projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
Merge branch 'mesa_7_5_branch'
[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 "util/u_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 regmem.mod != mod_REG) {
245 emit_1ub(p, 0x24); /* simplistic! */
246 }
247
248 switch (regmem.mod) {
249 case mod_REG:
250 case mod_INDIRECT:
251 break;
252 case mod_DISP8:
253 emit_1b(p, (char) regmem.disp);
254 break;
255 case mod_DISP32:
256 emit_1i(p, regmem.disp);
257 break;
258 default:
259 assert(0);
260 break;
261 }
262 }
263
264 /* Emits the "/0".."/7" specialized versions of the modrm ("/r") bytes.
265 */
266 static void emit_modrm_noreg( struct x86_function *p,
267 unsigned op,
268 struct x86_reg regmem )
269 {
270 struct x86_reg dummy = x86_make_reg(file_REG32, op);
271 emit_modrm(p, dummy, regmem);
272 }
273
274 /* Many x86 instructions have two opcodes to cope with the situations
275 * where the destination is a register or memory reference
276 * respectively. This function selects the correct opcode based on
277 * the arguments presented.
278 */
279 static void emit_op_modrm( struct x86_function *p,
280 unsigned char op_dst_is_reg,
281 unsigned char op_dst_is_mem,
282 struct x86_reg dst,
283 struct x86_reg src )
284 {
285 switch (dst.mod) {
286 case mod_REG:
287 emit_1ub(p, op_dst_is_reg);
288 emit_modrm(p, dst, src);
289 break;
290 case mod_INDIRECT:
291 case mod_DISP32:
292 case mod_DISP8:
293 assert(src.mod == mod_REG);
294 emit_1ub(p, op_dst_is_mem);
295 emit_modrm(p, src, dst);
296 break;
297 default:
298 assert(0);
299 break;
300 }
301 }
302
303
304
305
306
307
308
309 /* Create and manipulate registers and regmem values:
310 */
311 struct x86_reg x86_make_reg( enum x86_reg_file file,
312 enum x86_reg_name idx )
313 {
314 struct x86_reg reg;
315
316 reg.file = file;
317 reg.idx = idx;
318 reg.mod = mod_REG;
319 reg.disp = 0;
320
321 return reg;
322 }
323
324 struct x86_reg x86_make_disp( struct x86_reg reg,
325 int disp )
326 {
327 assert(reg.file == file_REG32);
328
329 if (reg.mod == mod_REG)
330 reg.disp = disp;
331 else
332 reg.disp += disp;
333
334 if (reg.disp == 0 && reg.idx != reg_BP)
335 reg.mod = mod_INDIRECT;
336 else if (reg.disp <= 127 && reg.disp >= -128)
337 reg.mod = mod_DISP8;
338 else
339 reg.mod = mod_DISP32;
340
341 return reg;
342 }
343
344 struct x86_reg x86_deref( struct x86_reg reg )
345 {
346 return x86_make_disp(reg, 0);
347 }
348
349 struct x86_reg x86_get_base_reg( struct x86_reg reg )
350 {
351 return x86_make_reg( reg.file, reg.idx );
352 }
353
354 int x86_get_label( struct x86_function *p )
355 {
356 return p->csr - p->store;
357 }
358
359
360
361 /***********************************************************************
362 * x86 instructions
363 */
364
365
366 void x86_jcc( struct x86_function *p,
367 enum x86_cc cc,
368 int label )
369 {
370 int offset = label - (x86_get_label(p) + 2);
371 DUMP_I(cc);
372
373 if (offset < 0) {
374 /*assert(p->csr - p->store > -offset);*/
375 if (p->csr - p->store <= -offset) {
376 /* probably out of memory (using the error_overflow buffer) */
377 return;
378 }
379 }
380
381 if (offset <= 127 && offset >= -128) {
382 emit_1ub(p, 0x70 + cc);
383 emit_1b(p, (char) offset);
384 }
385 else {
386 offset = label - (x86_get_label(p) + 6);
387 emit_2ub(p, 0x0f, 0x80 + cc);
388 emit_1i(p, offset);
389 }
390 }
391
392 /* Always use a 32bit offset for forward jumps:
393 */
394 int x86_jcc_forward( struct x86_function *p,
395 enum x86_cc cc )
396 {
397 DUMP_I(cc);
398 emit_2ub(p, 0x0f, 0x80 + cc);
399 emit_1i(p, 0);
400 return x86_get_label(p);
401 }
402
403 int x86_jmp_forward( struct x86_function *p)
404 {
405 DUMP();
406 emit_1ub(p, 0xe9);
407 emit_1i(p, 0);
408 return x86_get_label(p);
409 }
410
411 int x86_call_forward( struct x86_function *p)
412 {
413 DUMP();
414
415 emit_1ub(p, 0xe8);
416 emit_1i(p, 0);
417 return x86_get_label(p);
418 }
419
420 /* Fixup offset from forward jump:
421 */
422 void x86_fixup_fwd_jump( struct x86_function *p,
423 int fixup )
424 {
425 *(int *)(p->store + fixup - 4) = x86_get_label(p) - fixup;
426 }
427
428 void x86_jmp( struct x86_function *p, int label)
429 {
430 DUMP_I( label );
431 emit_1ub(p, 0xe9);
432 emit_1i(p, label - x86_get_label(p) - 4);
433 }
434
435 void x86_call( struct x86_function *p, struct x86_reg reg)
436 {
437 DUMP_R( reg );
438 emit_1ub(p, 0xff);
439 emit_modrm_noreg(p, 2, reg);
440 }
441
442
443 void x86_mov_reg_imm( struct x86_function *p, struct x86_reg dst, int imm )
444 {
445 DUMP_RI( dst, imm );
446 assert(dst.file == file_REG32);
447 assert(dst.mod == mod_REG);
448 emit_1ub(p, 0xb8 + dst.idx);
449 emit_1i(p, imm);
450 }
451
452 /**
453 * Immediate group 1 instructions.
