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/*
27848b8605Smrg * NOTE: Avoid using spaces in between '(' ')' and arguments, especially
28848b8605Smrg * with macros like CONST, LLBL that expand to CONCAT(...).  Putting spaces
29848b8605Smrg * in there will break the build on some platforms.
30848b8605Smrg */
31848b8605Smrg
32848b8605Smrg#include "assyntax.h"
33848b8605Smrg#include "matypes.h"
34848b8605Smrg#include "clip_args.h"
35848b8605Smrg
36848b8605Smrg#define SRC0		REGOFF(0, ESI)
37848b8605Smrg#define SRC1		REGOFF(4, ESI)
38848b8605Smrg#define SRC2		REGOFF(8, ESI)
39848b8605Smrg#define SRC3		REGOFF(12, ESI)
40848b8605Smrg#define DST0		REGOFF(0, EDI)
41848b8605Smrg#define DST1		REGOFF(4, EDI)
42848b8605Smrg#define DST2		REGOFF(8, EDI)
43848b8605Smrg#define DST3		REGOFF(12, EDI)
44848b8605Smrg#define MAT0		REGOFF(0, EDX)
45848b8605Smrg#define MAT1		REGOFF(4, EDX)
46848b8605Smrg#define MAT2		REGOFF(8, EDX)
47848b8605Smrg#define MAT3		REGOFF(12, EDX)
48848b8605Smrg
49848b8605Smrg
50848b8605Smrg/*
51848b8605Smrg * Table for clip test.
52848b8605Smrg *
53848b8605Smrg * 	bit6 = SRC3 < 0
54848b8605Smrg * 	bit5 = SRC2 < 0
55848b8605Smrg * 	bit4 = abs(S(2)) > abs(S(3))
56848b8605Smrg * 	bit3 = SRC1 < 0
57848b8605Smrg * 	bit2 = abs(S(1)) > abs(S(3))
58848b8605Smrg * 	bit1 = SRC0 < 0
59848b8605Smrg * 	bit0 = abs(S(0)) > abs(S(3))
60848b8605Smrg */
61848b8605Smrg
62848b8605Smrg	SEG_DATA
63848b8605Smrg
64848b8605Smrgclip_table:
65848b8605Smrg	D_BYTE 0x00, 0x01, 0x00, 0x02, 0x04, 0x05, 0x04, 0x06
66848b8605Smrg	D_BYTE 0x00, 0x01, 0x00, 0x02, 0x08, 0x09, 0x08, 0x0a
67848b8605Smrg	D_BYTE 0x20, 0x21, 0x20, 0x22, 0x24, 0x25, 0x24, 0x26
68848b8605Smrg	D_BYTE 0x20, 0x21, 0x20, 0x22, 0x28, 0x29, 0x28, 0x2a
69848b8605Smrg	D_BYTE 0x00, 0x01, 0x00, 0x02, 0x04, 0x05, 0x04, 0x06
70848b8605Smrg	D_BYTE 0x00, 0x01, 0x00, 0x02, 0x08, 0x09, 0x08, 0x0a
71848b8605Smrg	D_BYTE 0x10, 0x11, 0x10, 0x12, 0x14, 0x15, 0x14, 0x16
72848b8605Smrg	D_BYTE 0x10, 0x11, 0x10, 0x12, 0x18, 0x19, 0x18, 0x1a
73848b8605Smrg	D_BYTE 0x3f, 0x3d, 0x3f, 0x3e, 0x37, 0x35, 0x37, 0x36
74848b8605Smrg	D_BYTE 0x3f, 0x3d, 0x3f, 0x3e, 0x3b, 0x39, 0x3b, 0x3a
75848b8605Smrg	D_BYTE 0x2f, 0x2d, 0x2f, 0x2e, 0x27, 0x25, 0x27, 0x26
76848b8605Smrg	D_BYTE 0x2f, 0x2d, 0x2f, 0x2e, 0x2b, 0x29, 0x2b, 0x2a
77848b8605Smrg	D_BYTE 0x3f, 0x3d, 0x3f, 0x3e, 0x37, 0x35, 0x37, 0x36
78848b8605Smrg	D_BYTE 0x3f, 0x3d, 0x3f, 0x3e, 0x3b, 0x39, 0x3b, 0x3a
79848b8605Smrg	D_BYTE 0x1f, 0x1d, 0x1f, 0x1e, 0x17, 0x15, 0x17, 0x16
80848b8605Smrg	D_BYTE 0x1f, 0x1d, 0x1f, 0x1e, 0x1b, 0x19, 0x1b, 0x1a
81848b8605Smrg
82848b8605Smrg
83848b8605Smrg	SEG_TEXT
84848b8605Smrg
85848b8605Smrg/*
86848b8605Smrg * _mesa_x86_cliptest_points4
87848b8605Smrg *
