdef USE_3DNOW_ASM 27848b8605Smrg#include "assyntax.h" 28848b8605Smrg#include "matypes.h" 29848b8605Smrg#include "xform_args.h" 30848b8605Smrg 31848b8605Smrg SEG_TEXT 32848b8605Smrg 33848b8605Smrg#define FRAME_OFFSET 4 34848b8605Smrg 35848b8605Smrg 36848b8605SmrgALIGNTEXT16 37848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points4_general ) 38848b8605SmrgHIDDEN(_mesa_3dnow_transform_points4_general) 39848b8605SmrgGLNAME( _mesa_3dnow_transform_points4_general ): 40848b8605Smrg 41848b8605Smrg PUSH_L ( ESI ) 42848b8605Smrg 43848b8605Smrg MOV_L ( ARG_DEST, ECX ) 44848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 45848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 46848b8605Smrg MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) ) 47848b8605Smrg OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) ) 48848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 49848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 50848b8605Smrg 51848b8605Smrg PUSH_L ( EDI ) 52848b8605Smrg 53848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 54848b8605Smrg MOV_L ( ESI, ECX ) 55848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 56848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 57848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 58848b8605Smrg 59848b8605Smrg TEST_L ( ESI, ESI ) 60848b8605Smrg JZ ( LLBL( G3TPGR_2 ) ) 61848b8605Smrg 62848b8605Smrg PREFETCHW ( REGIND(EDX) ) 63848b8605Smrg 64848b8605SmrgALIGNTEXT16 65848b8605SmrgLLBL( G3TPGR_1 ): 66848b8605Smrg 67848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 68848b8605Smrg 69848b8605Smrg MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ 70848b8605Smrg MOVQ ( REGOFF(8, EAX), MM4 ) /* x3 | x2 */ 71848b8605Smrg 72848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 73848b8605Smrg PREFETCH ( REGIND(EAX) ) 74848b8605Smrg 75848b8605Smrg MOVQ ( MM0, MM2 ) /* x1 | x0 */ 76848b8605Smrg MOVQ ( MM4, MM6 ) /* x3 | x2 */ 77848b8605Smrg 78848b8605Smrg PUNPCKLDQ ( MM0, MM0 ) /* x0 | x0 */ 79848b8605Smrg PUNPCKHDQ ( MM2, MM2 ) /* x1 | x1 */ 80848b8605Smrg 81848b8605Smrg MOVQ ( MM0, MM1 ) /* x0 | x0 */ 82848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 83848b8605Smrg 84848b8605Smrg PFMUL ( REGIND(ECX), MM0 ) /* x0*m1 | x0*m0 */ 85848b8605Smrg MOVQ ( MM2, MM3 ) /* x1 | x1 */ 86848b8605Smrg 87848b8605Smrg PFMUL ( REGOFF(8, ECX), MM1 ) /* x0*m3 | x0*m2 */ 88848b8605Smrg PUNPCKLDQ ( MM4, MM4 ) /* x2 | x2 */ 89848b8605Smrg 90848b8605Smrg PFMUL ( REGOFF(16, ECX), MM2 ) /* x1*m5 | x1*m4 */ 91848b8605Smrg MOVQ ( MM4, MM5 ) /* x2 | x2 */ 92848b8605Smrg 93848b8605Smrg PFMUL ( REGOFF(24, ECX), MM3 ) /* x1*m7 | x1*m6 */ 94848b8605Smrg PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */ 95848b8605Smrg 96848b8605Smrg PFMUL ( REGOFF(32, ECX), MM4 ) /* x2*m9 | x2*m8 */ 97848b8605Smrg MOVQ ( MM6, MM7 ) /* x3 | x3 */ 98848b8605Smrg 99848b8605Smrg PFMUL ( REGOFF(40, ECX), MM5 ) /* x2*m11 | x2*m10 */ 100848b8605Smrg PFADD ( MM0, MM2 ) 101848b8605Smrg 102848b8605Smrg PFMUL ( REGOFF(48, ECX), MM6 ) /* x3*m13 | x3*m12 */ 103848b8605Smrg PFADD ( MM1, MM3 ) 104848b8605Smrg 105848b8605Smrg PFMUL ( REGOFF(56, ECX), MM7 ) /* x3*m15 | x3*m14 */ 106848b8605Smrg PFADD ( MM4, MM6 ) 107848b8605Smrg 108848b8605Smrg PFADD ( MM5, MM7 ) 109848b8605Smrg PFADD ( MM2, MM6 ) 110848b8605Smrg 111848b8605Smrg PFADD ( MM3, MM7 ) 112848b8605Smrg MOVQ ( MM6, REGOFF(-16, EDX) ) 113848b8605Smrg 114848b8605Smrg MOVQ ( MM7, REGOFF(-8, EDX) ) 115848b8605Smrg 116848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 117848b8605Smrg JNZ ( LLBL( G3TPGR_1 ) ) /* cnt > 0 ? -> process next vertex */ 118848b8605Smrg 119848b8605SmrgLLBL( G3TPGR_2 ): 120848b8605Smrg 121848b8605Smrg FEMMS 122848b8605Smrg