454 */
455 static INLINE void
456 x86_group1_imm( struct x86_function *p,
457 unsigned op, struct x86_reg dst, int imm )
458 {
459 assert(dst.file == file_REG32);
460 assert(dst.mod == mod_REG);
461 if(-0x80 <= imm && imm < 0x80) {
462 emit_1ub(p, 0x83);
463 emit_modrm_noreg(p, op, dst);
464 emit_1b(p, (char)imm);
465 }
466 else {
467 emit_1ub(p, 0x81);
468 emit_modrm_noreg(p, op, dst);
469 emit_1i(p, imm);
470 }
471 }
472
473 void x86_add_imm( struct x86_function *p, struct x86_reg dst, int imm )
474 {
475 DUMP_RI( dst, imm );
476 x86_group1_imm(p, 0, dst, imm);
477 }
478
479 void x86_or_imm( struct x86_function *p, struct x86_reg dst, int imm )
480 {
481 DUMP_RI( dst, imm );
482 x86_group1_imm(p, 1, dst, imm);
483 }
484
485 void x86_and_imm( struct x86_function *p, struct x86_reg dst, int imm )
486 {
487 DUMP_RI( dst, imm );
488 x86_group1_imm(p, 4, dst, imm);
489 }
490
491 void x86_sub_imm( struct x86_function *p, struct x86_reg dst, int imm )
492 {
493 DUMP_RI( dst, imm );
494 x86_group1_imm(p, 5, dst, imm);
495 }
496
497 void x86_xor_imm( struct x86_function *p, struct x86_reg dst, int imm )
498 {
499 DUMP_RI( dst, imm );
500 x86_group1_imm(p, 6, dst, imm);
501 }
502
503 void x86_cmp_imm( struct x86_function *p, struct x86_reg dst, int imm )
504 {
505 DUMP_RI( dst, imm );
506 x86_group1_imm(p, 7, dst, imm);
507 }
508
509
510 void x86_push( struct x86_function *p,
511 struct x86_reg reg )
512 {
513 DUMP_R( reg );
514 if (reg.mod == mod_REG)
515 emit_1ub(p, 0x50 + reg.idx);
516 else
517 {
518 emit_1ub(p, 0xff);
519 emit_modrm_noreg(p, 6, reg);
520 }
521
522
523 p->stack_offset += 4;
524 }
525
526 void x86_push_imm32( struct x86_function *p,
527 int imm32 )
528 {
529 DUMP_I( imm32 );
530 emit_1ub(p, 0x68);
531 emit_1i(p, imm32);
532
533 p->stack_offset += 4;
534 }
535
536
537 void x86_pop( struct x86_function *p,
538 struct x86_reg reg )
539 {
540 DUMP_R( reg );
541 assert(reg.mod == mod_REG);
542 emit_1ub(p, 0x58 + reg.idx);
543 p->stack_offset -= 4;
544 }
545
546 void x86_inc( struct x86_function *p,
547 struct x86_reg reg )
548 {
549 DUMP_R( reg );
550 assert(reg.mod == mod_REG);
551 emit_1ub(p, 0x40 + reg.idx);
552 }
553
554 void x86_dec( struct x86_function *p,
555 struct x86_reg reg )
556 {
557 DUMP_R( reg );
558 assert(reg.mod == mod_REG);
559 emit_1ub(p, 0x48 + reg.idx);
560 }
561
562 void x86_ret( struct x86_function *p )
563 {
564 DUMP();
565 assert(p->stack_offset == 0);
566 emit_1ub(p, 0xc3);
567 }
568
569 void x86_retw( struct x86_function *p, unsigned short imm )
570 {
571 DUMP();
572 emit_3ub(p, 0xc2, imm & 0xff, (imm >> 8) & 0xff);
573 }
574
575 void x86_sahf( struct x86_function *p )
576 {
577 DUMP();
578 emit_1ub(p, 0x9e);
579 }
580
581 void x86_mov( struct x86_function *p,
582 struct x86_reg dst,
583 struct x86_reg src )
584 {
585 DUMP_RR( dst, src );
586 emit_op_modrm( p, 0x8b, 0x89, dst, src );
587 }
588
589 void x86_xor( struct x86_function *p,
590 struct x86_reg dst,
591 struct x86_reg src )
592 {
593 DUMP_RR( dst, src );
594 emit_op_modrm( p, 0x33, 0x31, dst, src );
595 }
596
597 void x86_cmp( struct x86_function *p,
598 struct x86_reg dst,
599 struct x86_reg src )
600 {
601 DUMP_RR( dst, src );
602 emit_op_modrm( p, 0x3b, 0x39, dst, src );
603 }
604
605 void x86_lea( struct x86_function *p,
606 struct x86_reg dst,
607 struct x86_reg src )
608 {
609 DUMP_RR( dst, src );
610 emit_1ub(p, 0x8d);
611 emit_modrm( p, dst, src );
612 }
613
614 void x86_test( struct x86_function *p,
615 struct x86_reg dst,
616 struct x86_reg src )
617 {
618 DUMP_RR( dst, src );
619 emit_1ub(p, 0x85);
620 emit_modrm( p, dst, src );
621 }
622
623 void x86_add( struct x86_function *p,
624 struct x86_reg dst,
625 struct x86_reg src )
626 {
627 DUMP_RR( dst, src );
628 emit_op_modrm(p, 0x03, 0x01, dst, src );
629 }
630
631 /* Calculate EAX * src, results in EDX:EAX.
