Restore eda2c756edc4dca004ba217d5bf361235dd9de1f
[tinycc.git] / x86_64-gen.c
blob72709befe341b5cb8ae283c76ab0fe8ab0d2fd4b
1 /*
2 * x86-64 code generator for TCC
4 * Copyright (c) 2008 Shinichiro Hamaji
6 * Based on i386-gen.c by Fabrice Bellard
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2 of the License, or (at your option) any later version.
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 #ifdef TARGET_DEFS_ONLY
25 /* number of available registers */
26 #define NB_REGS 25
27 #define NB_ASM_REGS 8
29 /* a register can belong to several classes. The classes must be
30 sorted from more general to more precise (see gv2() code which does
31 assumptions on it). */
32 #define RC_INT 0x0001 /* generic integer register */
33 #define RC_FLOAT 0x0002 /* generic float register */
34 #define RC_RAX 0x0004
35 #define RC_RCX 0x0008
36 #define RC_RDX 0x0010
37 #define RC_ST0 0x0020 /* only for long double */
38 #define RC_R8 0x0040
39 #define RC_R9 0x0080
40 #define RC_XMM0 0x0100
41 #define RC_XMM1 0x0200
42 #define RC_XMM2 0x0400
43 #define RC_XMM3 0x0800
44 #define RC_XMM4 0x1000
45 #define RC_XMM5 0x2000
46 #define RC_XMM6 0x4000
47 #define RC_XMM7 0x8000
48 #define RC_RSI 0x10000
49 #define RC_RDI 0x20000
50 #define RC_INT1 0x40000 /* function_pointer */
51 #define RC_INT2 0x80000
52 #define RC_RBX 0x100000
53 #define RC_R10 0x200000
54 #define RC_R11 0x400000
55 #define RC_R12 0x800000
56 #define RC_R13 0x1000000
57 #define RC_R14 0x2000000
58 #define RC_R15 0x4000000
59 #define RC_IRET RC_RAX /* function return: integer register */
60 #define RC_LRET RC_RDX /* function return: second integer register */
61 #define RC_FRET RC_XMM0 /* function return: float register */
62 #define RC_QRET RC_XMM1 /* function return: second float register */
63 #define RC_MASK (RC_INT|RC_INT1|RC_INT2|RC_FLOAT)
65 /* pretty names for the registers */
66 enum {
67 TREG_RAX = 0,
68 TREG_RCX = 1,
69 TREG_RDX = 2,
70 TREG_RSP = 4,
71 TREG_ST0 = 5,
72 TREG_RSI = 6,
73 TREG_RDI = 7,
75 TREG_R8 = 8,
76 TREG_R9 = 9,
77 TREG_R10 = 10,
78 TREG_R11 = 11,
80 TREG_XMM0 = 16,
81 TREG_XMM1 = 17,
82 TREG_XMM2 = 18,
83 TREG_XMM3 = 19,
84 TREG_XMM4 = 20,
85 TREG_XMM5 = 21,
86 TREG_XMM6 = 22,
87 TREG_XMM7 = 23,
91 #define REX_BASE(reg) (((reg) >> 3) & 1)
92 #define REG_VALUE(reg) ((reg) & 7)
93 #define FLAG_GOT 0X01
95 /* return registers for function */
96 #define REG_IRET TREG_RAX /* single word int return register */
97 #define REG_LRET TREG_RDX /* second word return register (for long long) */
98 #define REG_FRET TREG_XMM0 /* float return register */
99 #define REG_QRET TREG_XMM1 /* second float return register */
101 /* defined if function parameters must be evaluated in reverse order */
102 #define INVERT_FUNC_PARAMS
104 /* pointer size, in bytes */
105 #define PTR_SIZE 8
107 /* long double size and alignment, in bytes */
108 #define LDOUBLE_SIZE 16
109 #define LDOUBLE_ALIGN 16
110 /* maximum alignment (for aligned attribute support) */
111 #define MAX_ALIGN 16
113 /******************************************************/
114 /* ELF defines */
116 #define EM_TCC_TARGET EM_X86_64
118 /* relocation type for 32 bit data relocation */
119 #define R_DATA_32 R_X86_64_32
120 #define R_DATA_PTR R_X86_64_64
121 #define R_JMP_SLOT R_X86_64_JUMP_SLOT
122 #define R_COPY R_X86_64_COPY
124 #define ELF_START_ADDR 0x400000
125 #define ELF_PAGE_SIZE 0x200000
127 /******************************************************/
128 #else /* ! TARGET_DEFS_ONLY */
129 /******************************************************/
130 #include "tcc.h"
131 #include <assert.h>
133 ST_DATA const int reg_classes[NB_REGS] = {
134 /* eax */ RC_INT|RC_RAX|RC_INT2,
135 /* ecx */ RC_INT|RC_RCX|RC_INT2,
136 /* edx */ RC_INT|RC_RDX,
137 RC_INT|RC_INT1|RC_INT2|RC_RBX,
139 /* st0 */ RC_ST0,
140 RC_RSI|RC_INT2,
141 RC_RDI|RC_INT2,
142 RC_INT|RC_R8|RC_INT2,
143 RC_INT|RC_R9|RC_INT2,
144 RC_INT|RC_INT1|RC_INT2|RC_R10,
145 RC_INT|RC_INT1|RC_INT2|RC_R11,
146 RC_INT|RC_INT1|RC_INT2|RC_R12,
147 RC_INT|RC_INT1|RC_INT2|RC_R13,
148 RC_INT|RC_INT1|RC_INT2|RC_R14,
149 RC_INT|RC_INT1|RC_INT2|RC_R15,
150 /* xmm0 */ RC_FLOAT | RC_XMM0,
151 RC_FLOAT|RC_XMM1,
152 RC_FLOAT|RC_XMM2,
153 RC_FLOAT|RC_XMM3,
154 RC_FLOAT|RC_XMM4,
155 RC_FLOAT|RC_XMM5,
156 RC_FLOAT|RC_XMM6,
157 RC_FLOAT|RC_XMM7,
160 static unsigned long func_sub_sp_offset;
161 static int func_ret_sub;
163 /* XXX: make it faster ? */
164 void g(int c)
166 int ind1;
167 ind1 = ind + 1;
168 if (ind1 > cur_text_section->data_allocated)
169 section_realloc(cur_text_section, ind1);
170 cur_text_section->data[ind] = c;
171 ind = ind1;
174 void o(unsigned int c)
176 while (c) {
177 g(c);
178 c = c >> 8;
182 void gen_le16(int v)
184 g(v);
185 g(v >> 8);
188 void gen_le32(int c)
190 g(c);
191 g(c >> 8);
192 g(c >> 16);
193 g(c >> 24);
196 void gen_le64(int64_t c)
198 g(c);
199 g(c >> 8);