88848b8605Smrg *   AL:  ormask
89848b8605Smrg *   AH:  andmask
90848b8605Smrg *   EBX: temp0
91848b8605Smrg *   ECX: temp1
92848b8605Smrg *   EDX: clipmask[]
93848b8605Smrg *   ESI: clip[]
94848b8605Smrg *   EDI: proj[]
95848b8605Smrg *   EBP: temp2
96848b8605Smrg */
97848b8605Smrg
98848b8605Smrg#if defined(__ELF__) && defined(__PIC__) && defined(GNU_ASSEMBLER) && !defined(ELFPIC)
99848b8605Smrg#define ELFPIC
100848b8605Smrg#endif
101848b8605Smrg
102848b8605SmrgALIGNTEXT16
103848b8605SmrgGLOBL GLNAME( _mesa_x86_cliptest_points4 )
104848b8605SmrgHIDDEN(_mesa_x86_cliptest_points4)
105848b8605SmrgGLNAME( _mesa_x86_cliptest_points4 ):
106848b8605Smrg
107848b8605Smrg#ifdef ELFPIC
108848b8605Smrg#define FRAME_OFFSET 20
109848b8605Smrg#else
110848b8605Smrg#define FRAME_OFFSET 16
111848b8605Smrg#endif
112848b8605Smrg	PUSH_L( ESI )
113848b8605Smrg	PUSH_L( EDI )
114848b8605Smrg	PUSH_L( EBP )
115848b8605Smrg	PUSH_L( EBX )
116848b8605Smrg
117848b8605Smrg#ifdef ELFPIC
118848b8605Smrg	/* store pointer to clip_table on stack */
119848b8605Smrg	CALL( LLBL(ctp4_get_eip) )
120848b8605Smrg	ADD_L( CONST(_GLOBAL_OFFSET_TABLE_), EBX )
121848b8605Smrg	MOV_L( REGOFF(clip_table@GOT, EBX), EBX )
122848b8605Smrg	PUSH_L( EBX )
123848b8605Smrg	JMP( LLBL(ctp4_clip_table_ready) )
124848b8605Smrg
125848b8605SmrgLLBL(ctp4_get_eip):
126848b8605Smrg	/* store eip in ebx */
127848b8605Smrg	MOV_L( REGIND(ESP), EBX )
128848b8605Smrg	RET
129848b8605Smrg
130848b8605SmrgLLBL(ctp4_clip_table_ready):
131848b8605Smrg#endif
132848b8605Smrg
133848b8605Smrg	MOV_L( ARG_SOURCE, ESI )
134848b8605Smrg	MOV_L( ARG_DEST, EDI )
135848b8605Smrg
136848b8605Smrg	MOV_L( ARG_CLIP, EDX )
137848b8605Smrg	MOV_L( ARG_OR, EBX )
138848b8605Smrg
139848b8605Smrg	MOV_L( ARG_AND, EBP )
140848b8605Smrg	MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
141848b8605Smrg
142848b8605Smrg	MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
143848b8605Smrg	MOV_L( REGOFF(V4F_START, ESI), ESI )
144848b8605Smrg
145848b8605Smrg	OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
146848b8605Smrg	MOV_L( EAX, ARG_SOURCE )	/* put stride in ARG_SOURCE */
147848b8605Smrg
148848b8605Smrg	MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
149848b8605Smrg	MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
150848b8605Smrg
151848b8605Smrg	MOV_L( REGOFF(V4F_START, EDI), EDI )
152848b8605Smrg	ADD_L( EDX, ECX )
153848b8605Smrg
154848b8605Smrg	MOV_L( ECX, ARG_CLIP )		/* put clipmask + count in ARG_CLIP */
155848b8605Smrg	CMP_L( ECX, EDX )
156848b8605Smrg
157848b8605Smrg	MOV_B( REGIND(EBX), AL )
158848b8605Smrg	MOV_B( REGIND(EBP), AH )
159848b8605Smrg
160848b8605Smrg	JZ( LLBL(ctp4_finish) )
161848b8605Smrg
162848b8605SmrgALIGNTEXT16