632 */
633 void x86_mul( struct x86_function *p,
634 struct x86_reg src )
635 {
636 DUMP_R( src );
637 emit_1ub(p, 0xf7);
638 emit_modrm_noreg(p, 4, src );
639 }
640
641
642 void x86_imul( struct x86_function *p,
643 struct x86_reg dst,
644 struct x86_reg src )
645 {
646 DUMP_RR( dst, src );
647 emit_2ub(p, X86_TWOB, 0xAF);
648 emit_modrm(p, dst, src);
649 }
650
651
652 void x86_sub( struct x86_function *p,
653 struct x86_reg dst,
654 struct x86_reg src )
655 {
656 DUMP_RR( dst, src );
657 emit_op_modrm(p, 0x2b, 0x29, dst, src );
658 }
659
660 void x86_or( struct x86_function *p,
661 struct x86_reg dst,
662 struct x86_reg src )
663 {
664 DUMP_RR( dst, src );
665 emit_op_modrm( p, 0x0b, 0x09, dst, src );
666 }
667
668 void x86_and( struct x86_function *p,
669 struct x86_reg dst,
670 struct x86_reg src )
671 {
672 DUMP_RR( dst, src );
673 emit_op_modrm( p, 0x23, 0x21, dst, src );
674 }
675
676
677
678 /***********************************************************************
679 * SSE instructions
680 */
681
682 void sse_prefetchnta( struct x86_function *p, struct x86_reg ptr)
683 {
684 DUMP_R( ptr );
685 assert(ptr.mod != mod_REG);
686 emit_2ub(p, 0x0f, 0x18);
687 emit_modrm_noreg(p, 0, ptr);
688 }
689
690 void sse_prefetch0( struct x86_function *p, struct x86_reg ptr)
691 {
692 DUMP_R( ptr );
693 assert(ptr.mod != mod_REG);
694 emit_2ub(p, 0x0f, 0x18);
695 emit_modrm_noreg(p, 1, ptr);
696 }
697
698 void sse_prefetch1( struct x86_function *p, struct x86_reg ptr)
699 {
700 DUMP_R( ptr );
701 assert(ptr.mod != mod_REG);
702 emit_2ub(p, 0x0f, 0x18);
703 emit_modrm_noreg(p, 2, ptr);
704 }
705
706 void sse_movntps( struct x86_function *p,
707 struct x86_reg dst,
708 struct x86_reg src)
709 {
710 DUMP_RR( dst, src );
711
712 assert(dst.mod != mod_REG);
713 assert(src.mod == mod_REG);
714 emit_2ub(p, 0x0f, 0x2b);
715 emit_modrm(p, src, dst);
716 }
717
718
719
720
721 void sse_movss( struct x86_function *p,
722 struct x86_reg dst,
723 struct x86_reg src )
724 {
725 DUMP_RR( dst, src );
726 emit_2ub(p, 0xF3, X86_TWOB);
727 emit_op_modrm( p, 0x10, 0x11, dst, src );
728 }
729
730 void sse_movaps( struct x86_function *p,
731 struct x86_reg dst,
732 struct x86_reg src )
733 {
734 DUMP_RR( dst, src );
735 emit_1ub(p, X86_TWOB);
736 emit_op_modrm( p, 0x28, 0x29, dst, src );
737 }
738
739 void sse_movups( struct x86_function *p,
740 struct x86_reg dst,
741 struct x86_reg src )
742 {
743 DUMP_RR( dst, src );
744 emit_1ub(p, X86_TWOB);
745 emit_op_modrm( p, 0x10, 0x11, dst, src );
746 }
747
748 void sse_movhps( struct x86_function *p,
749 struct x86_reg dst,
750 struct x86_reg src )
751 {
752 DUMP_RR( dst, src );
753 assert(dst.mod != mod_REG || src.mod != mod_REG);
754 emit_1ub(p, X86_TWOB);
755 emit_op_modrm( p, 0x16, 0x17, dst, src ); /* cf movlhps */
756 }
757
758 void sse_movlps( struct x86_function *p,
759 struct x86_reg dst,
760 struct x86_reg src )
761 {
762 DUMP_RR( dst, src );
763 assert(dst.mod != mod_REG || src.mod != mod_REG);
764 emit_1ub(p, X86_TWOB);
765 emit_op_modrm( p, 0x12, 0x13, dst, src ); /* cf movhlps */
766 }
767
768 void sse_maxps( struct x86_function *p,
769 struct x86_reg dst,
770 struct x86_reg src )
771 {
772 DUMP_RR( dst, src );
773 emit_2ub(p, X86_TWOB, 0x5F);
774 emit_modrm( p, dst, src );
775 }
776
777 void sse_maxss( struct x86_function *p,
778 struct x86_reg dst,
779 struct x86_reg src )
780 {
781 DUMP_RR( dst, src );
782 emit_3ub(p, 0xF3, X86_TWOB, 0x5F);
783 emit_modrm( p, dst, src );
784 }
785
786 void sse_divss( struct x86_function *p,
787 struct x86_reg dst,
788 struct x86_reg src )
789 {
790 DUMP_RR( dst, src );
791 emit_3ub(p, 0xF3, X86_TWOB, 0x5E);
792 emit_modrm( p, dst, src );
793 }
794
795 void sse_minps( struct x86_function *p,
796 struct x86_reg dst,
797 struct x86_reg src )
798 {
799 DUMP_RR( dst, src );
800 emit_2ub(p, X86_TWOB, 0x5D);
801 emit_modrm( p, dst, src );
802 }
803
804 void sse_subps( struct x86_function *p,
805 struct x86_reg dst,
806 struct x86_reg src )
807 {
808 DUMP_RR( dst, src );
809 emit_2ub(p, X86_TWOB, 0x5C);
810 emit_modrm( p, dst, src );
811 }
812
813 void sse_mulps( struct x86_function *p,
814 struct x86_reg dst,
815 struct x86_reg src )
816 {
817 DUMP_RR( dst, src );
818 emit_2ub(p, X86_TWOB, 0x59);
