Makefile: Add rules to create tags and TAGS.
[tinycc.git] / i386-asm.c
1 /*
2 * i386 specific functions for TCC assembler
3 *
4 * Copyright (c) 2001, 2002 Fabrice Bellard
5 * Copyright (c) 2009 Frédéric Feret (x86_64 support)
6 *
7 * This library is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2 of the License, or (at your option) any later version.
11 *
12 * This library is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with this library; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 */
21
22 #include "tcc.h"
23
24 /* #define NB_ASM_REGS 8 */
25 #define MAX_OPERANDS 3
26 #define NB_SAVED_REGS 3
27
28 #define TOK_ASM_first TOK_ASM_clc
29 #define TOK_ASM_last TOK_ASM_emms
30 #define TOK_ASM_alllast TOK_ASM_pxor
31
32 #define OPC_JMP 0x01 /* jmp operand */
33 #define OPC_B 0x02 /* only used with OPC_WL */
34 #define OPC_WL 0x04 /* accepts w, l or no suffix */
35 #define OPC_BWL (OPC_B | OPC_WL) /* accepts b, w, l or no suffix */
36 #define OPC_REG 0x08 /* register is added to opcode */
37 #define OPC_MODRM 0x10 /* modrm encoding */
38 #define OPC_FWAIT 0x20 /* add fwait opcode */
39 #define OPC_TEST 0x40 /* test opcodes */
40 #define OPC_SHIFT 0x80 /* shift opcodes */
41 #define OPC_D16 0x0100 /* generate data16 prefix */
42 #define OPC_ARITH 0x0200 /* arithmetic opcodes */
43 #define OPC_SHORTJMP 0x0400 /* short jmp operand */
44 #define OPC_FARITH 0x0800 /* FPU arithmetic opcodes */
45 #ifdef TCC_TARGET_X86_64
46 # define OPC_WLQ 0x1000 /* accepts w, l, q or no suffix */
47 # define OPC_BWLQ (OPC_B | OPC_WLQ) /* accepts b, w, l, q or no suffix */
48 # define OPC_WLX OPC_WLQ
49 #else
50 # define OPC_WLX OPC_WL
51 #endif
52
53 #define OPC_GROUP_SHIFT 13
54
55 /* in order to compress the operand type, we use specific operands and
56 we or only with EA */
57 enum {
58 OPT_REG8=0, /* warning: value is hardcoded from TOK_ASM_xxx */
59 OPT_REG16, /* warning: value is hardcoded from TOK_ASM_xxx */
60 OPT_REG32, /* warning: value is hardcoded from TOK_ASM_xxx */
61 #ifdef TCC_TARGET_X86_64
62 OPT_REG64, /* warning: value is hardcoded from TOK_ASM_xxx */
63 #endif
64 OPT_MMX, /* warning: value is hardcoded from TOK_ASM_xxx */
65 OPT_SSE, /* warning: value is hardcoded from TOK_ASM_xxx */
66 OPT_CR, /* warning: value is hardcoded from TOK_ASM_xxx */
67 OPT_TR, /* warning: value is hardcoded from TOK_ASM_xxx */
68 OPT_DB, /* warning: value is hardcoded from TOK_ASM_xxx */
69 OPT_SEG,
70 OPT_ST,
71 OPT_IM8,
72 OPT_IM8S,
73 OPT_IM16,
74 OPT_IM32,
75 #ifdef TCC_TARGET_X86_64
76 OPT_IM64,
77 #endif
78 OPT_EAX, /* %al, %ax, %eax or %rax register */
79 OPT_ST0, /* %st(0) register */
80 OPT_CL, /* %cl register */
81 OPT_DX, /* %dx register */
82 OPT_ADDR, /* OP_EA with only offset */
83 OPT_INDIR, /* *(expr) */
84 /* composite types */
85 OPT_COMPOSITE_FIRST,
86 OPT_IM, /* IM8 | IM16 | IM32 | IM64 */
87 OPT_REG, /* REG8 | REG16 | REG32 | REG64 */
88 OPT_REGW, /* REG16 | REG32 | REG64 */
89 OPT_IMW, /* IM16 | IM32 | IM64 */
90 #ifdef TCC_TARGET_X86_64
91 OPT_IMNO64, /* IM16 | IM32 */
92 #endif
93 /* can be ored with any OPT_xxx */
94 OPT_EA = 0x80
95 };
96
97 #define OP_REG8 (1 << OPT_REG8)
98 #define OP_REG16 (1 << OPT_REG16)
99 #define OP_REG32 (1 << OPT_REG32)
100 #define OP_MMX (1 << OPT_MMX)
101 #define OP_SSE (1 << OPT_SSE)
102 #define OP_CR (1 << OPT_CR)
103 #define OP_TR (1 << OPT_TR)
104 #define OP_DB (1 << OPT_DB)
105 #define OP_SEG (1 << OPT_SEG)
106 #define OP_ST (1 << OPT_ST)
107 #define OP_IM8 (1 << OPT_IM8)
108 #define OP_IM8S (1 << OPT_IM8S)
109 #define OP_IM16 (1 << OPT_IM16)
110 #define OP_IM32 (1 << OPT_IM32)
111 #define OP_EAX (1 << OPT_EAX)
112 #define OP_ST0 (1 << OPT_ST0)
113 #define OP_CL (1 << OPT_CL)
114 #define OP_DX (1 << OPT_DX)
115 #define OP_ADDR (1 << OPT_ADDR)
116 #define OP_INDIR (1 << OPT_INDIR)
117 #ifdef TCC_TARGET_X86_64
118 # define OP_REG64 (1 << OPT_REG64)
119 # define OP_IM64 (1 << OPT_IM64)
120 #else
121 # define OP_REG64 0
122 # define OP_IM64 0
123 #endif
124
125 #define OP_EA 0x40000000
126 #define OP_REG (OP_REG8 | OP_REG16 | OP_REG32 | OP_REG64)
127
128 #ifdef TCC_TARGET_X86_64
129 # define OP_IM OP_IM64
130 # define TREG_XAX TREG_RAX
131 # define TREG_XCX TREG_RCX
132 # define TREG_XDX TREG_RDX
133 #else
134 # define OP_IM OP_IM32
135 # define TREG_XAX TREG_EAX
136 # define TREG_XCX TREG_ECX
137 # define TREG_XDX TREG_EDX
138 #endif
139
140 typedef struct ASMInstr {
141 uint16_t sym;
142 uint16_t opcode;
143 uint16_t instr_type;
144 uint8_t nb_ops;
145 uint8_t op_type[MAX_OPERANDS]; /* see OP_xxx */
146 } ASMInstr;
147
148 typedef struct Operand {
149 uint32_t type;
150 int8_t reg; /* register, -1 if none */
151 int8_t reg2; /* second register, -1 if none */
152 uint8_t shift;
153 ExprValue e;
154 } Operand;
155
156 static const uint8_t reg_to_size[9] = {
157 /*
158 [OP_REG8] = 0,
159 [OP_REG16] = 1,