POP_L ( EDI ) 123848b8605Smrg POP_L ( ESI ) 124848b8605Smrg RET 125848b8605Smrg 126848b8605Smrg 127848b8605Smrg 128848b8605Smrg 129848b8605SmrgALIGNTEXT16 130848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points4_perspective ) 131848b8605SmrgHIDDEN(_mesa_3dnow_transform_points4_perspective) 132848b8605SmrgGLNAME( _mesa_3dnow_transform_points4_perspective ): 133848b8605Smrg 134848b8605Smrg PUSH_L ( ESI ) 135848b8605Smrg 136848b8605Smrg MOV_L ( ARG_DEST, ECX ) 137848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 138848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 139848b8605Smrg MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) ) 140848b8605Smrg OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) ) 141848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 142848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 143848b8605Smrg 144848b8605Smrg PUSH_L ( EDI ) 145848b8605Smrg 146848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 147848b8605Smrg MOV_L ( ESI, ECX ) 148848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 149848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 150848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 151848b8605Smrg 152848b8605Smrg TEST_L ( ESI, ESI ) 153848b8605Smrg JZ ( LLBL( G3TPPR_2 ) ) 154848b8605Smrg 155848b8605Smrg PREFETCH ( REGIND(EAX) ) 156848b8605Smrg PREFETCHW ( REGIND(EDX) ) 157848b8605Smrg 158848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 159848b8605Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */ 160848b8605Smrg 161848b8605Smrg MOVD ( REGOFF(40, ECX), MM1 ) /* | m22 */ 162848b8605Smrg PUNPCKLDQ ( REGOFF(56, ECX), MM1 ) /* m32 | m22 */ 163848b8605Smrg 164848b8605Smrg MOVQ ( REGOFF(32, ECX), MM2 ) /* m21 | m20 */ 165848b8605Smrg PXOR ( MM7, MM7 ) /* 0 | 0 */ 166848b8605Smrg 167848b8605SmrgALIGNTEXT16 168848b8605SmrgLLBL( G3TPPR_1 ): 169848b8605Smrg 170848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 171848b8605Smrg 172848b8605Smrg MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ 173848b8605Smrg MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ 174848b8605Smrg MOVD ( REGOFF(8, EAX), MM3 ) /* | x2 */ 175848b8605Smrg 176848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 177848b8605Smrg PREFETCH ( REGOFF(32, EAX) ) /* hopefully stride is zero */ 178848b8605Smrg 179848b8605Smrg MOVQ ( MM5, MM6 ) /* x3 | x2 */ 180848b8605Smrg PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */ 181848b8605Smrg 182848b8605Smrg PUNPCKLDQ ( MM5, MM5 ) /* x2 | x2 */ 183848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 184848b8605Smrg 185848b8605Smrg PFMUL ( MM2, MM5 ) /* x2*m21 | x2*m20 */ 186848b8605Smrg PFSUBR ( MM7, MM3 ) /* | -x2 */ 187848b8605Smrg 188848b8605Smrg PFMUL ( MM1, MM6 ) /* x3*m32 | x2*m22 */ 189848b8605Smrg PFADD ( MM4, MM5 ) /* x1*m11+x2*m21 | x0*m00+x2*m20 */ 190848b8605Smrg 191848b8605Smrg PFACC ( MM3, MM6 ) /* -x2 | x2*m22+x3*m32 */ 192848b8605Smrg MOVQ ( MM5, REGOFF(-16, EDX) ) /* write r0, r1 */ 193848b8605Smrg 194848b8605Smrg MOVQ ( MM6, REGOFF(-8, EDX) ) /* write r2, r3 */ 195848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 196848b8605Smrg 197848b8605Smrg JNZ ( LLBL( G3TPPR_1 ) ) /* cnt > 0 ? -> process next vertex */ 198848b8605Smrg 199848b8605SmrgLLBL( G3TPPR_2 ): 200848b8605Smrg 201848b8605Smrg FEMMS 202848b8605Smrg POP_L ( EDI ) 203848b8605Smrg POP_L ( ESI ) 204848b8605Smrg RET 205848b8605Smrg 206848b8605Smrg 207848b8605Smrg 208848b8605Smrg 209848b8605SmrgALIGNTEXT16 210848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points4_3d ) 211848b8605SmrgHIDDEN(_mesa_3dnow_transform_points4_3d) 212848b8605SmrgGLNAME( _mesa_3dnow_transform_points4_3d ): 213848b8605Smrg 214848b8605Smrg PUSH_L ( ESI ) 215848b8605Smrg 216848b8605Smrg MOV_L ( ARG_DEST, ECX ) 217848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 218848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 219848b8605Smrg MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) ) 220848b8605Smrg OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) ) 221848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 222848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 223848b8605Smrg 224848b8605Smrg PUSH_L ( EDI ) 225848b8605Smrg 226848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 227848b8605Smrg MOV_L ( ESI, ECX ) 228848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 229848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 230848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 231848b8605Smrg 232848b8605Smrg TEST_L ( ESI, ESI ) 233848b8605Smrg JZ ( LLBL( G3TP3R_2 ) ) 234848b8605Smrg 235848b8605Smrg MOVD ( REGOFF(8, ECX), MM6 ) /* | m2 */ 236848b8605Smrg PUNPCKLDQ ( REGOFF(24, ECX), MM6 ) /* m6 | m2 */ 237848b8605Smrg 238848b8605Smrg MOVD ( REGOFF(40, ECX), MM7 ) /* | m10 */ 239848b8605Smrg PUNPCKLDQ ( REGOFF(56, ECX), MM7 ) /* m14 | m10 */ 240848b8605Smrg 241848b8605SmrgALIGNTEXT16 242848b8605SmrgLLBL( G3TP3R_1 ): 243848b8605Smrg 244848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 245848b8605Smrg PREFETCH ( REGOFF(32, EAX) ) /* hopefully array is tightly packed */ 246848b8605Smrg 247848b8605Smrg MOVQ ( REGIND(EAX), MM2 ) /* x1 | x0 */ 248848b8605Smrg MOVQ ( REGOFF(8, EAX), MM3 ) /* x3 | x2 */ 249848b8605Smrg 250848b8605Smrg MOVQ ( MM2, MM0 ) /* x1 | x0 */ 251848b8605Smrg MOVQ ( MM3, MM4 ) /* x3 | x2 */ 252848b8605Smrg 253848b8605Smrg MOVQ ( MM0, MM1 ) /* x1 | x0 */ 254848b8605Smrg MOVQ ( MM4, MM5 ) /* x3 | x2 */ 255848b8605Smrg 256848b8605Smrg PUNPCKLDQ ( MM0, MM0 ) /* x0 | x0 */ 257848b8605Smrg PUNPCKHDQ ( MM1, MM1 ) /* x1 | x1 */ 258848b8605Smrg 259848b8605Smrg PFMUL ( REGIND(ECX), MM0 ) /* x0*m1 | x0*m0 */ 260848b8605Smrg PUNPCKLDQ ( MM3, MM3 ) /* x2 | x2 */ 261848b8605Smrg 262848b8605Smrg PFMUL ( REGOFF(16, ECX), MM1 ) /* x1*m5 | x1*m4 */ 263848b8605Smrg PUNPCKHDQ ( MM4, MM4 ) /* x3 | x3 */ 264848b8605Smrg 265848b8605Smrg PFMUL ( MM6, MM2 ) /* x1*m6 | x0*m2 */ 266848b8605Smrg PFADD ( MM0, MM1 ) /* x0*m1+x1*m5 | x0*m0+x1*m4 */ 267848b8605Smrg 268848b8605Smrg PFMUL ( REGOFF(32, ECX), MM3 ) /* x2*m9 | x2*m8 */ 269848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 270848b8605Smrg 271848b8605Smrg PFMUL ( REGOFF(48, ECX), MM4 ) /* x3*m13 | x3*m12 */ 272848b8605Smrg PFADD ( MM1, MM3 ) /* x0*m1+..+x2*m9 | x0*m0+...+x2*m8 */ 273848b8605Smrg 274848b8605Smrg PFMUL ( MM7, MM5 ) /* x3*m14 | x2*m10 */ 275848b8605Smrg PFADD ( MM3, MM4 ) /* r1 | r0 */ 276848b8605Smrg 277848b8605Smrg PFACC ( MM2, MM5 ) /* x0*m2+x1*m6 | x2*m10+x3*m14 */ 278848b8605Smrg MOVD ( REGOFF(12, EAX), MM0 ) /* | x3 */ 279848b8605Smrg 280848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 281848b8605Smrg PFACC ( MM0, MM5 ) /* r3 | r2 */ 282848b8605Smrg 283848b8605Smrg MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */ 284848b8605Smrg MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */ 285848b8605Smrg 286848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 287848b8605Smrg JNZ ( LLBL( G3TP3R_1 ) ) /* cnt > 0 ? -> process next vertex */ 288848b8605Smrg 289848b8605SmrgLLBL( G3TP3R_2 ): 290848b8605Smrg 291848b8605Smrg FEMMS 292848b8605Smrg POP_L ( EDI ) 293848b8605Smrg POP_L ( ESI ) 294848b8605Smrg RET 295848b8605Smrg 296848b8605Smrg 297848b8605Smrg 298848b8605Smrg 299848b8605SmrgALIGNTEXT16 300848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points4_3d_no_rot ) 301848b8605SmrgHIDDEN(_mesa_3dnow_transform_points4_3d_no_rot) 