projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
rtasm: fix debug build
[mesa.git] / src / gallium / auxiliary / rtasm / rtasm_x86sse.c
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 * BRIAN PAUL 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 #include "pipe/p_config.h"
25
26 #if defined(PIPE_ARCH_X86)
27
28 #include "pipe/p_compiler.h"
29 #include "pipe/p_debug.h"
30 #include "util/u_pointer.h"
31
32 #include "rtasm_execmem.h"
33 #include "rtasm_x86sse.h"
34
35 #define DISASSEM 0
36 #define X86_TWOB 0x0f
37
38
39 #define DUMP_SSE 0
40
41
42 void x86_print_reg( struct x86_reg reg )
43 {
44 if (reg.mod != mod_REG)
45 debug_printf( "[" );
46
47 switch( reg.file ) {
48 case file_REG32:
49 switch( reg.idx ) {
50 case reg_AX: debug_printf( "EAX" ); break;
51 case reg_CX: debug_printf( "ECX" ); break;
52 case reg_DX: debug_printf( "EDX" ); break;
53 case reg_BX: debug_printf( "EBX" ); break;
54 case reg_SP: debug_printf( "ESP" ); break;
55 case reg_BP: debug_printf( "EBP" ); break;
56 case reg_SI: debug_printf( "ESI" ); break;
57 case reg_DI: debug_printf( "EDI" ); break;
58 }
59 break;
60 case file_MMX:
61 debug_printf( "MMX%u", reg.idx );
62 break;
63 case file_XMM:
64 debug_printf( "XMM%u", reg.idx );
65 break;
66 case file_x87:
67 debug_printf( "fp%u", reg.idx );
68 break;
69 }
70
71 if (reg.mod == mod_DISP8 ||
72 reg.mod == mod_DISP32)
73 debug_printf("+%d", reg.disp);
74
75 if (reg.mod != mod_REG)
76 debug_printf( "]" );
77 }
78
79 #if DUMP_SSE
80
81 #define DUMP_START() debug_printf( "\n" )
82 #define DUMP_END() debug_printf( "\n" )
83
84 #define DUMP() do { \
85 const char *foo = __FUNCTION__; \
86 while (*foo && *foo != '_') \
87 foo++; \
88 if (*foo) \
89 foo++; \
90 debug_printf( "\n% 4x% 15s ", p->csr - p->store, foo ); \
91 } while (0)
92
93 #define DUMP_I( I ) do { \
94 DUMP(); \
95 debug_printf( "%u", I ); \
96 } while( 0 )
97
98 #define DUMP_R( R0 ) do { \
99 DUMP(); \
100 x86_print_reg( R0 ); \
101 } while( 0 )
102
103 #define DUMP_RR( R0, R1 ) do { \
104 DUMP(); \
105 x86_print_reg( R0 ); \
106 debug_printf( ", " ); \
107 x86_print_reg( R1 ); \
108 } while( 0 )
109
110 #define DUMP_RI( R0, I ) do { \
111 DUMP(); \
112 x86_print_reg( R0 ); \
113 debug_printf( ", %u", I ); \
114 } while( 0 )
115
116 #define DUMP_RRI( R0, R1, I ) do { \
117 DUMP(); \
118 x86_print_reg( R0 ); \
119 debug_printf( ", " ); \
120 x86_print_reg( R1 ); \
121 debug_printf( ", %u", I ); \
122 } while( 0 )
123
124 #else
125
126 #define DUMP_START()
127 #define DUMP_END()
128 #define DUMP( )
129 #define DUMP_I( I )
130 #define DUMP_R( R0 )
131 #define DUMP_RR( R0, R1 )
132 #define DUMP_RI( R0, I )
133 #define DUMP_RRI( R0, R1, I )
134
135 #endif
136
137
138 static void do_realloc( struct x86_function *p )
139 {
140 if (p->store == p->error_overflow) {
141 p->csr = p->store;
142 }
143 else if (p->size == 0) {
144 p->size = 1024;
145 p->store = rtasm_exec_malloc(p->size);
146 p->csr = p->store;
147 }
148 else {
149 uintptr_t used = pointer_to_uintptr( p->csr ) - pointer_to_uintptr( p->store );
150 unsigned char *tmp = p->store;
151 p->size *= 2;
152 p->store = rtasm_exec_malloc(p->size);
153
154 if (p->store) {
155 memcpy(p->store, tmp, used);
156 p->csr = p->store + used;
157 }
158 else {
159 p->csr = p->store;
160 }
161
162 rtasm_exec_free(tmp);
163 }
164
165 if (p->store == NULL) {
166 p->store = p->csr = p->error_overflow;
167 p->size = sizeof(p->error_overflow);
168 }
169 }
170
171 /* Emit bytes to the instruction stream:
172 */
173 static unsigned char *reserve( struct x86_function *p, int bytes )
174 {
175 if (p->csr + bytes - p->store > (int) p->size)
176 do_realloc(p);
177
178 {
179 unsigned char *csr = p->csr;
180 p->csr += bytes;
181 return csr;
182 }
183 }
184
185
186
187 static void emit_1b( struct x86_function *p, char b0 )
188 {
189 char *csr = (char *)reserve(p, 1);
190 *csr = b0;
191 }
192
193 static void emit_1i( struct x86_function *p, int i0 )
194 {
195 int *icsr = (int *)reserve(p, sizeof(i0));
196 *icsr = i0;
197 }
198
199 static void emit_1ub( struct x86_function *p, unsigned char b0 )
200 {
201 unsigned char *csr = reserve(p, 1);
202 *csr++ = b0;
203 }
204
205 static void emit_2ub( struct x86_function *p, unsigned char b0, unsigned char b1 )
206 {
207 unsigned char *csr = reserve(p, 2);
208 *csr++ = b0;
209 *csr++ = b1;
210 }
211
212 static void emit_3ub( struct x86_function *p, unsigned char b0, unsigned char b1, unsigned char b2 )
213 {
214 unsigned char *csr = reserve(p, 3);
215 *csr++ = b0;
216 *csr++ = b1;
217 *csr++ = b2;
218 }
219
220
221 /* Build a modRM byte + possible displacement. No treatment of SIB
222 * indexing. BZZT - no way to encode an absolute address.
223 *
224 * This is the "/r" field in the x86 manuals...
225 */
226 static void emit_modrm( struct x86_function *p,
227 struct x86_reg reg,
228 struct x86_reg regmem )
229 {
230 unsigned char val = 0;
231
232 assert(reg.mod == mod_REG);
233
234 val |= regmem.mod << 6; /* mod field */
235 val |= reg.idx << 3; /* reg field */
236 val |= regmem.idx; /* r/m field */
237
238 emit_1ub(p, val);
239
240 /* Oh-oh we've stumbled into the SIB thing.