160 [OP_REG32] = 2,
161 #ifdef TCC_TARGET_X86_64
162 [OP_REG64] = 3,
163 #endif
164 */
165 0, 0, 1, 0, 2, 0, 0, 0, 3
166 };
167
168 #define NB_TEST_OPCODES 30
169
170 static const uint8_t test_bits[NB_TEST_OPCODES] = {
171 0x00, /* o */
172 0x01, /* no */
173 0x02, /* b */
174 0x02, /* c */
175 0x02, /* nae */
176 0x03, /* nb */
177 0x03, /* nc */
178 0x03, /* ae */
179 0x04, /* e */
180 0x04, /* z */
181 0x05, /* ne */
182 0x05, /* nz */
183 0x06, /* be */
184 0x06, /* na */
185 0x07, /* nbe */
186 0x07, /* a */
187 0x08, /* s */
188 0x09, /* ns */
189 0x0a, /* p */
190 0x0a, /* pe */
191 0x0b, /* np */
192 0x0b, /* po */
193 0x0c, /* l */
194 0x0c, /* nge */
195 0x0d, /* nl */
196 0x0d, /* ge */
197 0x0e, /* le */
198 0x0e, /* ng */
199 0x0f, /* nle */
200 0x0f, /* g */
201 };
202
203 static const uint8_t segment_prefixes[] = {
204 0x26, /* es */
205 0x2e, /* cs */
206 0x36, /* ss */
207 0x3e, /* ds */
208 0x64, /* fs */
209 0x65 /* gs */
210 };
211
212 static const ASMInstr asm_instrs[] = {
213 #define ALT(x) x
214 #define DEF_ASM_OP0(name, opcode)
215 #define DEF_ASM_OP0L(name, opcode, group, instr_type) { TOK_ASM_ ## name, opcode, (instr_type | group << OPC_GROUP_SHIFT), 0 },
216 #define DEF_ASM_OP1(name, opcode, group, instr_type, op0) { TOK_ASM_ ## name, opcode, (instr_type | group << OPC_GROUP_SHIFT), 1, { op0 }},
217 #define DEF_ASM_OP2(name, opcode, group, instr_type, op0, op1) { TOK_ASM_ ## name, opcode, (instr_type | group << OPC_GROUP_SHIFT), 2, { op0, op1 }},
218 #define DEF_ASM_OP3(name, opcode, group, instr_type, op0, op1, op2) { TOK_ASM_ ## name, opcode, (instr_type | group << OPC_GROUP_SHIFT), 3, { op0, op1, op2 }},
219 #ifdef TCC_TARGET_X86_64
220 # include "x86_64-asm.h"
221 #else
222 # include "i386-asm.h"
223 #endif
224 /* last operation */
225 { 0, },
226 };
227
228 static const uint16_t op0_codes[] = {
229 #define ALT(x)
230 #define DEF_ASM_OP0(x, opcode) opcode,
231 #define DEF_ASM_OP0L(name, opcode, group, instr_type)
232 #define DEF_ASM_OP1(name, opcode, group, instr_type, op0)
233 #define DEF_ASM_OP2(name, opcode, group, instr_type, op0, op1)
234 #define DEF_ASM_OP3(name, opcode, group, instr_type, op0, op1, op2)
235 #ifdef TCC_TARGET_X86_64
236 # include "x86_64-asm.h"
237 #else
238 # include "i386-asm.h"
239 #endif
240 };
241
242 static inline int get_reg_shift(TCCState *s1)
243 {
244 int shift, v;
245 #ifdef I386_ASM_16
246 if (s1->seg_size == 16)
247 tcc_error("invalid effective address");
248 #endif
249 v = asm_int_expr(s1);
250 switch(v) {
251 case 1:
252 shift = 0;
253 break;
254 case 2:
255 shift = 1;
256 break;
257 case 4:
258 shift = 2;
259 break;
260 case 8:
261 shift = 3;
262 break;
263 default:
264 expect("1, 2, 4 or 8 constant");
265 shift = 0;
266 break;
267 }
268 return shift;
269 }
270
271 static int asm_parse_reg(void)
272 {
273 int reg = 0;
274 if (tok != '%')
275 goto error_32;
276 next();
277 if (tok >= TOK_ASM_eax && tok <= TOK_ASM_edi) {
278 reg = tok - TOK_ASM_eax;
279 #ifdef TCC_TARGET_X86_64
280 } else if (tok >= TOK_ASM_rax && tok <= TOK_ASM_rdi) {
281 reg = tok - TOK_ASM_rax;
282 #endif
283 #ifdef I386_ASM_16
284 } else if (tok >= TOK_ASM_ax && tok <= TOK_ASM_di) {
285 reg = tok - TOK_ASM_ax;
286 #endif
287 } else {
288 error_32:
289 expect("register");
290 }
291 next();
292 return reg;
293 }
294
295 static void parse_operand(TCCState *s1, Operand *op)
296 {
297 ExprValue e;
298 int reg, indir;
299 const char *p;
300
301 indir = 0;
302 if (tok == '*') {
303 next();
304 indir = OP_INDIR;
305 }
306
307 if (tok == '%') {
308 next();
309 if (tok >= TOK_ASM_al && tok <= TOK_ASM_db7) {
310 reg = tok - TOK_ASM_al;
311 op->type = 1 << (reg >> 3); /* WARNING: do not change constant order */
312 op->reg = reg & 7;
313 if ((op->type & OP_REG) && op->reg == TREG_XAX)
314 op->type |= OP_EAX;
315 else if (op->type == OP_REG8 && op->reg == TREG_XCX)
316 op->type |= OP_CL;
317 else if (op->type == OP_REG16 && op->reg == TREG_XDX)
318 op->type |= OP_DX;
319 } else if (tok >= TOK_ASM_dr0 && tok <= TOK_ASM_dr7) {
320 op->type = OP_DB;
321 op->reg = tok - TOK_ASM_dr0;
322 } else if (tok >= TOK_ASM_es && tok <= TOK_ASM_gs) {
323 op->type = OP_SEG;
324 op->reg = tok - TOK_ASM_es;
325 } else if (tok == TOK_ASM_st) {
326 op->type = OP_ST;
327 op->reg = 0;
328 next();
329 if (tok == '(') {
330 next();
331 if (tok != TOK_PPNUM)
332 goto reg_error;
333 p = tokc.cstr->data;
334 reg = p[0] - '0';
335 if ((unsigned)reg >= 8 || p[1] != '\0')
336 goto reg_error;
337 op->reg = reg;
338 next();
339 skip(')');
340 }
341 if (op->reg == 0)
342 op->type |= OP_ST0;
343 goto no_skip;
344 } else {
345 reg_error:
346 tcc_error("unknown register");
347 }
348 next();
349 no_skip: ;
350 } else if (tok == '$') {
351 /* constant value */
352 next();
353 asm_expr(s1, &e);
354 op->type = OP_IM;
355 op->e.v = e.v;
356 op->e.sym = e.sym;
