mesa: Change "BRIAN PAUL" to "THE AUTHORS" in license text.
[mesa.git] / src / gallium / auxiliary / rtasm / rtasm_x86sse.h
1 /**************************************************************************
2 *
3 * Copyright (C) 1999-2005 Brian Paul All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice shall be included
13 * in all copies or substantial portions of the Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
16 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
19 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
20 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
21 *
22 **************************************************************************/
23
24 #ifndef _RTASM_X86SSE_H_
25 #define _RTASM_X86SSE_H_
26
27 #include "pipe/p_compiler.h"
28 #include "pipe/p_config.h"
29
30 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
31
32 /* It is up to the caller to ensure that instructions issued are
33 * suitable for the host cpu. There are no checks made in this module
34 * for mmx/sse/sse2 support on the cpu.
35 */
36 struct x86_reg {
37 unsigned file:2;
38 unsigned idx:4;
39 unsigned mod:2; /* mod_REG if this is just a register */
40 int disp:24; /* only +/- 23bits of offset - should be enough... */
41 };
42
43 #define X86_MMX 1
44 #define X86_MMX2 2
45 #define X86_SSE 4
46 #define X86_SSE2 8
47 #define X86_SSE3 0x10
48 #define X86_SSE4_1 0x20
49
50 struct x86_function {
51 unsigned caps;
52 unsigned size;
53 unsigned char *store;
54 unsigned char *csr;
55
56 unsigned stack_offset:16;
57 unsigned need_emms:8;
58 int x87_stack:8;
59
60 unsigned char error_overflow[4];
61 };
62
63 enum x86_reg_file {
64 file_REG32,
65 file_MMX,
66 file_XMM,
67 file_x87
68 };
69
70 /* Values for mod field of modr/m byte
71 */
72 enum x86_reg_mod {
73 mod_INDIRECT,
74 mod_DISP8,
75 mod_DISP32,
76 mod_REG
77 };
78
79 enum x86_reg_name {
80 reg_AX,
81 reg_CX,
82 reg_DX,
83 reg_BX,
84 reg_SP,
85 reg_BP,
86 reg_SI,
87 reg_DI,
88 reg_R8,
89 reg_R9,
90 reg_R10,
91 reg_R11,
92 reg_R12,
93 reg_R13,
94 reg_R14,
95 reg_R15
96 };
97
98
99 enum x86_cc {
100 cc_O, /* overflow */
101 cc_NO, /* not overflow */
102 cc_NAE, /* not above or equal / carry */
103 cc_AE, /* above or equal / not carry */
104 cc_E, /* equal / zero */
105 cc_NE /* not equal / not zero */
106 };
107
108 enum sse_cc {
109 cc_Equal,
110 cc_LessThan,
111 cc_LessThanEqual,
112 cc_Unordered,
113 cc_NotEqual,
114 cc_NotLessThan,
115 cc_NotLessThanEqual,
116 cc_Ordered
117 };
118
119 #define cc_Z cc_E
120 #define cc_NZ cc_NE
121
122
123 /** generic pointer to function */
124 typedef void (*x86_func)(void);
125
126
127 /* Begin/end/retrieve function creation:
128 */
129
130 enum x86_target
131 {
132 X86_32,
133 X86_64_STD_ABI,
134 X86_64_WIN64_ABI
135 };
136
137 /* make this read a member of x86_function if target != host is desired */
138 static INLINE enum x86_target x86_target( struct x86_function* p )
139 {
140 #ifdef PIPE_ARCH_X86
141 return X86_32;
142 #elif defined(_WIN64)
143 return X86_64_WIN64_ABI;
144 #elif defined(PIPE_ARCH_X86_64)
145 return X86_64_STD_ABI;
146 #endif
147 }
148
149 static INLINE unsigned x86_target_caps( struct x86_function* p )
150 {
151 return p->caps;
152 }
153
154 void x86_init_func( struct x86_function *p );
155 void x86_init_func_size( struct x86_function *p, unsigned code_size );
156 void x86_release_func( struct x86_function *p );
157 x86_func x86_get_func( struct x86_function *p );
158
159 /* Debugging:
160 */
161 void x86_print_reg( struct x86_reg reg );
162
163
164 /* Create and manipulate registers and regmem values:
165 */
166 struct x86_reg x86_make_reg( enum x86_reg_file file,
167 enum x86_reg_name idx );
168
169 struct x86_reg x86_make_disp( struct x86_reg reg,
170 int disp );
171
172 struct x86_reg x86_deref( struct x86_reg reg );
173
174 struct x86_reg x86_get_base_reg( struct x86_reg reg );
175
176
177 /* Labels, jumps and fixup:
178 */
179 int x86_get_label( struct x86_function *p );
180
181 void x64_rexw(struct x86_function *p);
182
183 void x86_jcc( struct x86_function *p,
184 enum x86_cc cc,
185 int label );
186
187 int x86_jcc_forward( struct x86_function *p,
188 enum x86_cc cc );
189
190 int x86_jmp_forward( struct x86_function *p);
191
192 int x86_call_forward( struct x86_function *p);
193
194 void x86_fixup_fwd_jump( struct x86_function *p,
195 int fixup );
196
197 void x86_jmp( struct x86_function *p, int label );