302848b8605SmrgGLNAME( _mesa_3dnow_transform_points4_3d_no_rot ): 303848b8605Smrg 304848b8605Smrg PUSH_L ( ESI ) 305848b8605Smrg MOV_L ( ARG_DEST, ECX ) 306848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 307848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 308848b8605Smrg MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) ) 309848b8605Smrg OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) ) 310848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 311848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 312848b8605Smrg 313848b8605Smrg PUSH_L ( EDI ) 314848b8605Smrg 315848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 316848b8605Smrg MOV_L ( ESI, ECX ) 317848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 318848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 319848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 320848b8605Smrg 321848b8605Smrg TEST_L ( ESI, ESI ) 322848b8605Smrg JZ ( LLBL( G3TP3NRR_2 ) ) 323848b8605Smrg 324848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 325848b8605Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */ 326848b8605Smrg 327848b8605Smrg MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */ 328848b8605Smrg PUNPCKLDQ ( REGOFF(56, ECX), MM2 ) /* m32 | m22 */ 329848b8605Smrg 330848b8605Smrg MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */ 331848b8605Smrg 332848b8605SmrgALIGNTEXT16 333848b8605SmrgLLBL( G3TP3NRR_1 ): 334848b8605Smrg 335848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 336848b8605Smrg 337848b8605Smrg MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ 338848b8605Smrg MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ 339848b8605Smrg MOVD ( REGOFF(12, EAX), MM7 ) /* | x3 */ 340848b8605Smrg 341848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 342848b8605Smrg PREFETCH ( REGOFF(32, EAX) ) /* hopefully stride is zero */ 343848b8605Smrg 344848b8605Smrg MOVQ ( MM5, MM6 ) /* x3 | x2 */ 345848b8605Smrg PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */ 346848b8605Smrg 347848b8605Smrg PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */ 348848b8605Smrg PFMUL ( MM2, MM5 ) /* x3*m32 | x2*m22 */ 349848b8605Smrg 350848b8605Smrg PFMUL ( MM1, MM6 ) /* x3*m31 | x3*m30 */ 351848b8605Smrg PFACC ( MM7, MM5 ) /* x3 | x2*m22+x3*m32 */ 352848b8605Smrg 353848b8605Smrg PFADD ( MM6, MM4 ) /* x1*m11+x3*m31 | x0*m00+x3*m30 */ 354848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 355848b8605Smrg 356848b8605Smrg MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */ 357848b8605Smrg MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */ 358848b8605Smrg 359848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 360848b8605Smrg JNZ ( LLBL( G3TP3NRR_1 ) ) /* cnt > 0 ? -> process next vertex */ 361848b8605Smrg 362848b8605SmrgLLBL( G3TP3NRR_2 ): 363848b8605Smrg 364848b8605Smrg FEMMS 365848b8605Smrg POP_L ( EDI ) 366848b8605Smrg POP_L ( ESI ) 367848b8605Smrg RET 368848b8605Smrg 369848b8605Smrg 370848b8605Smrg 371848b8605Smrg 372848b8605SmrgALIGNTEXT16 373848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points4_2d ) 374848b8605SmrgHIDDEN(_mesa_3dnow_transform_points4_2d) 375848b8605SmrgGLNAME( _mesa_3dnow_transform_points4_2d ): 376848b8605Smrg 377848b8605Smrg PUSH_L ( ESI ) 378848b8605Smrg 379848b8605Smrg MOV_L ( ARG_DEST, ECX ) 380848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 381848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 382848b8605Smrg MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) ) 383848b8605Smrg OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) ) 384848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 385848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 386848b8605Smrg 387848b8605Smrg