Home | History | Annotate | Download | only in x86
      1 
      2 /*
      3  * Mesa 3-D graphics library
      4  *
      5  * Copyright (C) 1999-2001  Brian Paul   All Rights Reserved.
      6  *
      7  * Permission is hereby granted, free of charge, to any person obtaining a
      8  * copy of this software and associated documentation files (the "Software"),
      9  * to deal in the Software without restriction, including without limitation
     10  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
     11  * and/or sell copies of the Software, and to permit persons to whom the
     12  * Software is furnished to do so, subject to the following conditions:
     13  *
     14  * The above copyright notice and this permission notice shall be included
     15  * in all copies or substantial portions of the Software.
     16  *
     17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
     18  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
     19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
     20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
     21  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
     22  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
     23  * OTHER DEALINGS IN THE SOFTWARE.
     24  */
     25 
     26 /** TODO:
     27   * - insert PREFETCH instructions to avoid cache-misses !
     28   * - some more optimizations are possible...
     29   * - for 40-50% more performance in the SSE-functions, the
     30   *   data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned !
     31   */
     32 
     33 #ifdef USE_SSE_ASM
     34 #include "assyntax.h"
     35 #include "matypes.h"
     36 #include "xform_args.h"
     37 
     38    SEG_TEXT
     39 
     40 #define S(i) 	REGOFF(i * 4, ESI)
     41 #define D(i) 	REGOFF(i * 4, EDI)
     42 #define M(i) 	REGOFF(i * 4, EDX)
     43 
     44 
     45 ALIGNTEXT4
     46 GLOBL GLNAME(_mesa_sse_transform_points2_general)
     47 HIDDEN (_mesa_sse_transform_points2_general)
     48 GLNAME( _mesa_sse_transform_points2_general ):
     49 
     50 #define FRAME_OFFSET 8
     51     PUSH_L    ( ESI )
     52     PUSH_L    ( EDI )
     53 
     54     MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
     55     MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
     56 
     57     MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
     58     MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
     59 
     60     TEST_L( ECX, ECX )
     61     JZ( LLBL(K_GTP2GR_finish) )			/* count was zero; go to finish */
     62 
     63     MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
     64     OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
     65 
     66     MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
     67     MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
     68 
     69     SHL_L( CONST(4), ECX ) 			/* count *= 16 */
     70     MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
     71 
     72     MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
     73     ADD_L( EDI, ECX ) 				/* count += dest ptr */
     74 
     75 ALIGNTEXT32
     76     MOVAPS( M(0), XMM0 )			/* m3  | m2  | m1  | m0 */
     77     MOVAPS( M(4), XMM1 )			/* m7  | m6  | m5  | m4 */
     78     MOVAPS( M(12), XMM2 )			/* m15 | m14 | m13 | m12 */
     79 
     80 ALIGNTEXT32
     81 LLBL(K_GTP2GR_top):
     82     MOVSS( S(0), XMM3 )				/* ox */
     83     SHUFPS( CONST(0x0), XMM3, XMM3 )		/* ox | ox | ox | ox */
     84     MULPS( XMM0, XMM3 )				/* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
     85     MOVSS( S(1), XMM4 )				/* oy */
     86     SHUFPS( CONST(0x0), XMM4, XMM4 )		/* oy | oy | oy | oy */
     87     MULPS( XMM1, XMM4 )				/* oy*m7 | oy*m6 | oy*m5 | oy*m4 */
     88 
     89     ADDPS( XMM4, XMM3 )
     90     ADDPS( XMM2, XMM3 )
     91     MOVAPS( XMM3, D(0) )
     92 
     93 LLBL(K_GTP2GR_skip):
     94     ADD_L     ( CONST(16), EDI )
     95     ADD_L     ( EAX, ESI )
     96     CMP_L     ( ECX, EDI )
     97     JNE       ( LLBL(K_GTP2GR_top) )
     98 
     99 LLBL(K_GTP2GR_finish):
    100     POP_L     ( EDI )
    101     POP_L     ( ESI )
    102     RET
    103 #undef FRAME_OFFSET
    104 
    105 
    106 ALIGNTEXT4
    107 GLOBL GLNAME(_mesa_sse_transform_points2_identity)