200 g(c >> 16);
201 g(c >> 24);
202 g(c >> 32);
203 g(c >> 40);
204 g(c >> 48);
205 g(c >> 56);
208 void orex(int ll, int r, int r2, int b)
210 if ((r & VT_VALMASK) >= VT_CONST)
211 r = 0;
212 if ((r2 & VT_VALMASK) >= VT_CONST)
213 r2 = 0;
214 if (ll || REX_BASE(r) || REX_BASE(r2))
215 o(0x40 | REX_BASE(r) | (REX_BASE(r2) << 2) | (ll << 3));
216 o(b);
219 /* output a symbol and patch all calls to it */
220 void gsym_addr(int t, int a)
222 int n, *ptr;
223 while (t) {
224 ptr = (int *)(cur_text_section->data + t);
225 n = *ptr; /* next value */
226 *ptr = a - t - 4;
227 t = n;
231 void gsym(int t)
233 gsym_addr(t, ind);
236 /* psym is used to put an instruction with a data field which is a
237 reference to a symbol. It is in fact the same as oad ! */
238 #define psym oad
240 static int is64_type(int t)
242 return ((t & VT_BTYPE) == VT_PTR ||
243 (t & VT_BTYPE) == VT_FUNC ||
244 (t & VT_BTYPE) == VT_LLONG);
247 /* instruction + 4 bytes data. Return the address of the data */
248 ST_FUNC int oad(int c, int s)
250 int ind1;
252 o(c);
253 ind1 = ind + 4;
254 if (ind1 > cur_text_section->data_allocated)
255 section_realloc(cur_text_section, ind1);
256 *(int *)(cur_text_section->data + ind) = s;
257 s = ind;
258 ind = ind1;
259 return s;
262 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
264 if (r & VT_SYM)
265 greloc(cur_text_section, sym, ind, R_X86_64_32);
266 gen_le32(c);
269 /* output constant with relocation if 'r & VT_SYM' is true */
270 ST_FUNC void gen_addr64(int r, Sym *sym, int64_t c)
272 if (r & VT_SYM)
273 greloc(cur_text_section, sym, ind, R_X86_64_64);
274 gen_le64(c);
277 /* output constant with relocation if 'r & VT_SYM' is true */
278 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
280 if (r & VT_SYM)
281 greloc(cur_text_section, sym, ind, R_X86_64_PC32);
282 gen_le32(c-4);
285 /* output got address with relocation */
286 static void gen_gotpcrel(int r, Sym *sym, int c)
288 #ifndef TCC_TARGET_PE
289 Section *sr;
290 ElfW(Rela) *rel;
291 greloc(cur_text_section, sym, ind, R_X86_64_GOTPCREL);
292 sr = cur_text_section->reloc;
293 rel = (ElfW(Rela) *)(sr->data + sr->data_offset - sizeof(ElfW(Rela)));
294 rel->r_addend = -4;
295 #else
296 printf("picpic: %s %x %x | %02x %02x %02x\n", get_tok_str(sym->v, NULL), c, r,
297 cur_text_section->data[ind-3],
298 cur_text_section->data[ind-2],
299 cur_text_section->data[ind-1]
301 greloc(cur_text_section, sym, ind, R_X86_64_PC32);
302 #endif
303 gen_le32(0);
304 if (c) {
305 /* we use add c, %xxx for displacement */
306 orex(1, r, 0, 0x81);
307 o(0xc0 + REG_VALUE(r));
308 gen_le32(c);
312 static void gen_modrm_impl(int op_reg, int r, Sym *sym, int c, int is_got)
314 op_reg = REG_VALUE(op_reg) << 3;
315 if ((r & VT_VALMASK) == VT_CONST) {
316 /* constant memory reference */
317 o(0x05 | op_reg);
318 if (is_got) {
319 gen_gotpcrel(r, sym, c);
320 } else {
321 gen_addrpc32(r, sym, c);
323 } else if ((r & VT_VALMASK) == VT_LOCAL) {
324 /* currently, we use only ebp as base */
325 if (c == (char)c) {
326 /* short reference */
327 o(0x45 | op_reg);
328 g(c);
329 } else {
330 oad(0x85 | op_reg, c);
332 } else if (r & TREG_MEM) {
333 if (c) {
334 g(0x80 | op_reg | REG_VALUE(r));
335 gen_le32(c);
336 } else {
337 g(0x00 | op_reg | REG_VALUE(r));
339 } else {
340 g(0x00 | op_reg | REG_VALUE(r));
344 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
345 opcode bits */
346 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
348 gen_modrm_impl(op_reg, r, sym, c, 0);
351 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
352 opcode bits */
353 static void gen_modrm64(int opcode, int op_reg, int r, Sym *sym, int c)
355 int is_got;
356 is_got = (op_reg & TREG_MEM) && !(sym->type.t & VT_STATIC);
357 orex(1, r, op_reg, opcode);
358 gen_modrm_impl(op_reg, r, sym, c, is_got);
362 /* load 'r' from value 'sv' */
363 void load(int r, SValue *sv)
365 int v, t, ft, fc, fr;
366 SValue v1;
368 #ifdef TCC_TARGET_PE
369 SValue v2;
370 sv = pe_getimport(sv, &v2);
371 #endif
373 fr = sv->r;
374 ft = sv->type.t & ~VT_DEFSIGN;
375 fc = sv->c.ul;
377 #ifndef TCC_TARGET_PE
378 /* we use indirect access via got */
379 if ((fr & VT_VALMASK) == VT_CONST && (fr & VT_SYM) &&
380 (fr & VT_LVAL) && !(sv->sym->type.t & VT_STATIC)) {
381 /* use the result register as a temporal register */
382 int tr = r | TREG_MEM;
383 if (is_float(ft)) {
384 /* we cannot use float registers as a temporal register */
385 tr = get_reg(RC_INT) | TREG_MEM;
387 gen_modrm64(0x8b, tr, fr, sv->sym, 0);
389 /* load from the temporal register */
390 fr = tr | VT_LVAL;
392 #endif
394 v = fr & VT_VALMASK;
395 if (fr & VT_LVAL) {
396 int b, ll;
397 if (v == VT_LLOCAL) {
398 v1.type.t = VT_PTR;
399 v1.r = VT_LOCAL | VT_LVAL;
400 v1.c.ul = fc;
401 fr = r;
402 if (!(reg_classes[fr] & RC_INT))
403 fr = get_reg(RC_INT);
404 load(fr, &v1);
406 ll = 0;