819 emit_modrm( p, dst, src );
820 }
821
822 void sse_mulss( struct x86_function *p,
823 struct x86_reg dst,
824 struct x86_reg src )
825 {
826 DUMP_RR( dst, src );
827 emit_3ub(p, 0xF3, X86_TWOB, 0x59);
828 emit_modrm( p, dst, src );
829 }
830
831 void sse_addps( struct x86_function *p,
832 struct x86_reg dst,
833 struct x86_reg src )
834 {
835 DUMP_RR( dst, src );
836 emit_2ub(p, X86_TWOB, 0x58);
837 emit_modrm( p, dst, src );
838 }
839
840 void sse_addss( struct x86_function *p,
841 struct x86_reg dst,
842 struct x86_reg src )
843 {
844 DUMP_RR( dst, src );
845 emit_3ub(p, 0xF3, X86_TWOB, 0x58);
846 emit_modrm( p, dst, src );
847 }
848
849 void sse_andnps( struct x86_function *p,
850 struct x86_reg dst,
851 struct x86_reg src )
852 {
853 DUMP_RR( dst, src );
854 emit_2ub(p, X86_TWOB, 0x55);
855 emit_modrm( p, dst, src );
856 }
857
858 void sse_andps( struct x86_function *p,
859 struct x86_reg dst,
860 struct x86_reg src )
861 {
862 DUMP_RR( dst, src );
863 emit_2ub(p, X86_TWOB, 0x54);
864 emit_modrm( p, dst, src );
865 }
866
867 void sse_rsqrtps( struct x86_function *p,
868 struct x86_reg dst,
869 struct x86_reg src )
870 {
871 DUMP_RR( dst, src );
872 emit_2ub(p, X86_TWOB, 0x52);
873 emit_modrm( p, dst, src );
874 }
875
876 void sse_rsqrtss( struct x86_function *p,
877 struct x86_reg dst,
878 struct x86_reg src )
879 {
880 DUMP_RR( dst, src );
881 emit_3ub(p, 0xF3, X86_TWOB, 0x52);
882 emit_modrm( p, dst, src );
883
884 }
885
886 void sse_movhlps( struct x86_function *p,
887 struct x86_reg dst,
888 struct x86_reg src )
889 {
890 DUMP_RR( dst, src );
891 assert(dst.mod == mod_REG && src.mod == mod_REG);
892 emit_2ub(p, X86_TWOB, 0x12);
893 emit_modrm( p, dst, src );
894 }
895
896 void sse_movlhps( struct x86_function *p,
897 struct x86_reg dst,
898 struct x86_reg src )
899 {
900 DUMP_RR( dst, src );
901 assert(dst.mod == mod_REG && src.mod == mod_REG);
902 emit_2ub(p, X86_TWOB, 0x16);
903 emit_modrm( p, dst, src );
904 }
905
906 void sse_orps( struct x86_function *p,
907 struct x86_reg dst,
908 struct x86_reg src )
909 {
910 DUMP_RR( dst, src );
911 emit_2ub(p, X86_TWOB, 0x56);
912 emit_modrm( p, dst, src );
913 }
914
915 void sse_xorps( struct x86_function *p,
916 struct x86_reg dst,
917 struct x86_reg src )
918 {
919 DUMP_RR( dst, src );
920 emit_2ub(p, X86_TWOB, 0x57);
921 emit_modrm( p, dst, src );
922 }
923
924 void sse_cvtps2pi( struct x86_function *p,
925 struct x86_reg dst,
926 struct x86_reg src )
927 {
928 DUMP_RR( dst, src );
929 assert(dst.file == file_MMX &&
930 (src.file == file_XMM || src.mod != mod_REG));
931
932 p->need_emms = 1;
933
934 emit_2ub(p, X86_TWOB, 0x2d);
935 emit_modrm( p, dst, src );
936 }
937
938 void sse2_cvtdq2ps( struct x86_function *p,
939 struct x86_reg dst,
940 struct x86_reg src )
941 {
942 DUMP_RR( dst, src );
943 emit_2ub(p, X86_TWOB, 0x5b);
944 emit_modrm( p, dst, src );
945 }
946
947
948 /* Shufps can also be used to implement a reduced swizzle when dest ==
949 * arg0.
950 */
951 void sse_shufps( struct x86_function *p,
952 struct x86_reg dst,
953 struct x86_reg src,
954 unsigned char shuf)
955 {
956 DUMP_RRI( dst, src, shuf );
957 emit_2ub(p, X86_TWOB, 0xC6);
958 emit_modrm(p, dst, src);
959 emit_1ub(p, shuf);
960 }
961
962 void sse_unpckhps( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
963 {
964 DUMP_RR( dst, src );
965 emit_2ub( p, X86_TWOB, 0x15 );
966 emit_modrm( p, dst, src );
967 }
968
969 void sse_unpcklps( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
970 {
971 DUMP_RR( dst, src );
972 emit_2ub( p, X86_TWOB, 0x14 );
973 emit_modrm( p, dst, src );
974 }
975
976 void sse_cmpps( struct x86_function *p,
977 struct x86_reg dst,
978 struct x86_reg src,
979 enum sse_cc cc)
980 {
981 DUMP_RRI( dst, src, cc );
982 emit_2ub(p, X86_TWOB, 0xC2);
983 emit_modrm(p, dst, src);
984 emit_1ub(p, cc);
985 }
986
987 void sse_pmovmskb( struct x86_function *p,
988 struct x86_reg dst,
989 struct x86_reg src)
990 {
991 DUMP_RR( dst, src );
992 emit_3ub(p, 0x66, X86_TWOB, 0xD7);
993 emit_modrm(p, dst, src);
994 }
995
996 void sse_movmskps( struct x86_function *p,
997 struct x86_reg dst,
998 struct x86_reg src)
999 {
1000 DUMP_RR( dst, src );
1001 emit_2ub(p, X86_TWOB, 0x50);
1002 emit_modrm(p, dst, src);
1003 }
1004