357 if (!op->e.sym) {
358 if (op->e.v == (uint8_t)op->e.v)
359 op->type |= OP_IM8;
360 if (op->e.v == (int8_t)op->e.v)
361 op->type |= OP_IM8S;
362 if (op->e.v == (uint16_t)op->e.v)
363 op->type |= OP_IM16;
364 #ifdef TCC_TARGET_X86_64
365 if (op->e.v == (uint32_t)op->e.v)
366 op->type |= OP_IM32;
367 #endif
368 }
369 } else {
370 /* address(reg,reg2,shift) with all variants */
371 op->type = OP_EA;
372 op->reg = -1;
373 op->reg2 = -1;
374 op->shift = 0;
375 if (tok != '(') {
376 asm_expr(s1, &e);
377 op->e.v = e.v;
378 op->e.sym = e.sym;
379 } else {
380 next();
381 if (tok == '%') {
382 unget_tok('(');
383 op->e.v = 0;
384 op->e.sym = NULL;
385 } else {
386 /* bracketed offset expression */
387 asm_expr(s1, &e);
388 if (tok != ')')
389 expect(")");
390 next();
391 op->e.v = e.v;
392 op->e.sym = e.sym;
393 }
394 }
395 if (tok == '(') {
396 next();
397 if (tok != ',') {
398 op->reg = asm_parse_reg();
399 }
400 if (tok == ',') {
401 next();
402 if (tok != ',') {
403 op->reg2 = asm_parse_reg();
404 }
405 if (tok == ',') {
406 next();
407 op->shift = get_reg_shift(s1);
408 }
409 }
410 skip(')');
411 }
412 if (op->reg == -1 && op->reg2 == -1)
413 op->type |= OP_ADDR;
414 }
415 op->type |= indir;
416 }
417
418 /* XXX: unify with C code output ? */
419 ST_FUNC void gen_expr32(ExprValue *pe)
420 {
421 gen_addr32(pe->sym ? VT_SYM : 0, pe->sym, pe->v);
422 }
423
424 #ifdef TCC_TARGET_X86_64
425 static void gen_expr64(ExprValue *pe)
426 {
427 gen_addr64(pe->sym ? VT_SYM : 0, pe->sym, pe->v);
428 }
429 #endif
430
431 /* XXX: unify with C code output ? */
432 static void gen_disp32(ExprValue *pe)
433 {
434 Sym *sym = pe->sym;
435 if (sym && sym->r == cur_text_section->sh_num) {
436 /* same section: we can output an absolute value. Note
437 that the TCC compiler behaves differently here because
438 it always outputs a relocation to ease (future) code
439 elimination in the linker */
440 gen_le32(pe->v + sym->jnext - ind - 4);
441 } else {
442 if (sym && sym->type.t == VT_VOID) {
443 sym->type.t = VT_FUNC;
444 sym->type.ref = NULL;
445 }
446 gen_addrpc32(VT_SYM, sym, pe->v);
447 }
448 }
449
450 #ifdef I386_ASM_16
451 static void gen_expr16(ExprValue *pe)
452 {
453 if (pe->sym)
454 greloc(cur_text_section, pe->sym, ind, R_386_16);
455 gen_le16(pe->v);
456 }
457 static void gen_disp16(ExprValue *pe)
458 {
459 Sym *sym;
460 sym = pe->sym;
461 if (sym) {
462 if (sym->r == cur_text_section->sh_num) {
463 /* same section: we can output an absolute value. Note
464 that the TCC compiler behaves differently here because
465 it always outputs a relocation to ease (future) code
466 elimination in the linker */
467 gen_le16(pe->v + sym->jnext - ind - 2);
468 } else {
469 greloc(cur_text_section, sym, ind, R_386_PC16);
470 gen_le16(pe->v - 2);
471 }
472 } else {
473 /* put an empty PC32 relocation */
474 put_elf_reloc(symtab_section, cur_text_section,
475 ind, R_386_PC16, 0);
476 gen_le16(pe->v - 2);
477 }
478 }
479 #endif
480
481 /* generate the modrm operand */
482 static inline void asm_modrm(int reg, Operand *op)
483 {
484 int mod, reg1, reg2, sib_reg1;
485
486 if (op->type & (OP_REG | OP_MMX | OP_SSE)) {
487 g(0xc0 + (reg << 3) + op->reg);
488 } else if (op->reg == -1 && op->reg2 == -1) {
489 /* displacement only */
490 #ifdef I386_ASM_16
491 if (tcc_state->seg_size == 16) {
492 g(0x06 + (reg << 3));
493 gen_expr16(&op->e);
494 } else if (tcc_state->seg_size == 32)
495 #endif
496 {
497 g(0x05 + (reg << 3));
498 gen_expr32(&op->e);
499 }
500 } else {
501 sib_reg1 = op->reg;
502 /* fist compute displacement encoding */
503 if (sib_reg1 == -1) {
504 sib_reg1 = 5;
505 mod = 0x00;
506 } else if (op->e.v == 0 && !op->e.sym && op->reg != 5) {
507 mod = 0x00;
508 } else if (op->e.v == (int8_t)op->e.v && !op->e.sym) {
509 mod = 0x40;
510 } else {
511 mod = 0x80;
512 }
513 /* compute if sib byte needed */
514 reg1 = op->reg;
515 if (op->reg2 != -1)
516 reg1 = 4;
517 #ifdef I386_ASM_16
518 if (tcc_state->seg_size == 32) {
519 #endif
520 g(mod + (reg << 3) + reg1);
521 if (reg1 == 4) {
522 /* add sib byte */
523 reg2 = op->reg2;
524 if (reg2 == -1)
525 reg2 = 4; /* indicate no index */
526 g((op->shift << 6) + (reg2 << 3) + sib_reg1);
527 }
528 #ifdef I386_ASM_16
529 } else if (tcc_state->seg_size == 16) {
530 /* edi = 7, esi = 6 --> di = 5, si = 4 */
531 if ((reg1 == 6) || (reg1 == 7)) {
532 reg1 -= 2;
533 /* ebx = 3 --> bx = 7 */
534 } else if (reg1 == 3) {
535 reg1 = 7;
536 /* o32 = 5 --> o16 = 6 */
537 } else if (reg1 == 5) {
538 reg1 = 6;
539 /* sib not valid in 16-bit mode */
540 } else if (reg1 == 4) {
541 reg2 = op->reg2;
542 /* bp + si + offset */
543 if ((sib_reg1 == 5) && (reg2 == 6)) {
544 reg1 = 2;
545 /* bp + di + offset */
546 } else if ((sib_reg1 == 5) && (reg2 == 7)) {
547 reg1 = 3;
548 /* bx + si + offset */
549 } else if ((sib_reg1 == 3) && (reg2 == 6)) {