241 */
242 if (regmem.file == file_REG32 &&
243 regmem.idx == reg_SP) {
244 emit_1ub(p, 0x24); /* simplistic! */
245 }
246
247 switch (regmem.mod) {
248 case mod_REG:
249 case mod_INDIRECT:
250 break;
251 case mod_DISP8:
252 emit_1b(p, (char) regmem.disp);
253 break;
254 case mod_DISP32:
255 emit_1i(p, regmem.disp);
256 break;
257 default:
258 assert(0);
259 break;
260 }
261 }
262
263 /* Emits the "/0".."/7" specialized versions of the modrm ("/r") bytes.
264 */
265 static void emit_modrm_noreg( struct x86_function *p,
266 unsigned op,
267 struct x86_reg regmem )
268 {
269 struct x86_reg dummy = x86_make_reg(file_REG32, op);
270 emit_modrm(p, dummy, regmem);
271 }
272
273 /* Many x86 instructions have two opcodes to cope with the situations
274 * where the destination is a register or memory reference
275 * respectively. This function selects the correct opcode based on
276 * the arguments presented.
277 */
278 static void emit_op_modrm( struct x86_function *p,
279 unsigned char op_dst_is_reg,
280 unsigned char op_dst_is_mem,
281 struct x86_reg dst,
282 struct x86_reg src )
283 {
284 switch (dst.mod) {
285 case mod_REG:
286 emit_1ub(p, op_dst_is_reg);
287 emit_modrm(p, dst, src);
288 break;
289 case mod_INDIRECT:
290 case mod_DISP32:
291 case mod_DISP8:
292 assert(src.mod == mod_REG);
293 emit_1ub(p, op_dst_is_mem);
294 emit_modrm(p, src, dst);
295 break;
296 default:
297 assert(0);
298 break;
299 }
300 }
301
302
303
304
305
306
307
308 /* Create and manipulate registers and regmem values:
309 */
310 struct x86_reg x86_make_reg( enum x86_reg_file file,
311 enum x86_reg_name idx )
312 {
313 struct x86_reg reg;
314
315 reg.file = file;
316 reg.idx = idx;
317 reg.mod = mod_REG;
318 reg.disp = 0;
319
320 return reg;
321 }
322
323 struct x86_reg x86_make_disp( struct x86_reg reg,
324 int disp )
325 {
326 assert(reg.file == file_REG32);
327
328 if (reg.mod == mod_REG)
329 reg.disp = disp;
330 else
331 reg.disp += disp;
332
333 if (reg.disp == 0 && reg.idx != reg_BP)
334 reg.mod = mod_INDIRECT;
335 else if (reg.disp <= 127 && reg.disp >= -128)
336 reg.mod = mod_DISP8;
337 else
338 reg.mod = mod_DISP32;
339
340 return reg;
341 }
342
343 struct x86_reg x86_deref( struct x86_reg reg )
344 {
345 return x86_make_disp(reg, 0);
346 }
347
348 struct x86_reg x86_get_base_reg( struct x86_reg reg )
349 {
350 return x86_make_reg( reg.file, reg.idx );
351 }
352
353 int x86_get_label( struct x86_function *p )
354 {
355 return p->csr - p->store;
356 }
357
358
359
360 /***********************************************************************
361 * x86 instructions
362 */
363
364
365 void x86_jcc( struct x86_function *p,
366 enum x86_cc cc,
367 int label )
368 {
369 int offset = label - (x86_get_label(p) + 2);
370 DUMP_I(cc);
371
372 if (offset < 0) {
373 assert(p->csr - p->store > -offset);
374 }
375
376 if (offset <= 127 && offset >= -128) {
377 emit_1ub(p, 0x70 + cc);
378 emit_1b(p, (char) offset);
379 }
380 else {
381 offset = label - (x86_get_label(p) + 6);
382 emit_2ub(p, 0x0f, 0x80 + cc);
383 emit_1i(p, offset);
384 }
385 }
386
387 /* Always use a 32bit offset for forward jumps:
388 */
389 int x86_jcc_forward( struct x86_function *p,
390 enum x86_cc cc )
391 {
392 DUMP_I(cc);
393 emit_2ub(p, 0x0f, 0x80 + cc);
394 emit_1i(p, 0);
395 return x86_get_label(p);
396 }
397
398 int x86_jmp_forward( struct x86_function *p)
399 {
400 DUMP();
401 emit_1ub(p, 0xe9);
402 emit_1i(p, 0);
403 return x86_get_label(p);
404 }
405
406 int x86_call_forward( struct x86_function *p)
407 {
408 DUMP();
409
410 emit_1ub(p, 0xe8);
411 emit_1i(p, 0);
412 return x86_get_label(p);
413 }
414
415 /* Fixup offset from forward jump:
416 */
417 void x86_fixup_fwd_jump( struct x86_function *p,
418 int fixup )
419 {
420 *(int *)(p->store + fixup - 4) = x86_get_label(p) - fixup;
421 }
422
423 void x86_jmp( struct x86_function *p, int label)
424 {
425 DUMP_I( label );
426 emit_1ub(p, 0xe9);
427 emit_1i(p, label - x86_get_label(p) - 4);
428 }
429
430 void x86_call( struct x86_function *p, struct x86_reg reg)
431 {
432 DUMP_R( reg );
433 emit_1ub(p, 0xff);
434 emit_modrm_noreg(p, 2, reg);
435 }
436
437
438 /* michal:
439 * Temporary. As I need immediate operands, and dont want to mess with the codegen,
440 * I load the immediate into general purpose register and use it.
