Remove CVS keywords.
[mesa.git] / src / mesa / x86 / x86_cliptest.S
1
2 /*
3 * Mesa 3-D graphics library
4 * Version: 3.5
5 *
6 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice shall be included
16 * in all copies or substantial portions of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
22 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
23 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
24 */
25
26 /*
27 * NOTE: Avoid using spaces in between '(' ')' and arguments, especially
28 * with macros like CONST, LLBL that expand to CONCAT(...). Putting spaces
29 * in there will break the build on some platforms.
30 */
31
32 #include "matypes.h"
33 #include "clip_args.h"
34
35 #define SRC0 REGOFF(0, ESI)
36 #define SRC1 REGOFF(4, ESI)
37 #define SRC2 REGOFF(8, ESI)
38 #define SRC3 REGOFF(12, ESI)
39 #define DST0 REGOFF(0, EDI)
40 #define DST1 REGOFF(4, EDI)
41 #define DST2 REGOFF(8, EDI)
42 #define DST3 REGOFF(12, EDI)
43 #define MAT0 REGOFF(0, EDX)
44 #define MAT1 REGOFF(4, EDX)
45 #define MAT2 REGOFF(8, EDX)
46 #define MAT3 REGOFF(12, EDX)
47
48
49 /*
50 * Table for clip test.
51 *
52 * bit6 = SRC3 < 0
53 * bit5 = SRC2 < 0
54 * bit4 = abs(S(2)) > abs(S(3))
55 * bit3 = SRC1 < 0
56 * bit2 = abs(S(1)) > abs(S(3))
57 * bit1 = SRC0 < 0
58 * bit0 = abs(S(0)) > abs(S(3))
59 */
60
61 SEG_DATA
62
63 clip_table:
64 D_BYTE 0x00, 0x01, 0x00, 0x02, 0x04, 0x05, 0x04, 0x06
65 D_BYTE 0x00, 0x01, 0x00, 0x02, 0x08, 0x09, 0x08, 0x0a
66 D_BYTE 0x20, 0x21, 0x20, 0x22, 0x24, 0x25, 0x24, 0x26
67 D_BYTE 0x20, 0x21, 0x20, 0x22, 0x28, 0x29, 0x28, 0x2a
68 D_BYTE 0x00, 0x01, 0x00, 0x02, 0x04, 0x05, 0x04, 0x06
69 D_BYTE 0x00, 0x01, 0x00, 0x02, 0x08, 0x09, 0x08, 0x0a
70 D_BYTE 0x10, 0x11, 0x10, 0x12, 0x14, 0x15, 0x14, 0x16
71 D_BYTE 0x10, 0x11, 0x10, 0x12, 0x18, 0x19, 0x18, 0x1a
72 D_BYTE 0x3f, 0x3d, 0x3f, 0x3e, 0x37, 0x35, 0x37, 0x36
73 D_BYTE 0x3f, 0x3d, 0x3f, 0x3e, 0x3b, 0x39, 0x3b, 0x3a
74 D_BYTE 0x2f, 0x2d, 0x2f, 0x2e, 0x27, 0x25, 0x27, 0x26
75 D_BYTE 0x2f, 0x2d, 0x2f, 0x2e, 0x2b, 0x29, 0x2b, 0x2a
76 D_BYTE 0x3f, 0x3d, 0x3f, 0x3e, 0x37, 0x35, 0x37, 0x36
77 D_BYTE 0x3f, 0x3d, 0x3f, 0x3e, 0x3b, 0x39, 0x3b, 0x3a
78 D_BYTE 0x1f, 0x1d, 0x1f, 0x1e, 0x17, 0x15, 0x17, 0x16
79 D_BYTE 0x1f, 0x1d, 0x1f, 0x1e, 0x1b, 0x19, 0x1b, 0x1a
80
81
82 SEG_TEXT
83
84 /*
85 * _mesa_x86_cliptest_points4
86 *
87 * AL: ormask
88 * AH: andmask
89 * EBX: temp0
90 * ECX: temp1
91 * EDX: clipmask[]
92 * ESI: clip[]
93 * EDI: proj[]
94 * EBP: temp2
95 */
96
97 #if defined(__ELF__) && defined(__PIC__) && defined(GNU_ASSEMBLER) && !defined(ELFPIC)
98 #define ELFPIC
99 #endif
100
101 ALIGNTEXT16
102 GLOBL GLNAME( _mesa_x86_cliptest_points4 )
103 HIDDEN(_mesa_x86_cliptest_points4)
104 GLNAME( _mesa_x86_cliptest_points4 ):
105
106 #ifdef ELFPIC
107 #define FRAME_OFFSET 20
108 #else
109 #define FRAME_OFFSET 16
110 #endif
111 PUSH_L( ESI )
112 PUSH_L( EDI )
113 PUSH_L( EBP )
114 PUSH_L( EBX )
115
116 #ifdef ELFPIC
117 /* store pointer to clip_table on stack */
118 CALL( LLBL(ctp4_get_eip) )
119 ADD_L( CONST(_GLOBAL_OFFSET_TABLE_), EBX )
120 MOV_L( REGOFF(clip_table@GOT, EBX), EBX )
121 PUSH_L( EBX )