550 reg1 = 0;
551 /* bx + di + offset */
552 } else if ((sib_reg1 == 3) && (reg2 == 7)) {
553 reg1 = 1;
554 } else {
555 tcc_error("invalid effective address");
556 }
557 if (op->e.v == 0)
558 mod = 0;
559 } else {
560 tcc_error("invalid register");
561 }
562 g(mod + (reg << 3) + reg1);
563 }
564 #endif
565 /* add offset */
566 if (mod == 0x40) {
567 g(op->e.v);
568 } else if (mod == 0x80 || op->reg == -1) {
569 #ifdef I386_ASM_16
570 if (tcc_state->seg_size == 16)
571 gen_expr16(&op->e);
572 else if (tcc_state->seg_size == 32)
573 #endif
574 gen_expr32(&op->e);
575 }
576 }
577 }
578
579 ST_FUNC void asm_opcode(TCCState *s1, int opcode)
580 {
581 const ASMInstr *pa;
582 int i, modrm_index, reg, v, op1, is_short_jmp, seg_prefix;
583 int nb_ops, s;
584 Operand ops[MAX_OPERANDS], *pop;
585 int op_type[3]; /* decoded op type */
586 #ifdef I386_ASM_16
587 static int a32 = 0, o32 = 0, addr32 = 0, data32 = 0;
588 #endif
589
590 /* force synthetic ';' after prefix instruction, so we can handle */
591 /* one-line things like "rep stosb" instead of only "rep\nstosb" */
592 if (opcode >= TOK_ASM_wait && opcode <= TOK_ASM_repnz)
593 unget_tok(';');
594
595 /* get operands */
596 pop = ops;
597 nb_ops = 0;
598 seg_prefix = 0;
599 for(;;) {
600 if (tok == ';' || tok == TOK_LINEFEED)
601 break;
602 if (nb_ops >= MAX_OPERANDS) {
603 tcc_error("incorrect number of operands");
604 }
605 parse_operand(s1, pop);
606 if (tok == ':') {
607 if (pop->type != OP_SEG || seg_prefix)
608 tcc_error("incorrect prefix");
609 seg_prefix = segment_prefixes[pop->reg];
610 next();
611 parse_operand(s1, pop);
612 #ifndef I386_ASM_16
613 if (!(pop->type & OP_EA)) {
614 tcc_error("segment prefix must be followed by memory reference");
615 }
616 #endif
617 }
618 pop++;
619 nb_ops++;
620 if (tok != ',')
621 break;
622 next();
623 }
624
625 is_short_jmp = 0;
626 s = 0; /* avoid warning */
627
628 /* optimize matching by using a lookup table (no hashing is needed
629 !) */
630 for(pa = asm_instrs; pa->sym != 0; pa++) {
631 s = 0;
632 if (pa->instr_type & OPC_FARITH) {
633 v = opcode - pa->sym;
634 if (!((unsigned)v < 8 * 6 && (v % 6) == 0))
635 continue;
636 } else if (pa->instr_type & OPC_ARITH) {
637 if (!(opcode >= pa->sym && opcode < pa->sym + 8*NBWLX))
638 continue;
639 s = (opcode - pa->sym) % NBWLX;
640 } else if (pa->instr_type & OPC_SHIFT) {
641 if (!(opcode >= pa->sym && opcode < pa->sym + 7*NBWLX))
642 continue;
643 s = (opcode - pa->sym) % NBWLX;
644 } else if (pa->instr_type & OPC_TEST) {
645 if (!(opcode >= pa->sym && opcode < pa->sym + NB_TEST_OPCODES))
646 continue;
647 } else if (pa->instr_type & OPC_B) {
648 if (!(opcode >= pa->sym && opcode < pa->sym + NBWLX))
649 continue;
650 s = opcode - pa->sym;
651 } else if (pa->instr_type & OPC_WLX) {
652 if (!(opcode >= pa->sym && opcode < pa->sym + NBWLX-1))
653 continue;
654 s = opcode - pa->sym + 1;
655 } else {
656 if (pa->sym != opcode)
657 continue;
658 }
659 if (pa->nb_ops != nb_ops)
660 continue;
661 /* now decode and check each operand */
662 for(i = 0; i < nb_ops; i++) {
663 int op1, op2;
664 op1 = pa->op_type[i];
665 op2 = op1 & 0x1f;
666 switch(op2) {
667 case OPT_IM:
668 v = OP_IM8 | OP_IM16 | OP_IM32 | OP_IM64;
669 break;
670 case OPT_REG:
671 v = OP_REG8 | OP_REG16 | OP_REG32 | OP_REG64;
672 break;
673 case OPT_REGW:
674 v = OP_REG16 | OP_REG32 | OP_REG64;
675 break;
676 case OPT_IMW:
677 v = OP_IM16 | OP_IM32 | OP_IM64;
678 break;
679 #ifdef TCC_TARGET_X86_64
680 case OPT_IMNO64:
681 v = OP_IM16 | OP_IM32;
682 break;
683 #endif
684 default:
685 v = 1 << op2;
686 break;
687 }
688 if (op1 & OPT_EA)
689 v |= OP_EA;
690 op_type[i] = v;
691 if ((ops[i].type & v) == 0)
692 goto next;
693 }
694 /* all is matching ! */
695 break;
696 next: ;
697 }
698 if (pa->sym == 0) {
699 if (opcode >= TOK_ASM_first && opcode <= TOK_ASM_last) {
700 int b;
701 b = op0_codes[opcode - TOK_ASM_first];
702 #ifdef I386_ASM_16
703 if (opcode == TOK_ASM_o32) {
704 if (s1->seg_size == 32)
705 tcc_error("incorrect prefix");
706 else
707 o32 = data32 = 1;
708 } else if (opcode == TOK_ASM_a32) {
709 if (s1->seg_size == 32)
710 tcc_error("incorrect prefix");
711 else
712 a32 = addr32 = 1;
713 }
714 #endif
715 if (b & 0xff00)
716 g(b >> 8);
717 g(b);
718 return;
719 } else if (opcode <= TOK_ASM_alllast) {
720 tcc_error("bad operand with opcode '%s'",
721 get_tok_str(opcode, NULL));
722 } else {
723 tcc_error("unknown opcode '%s'",
724 get_tok_str(opcode, NULL));
725 }
726 }
727 /* if the size is unknown, then evaluate it (OPC_B or OPC_WL case) */
728 if (s == NBWLX-1) {
729 for(i = 0; s == NBWLX-1 && i < nb_ops; i++) {
730 if ((ops[i].type & OP_REG) && !(op_type[i] & (OP_CL | OP_DX)))
731 s = reg_to_size[ops[i].type & OP_REG];
732 }
733 if (s == NBWLX-1) {
734 if ((opcode == TOK_ASM_push || opcode == TOK_ASM_pop) &&
735 (ops[0].type & (OP_SEG | OP_IM8S | OP_IM32 | OP_IM64)))
