int: do not store output values in ic struct
[neatcc.git] / x86.c
blobb57845a891629b8fb08d1f41d9befd7e43d5f89b
1 /* architecture-dependent code generation for x86 */
2 #include <stdlib.h>
3 #include "ncc.h"
5 /* x86-64 registers, without r8-r15 */
6 #define R_RAX 0x00
7 #define R_RCX 0x01
8 #define R_RDX 0x02
9 #define R_RBX 0x03
10 #define R_RSP 0x04
11 #define R_RBP 0x05
12 #define R_RSI 0x06
13 #define R_RDI 0x07
15 #define REG_RET R_RAX
16 #define R_BYTE 0x0007
18 /* x86 opcodes */
19 #define I_MOV 0x89
20 #define I_MOVI 0xc7
21 #define I_MOVIR 0xb8
22 #define I_MOVR 0x8b
23 #define I_MOVSXD 0x63
24 #define I_SHX 0xd3
25 #define I_CMP 0x3b
26 #define I_TST 0x85
27 #define I_LEA 0x8d
28 #define I_NOT 0xf7
29 #define I_CALL 0xff
30 #define I_MUL 0xf7
31 #define I_XOR 0x33
32 #define I_CQO 0x99
33 #define I_PUSH 0x50
34 #define I_POP 0x58
36 #define MIN(a, b) ((a) < (b) ? (a) : (b))
37 #define ALIGN(x, a) (((x) + (a) - 1) & ~((a) - 1))
39 int tmpregs[] = {0, 1, 2, 6, 7, 3};
40 int argregs[] = {0};
42 #define OP2(o2, o1) (0x010000 | ((o2) << 8) | (o1))
43 #define O2(op) (((op) >> 8) & 0xff)
44 #define O1(op) ((op) & 0xff)
45 #define MODRM(m, r1, r2) ((m) << 6 | (r1) << 3 | (r2))
47 static struct mem cs; /* generated code */
49 /* code generation functions */
50 static void os(void *s, int n)
52 mem_put(&cs, s, n);
55 static char *ointbuf(long n, int l)
57 static char buf[16];
58 int i;
59 for (i = 0; i < l; i++) {
60 buf[i] = n & 0xff;
61 n >>= 8;
63 return buf;
66 static void oi(long n, int l)
68 mem_put(&cs, ointbuf(n, l), l);
71 static void oi_at(long pos, long n, int l)
73 mem_cpy(&cs, pos, ointbuf(n, l), l);
76 static long opos(void)
78 return mem_len(&cs);
81 static void op_x(int op, int r1, int r2, int bt)
83 int sz = T_SZ(bt);
84 if (sz == 2)
85 oi(0x66, 1);
86 if (op & 0x10000)
87 oi(O2(op), 1);
88 oi(sz == 1 ? O1(op) & ~0x1 : O1(op), 1);
91 #define op_mr op_rm
93 /* op_*(): r=reg, m=mem, i=imm, s=sym */
94 static void op_rm(int op, int src, int base, int off, int bt)
96 int dis = off == (char) off ? 1 : 4;
97 int mod = dis == 4 ? 2 : 1;
98 if (!off && (base & 7) != R_RBP)
99 mod = 0;
100 op_x(op, src, base, bt);
101 oi(MODRM(mod, src & 0x07, base & 0x07), 1);
102 if ((base & 7) == R_RSP)
103 oi(0x24, 1);
104 if (mod)
105 oi(off, dis);
108 static void op_rr(int op, int src, int dst, int bt)
110 op_x(op, src, dst, bt);
111 oi(MODRM(3, src & 0x07, dst & 0x07), 1);
114 #define movrx_bt(bt) (LONGSZ)
116 static int movrx_op(int bt, int mov)
118 int sz = T_SZ(bt);
119 if (sz == 2)
120 return OP2(0x0f, bt & T_MSIGN ? 0xbf : 0xb7);
121 if (sz == 1)
122 return OP2(0x0f, bt & T_MSIGN ? 0xbe : 0xb6);
123 return mov;
126 static void mov_r2r(int rd, int r1, unsigned bt)
128 if (rd != r1 || T_SZ(bt) != LONGSZ)
129 op_rr(movrx_op(bt, I_MOVR), rd, r1, movrx_bt(bt));