198
199 /* void x86_call( struct x86_function *p, void (*label)() ); */
200 void x86_call( struct x86_function *p, struct x86_reg reg);
201
202 void x86_mov_reg_imm( struct x86_function *p, struct x86_reg dst, int imm );
203 void x86_add_imm( struct x86_function *p, struct x86_reg dst, int imm );
204 void x86_or_imm( struct x86_function *p, struct x86_reg dst, int imm );
205 void x86_and_imm( struct x86_function *p, struct x86_reg dst, int imm );
206 void x86_sub_imm( struct x86_function *p, struct x86_reg dst, int imm );
207 void x86_xor_imm( struct x86_function *p, struct x86_reg dst, int imm );
208 void x86_cmp_imm( struct x86_function *p, struct x86_reg dst, int imm );
209
210
211 /* Macro for sse_shufps() and sse2_pshufd():
212 */
213 #define SHUF(_x,_y,_z,_w) (((_x)<<0) | ((_y)<<2) | ((_z)<<4) | ((_w)<<6))
214 #define SHUF_NOOP RSW(0,1,2,3)
215 #define GET_SHUF(swz, idx) (((swz) >> ((idx)*2)) & 0x3)
216
217 void mmx_emms( struct x86_function *p );
218 void mmx_movd( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
219 void mmx_movq( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
220 void mmx_packssdw( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
221 void mmx_packuswb( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
222
223 void sse2_movd( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
224 void sse2_movq( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
225 void sse2_movdqu( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
226 void sse2_movdqa( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
227 void sse2_movsd( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
228 void sse2_movupd( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
229 void sse2_movapd( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
230
231 void sse2_cvtps2dq( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
232 void sse2_cvttps2dq( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
233 void sse2_cvtdq2ps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
234 void sse2_cvtsd2ss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
235 void sse2_cvtpd2ps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
236
237 void sse2_movd( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
238 void sse2_packssdw( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
239 void sse2_packsswb( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
240 void sse2_packuswb( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
241 void sse2_pshufd( struct x86_function *p, struct x86_reg dest, struct x86_reg arg0,
242 unsigned char shuf );
243 void sse2_pshuflw( struct x86_function *p, struct x86_reg dest, struct x86_reg arg0,
244 unsigned char shuf );
245 void sse2_pshufhw( struct x86_function *p, struct x86_reg dest, struct x86_reg arg0,
246 unsigned char shuf );
247 void sse2_rcpps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
248 void sse2_rcpss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
249
250 void sse2_punpcklbw( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
251 void sse2_punpcklwd( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
252 void sse2_punpckldq( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
253 void sse2_punpcklqdq( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
254
255 void sse2_psllw_imm( struct x86_function *p, struct x86_reg dst, unsigned imm );
256 void sse2_pslld_imm( struct x86_function *p, struct x86_reg dst, unsigned imm );
257 void sse2_psllq_imm( struct x86_function *p, struct x86_reg dst, unsigned imm );
258
259 void sse2_psrlw_imm( struct x86_function *p, struct x86_reg dst, unsigned imm );
260 void sse2_psrld_imm( struct x86_function *p, struct x86_reg dst, unsigned imm );
261 void sse2_psrlq_imm( struct x86_function *p, struct x86_reg dst, unsigned imm );
262
263 void sse2_psraw_imm( struct x86_function *p, struct x86_reg dst, unsigned imm );
264 void sse2_psrad_imm( struct x86_function *p, struct x86_reg dst, unsigned imm );
265
266 void sse2_por( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
267
268 void sse2_pshuflw( struct x86_function *p, struct x86_reg dst, struct x86_reg src, uint8_t imm );
269 void sse2_pshufhw( struct x86_function *p, struct x86_reg dst, struct x86_reg src, uint8_t imm );
270 void sse2_pshufd( struct x86_function *p, struct x86_reg dst, struct x86_reg src, uint8_t imm );