407 if ((ft & VT_BTYPE) == VT_FLOAT) {
408 b = 0x6e0f66;
409 r = REG_VALUE(r); /* movd */
410 } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
411 b = 0x7e0ff3; /* movq */
412 r = REG_VALUE(r);
413 } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
414 b = 0xdb, r = 5; /* fldt */
415 } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
416 b = 0xbe0f; /* movsbl */
417 } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
418 b = 0xb60f; /* movzbl */
419 } else if ((ft & VT_TYPE) == VT_SHORT) {
420 b = 0xbf0f; /* movswl */
421 } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
422 b = 0xb70f; /* movzwl */
423 } else {
424 assert(((ft & VT_BTYPE) == VT_INT) || ((ft & VT_BTYPE) == VT_LLONG)
425 || ((ft & VT_BTYPE) == VT_PTR) || ((ft & VT_BTYPE) == VT_ENUM)
426 || ((ft & VT_BTYPE) == VT_FUNC));
427 ll = is64_type(ft);
428 b = 0x8b;
430 if (ll) {
431 gen_modrm64(b, r, fr, sv->sym, fc);
432 } else {
433 orex(ll, fr, r, b);
434 gen_modrm(r, fr, sv->sym, fc);
436 } else {
437 if (v == VT_CONST) {
438 if (fr & VT_SYM) {
439 #ifdef TCC_TARGET_PE
440 orex(1,0,r,0x8d);
441 o(0x05 + REG_VALUE(r) * 8); /* lea xx(%rip), r */
442 gen_addrpc32(fr, sv->sym, fc);
443 #else
444 if (sv->sym->type.t & VT_STATIC) {
445 orex(1,0,r,0x8d);
446 o(0x05 + REG_VALUE(r) * 8); /* lea xx(%rip), r */
447 gen_addrpc32(fr, sv->sym, fc);
448 } else {
449 orex(1,0,r,0x8b);
450 o(0x05 + REG_VALUE(r) * 8); /* mov xx(%rip), r */
451 gen_gotpcrel(r, sv->sym, fc);
453 #endif
454 } else if (is64_type(ft)) {
455 orex(1,r,0, 0xb8 + REG_VALUE(r)); /* mov $xx, r */
456 gen_le64(sv->c.ull);
457 } else {
458 orex(0,r,0, 0xb8 + REG_VALUE(r)); /* mov $xx, r */
459 gen_le32(fc);
461 } else if (v == VT_LOCAL) {
462 orex(1,0,r,0x8d); /* lea xxx(%ebp), r */
463 gen_modrm(r, VT_LOCAL, sv->sym, fc);
464 } else if (v == VT_CMP) {
465 orex(0,r,0,0);
466 if ((fc & ~0x100) != TOK_NE)
467 oad(0xb8 + REG_VALUE(r), 0); /* mov $0, r */
468 else
469 oad(0xb8 + REG_VALUE(r), 1); /* mov $1, r */
470 if (fc & 0x100)
472 /* This was a float compare. If the parity bit is
473 set the result was unordered, meaning false for everything
474 except TOK_NE, and true for TOK_NE. */
475 fc &= ~0x100;
476 o(0x037a + (REX_BASE(r) << 8));
478 orex(0,r,0, 0x0f); /* setxx %br */
479 o(fc);
480 o(0xc0 + REG_VALUE(r));
481 } else if (v == VT_JMP || v == VT_JMPI) {
482 t = v & 1;
483 orex(0,r,0,0);
484 oad(0xb8 + REG_VALUE(r), t); /* mov $1, r */
485 o(0x05eb + (REX_BASE(r) << 8)); /* jmp after */
486 gsym(fc);
487 orex(0,r,0,0);
488 oad(0xb8 + REG_VALUE(r), t ^ 1); /* mov $0, r */
489 } else if (v != r) {
490 if ((r >= TREG_XMM0) && (r <= TREG_XMM7)) {
491 if (v == TREG_ST0) {
492 /* gen_cvt_ftof(VT_DOUBLE); */
493 o(0xf0245cdd); /* fstpl -0x10(%rsp) */
494 /* movsd -0x10(%rsp),%xmmN */
495 o(0x100ff2);
496 o(0x44 + REG_VALUE(r)*8); /* %xmmN */
497 o(0xf024);
498 } else {
499 assert((v >= TREG_XMM0) && (v <= TREG_XMM7));
500 if ((ft & VT_BTYPE) == VT_FLOAT) {
501 o(0x100ff3);
502 } else {
503 assert((ft & VT_BTYPE) == VT_DOUBLE);
504 o(0x100ff2);
506 o(0xc0 + REG_VALUE(v) + REG_VALUE(r)*8);
508 } else if (r == TREG_ST0) {
509 assert((v >= TREG_XMM0) && (v <= TREG_XMM7));
510 /* gen_cvt_ftof(VT_LDOUBLE); */
511 /* movsd %xmmN,-0x10(%rsp) */
512 o(0x110ff2);
513 o(0x44 + REG_VALUE(r)*8); /* %xmmN */
514 o(0xf024);
515 o(0xf02444dd); /* fldl -0x10(%rsp) */
516 } else {
517 orex(1,r,v, 0x89);
518 o(0xc0 + REG_VALUE(r) + REG_VALUE(v) * 8); /* mov v, r */
524 /* store register 'r' in lvalue 'v' */
525 void store(int r, SValue *v)
527 int fr, bt, ft, fc;
528 int op64 = 0;
529 /* store the REX prefix in this variable when PIC is enabled */
530 int pic = 0;
532 #ifdef TCC_TARGET_PE
533 SValue v2;
534 v = pe_getimport(v, &v2);
535 #endif
537 ft = v->type.t;
538 fc = v->c.ul;
539 fr = v->r & VT_VALMASK;
540 bt = ft & VT_BTYPE;
542 #ifndef TCC_TARGET_PE
543 /* we need to access the variable via got */
544 if (fr == VT_CONST && (v->r & VT_SYM)) {
545 /* mov xx(%rip), %r11 */
546 o(0x1d8b4c);
547 gen_gotpcrel(TREG_R11, v->sym, v->c.ul);
548 pic = is64_type(bt) ? 0x49 : 0x41;
550 #endif
552 /* XXX: incorrect if float reg to reg */
553 if (bt == VT_FLOAT) {
554 o(0x66);
555 o(pic);
556 o(0x7e0f); /* movd */
557 r = REG_VALUE(r);
558 } else if (bt == VT_DOUBLE) {
559 o(0x66);
560 o(pic);
561 o(0xd60f); /* movq */
562 r = REG_VALUE(r);
563 } else if (bt == VT_LDOUBLE) {
564 o(0xc0d9); /* fld %st(0) */
565 o(pic);
566 o(0xdb); /* fstpt */
567 r = 7;
568 } else {
569 if (bt == VT_SHORT)
570 o(0x66);
571 o(pic);
572 if (bt == VT_BYTE || bt == VT_BOOL)
573 orex(0, 0, r, 0x88);
574 else if (is64_type(bt))
575 op64 = 0x89;
576 else
577 orex(0, 0, r, 0x89);
579 if (pic) {
580 /* xxx r, (%r11) where xxx is mov, movq, fld, or etc */
581 if (op64)
582 o(op64);
583 o(3 + (r << 3));
584 } else if (op64) {
585 if (fr == VT_CONST || fr == VT_LOCAL || (v->r & VT_LVAL)) {