441 */
442 void x86_mov_reg_imm( struct x86_function *p, struct x86_reg dst, int imm )
443 {
444 DUMP_RI( dst, imm );
445 assert(dst.mod == mod_REG);
446 emit_1ub(p, 0xb8 + dst.idx);
447 emit_1i(p, imm);
448 }
449
450 void x86_add_reg_imm8( struct x86_function *p, struct x86_reg dst, ubyte imm )
451 {
452 DUMP_RI( dst, imm );
453 assert(dst.mod == mod_REG);
454 emit_1ub(p, 0x80);
455 emit_modrm_noreg(p, 0, dst);
456 emit_1ub(p, imm);
457 }
458
459
460 void x86_push( struct x86_function *p,
461 struct x86_reg reg )
462 {
463 DUMP_R( reg );
464 if (reg.mod == mod_REG)
465 emit_1ub(p, 0x50 + reg.idx);
466 else
467 {
468 emit_1ub(p, 0xff);
469 emit_modrm_noreg(p, 6, reg);
470 }
471
472
473 p->stack_offset += 4;
474 }
475
476 void x86_push_imm32( struct x86_function *p,
477 int imm32 )
478 {
479 DUMP_I( imm32 );
480 emit_1ub(p, 0x68);
481 emit_1i(p, imm32);
482
483 p->stack_offset += 4;
484 }
485
486
487 void x86_pop( struct x86_function *p,
488 struct x86_reg reg )
489 {
490 DUMP_R( reg );
491 assert(reg.mod == mod_REG);
492 emit_1ub(p, 0x58 + reg.idx);
493 p->stack_offset -= 4;
494 }
495
496 void x86_inc( struct x86_function *p,
497 struct x86_reg reg )
498 {
499 DUMP_R( reg );
500 assert(reg.mod == mod_REG);
501 emit_1ub(p, 0x40 + reg.idx);
502 }
503
504 void x86_dec( struct x86_function *p,
505 struct x86_reg reg )
506 {
507 DUMP_R( reg );
508 assert(reg.mod == mod_REG);
509 emit_1ub(p, 0x48 + reg.idx);
510 }
511
512 void x86_ret( struct x86_function *p )
513 {
514 DUMP();
515 assert(p->stack_offset == 0);
516 emit_1ub(p, 0xc3);
517 }
518
519 void x86_retw( struct x86_function *p, unsigned short imm )
520 {
521 DUMP();
522 emit_3ub(p, 0xc2, imm & 0xff, (imm >> 8) & 0xff);
523 }
524
525 void x86_sahf( struct x86_function *p )
526 {
527 DUMP();
528 emit_1ub(p, 0x9e);
529 }
530
531 void x86_mov( struct x86_function *p,
532 struct x86_reg dst,
533 struct x86_reg src )
534 {
535 DUMP_RR( dst, src );
536 emit_op_modrm( p, 0x8b, 0x89, dst, src );
537 }
538
539 void x86_xor( struct x86_function *p,
540 struct x86_reg dst,
541 struct x86_reg src )
542 {
543 DUMP_RR( dst, src );
544 emit_op_modrm( p, 0x33, 0x31, dst, src );
545 }
546
547 void x86_cmp( struct x86_function *p,
548 struct x86_reg dst,
549 struct x86_reg src )
550 {
551 DUMP_RR( dst, src );
552 emit_op_modrm( p, 0x3b, 0x39, dst, src );
553 }
554
555 void x86_lea( struct x86_function *p,
556 struct x86_reg dst,
557 struct x86_reg src )
558 {
559 DUMP_RR( dst, src );
560 emit_1ub(p, 0x8d);
561 emit_modrm( p, dst, src );
562 }
563
564 void x86_test( struct x86_function *p,
565 struct x86_reg dst,
566 struct x86_reg src )
567 {
568 DUMP_RR( dst, src );
569 emit_1ub(p, 0x85);
570 emit_modrm( p, dst, src );
571 }
572
573 void x86_add( struct x86_function *p,
574 struct x86_reg dst,
575 struct x86_reg src )
576 {
577 DUMP_RR( dst, src );
578 emit_op_modrm(p, 0x03, 0x01, dst, src );
579 }
580
581 /* Calculate EAX * src, results in EDX:EAX.
582 */
583 void x86_mul( struct x86_function *p,
584 struct x86_reg src )
585 {
586 DUMP_R( src );
587 emit_1ub(p, 0xf7);
588 emit_modrm_noreg(p, 4, src );
589 }
590
591
592 void x86_imul( struct x86_function *p,
593 struct x86_reg dst,
594 struct x86_reg src )
595 {
596 DUMP_RR( dst, src );
597 emit_2ub(p, X86_TWOB, 0xAF);
598 emit_modrm(p, dst, src);
599 }
600
601
602 void x86_sub( struct x86_function *p,
603 struct x86_reg dst,
604 struct x86_reg src )
605 {
606 DUMP_RR( dst, src );
607 emit_op_modrm(p, 0x2b, 0x29, dst, src );
608 }
609
610 void x86_or( struct x86_function *p,
611 struct x86_reg dst,
612 struct x86_reg src )
613 {
614 DUMP_RR( dst, src );
615 emit_op_modrm( p, 0x0b, 0x09, dst, src );
616 }
617
618 void x86_and( struct x86_function *p,
619 struct x86_reg dst,
620 struct x86_reg src )
621 {
622 DUMP_RR( dst, src );
623 emit_op_modrm( p, 0x23, 0x21, dst, src );
624 }
625
626
627
628 /***********************************************************************
629 * SSE instructions
630 */
631
632 void sse_prefetchnta( struct x86_function *p, struct x86_reg ptr)
633 {
634 DUMP_R( ptr );
635 assert(ptr.mod != mod_REG);
636 emit_2ub(p, 0x0f, 0x18);
637 emit_modrm_noreg(p, 0, ptr);
638 }
639
640 void sse_prefetch0( struct x86_function *p, struct x86_reg ptr)
641 {
642 DUMP_R( ptr );
643 assert(ptr.mod != mod_REG);
644 emit_2ub(p, 0x0f, 0x18);
645 emit_modrm_noreg(p, 1, ptr);
646 }
647
648 void sse_prefetch1( struct x86_function *p, struct x86_reg ptr)
649 {
650 DUMP_R( ptr );
651 assert(ptr.mod != mod_REG);
652 emit_2ub(p, 0x0f, 0x18);
653 emit_modrm_noreg(p, 2, ptr);
654 }
655
656 void sse_movntps( struct x86_function *p,
657 struct x86_reg dst,
658 struct x86_reg src)
659 {
660 DUMP_RR( dst, src );
661
662 assert(dst.mod != mod_REG);
663 assert(src.mod == mod_REG);
664 emit_2ub(p, 0x0f, 0x2b);
665 emit_modrm(p, src, dst);
666 }
667
668
669
670
671 void sse_movss( struct x86_function *p,
672 struct x86_reg dst,
673 struct x86_reg src )
674 {
675 DUMP_RR( dst, src );
676 emit_2ub(p, 0xF3, X86_TWOB);
677 emit_op_modrm( p, 0x10, 0x11, dst, src );
678 }
679
680 void sse_movaps( struct x86_function *p,