132 static void i_mov(int rd, int rn)
134 op_rr(movrx_op(LONGSZ, I_MOVR), rd, rn, movrx_bt(LONGSZ));
137 static void i_add(int op, int rd, int r1, int r2)
139 /* opcode for O_ADD, O_SUB, O_AND, O_OR, O_XOR */
140 static int rx[] = {0003, 0053, 0043, 0013, 0063};
141 op_rr(rx[op & 0x0f], rd, r2, LONGSZ);
144 static void i_add_imm(int op, int rd, int rn, long n)
146 /* opcode for O_ADD, O_SUB, O_AND, O_OR, O_XOR */
147 static int rx[] = {0xc0, 0xe8, 0xe0, 0xc8, 0xf0};
148 unsigned char s[4] = {0x83, rx[op & 0x0f] | rd, n & 0xff};
149 os((void *) s, 3);
152 static void i_num(int rd, long n)
154 if (!n) {
155 op_rr(I_XOR, rd, rd, 4);
156 return;
157 } else {
158 op_x(I_MOVIR + (rd & 7), 0, rd, LONGSZ);
159 oi(n, LONGSZ);
163 static void i_mul(int rd, int r1, int r2)
165 if (r2 != R_RDX)
166 i_num(R_RDX, 0);
167 op_rr(I_MUL, 4, r2, LONGSZ);
170 static void i_div(int op, int rd, int r1, int r2)
172 long bt = O_T(op);
173 if (r2 != R_RDX) {
174 if (bt & T_MSIGN)
175 op_x(I_CQO, R_RAX, R_RDX, LONGSZ);
176 else
177 i_num(R_RDX, 0);
179 op_rr(I_MUL, bt & T_MSIGN ? 7 : 6, r2, LONGSZ);
182 static void i_tst(int rn, int rm)
184 op_rr(I_TST, rn, rm, LONGSZ);
187 static void i_cmp(int rn, int rm)
189 op_rr(I_CMP, rn, rm, LONGSZ);
192 static void i_cmp_imm(int rn, long n)
194 unsigned char s[4] = {0x83, 0xf8 | rn, n & 0xff};
195 os(s, 3);
198 static void i_shl(int op, int rd, int r1, int rs)
200 long bt = O_T(op);
201 int sm = 4;
202 if ((op & 0x0f) == 1)
203 sm = bt & T_MSIGN ? 7 : 5;
204 op_rr(I_SHX, sm, rd, LONGSZ);
207 static void i_shl_imm(int op, int rd, int rn, long n)
209 long bt = O_T(op);
210 int sm = (op & 0x1) ? (bt & T_MSIGN ? 0xf8 : 0xe8) : 0xe0;
211 char s[4] = {0xc1, sm | rn, n & 0xff};
212 os(s, 3);
215 static void i_neg(int rd)
217 op_rr(I_NOT, 3, rd, LONGSZ);
220 static void i_not(int rd)
222 op_rr(I_NOT, 2, rd, LONGSZ);
225 static int i_cond(long op)
227 /* lt, ge, eq, ne, le, gt */
228 static int ucond[] = {0x92, 0x93, 0x94, 0x95, 0x96, 0x97};
229 static int scond[] = {0x9c, 0x9d, 0x94, 0x95, 0x9e, 0x9f};
230 long bt = O_T(op);
231 return bt & T_MSIGN ? scond[op & 0x0f] : ucond[op & 0x0f];
234 static void i_set(long op, int rd)
236 char set[] = "\x0f\x00\xc0";
237 set[1] = i_cond(op);
238 os(set, 3); /* setl al */
239 os("\x0f\xb6\xc0", 3); /* movzx rax, al */
242 static void i_lnot(int rd)
244 char cmp[] = "\x83\xf8\x00";
245 cmp[1] |= rd;
246 os(cmp, 3); /* cmp rax, 0 */
247 i_set(O_EQ, rd);
250 static void jx(int x, int nbytes)
252 char op[2] = {0x0f};
253 if (nbytes == 1) {
254 oi(0x70 | (x & 0x0f), 1); /* jx $addr */
255 } else {
256 op[1] = x;
257 os(op, 2); /* jx $addr */
261 static long i_jmp(long op, long rn, long rm, int nbytes)
263 long ret;
264 if (nbytes > 1)
265 nbytes = 4;
266 if (op & (O_JZ | O_JCC)) {
267 if (op & O_JZ) {
268 i_tst(rn, rn);