122 JMP( LLBL(ctp4_clip_table_ready) )
123
124 LLBL(ctp4_get_eip):
125 /* store eip in ebx */
126 MOV_L( REGIND(ESP), EBX )
127 RET
128
129 LLBL(ctp4_clip_table_ready):
130 #endif
131
132 MOV_L( ARG_SOURCE, ESI )
133 MOV_L( ARG_DEST, EDI )
134
135 MOV_L( ARG_CLIP, EDX )
136 MOV_L( ARG_OR, EBX )
137
138 MOV_L( ARG_AND, EBP )
139 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
140
141 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
142 MOV_L( REGOFF(V4F_START, ESI), ESI )
143
144 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
145 MOV_L( EAX, ARG_SOURCE ) /* put stride in ARG_SOURCE */
146
147 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
148 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
149
150 MOV_L( REGOFF(V4F_START, EDI), EDI )
151 ADD_L( EDX, ECX )
152
153 MOV_L( ECX, ARG_CLIP ) /* put clipmask + count in ARG_CLIP */
154 CMP_L( ECX, EDX )
155
156 MOV_B( REGIND(EBX), AL )
157 MOV_B( REGIND(EBP), AH )
158
159 JZ( LLBL(ctp4_finish) )
160
161 ALIGNTEXT16
162 LLBL(ctp4_top):
163
164 FLD1 /* F3 */
165 FDIV_S( SRC3 ) /* GH: don't care about div-by-zero */
166
167 MOV_L( SRC3, EBP )
168 MOV_L( SRC2, EBX )
169
170 XOR_L( ECX, ECX )
171 ADD_L( EBP, EBP ) /* ebp = abs(S(3))*2 ; carry = sign of S(3) */
172
173 ADC_L( ECX, ECX )
174 ADD_L( EBX, EBX ) /* ebx = abs(S(2))*2 ; carry = sign of S(2) */
175
176 ADC_L( ECX, ECX )
177 CMP_L( EBX, EBP ) /* carry = abs(S(2))*2 > abs(S(3))*2 */
178
179 ADC_L( ECX, ECX )
180 MOV_L( SRC1, EBX )
181
182 ADD_L( EBX, EBX ) /* ebx = abs(S(1))*2 ; carry = sign of S(1) */
183
184 ADC_L( ECX, ECX )
185 CMP_L( EBX, EBP ) /* carry = abs(S(1))*2 > abs(S(3))*2 */
186
187 ADC_L( ECX, ECX )
188 MOV_L( SRC0, EBX )
189
190 ADD_L( EBX, EBX ) /* ebx = abs(S(0))*2 ; carry = sign of S(0) */
191
192 ADC_L( ECX, ECX )
193 CMP_L( EBX, EBP ) /* carry = abs(S(0))*2 > abs(S(3))*2 */
194
195 ADC_L( ECX, ECX )
196
197 #ifdef ELFPIC
198 MOV_L( REGIND(ESP), EBP ) /* clip_table */
199
200 MOV_B( REGBI(EBP, ECX), CL )
201 #else
202 MOV_B( REGOFF(clip_table,ECX), CL )
203 #endif
204
205 OR_B( CL, AL )
206 AND_B( CL, AH )
207
208 TEST_B( CL, CL )
209 MOV_B( CL, REGIND(EDX) )
210
211 JZ( LLBL(ctp4_proj) )
212
213 LLBL(ctp4_noproj):
214
215 FSTP( ST(0) ) /* */
216
217 MOV_L( CONST(0), DST0 )
218 MOV_L( CONST(0), DST1 )
219 MOV_L( CONST(0), DST2 )
220 MOV_L( CONST(0x3f800000), DST3 )
221
222 JMP( LLBL(ctp4_next) )
223
224 LLBL(ctp4_proj):
225
226 FLD_S( SRC0 ) /* F0 F3 */
227 FMUL2( ST(1), ST0 )
228
229 FLD_S( SRC1 ) /* F1 F0 F3 */
230 FMUL2( ST(2), ST0 )
231
232 FLD_S( SRC2 ) /* F2 F1 F0 F3 */
233 FMUL2( ST(3), ST0 )
234
235 FXCH( ST(2) ) /* F0 F1 F2 F3 */
236 FSTP_S( DST0 ) /* F1 F2 F3 */
237 FSTP_S( DST1 ) /* F2 F3 */
238 FSTP_S( DST2 ) /* F3 */
239 FSTP_S( DST3 ) /* */
240
241 LLBL(ctp4_next):
242
243 INC_L( EDX )
244 ADD_L( CONST(16), EDI )
245
246 ADD_L( ARG_SOURCE, ESI )
247 CMP_L( EDX, ARG_CLIP )
248
249 JNZ( LLBL(ctp4_top) )
250
251 MOV_L( ARG_OR, ECX )
252 MOV_L( ARG_AND, EDX )
253
254 MOV_B( AL, REGIND(ECX) )
255 MOV_B( AH, REGIND(EDX) )
256
257 LLBL(ctp4_finish):
258
259 MOV_L( ARG_DEST, EAX )
260 #ifdef ELFPIC