681 struct x86_reg dst,
682 struct x86_reg src )
683 {
684 DUMP_RR( dst, src );
685 emit_1ub(p, X86_TWOB);
686 emit_op_modrm( p, 0x28, 0x29, dst, src );
687 }
688
689 void sse_movups( struct x86_function *p,
690 struct x86_reg dst,
691 struct x86_reg src )
692 {
693 DUMP_RR( dst, src );
694 emit_1ub(p, X86_TWOB);
695 emit_op_modrm( p, 0x10, 0x11, dst, src );
696 }
697
698 void sse_movhps( struct x86_function *p,
699 struct x86_reg dst,
700 struct x86_reg src )
701 {
702 DUMP_RR( dst, src );
703 assert(dst.mod != mod_REG || src.mod != mod_REG);
704 emit_1ub(p, X86_TWOB);
705 emit_op_modrm( p, 0x16, 0x17, dst, src ); /* cf movlhps */
706 }
707
708 void sse_movlps( struct x86_function *p,
709 struct x86_reg dst,
710 struct x86_reg src )
711 {
712 DUMP_RR( dst, src );
713 assert(dst.mod != mod_REG || src.mod != mod_REG);
714 emit_1ub(p, X86_TWOB);
715 emit_op_modrm( p, 0x12, 0x13, dst, src ); /* cf movhlps */
716 }
717
718 void sse_maxps( struct x86_function *p,
719 struct x86_reg dst,
720 struct x86_reg src )
721 {
722 DUMP_RR( dst, src );
723 emit_2ub(p, X86_TWOB, 0x5F);
724 emit_modrm( p, dst, src );
725 }
726
727 void sse_maxss( struct x86_function *p,
728 struct x86_reg dst,
729 struct x86_reg src )
730 {
731 DUMP_RR( dst, src );
732 emit_3ub(p, 0xF3, X86_TWOB, 0x5F);
733 emit_modrm( p, dst, src );
734 }
735
736 void sse_divss( struct x86_function *p,
737 struct x86_reg dst,
738 struct x86_reg src )
739 {
740 DUMP_RR( dst, src );
741 emit_3ub(p, 0xF3, X86_TWOB, 0x5E);
742 emit_modrm( p, dst, src );
743 }
744
745 void sse_minps( struct x86_function *p,
746 struct x86_reg dst,
747 struct x86_reg src )
748 {
749 DUMP_RR( dst, src );
750 emit_2ub(p, X86_TWOB, 0x5D);
751 emit_modrm( p, dst, src );
752 }
753
754 void sse_subps( struct x86_function *p,
755 struct x86_reg dst,
756 struct x86_reg src )
757 {
758 DUMP_RR( dst, src );
759 emit_2ub(p, X86_TWOB, 0x5C);
760 emit_modrm( p, dst, src );
761 }
762
763 void sse_mulps( struct x86_function *p,
764 struct x86_reg dst,
765 struct x86_reg src )
766 {
767 DUMP_RR( dst, src );
768 emit_2ub(p, X86_TWOB, 0x59);
769 emit_modrm( p, dst, src );
770 }
771
772 void sse_mulss( struct x86_function *p,
773 struct x86_reg dst,
774 struct x86_reg src )
775 {
776 DUMP_RR( dst, src );
777 emit_3ub(p, 0xF3, X86_TWOB, 0x59);
778 emit_modrm( p, dst, src );
779 }
780
781 void sse_addps( struct x86_function *p,
782 struct x86_reg dst,
783 struct x86_reg src )
784 {
785 DUMP_RR( dst, src );
786 emit_2ub(p, X86_TWOB, 0x58);
787 emit_modrm( p, dst, src );
788 }
789
790 void sse_addss( struct x86_function *p,
791 struct x86_reg dst,
792 struct x86_reg src )
793 {
794 DUMP_RR( dst, src );
795 emit_3ub(p, 0xF3, X86_TWOB, 0x58);
796 emit_modrm( p, dst, src );
797 }
798
799 void sse_andnps( struct x86_function *p,
800 struct x86_reg dst,
801 struct x86_reg src )
802 {
803 DUMP_RR( dst, src );
804 emit_2ub(p, X86_TWOB, 0x55);
805 emit_modrm( p, dst, src );
806 }
807
808 void sse_andps( struct x86_function *p,
809 struct x86_reg dst,
810 struct x86_reg src )
811 {
812 DUMP_RR( dst, src );
813 emit_2ub(p, X86_TWOB, 0x54);
814 emit_modrm( p, dst, src );
815 }
816
817 void sse_rsqrtps( struct x86_function *p,
818 struct x86_reg dst,
819 struct x86_reg src )
820 {
821 DUMP_RR( dst, src );
822 emit_2ub(p, X86_TWOB, 0x52);
823 emit_modrm( p, dst, src );
824 }
825
826 void sse_rsqrtss( struct x86_function *p,
827 struct x86_reg dst,
828 struct x86_reg src )
829 {
830 DUMP_RR( dst, src );
831 emit_3ub(p, 0xF3, X86_TWOB, 0x52);
832 emit_modrm( p, dst, src );
833
834 }
835
836 void sse_movhlps( struct x86_function *p,
837 struct x86_reg dst,
838 struct x86_reg src )
839 {
840 DUMP_RR( dst, src );
841 assert(dst.mod == mod_REG && src.mod == mod_REG);
842 emit_2ub(p, X86_TWOB, 0x12);
843 emit_modrm( p, dst, src );
844 }
845
846 void sse_movlhps( struct x86_function *p,
847 struct x86_reg dst,
848 struct x86_reg src )
849 {
850 DUMP_RR( dst, src );
851 assert(dst.mod == mod_REG && src.mod == mod_REG);
852 emit_2ub(p, X86_TWOB, 0x16);
853 emit_modrm( p, dst, src );
854 }
855
856 void sse_orps( struct x86_function *p,
857 struct x86_reg dst,
858 struct x86_reg src )
859 {
860 DUMP_RR( dst, src );
861 emit_2ub(p, X86_TWOB, 0x56);
862 emit_modrm( p, dst, src );
863 }
864
865 void sse_xorps( struct x86_function *p,
866 struct x86_reg dst,
867 struct x86_reg src )
868 {
869 DUMP_RR( dst, src );
870 emit_2ub(p, X86_TWOB, 0x57);
871 emit_modrm( p, dst, src );
872 }
873
874 void sse_cvtps2pi( struct x86_function *p,
875 struct x86_reg dst,
876 struct x86_reg src )
877 {
878 DUMP_RR( dst, src );
879 assert(dst.file == file_MMX &&
880 (src.file == file_XMM || src.mod != mod_REG));
881
882 p->need_emms = 1;
883
884 emit_2ub(p, X86_TWOB, 0x2d);
885 emit_modrm( p, dst, src );
886 }
887
888 void sse2_cvtdq2ps( struct x86_function *p,
889 struct x86_reg dst,
890 struct x86_reg src )
891 {
892 DUMP_RR( dst, src );
893 emit_2ub(p, X86_TWOB, 0x5b);
894 emit_modrm( p, dst, src );
895 }
896
897
898 /* Shufps can also be used to implement a reduced swizzle when dest ==
899 * arg0.