1005 /***********************************************************************
1006 * SSE2 instructions
1007 */
1008
1009 /**
1010 * Perform a reduced swizzle:
1011 */
1012 void sse2_pshufd( struct x86_function *p,
1013 struct x86_reg dst,
1014 struct x86_reg src,
1015 unsigned char shuf)
1016 {
1017 DUMP_RRI( dst, src, shuf );
1018 emit_3ub(p, 0x66, X86_TWOB, 0x70);
1019 emit_modrm(p, dst, src);
1020 emit_1ub(p, shuf);
1021 }
1022
1023 void sse2_cvttps2dq( struct x86_function *p,
1024 struct x86_reg dst,
1025 struct x86_reg src )
1026 {
1027 DUMP_RR( dst, src );
1028 emit_3ub( p, 0xF3, X86_TWOB, 0x5B );
1029 emit_modrm( p, dst, src );
1030 }
1031
1032 void sse2_cvtps2dq( struct x86_function *p,
1033 struct x86_reg dst,
1034 struct x86_reg src )
1035 {
1036 DUMP_RR( dst, src );
1037 emit_3ub(p, 0x66, X86_TWOB, 0x5B);
1038 emit_modrm( p, dst, src );
1039 }
1040
1041 void sse2_packssdw( struct x86_function *p,
1042 struct x86_reg dst,
1043 struct x86_reg src )
1044 {
1045 DUMP_RR( dst, src );
1046 emit_3ub(p, 0x66, X86_TWOB, 0x6B);
1047 emit_modrm( p, dst, src );
1048 }
1049
1050 void sse2_packsswb( struct x86_function *p,
1051 struct x86_reg dst,
1052 struct x86_reg src )
1053 {
1054 DUMP_RR( dst, src );
1055 emit_3ub(p, 0x66, X86_TWOB, 0x63);
1056 emit_modrm( p, dst, src );
1057 }
1058
1059 void sse2_packuswb( struct x86_function *p,
1060 struct x86_reg dst,
1061 struct x86_reg src )
1062 {
1063 DUMP_RR( dst, src );
1064 emit_3ub(p, 0x66, X86_TWOB, 0x67);
1065 emit_modrm( p, dst, src );
1066 }
1067
1068 void sse2_punpcklbw( struct x86_function *p,
1069 struct x86_reg dst,
1070 struct x86_reg src )
1071 {
1072 DUMP_RR( dst, src );
1073 emit_3ub(p, 0x66, X86_TWOB, 0x60);
1074 emit_modrm( p, dst, src );
1075 }
1076
1077
1078 void sse2_rcpps( struct x86_function *p,
1079 struct x86_reg dst,
1080 struct x86_reg src )
1081 {
1082 DUMP_RR( dst, src );
1083 emit_2ub(p, X86_TWOB, 0x53);
1084 emit_modrm( p, dst, src );
1085 }
1086
1087 void sse2_rcpss( struct x86_function *p,
1088 struct x86_reg dst,
1089 struct x86_reg src )
1090 {
1091 DUMP_RR( dst, src );
1092 emit_3ub(p, 0xF3, X86_TWOB, 0x53);
1093 emit_modrm( p, dst, src );
1094 }
1095
1096 void sse2_movd( struct x86_function *p,
1097 struct x86_reg dst,
1098 struct x86_reg src )
1099 {
1100 DUMP_RR( dst, src );
1101 emit_2ub(p, 0x66, X86_TWOB);
1102 emit_op_modrm( p, 0x6e, 0x7e, dst, src );
1103 }
1104
1105
1106
1107
1108 /***********************************************************************
1109 * x87 instructions
1110 */
1111 static void note_x87_pop( struct x86_function *p )
1112 {
1113 p->x87_stack--;
1114 assert(p->x87_stack >= 0);
1115 }
1116
1117 static void note_x87_push( struct x86_function *p )
1118 {
1119 p->x87_stack++;
1120 assert(p->x87_stack <= 7);
1121 }
1122
1123 void x87_assert_stack_empty( struct x86_function *p )
1124 {
1125 assert (p->x87_stack == 0);
1126 }
1127
1128
1129 void x87_fist( struct x86_function *p, struct x86_reg dst )
1130 {
1131 DUMP_R( dst );
1132 emit_1ub(p, 0xdb);
1133 emit_modrm_noreg(p, 2, dst);
1134 }
1135
1136 void x87_fistp( struct x86_function *p, struct x86_reg dst )
1137 {
1138 DUMP_R( dst );
1139 emit_1ub(p, 0xdb);
1140 emit_modrm_noreg(p, 3, dst);
1141 note_x87_pop(p);
1142 }
1143
1144 void x87_fild( struct x86_function *p, struct x86_reg arg )
1145 {
1146 DUMP_R( arg );
1147 emit_1ub(p, 0xdf);
1148 emit_modrm_noreg(p, 0, arg);
1149 note_x87_push(p);
1150 }
1151
1152 void x87_fldz( struct x86_function *p )
1153 {
1154 DUMP();
1155 emit_2ub(p, 0xd9, 0xee);
1156 note_x87_push(p);
1157 }
1158
1159
1160 void x87_fldcw( struct x86_function *p, struct x86_reg arg )
1161 {
1162 DUMP_R( arg );
1163 assert(arg.file == file_REG32);
1164 assert(arg.mod != mod_REG);
1165 emit_1ub(p, 0xd9);
1166 emit_modrm_noreg(p, 5, arg);
1167 }
1168
1169 void x87_fld1( struct x86_function *p )
1170 {
1171 DUMP();
1172 emit_2ub(p, 0xd9, 0xe8);
1173 note_x87_push(p);
1174 }
1175
1176 void x87_fldl2e( struct x86_function *p )
1177 {
1178 DUMP();
1179 emit_2ub(p, 0xd9, 0xea);
1180 note_x87_push(p);
1181 }
1182
1183 void x87_fldln2( struct x86_function *p )
1184 {
1185 DUMP();
1186 emit_2ub(p, 0xd9, 0xed);
1187 note_x87_push(p);
1188 }
1189
1190 void x87_fwait( struct x86_function *p )
1191 {
1192 DUMP();