586 gen_modrm64(op64, r, v->r, v->sym, fc);
587 } else if (fr != r) {
588 /* XXX: don't we really come here? */
589 abort();
590 o(0xc0 + fr + r * 8); /* mov r, fr */
592 } else {
593 if (fr == VT_CONST || fr == VT_LOCAL || (v->r & VT_LVAL)) {
594 gen_modrm(r, v->r, v->sym, fc);
595 } else if (fr != r) {
596 /* XXX: don't we really come here? */
597 abort();
598 o(0xc0 + fr + r * 8); /* mov r, fr */
603 /* 'is_jmp' is '1' if it is a jump */
604 static void gcall_or_jmp(int is_jmp)
606 int r;
607 if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
608 /* constant case */
609 if (vtop->r & VT_SYM) {
610 /* relocation case */
611 greloc(cur_text_section, vtop->sym,
612 ind + 1, R_X86_64_PLT32);
613 } else {
614 /* put an empty PC32 relocation */
615 put_elf_reloc(symtab_section, cur_text_section,
616 ind + 1, R_X86_64_PC32, 0);
618 oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
619 } else {
620 /* otherwise, indirect call */
621 r = TREG_R11;
622 load(r, vtop);
623 o(0x41); /* REX */
624 o(0xff); /* call/jmp *r */
625 o(0xd0 + REG_VALUE(r) + (is_jmp << 4));
629 #ifdef TCC_TARGET_PE
631 #define REGN 4
632 static const uint8_t arg_regs[REGN] = {
633 TREG_RCX, TREG_RDX, TREG_R8, TREG_R9
636 /* Prepare arguments in R10 and R11 rather than RCX and RDX
637 because gv() will not ever use these */
638 static int arg_prepare_reg(int idx) {
639 if (idx == 0 || idx == 1)
640 /* idx=0: r10, idx=1: r11 */
641 return idx + 10;
642 else
643 return arg_regs[idx];
646 static int func_scratch;
648 /* Generate function call. The function address is pushed first, then
649 all the parameters in call order. This functions pops all the
650 parameters and the function address. */
652 void gen_offs_sp(int b, int r, int d)
654 orex(1,0,r & 0x100 ? 0 : r, b);
655 if (d == (char)d) {
656 o(0x2444 | (REG_VALUE(r) << 3));
657 g(d);
658 } else {
659 o(0x2484 | (REG_VALUE(r) << 3));
660 gen_le32(d);
664 /* Return the number of registers needed to return the struct, or 0 if
665 returning via struct pointer. */
666 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align)
668 int size, align;
669 *ret_align = 1; // Never have to re-align return values for x86-64
670 size = type_size(vt, &align);
671 ret->ref = NULL;
672 if (size > 8) {
673 return 0;
674 } else if (size > 4) {
675 ret->t = VT_LLONG;
676 return 1;
677 } else if (size > 2) {
678 ret->t = VT_INT;
679 return 1;
680 } else if (size > 1) {
681 ret->t = VT_SHORT;
682 return 1;
683 } else {
684 ret->t = VT_BYTE;
685 return 1;
689 static int is_sse_float(int t) {
690 int bt;
691 bt = t & VT_BTYPE;
692 return bt == VT_DOUBLE || bt == VT_FLOAT;
695 int gfunc_arg_size(CType *type) {
696 int align;
697 if (type->t & (VT_ARRAY|VT_BITFIELD))
698 return 8;
699 return type_size(type, &align);
702 void gfunc_call(int nb_args)
704 int size, r, args_size, i, d, bt, struct_size;
705 int arg;
707 args_size = (nb_args < REGN ? REGN : nb_args) * PTR_SIZE;
708 arg = nb_args;
710 /* for struct arguments, we need to call memcpy and the function
711 call breaks register passing arguments we are preparing.
712 So, we process arguments which will be passed by stack first. */
713 struct_size = args_size;
714 for(i = 0; i < nb_args; i++) {
715 SValue *sv;
717 --arg;
718 sv = &vtop[-i];
719 bt = (sv->type.t & VT_BTYPE);
720 size = gfunc_arg_size(&sv->type);
722 if (size <= 8)
723 continue; /* arguments smaller than 8 bytes passed in registers or on stack */
725 if (bt == VT_STRUCT) {
726 /* align to stack align size */
727 size = (size + 15) & ~15;
728 /* generate structure store */
729 r = get_reg(RC_INT);
730 gen_offs_sp(0x8d, r, struct_size);
731 struct_size += size;
733 /* generate memcpy call */
734 vset(&sv->type, r | VT_LVAL, 0);
735 vpushv(sv);
736 vstore();
737 --vtop;
738 } else if (bt == VT_LDOUBLE) {
739 gv(RC_ST0);
740 gen_offs_sp(0xdb, 0x107, struct_size);
741 struct_size += 16;
745 if (func_scratch < struct_size)
746 func_scratch = struct_size;
748 arg = nb_args;
749 struct_size = args_size;
751 for(i = 0; i < nb_args; i++) {
752 --arg;
753 bt = (vtop->type.t & VT_BTYPE);
755 size = gfunc_arg_size(&vtop->type);
756 if (size > 8) {
757 /* align to stack align size */
758 size = (size + 15) & ~15;
759 if (arg >= REGN) {
760 d = get_reg(RC_INT);
761 gen_offs_sp(0x8d, d, struct_size);
762 gen_offs_sp(0x89, d, arg*8);
763 } else {
764 d = arg_prepare_reg(arg);
765 gen_offs_sp(0x8d, d, struct_size);
767 struct_size += size;
768 } else {
769 if (is_sse_float(vtop->type.t)) {
770 gv(RC_XMM0); /* only use one float register */
771 if (arg >= REGN) {
772 /* movq %xmm0, j*8(%rsp) */
773 gen_offs_sp(0xd60f66, 0x100, arg*8);
774 } else {
775 /* movaps %xmm0, %xmmN */
776 o(0x280f);
777 o(0xc0 + (arg << 3));
778 d = arg_prepare_reg(arg);
779 /* mov %xmm0, %rxx */
780 o(0x66);
781 orex(1,d,0, 0x7e0f);
