ncc: new intermediate code
[neatcc.git] / x86.c
blob0d838ea6196532c99ad2320521bda1b4ef940889
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 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 void oi(long n, int l)
68 mem_put(&cs, ointbuf(n, l), l);
71 void oi_at(long pos, long n, int l)
73 mem_cpy(&cs, pos, ointbuf(n, l), l);
76 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_push(int reg)
134 op_x(I_PUSH | (reg & 0x7), 0, reg, LONGSZ);
137 static void i_pop(int reg)
139 op_x(I_POP | (reg & 0x7), 0, reg, LONGSZ);
142 void i_mov(int rd, int rn)
144 op_rr(movrx_op(LONGSZ, I_MOVR), rd, rn, movrx_bt(LONGSZ));
147 static void i_add(int op, int rd, int r1, int r2)
149 /* opcode for O_ADD, O_SUB, O_AND, O_OR, O_XOR */
150 static int rx[] = {0003, 0053, 0043, 0013, 0063};
151 op_rr(rx[op & 0x0f], rd, r2, LONGSZ);
154 static void i_add_imm(int op, int rd, int rn, long n)
156 /* opcode for O_ADD, O_SUB, O_AND, O_OR, O_XOR */
157 static int rx[] = {0xc0, 0xe8, 0xe0, 0xc8, 0xf0};
158 unsigned char s[4] = {0x83, rx[op & 0x0f] | rd, n & 0xff};
159 os((void *) s, 3);
162 static void i_num(int rd, long n)
164 if (!n) {
165 op_rr(I_XOR, rd, rd, 4);
166 return;
167 } else {
168 op_x(I_MOVIR + (rd & 7), 0, rd, LONGSZ);
169 oi(n, LONGSZ);
173 static void i_mul(int rd, int r1, int r2)
175 if (r2 != R_RDX)
176 i_num(R_RDX, 0);
177 op_rr(I_MUL, 4, r2, LONGSZ);
180 static void i_div(int op, int rd, int r1, int r2)
182 long bt = O_T(op);
183 if (r2 != R_RDX) {
184 if (bt & T_MSIGN)
185 op_x(I_CQO, R_RAX, R_RDX, LONGSZ);
186 else
187 i_num(R_RDX, 0);
189 op_rr(I_MUL, bt & T_MSIGN ? 7 : 6, r2, LONGSZ);
192 static void i_tst(int rn, int rm)
194 op_rr(I_TST, rn, rm, LONGSZ);
197 static void i_cmp(int rn, int rm)
199 op_rr(I_CMP, rn, rm, LONGSZ);
202 static void i_cmp_imm(int rn, long n)
204 unsigned char s[4] = {0x83, 0xf8 | rn, n & 0xff};
205 os(s, 3);
208 static void i_shl(int op, int rd, int r1, int rs)
210 long bt = O_T(op);
211 int sm = 4;
212 if ((op & 0x0f) == 1)
213 sm = bt & T_MSIGN ? 7 : 5;
214 op_rr(I_SHX, sm, rd, LONGSZ);
217 static void i_shl_imm(int op, int rd, int rn, long n)
219 long bt = O_T(op);
220 int sm = (op & 0x1) ? (bt & T_MSIGN ? 0xf8 : 0xe8) : 0xe0;
221 char s[4] = {0xc1, sm | rn, n & 0xff};
222 os(s, 3);
225 static void i_neg(int rd)
227 op_rr(I_NOT, 3, rd, LONGSZ);
230 static void i_not(int rd)
232 op_rr(I_NOT, 2, rd, LONGSZ);
235 static int i_cond(long op)
237 /* lt, ge, eq, ne, le, gt */
238 static int ucond[] = {0x92, 0x93, 0x94, 0x95, 0x96, 0x97};
239 static int scond[] = {0x9c, 0x9d, 0x94, 0x95, 0x9e, 0x9f};
240 long bt = O_T(op);
241 return bt & T_MSIGN ? scond[op & 0x0f] : ucond[op & 0x0f];
244 static void i_set(long op, int rd)
246 char set[] = "\x0f\x00\xc0";
247 set[1] = i_cond(op);
248 os(set, 3); /* setl al */
249 os("\x0f\xb6\xc0", 3); /* movzx rax, al */
252 static void i_lnot(int rd)
254 char cmp[] = "\x83\xf8\x00";