900 */
901 void sse_shufps( struct x86_function *p,
902 struct x86_reg dst,
903 struct x86_reg src,
904 unsigned char shuf)
905 {
906 DUMP_RRI( dst, src, shuf );
907 emit_2ub(p, X86_TWOB, 0xC6);
908 emit_modrm(p, dst, src);
909 emit_1ub(p, shuf);
910 }
911
912 void sse_unpckhps( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
913 {
914 DUMP_RR( dst, src );
915 emit_2ub( p, X86_TWOB, 0x15 );
916 emit_modrm( p, dst, src );
917 }
918
919 void sse_unpcklps( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
920 {
921 DUMP_RR( dst, src );
922 emit_2ub( p, X86_TWOB, 0x14 );
923 emit_modrm( p, dst, src );
924 }
925
926 void sse_cmpps( struct x86_function *p,
927 struct x86_reg dst,
928 struct x86_reg src,
929 enum sse_cc cc)
930 {
931 DUMP_RRI( dst, src, cc );
932 emit_2ub(p, X86_TWOB, 0xC2);
933 emit_modrm(p, dst, src);
934 emit_1ub(p, cc);
935 }
936
937 void sse_pmovmskb( struct x86_function *p,
938 struct x86_reg dst,
939 struct x86_reg src)
940 {
941 DUMP_RR( dst, src );
942 emit_3ub(p, 0x66, X86_TWOB, 0xD7);
943 emit_modrm(p, dst, src);
944 }
945
946 /***********************************************************************
947 * SSE2 instructions
948 */
949
950 /**
951 * Perform a reduced swizzle:
952 */
953 void sse2_pshufd( struct x86_function *p,
954 struct x86_reg dst,
955 struct x86_reg src,
956 unsigned char shuf)
957 {
958 DUMP_RRI( dst, src, shuf );
959 emit_3ub(p, 0x66, X86_TWOB, 0x70);
960 emit_modrm(p, dst, src);
961 emit_1ub(p, shuf);
962 }
963
964 void sse2_cvttps2dq( struct x86_function *p,
965 struct x86_reg dst,
966 struct x86_reg src )
967 {
968 DUMP_RR( dst, src );
969 emit_3ub( p, 0xF3, X86_TWOB, 0x5B );
970 emit_modrm( p, dst, src );
971 }
972
973 void sse2_cvtps2dq( struct x86_function *p,
974 struct x86_reg dst,
975 struct x86_reg src )
976 {
977 DUMP_RR( dst, src );
978 emit_3ub(p, 0x66, X86_TWOB, 0x5B);
979 emit_modrm( p, dst, src );
980 }
981
982 void sse2_packssdw( struct x86_function *p,
983 struct x86_reg dst,
984 struct x86_reg src )
985 {
986 DUMP_RR( dst, src );
987 emit_3ub(p, 0x66, X86_TWOB, 0x6B);
988 emit_modrm( p, dst, src );
989 }
990
991 void sse2_packsswb( struct x86_function *p,
992 struct x86_reg dst,
993 struct x86_reg src )
994 {
995 DUMP_RR( dst, src );
996 emit_3ub(p, 0x66, X86_TWOB, 0x63);
997 emit_modrm( p, dst, src );
998 }
999
1000 void sse2_packuswb( struct x86_function *p,
1001 struct x86_reg dst,
1002 struct x86_reg src )
1003 {
1004 DUMP_RR( dst, src );
1005 emit_3ub(p, 0x66, X86_TWOB, 0x67);
1006 emit_modrm( p, dst, src );
1007 }
1008
1009 void sse2_punpcklbw( struct x86_function *p,
1010 struct x86_reg dst,
1011 struct x86_reg src )
1012 {
1013 DUMP_RR( dst, src );
1014 emit_3ub(p, 0x66, X86_TWOB, 0x60);
1015 emit_modrm( p, dst, src );
1016 }
1017
1018
1019 void sse2_rcpps( struct x86_function *p,
1020 struct x86_reg dst,
1021 struct x86_reg src )
1022 {
1023 DUMP_RR( dst, src );
1024 emit_2ub(p, X86_TWOB, 0x53);
1025 emit_modrm( p, dst, src );
1026 }
1027
1028 void sse2_rcpss( struct x86_function *p,
1029 struct x86_reg dst,
1030 struct x86_reg src )
1031 {
1032 DUMP_RR( dst, src );
1033 emit_3ub(p, 0xF3, X86_TWOB, 0x53);
1034 emit_modrm( p, dst, src );
1035 }
1036
1037 void sse2_movd( struct x86_function *p,
1038 struct x86_reg dst,
1039 struct x86_reg src )
1040 {
1041 DUMP_RR( dst, src );
1042 emit_2ub(p, 0x66, X86_TWOB);
1043 emit_op_modrm( p, 0x6e, 0x7e, dst, src );
1044 }
1045
1046
1047
1048
1049 /***********************************************************************
1050 * x87 instructions
1051 */
1052 static void note_x87_pop( struct x86_function *p )
1053 {
1054 p->x87_stack--;
1055 assert(p->x87_stack >= 0);
1056 }
1057
1058 static void note_x87_push( struct x86_function *p )
1059 {
1060 p->x87_stack++;
1061 assert(p->x87_stack <= 7);
1062 }
1063
1064 void x87_assert_stack_empty( struct x86_function *p )
1065 {
1066 assert (p->x87_stack == 0);
1067 }
1068
1069
1070 void x87_fist( struct x86_function *p, struct x86_reg dst )
1071 {
1072 DUMP_R( dst );
1073 emit_1ub(p, 0xdb);
1074 emit_modrm_noreg(p, 2, dst);
1075 }
1076
1077 void x87_fistp( struct x86_function *p, struct x86_reg dst )
1078 {
1079 DUMP_R( dst );
1080 emit_1ub(p, 0xdb);
1081 emit_modrm_noreg(p, 3, dst);
1082 note_x87_pop(p);
1083 }
1084
1085 void x87_fild( struct x86_function *p, struct x86_reg arg )
1086 {
1087 DUMP_R( arg );
1088 emit_1ub(p, 0xdf);
1089 emit_modrm_noreg(p, 0, arg);
1090 note_x87_push(p);
1091 }
1092
1093 void x87_fldz( struct x86_function *p )
1094 {
1095 DUMP();