782 o(0xc0 + REG_VALUE(d));
784 } else {
785 if (bt == VT_STRUCT) {
786 vtop->type.ref = NULL;
787 vtop->type.t = size > 4 ? VT_LLONG : size > 2 ? VT_INT
788 : size > 1 ? VT_SHORT : VT_BYTE;
791 r = gv(RC_INT);
792 if (arg >= REGN) {
793 gen_offs_sp(0x89, r, arg*8);
794 } else {
795 d = arg_prepare_reg(arg);
796 orex(1,d,r,0x89); /* mov */
797 o(0xc0 + REG_VALUE(r) * 8 + REG_VALUE(d));
801 vtop--;
803 save_regs(0);
805 /* Copy R10 and R11 into RCX and RDX, respectively */
806 if (nb_args > 0) {
807 o(0xd1894c); /* mov %r10, %rcx */
808 if (nb_args > 1) {
809 o(0xda894c); /* mov %r11, %rdx */
813 gcall_or_jmp(0);
814 vtop--;
818 #define FUNC_PROLOG_SIZE 11
820 /* generate function prolog of type 't' */
821 void gfunc_prolog(CType *func_type)
823 int addr, reg_param_index, bt, size;
824 Sym *sym;
825 CType *type;
827 func_ret_sub = 0;
828 func_scratch = 0;
829 loc = 0;
831 addr = PTR_SIZE * 2;
832 ind += FUNC_PROLOG_SIZE;
833 func_sub_sp_offset = ind;
834 reg_param_index = 0;
836 sym = func_type->ref;
838 /* if the function returns a structure, then add an
839 implicit pointer parameter */
840 func_vt = sym->type;
841 func_var = (sym->c == FUNC_ELLIPSIS);
842 size = gfunc_arg_size(&func_vt);
843 if (size > 8) {
844 gen_modrm64(0x89, arg_regs[reg_param_index], VT_LOCAL, NULL, addr);
845 func_vc = addr;
846 reg_param_index++;
847 addr += 8;
850 /* define parameters */
851 while ((sym = sym->next) != NULL) {
852 type = &sym->type;
853 bt = type->t & VT_BTYPE;
854 size = gfunc_arg_size(type);
855 if (size > 8) {
856 if (reg_param_index < REGN) {
857 gen_modrm64(0x89, arg_regs[reg_param_index], VT_LOCAL, NULL, addr);
859 sym_push(sym->v & ~SYM_FIELD, type, VT_LOCAL | VT_LVAL | VT_REF, addr);
860 } else {
861 if (reg_param_index < REGN) {
862 /* save arguments passed by register */
863 if ((bt == VT_FLOAT) || (bt == VT_DOUBLE)) {
864 o(0xd60f66); /* movq */
865 gen_modrm(reg_param_index, VT_LOCAL, NULL, addr);
866 } else {
867 gen_modrm64(0x89, arg_regs[reg_param_index], VT_LOCAL, NULL, addr);
870 sym_push(sym->v & ~SYM_FIELD, type, VT_LOCAL | VT_LVAL, addr);
872 addr += 8;
873 reg_param_index++;
876 while (reg_param_index < REGN) {
877 if (func_type->ref->c == FUNC_ELLIPSIS) {
878 gen_modrm64(0x89, arg_regs[reg_param_index], VT_LOCAL, NULL, addr);
879 addr += 8;
881 reg_param_index++;
885 /* generate function epilog */
886 void gfunc_epilog(void)
888 int v, saved_ind;
890 o(0xc9); /* leave */
891 if (func_ret_sub == 0) {
892 o(0xc3); /* ret */
893 } else {
894 o(0xc2); /* ret n */
895 g(func_ret_sub);
896 g(func_ret_sub >> 8);
899 saved_ind = ind;
900 ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
901 /* align local size to word & save local variables */
902 v = (func_scratch + -loc + 15) & -16;
904 if (v >= 4096) {
905 Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
906 oad(0xb8, v); /* mov stacksize, %eax */
907 oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
908 greloc(cur_text_section, sym, ind-4, R_X86_64_PC32);
909 o(0x90); /* fill for FUNC_PROLOG_SIZE = 11 bytes */
910 } else {
911 o(0xe5894855); /* push %rbp, mov %rsp, %rbp */
912 o(0xec8148); /* sub rsp, stacksize */
913 gen_le32(v);
916 cur_text_section->data_offset = saved_ind;
917 pe_add_unwind_data(ind, saved_ind, v);
918 ind = cur_text_section->data_offset;
921 #else
923 static void gadd_sp(int val)
925 if (val == (char)val) {
926 o(0xc48348);
927 g(val);
928 } else {
929 oad(0xc48148, val); /* add $xxx, %rsp */
933 typedef enum X86_64_Mode {
934 x86_64_mode_none,
935 x86_64_mode_memory,
936 x86_64_mode_integer,
937 x86_64_mode_sse,
938 x86_64_mode_x87
939 } X86_64_Mode;
941 static X86_64_Mode classify_x86_64_merge(X86_64_Mode a, X86_64_Mode b)
943 if (a == b)
944 return a;
945 else if (a == x86_64_mode_none)
946 return b;
947 else if (b == x86_64_mode_none)
948 return a;
949 else if ((a == x86_64_mode_memory) || (b == x86_64_mode_memory))
950 return x86_64_mode_memory;
951 else if ((a == x86_64_mode_integer) || (b == x86_64_mode_integer))
952 return x86_64_mode_integer;
953 else if ((a == x86_64_mode_x87) || (b == x86_64_mode_x87))
954 return x86_64_mode_memory;
955 else
956 return x86_64_mode_sse;
959 static X86_64_Mode classify_x86_64_inner(CType *ty)
961 X86_64_Mode mode;
962 Sym *f;
964 switch (ty->t & VT_BTYPE) {
965 case VT_VOID: return x86_64_mode_none;
967 case VT_INT:
968 case VT_BYTE:
969 case VT_SHORT:
970 case VT_LLONG:
971 case VT_BOOL:
972 case VT_PTR:
973 case VT_FUNC:
974 case VT_ENUM: return x86_64_mode_integer;
976 case VT_FLOAT:
977 case VT_DOUBLE: return x86_64_mode_sse;
979 case VT_LDOUBLE: return x86_64_mode_x87;
981 case VT_STRUCT:
982 f = ty->ref;
984 // Detect union
985 if (f->next && (f->c == f->next->c))
986 return x86_64_mode_memory;
988 mode = x86_64_mode_none;
989 for (f = f->next; f; f = f->next)
990 mode = classify_x86_64_merge(mode, classify_x86_64_inner(&f->type));