271
272 void sse_prefetchnta( struct x86_function *p, struct x86_reg ptr);
273 void sse_prefetch0( struct x86_function *p, struct x86_reg ptr);
274 void sse_prefetch1( struct x86_function *p, struct x86_reg ptr);
275
276 void sse_movntps( struct x86_function *p, struct x86_reg dst, struct x86_reg src);
277
278 void sse_addps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
279 void sse_addss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
280 void sse_cvtps2pi( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
281 void sse_divss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
282 void sse_andnps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
283 void sse_andps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
284 void sse_cmpps( struct x86_function *p, struct x86_reg dst, struct x86_reg src,
285 enum sse_cc cc );
286 void sse_maxps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
287 void sse_maxss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
288 void sse_minps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
289 void sse_movaps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
290 void sse_movhlps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
291 void sse_movhps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
292 void sse_movlhps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
293 void sse_movlps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
294 void sse_movss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
295 void sse_movups( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
296 void sse_mulps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
297 void sse_mulss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
298 void sse_orps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
299 void sse_xorps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
300 void sse_subps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
301 void sse_rsqrtps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
302 void sse_rsqrtss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
303 void sse_shufps( struct x86_function *p, struct x86_reg dest, struct x86_reg arg0,
304 unsigned char shuf );
305 void sse_unpckhps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
306 void sse_unpcklps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
307 void sse_pmovmskb( struct x86_function *p, struct x86_reg dest, struct x86_reg src );
308 void sse_movmskps( struct x86_function *p, struct x86_reg dst, struct x86_reg src);
309
310 void x86_add( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
311 void x86_and( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
312 void x86_cmovcc( struct x86_function *p, struct x86_reg dst, struct x86_reg src, enum x86_cc cc );
313 void x86_cmp( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
314 void x86_dec( struct x86_function *p, struct x86_reg reg );
315 void x86_inc( struct x86_function *p, struct x86_reg reg );
316 void x86_lea( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
317 void x86_mov( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
318 void x64_mov64( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
319 void x86_mov8( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
320 void x86_mov16( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
321 void x86_movzx8(struct x86_function *p, struct x86_reg dst, struct x86_reg src );
322 void x86_movzx16(struct x86_function *p, struct x86_reg dst, struct x86_reg src );
323 void x86_mov_imm(struct x86_function *p, struct x86_reg dst, int imm );
324 void x86_mov8_imm(struct x86_function *p, struct x86_reg dst, uint8_t imm );
325 void x86_mov16_imm(struct x86_function *p, struct x86_reg dst, uint16_t imm );
326 void x86_mul( struct x86_function *p, struct x86_reg src );
327 void x86_imul( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
328 void x86_or( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
329 void x86_pop( struct x86_function *p, struct x86_reg reg );
330 void x86_push( struct x86_function *p, struct x86_reg reg );
331 void x86_push_imm32( struct x86_function *p, int imm );
332 void x86_ret( struct x86_function *p );
333 void x86_retw( struct x86_function *p, unsigned short imm );
334 void x86_sub( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
335 void x86_test( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