1096 emit_2ub(p, 0xd9, 0xee);
1097 note_x87_push(p);
1098 }
1099
1100
1101 void x87_fldcw( struct x86_function *p, struct x86_reg arg )
1102 {
1103 DUMP_R( arg );
1104 assert(arg.file == file_REG32);
1105 assert(arg.mod != mod_REG);
1106 emit_1ub(p, 0xd9);
1107 emit_modrm_noreg(p, 5, arg);
1108 }
1109
1110 void x87_fld1( struct x86_function *p )
1111 {
1112 DUMP();
1113 emit_2ub(p, 0xd9, 0xe8);
1114 note_x87_push(p);
1115 }
1116
1117 void x87_fldl2e( struct x86_function *p )
1118 {
1119 DUMP();
1120 emit_2ub(p, 0xd9, 0xea);
1121 note_x87_push(p);
1122 }
1123
1124 void x87_fldln2( struct x86_function *p )
1125 {
1126 DUMP();
1127 emit_2ub(p, 0xd9, 0xed);
1128 note_x87_push(p);
1129 }
1130
1131 void x87_fwait( struct x86_function *p )
1132 {
1133 DUMP();
1134 emit_1ub(p, 0x9b);
1135 }
1136
1137 void x87_fnclex( struct x86_function *p )
1138 {
1139 DUMP();
1140 emit_2ub(p, 0xdb, 0xe2);
1141 }
1142
1143 void x87_fclex( struct x86_function *p )
1144 {
1145 x87_fwait(p);
1146 x87_fnclex(p);
1147 }
1148
1149 void x87_fcmovb( struct x86_function *p, struct x86_reg arg )
1150 {
1151 DUMP_R( arg );
1152 assert(arg.file == file_x87);
1153 emit_2ub(p, 0xda, 0xc0+arg.idx);
1154 }
1155
1156 void x87_fcmove( struct x86_function *p, struct x86_reg arg )
1157 {
1158 DUMP_R( arg );
1159 assert(arg.file == file_x87);
1160 emit_2ub(p, 0xda, 0xc8+arg.idx);
1161 }
1162
1163 void x87_fcmovbe( struct x86_function *p, struct x86_reg arg )
1164 {
1165 DUMP_R( arg );
1166 assert(arg.file == file_x87);
1167 emit_2ub(p, 0xda, 0xd0+arg.idx);
1168 }
1169
1170 void x87_fcmovnb( struct x86_function *p, struct x86_reg arg )
1171 {
1172 DUMP_R( arg );
1173 assert(arg.file == file_x87);
1174 emit_2ub(p, 0xdb, 0xc0+arg.idx);
1175 }
1176
1177 void x87_fcmovne( struct x86_function *p, struct x86_reg arg )
1178 {
1179 DUMP_R( arg );
1180 assert(arg.file == file_x87);
1181 emit_2ub(p, 0xdb, 0xc8+arg.idx);
1182 }
1183
1184 void x87_fcmovnbe( struct x86_function *p, struct x86_reg arg )
1185 {
1186 DUMP_R( arg );
1187 assert(arg.file == file_x87);
1188 emit_2ub(p, 0xdb, 0xd0+arg.idx);
1189 }
1190
1191
1192
1193 static void x87_arith_op( struct x86_function *p, struct x86_reg dst, struct x86_reg arg,
1194 unsigned char dst0ub0,
1195 unsigned char dst0ub1,
1196 unsigned char arg0ub0,
1197 unsigned char arg0ub1,
1198 unsigned char argmem_noreg)
1199 {
1200 assert(dst.file == file_x87);
1201
1202 if (arg.file == file_x87) {
1203 if (dst.idx == 0)
1204 emit_2ub(p, dst0ub0, dst0ub1+arg.idx);
1205 else if (arg.idx == 0)
1206 emit_2ub(p, arg0ub0, arg0ub1+arg.idx);
1207 else
1208 assert(0);
1209 }
1210 else if (dst.idx == 0) {
1211 assert(arg.file == file_REG32);
1212 emit_1ub(p, 0xd8);
1213 emit_modrm_noreg(p, argmem_noreg, arg);
1214 }
1215 else
1216 assert(0);
1217 }
1218
1219 void x87_fmul( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
1220 {
1221 DUMP_RR( dst, src );
1222 x87_arith_op(p, dst, src,
1223 0xd8, 0xc8,
1224 0xdc, 0xc8,
1225 4);
1226 }
1227
1228 void x87_fsub( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
1229 {
1230 DUMP_RR( dst, src );
1231 x87_arith_op(p, dst, src,
1232 0xd8, 0xe0,
1233 0xdc, 0xe8,
1234 4);
1235 }
1236
1237 void x87_fsubr( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
1238 {
1239 DUMP_RR( dst, src );
1240 x87_arith_op(p, dst, src,
1241 0xd8, 0xe8,
1242 0xdc, 0xe0,
1243 5);
1244 }
1245
1246 void x87_fadd( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
1247 {
1248 DUMP_RR( dst, src );
1249 x87_arith_op(p, dst, src,
1250 0xd8, 0xc0,
1251 0xdc, 0xc0,
1252 0);
1253 }
1254
1255 void x87_fdiv( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
1256 {
1257 DUMP_RR( dst, src );
1258 x87_arith_op(p, dst, src,
1259 0xd8, 0xf0,
1260 0xdc, 0xf8,
1261 6);
1262 }
1263
1264 void x87_fdivr( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
1265 {
1266 DUMP_RR( dst, src );
1267 x87_arith_op(p, dst, src,
1268 0xd8, 0xf8,
1269 0xdc, 0xf0,
1270 7);
1271 }
1272
1273 void x87_fmulp( struct x86_function *p, struct x86_reg dst )
1274 {
1275 DUMP_R( dst );
1276 assert(dst.file == file_x87);
1277 assert(dst.idx >= 1);
1278 emit_2ub(p, 0xde, 0xc8+dst.idx);
1279 note_x87_pop(p);
1280 }
1281
1282 void x87_fsubp( struct x86_function *p, struct x86_reg dst )
1283 {
1284 DUMP_R( dst );
1285 assert(dst.file == file_x87);
1286 assert(dst.idx >= 1);
1287 emit_2ub(p, 0xde, 0xe8+dst.idx);
1288 note_x87_pop(p);
1289 }
1290
1291 void x87_fsubrp( struct x86_function *p, struct x86_reg dst )
1292 {
1293 DUMP_R( dst );
1294 assert(dst.file == file_x87);