992 return mode;
995 assert(0);
998 static X86_64_Mode classify_x86_64_arg(CType *ty, CType *ret, int *psize, int *palign, int *reg_count)
1000 X86_64_Mode mode;
1001 int size, align, ret_t = 0;
1003 if (ty->t & (VT_BITFIELD|VT_ARRAY)) {
1004 *psize = 8;
1005 *palign = 8;
1006 *reg_count = 1;
1007 ret_t = ty->t;
1008 mode = x86_64_mode_integer;
1009 } else {
1010 size = type_size(ty, &align);
1011 *psize = (size + 7) & ~7;
1012 *palign = (align + 7) & ~7;
1014 if (size > 16) {
1015 mode = x86_64_mode_memory;
1016 } else {
1017 mode = classify_x86_64_inner(ty);
1018 switch (mode) {
1019 case x86_64_mode_integer:
1020 if (size > 8) {
1021 *reg_count = 2;
1022 ret_t = VT_QLONG;
1023 } else {
1024 *reg_count = 1;
1025 ret_t = (size > 4) ? VT_LLONG : VT_INT;
1027 break;
1029 case x86_64_mode_x87:
1030 *reg_count = 1;
1031 ret_t = VT_LDOUBLE;
1032 break;
1034 case x86_64_mode_sse:
1035 if (size > 8) {
1036 *reg_count = 2;
1037 ret_t = VT_QFLOAT;
1038 } else {
1039 *reg_count = 1;
1040 ret_t = (size > 4) ? VT_DOUBLE : VT_FLOAT;
1042 break;
1043 default: break; /* nothing to be done for x86_64_mode_memory and x86_64_mode_none*/
1048 if (ret) {
1049 ret->ref = NULL;
1050 ret->t = ret_t;
1053 return mode;
1056 ST_FUNC int classify_x86_64_va_arg(CType *ty)
1058 /* This definition must be synced with stdarg.h */
1059 enum __va_arg_type {
1060 __va_gen_reg, __va_float_reg, __va_stack
1062 int size, align, reg_count;
1063 X86_64_Mode mode = classify_x86_64_arg(ty, NULL, &size, &align, &reg_count);
1064 switch (mode) {
1065 default: return __va_stack;
1066 case x86_64_mode_integer: return __va_gen_reg;
1067 case x86_64_mode_sse: return __va_float_reg;
1071 /* Return the number of registers needed to return the struct, or 0 if
1072 returning via struct pointer. */
1073 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align)
1075 int size, align, reg_count;
1076 *ret_align = 1; // Never have to re-align return values for x86-64
1077 return (classify_x86_64_arg(vt, ret, &size, &align, &reg_count) != x86_64_mode_memory);
1080 #define REGN 6
1081 static const uint8_t arg_regs[REGN] = {
1082 TREG_RDI, TREG_RSI, TREG_RDX, TREG_RCX, TREG_R8, TREG_R9
1085 static int arg_prepare_reg(int idx) {
1086 if (idx == 2 || idx == 3)
1087 /* idx=2: r10, idx=3: r11 */
1088 return idx + 8;
1089 else
1090 return arg_regs[idx];
1093 /* Generate function call. The function address is pushed first, then
1094 all the parameters in call order. This functions pops all the
1095 parameters and the function address. */
1096 void gfunc_call(int nb_args)
1098 X86_64_Mode mode;
1099 CType type;
1100 int size, align, r, args_size, stack_adjust, run_start, run_end, i, reg_count;
1101 int nb_reg_args = 0;
1102 int nb_sse_args = 0;
1103 int sse_reg, gen_reg;
1105 /* calculate the number of integer/float register arguments */
1106 for(i = 0; i < nb_args; i++) {
1107 mode = classify_x86_64_arg(&vtop[-i].type, NULL, &size, &align, &reg_count);
1108 if (mode == x86_64_mode_sse)
1109 nb_sse_args += reg_count;
1110 else if (mode == x86_64_mode_integer)
1111 nb_reg_args += reg_count;
1114 /* arguments are collected in runs. Each run is a collection of 8-byte aligned arguments
1115 and ended by a 16-byte aligned argument. This is because, from the point of view of
1116 the callee, argument alignment is computed from the bottom up. */
1117 /* for struct arguments, we need to call memcpy and the function
1118 call breaks register passing arguments we are preparing.
1119 So, we process arguments which will be passed by stack first. */
1120 gen_reg = nb_reg_args;
1121 sse_reg = nb_sse_args;
1122 run_start = 0;
1123 args_size = 0;
1124 while (run_start != nb_args) {
1125 int run_gen_reg = gen_reg, run_sse_reg = sse_reg;
1127 run_end = nb_args;
1128 stack_adjust = 0;
1129 for(i = run_start; (i < nb_args) && (run_end == nb_args); i++) {
1130 mode = classify_x86_64_arg(&vtop[-i].type, NULL, &size, &align, &reg_count);
1131 switch (mode) {
1132 case x86_64_mode_memory:
1133 case x86_64_mode_x87:
1134 stack_arg:
1135 if (align == 16)
1136 run_end = i;
1137 else
1138 stack_adjust += size;
1139 break;
1141 case x86_64_mode_sse:
1142 sse_reg -= reg_count;
1143 if (sse_reg + reg_count > 8) goto stack_arg;
1144 break;
1146 case x86_64_mode_integer:
1147 gen_reg -= reg_count;
1148 if (gen_reg + reg_count > REGN) goto stack_arg;
1149 break;
1150 default: break; /* nothing to be done for x86_64_mode_none */
1154 gen_reg = run_gen_reg;
1155 sse_reg = run_sse_reg;
1157 /* adjust stack to align SSE boundary */
1158 if (stack_adjust &= 15) {
1159 /* fetch cpu flag before the following sub will change the value */
1160 if (vtop >= vstack && (vtop->r & VT_VALMASK) == VT_CMP)
1161 gv(RC_INT);
1163 stack_adjust = 16 - stack_adjust;
1164 o(0x48);
1165 oad(0xec81, stack_adjust); /* sub $xxx, %rsp */