1193 emit_1ub(p, 0x9b);
1194 }
1195
1196 void x87_fnclex( struct x86_function *p )
1197 {
1198 DUMP();
1199 emit_2ub(p, 0xdb, 0xe2);
1200 }
1201
1202 void x87_fclex( struct x86_function *p )
1203 {
1204 x87_fwait(p);
1205 x87_fnclex(p);
1206 }
1207
1208 void x87_fcmovb( struct x86_function *p, struct x86_reg arg )
1209 {
1210 DUMP_R( arg );
1211 assert(arg.file == file_x87);
1212 emit_2ub(p, 0xda, 0xc0+arg.idx);
1213 }
1214
1215 void x87_fcmove( struct x86_function *p, struct x86_reg arg )
1216 {
1217 DUMP_R( arg );
1218 assert(arg.file == file_x87);
1219 emit_2ub(p, 0xda, 0xc8+arg.idx);
1220 }
1221
1222 void x87_fcmovbe( struct x86_function *p, struct x86_reg arg )
1223 {
1224 DUMP_R( arg );
1225 assert(arg.file == file_x87);
1226 emit_2ub(p, 0xda, 0xd0+arg.idx);
1227 }
1228
1229 void x87_fcmovnb( struct x86_function *p, struct x86_reg arg )
1230 {
1231 DUMP_R( arg );
1232 assert(arg.file == file_x87);
1233 emit_2ub(p, 0xdb, 0xc0+arg.idx);
1234 }
1235
1236 void x87_fcmovne( struct x86_function *p, struct x86_reg arg )
1237 {
1238 DUMP_R( arg );
1239 assert(arg.file == file_x87);
1240 emit_2ub(p, 0xdb, 0xc8+arg.idx);
1241 }
1242
1243 void x87_fcmovnbe( struct x86_function *p, struct x86_reg arg )
1244 {
1245 DUMP_R( arg );
1246 assert(arg.file == file_x87);
1247 emit_2ub(p, 0xdb, 0xd0+arg.idx);
1248 }
1249
1250
1251
1252 static void x87_arith_op( struct x86_function *p, struct x86_reg dst, struct x86_reg arg,
1253 unsigned char dst0ub0,
1254 unsigned char dst0ub1,
1255 unsigned char arg0ub0,
1256 unsigned char arg0ub1,
1257 unsigned char argmem_noreg)
1258 {
1259 assert(dst.file == file_x87);
1260
1261 if (arg.file == file_x87) {
1262 if (dst.idx == 0)
1263 emit_2ub(p, dst0ub0, dst0ub1+arg.idx);
1264 else if (arg.idx == 0)
1265 emit_2ub(p, arg0ub0, arg0ub1+arg.idx);
1266 else
1267 assert(0);
1268 }
1269 else if (dst.idx == 0) {
1270 assert(arg.file == file_REG32);
1271 emit_1ub(p, 0xd8);
1272 emit_modrm_noreg(p, argmem_noreg, arg);
1273 }
1274 else
1275 assert(0);
1276 }
1277
1278 void x87_fmul( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
1279 {
1280 DUMP_RR( dst, src );
1281 x87_arith_op(p, dst, src,
1282 0xd8, 0xc8,
1283 0xdc, 0xc8,
1284 4);
1285 }
1286
1287 void x87_fsub( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
1288 {
1289 DUMP_RR( dst, src );
1290 x87_arith_op(p, dst, src,
1291 0xd8, 0xe0,
1292 0xdc, 0xe8,
1293 4);
1294 }
1295
1296 void x87_fsubr( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
1297 {
1298 DUMP_RR( dst, src );
1299 x87_arith_op(p, dst, src,
1300 0xd8, 0xe8,
1301 0xdc, 0xe0,
1302 5);
1303 }
1304
1305 void x87_fadd( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
1306 {
1307 DUMP_RR( dst, src );
1308 x87_arith_op(p, dst, src,
1309 0xd8, 0xc0,
1310 0xdc, 0xc0,
1311 0);
1312 }
1313
1314 void x87_fdiv( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
1315 {
1316 DUMP_RR( dst, src );
1317 x87_arith_op(p, dst, src,
1318 0xd8, 0xf0,
1319 0xdc, 0xf8,
1320 6);
1321 }
1322
1323 void x87_fdivr( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
1324 {
1325 DUMP_RR( dst, src );
1326 x87_arith_op(p, dst, src,
1327 0xd8, 0xf8,
1328 0xdc, 0xf0,
1329 7);
1330 }
1331
1332 void x87_fmulp( struct x86_function *p, struct x86_reg dst )
1333 {
1334 DUMP_R( dst );
1335 assert(dst.file == file_x87);
1336 assert(dst.idx >= 1);
1337 emit_2ub(p, 0xde, 0xc8+dst.idx);
1338 note_x87_pop(p);
1339 }
1340
1341 void x87_fsubp( struct x86_function *p, struct x86_reg dst )
1342 {
1343 DUMP_R( dst );
1344 assert(dst.file == file_x87);
1345 assert(dst.idx >= 1);
1346 emit_2ub(p, 0xde, 0xe8+dst.idx);
1347 note_x87_pop(p);
1348 }
1349
1350 void x87_fsubrp( struct x86_function *p, struct x86_reg dst )
1351 {
1352 DUMP_R( dst );
1353 assert(dst.file == file_x87);
1354 assert(dst.idx >= 1);
1355 emit_2ub(p, 0xde, 0xe0+dst.idx);
1356 note_x87_pop(p);
1357 }
1358
1359 void x87_faddp( struct x86_function *p, struct x86_reg dst )
1360 {
1361 DUMP_R( dst );
1362 assert(dst.file == file_x87);
1363 assert(dst.idx >= 1);
1364 emit_2ub(p, 0xde, 0xc0+dst.idx);
1365 note_x87_pop(p);
1366 }
1367
1368 void x87_fdivp( struct x86_function *p, struct x86_reg dst )
1369 {
1370 DUMP_R( dst );
1371 assert(dst.file == file_x87);
1372 assert(dst.idx >= 1);