163848b8605SmrgLLBL(ctp4_top):
164848b8605Smrg
165848b8605Smrg	FLD1				/* F3 */
166848b8605Smrg	FDIV_S( SRC3 )		/* GH: don't care about div-by-zero */
167848b8605Smrg
168848b8605Smrg	MOV_L( SRC3, EBP )
169848b8605Smrg	MOV_L( SRC2, EBX )
170848b8605Smrg
171848b8605Smrg	XOR_L( ECX, ECX )
172848b8605Smrg	ADD_L( EBP, EBP )	/* ebp = abs(S(3))*2 ; carry = sign of S(3) */
173848b8605Smrg
174848b8605Smrg	ADC_L( ECX, ECX )
175848b8605Smrg	ADD_L( EBX, EBX )	/* ebx = abs(S(2))*2 ; carry = sign of S(2) */
176848b8605Smrg
177848b8605Smrg	ADC_L( ECX, ECX )
178848b8605Smrg	CMP_L( EBX, EBP )	/* carry = abs(S(2))*2 > abs(S(3))*2 */
179848b8605Smrg
180848b8605Smrg	ADC_L( ECX, ECX )
181848b8605Smrg	MOV_L( SRC1, EBX )
182848b8605Smrg
183848b8605Smrg	ADD_L( EBX, EBX )	/* ebx = abs(S(1))*2 ; carry = sign of S(1) */
184848b8605Smrg
185848b8605Smrg	ADC_L( ECX, ECX )
186848b8605Smrg	CMP_L( EBX, EBP )	/* carry = abs(S(1))*2 > abs(S(3))*2 */
187848b8605Smrg
188848b8605Smrg	ADC_L( ECX, ECX )
189848b8605Smrg	MOV_L( SRC0, EBX )
190848b8605Smrg
191848b8605Smrg	ADD_L( EBX, EBX )	/* ebx = abs(S(0))*2 ; carry = sign of S(0) */
192848b8605Smrg
193848b8605Smrg	ADC_L( ECX, ECX )
194848b8605Smrg	CMP_L( EBX, EBP )	/* carry = abs(S(0))*2 > abs(S(3))*2 */
195848b8605Smrg
196848b8605Smrg	ADC_L( ECX, ECX )
197848b8605Smrg
198848b8605Smrg#ifdef ELFPIC
199848b8605Smrg	MOV_L( REGIND(ESP), EBP )	/* clip_table */
200848b8605Smrg
201848b8605Smrg	MOV_B( REGBI(EBP, ECX), CL )
202848b8605Smrg#else
203848b8605Smrg	MOV_B( REGOFF(clip_table,ECX), CL )
204848b8605Smrg#endif
205848b8605Smrg
206848b8605Smrg	OR_B( CL, AL )
207848b8605Smrg	AND_B( CL, AH )
208848b8605Smrg
209848b8605Smrg	TEST_B( CL, CL )
210848b8605Smrg	MOV_B( CL, REGIND(EDX) )
211848b8605Smrg
212848b8605Smrg	JZ( LLBL(ctp4_proj) )
213848b8605Smrg
214848b8605SmrgLLBL(ctp4_noproj):
215848b8605Smrg
216848b8605Smrg	FSTP( ST(0) )			/* */
217848b8605Smrg
218848b8605Smrg	MOV_L( CONST(0), DST0 )
219848b8605Smrg	MOV_L( CONST(0), DST1 )
220848b8605Smrg	MOV_L( CONST(0), DST2 )
221848b8605Smrg	MOV_L( CONST(0x3f800000), DST3 )
222848b8605Smrg
223848b8605Smrg	JMP( LLBL(ctp4_next) )
224848b8605Smrg
225848b8605SmrgLLBL(ctp4_proj):
226848b8605Smrg
227848b8605Smrg	FLD_S( SRC0 )			/* F0 F3 */
228848b8605Smrg	FMUL2( ST(1), ST0 )
229848b8605Smrg
230848b8605Smrg	FLD_S( SRC1 )			/* F1 F0 F3 */
231848b8605Smrg	FMUL2( ST(2), ST0 )
232848b8605Smrg
233848b8605Smrg	FLD_S( SRC2 )			/* F2 F1 F0 F3 */
234848b8605Smrg	FMUL2( ST(3), ST0 )
235848b8605Smrg
236848b8605Smrg	FXCH( ST(2) )			/* F0 F1 F2 F3 */
237848b8605Smrg	FSTP_S( DST0 )		/* F1 F2 F3 */
238848b8605Smrg	FSTP_S( DST1 )		/* F2 F3 */
239848b8605Smrg	FSTP_S( DST2 )		/* F3 */
240848b8605Smrg	FSTP_S( DST3 )		/* */
241848b8605Smrg