269 jx(O_C(op) == O_JZ ? 0x84 : 0x85, nbytes);
270 } else {
271 if (op & O_NUM)
272 i_cmp_imm(rn, rm);
273 else
274 i_cmp(rn, rm);
275 jx(i_cond(op) & ~0x10, nbytes);
277 } else {
278 os(nbytes == 1 ? "\xeb" : "\xe9", 1); /* jmp $addr */
280 ret = opos();
281 oi(0, nbytes);
282 return ret;
285 static void i_zx(int rd, int r1, int bits)
287 if (bits & 0x07) {
288 i_shl_imm(O_SHL, rd, rd, LONGSZ * 8 - bits);
289 i_shl_imm(O_SHR, rd, rd, LONGSZ * 8 - bits);
290 } else {
291 mov_r2r(rd, r1, bits >> 3);
295 static void i_sx(int rd, int r1, int bits)
297 mov_r2r(rd, r1, T_MSIGN | (bits >> 3));
300 static void i_cast(int rd, int rn, int bt)
302 if (T_SZ(bt) == 8) {
303 if (rd != rn)
304 i_mov(rd, rn);
305 } else {
306 if (bt & T_MSIGN)
307 i_sx(rd, rn, T_SZ(bt) * 8);
308 else
309 i_zx(rd, rn, T_SZ(bt) * 8);
313 static void i_add_anyimm(int rd, int rn, long n)
315 op_rm(I_LEA, rd, rn, n, LONGSZ);
318 static long *rel_sym; /* relocation symbols */
319 static long *rel_flg; /* relocation flags */
320 static long *rel_off; /* relocation offsets */
321 static long rel_n, rel_sz; /* relocation count */
323 static long lab_sz; /* label count */
324 static long *lab_loc; /* label offsets in cs */
325 static long jmp_n, jmp_sz; /* jump count */
326 static long *jmp_off; /* jump offsets */
327 static long *jmp_dst; /* jump destinations */
328 static long jmp_ret; /* the position of the last return jmp */
330 static void lab_add(long id)
332 while (id >= lab_sz) {
333 int lab_n = lab_sz;
334 lab_sz = MAX(128, lab_sz * 2);
335 lab_loc = mextend(lab_loc, lab_n, lab_sz, sizeof(*lab_loc));
337 lab_loc[id] = opos();
340 static void jmp_add(long off, long dst)
342 if (jmp_n == jmp_sz) {
343 jmp_sz = MAX(128, jmp_sz * 2);
344 jmp_off = mextend(jmp_off, jmp_n, jmp_sz, sizeof(*jmp_off));
345 jmp_dst = mextend(jmp_dst, jmp_n, jmp_sz, sizeof(*jmp_dst));
347 jmp_off[jmp_n] = off;
348 jmp_dst[jmp_n] = dst;
349 jmp_n++;
352 void i_label(long id)
354 lab_add(id + 1);
357 static void i_rel(long sym, long flg, long off)
359 if (rel_n == rel_sz) {
360 rel_sz = MAX(128, rel_sz * 2);
361 rel_sym = mextend(rel_sym, rel_n, rel_sz, sizeof(*rel_sym));
362 rel_flg = mextend(rel_flg, rel_n, rel_sz, sizeof(*rel_flg));
363 rel_off = mextend(rel_off, rel_n, rel_sz, sizeof(*rel_off));
365 rel_sym[rel_n] = sym;
366 rel_flg[rel_n] = flg;
367 rel_off[rel_n] = off;
368 rel_n++;
371 static void i_sym(int rd, int sym, int off)
373 op_x(I_MOVIR + (rd & 7), 0, rd, LONGSZ);
374 i_rel(sym, OUT_CS, opos());
375 oi(off, LONGSZ);
378 static void i_saveregs(long sregs, long sregs_pos, int st)
380 int nsregs = 0;
381 int i;
382 for (i = 0; i < N_TMPS; i++)
383 if ((1 << tmpregs[i]) & sregs)
384 op_rm(st ? I_MOV : I_MOVR, tmpregs[i], REG_FP,
385 sregs_pos + nsregs++ * ULNG, ULNG);
388 void i_wrap(int argc, long sargs, long spsub, int initfp, long sregs, long sregs_pos)