261 POP_L( ESI ) /* discard ptr to clip_table */
262 #endif
263 POP_L( EBX )
264 POP_L( EBP )
265 POP_L( EDI )
266 POP_L( ESI )
267
268 RET
269
270
271
272
273
274
275
276 ALIGNTEXT16
277 GLOBL GLNAME( _mesa_x86_cliptest_points4_np )
278 HIDDEN(_mesa_x86_cliptest_points4_np)
279 GLNAME( _mesa_x86_cliptest_points4_np ):
280
281 #ifdef ELFPIC
282 #define FRAME_OFFSET 20
283 #else
284 #define FRAME_OFFSET 16
285 #endif
286 PUSH_L( ESI )
287 PUSH_L( EDI )
288 PUSH_L( EBP )
289 PUSH_L( EBX )
290
291 #ifdef ELFPIC
292 /* store pointer to clip_table on stack */
293 CALL( LLBL(ctp4_np_get_eip) )
294 ADD_L( CONST(_GLOBAL_OFFSET_TABLE_), EBX )
295 MOV_L( REGOFF(clip_table@GOT, EBX), EBX )
296 PUSH_L( EBX )
297 JMP( LLBL(ctp4_np_clip_table_ready) )
298
299 LLBL(ctp4_np_get_eip):
300 /* store eip in ebx */
301 MOV_L( REGIND(ESP), EBX )
302 RET
303
304 LLBL(ctp4_np_clip_table_ready):
305 #endif
306
307 MOV_L( ARG_SOURCE, ESI )
308 /* slot */
309
310 MOV_L( ARG_CLIP, EDX )
311 MOV_L( ARG_OR, EBX )
312
313 MOV_L( ARG_AND, EBP )
314 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
315
316 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
317 MOV_L( REGOFF(V4F_START, ESI), ESI )
318
319 MOV_L( EAX, ARG_DEST ) /* put stride in ARG_DEST */
320 ADD_L( EDX, ECX )
321
322 MOV_L( ECX, EDI ) /* put clipmask + count in EDI */
323 CMP_L( ECX, EDX )
324
325 MOV_B( REGIND(EBX), AL )
326 MOV_B( REGIND(EBP), AH )
327
328 JZ( LLBL(ctp4_np_finish) )
329
330 ALIGNTEXT16
331 LLBL(ctp4_np_top):
332
333 MOV_L( SRC3, EBP )
334 MOV_L( SRC2, EBX )
335
336 XOR_L( ECX, ECX )
337 ADD_L( EBP, EBP ) /* ebp = abs(S(3))*2 ; carry = sign of S(3) */
338
339 ADC_L( ECX, ECX )
340 ADD_L( EBX, EBX ) /* ebx = abs(S(2))*2 ; carry = sign of S(2) */
341
342 ADC_L( ECX, ECX )
343 CMP_L( EBX, EBP ) /* carry = abs(S(2))*2 > abs(S(3))*2 */
344
345 ADC_L( ECX, ECX )
346 MOV_L( SRC1, EBX )
347
348 ADD_L( EBX, EBX ) /* ebx = abs(S(1))*2 ; carry = sign of S(1) */
349
350 ADC_L( ECX, ECX )
351 CMP_L( EBX, EBP ) /* carry = abs(S(1))*2 > abs(S(3))*2 */
352
353 ADC_L( ECX, ECX )
354 MOV_L( SRC0, EBX )
355
356 ADD_L( EBX, EBX ) /* ebx = abs(S(0))*2 ; carry = sign of S(0) */
357
358 ADC_L( ECX, ECX )
359 CMP_L( EBX, EBP ) /* carry = abs(S(0))*2 > abs(S(3))*2 */
360
361 ADC_L( ECX, ECX )
362
363 #ifdef ELFPIC
364 MOV_L( REGIND(ESP), EBP ) /* clip_table */
365
366 MOV_B( REGBI(EBP, ECX), CL )
367 #else
368 MOV_B( REGOFF(clip_table,ECX), CL )
369 #endif
370
371 OR_B( CL, AL )
372 AND_B( CL, AH )
373
374 TEST_B( CL, CL )
375 MOV_B( CL, REGIND(EDX) )
376
377 INC_L( EDX )
378 /* slot */
379
380 ADD_L( ARG_DEST, ESI )
381 CMP_L( EDX, EDI )
382
383 JNZ( LLBL(ctp4_np_top) )
384
385 MOV_L( ARG_OR, ECX )
386 MOV_L( ARG_AND, EDX )
387
388 MOV_B( AL, REGIND(ECX) )
389 MOV_B( AH, REGIND(EDX) )
390
391 LLBL(ctp4_np_finish):
392
393 MOV_L( ARG_SOURCE, EAX )
394 #ifdef ELFPIC
395 POP_L( ESI ) /* discard ptr to clip_table */
396 #endif
397 POP_L( EBX )
398 POP_L( EBP )
399 POP_L( EDI )
400 POP_L( ESI )
401
402 RET
403
404 #if defined (__ELF__) && defined (__linux__)
405 .section .note.GNU-stack,"",%progbits
406 #endif