1166 args_size += stack_adjust;
1169 for(i = run_start; i < run_end;) {
1170 /* Swap argument to top, it will possibly be changed here,
1171 and might use more temps. At the end of the loop we keep
1172 in on the stack and swap it back to its original position
1173 if it is a register. */
1174 SValue tmp = vtop[0];
1175 vtop[0] = vtop[-i];
1176 vtop[-i] = tmp;
1178 mode = classify_x86_64_arg(&vtop->type, NULL, &size, &align, &reg_count);
1180 int arg_stored = 1;
1181 switch (vtop->type.t & VT_BTYPE) {
1182 case VT_STRUCT:
1183 if (mode == x86_64_mode_sse) {
1184 if (sse_reg > 8)
1185 sse_reg -= reg_count;
1186 else
1187 arg_stored = 0;
1188 } else if (mode == x86_64_mode_integer) {
1189 if (gen_reg > REGN)
1190 gen_reg -= reg_count;
1191 else
1192 arg_stored = 0;
1195 if (arg_stored) {
1196 /* allocate the necessary size on stack */
1197 o(0x48);
1198 oad(0xec81, size); /* sub $xxx, %rsp */
1199 /* generate structure store */
1200 r = get_reg(RC_INT);
1201 orex(1, r, 0, 0x89); /* mov %rsp, r */
1202 o(0xe0 + REG_VALUE(r));
1203 vset(&vtop->type, r | VT_LVAL, 0);
1204 vswap();
1205 vstore();
1206 args_size += size;
1208 break;
1210 case VT_LDOUBLE:
1211 assert(0);
1212 break;
1214 case VT_FLOAT:
1215 case VT_DOUBLE:
1216 assert(mode == x86_64_mode_sse);
1217 if (sse_reg > 8) {
1218 --sse_reg;
1219 r = gv(RC_FLOAT);
1220 o(0x50); /* push $rax */
1221 /* movq %xmmN, (%rsp) */
1222 o(0xd60f66);
1223 o(0x04 + REG_VALUE(r)*8);
1224 o(0x24);
1225 args_size += size;
1226 } else {
1227 arg_stored = 0;
1229 break;
1231 default:
1232 assert(mode == x86_64_mode_integer);
1233 /* simple type */
1234 /* XXX: implicit cast ? */
1235 if (gen_reg > REGN) {
1236 --gen_reg;
1237 r = gv(RC_INT);
1238 orex(0,r,0,0x50 + REG_VALUE(r)); /* push r */
1239 args_size += size;
1240 } else {
1241 arg_stored = 0;
1243 break;
1246 /* And swap the argument back to it's original position. */
1247 tmp = vtop[0];
1248 vtop[0] = vtop[-i];
1249 vtop[-i] = tmp;
1251 if (arg_stored) {
1252 vrotb(i+1);
1253 assert((vtop->type.t == tmp.type.t) && (vtop->r == tmp.r));
1254 vpop();
1255 --nb_args;
1256 --run_end;
1257 } else {
1258 ++i;
1262 /* handle 16 byte aligned arguments at end of run */
1263 run_start = i = run_end;
1264 while (i < nb_args) {
1265 /* Rotate argument to top since it will always be popped */
1266 mode = classify_x86_64_arg(&vtop[-i].type, NULL, &size, &align, &reg_count);
1267 if (align != 16)
1268 break;
1270 vrotb(i+1);
1272 if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
1273 gv(RC_ST0);
1274 oad(0xec8148, size); /* sub $xxx, %rsp */
1275 o(0x7cdb); /* fstpt 0(%rsp) */
1276 g(0x24);
1277 g(0x00);
1278 args_size += size;
1279 } else {
1280 //assert(mode == x86_64_mode_memory);
1282 /* allocate the necessary size on stack */
1283 o(0x48);
1284 oad(0xec81, size); /* sub $xxx, %rsp */
1285 /* generate structure store */
1286 r = get_reg(RC_INT);
1287 orex(1, r, 0, 0x89); /* mov %rsp, r */
1288 o(0xe0 + REG_VALUE(r));
1289 vset(&vtop->type, r | VT_LVAL, 0);
1290 vswap();
1291 vstore();
1292 args_size += size;
1295 vpop();
1296 --nb_args;
1300 /* XXX This should be superfluous. */
1301 save_regs(0); /* save used temporary registers */
1303 /* then, we prepare register passing arguments.
1304 Note that we cannot set RDX and RCX in this loop because gv()
1305 may break these temporary registers. Let's use R10 and R11
1306 instead of them */
1307 assert(gen_reg <= REGN);
1308 assert(sse_reg <= 8);
1309 for(i = 0; i < nb_args; i++) {
1310 mode = classify_x86_64_arg(&vtop->type, &type, &size, &align, &reg_count);
1311 /* Alter stack entry type so that gv() knows how to treat it */
1312 vtop->type = type;
1313 if (mode == x86_64_mode_sse) {
1314 if (reg_count == 2) {
1315 sse_reg -= 2;
1316 gv(RC_FRET); /* Use pair load into xmm0 & xmm1 */
1317 if (sse_reg) { /* avoid redundant movaps %xmm0, %xmm0 */
1318 /* movaps %xmm0, %xmmN */
1319 o(0x280f);
1320 o(0xc0 + (sse_reg << 3));
1321 /* movaps %xmm1, %xmmN */
1322 o(0x280f);
1323 o(0xc1 + ((sse_reg+1) << 3));
1325 } else {
1326 assert(reg_count == 1);
1327 --sse_reg;
1328 /* Load directly to register */
1329 gv(RC_XMM0 << sse_reg);
1331 } else if (mode == x86_64_mode_integer) {
1332 /* simple type */
1333 /* XXX: implicit cast ? */
1334 gen_reg -= reg_count;
1335 r = gv(RC_INT);
1336 int d = arg_prepare_reg(gen_reg);
1337 orex(1,d,r,0x89); /* mov */
1338 o(0xc0 + REG_VALUE(r) * 8 + REG_VALUE(d));
1339 if (reg_count == 2) {
1340 d = arg_prepare_reg(gen_reg+1);
1341 orex(1,d,vtop->r2,0x89); /* mov */
1342 o(0xc0 + REG_VALUE(vtop->r2) * 8 + REG_VALUE(d));
1345 vtop--;
1347 assert(gen_reg == 0);
1348 assert(sse_reg == 0);
1350 /* We shouldn't have many operands on the stack anymore, but the
1351 call address itself is still there, and it might be in %eax
1352 (or edx/ecx) currently, which the below writes would clobber.