390 long body_n;
391 void *body;
392 long diff; /* prologue length */
393 int i;
394 /* removing the last jmp to the epilogue */
395 if (jmp_ret + 5 == opos()) {
396 mem_cut(&cs, jmp_ret);
397 jmp_n--;
399 lab_add(0); /* the return label */
400 body_n = mem_len(&cs);
401 body = mem_get(&cs);
402 /* generating function prologue */
403 if (initfp) {
404 os("\x55", 1); /* push rbp */
405 os("\x89\xe5", 2); /* mov rbp, rsp */
407 if (spsub) {
408 os("\x81\xec", 2);
409 spsub = ALIGN(spsub, 8);
410 oi(spsub, 4);
412 i_saveregs(sregs, sregs_pos, 1); /* saving registers */
413 diff = mem_len(&cs);
414 mem_put(&cs, body, body_n);
415 free(body);
416 /* generating function epilogue */
417 i_saveregs(sregs, sregs_pos, 0); /* restoring saved registers */
418 if (initfp)
419 os("\xc9", 1); /* leave */
420 os("\xc3", 1); /* ret */
421 /* adjusting code offsets */
422 for (i = 0; i < rel_n; i++)
423 rel_off[i] += diff;
424 for (i = 0; i < jmp_n; i++)
425 jmp_off[i] += diff;
426 for (i = 0; i < lab_sz; i++)
427 lab_loc[i] += diff;
430 void i_code(char **c, long *c_len, long **rsym, long **rflg, long **roff, long *rcnt)
432 int i;
433 for (i = 0; i < jmp_n; i++) /* filling jmp destinations */
434 oi_at(jmp_off[i], lab_loc[jmp_dst[i]] - jmp_off[i] - 4, 4);
435 *c_len = mem_len(&cs);
436 *c = mem_get(&cs);
437 *rsym = rel_sym;
438 *rflg = rel_flg;
439 *roff = rel_off;
440 *rcnt = rel_n;
441 rel_sym = NULL;
442 rel_flg = NULL;
443 rel_off = NULL;
444 rel_n = 0;
445 rel_sz = 0;
446 jmp_n = 0;
449 void i_done(void)
451 free(jmp_off);
452 free(jmp_dst);
453 free(lab_loc);
456 long i_reg(long op, long *rd, long *r1, long *r2, long *r3, long *tmp)
458 long oc = O_C(op);
459 long bt = O_T(op);
460 *rd = 0;
461 *r1 = 0;
462 *r2 = 0;
463 *r3 = 0;
464 *tmp = 0;
465 if (oc & O_MOV) {
466 *rd = R_TMPS;
467 if (oc & (O_NUM | O_SYM))
468 *r1 = oc & (O_NUM | O_SYM) ? LONGSZ * 8 : R_TMPS;
469 else
470 *r1 = T_SZ(bt) == 1 ? R_BYTE : R_TMPS;
471 return 0;
473 if (oc & O_ADD) {
474 *r1 = R_TMPS;
475 *r2 = oc & O_NUM ? (oc == O_ADD ? 32 : 8) : R_TMPS;
476 return 0;
478 if (oc & O_SHL) {
479 if (oc & O_NUM) {
480 *r1 = R_TMPS;
481 *r2 = 8;
482 } else {
483 *r2 = 1 << R_RCX;
484 *r1 = R_TMPS & ~*r2;
486 return 0;
488 if (oc & O_MUL) {
489 if (oc & O_NUM)
490 return 1;
491 *rd = oc == O_MOD ? (1 << R_RDX) : (1 << R_RAX);
492 *r1 = (1 << R_RAX);
493 *r2 = R_TMPS & ~*rd & ~*r1;
494 if (oc == O_DIV)
495 *r2 &= ~(1 << R_RDX);
496 *tmp = (1 << R_RDX) | (1 << R_RAX);
497 return 0;
499 if (oc & O_CMP) {
500 *rd = 1 << R_RAX;
501 *r1 = R_TMPS;
502 *r2 = oc & O_NUM ? 8 : R_TMPS;
503 return 0;
505 if (oc & O_UOP) {
506 if (oc == O_LNOT)
507 *r1 = 1 << R_RAX;
508 else
509 *r1 = R_TMPS;
510 return 0;
512 if (oc == O_MSET) {
513 *r1 = 1 << R_RDI;
514 *r2 = 1 << R_RAX;