1373 emit_2ub(p, 0xde, 0xf8+dst.idx);
1374 note_x87_pop(p);
1375 }
1376
1377 void x87_fdivrp( struct x86_function *p, struct x86_reg dst )
1378 {
1379 DUMP_R( dst );
1380 assert(dst.file == file_x87);
1381 assert(dst.idx >= 1);
1382 emit_2ub(p, 0xde, 0xf0+dst.idx);
1383 note_x87_pop(p);
1384 }
1385
1386 void x87_ftst( struct x86_function *p )
1387 {
1388 DUMP();
1389 emit_2ub(p, 0xd9, 0xe4);
1390 }
1391
1392 void x87_fucom( struct x86_function *p, struct x86_reg arg )
1393 {
1394 DUMP_R( arg );
1395 assert(arg.file == file_x87);
1396 emit_2ub(p, 0xdd, 0xe0+arg.idx);
1397 }
1398
1399 void x87_fucomp( struct x86_function *p, struct x86_reg arg )
1400 {
1401 DUMP_R( arg );
1402 assert(arg.file == file_x87);
1403 emit_2ub(p, 0xdd, 0xe8+arg.idx);
1404 note_x87_pop(p);
1405 }
1406
1407 void x87_fucompp( struct x86_function *p )
1408 {
1409 DUMP();
1410 emit_2ub(p, 0xda, 0xe9);
1411 note_x87_pop(p); /* pop twice */
1412 note_x87_pop(p); /* pop twice */
1413 }
1414
1415 void x87_fxch( struct x86_function *p, struct x86_reg arg )
1416 {
1417 DUMP_R( arg );
1418 assert(arg.file == file_x87);
1419 emit_2ub(p, 0xd9, 0xc8+arg.idx);
1420 }
1421
1422 void x87_fabs( struct x86_function *p )
1423 {
1424 DUMP();
1425 emit_2ub(p, 0xd9, 0xe1);
1426 }
1427
1428 void x87_fchs( struct x86_function *p )
1429 {
1430 DUMP();
1431 emit_2ub(p, 0xd9, 0xe0);
1432 }
1433
1434 void x87_fcos( struct x86_function *p )
1435 {
1436 DUMP();
1437 emit_2ub(p, 0xd9, 0xff);
1438 }
1439
1440
1441 void x87_fprndint( struct x86_function *p )
1442 {
1443 DUMP();
1444 emit_2ub(p, 0xd9, 0xfc);
1445 }
1446
1447 void x87_fscale( struct x86_function *p )
1448 {
1449 DUMP();
1450 emit_2ub(p, 0xd9, 0xfd);
1451 }
1452
1453 void x87_fsin( struct x86_function *p )
1454 {
1455 DUMP();
1456 emit_2ub(p, 0xd9, 0xfe);
1457 }
1458
1459 void x87_fsincos( struct x86_function *p )
1460 {
1461 DUMP();
1462 emit_2ub(p, 0xd9, 0xfb);
1463 }
1464
1465 void x87_fsqrt( struct x86_function *p )
1466 {
1467 DUMP();
1468 emit_2ub(p, 0xd9, 0xfa);
1469 }
1470
1471 void x87_fxtract( struct x86_function *p )
1472 {
1473 DUMP();
1474 emit_2ub(p, 0xd9, 0xf4);
1475 }
1476
1477 /* st0 = (2^st0)-1
1478 *
1479 * Restrictions: -1.0 <= st0 <= 1.0
1480 */
1481 void x87_f2xm1( struct x86_function *p )
1482 {
1483 DUMP();
1484 emit_2ub(p, 0xd9, 0xf0);
1485 }
1486
1487 /* st1 = st1 * log2(st0);
1488 * pop_stack;
1489 */
1490 void x87_fyl2x( struct x86_function *p )
1491 {
1492 DUMP();
1493 emit_2ub(p, 0xd9, 0xf1);
1494 note_x87_pop(p);
1495 }
1496
1497 /* st1 = st1 * log2(st0 + 1.0);
1498 * pop_stack;
1499 *
1500 * A fast operation, with restrictions: -.29 < st0 < .29
1501 */
1502 void x87_fyl2xp1( struct x86_function *p )
1503 {
1504 DUMP();
1505 emit_2ub(p, 0xd9, 0xf9);
1506 note_x87_pop(p);
1507 }
1508
1509
1510 void x87_fld( struct x86_function *p, struct x86_reg arg )
1511 {
1512 DUMP_R( arg );
1513 if (arg.file == file_x87)
1514 emit_2ub(p, 0xd9, 0xc0 + arg.idx);
1515 else {
1516 emit_1ub(p, 0xd9);
1517 emit_modrm_noreg(p, 0, arg);
1518 }
1519 note_x87_push(p);
1520 }
1521
1522 void x87_fst( struct x86_function *p, struct x86_reg dst )
1523 {
1524 DUMP_R( dst );
1525 if (dst.file == file_x87)
1526 emit_2ub(p, 0xdd, 0xd0 + dst.idx);
1527 else {
1528 emit_1ub(p, 0xd9);
1529 emit_modrm_noreg(p, 2, dst);
1530 }
1531 }
1532
1533 void x87_fstp( struct x86_function *p, struct x86_reg dst )
1534 {
1535 DUMP_R( dst );
1536 if (dst.file == file_x87)
1537 emit_2ub(p, 0xdd, 0xd8 + dst.idx);
1538 else {
1539 emit_1ub(p, 0xd9);
1540 emit_modrm_noreg(p, 3, dst);
1541 }
1542 note_x87_pop(p);
1543 }
1544
1545 void x87_fpop( struct x86_function *p )
1546 {
1547 x87_fstp( p, x86_make_reg( file_x87, 0 ));
1548 }
1549
1550
1551 void x87_fcom( struct x86_function *p, struct x86_reg dst )
1552 {
1553 DUMP_R( dst );
1554 if (dst.file == file_x87)
1555 emit_2ub(p, 0xd8, 0xd0 + dst.idx);
1556 else {
1557 emit_1ub(p, 0xd8);
1558 emit_modrm_noreg(p, 2, dst);
1559 }
1560 }
1561
1562
1563 void x87_fcomp( struct x86_function *p, struct x86_reg dst )
1564 {
1565 DUMP_R( dst );
1566 if (dst.file == file_x87)
1567 emit_2ub(p, 0xd8, 0xd8 + dst.idx);
1568 else {
1569 emit_1ub(p, 0xd8);
1570 emit_modrm_noreg(p, 3, dst);
1571 }
1572 note_x87_pop(p);
1573 }
1574
1575 void x87_fcomi( struct x86_function *p, struct x86_reg arg )