736 s = 2;
737 else
738 tcc_error("cannot infer opcode suffix");
739 }
740 }
741
742 #ifdef I386_ASM_16
743 for(i = 0; i < nb_ops; i++) {
744 if (ops[i].type & OP_REG32) {
745 if (s1->seg_size == 16)
746 o32 = 1;
747 } else if (!(ops[i].type & OP_REG32)) {
748 if (s1->seg_size == 32)
749 o32 = 1;
750 }
751 }
752
753
754 if (s == 1 || (pa->instr_type & OPC_D16)) {
755 if (s1->seg_size == 32)
756 o32 = 1;
757 } else if (s == 2) {
758 if (s1->seg_size == 16) {
759 if (!(pa->instr_type & OPC_D16))
760 o32 = 1;
761 }
762 }
763
764 /* generate a16/a32 prefix if needed */
765 if ((a32 == 1) && (addr32 == 0))
766 g(0x67);
767 /* generate o16/o32 prefix if needed */
768 if ((o32 == 1) && (data32 == 0))
769 g(0x66);
770
771 addr32 = data32 = 0;
772 #else
773 /* generate data16 prefix if needed */
774 if (s == 1 || (pa->instr_type & OPC_D16))
775 g(0x66);
776 #ifdef TCC_TARGET_X86_64
777 else if (s == 3) {
778 /* generate REX prefix */
779 if ((opcode != TOK_ASM_push && opcode != TOK_ASM_pop)
780 || !(ops[0].type & OP_REG64))
781 g(0x48);
782 }
783 #endif
784 #endif
785
786 /* now generates the operation */
787 if (pa->instr_type & OPC_FWAIT)
788 g(0x9b);
789 if (seg_prefix)
790 g(seg_prefix);
791
792 v = pa->opcode;
793 if ((v == 0x69 || v == 0x6b) && nb_ops == 2) {
794 /* kludge for imul $im, %reg */
795 nb_ops = 3;
796 ops[2] = ops[1];
797 op_type[2] = op_type[1];
798 } else if (v == 0xcd && ops[0].e.v == 3 && !ops[0].e.sym) {
799 v--; /* int $3 case */
800 nb_ops = 0;
801 } else if ((v == 0x06 || v == 0x07)) {
802 if (ops[0].reg >= 4) {
803 /* push/pop %fs or %gs */
804 v = 0x0fa0 + (v - 0x06) + ((ops[0].reg - 4) << 3);
805 } else {
806 v += ops[0].reg << 3;
807 }
808 nb_ops = 0;
809 } else if (v <= 0x05) {
810 /* arith case */
811 v += ((opcode - TOK_ASM_addb) / NBWLX) << 3;
812 } else if ((pa->instr_type & (OPC_FARITH | OPC_MODRM)) == OPC_FARITH) {
813 /* fpu arith case */
814 v += ((opcode - pa->sym) / 6) << 3;
815 }
816 if (pa->instr_type & OPC_REG) {
817 for(i = 0; i < nb_ops; i++) {
818 if (op_type[i] & (OP_REG | OP_ST)) {
819 v += ops[i].reg;
820 break;
821 }
822 }
823 /* mov $im, %reg case */
824 if (pa->opcode == 0xb0 && s >= 1)
825 v += 7;
826 }
827 if (pa->instr_type & OPC_B)
828 v += s >= 1;
829 if (pa->instr_type & OPC_TEST)
830 v += test_bits[opcode - pa->sym];
831 if (pa->instr_type & OPC_SHORTJMP) {
832 Sym *sym;
833 int jmp_disp;
834
835 /* see if we can really generate the jump with a byte offset */
836 sym = ops[0].e.sym;
837 if (!sym)
838 goto no_short_jump;
839 if (sym->r != cur_text_section->sh_num)
840 goto no_short_jump;
841 jmp_disp = ops[0].e.v + sym->jnext - ind - 2;
842 if (jmp_disp == (int8_t)jmp_disp) {
843 /* OK to generate jump */
844 is_short_jmp = 1;
845 ops[0].e.v = jmp_disp;
846 } else {
847 no_short_jump:
848 if (pa->instr_type & OPC_JMP) {
849 /* long jump will be allowed. need to modify the
850 opcode slightly */
851 if (v == 0xeb)
852 v = 0xe9;
853 else
854 v += 0x0f10;
855 } else {
856 tcc_error("invalid displacement");
857 }
858 }
859 }
860 op1 = v >> 8;
861 if (op1)
862 g(op1);
863 g(v);
864
865 /* search which operand will used for modrm */
866 modrm_index = 0;
867 if (pa->instr_type & OPC_SHIFT) {
868 reg = (opcode - pa->sym) / NBWLX;
869 if (reg == 6)
870 reg = 7;
871 } else if (pa->instr_type & OPC_ARITH) {
872 reg = (opcode - pa->sym) / NBWLX;
873 } else if (pa->instr_type & OPC_FARITH) {
874 reg = (opcode - pa->sym) / 6;
875 } else {
876 reg = (pa->instr_type >> OPC_GROUP_SHIFT) & 7;
877 }
878 if (pa->instr_type & OPC_MODRM) {
879 /* first look for an ea operand */
880 for(i = 0;i < nb_ops; i++) {
881 if (op_type[i] & OP_EA)
882 goto modrm_found;
883 }
884 /* then if not found, a register or indirection (shift instructions) */
885 for(i = 0;i < nb_ops; i++) {
886 if (op_type[i] & (OP_REG | OP_MMX | OP_SSE | OP_INDIR))
887 goto modrm_found;
888 }
889 #ifdef ASM_DEBUG
890 tcc_error("bad op table");
891 #endif
892 modrm_found:
893 modrm_index = i;
894 /* if a register is used in another operand then it is
895 used instead of group */
896 for(i = 0;i < nb_ops; i++) {
897 v = op_type[i];
898 if (i != modrm_index &&
899 (v & (OP_REG | OP_MMX | OP_SSE | OP_CR | OP_TR | OP_DB | OP_SEG))) {
900 reg = ops[i].reg;
901 break;
902 }
903 }
904
905 asm_modrm(reg, &ops[modrm_index]);
906 }
907
908 /* emit constants */
909 #ifndef TCC_TARGET_X86_64
910 if (pa->opcode == 0x9a || pa->opcode == 0xea) {
911 /* ljmp or lcall kludge */
912 #ifdef I386_ASM_16
913 if (s1->seg_size == 16 && o32 == 0)
914 gen_expr16(&ops[1].e);
915 else
916 #endif
917 gen_expr32(&ops[1].e);
918 if (ops[0].e.sym)
919 tcc_error("cannot relocate");
920 gen_le16(ops[0].e.v);
921 return;
922 }
923 #endif
924 for(i = 0;i < nb_ops; i++) {
925 v = op_type[i];
926 if (v & (OP_IM8 | OP_IM16 | OP_IM32 | OP_IM64 | OP_IM8S | OP_ADDR)) {