    108 HIDDEN(_mesa_sse_transform_points2_identity)
    109 GLNAME( _mesa_sse_transform_points2_identity ):
    110 
    111 #define FRAME_OFFSET 8
    112     PUSH_L    ( ESI )
    113     PUSH_L    ( EDI )
    114 
    115     MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
    116     MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
    117 
    118     MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
    119 
    120     TEST_L( ECX, ECX)
    121     JZ( LLBL(K_GTP2IR_finish) )			/* count was zero; go to finish */
    122 
    123     MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
    124     OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
    125 
    126     MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
    127     MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
    128 
    129     SHL_L( CONST(4), ECX ) 			/* count *= 16 */
    130     MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
    131 
    132     MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
    133     ADD_L( EDI, ECX ) 				/* count += dest ptr */
    134 
    135     CMP_L( ESI, EDI )
    136     JE( LLBL(K_GTP2IR_finish) )
    137 
    138 
    139 ALIGNTEXT32
    140 LLBL(K_GTP2IR_top):
    141     MOV_L     ( S(0), EDX )
    142     MOV_L     ( EDX, D(0) )
    143     MOV_L     ( S(1), EDX )
    144     MOV_L     ( EDX, D(1) )
    145 
    146 LLBL(K_GTP2IR_skip):
    147     ADD_L     ( CONST(16), EDI )
    148     ADD_L     ( EAX, ESI )
    149     CMP_L     ( ECX, EDI )
    150     JNE       ( LLBL(K_GTP2IR_top) )
    151 
    152 LLBL(K_GTP2IR_finish):
    153     POP_L     ( EDI )
    154     POP_L     ( ESI )
    155     RET
    156 #undef FRAME_OFFSET
    157 
    158 
    159 ALIGNTEXT4
    160 GLOBL GLNAME(_mesa_sse_transform_points2_3d_no_rot)
    161 HIDDEN(_mesa_sse_transform_points2_3d_no_rot)
    162 GLNAME(_mesa_sse_transform_points2_3d_no_rot):
    163 
    164 #define FRAME_OFFSET 8
    165     PUSH_L( ESI )
    166     PUSH_L( EDI )
    167 
    168     MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
    169     MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
    170 
    171     MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
    172     MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
    173 
    174     TEST_L( ECX, ECX)
    175     JZ( LLBL(K_GTP23DNRR_finish) ) 		/* count was zero; go to finish */
    176 
    177     MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
    178     OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
    179 
    180     MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
    181     MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
    182 
    183     SHL_L( CONST(4), ECX ) 			/* count *= 16 */
    184     MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
    185 
    186     MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
    187     ADD_L( EDI, ECX ) 				/* count += dest ptr */
    188 
    189     XORPS( XMM0, XMM0 )                         /* clean the working register */
    190 
    191 ALIGNTEXT32
    192     MOVSS    ( M(0), XMM1 )			/* - | - |  -  | m0  */
    193     MOVSS    ( M(5), XMM2 )			/* - | - |  -  | m5  */
    194     UNPCKLPS ( XMM2, XMM1 )			/* - | - | m5  | m0  */
    195     MOVLPS   ( M(12), XMM2 )			/* - | - | m13 | m12 */
    196     MOVSS    ( M(14), XMM3 )			/* - | - |  -  | m14 */
    197 
    198 ALIGNTEXT32
    199 LLBL(K_GTP23DNRR_top):
    200     MOVLPS   ( S(0), XMM0 )			/* - | - |  oy   | ox */
    201     MULPS    ( XMM1, XMM0 )			/* - | - | oy*m5 | ox*m0 */
    202     ADDPS    ( XMM2, XMM0 )			/* - | - | +m13  | +m12 */
    203     MOVLPS   ( XMM0, D(0) )			/* -> D(1) | -> D(0) */
    204 
    205     MOVSS    ( XMM3, D(2) )			/* -> D(2) */
    206 
    207 LLBL(K_GTP23DNRR_skip):
    208     ADD_L    ( CONST(16), EDI )
    209     ADD_L    ( EAX, ESI )
    210     CMP_L    ( ECX, EDI )
    211     JNE      ( LLBL(K_GTP23DNRR_top) )
    212 
    213 LLBL(K_GTP23DNRR_finish):
    214     POP_L    ( EDI )
    215     POP_L    ( ESI )
    216     RET
    217 #undef FRAME_OFFSET
    218 
    219 
    220 ALIGNTEXT4
    221 GLOBL GLNAME(_mesa_sse_transform_points2_perspective)
    222 HIDDEN(_mesa_sse_transform_points2_perspective)
    223 GLNAME(_mesa_sse_transform_points2_perspective):
    224 
    225 #define FRAME_OFFSET 8