1576 {
1577 DUMP_R( arg );
1578 emit_2ub(p, 0xdb, 0xf0+arg.idx);
1579 }
1580
1581 void x87_fcomip( struct x86_function *p, struct x86_reg arg )
1582 {
1583 DUMP_R( arg );
1584 emit_2ub(p, 0xdb, 0xf0+arg.idx);
1585 note_x87_pop(p);
1586 }
1587
1588
1589 void x87_fnstsw( struct x86_function *p, struct x86_reg dst )
1590 {
1591 DUMP_R( dst );
1592 assert(dst.file == file_REG32);
1593
1594 if (dst.idx == reg_AX &&
1595 dst.mod == mod_REG)
1596 emit_2ub(p, 0xdf, 0xe0);
1597 else {
1598 emit_1ub(p, 0xdd);
1599 emit_modrm_noreg(p, 7, dst);
1600 }
1601 }
1602
1603
1604 void x87_fnstcw( struct x86_function *p, struct x86_reg dst )
1605 {
1606 DUMP_R( dst );
1607 assert(dst.file == file_REG32);
1608
1609 emit_1ub(p, 0x9b); /* WAIT -- needed? */
1610 emit_1ub(p, 0xd9);
1611 emit_modrm_noreg(p, 7, dst);
1612 }
1613
1614
1615
1616
1617 /***********************************************************************
1618 * MMX instructions
1619 */
1620
1621 void mmx_emms( struct x86_function *p )
1622 {
1623 DUMP();
1624 assert(p->need_emms);
1625 emit_2ub(p, 0x0f, 0x77);
1626 p->need_emms = 0;
1627 }
1628
1629 void mmx_packssdw( struct x86_function *p,
1630 struct x86_reg dst,
1631 struct x86_reg src )
1632 {
1633 DUMP_RR( dst, src );
1634 assert(dst.file == file_MMX &&
1635 (src.file == file_MMX || src.mod != mod_REG));
1636
1637 p->need_emms = 1;
1638
1639 emit_2ub(p, X86_TWOB, 0x6b);
1640 emit_modrm( p, dst, src );
1641 }
1642
1643 void mmx_packuswb( struct x86_function *p,
1644 struct x86_reg dst,
1645 struct x86_reg src )
1646 {
1647 DUMP_RR( dst, src );
1648 assert(dst.file == file_MMX &&
1649 (src.file == file_MMX || src.mod != mod_REG));
1650
1651 p->need_emms = 1;
1652
1653 emit_2ub(p, X86_TWOB, 0x67);
1654 emit_modrm( p, dst, src );
1655 }
1656
1657 void mmx_movd( struct x86_function *p,
1658 struct x86_reg dst,
1659 struct x86_reg src )
1660 {
1661 DUMP_RR( dst, src );
1662 p->need_emms = 1;
1663 emit_1ub(p, X86_TWOB);
1664 emit_op_modrm( p, 0x6e, 0x7e, dst, src );
1665 }
1666
1667 void mmx_movq( struct x86_function *p,
1668 struct x86_reg dst,
1669 struct x86_reg src )
1670 {
1671 DUMP_RR( dst, src );
1672 p->need_emms = 1;
1673 emit_1ub(p, X86_TWOB);
1674 emit_op_modrm( p, 0x6f, 0x7f, dst, src );
1675 }
1676
1677
1678 /***********************************************************************
1679 * Helper functions
1680 */
1681
1682
1683 void x86_cdecl_caller_push_regs( struct x86_function *p )
1684 {
1685 x86_push(p, x86_make_reg(file_REG32, reg_AX));
1686 x86_push(p, x86_make_reg(file_REG32, reg_CX));
1687 x86_push(p, x86_make_reg(file_REG32, reg_DX));
1688 }
1689
1690 void x86_cdecl_caller_pop_regs( struct x86_function *p )
1691 {
1692 x86_pop(p, x86_make_reg(file_REG32, reg_DX));
1693 x86_pop(p, x86_make_reg(file_REG32, reg_CX));
1694 x86_pop(p, x86_make_reg(file_REG32, reg_AX));
1695 }
1696
1697
1698 /* Retreive a reference to one of the function arguments, taking into
1699 * account any push/pop activity:
1700 */
1701 struct x86_reg x86_fn_arg( struct x86_function *p,
1702 unsigned arg )
1703 {
1704 return x86_make_disp(x86_make_reg(file_REG32, reg_SP),
1705 p->stack_offset + arg * 4); /* ??? */
1706 }
1707
1708
1709 void x86_init_func( struct x86_function *p )
1710 {
1711 p->size = 0;
1712 p->store = NULL;
1713 p->csr = p->store;
1714 DUMP_START();
1715 }
1716
1717 void x86_init_func_size( struct x86_function *p, unsigned code_size )
1718 {
1719 p->size = code_size;
1720 p->store = rtasm_exec_malloc(code_size);
1721 if (p->store == NULL) {
1722 p->store = p->error_overflow;
1723 }
1724 p->csr = p->store;
1725 DUMP_START();
1726 }
1727
1728 void x86_release_func( struct x86_function *p )
1729 {
1730 if (p->store && p->store != p->error_overflow)
1731 rtasm_exec_free(p->store);
1732
1733 p->store = NULL;
1734 p->csr = NULL;
1735 p->size = 0;
1736 }
1737
1738
1739 void (*x86_get_func( struct x86_function *p ))(void)
1740 {
1741 DUMP_END();
1742 if (DISASSEM && p->store)
1743 debug_printf("disassemble %p %p\n", p->store, p->csr);
1744
1745 if (p->store == p->error_overflow)
1746 return (void (*)(void)) NULL;
1747 else
1748 return (void (*)(void)) p->store;
1749 }
1750
1751 #else
1752
1753 void x86sse_dummy( void )
1754 {
1755 }
1756
1757 #endif