927 /* if multiple sizes are given it means we must look
928 at the op size */
929 if ((v | OP_IM8 | OP_IM64) == (OP_IM8 | OP_IM16 | OP_IM32 | OP_IM64)) {
930 if (s == 0)
931 v = OP_IM8;
932 else if (s == 1)
933 v = OP_IM16;
934 else if (s == 2 || (v & OP_IM64) == 0)
935 v = OP_IM32;
936 else
937 v = OP_IM64;
938 }
939 if (v & (OP_IM8 | OP_IM8S)) {
940 if (ops[i].e.sym)
941 goto error_relocate;
942 g(ops[i].e.v);
943 } else if (v & OP_IM16) {
944 #ifdef I386_ASM_16
945 if (s1->seg_size == 16)
946 gen_expr16(&ops[i].e);
947 else
948 #endif
949 if (ops[i].e.sym)
950 error_relocate:
951 tcc_error("cannot relocate");
952 else
953 gen_le16(ops[i].e.v);
954 } else {
955 if (pa->instr_type & (OPC_JMP | OPC_SHORTJMP)) {
956 if (is_short_jmp)
957 g(ops[i].e.v);
958 #ifdef I386_ASM_16
959 else if (s1->seg_size == 16)
960 gen_disp16(&ops[i].e);
961 #endif
962 else
963 gen_disp32(&ops[i].e);
964 } else {
965 #ifdef I386_ASM_16
966 if (s1->seg_size == 16 && !((o32 == 1) && (v & OP_IM32)))
967 gen_expr16(&ops[i].e);
968 else
969 #endif
970 #ifdef TCC_TARGET_X86_64
971 if (v & OP_IM64)
972 gen_expr64(&ops[i].e);
973 else
974 #endif
975 gen_expr32(&ops[i].e);
976 }
977 }
978 #ifdef I386_ASM_16
979 } else if (v & (OP_REG16 | OP_REG32)) {
980 if (pa->instr_type & (OPC_JMP | OPC_SHORTJMP)) {
981 /* jmp $r */
982 g(0xE0 + ops[i].reg);
983 }
984 #endif
985 #ifdef TCC_TARGET_X86_64
986 } else if (v & (OP_REG32 | OP_REG64)) {
987 if (pa->instr_type & (OPC_JMP | OPC_SHORTJMP)) {
988 /* jmp $r */
989 g(0xE0 + ops[i].reg);
990 }
991 #endif
992 }
993 }
994 #ifdef I386_ASM_16
995 a32 = o32 = 0;
996 #endif
997 }
998
999 /* return the constraint priority (we allocate first the lowest
1000 numbered constraints) */
1001 static inline int constraint_priority(const char *str)
1002 {
1003 int priority, c, pr;
1004
1005 /* we take the lowest priority */
1006 priority = 0;
1007 for(;;) {
1008 c = *str;
1009 if (c == '\0')
1010 break;
1011 str++;
1012 switch(c) {
1013 case 'A':
1014 pr = 0;
1015 break;
1016 case 'a':
1017 case 'b':
1018 case 'c':
1019 case 'd':
1020 case 'S':
1021 case 'D':
1022 pr = 1;
1023 break;
1024 case 'q':
1025 pr = 2;
1026 break;
1027 case 'r':
1028 pr = 3;
1029 break;
1030 case 'N':
1031 case 'M':
1032 case 'I':
1033 case 'i':
1034 case 'm':
1035 case 'g':
1036 pr = 4;
1037 break;
1038 default:
1039 tcc_error("unknown constraint '%c'", c);
1040 pr = 0;
1041 }
1042 if (pr > priority)
1043 priority = pr;
1044 }
1045 return priority;
1046 }
1047
1048 static const char *skip_constraint_modifiers(const char *p)
1049 {
1050 while (*p == '=' || *p == '&' || *p == '+' || *p == '%')
1051 p++;
1052 return p;
1053 }
1054
1055 #define REG_OUT_MASK 0x01
1056 #define REG_IN_MASK 0x02
1057
1058 #define is_reg_allocated(reg) (regs_allocated[reg] & reg_mask)
1059
1060 ST_FUNC void asm_compute_constraints(ASMOperand *operands,
1061 int nb_operands, int nb_outputs,
1062 const uint8_t *clobber_regs,
1063 int *pout_reg)
1064 {
1065 ASMOperand *op;
1066 int sorted_op[MAX_ASM_OPERANDS];
1067 int i, j, k, p1, p2, tmp, reg, c, reg_mask;
1068 const char *str;
1069 uint8_t regs_allocated[NB_ASM_REGS];
1070
1071 /* init fields */
1072 for(i=0;i<nb_operands;i++) {
1073 op = &operands[i];
1074 op->input_index = -1;
1075 op->ref_index = -1;
1076 op->reg = -1;
1077 op->is_memory = 0;
1078 op->is_rw = 0;
1079 }
1080 /* compute constraint priority and evaluate references to output
1081 constraints if input constraints */
1082 for(i=0;i<nb_operands;i++) {
1083 op = &operands[i];
1084 str = op->constraint;
1085 str = skip_constraint_modifiers(str);
1086 if (isnum(*str) || *str == '[') {
1087 /* this is a reference to another constraint */
1088 k = find_constraint(operands, nb_operands, str, NULL);
1089 if ((unsigned)k >= i || i < nb_outputs)
1090 tcc_error("invalid reference in constraint %d ('%s')",
1091 i, str);
1092 op->ref_index = k;
1093 if (operands[k].input_index >= 0)
1094 tcc_error("cannot reference twice the same operand");
1095 operands[k].input_index = i;
1096 op->priority = 5;
1097 } else {
1098 op->priority = constraint_priority(str);
1099 }
1100 }
1101
1102 /* sort operands according to their priority */
1103 for(i=0;i<nb_operands;i++)
1104 sorted_op[i] = i;
1105 for(i=0;i<nb_operands - 1;i++) {
1106 for(j=i+1;j<nb_operands;j++) {
1107 p1 = operands[sorted_op[i]].priority;
1108 p2 = operands[sorted_op[j]].priority;
1109 if (p2 < p1) {
1110 tmp = sorted_op[i];
1111 sorted_op[i] = sorted_op[j];
1112 sorted_op[j] = tmp;
1113 }
1114 }
1115 }
1116
1117 for(i = 0;i < NB_ASM_REGS; i++) {
1118 if (clobber_regs[i])
1119 regs_allocated[i] = REG_IN_MASK | REG_OUT_MASK;
1120 else
1121 regs_allocated[i] = 0;
1122 }
1123 /* esp cannot be used */
1124 regs_allocated[4] = REG_IN_MASK | REG_OUT_MASK;
1125 /* ebp cannot be used yet */
1126 regs_allocated[5] = REG_IN_MASK | REG_OUT_MASK;