242848b8605SmrgLLBL(ctp4_next):
243848b8605Smrg
244848b8605Smrg	INC_L( EDX )
245848b8605Smrg	ADD_L( CONST(16), EDI )
246848b8605Smrg
247848b8605Smrg	ADD_L( ARG_SOURCE, ESI )
248848b8605Smrg	CMP_L( EDX, ARG_CLIP )
249848b8605Smrg
250848b8605Smrg	JNZ( LLBL(ctp4_top) )
251848b8605Smrg
252848b8605Smrg	MOV_L( ARG_OR, ECX )
253848b8605Smrg	MOV_L( ARG_AND, EDX )
254848b8605Smrg
255848b8605Smrg	MOV_B( AL, REGIND(ECX) )
256848b8605Smrg	MOV_B( AH, REGIND(EDX) )
257848b8605Smrg
258848b8605SmrgLLBL(ctp4_finish):
259848b8605Smrg
260848b8605Smrg	MOV_L( ARG_DEST, EAX )
261848b8605Smrg#ifdef ELFPIC
262848b8605Smrg	POP_L( ESI )			/* discard ptr to clip_table */
263848b8605Smrg#endif
264848b8605Smrg	POP_L( EBX )
265848b8605Smrg	POP_L( EBP )
266848b8605Smrg	POP_L( EDI )
267848b8605Smrg	POP_L( ESI )
268848b8605Smrg
269848b8605Smrg	RET
270848b8605Smrg
271848b8605Smrg
272848b8605Smrg
273848b8605Smrg
274848b8605Smrg
275848b8605Smrg
276848b8605Smrg
277848b8605SmrgALIGNTEXT16
278848b8605SmrgGLOBL GLNAME( _mesa_x86_cliptest_points4_np )
279848b8605SmrgHIDDEN(_mesa_x86_cliptest_points4_np)
280848b8605SmrgGLNAME( _mesa_x86_cliptest_points4_np ):
281848b8605Smrg
282848b8605Smrg#ifdef ELFPIC
283848b8605Smrg#define FRAME_OFFSET 20
284848b8605Smrg#else
285848b8605Smrg#define FRAME_OFFSET 16
286848b8605Smrg#endif
287848b8605Smrg	PUSH_L( ESI )
288848b8605Smrg	PUSH_L( EDI )
289848b8605Smrg	PUSH_L( EBP )
290848b8605Smrg	PUSH_L( EBX )
291848b8605Smrg
292848b8605Smrg#ifdef ELFPIC
293848b8605Smrg	/* store pointer to clip_table on stack */
294848b8605Smrg	CALL( LLBL(ctp4_np_get_eip) )
295848b8605Smrg	ADD_L( CONST(_GLOBAL_OFFSET_TABLE_), EBX )
296848b8605Smrg	MOV_L( REGOFF(clip_table@GOT, EBX), EBX )
297848b8605Smrg	PUSH_L( EBX )
298848b8605Smrg	JMP( LLBL(ctp4_np_clip_table_ready) )
299848b8605Smrg
300848b8605SmrgLLBL(ctp4_np_get_eip):
301848b8605Smrg	/* store eip in ebx */
302848b8605Smrg	MOV_L( REGIND(ESP), EBX )
303848b8605Smrg	RET
304848b8605Smrg
305848b8605SmrgLLBL(ctp4_np_clip_table_ready):
306848b8605Smrg#endif
307848b8605Smrg
308848b8605Smrg	MOV_L( ARG_SOURCE, ESI )
309848b8605Smrg	/* slot */
310848b8605Smrg
311848b8605Smrg	MOV_L( ARG_CLIP, EDX )
312848b8605Smrg	MOV_L( ARG_OR, EBX )
313848b8605Smrg
314848b8605Smrg	MOV_L( ARG_AND, EBP )
315848b8605Smrg	MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
316848b8605Smrg
317848b8605Smrg	MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
318848b8605Smrg	MOV_L( REGOFF(V4F_START, ESI), ESI )
319848b8605Smrg
320848b8605Smrg	MOV_L( EAX, ARG_DEST )   	/* put stride in ARG_DEST */
321848b8605Smrg	ADD_L( EDX, ECX )
322848b8605Smrg
323848b8605Smrg	MOV_L( ECX, EDI )		/* put clipmask + count in EDI */
324848b8605Smrg	CMP_L( ECX, EDX )
325848b8605Smrg