255 cmp[1] |= rd;
256 os(cmp, 3); /* cmp rax, 0 */
257 i_set(O_EQ, rd);
260 static void jx(int x, int nbytes)
262 char op[2] = {0x0f};
263 if (nbytes == 1) {
264 oi(0x70 | (x & 0x0f), 1); /* jx $addr */
265 } else {
266 op[1] = x;
267 os(op, 2); /* jx $addr */
271 static long i_jmp(long op, long rn, long rm, int nbytes)
273 long ret;
274 if (nbytes > 1)
275 nbytes = 4;
276 if (op & (O_JZ | O_JCC)) {
277 if (op & O_JZ) {
278 i_tst(rn, rn);
279 jx(O_C(op) == O_JZ ? 0x84 : 0x85, nbytes);
280 } else {
281 if (op & O_NUM)
282 i_cmp_imm(rn, rm);
283 else
284 i_cmp(rn, rm);
285 jx(i_cond(op) & ~0x10, nbytes);
287 } else {
288 os(nbytes == 1 ? "\xeb" : "\xe9", 1); /* jmp $addr */
290 ret = opos();
291 oi(0, nbytes);
292 return ret;
295 void i_fill(long src, long dst, long nbytes)
297 if (nbytes > 1)
298 nbytes = 4;
299 oi_at(src, dst - src - nbytes, nbytes);
302 static void i_zx(int rd, int r1, int bits)
304 if (bits & 0x07) {
305 i_shl_imm(O_SHL, rd, rd, LONGSZ * 8 - bits);
306 i_shl_imm(O_SHR, rd, rd, LONGSZ * 8 - bits);
307 } else {
308 mov_r2r(rd, r1, bits >> 3);
312 static void i_sx(int rd, int r1, int bits)
314 mov_r2r(rd, r1, T_MSIGN | (bits >> 3));
317 static void i_cast(int rd, int rn, int bt)
319 if (T_SZ(bt) == 8) {
320 if (rd != rn)
321 i_mov(rd, rn);
322 } else {
323 if (bt & T_MSIGN)
324 i_sx(rd, rn, T_SZ(bt) * 8);
325 else
326 i_zx(rd, rn, T_SZ(bt) * 8);
330 static void i_add_anyimm(int rd, int rn, long n)
332 op_rm(I_LEA, rd, rn, n, LONGSZ);
335 static long *rel_sym; /* relocation symbols */
336 static long *rel_flg; /* relocation flags */
337 static long *rel_off; /* relocation offsets */
338 static long rel_n, rel_sz; /* relocation count */
340 static long lab_sz; /* label count */
341 static long *lab_loc; /* label offsets in cs */
342 static long jmp_n, jmp_sz; /* jump count */
343 static long *jmp_off; /* jump offsets */
344 static long *jmp_dst; /* jump destinations */
345 static long jmp_ret; /* the position of the last return jmp */
347 static void lab_add(long id)
349 while (id >= lab_sz) {
350 int lab_n = lab_sz;
351 lab_sz = MAX(128, lab_sz * 2);
352 lab_loc = mextend(lab_loc, lab_n, lab_sz, sizeof(*lab_loc));
354 lab_loc[id] = opos();
357 static void jmp_add(long off, long dst)
359 if (jmp_n == jmp_sz) {
360 jmp_sz = MAX(128, jmp_sz * 2);
361 jmp_off = mextend(jmp_off, jmp_n, jmp_sz, sizeof(*jmp_off));
362 jmp_dst = mextend(jmp_dst, jmp_n, jmp_sz, sizeof(*jmp_dst));
364 jmp_off[jmp_n] = off;
365 jmp_dst[jmp_n] = dst;
366 jmp_n++;
369 void i_label(long id)
371 lab_add(id + 1);
374 static void i_rel(long sym, long flg, long off)
376 if (rel_n == rel_sz) {
377 rel_sz = MAX(128, rel_sz * 2);
378 rel_sym = mextend(rel_sym, rel_n, rel_sz, sizeof(*rel_sym));
379 rel_flg = mextend(rel_flg, rel_n, rel_sz, sizeof(*rel_flg));
380 rel_off = mextend(rel_off, rel_n, rel_sz, sizeof(*rel_off));
382 rel_sym[rel_n] = sym;
383 rel_flg[rel_n] = flg;
384 rel_off[rel_n] = off;