1127
1128 /* allocate registers and generate corresponding asm moves */
1129 for(i=0;i<nb_operands;i++) {
1130 j = sorted_op[i];
1131 op = &operands[j];
1132 str = op->constraint;
1133 /* no need to allocate references */
1134 if (op->ref_index >= 0)
1135 continue;
1136 /* select if register is used for output, input or both */
1137 if (op->input_index >= 0) {
1138 reg_mask = REG_IN_MASK | REG_OUT_MASK;
1139 } else if (j < nb_outputs) {
1140 reg_mask = REG_OUT_MASK;
1141 } else {
1142 reg_mask = REG_IN_MASK;
1143 }
1144 try_next:
1145 c = *str++;
1146 switch(c) {
1147 case '=':
1148 goto try_next;
1149 case '+':
1150 op->is_rw = 1;
1151 /* FALL THRU */
1152 case '&':
1153 if (j >= nb_outputs)
1154 tcc_error("'%c' modifier can only be applied to outputs", c);
1155 reg_mask = REG_IN_MASK | REG_OUT_MASK;
1156 goto try_next;
1157 case 'A':
1158 /* allocate both eax and edx */
1159 if (is_reg_allocated(TREG_XAX) ||
1160 is_reg_allocated(TREG_XDX))
1161 goto try_next;
1162 op->is_llong = 1;
1163 op->reg = TREG_XAX;
1164 regs_allocated[TREG_XAX] |= reg_mask;
1165 regs_allocated[TREG_XDX] |= reg_mask;
1166 break;
1167 case 'a':
1168 reg = TREG_XAX;
1169 goto alloc_reg;
1170 case 'b':
1171 reg = 3;
1172 goto alloc_reg;
1173 case 'c':
1174 reg = TREG_XCX;
1175 goto alloc_reg;
1176 case 'd':
1177 reg = TREG_XDX;
1178 goto alloc_reg;
1179 case 'S':
1180 reg = 6;
1181 goto alloc_reg;
1182 case 'D':
1183 reg = 7;
1184 alloc_reg:
1185 if (is_reg_allocated(reg))
1186 goto try_next;
1187 goto reg_found;
1188 case 'q':
1189 /* eax, ebx, ecx or edx */
1190 for(reg = 0; reg < 4; reg++) {
1191 if (!is_reg_allocated(reg))
1192 goto reg_found;
1193 }
1194 goto try_next;
1195 case 'r':
1196 /* any general register */
1197 for(reg = 0; reg < 8; reg++) {
1198 if (!is_reg_allocated(reg))
1199 goto reg_found;
1200 }
1201 goto try_next;
1202 reg_found:
1203 /* now we can reload in the register */
1204 op->is_llong = 0;
1205 op->reg = reg;
1206 regs_allocated[reg] |= reg_mask;
1207 break;
1208 case 'i':
1209 if (!((op->vt->r & (VT_VALMASK | VT_LVAL)) == VT_CONST))
1210 goto try_next;
1211 break;
1212 case 'I':
1213 case 'N':
1214 case 'M':
1215 if (!((op->vt->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST))
1216 goto try_next;
1217 break;
1218 case 'm':
1219 case 'g':
1220 /* nothing special to do because the operand is already in
1221 memory, except if the pointer itself is stored in a
1222 memory variable (VT_LLOCAL case) */
1223 /* XXX: fix constant case */
1224 /* if it is a reference to a memory zone, it must lie
1225 in a register, so we reserve the register in the
1226 input registers and a load will be generated
1227 later */
1228 if (j < nb_outputs || c == 'm') {
1229 if ((op->vt->r & VT_VALMASK) == VT_LLOCAL) {
1230 /* any general register */
1231 for(reg = 0; reg < 8; reg++) {
1232 if (!(regs_allocated[reg] & REG_IN_MASK))
1233 goto reg_found1;
1234 }
1235 goto try_next;
1236 reg_found1:
1237 /* now we can reload in the register */
1238 regs_allocated[reg] |= REG_IN_MASK;
1239 op->reg = reg;
1240 op->is_memory = 1;
1241 }
1242 }
1243 break;
1244 default:
1245 tcc_error("asm constraint %d ('%s') could not be satisfied",
1246 j, op->constraint);
1247 break;
1248 }
1249 /* if a reference is present for that operand, we assign it too */
1250 if (op->input_index >= 0) {
1251 operands[op->input_index].reg = op->reg;
1252 operands[op->input_index].is_llong = op->is_llong;
1253 }
1254 }
1255
1256 /* compute out_reg. It is used to store outputs registers to memory
1257 locations references by pointers (VT_LLOCAL case) */
1258 *pout_reg = -1;
1259 for(i=0;i<nb_operands;i++) {
1260 op = &operands[i];
1261 if (op->reg >= 0 &&
1262 (op->vt->r & VT_VALMASK) == VT_LLOCAL &&
1263 !op->is_memory) {
1264 for(reg = 0; reg < 8; reg++) {
1265 if (!(regs_allocated[reg] & REG_OUT_MASK))
1266 goto reg_found2;
1267 }
1268 tcc_error("could not find free output register for reloading");
1269 reg_found2:
1270 *pout_reg = reg;
1271 break;
1272 }
1273 }
1274
1275 /* print sorted constraints */
1276 #ifdef ASM_DEBUG
1277 for(i=0;i<nb_operands;i++) {
1278 j = sorted_op[i];
1279 op = &operands[j];
1280 printf("%%%d [%s]: \"%s\" r=0x%04x reg=%d\n",
1281 j,
1282 op->id ? get_tok_str(op->id, NULL) : "",
1283 op->constraint,
1284 op->vt->r,
1285 op->reg);
1286 }
1287 if (*pout_reg >= 0)
1288 printf("out_reg=%d\n", *pout_reg);
1289 #endif
1290 }
1291
1292 ST_FUNC void subst_asm_operand(CString *add_str,
1293 SValue *sv, int modifier)
1294 {
1295 int r, reg, size, val;
1296 char buf[64];
1297
1298 r = sv->r;
1299 if ((r & VT_VALMASK) == VT_CONST) {
1300 if (!(r & VT_LVAL) && modifier != 'c' && modifier != 'n')
1301 cstr_ccat(add_str, '$');
1302 if (r & VT_SYM) {
1303 cstr_cat(add_str, get_tok_str(sv->sym->v, NULL));
1304 if (sv->c.i != 0) {
1305 cstr_ccat(add_str, '+');
1306 } else {
1307 return;