1295 assert(dst.idx >= 1);
1296 emit_2ub(p, 0xde, 0xe0+dst.idx);
1297 note_x87_pop(p);
1298 }
1299
1300 void x87_faddp( struct x86_function *p, struct x86_reg dst )
1301 {
1302 DUMP_R( dst );
1303 assert(dst.file == file_x87);
1304 assert(dst.idx >= 1);
1305 emit_2ub(p, 0xde, 0xc0+dst.idx);
1306 note_x87_pop(p);
1307 }
1308
1309 void x87_fdivp( struct x86_function *p, struct x86_reg dst )
1310 {
1311 DUMP_R( dst );
1312 assert(dst.file == file_x87);
1313 assert(dst.idx >= 1);
1314 emit_2ub(p, 0xde, 0xf8+dst.idx);
1315 note_x87_pop(p);
1316 }
1317
1318 void x87_fdivrp( struct x86_function *p, struct x86_reg dst )
1319 {
1320 DUMP_R( dst );
1321 assert(dst.file == file_x87);
1322 assert(dst.idx >= 1);
1323 emit_2ub(p, 0xde, 0xf0+dst.idx);
1324 note_x87_pop(p);
1325 }
1326
1327 void x87_ftst( struct x86_function *p )
1328 {
1329 DUMP();
1330 emit_2ub(p, 0xd9, 0xe4);
1331 }
1332
1333 void x87_fucom( struct x86_function *p, struct x86_reg arg )
1334 {
1335 DUMP_R( arg );
1336 assert(arg.file == file_x87);
1337 emit_2ub(p, 0xdd, 0xe0+arg.idx);
1338 }
1339
1340 void x87_fucomp( struct x86_function *p, struct x86_reg arg )
1341 {
1342 DUMP_R( arg );
1343 assert(arg.file == file_x87);
1344 emit_2ub(p, 0xdd, 0xe8+arg.idx);
1345 note_x87_pop(p);
1346 }
1347
1348 void x87_fucompp( struct x86_function *p )
1349 {
1350 DUMP();
1351 emit_2ub(p, 0xda, 0xe9);
1352 note_x87_pop(p); /* pop twice */
1353 note_x87_pop(p); /* pop twice */
1354 }
1355
1356 void x87_fxch( struct x86_function *p, struct x86_reg arg )
1357 {
1358 DUMP_R( arg );
1359 assert(arg.file == file_x87);
1360 emit_2ub(p, 0xd9, 0xc8+arg.idx);
1361 }
1362
1363 void x87_fabs( struct x86_function *p )
1364 {
1365 DUMP();
1366 emit_2ub(p, 0xd9, 0xe1);
1367 }
1368
1369 void x87_fchs( struct x86_function *p )
1370 {
1371 DUMP();
1372 emit_2ub(p, 0xd9, 0xe0);
1373 }
1374
1375 void x87_fcos( struct x86_function *p )
1376 {
1377 DUMP();
1378 emit_2ub(p, 0xd9, 0xff);
1379 }
1380
1381
1382 void x87_fprndint( struct x86_function *p )
1383 {
1384 DUMP();
1385 emit_2ub(p, 0xd9, 0xfc);
1386 }
1387
1388 void x87_fscale( struct x86_function *p )
1389 {
1390 DUMP();
1391 emit_2ub(p, 0xd9, 0xfd);
1392 }
1393
1394 void x87_fsin( struct x86_function *p )
1395 {
1396 DUMP();
1397 emit_2ub(p, 0xd9, 0xfe);
1398 }
1399
1400 void x87_fsincos( struct x86_function *p )
1401 {
1402 DUMP();
1403 emit_2ub(p, 0xd9, 0xfb);
1404 }
1405
1406 void x87_fsqrt( struct x86_function *p )
1407 {
1408 DUMP();
1409 emit_2ub(p, 0xd9, 0xfa);
1410 }
1411
1412 void x87_fxtract( struct x86_function *p )
1413 {
1414 DUMP();
1415 emit_2ub(p, 0xd9, 0xf4);
1416 }
1417
1418 /* st0 = (2^st0)-1
1419 *
1420 * Restrictions: -1.0 <= st0 <= 1.0
1421 */
1422 void x87_f2xm1( struct x86_function *p )
1423 {
1424 DUMP();
1425 emit_2ub(p, 0xd9, 0xf0);
1426 }
1427
1428 /* st1 = st1 * log2(st0);
1429 * pop_stack;
1430 */
1431 void x87_fyl2x( struct x86_function *p )
1432 {
1433 DUMP();
1434 emit_2ub(p, 0xd9, 0xf1);
1435 note_x87_pop(p);
1436 }
1437
1438 /* st1 = st1 * log2(st0 + 1.0);
1439 * pop_stack;
1440 *
1441 * A fast operation, with restrictions: -.29 < st0 < .29
1442 */
1443 void x87_fyl2xp1( struct x86_function *p )
1444 {
1445 DUMP();
1446 emit_2ub(p, 0xd9, 0xf9);
1447 note_x87_pop(p);
1448 }
1449
1450
1451 void x87_fld( struct x86_function *p, struct x86_reg arg )
1452 {
1453 DUMP_R( arg );
1454 if (arg.file == file_x87)
1455 emit_2ub(p, 0xd9, 0xc0 + arg.idx);
1456 else {
1457 emit_1ub(p, 0xd9);
1458 emit_modrm_noreg(p, 0, arg);
1459 }
1460 note_x87_push(p);
1461 }
1462
1463 void x87_fst( struct x86_function *p, struct x86_reg dst )
1464 {
1465 DUMP_R( dst );
1466 if (dst.file == file_x87)
1467 emit_2ub(p, 0xdd, 0xd0 + dst.idx);
1468 else {
1469 emit_1ub(p, 0xd9);
1470 emit_modrm_noreg(p, 2, dst);
1471 }
1472 }
1473
1474 void x87_fstp( struct x86_function *p, struct x86_reg dst )
1475 {
1476 DUMP_R( dst );
1477 if (dst.file == file_x87)
1478 emit_2ub(p, 0xdd, 0xd8 + dst.idx);
1479 else {
1480 emit_1ub(p, 0xd9);
1481 emit_modrm_noreg(p, 3, dst);
1482 }
1483 note_x87_pop(p);
1484 }
1485
1486 void x87_fpop( struct x86_function *p )
1487 {
1488 x87_fstp( p, x86_make_reg( file_x87, 0 ));
1489 }
1490
1491
1492 void x87_fcom( struct x86_function *p, struct x86_reg dst )
1493 {
1494 DUMP_R( dst );
1495 if (dst.file == file_x87)
1496 emit_2ub(p, 0xd8, 0xd0 + dst.idx);
1497 else {
1498 emit_1ub(p, 0xd8);
1499 emit_modrm_noreg(p, 2, dst);
1500 }
1501 }
1502
1503
1504 void x87_fcomp( struct x86_function *p, struct x86_reg dst )