385 rel_n++;
388 static void i_sym(int rd, int sym, int off)
390 op_x(I_MOVIR + (rd & 7), 0, rd, LONGSZ);
391 i_rel(sym, OUT_CS, opos());
392 oi(off, LONGSZ);
395 static void i_saveregs(long sregs, long sregs_pos, int st)
397 int nsregs = 0;
398 int i;
399 for (i = 0; i < N_TMPS; i++)
400 if ((1 << tmpregs[i]) & sregs)
401 op_rm(st ? I_MOV : I_MOVR, tmpregs[i], REG_FP,
402 sregs_pos + nsregs++ * ULNG, ULNG);
405 void i_wrap(int argc, long sargs, long spsub, int initfp, long sregs, long sregs_pos)
407 long body_n;
408 void *body;
409 long diff; /* prologue length */
410 int i;
411 /* removing the last jmp to the epilogue */
412 if (jmp_ret + 5 == opos()) {
413 mem_cut(&cs, jmp_ret);
414 jmp_n--;
416 lab_add(0); /* the return label */
417 body_n = mem_len(&cs);
418 body = mem_get(&cs);
419 /* generating function prologue */
420 if (initfp) {
421 os("\x55", 1); /* push rbp */
422 os("\x89\xe5", 2); /* mov rbp, rsp */
424 if (spsub) {
425 os("\x81\xec", 2);
426 spsub = ALIGN(spsub, 8);
427 oi(spsub, 4);
429 i_saveregs(sregs, sregs_pos, 1); /* saving registers */
430 diff = mem_len(&cs);
431 mem_put(&cs, body, body_n);
432 free(body);
433 /* generating function epilogue */
434 i_saveregs(sregs, sregs_pos, 0); /* restoring saved registers */
435 if (initfp)
436 os("\xc9", 1); /* leave */
437 os("\xc3", 1); /* ret */
438 /* adjusting code offsets */
439 for (i = 0; i < rel_n; i++)
440 rel_off[i] += diff;
441 for (i = 0; i < jmp_n; i++)
442 jmp_off[i] += diff;
443 for (i = 0; i < lab_sz; i++)
444 lab_loc[i] += diff;
447 void i_code(char **c, long *c_len, long **rsym, long **rflg, long **roff, long *rcnt)
449 int i;
450 for (i = 0; i < jmp_n; i++) /* filling jmp destinations */
451 oi_at(jmp_off[i], lab_loc[jmp_dst[i]] - jmp_off[i] - 4, 4);
452 *c_len = mem_len(&cs);
453 *c = mem_get(&cs);
454 *rsym = rel_sym;
455 *rflg = rel_flg;
456 *roff = rel_off;
457 *rcnt = rel_n;
458 rel_sym = NULL;
459 rel_flg = NULL;
460 rel_off = NULL;
461 rel_n = 0;
462 rel_sz = 0;
463 jmp_n = 0;
466 void i_done(void)
468 free(jmp_off);
469 free(jmp_dst);
470 free(lab_loc);
473 long i_reg(long op, long *rd, long *r1, long *r2, long *tmp)
475 long oc = O_C(op);
476 long bt = O_T(op);
477 *rd = 0;
478 *r1 = 0;
479 *r2 = 0;
480 *tmp = 0;
481 if (oc & O_MOV) {
482 *rd = R_TMPS;
483 if (oc & (O_NUM | O_SYM))
484 *r1 = oc & (O_NUM | O_SYM) ? LONGSZ * 8 : R_TMPS;
485 else
486 *r1 = T_SZ(bt) == 1 ? R_BYTE : R_TMPS;
487 return 0;
489 if (oc & O_ADD) {
490 *r1 = R_TMPS;
491 *r2 = oc & O_NUM ? (oc == O_ADD ? 32 : 8) : R_TMPS;
492 return 0;
494 if (oc & O_SHL) {
495 if (oc & O_NUM) {
496 *r1 = R_TMPS;
497 *r2 = 8;
498 } else {
499 *r2 = 1 << R_RCX;
500 *r1 = R_TMPS & ~*r2;
502 return 0;
504 if (oc & O_MUL) {
505 if (oc & O_NUM)
506 return 1;
507 *rd = oc == O_MOD ? (1 << R_RDX) : (1 << R_RAX);
508 *r1 = (1 << R_RAX);
509 *r2 = R_TMPS & ~*rd & ~*r1;
510 if (oc == O_DIV)
511 *r2 &= ~(1 << R_RDX);