    226     PUSH_L   ( ESI )
    227     PUSH_L   ( EDI )
    228 
    229     MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI )	/* ptr to source GLvector4f */
    230     MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
    231 
    232     MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
    233     MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
    234 
    235     TEST_L( ECX, ECX)
    236     JZ( LLBL(K_GTP23PR_finish) )		/* count was zero; go to finish */
    237 
    238     MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
    239     OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
    240 
    241     MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
    242     MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
    243 
    244     SHL_L( CONST(4), ECX ) 			/* count *= 16 */
    245     MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
    246 
    247     MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
    248     ADD_L( EDI, ECX ) 				/* count += dest ptr */
    249 
    250 ALIGNTEXT32
    251     MOVSS    ( M(0), XMM1 )			/* -  | -  |  -  | m0  */
    252     MOVSS    ( M(5), XMM2 )			/* -  | -  |  -  | m5  */
    253     UNPCKLPS ( XMM2, XMM1 )			/* -  | -  | m5  | m0  */
    254     MOVSS    ( M(14), XMM3 )			/* m14 */
    255     XORPS    ( XMM0, XMM0 )			/* 0 | 0 | 0 | 0 */
    256 
    257 ALIGNTEXT32
    258 LLBL(K_GTP23PR_top):
    259     MOVLPS( S(0), XMM4 )			/* oy | ox */
    260     MULPS( XMM1, XMM4 )				/* oy*m5 | ox*m0 */
    261     MOVLPS( XMM4, D(0) )			/* ->D(1) | ->D(0) */
    262     MOVSS( XMM3, D(2) )				/* ->D(2) */
    263     MOVSS( XMM0, D(3) )				/* ->D(3) */
    264 
    265 LLBL(K_GTP23PR_skip):
    266     ADD_L( CONST(16), EDI )
    267     ADD_L( EAX, ESI )
    268     CMP_L( ECX, EDI )
    269     JNE( LLBL(K_GTP23PR_top) )
    270 
    271 LLBL(K_GTP23PR_finish):
    272     POP_L    ( EDI )
    273     POP_L    ( ESI )
    274     RET
    275 #undef FRAME_OFFSET
    276 
    277 
    278 
    279 ALIGNTEXT4
    280 GLOBL GLNAME(_mesa_sse_transform_points2_2d)
    281 HIDDEN(_mesa_sse_transform_points2_2d)
    282 GLNAME(_mesa_sse_transform_points2_2d):
    283 
    284 #define FRAME_OFFSET 8
    285     PUSH_L( ESI )
    286     PUSH_L( EDI )
    287 
    288     MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
    289     MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
    290 
    291     MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
    292     MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
    293 
    294     TEST_L( ECX, ECX)
    295     JZ( LLBL(K_GTP23P2DR_finish) ) 		/* count was zero; go to finish */
    296 
    297     MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
    298     OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
    299 
    300     MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
    301     MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
    302 
    303     SHL_L( CONST(4), ECX ) 			/* count *= 16 */
    304     MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
    305 
    306     MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
    307     ADD_L( EDI, ECX ) 				/* count += dest ptr */
    308 
    309 ALIGNTEXT32
    310     MOVLPS( M(0), XMM0 )			/* m1  | m0 */
    311     MOVLPS( M(4), XMM1 )			/* m5  | m4 */
    312     MOVLPS( M(12), XMM2 )			/* m13 | m12 */
    313 
    314 ALIGNTEXT32
    315 LLBL(K_GTP23P2DR_top):
    316     MOVSS( S(0), XMM3 )				/* ox */
    317     SHUFPS( CONST(0x0), XMM3, XMM3 )		/* ox | ox */
    318     MULPS( XMM0, XMM3 )				/* ox*m1 | ox*m0 */
    319 
    320     MOVSS( S(1), XMM4 )				/* oy */
    321     SHUFPS( CONST(0x0), XMM4, XMM4 )		/* oy | oy */
    322     MULPS( XMM1, XMM4 )				/* oy*m5 | oy*m4 */
    323 
    324     ADDPS( XMM4, XMM3 )
    325     ADDPS( XMM2, XMM3 )
    326     MOVLPS( XMM3, D(0) )			/* ->D(1) | ->D(0) */
    327 
    328 LLBL(K_GTP23P2DR_skip):
    329     ADD_L    ( CONST(16), EDI )
    330     ADD_L    ( EAX, ESI )
    331     CMP_L    ( ECX, EDI )
    332     JNE      ( LLBL(K_GTP23P2DR_top) )
    333 
    334 LLBL(K_GTP23P2DR_finish):
    335     POP_L    ( EDI )
    336     POP_L    ( ESI )
    337     RET
    338 #undef FRAME_OFFSET
    339 
    340 
    341 
    342 ALIGNTEXT4