515 *r3 = 1 << R_RCX;
516 *tmp = (1 << R_RDI) | (1 << R_RCX);
517 return 0;
519 if (oc == O_MCPY) {
520 *r1 = 1 << R_RDI;
521 *r2 = 1 << R_RSI;
522 *r3 = 1 << R_RCX;
523 *tmp = (1 << R_RDI) | (1 << R_RSI) | (1 << R_RCX);
524 return 0;
526 if (oc == O_RET) {
527 *r1 = (1 << REG_RET);
528 return 0;
530 if (oc & O_CALL) {
531 *rd = (1 << REG_RET);
532 *r1 = oc & O_SYM ? 0 : R_TMPS;
533 *tmp = R_TMPS & ~R_PERM;
534 return 0;
536 if (oc & O_LD) {
537 *rd = T_SZ(bt) == 1 ? R_BYTE : R_TMPS;
538 *r1 = R_TMPS;
539 *r2 = oc & O_NUM ? 0 : R_TMPS;
540 return 0;
542 if (oc & O_ST) {
543 *r1 = T_SZ(bt) == 1 ? R_BYTE : R_TMPS;
544 *r2 = R_TMPS;
545 *r3 = oc & O_NUM ? 0 : R_TMPS;
546 return 0;
548 if (oc & O_JZ) {
549 *r1 = R_TMPS;
550 return 0;
552 if (oc & O_JCC) {
553 *r1 = R_TMPS;
554 *r2 = oc & O_NUM ? 8 : R_TMPS;
555 return 0;
557 if (oc == O_JMP)
558 return 0;
559 return 1;
562 int i_imm(long lim, long n)
564 long max = (1 << (lim - 1)) - 1;
565 return n <= max && n + 1 >= -max;
568 long i_ins(long op, long rd, long r1, long r2, long r3)
570 long oc = O_C(op);
571 long bt = O_T(op);
572 if (oc & O_ADD) {
573 if (oc & O_NUM) {
574 if (rd == r1 && r2 <= 127 && r2 >= -128)
575 i_add_imm(op, r1, r1, r2);
576 else
577 i_add_anyimm(rd, r1, r2);
578 } else {
579 i_add(op, r1, r1, r2);
582 if (oc & O_SHL) {
583 if (oc & O_NUM)
584 i_shl_imm(op, r1, r1, r2);
585 else
586 i_shl(op, r1, r1, r2);
588 if (oc & O_MUL) {
589 if (oc == O_MUL)
590 i_mul(R_RAX, r1, r2);
591 if (oc == O_DIV)
592 i_div(op, R_RAX, r1, r2);
593 if (oc == O_MOD)
594 i_div(op, R_RDX, r1, r2);
595 return 0;
597 if (oc & O_CMP) {
598 if (oc & O_NUM)
599 i_cmp_imm(r1, r2);
600 else
601 i_cmp(r1, r2);
602 i_set(op, rd);
603 return 0;
605 if (oc & O_UOP) { /* uop */
606 if (oc == O_NEG)
607 i_neg(r1);
608 if (oc == O_NOT)
609 i_not(r1);
610 if (oc == O_LNOT)
611 i_lnot(r1);
612 return 0;
614 if (oc == O_CALL) {
615 op_rr(I_CALL, 2, r1, LONGSZ);
616 return 0;
618 if (oc == (O_CALL | O_SYM)) {
619 os("\xe8", 1); /* call $x */
620 i_rel(r1, OUT_CS | OUT_RLREL, opos());
621 oi(-4, 4);
622 return 0;
624 if (oc == (O_MOV | O_SYM)) {
625 i_sym(rd, r1, r2);
626 return 0;
628 if (oc == (O_MOV | O_NUM)) {
629 i_num(rd, r1);
630 return 0;
632 if (oc == O_MSET) {
633 os("\xfc\xf3\xaa", 3); /* cld; rep stosb */
634 return 0;
636 if (oc == O_MCPY) {
637 os("\xfc\xf3\xa4", 3); /* cld; rep movs */
638 return 0;
640 if (oc == O_RET) {
641 jmp_ret = opos();
642 jmp_add(i_jmp(O_JMP, 0, 0, 4), 0);
643 return 0;
645 if (oc == (O_LD | O_NUM)) {
646 op_rm(movrx_op(bt, I_MOVR), rd, r1, r2, movrx_bt(bt));
647 return 0;
649 if (oc == (O_ST | O_NUM)) {
650 op_rm(I_MOV, r1, r2, r3, bt);
651 return 0;
653 if (oc == O_MOV) {
654 i_cast(rd, r1, bt);
655 return 0;
657 if (oc & O_JXX) {
658 jmp_add(i_jmp(op, r1, r2, 4), r3 + 1);
659 return 0;
661 return 1;