512 *tmp = (1 << R_RDX) | (1 << R_RAX);
513 return 0;
515 if (oc & O_CMP) {
516 *rd = 1 << R_RAX;
517 *r1 = R_TMPS;
518 *r2 = oc & O_NUM ? 8 : R_TMPS;
519 return 0;
521 if (oc & O_UOP) {
522 *rd = R_TMPS;
523 if (oc == O_LNOT)
524 *r1 = 1 << R_RAX;
525 else
526 *r1 = R_TMPS;
527 return 0;
529 if (oc == O_MSET) {
530 *rd = 1 << R_RDI;
531 *r1 = 1 << R_RAX;
532 *r2 = 1 << R_RCX;
533 return 0;
535 if (oc == O_MCPY) {
536 *rd = 1 << R_RDI;
537 *r1 = 1 << R_RSI;
538 *r2 = 1 << R_RCX;
539 return 0;
541 if (oc == O_RET) {
542 *rd = (1 << REG_RET);
543 return 0;
545 if (oc & O_CALL) {
546 *rd = (1 << REG_RET);
547 *r1 = oc & O_SYM ? 0 : R_TMPS;
548 return 0;
550 if (oc & (O_LD | O_ST)) {
551 *rd = R_TMPS;
552 *r1 = R_TMPS;
553 *r2 = oc & O_NUM ? 0 : R_TMPS;
554 return 0;
556 if (oc & O_JZ) {
557 *rd = R_TMPS;
558 return 0;
560 if (oc & O_JCC) {
561 *rd = R_TMPS;
562 *r1 = oc & O_NUM ? 8 : R_TMPS;
563 return 0;
565 if (oc == O_JMP)
566 return 0;
567 return 1;
570 int i_imm(long lim, long n)
572 long max = (1 << (lim - 1)) - 1;
573 return n <= max && n + 1 >= -max;
576 long i_ins(long op, long r0, long r1, long r2)
578 long oc = O_C(op);
579 long bt = O_T(op);
580 if (oc & O_ADD) {
581 if (oc & O_NUM) {
582 if (r0 == r1 && r2 <= 127 && r2 >= -128)
583 i_add_imm(op, r1, r1, r2);
584 else
585 i_add_anyimm(r0, r1, r2);
586 } else {
587 i_add(op, r1, r1, r2);
590 if (oc & O_SHL) {
591 if (oc & O_NUM)
592 i_shl_imm(op, r1, r1, r2);
593 else
594 i_shl(op, r1, r1, r2);
596 if (oc & O_MUL) {
597 if (oc == O_MUL)
598 i_mul(R_RAX, r1, r2);
599 if (oc == O_DIV)
600 i_div(op, R_RAX, r1, r2);
601 if (oc == O_MOD)
602 i_div(op, R_RDX, r1, r2);
603 return 0;
605 if (oc & O_CMP) {
606 if (oc & O_NUM)
607 i_cmp_imm(r1, r2);
608 else
609 i_cmp(r1, r2);
610 i_set(op, r0);
611 return 0;
613 if (oc & O_UOP) { /* uop */
614 if (oc == O_NEG)
615 i_neg(r1);
616 if (oc == O_NOT)
617 i_not(r1);
618 if (oc == O_LNOT)
619 i_lnot(r1);
620 return 0;
622 if (oc == O_CALL) {
623 op_rr(I_CALL, 2, r1, LONGSZ);
624 return 0;
626 if (oc == (O_CALL | O_SYM)) {
627 os("\xe8", 1); /* call $x */
628 i_rel(r1, OUT_CS | OUT_RLREL, opos());
629 oi(-4, 4);
630 return 0;
632 if (oc == (O_MOV | O_SYM)) {
633 i_sym(r0, r1, r2);
634 return 0;
636 if (oc == (O_MOV | O_NUM)) {
637 i_num(r0, r1);
638 return 0;
640 if (oc == O_MSET) {
641 os("\xfc\xf3\xaa", 3); /* cld; rep stosb */
642 return 0;
644 if (oc == O_MCPY) {
645 os("\xfc\xf3\xa4", 3); /* cld; rep movs */
646 return 0;
648 if (oc == O_RET) {
649 jmp_ret = opos();
650 jmp_add(i_jmp(O_JMP, 0, 0, 4), 0);
651 return 0;
653 if (oc == (O_LD | O_NUM)) {
654 op_rm(movrx_op(bt, I_MOVR), r0, r1, r2, movrx_bt(bt));
655 return 0;
657 if (oc == (O_ST | O_NUM)) {
658 op_rm(I_MOV, r0, r1, r2, bt);
659 return 0;
661 if (oc == O_MOV) {
662 i_cast(r0, r1, bt);
663 return 0;
665 if (oc & O_JXX) {
666 jmp_add(i_jmp(op, r0, r1, 4), r2 + 1);
667 return 0;
669 return 1;