?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
align emitted code
[qbe.git] / arm64 / emit.c
blobffdc178d4207abd33e53371d35cad9b14e47d4ee
1 #include "all.h"
2
3 typedef struct E E;
4
5 struct E {
6 FILE *f;
7 Fn *fn;
8 uint64_t frame;
9 uint padding;
10 };
11
12 #define CMP(X) \
13 X(Cieq, "eq") \
14 X(Cine, "ne") \
15 X(Cisge, "ge") \
16 X(Cisgt, "gt") \
17 X(Cisle, "le") \
18 X(Cislt, "lt") \
19 X(Ciuge, "cs") \
20 X(Ciugt, "hi") \
21 X(Ciule, "ls") \
22 X(Ciult, "cc") \
23 X(NCmpI+Cfeq, "eq") \
24 X(NCmpI+Cfge, "ge") \
25 X(NCmpI+Cfgt, "gt") \
26 X(NCmpI+Cfle, "ls") \
27 X(NCmpI+Cflt, "mi") \
28 X(NCmpI+Cfne, "ne") \
29 X(NCmpI+Cfo, "vc") \
30 X(NCmpI+Cfuo, "vs")
31
32 enum {
33 Ki = -1, /* matches Kw and Kl */
34 Ka = -2, /* matches all classes */
35 };
36
37 static struct {
38 short op;
39 short cls;
40 char *fmt;
41 } omap[] = {
42 { Oadd, Ki, "add %=, %0, %1" },
43 { Oadd, Ka, "fadd %=, %0, %1" },
44 { Osub, Ki, "sub %=, %0, %1" },
45 { Osub, Ka, "fsub %=, %0, %1" },
46 { Oneg, Ki, "neg %=, %0" },
47 { Oneg, Ka, "fneg %=, %0" },
48 { Oand, Ki, "and %=, %0, %1" },
49 { Oor, Ki, "orr %=, %0, %1" },
50 { Oxor, Ki, "eor %=, %0, %1" },
51 { Osar, Ki, "asr %=, %0, %1" },
52 { Oshr, Ki, "lsr %=, %0, %1" },
53 { Oshl, Ki, "lsl %=, %0, %1" },
54 { Omul, Ki, "mul %=, %0, %1" },
55 { Omul, Ka, "fmul %=, %0, %1" },
56 { Odiv, Ki, "sdiv %=, %0, %1" },
57 { Odiv, Ka, "fdiv %=, %0, %1" },
58 { Oudiv, Ki, "udiv %=, %0, %1" },
59 { Orem, Ki, "sdiv %?, %0, %1\n\tmsub\t%=, %?, %1, %0" },
60 { Ourem, Ki, "udiv %?, %0, %1\n\tmsub\t%=, %?, %1, %0" },
61 { Ocopy, Ki, "mov %=, %0" },
62 { Ocopy, Ka, "fmov %=, %0" },
63 { Oswap, Ki, "mov %?, %0\n\tmov\t%0, %1\n\tmov\t%1, %?" },
64 { Oswap, Ka, "fmov %?, %0\n\tfmov\t%0, %1\n\tfmov\t%1, %?" },
65 { Ostoreb, Kw, "strb %W0, %M1" },
66 { Ostoreh, Kw, "strh %W0, %M1" },
67 { Ostorew, Kw, "str %W0, %M1" },
68 { Ostorel, Kw, "str %L0, %M1" },
69 { Ostores, Kw, "str %S0, %M1" },
70 { Ostored, Kw, "str %D0, %M1" },
71 { Oloadsb, Ki, "ldrsb %=, %M0" },
72 { Oloadub, Ki, "ldrb %W=, %M0" },
73 { Oloadsh, Ki, "ldrsh %=, %M0" },
74 { Oloaduh, Ki, "ldrh %W=, %M0" },
75 { Oloadsw, Kw, "ldr %=, %M0" },
76 { Oloadsw, Kl, "ldrsw %=, %M0" },
77 { Oloaduw, Ki, "ldr %W=, %M0" },
78 { Oload, Ka, "ldr %=, %M0" },
79 { Oextsb, Ki, "sxtb %=, %W0" },
80 { Oextub, Ki, "uxtb %W=, %W0" },
81 { Oextsh, Ki, "sxth %=, %W0" },
82 { Oextuh, Ki, "uxth %W=, %W0" },
83 { Oextsw, Ki, "sxtw %L=, %W0" },
84 { Oextuw, Ki, "mov %W=, %W0" },
85 { Oexts, Kd, "fcvt %=, %S0" },
86 { Otruncd, Ks, "fcvt %=, %D0" },
87 { Ocast, Kw, "fmov %=, %S0" },
88 { Ocast, Kl, "fmov %=, %D0" },
89 { Ocast, Ks, "fmov %=, %W0" },
90 { Ocast, Kd, "fmov %=, %L0" },
91 { Ostosi, Ka, "fcvtzs %=, %S0" },
92 { Ostoui, Ka, "fcvtzu %=, %S0" },
93 { Odtosi, Ka, "fcvtzs %=, %D0" },
94 { Odtoui, Ka, "fcvtzu %=, %D0" },
95 { Oswtof, Ka, "scvtf %=, %W0" },