PUSH_L ( EDI ) 388848b8605Smrg 389848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 390848b8605Smrg MOV_L ( ESI, ECX ) 391848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 392848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 393848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 394848b8605Smrg 395848b8605Smrg TEST_L ( ESI, ESI ) 396848b8605Smrg JZ ( LLBL( G3TP2R_2 ) ) 397848b8605Smrg 398848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 399848b8605Smrg PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */ 400848b8605Smrg 401848b8605Smrg MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */ 402848b8605Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */ 403848b8605Smrg 404848b8605Smrg MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */ 405848b8605Smrg 406848b8605SmrgALIGNTEXT16 407848b8605SmrgLLBL( G3TP2R_1 ): 408848b8605Smrg 409848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 410848b8605Smrg 411848b8605Smrg MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */ 412848b8605Smrg MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ 413848b8605Smrg 414848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 415848b8605Smrg PREFETCH ( REGIND(EAX) ) 416848b8605Smrg 417848b8605Smrg MOVQ ( MM3, MM4 ) /* x1 | x0 */ 418848b8605Smrg MOVQ ( MM5, MM6 ) /* x3 | x2 */ 419848b8605Smrg 420848b8605Smrg PFMUL ( MM1, MM4 ) /* x1*m11 | x0*m01 */ 421848b8605Smrg PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */ 422848b8605Smrg 423848b8605Smrg PFMUL ( MM0, MM3 ) /* x1*m10 | x0*m00 */ 424848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 425848b8605Smrg 426848b8605Smrg PFACC ( MM4, MM3 ) /* x0*m01+x1*m11 | x0*m00+x1*m10 */ 427848b8605Smrg PFMUL ( MM2, MM6 ) /* x3*m31 | x3*m30 */ 428848b8605Smrg 429848b8605Smrg PFADD ( MM6, MM3 ) /* r1 | r0 */ 430848b8605Smrg MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */ 431848b8605Smrg 432848b8605Smrg MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */ 433848b8605Smrg 434848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 435848b8605Smrg JNZ ( LLBL( G3TP2R_1 ) ) /* cnt > 0 ? -> process next vertex */ 436848b8605Smrg 437848b8605SmrgLLBL( G3TP2R_2 ): 438848b8605Smrg 439848b8605Smrg FEMMS 440848b8605Smrg POP_L ( EDI ) 441848b8605Smrg POP_L ( ESI ) 442848b8605Smrg RET 443848b8605Smrg 444848b8605Smrg 445848b8605Smrg 446848b8605Smrg 447848b8605SmrgALIGNTEXT16 448848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points4_2d_no_rot ) 449848b8605SmrgHIDDEN(_mesa_3dnow_transform_points4_2d_no_rot) 450848b8605SmrgGLNAME( _mesa_3dnow_transform_points4_2d_no_rot ): 451848b8605Smrg 452848b8605Smrg PUSH_L ( ESI ) 453848b8605Smrg 454848b8605Smrg MOV_L ( ARG_DEST, ECX ) 455848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 456848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 457848b8605Smrg MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) ) 458848b8605Smrg OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) ) 459848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 460848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 461848b8605Smrg 462848b8605Smrg PUSH_L ( EDI ) 463848b8605Smrg 464848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 465848b8605Smrg MOV_L ( ESI, ECX ) 466848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 467848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 468848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 469848b8605Smrg 470848b8605Smrg TEST_L ( ESI, ESI ) 471848b8605Smrg JZ ( LLBL( G3TP2NRR_3 ) ) 472848b8605Smrg 473848b8605Smrg MOVD ( REGIND(ECX), MM0 ) /* | m00 */ 474848b8605Smrg PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */ 475848b8605Smrg 476848b8605Smrg MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */ 477848b8605Smrg 478848b8605SmrgALIGNTEXT16 479848b8605SmrgLLBL( G3TP2NRR_2 ): 480848b8605Smrg 481848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 482848b8605Smrg 483848b8605Smrg MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ 484848b8605Smrg MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ 485848b8605Smrg 486848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 487848b8605Smrg PREFETCH ( REGIND(EAX) ) 488848b8605Smrg 489848b8605Smrg PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */ 490848b8605Smrg MOVQ ( MM5, MM6 ) /* x3 | x2 */ 491848b8605Smrg 492848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 493848b8605Smrg PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */ 494848b8605Smrg 495848b8605Smrg PFMUL ( MM1, MM6 ) /* x3*m31 | x3*m30 */ 496848b8605Smrg PFADD ( MM4, MM6 ) /* x1*m11+x3*m31 | x0*m00+x3*m30 */ 497848b8605Smrg 498848b8605Smrg MOVQ ( MM6, REGOFF(-16, EDX) ) /* write r0, r1 */ 499848b8605Smrg MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */ 500848b8605Smrg 501848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 502848b8605Smrg 503848b8605Smrg JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */ 504848b8605Smrg 505848b8605SmrgLLBL( G3TP2NRR_3 ): 506848b8605Smrg 507848b8605Smrg FEMMS 508848b8605Smrg POP_L ( EDI ) 509848b8605Smrg POP_L ( ESI ) 510848b8605Smrg RET 511848b8605Smrg 512848b8605Smrg 513848b8605Smrg 514848b8605Smrg 515848b8605SmrgALIGNTEXT16 516848b8605SmrgGLOBL GLNAME( _mesa_3dnow_transform_points4_identity ) 517848b8605SmrgHIDDEN(_mesa_3dnow_transform_points4_identity) 518848b8605SmrgGLNAME( _mesa_3dnow_transform_points4_identity ): 519848b8605Smrg 520848b8605Smrg PUSH_L ( ESI ) 521848b8605Smrg 522848b8605Smrg MOV_L ( ARG_DEST, ECX ) 523848b8605Smrg MOV_L ( ARG_MATRIX, ESI ) 524848b8605Smrg MOV_L ( ARG_SOURCE, EAX ) 525848b8605Smrg MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) ) 526848b8605Smrg OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) ) 527848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), EDX ) 528848b8605Smrg MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) ) 529848b8605Smrg 530848b8605Smrg PUSH_L ( EDI ) 531848b8605Smrg 532848b8605Smrg MOV_L ( REGOFF(V4F_START, ECX), EDX ) 533848b8605Smrg MOV_L ( ESI, ECX ) 534848b8605Smrg MOV_L ( REGOFF(V4F_COUNT, EAX), ESI ) 535848b8605Smrg MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI ) 536848b8605Smrg MOV_L ( REGOFF(V4F_START, EAX), EAX ) 537848b8605Smrg 538848b8605Smrg TEST_L ( ESI, ESI ) 539848b8605Smrg JZ ( LLBL( G3TPIR_2 ) ) 540848b8605Smrg 541848b8605SmrgALIGNTEXT16 542848b8605SmrgLLBL( G3TPIR_1 ): 543848b8605Smrg 544848b8605Smrg PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ 545848b8605Smrg 546848b8605Smrg MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ 547848b8605Smrg MOVQ ( REGOFF(8, EAX), MM1 ) /* x3 | x2 */ 548848b8605Smrg 549848b8605Smrg ADD_L ( EDI, EAX ) /* next vertex */ 550848b8605Smrg PREFETCH ( REGIND(EAX) ) 551848b8605Smrg 552848b8605Smrg ADD_L ( CONST(16), EDX ) /* next r */ 553848b8605Smrg MOVQ ( MM0, REGOFF(-16, EDX) ) /* r1 | r0 */ 554848b8605Smrg 555848b8605Smrg MOVQ ( MM1, REGOFF(-8, EDX) ) /* r3 | r2 */ 556848b8605Smrg 557848b8605Smrg DEC_L ( ESI ) /* decrement vertex counter */ 558848b8605Smrg JNZ ( LLBL( G3TPIR_1 ) ) /* cnt > 0 ? -> process next vertex */ 559848b8605Smrg 560848b8605SmrgLLBL( G3TPIR_2 ): 561848b8605Smrg 562848b8605Smrg FEMMS 563848b8605Smrg POP_L ( EDI ) 564848b8605Smrg POP_L ( ESI ) 565848b8605Smrg RET 566848b8605Smrg#endif 567848b8605Smrg 568848b8605Smrg#if defined (__ELF__) && defined (__linux__) 569848b8605Smrg .section .note.GNU-stack,"",%progbits 570848b8605Smrg#endif 571