    343 GLOBL GLNAME(_mesa_sse_transform_points2_2d_no_rot)
    344 HIDDEN(_mesa_sse_transform_points2_2d_no_rot)
    345 GLNAME(_mesa_sse_transform_points2_2d_no_rot):
    346 
    347 #define FRAME_OFFSET 8
    348 	PUSH_L( ESI )
    349 	PUSH_L( EDI )
    350 
    351 	MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
    352 	MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
    353 
    354 	MOV_L( ARG_MATRIX, EDX ) 		/* ptr to matrix */
    355 	MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
    356 
    357 	TEST_L( ECX, ECX)
    358 	JZ( LLBL(K_GTP23P2DNRR_finish) ) 	/* count was zero; go to finish */
    359 
    360 	MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
    361 	OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
    362 
    363 	MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
    364 	MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
    365 
    366 	SHL_L( CONST(4), ECX ) 			/* count *= 16 */
    367 	MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
    368 
    369 	MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
    370 	ADD_L( EDI, ECX ) 			/* count += dest ptr */
    371 
    372 ALIGNTEXT32
    373 	MOVSS    ( M(0), XMM1 )			/* m0 */
    374 	MOVSS    ( M(5), XMM2 )			/* m5 */
    375 	UNPCKLPS ( XMM2, XMM1 )			/* m5 | m0 */
    376 	MOVLPS   ( M(12), XMM2 )		/* m13 | m12 */
    377 
    378 ALIGNTEXT32
    379 LLBL(K_GTP23P2DNRR_top):
    380 	MOVLPS( S(0), XMM0 )			/* oy | ox */
    381 	MULPS( XMM1, XMM0 )			/* oy*m5 | ox*m0 */
    382 	ADDPS( XMM2, XMM0 )			/* +m13 | +m12 */
    383 	MOVLPS( XMM0, D(0) )			/* ->D(1) | ->D(0) */
    384 
    385 LLBL(K_GTP23P2DNRR_skip):
    386 	ADD_L( CONST(16), EDI )
    387 	ADD_L( EAX, ESI )
    388 	CMP_L( ECX, EDI )
    389 	JNE( LLBL(K_GTP23P2DNRR_top) )
    390 
    391 LLBL(K_GTP23P2DNRR_finish):
    392 	POP_L( EDI )
    393 	POP_L( ESI )
    394 	RET
    395 #undef FRAME_OFFSET
    396 
    397 
    398 
    399 ALIGNTEXT4
    400 GLOBL GLNAME(_mesa_sse_transform_points2_3d)
    401 HIDDEN(_mesa_sse_transform_points2_3d)
    402 GLNAME(_mesa_sse_transform_points2_3d):
    403 
    404 #define FRAME_OFFSET 8
    405 	PUSH_L( ESI )
    406 	PUSH_L( EDI )
    407 
    408 	MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
    409 	MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */
    410 
    411 	MOV_L( ARG_MATRIX, EDX ) 		/* ptr to matrix */
    412 	MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */
    413 
    414 	TEST_L( ECX, ECX)
    415 	JZ( LLBL(K_GTP23P3DR_finish) ) 	/* count was zero; go to finish */
    416 
    417 	MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
    418 	OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */
    419 
    420 	MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
    421 	MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */
    422 
    423 	SHL_L( CONST(4), ECX ) 			/* count *= 16 */
    424 	MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */
    425 
    426 	MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
    427 	ADD_L( EDI, ECX ) 			/* count += dest ptr */
    428 
    429 ALIGNTEXT32
    430 	MOVAPS( M(0), XMM0 )			/* m2  | m1  | m0 */
    431 	MOVAPS( M(4), XMM1 )			/* m6  | m5  | m4 */
    432 	MOVAPS( M(12), XMM2 )			/* m14 | m13 | m12 */
    433 
    434 ALIGNTEXT32
    435 LLBL(K_GTP23P3DR_top):
    436 	MOVSS( S(0), XMM3 )			/* ox */
    437 	SHUFPS( CONST(0x0), XMM3, XMM3 )	/* ox | ox | ox */
    438 	MULPS( XMM0, XMM3 )			/* ox*m2 | ox*m1 | ox*m0 */
    439 
    440 	MOVSS( S(1), XMM4 )			/* oy */
    441 	SHUFPS( CONST(0x0), XMM4, XMM4 )	/* oy | oy | oy */
    442 	MULPS( XMM1, XMM4 )			/* oy*m6 | oy*m5 | oy*m4 */
    443 
    444 	ADDPS( XMM4, XMM3 )
    445 	ADDPS( XMM2, XMM3 )
    446 
    447 	MOVLPS( XMM3, D(0) )			/* ->D(1) | ->D(0) */
    448 	UNPCKHPS( XMM3, XMM3 )
    449 	MOVSS( XMM3, D(2) )			/* ->D(2) */
    450 
    451 LLBL(K_GTP23P3DR_skip):
    452 	ADD_L( CONST(16), EDI )
    453 	ADD_L( EAX, ESI )
    454 	CMP_L( ECX, EDI )
    455 	JNE( LLBL(K_GTP23P3DR_top) )
    456 
    457 LLBL(K_GTP23P3DR_finish):
    458 	POP_L( EDI )
    459 	POP_L( ESI )
    460 	RET
    461 #undef FRAME_OFFSET
    462 #endif
    463 
    464 #if defined (__ELF__) && defined (__linux__)
    465 	.section .note.GNU-stack,"",%progbits
    466 #endif
    467