96 { Ouwtof, Ka, "ucvtf %=, %W0" },
97 { Osltof, Ka, "scvtf %=, %L0" },
98 { Oultof, Ka, "ucvtf %=, %L0" },
99 { Ocall, Kw, "blr %L0" },
100
101 { Oacmp, Ki, "cmp %0, %1" },
102 { Oacmn, Ki, "cmn %0, %1" },
103 { Oafcmp, Ka, "fcmpe %0, %1" },
104
105 #define X(c, str) \
106 { Oflag+c, Ki, "cset %=, " str },
107 CMP(X)
108 #undef X
109 { NOp, 0, 0 }
110 };
111
112 static char *
113 rname(int r, int k)
114 {
115 static char buf[4];
116
117 if (r == SP) {
118 assert(k == Kl);
119 sprintf(buf, "sp");
120 }
121 else if (R0 <= r && r <= LR)
122 switch (k) {
123 default: die("invalid class");
124 case Kw: sprintf(buf, "w%d", r-R0); break;
125 case Kx:
126 case Kl: sprintf(buf, "x%d", r-R0); break;
127 }
128 else if (V0 <= r && r <= V30)
129 switch (k) {
130 default: die("invalid class");
131 case Ks: sprintf(buf, "s%d", r-V0); break;
132 case Kx:
133 case Kd: sprintf(buf, "d%d", r-V0); break;
134 }
135 else
136 die("invalid register");
137 return buf;
138 }
139
140 static uint64_t
141 slot(Ref r, E *e)
142 {
143 int s;
144
145 s = rsval(r);
146 if (s == -1)
147 return 16 + e->frame;
148 if (s < 0) {
149 if (e->fn->vararg && !T.apple)
150 return 16 + e->frame + 192 - (s+2);
151 else
152 return 16 + e->frame - (s+2);
153 } else
154 return 16 + e->padding + 4 * s;
155 }
156
157 static void
158 emitf(char *s, Ins *i, E *e)
159 {
160 Ref r;
161 int k, c;
162 Con *pc;
163 uint n, sp;
164
165 fputc('\t', e->f);
166
167 sp = 0;
168 for (;;) {
169 k = i->cls;
170 while ((c = *s++) != '%')
171 if (c == ' ' && !sp) {
172 fputc('\t', e->f);
173 sp = 1;
174 } else if ( !c) {
175 fputc('\n', e->f);
176 return;
177 } else
178 fputc(c, e->f);
179 Switch:
180 switch ((c = *s++)) {
181 default:
182 die("invalid escape");
183 case 'W':
184 k = Kw;
185 goto Switch;
186 case 'L':
187 k = Kl;
188 goto Switch;
189 case 'S':
190 k = Ks;
191 goto Switch;
192 case 'D':
193 k = Kd;
194 goto Switch;
195 case '?':
196 if (KBASE(k) == 0)
197 fputs(rname(R18, k), e->f);
198 else
199 fputs(k==Ks ? "s31" : "d31", e->f);
200 break;
201 case '=':
202 case '0':
203 r = c == '=' ? i->to : i->arg[0];
204 assert(isreg(r));
205 fputs(rname(r.val, k), e->f);
206 break;
207 case '1':
208 r = i->arg[1];
209 switch (rtype(r)) {
210 default:
211 die("invalid second argument");
212 case RTmp:
213 assert(isreg(r));
214 fputs(rname(r.val, k), e->f);
215 break;
216 case RCon:
217 pc = &e->fn->con[r.val];
218 n = pc->bits.i;
219 assert(pc->type == CBits);
220 if (n & 0xfff000)
221 fprintf(e->f, "#%u, lsl #12", n>>12);
222 else
223 fprintf(e->f, "#%u", n);
224 break;
225 }
226 break;
227 case 'M':
228 c = *s++;
229 assert(c == '0' || c == '1' || c == '=');
230 r = c == '=' ? i->to : i->arg[c - '0'];
231 switch (rtype(r)) {
232 default:
233 die("todo (arm emit): unhandled ref");
234 case RTmp:
235 assert(isreg(r));
236 fprintf(e->f, "[%s]", rname(r.val, Kl));
237 break;
238 case RSlot:
239 fprintf(e->f, "[x29, %"PRIu64"]", slot(r, e));
240 break;
241 }
242 break;
243 }
244 }
245 }
246
247 static void
248 loadaddr(Con *c, char *rn, E *e)
249 {
250 char *p, *l, *s;
251
252 switch (c->sym.type) {
253 default:
254 die("unreachable");
255 case SGlo:
256 if (T.apple)
257 s = "\tadrp\tR, S@pageO\n"
258 "\tadd\tR, R, S@pageoffO\n";
259 else
260 s = "\tadrp\tR, SO\n"
261 "\tadd\tR, R, #:lo12:SO\n";
262 break;
263 case SThr:
264 if (T.apple)
265 s = "\tadrp\tR, S@tlvppage\n"
266 "\tldr\tR, [R, S@tlvppageoff]\n";
267 else
268 s = "\tmrs\tR, tpidr_el0\n"
269 "\tadd\tR, R, #:tprel_hi12:SO, lsl #12\n"
270 "\tadd\tR, R, #:tprel_lo12_nc:SO\n";
271 break;
272 }
273
274 l = str(c->sym.id);
275 p = l[0] == '"' ? "" : T.assym;
276 for (; *s; s++)
277 switch (*s) {
278 default:
279 fputc(*s, e->f);
280 break;
281 case 'R':
282 fputs(rn, e->f);
283 break;
284 case 'S':
285 fputs(p, e->f);
286 fputs(l, e->f);
287 break;
288 case 'O':
289 if (c->bits.i)
290 /* todo, handle large offsets */
291 fprintf(e->f, "+%"PRIi64, c->bits.i);
292 break;
293 }
294 }
295
296 static void
297 loadcon(Con *c, int r, int k, E *e)
298 {
299 char *rn;
300 int64_t n;
301 int w, sh;
302
303 w = KWIDE(k);
304 rn = rname(r, k);
305 n = c->bits.i;
306 if (c->type == CAddr) {
307 loadaddr(c, rn, e);
308 return;
309 }
310 assert(c->type == CBits);
311 if (!w)
312 n = (int32_t)n;
313 if ((n | 0xffff) == -1 || arm64_logimm(n, k)) {
314 fprintf(e->f, "\tmov\t%s, #%"PRIi64"\n", rn, n);
315 } else {
316 fprintf(e->f, "\tmov\t%s, #%d\n",
317 rn, (int)(n & 0xffff));
318 for (sh=16; n>>=16; sh+=16) {
319 if ((!w && sh == 32) || sh == 64)
320 break;
321 fprintf(e->f, "\tmovk\t%s, #0x%x, lsl #%d\n",
322 rn, (uint)(n & 0xffff), sh);
323 }
324 }
325 }
326
327 static void emitins(Ins *, E *);
328
329 static void
330 fixarg(Ref *pr, int sz, E *e)
331 {
332 Ins *i;
333 Ref r;
334 uint64_t s;
335
336 r = *pr;
337 if (rtype(r) == RSlot) {
338 s = slot(r, e);
339 if (s > sz * 4095u) {
340 i = &(Ins){Oaddr, Kl, TMP(IP0), {r}};
341 emitins(i, e);
342 *pr = TMP(IP0);
343 }
344 }
345 }
346
347 static void
348 emitins(Ins *i, E *e)
349 {
350 char *l, *p, *rn;
351 uint64_t s;
352 int o;
353 Ref r;
354 Con *c;
355
356 switch (i->op) {
357 default:
358 if (isload(i->op))
359 fixarg(&i->arg[0], loadsz(i), e);
360 if (isstore(i->op))
361 fixarg(&i->arg[1], storesz(i), e);
362 Table:
363 /* most instructions are just pulled out of
364 * the table omap[], some special cases are
365 * detailed below */
366 for (o=0;; o++) {
367 /* this linear search should really be a binary
368 * search */
369 if (omap[o].op == NOp)
370 die("no match for %s(%c)",
371 optab[i->op].name, "wlsd"[i->cls]);
372 if (omap[o].op == i->op)
373 if (omap[o].cls == i->cls || omap[o].cls == Ka
374 || (omap[o].cls == Ki && KBASE(i->cls) == 0))
375 break;
376 }
377 emitf(omap[o].fmt, i, e);
378 break;
379 case Onop:
380 break;
381 case Ocopy:
382 if (req(i->to, i->arg[0]))
383 break;
384 if (rtype(i->to) == RSlot) {
385 r = i->to;
386 if (!isreg(i->arg[0])) {
387 i->to = TMP(R18);
388 emitins(i, e);
389 i->arg[0] = i->to;
390 }
391 i->op = Ostorew + i->cls;
392 i->cls = Kw;
393 i->arg[1] = r;
394 emitins(i, e);
395 break;
396 }
397 assert(isreg(i->to));
398 switch (rtype(i->arg[0])) {
399 case RCon:
400 c = &e->fn->con[i->arg[0].val];
401 loadcon(c, i->to.val, i->cls, e);
402 break;
403 case RSlot:
404 i->op = Oload;
405 emitins(i, e);
406 break;
407 default:
408 assert(i->to.val != R18);
409 goto Table;
410 }
411 break;
412 case Oaddr:
413 assert(rtype(i->arg[0]) == RSlot);
414 rn = rname(i->to.val, Kl);
415 s = slot(i->arg[0], e);
416 if (s <= 4095)
417 fprintf(e->f, "\tadd\t%s, x29, #%"PRIu64"\n", rn, s);
418 else if (s <= 65535)
419 fprintf(e->f,
420 "\tmov\t%s, #%"PRIu64"\n"
421 "\tadd\t%s, x29, %s\n",
422 rn, s, rn, rn
423 );
424 else
425 fprintf(e->f,
426 "\tmov\t%s, #%"PRIu64"\n"
427 "\tmovk\t%s, #%"PRIu64", lsl #16\n"
428 "\tadd\t%s, x29, %s\n",
429 rn, s & 0xFFFF, rn, s >> 16, rn, rn
430 );
431 break;
432 case Ocall:
433 if (rtype(i->arg[0]) != RCon)
434 goto Table;
435 c = &e->fn->con[i->arg[0].val];
436 if (c->type != CAddr
437 || c->sym.type != SGlo
438 || c->bits.i)
439 die("invalid call argument");
440 l = str(c->sym.id);
441 p = l[0] == '"' ? "" : T.assym;
442 fprintf(e->f, "\tbl\t%s%s\n", p, l);
443 break;
444 case Osalloc:
445 emitf("sub sp, sp, %0", i, e);
446 if (!req(i->to, R))
447 emitf("mov %=, sp", i, e);
448 break;
449 case Odbgloc:
450 emitdbgloc(i->arg[0].val, i->arg[1].val, e->f);
451 break;
452 }
453 }
454
455 static void
456 framelayout(E *e)
457 {
458 int *r;
459 uint o;
460 uint64_t f;
461
462 for (o=0, r=arm64_rclob; *r>=0; r++)
463 o += 1 & (e->fn->reg >> *r);
464 f = e->fn->slot;
465 f = (f + 3) & -4;
466 o += o & 1;
467 e->padding = 4*(f-e->fn->slot);
468 e->frame = 4*f + 8*o;
469 }
470
471 /*
472
473 Stack-frame layout:
474
475 +=============+
476 | varargs |
477 | save area |
478 +-------------+
479 | callee-save | ^
480 | registers | |
481 +-------------+ |
482 | ... | |
483 | spill slots | |
484 | ... | | e->frame
485 +-------------+ |
486 | ... | |
487 | locals | |
488 | ... | |
489 +-------------+ |
490 | e->padding | v
491 +-------------+
492 | saved x29 |
493 | saved x30 |
494 +=============+ <- x29
495
496 */
497
498 void
499 arm64_emitfn(Fn *fn, FILE *out)
500 {
501 static char *ctoa[] = {
502 #define X(c, s) [c] = s,
503 CMP(X)
504 #undef X
505 };
506 static int id0;
507 int s, n, c, lbl, *r;
508 uint64_t o;
509 Blk *b, *t;
510 Ins *i;
511 E *e;
512
513 e = &(E){.f = out, .fn = fn};
514 if (T.apple)
515 e->fn->lnk.align = 4;
516 emitfnlnk(e->fn->name, &e->fn->lnk, e->f);
517 fputs("\thint\t#34\n", e->f);
518 framelayout(e);
519
520 if (e->fn->vararg && !T.apple) {
521 for (n=7; n>=0; n--)
522 fprintf(e->f, "\tstr\tq%d, [sp, -16]!\n", n);