1505 {
1506 DUMP_R( dst );
1507 if (dst.file == file_x87)
1508 emit_2ub(p, 0xd8, 0xd8 + dst.idx);
1509 else {
1510 emit_1ub(p, 0xd8);
1511 emit_modrm_noreg(p, 3, dst);
1512 }
1513 note_x87_pop(p);
1514 }
1515
1516 void x87_fcomi( struct x86_function *p, struct x86_reg arg )
1517 {
1518 DUMP_R( arg );
1519 emit_2ub(p, 0xdb, 0xf0+arg.idx);
1520 }
1521
1522 void x87_fcomip( struct x86_function *p, struct x86_reg arg )
1523 {
1524 DUMP_R( arg );
1525 emit_2ub(p, 0xdb, 0xf0+arg.idx);
1526 note_x87_pop(p);
1527 }
1528
1529
1530 void x87_fnstsw( struct x86_function *p, struct x86_reg dst )
1531 {
1532 DUMP_R( dst );
1533 assert(dst.file == file_REG32);
1534
1535 if (dst.idx == reg_AX &&
1536 dst.mod == mod_REG)
1537 emit_2ub(p, 0xdf, 0xe0);
1538 else {
1539 emit_1ub(p, 0xdd);
1540 emit_modrm_noreg(p, 7, dst);
1541 }
1542 }
1543
1544
1545 void x87_fnstcw( struct x86_function *p, struct x86_reg dst )
1546 {
1547 DUMP_R( dst );
1548 assert(dst.file == file_REG32);
1549
1550 emit_1ub(p, 0x9b); /* WAIT -- needed? */
1551 emit_1ub(p, 0xd9);
1552 emit_modrm_noreg(p, 7, dst);
1553 }
1554
1555
1556
1557
1558 /***********************************************************************
1559 * MMX instructions
1560 */
1561
1562 void mmx_emms( struct x86_function *p )
1563 {
1564 DUMP();
1565 assert(p->need_emms);
1566 emit_2ub(p, 0x0f, 0x77);
1567 p->need_emms = 0;
1568 }
1569
1570 void mmx_packssdw( struct x86_function *p,
1571 struct x86_reg dst,
1572 struct x86_reg src )
1573 {
1574 DUMP_RR( dst, src );
1575 assert(dst.file == file_MMX &&
1576 (src.file == file_MMX || src.mod != mod_REG));
1577
1578 p->need_emms = 1;
1579
1580 emit_2ub(p, X86_TWOB, 0x6b);
1581 emit_modrm( p, dst, src );
1582 }
1583
1584 void mmx_packuswb( struct x86_function *p,
1585 struct x86_reg dst,
1586 struct x86_reg src )
1587 {
1588 DUMP_RR( dst, src );
1589 assert(dst.file == file_MMX &&
1590 (src.file == file_MMX || src.mod != mod_REG));
1591
1592 p->need_emms = 1;
1593
1594 emit_2ub(p, X86_TWOB, 0x67);
1595 emit_modrm( p, dst, src );
1596 }
1597
1598 void mmx_movd( struct x86_function *p,
1599 struct x86_reg dst,
1600 struct x86_reg src )
1601 {
1602 DUMP_RR( dst, src );
1603 p->need_emms = 1;
1604 emit_1ub(p, X86_TWOB);
1605 emit_op_modrm( p, 0x6e, 0x7e, dst, src );
1606 }
1607
1608 void mmx_movq( struct x86_function *p,
1609 struct x86_reg dst,
1610 struct x86_reg src )
1611 {
1612 DUMP_RR( dst, src );
1613 p->need_emms = 1;
1614 emit_1ub(p, X86_TWOB);
1615 emit_op_modrm( p, 0x6f, 0x7f, dst, src );
1616 }
1617
1618
1619 /***********************************************************************
1620 * Helper functions
1621 */
1622
1623
1624 void x86_cdecl_caller_push_regs( struct x86_function *p )
1625 {
1626 x86_push(p, x86_make_reg(file_REG32, reg_AX));
1627 x86_push(p, x86_make_reg(file_REG32, reg_CX));
1628 x86_push(p, x86_make_reg(file_REG32, reg_DX));
1629 }
1630
1631 void x86_cdecl_caller_pop_regs( struct x86_function *p )
1632 {
1633 x86_pop(p, x86_make_reg(file_REG32, reg_DX));
1634 x86_pop(p, x86_make_reg(file_REG32, reg_CX));
1635 x86_pop(p, x86_make_reg(file_REG32, reg_AX));
1636 }
1637
1638
1639 /* Retreive a reference to one of the function arguments, taking into
1640 * account any push/pop activity:
1641 */
1642 struct x86_reg x86_fn_arg( struct x86_function *p,
1643 unsigned arg )
1644 {
1645 return x86_make_disp(x86_make_reg(file_REG32, reg_SP),
1646 p->stack_offset + arg * 4); /* ??? */
1647 }
1648
1649
1650 void x86_init_func( struct x86_function *p )
1651 {
1652 p->size = 0;
1653 p->store = NULL;
1654 p->csr = p->store;
1655 DUMP_START();
1656 }
1657
1658 void x86_init_func_size( struct x86_function *p, unsigned code_size )
1659 {
1660 p->size = code_size;
1661 p->store = rtasm_exec_malloc(code_size);
1662 if (p->store == NULL) {
1663 p->store = p->error_overflow;
1664 }
1665 p->csr = p->store;
1666 DUMP_START();
1667 }
1668
1669 void x86_release_func( struct x86_function *p )
1670 {
1671 if (p->store && p->store != p->error_overflow)
1672 rtasm_exec_free(p->store);
1673
1674 p->store = NULL;
1675 p->csr = NULL;
1676 p->size = 0;
1677 }
1678
1679
1680 void (*x86_get_func( struct x86_function *p ))(void)
1681 {
1682 DUMP_END();
1683 if (DISASSEM && p->store)
1684 debug_printf("disassemble %p %p\n", p->store, p->csr);
1685
1686 if (p->store == p->error_overflow)
1687 return (void (*)(void)) NULL;
1688 else
1689 return (void (*)(void)) p->store;
1690 }
1691
1692 #else
1693
1694 void x86sse_dummy( void )
1695 {
1696 }
1697
1698 #endif