326848b8605Smrg	MOV_B( REGIND(EBX), AL )
327848b8605Smrg	MOV_B( REGIND(EBP), AH )
328848b8605Smrg
329848b8605Smrg	JZ( LLBL(ctp4_np_finish) )
330848b8605Smrg
331848b8605SmrgALIGNTEXT16
332848b8605SmrgLLBL(ctp4_np_top):
333848b8605Smrg
334848b8605Smrg	MOV_L( SRC3, EBP )
335848b8605Smrg	MOV_L( SRC2, EBX )
336848b8605Smrg
337848b8605Smrg	XOR_L( ECX, ECX )
338848b8605Smrg	ADD_L( EBP, EBP )	/* ebp = abs(S(3))*2 ; carry = sign of S(3) */
339848b8605Smrg
340848b8605Smrg	ADC_L( ECX, ECX )
341848b8605Smrg	ADD_L( EBX, EBX )	/* ebx = abs(S(2))*2 ; carry = sign of S(2) */
342848b8605Smrg
343848b8605Smrg	ADC_L( ECX, ECX )
344848b8605Smrg	CMP_L( EBX, EBP )	/* carry = abs(S(2))*2 > abs(S(3))*2 */
345848b8605Smrg
346848b8605Smrg	ADC_L( ECX, ECX )
347848b8605Smrg	MOV_L( SRC1, EBX )
348848b8605Smrg
349848b8605Smrg	ADD_L( EBX, EBX )	/* ebx = abs(S(1))*2 ; carry = sign of S(1) */
350848b8605Smrg
351848b8605Smrg	ADC_L( ECX, ECX )
352848b8605Smrg	CMP_L( EBX, EBP )	/* carry = abs(S(1))*2 > abs(S(3))*2 */
353848b8605Smrg
354848b8605Smrg	ADC_L( ECX, ECX )
355848b8605Smrg	MOV_L( SRC0, EBX )
356848b8605Smrg
357848b8605Smrg	ADD_L( EBX, EBX )	/* ebx = abs(S(0))*2 ; carry = sign of S(0) */
358848b8605Smrg
359848b8605Smrg	ADC_L( ECX, ECX )
360848b8605Smrg	CMP_L( EBX, EBP )	/* carry = abs(S(0))*2 > abs(S(3))*2 */
361848b8605Smrg
362848b8605Smrg	ADC_L( ECX, ECX )
363848b8605Smrg
364848b8605Smrg#ifdef ELFPIC
365848b8605Smrg	MOV_L( REGIND(ESP), EBP )	/* clip_table */
366848b8605Smrg
367848b8605Smrg	MOV_B( REGBI(EBP, ECX), CL )
368848b8605Smrg#else
369848b8605Smrg	MOV_B( REGOFF(clip_table,ECX), CL )
370848b8605Smrg#endif
371848b8605Smrg
372848b8605Smrg	OR_B( CL, AL )
373848b8605Smrg	AND_B( CL, AH )
374848b8605Smrg
375848b8605Smrg	TEST_B( CL, CL )
376848b8605Smrg	MOV_B( CL, REGIND(EDX) )
377848b8605Smrg
378848b8605Smrg	INC_L( EDX )
379848b8605Smrg	/* slot */
380848b8605Smrg
381848b8605Smrg	ADD_L( ARG_DEST, ESI )
382848b8605Smrg	CMP_L( EDX, EDI )
383848b8605Smrg
384848b8605Smrg	JNZ( LLBL(ctp4_np_top) )
385848b8605Smrg
386848b8605Smrg	MOV_L( ARG_OR, ECX )
387848b8605Smrg	MOV_L( ARG_AND, EDX )
388848b8605Smrg
389848b8605Smrg	MOV_B( AL, REGIND(ECX) )
390848b8605Smrg	MOV_B( AH, REGIND(EDX) )
391848b8605Smrg
392848b8605SmrgLLBL(ctp4_np_finish):
393848b8605Smrg
394848b8605Smrg	MOV_L( ARG_SOURCE, EAX )
395848b8605Smrg#ifdef ELFPIC
396848b8605Smrg	POP_L( ESI )			/* discard ptr to clip_table */
397848b8605Smrg#endif
398848b8605Smrg	POP_L( EBX )
399848b8605Smrg	POP_L( EBP )
400848b8605Smrg	POP_L( EDI )
401848b8605Smrg	POP_L( ESI )
402848b8605Smrg
403848b8605Smrg	RET
404848b8605Smrg
405848b8605Smrg#if defined (__ELF__) && defined (__linux__)
406848b8605Smrg	.section .note.GNU-stack,"",%progbits
407848b8605Smrg#endif
408