336 void x86_xor( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
337 void x86_sahf( struct x86_function *p );
338 void x86_div( struct x86_function *p, struct x86_reg src );
339 void x86_bswap( struct x86_function *p, struct x86_reg src );
340 void x86_shr_imm( struct x86_function *p, struct x86_reg reg, unsigned imm );
341 void x86_sar_imm( struct x86_function *p, struct x86_reg reg, unsigned imm );
342 void x86_shl_imm( struct x86_function *p, struct x86_reg reg, unsigned imm );
343
344 void x86_cdecl_caller_push_regs( struct x86_function *p );
345 void x86_cdecl_caller_pop_regs( struct x86_function *p );
346
347 void x87_assert_stack_empty( struct x86_function *p );
348
349 void x87_f2xm1( struct x86_function *p );
350 void x87_fabs( struct x86_function *p );
351 void x87_fadd( struct x86_function *p, struct x86_reg dst, struct x86_reg arg );
352 void x87_faddp( struct x86_function *p, struct x86_reg dst );
353 void x87_fchs( struct x86_function *p );
354 void x87_fclex( struct x86_function *p );
355 void x87_fcmovb( struct x86_function *p, struct x86_reg src );
356 void x87_fcmovbe( struct x86_function *p, struct x86_reg src );
357 void x87_fcmove( struct x86_function *p, struct x86_reg src );
358 void x87_fcmovnb( struct x86_function *p, struct x86_reg src );
359 void x87_fcmovnbe( struct x86_function *p, struct x86_reg src );
360 void x87_fcmovne( struct x86_function *p, struct x86_reg src );
361 void x87_fcom( struct x86_function *p, struct x86_reg dst );
362 void x87_fcomi( struct x86_function *p, struct x86_reg dst );
363 void x87_fcomip( struct x86_function *p, struct x86_reg dst );
364 void x87_fcomp( struct x86_function *p, struct x86_reg dst );
365 void x87_fcos( struct x86_function *p );
366 void x87_fdiv( struct x86_function *p, struct x86_reg dst, struct x86_reg arg );
367 void x87_fdivp( struct x86_function *p, struct x86_reg dst );
368 void x87_fdivr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg );
369 void x87_fdivrp( struct x86_function *p, struct x86_reg dst );
370 void x87_fild( struct x86_function *p, struct x86_reg arg );
371 void x87_fist( struct x86_function *p, struct x86_reg dst );
372 void x87_fistp( struct x86_function *p, struct x86_reg dst );
373 void x87_fld( struct x86_function *p, struct x86_reg arg );
374 void x87_fld1( struct x86_function *p );
375 void x87_fldcw( struct x86_function *p, struct x86_reg arg );
376 void x87_fldl2e( struct x86_function *p );
377 void x87_fldln2( struct x86_function *p );
378 void x87_fldz( struct x86_function *p );
379 void x87_fmul( struct x86_function *p, struct x86_reg dst, struct x86_reg arg );
380 void x87_fmulp( struct x86_function *p, struct x86_reg dst );
381 void x87_fnclex( struct x86_function *p );
382 void x87_fprndint( struct x86_function *p );
383 void x87_fpop( struct x86_function *p );
384 void x87_fscale( struct x86_function *p );
385 void x87_fsin( struct x86_function *p );
386 void x87_fsincos( struct x86_function *p );
387 void x87_fsqrt( struct x86_function *p );
388 void x87_fst( struct x86_function *p, struct x86_reg dst );
389 void x87_fstp( struct x86_function *p, struct x86_reg dst );
390 void x87_fsub( struct x86_function *p, struct x86_reg dst, struct x86_reg arg );
391 void x87_fsubp( struct x86_function *p, struct x86_reg dst );
392 void x87_fsubr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg );
393 void x87_fsubrp( struct x86_function *p, struct x86_reg dst );
394 void x87_ftst( struct x86_function *p );
395 void x87_fxch( struct x86_function *p, struct x86_reg dst );
396 void x87_fxtract( struct x86_function *p );
397 void x87_fyl2x( struct x86_function *p );
398 void x87_fyl2xp1( struct x86_function *p );
399 void x87_fwait( struct x86_function *p );
400 void x87_fnstcw( struct x86_function *p, struct x86_reg dst );
401 void x87_fnstsw( struct x86_function *p, struct x86_reg dst );
402 void x87_fucompp( struct x86_function *p );
403 void x87_fucomp( struct x86_function *p, struct x86_reg arg );
404 void x87_fucom( struct x86_function *p, struct x86_reg arg );
405
406
407
408 /* Retrieve a reference to one of the function arguments, taking into
409 * account any push/pop activity. Note - doesn't track explicit
410 * manipulation of ESP by other instructions.
411 */
412 struct x86_reg x86_fn_arg( struct x86_function *p, unsigned arg );
413
414 #endif
415 #endif