523 for (n=7; n>=0; n-=2)
524 fprintf(e->f, "\tstp\tx%d, x%d, [sp, -16]!\n", n-1, n);
525 }
526
527 if (e->frame + 16 <= 512)
528 fprintf(e->f,
529 "\tstp\tx29, x30, [sp, -%"PRIu64"]!\n",
530 e->frame + 16
531 );
532 else if (e->frame <= 4095)
533 fprintf(e->f,
534 "\tsub\tsp, sp, #%"PRIu64"\n"
535 "\tstp\tx29, x30, [sp, -16]!\n",
536 e->frame
537 );
538 else if (e->frame <= 65535)
539 fprintf(e->f,
540 "\tmov\tx16, #%"PRIu64"\n"
541 "\tsub\tsp, sp, x16\n"
542 "\tstp\tx29, x30, [sp, -16]!\n",
543 e->frame
544 );
545 else
546 fprintf(e->f,
547 "\tmov\tx16, #%"PRIu64"\n"
548 "\tmovk\tx16, #%"PRIu64", lsl #16\n"
549 "\tsub\tsp, sp, x16\n"
550 "\tstp\tx29, x30, [sp, -16]!\n",
551 e->frame & 0xFFFF, e->frame >> 16
552 );
553 fputs("\tmov\tx29, sp\n", e->f);
554 s = (e->frame - e->padding) / 4;
555 for (r=arm64_rclob; *r>=0; r++)
556 if (e->fn->reg & BIT(*r)) {
557 s -= 2;
558 i = &(Ins){.arg = {TMP(*r), SLOT(s)}};
559 i->op = *r >= V0 ? Ostored : Ostorel;
560 emitins(i, e);
561 }
562
563 for (lbl=0, b=e->fn->start; b; b=b->link) {
564 if (lbl || b->npred > 1)
565 fprintf(e->f, "%s%d:\n", T.asloc, id0+b->id);
566 for (i=b->ins; i!=&b->ins[b->nins]; i++)
567 emitins(i, e);
568 lbl = 1;
569 switch (b->jmp.type) {
570 case Jhlt:
571 fprintf(e->f, "\tbrk\t#1000\n");
572 break;
573 case Jret0:
574 s = (e->frame - e->padding) / 4;
575 for (r=arm64_rclob; *r>=0; r++)
576 if (e->fn->reg & BIT(*r)) {
577 s -= 2;
578 i = &(Ins){Oload, 0, TMP(*r), {SLOT(s)}};
579 i->cls = *r >= V0 ? Kd : Kl;
580 emitins(i, e);
581 }
582 if (e->fn->dynalloc)
583 fputs("\tmov sp, x29\n", e->f);
584 o = e->frame + 16;
585 if (e->fn->vararg && !T.apple)
586 o += 192;
587 if (o <= 504)
588 fprintf(e->f,
589 "\tldp\tx29, x30, [sp], %"PRIu64"\n",
590 o
591 );
592 else if (o - 16 <= 4095)
593 fprintf(e->f,
594 "\tldp\tx29, x30, [sp], 16\n"
595 "\tadd\tsp, sp, #%"PRIu64"\n",
596 o - 16
597 );
598 else if (o - 16 <= 65535)
599 fprintf(e->f,
600 "\tldp\tx29, x30, [sp], 16\n"
601 "\tmov\tx16, #%"PRIu64"\n"
602 "\tadd\tsp, sp, x16\n",
603 o - 16
604 );
605 else
606 fprintf(e->f,
607 "\tldp\tx29, x30, [sp], 16\n"
608 "\tmov\tx16, #%"PRIu64"\n"
609 "\tmovk\tx16, #%"PRIu64", lsl #16\n"
610 "\tadd\tsp, sp, x16\n",
611 (o - 16) & 0xFFFF, (o - 16) >> 16
612 );
613 fprintf(e->f, "\tret\n");
614 break;
615 case Jjmp:
616 Jmp:
617 if (b->s1 != b->link)
618 fprintf(e->f,
619 "\tb\t%s%d\n",
620 T.asloc, id0+b->s1->id
621 );
622 else
623 lbl = 0;
624 break;
625 default:
626 c = b->jmp.type - Jjf;
627 if (c < 0 || c > NCmp)
628 die("unhandled jump %d", b->jmp.type);
629 if (b->link == b->s2) {
630 t = b->s1;
631 b->s1 = b->s2;
632 b->s2 = t;
633 } else
634 c = cmpneg(c);
635 fprintf(e->f,
636 "\tb%s\t%s%d\n",
637 ctoa[c], T.asloc, id0+b->s2->id
638 );
639 goto Jmp;
640 }
641 }
642 id0 += e->fn->nblk;
643 if (!T.apple)
644 elf_emitfnfin(fn->name, out);
645 }
QBE - Compiler Backend