1308 }
1309 }
1310 val = sv->c.i;
1311 if (modifier == 'n')
1312 val = -val;
1313 snprintf(buf, sizeof(buf), "%d", sv->c.i);
1314 cstr_cat(add_str, buf);
1315 } else if ((r & VT_VALMASK) == VT_LOCAL) {
1316 snprintf(buf, sizeof(buf), "%d(%%ebp)", sv->c.i);
1317 cstr_cat(add_str, buf);
1318 } else if (r & VT_LVAL) {
1319 reg = r & VT_VALMASK;
1320 if (reg >= VT_CONST)
1321 tcc_error("internal compiler error");
1322 snprintf(buf, sizeof(buf), "(%%%s)",
1323 get_tok_str(TOK_ASM_eax + reg, NULL));
1324 cstr_cat(add_str, buf);
1325 } else {
1326 /* register case */
1327 reg = r & VT_VALMASK;
1328 if (reg >= VT_CONST)
1329 tcc_error("internal compiler error");
1330
1331 /* choose register operand size */
1332 if ((sv->type.t & VT_BTYPE) == VT_BYTE)
1333 size = 1;
1334 else if ((sv->type.t & VT_BTYPE) == VT_SHORT)
1335 size = 2;
1336 #ifdef TCC_TARGET_X86_64
1337 else if ((sv->type.t & VT_BTYPE) == VT_LLONG)
1338 size = 8;
1339 #endif
1340 else
1341 size = 4;
1342 if (size == 1 && reg >= 4)
1343 size = 4;
1344
1345 if (modifier == 'b') {
1346 if (reg >= 4)
1347 tcc_error("cannot use byte register");
1348 size = 1;
1349 } else if (modifier == 'h') {
1350 if (reg >= 4)
1351 tcc_error("cannot use byte register");
1352 size = -1;
1353 } else if (modifier == 'w') {
1354 size = 2;
1355 #ifdef TCC_TARGET_X86_64
1356 } else if (modifier == 'q') {
1357 size = 8;
1358 #endif
1359 }
1360
1361 switch(size) {
1362 case -1:
1363 reg = TOK_ASM_ah + reg;
1364 break;
1365 case 1:
1366 reg = TOK_ASM_al + reg;
1367 break;
1368 case 2:
1369 reg = TOK_ASM_ax + reg;
1370 break;
1371 default:
1372 reg = TOK_ASM_eax + reg;
1373 break;
1374 #ifdef TCC_TARGET_X86_64
1375 case 8:
1376 reg = TOK_ASM_rax + reg;
1377 break;
1378 #endif
1379 }
1380 snprintf(buf, sizeof(buf), "%%%s", get_tok_str(reg, NULL));
1381 cstr_cat(add_str, buf);
1382 }
1383 }
1384
1385 /* generate prolog and epilog code for asm statement */
1386 ST_FUNC void asm_gen_code(ASMOperand *operands, int nb_operands,
1387 int nb_outputs, int is_output,
1388 uint8_t *clobber_regs,
1389 int out_reg)
1390 {
1391 uint8_t regs_allocated[NB_ASM_REGS];
1392 ASMOperand *op;
1393 int i, reg;
1394 static uint8_t reg_saved[NB_SAVED_REGS] = { 3, 6, 7 };
1395
1396 /* mark all used registers */
1397 memcpy(regs_allocated, clobber_regs, sizeof(regs_allocated));
1398 for(i = 0; i < nb_operands;i++) {
1399 op = &operands[i];
1400 if (op->reg >= 0)
1401 regs_allocated[op->reg] = 1;
1402 }
1403 if (!is_output) {
1404 /* generate reg save code */
1405 for(i = 0; i < NB_SAVED_REGS; i++) {
1406 reg = reg_saved[i];
1407 if (regs_allocated[reg]) {
1408 #ifdef I386_ASM_16
1409 if (tcc_state->seg_size == 16)
1410 g(0x66);
1411 #endif
1412 g(0x50 + reg);
1413 }
1414 }
1415
1416 /* generate load code */
1417 for(i = 0; i < nb_operands; i++) {
1418 op = &operands[i];
1419 if (op->reg >= 0) {
1420 if ((op->vt->r & VT_VALMASK) == VT_LLOCAL &&
1421 op->is_memory) {
1422 /* memory reference case (for both input and
1423 output cases) */
1424 SValue sv;
1425 sv = *op->vt;
1426 sv.r = (sv.r & ~VT_VALMASK) | VT_LOCAL;
1427 load(op->reg, &sv);
1428 } else if (i >= nb_outputs || op->is_rw) {
1429 /* load value in register */
1430 load(op->reg, op->vt);
1431 if (op->is_llong) {
1432 SValue sv;
1433 sv = *op->vt;
1434 sv.c.ul += 4;
1435 load(TREG_XDX, &sv);
1436 }
1437 }
1438 }
1439 }
1440 } else {
1441 /* generate save code */
1442 for(i = 0 ; i < nb_outputs; i++) {
1443 op = &operands[i];
1444 if (op->reg >= 0) {
1445 if ((op->vt->r & VT_VALMASK) == VT_LLOCAL) {
1446 if (!op->is_memory) {
1447 SValue sv;
1448 sv = *op->vt;
1449 sv.r = (sv.r & ~VT_VALMASK) | VT_LOCAL;
1450 load(out_reg, &sv);
1451
1452 sv.r = (sv.r & ~VT_VALMASK) | out_reg;
1453 store(op->reg, &sv);
1454 }
1455 } else {
1456 store(op->reg, op->vt);
1457 if (op->is_llong) {
1458 SValue sv;
1459 sv = *op->vt;
1460 sv.c.ul += 4;
1461 store(TREG_XDX, &sv);
1462 }
1463 }
1464 }
1465 }
1466 /* generate reg restore code */
1467 for(i = NB_SAVED_REGS - 1; i >= 0; i--) {
1468 reg = reg_saved[i];
1469 if (regs_allocated[reg]) {
1470 #ifdef I386_ASM_16
1471 if (tcc_state->seg_size == 16)
1472 g(0x66);
1473 #endif
1474 g(0x58 + reg);
1475 }
1476 }
1477 }
1478 }
1479
1480 ST_FUNC void asm_clobber(uint8_t *clobber_regs, const char *str)
1481 {
1482 int reg;
1483 TokenSym *ts;
1484
1485 if (!strcmp(str, "memory") ||
1486 !strcmp(str, "cc"))
1487 return;
1488 ts = tok_alloc(str, strlen(str));
1489 reg = ts->tok;
1490 if (reg >= TOK_ASM_eax && reg <= TOK_ASM_edi) {
1491 reg -= TOK_ASM_eax;
1492 } else if (reg >= TOK_ASM_ax && reg <= TOK_ASM_di) {
1493 reg -= TOK_ASM_ax;
1494 #ifdef TCC_TARGET_X86_64
1495 } else if (reg >= TOK_ASM_rax && reg <= TOK_ASM_rdi) {
1496 reg -= TOK_ASM_rax;
1497 #endif
1498 } else {
1499 tcc_error("invalid clobber register '%s'", str);
1500 }
1501 clobber_regs[reg] = 1;
1502 }