1353 So evict all remaining operands here. */
1354 save_regs(0);
1356 /* Copy R10 and R11 into RDX and RCX, respectively */
1357 if (nb_reg_args > 2) {
1358 o(0xd2894c); /* mov %r10, %rdx */
1359 if (nb_reg_args > 3) {
1360 o(0xd9894c); /* mov %r11, %rcx */
1364 oad(0xb8, nb_sse_args < 8 ? nb_sse_args : 8); /* mov nb_sse_args, %eax */
1365 gcall_or_jmp(0);
1366 if (args_size)
1367 gadd_sp(args_size);
1368 vtop--;
1372 #define FUNC_PROLOG_SIZE 11
1374 static void push_arg_reg(int i) {
1375 loc -= 8;
1376 gen_modrm64(0x89, arg_regs[i], VT_LOCAL, NULL, loc);
1379 /* generate function prolog of type 't' */
1380 void gfunc_prolog(CType *func_type)
1382 X86_64_Mode mode;
1383 int i, addr, align, size, reg_count;
1384 int param_addr = 0, reg_param_index, sse_param_index;
1385 Sym *sym;
1386 CType *type;
1388 sym = func_type->ref;
1389 addr = PTR_SIZE * 2;
1390 loc = 0;
1391 ind += FUNC_PROLOG_SIZE;
1392 func_sub_sp_offset = ind;
1393 func_ret_sub = 0;
1395 if (func_type->ref->c == FUNC_ELLIPSIS) {
1396 int seen_reg_num, seen_sse_num, seen_stack_size;
1397 seen_reg_num = seen_sse_num = 0;
1398 /* frame pointer and return address */
1399 seen_stack_size = PTR_SIZE * 2;
1400 /* count the number of seen parameters */
1401 sym = func_type->ref;
1402 while ((sym = sym->next) != NULL) {
1403 type = &sym->type;
1404 mode = classify_x86_64_arg(type, NULL, &size, &align, &reg_count);
1405 switch (mode) {
1406 default:
1407 stack_arg:
1408 seen_stack_size = ((seen_stack_size + align - 1) & -align) + size;
1409 break;
1411 case x86_64_mode_integer:
1412 if (seen_reg_num + reg_count <= 8) {
1413 seen_reg_num += reg_count;
1414 } else {
1415 seen_reg_num = 8;
1416 goto stack_arg;
1418 break;
1420 case x86_64_mode_sse:
1421 if (seen_sse_num + reg_count <= 8) {
1422 seen_sse_num += reg_count;
1423 } else {
1424 seen_sse_num = 8;
1425 goto stack_arg;
1427 break;
1431 loc -= 16;
1432 /* movl $0x????????, -0x10(%rbp) */
1433 o(0xf045c7);
1434 gen_le32(seen_reg_num * 8);
1435 /* movl $0x????????, -0xc(%rbp) */
1436 o(0xf445c7);
1437 gen_le32(seen_sse_num * 16 + 48);
1438 /* movl $0x????????, -0x8(%rbp) */
1439 o(0xf845c7);
1440 gen_le32(seen_stack_size);
1442 /* save all register passing arguments */
1443 for (i = 0; i < 8; i++) {
1444 loc -= 16;
1445 o(0xd60f66); /* movq */
1446 gen_modrm(7 - i, VT_LOCAL, NULL, loc);
1447 /* movq $0, loc+8(%rbp) */
1448 o(0x85c748);
1449 gen_le32(loc + 8);
1450 gen_le32(0);
1452 for (i = 0; i < REGN; i++) {
1453 push_arg_reg(REGN-1-i);
1457 sym = func_type->ref;
1458 reg_param_index = 0;
1459 sse_param_index = 0;
1461 /* if the function returns a structure, then add an
1462 implicit pointer parameter */
1463 func_vt = sym->type;
1464 mode = classify_x86_64_arg(&func_vt, NULL, &size, &align, &reg_count);
1465 if (mode == x86_64_mode_memory) {
1466 push_arg_reg(reg_param_index);
1467 func_vc = loc;
1468 reg_param_index++;
1470 /* define parameters */
1471 while ((sym = sym->next) != NULL) {
1472 type = &sym->type;
1473 mode = classify_x86_64_arg(type, NULL, &size, &align, &reg_count);
1474 switch (mode) {
1475 case x86_64_mode_sse:
1476 if (sse_param_index + reg_count <= 8) {
1477 /* save arguments passed by register */
1478 loc -= reg_count * 8;
1479 param_addr = loc;
1480 for (i = 0; i < reg_count; ++i) {
1481 o(0xd60f66); /* movq */
1482 gen_modrm(sse_param_index, VT_LOCAL, NULL, param_addr + i*8);
1483 ++sse_param_index;
1485 } else {
1486 addr = (addr + align - 1) & -align;
1487 param_addr = addr;
1488 addr += size;
1489 sse_param_index += reg_count;
1491 break;
1493 case x86_64_mode_memory:
1494 case x86_64_mode_x87:
1495 addr = (addr + align - 1) & -align;
1496 param_addr = addr;
1497 addr += size;
1498 break;
1500 case x86_64_mode_integer: {
1501 if (reg_param_index + reg_count <= REGN) {
1502 /* save arguments passed by register */
1503 loc -= reg_count * 8;
1504 param_addr = loc;
1505 for (i = 0; i < reg_count; ++i) {
1506 gen_modrm64(0x89, arg_regs[reg_param_index], VT_LOCAL, NULL, param_addr + i*8);
1507 ++reg_param_index;
1509 } else {
1510 addr = (addr + align - 1) & -align;
1511 param_addr = addr;
1512 addr += size;
1513 reg_param_index += reg_count;
1515 break;
1517 default: break; /* nothing to be done for x86_64_mode_none */
1519 sym_push(sym->v & ~SYM_FIELD, type,
1520 VT_LOCAL | VT_LVAL, param_addr);
1524 /* generate function epilog */
1525 void gfunc_epilog(void)
1527 int v, saved_ind;
1529 o(0xc9); /* leave */
1530 if (func_ret_sub == 0) {
1531 o(0xc3); /* ret */
1532 } else {
1533 o(0xc2); /* ret n */
1534 g(func_ret_sub);
1535 g(func_ret_sub >> 8);
1537 /* align local size to word & save local variables */
1538 v = (-loc + 15) & -16;
1539 saved_ind = ind;
1540 ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
1541 o(0xe5894855); /* push %rbp, mov %rsp, %rbp */
1542 o(0xec8148); /* sub rsp, stacksize */
1543 gen_le32(v);
1544 ind = saved_ind;
1547 #endif /* not PE */
1549 /* generate a jump to a label */
1550 int gjmp(int t)
1552 return psym(0xe9, t);
1555 /* generate a jump to a fixed address */
1556 void gjmp_addr(int a)
1558 int r;
1559 r = a - ind - 2;
1560 if (r == (char)r) {
1561 g(0xeb);
1562 g(r);
1563 } else {
1564 oad(0xe9, a - ind - 5);
1568 /* generate a test. set 'inv' to invert test. Stack entry is popped */
1569 int gtst(int inv, int t)
1571 int v, *p;
1573 v = vtop->r & VT_VALMASK;
1574 if (v == VT_CMP) {
1575 /* fast case : can jump directly since flags are set */
1576 if (vtop->c.i & 0x100)
1578 /* This was a float compare. If the parity flag is set
1579 the result was unordered. For anything except != this
1580 means false and we don't jump (anding both conditions).
1581 For != this means true (oring both).
1582 Take care about inverting the test. We need to jump
1583 to our target if the result was unordered and test wasn't NE,
1584 otherwise if unordered we don't want to jump. */
1585 vtop->c.i &= ~0x100;
1586 if (!inv == (vtop->c.i != TOK_NE))
1587 o(0x067a); /* jp +6 */
1588 else
1590 g(0x0f);
1591 t = psym(0x8a, t); /* jp t */
1594 g(0x0f);
1595 t = psym((vtop->c.i - 16) ^ inv, t);
1596 } else if (v == VT_JMP || v == VT_JMPI) {
1597 /* && or || optimization */
1598 if ((v & 1) == inv) {
1599 /* insert vtop->c jump list in t */
1600 p = &vtop->c.i;
1601 while (*p != 0)
1602 p = (int *)(cur_text_section->data + *p);
1603 *p = t;
1604 t = vtop->c.i;
1605 } else {
1606 t = gjmp(t);
1607 gsym(vtop->c.i);
1609 } else {
1610 if (is_float(vtop->type.t) ||
1611 (vtop->type.t & VT_BTYPE) == VT_LLONG) {
1612 vpushi(0);
1613 gen_op(TOK_NE);
1615 if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1616 /* constant jmp optimization */
1617 if ((vtop->c.i != 0) != inv)
1618 t = gjmp(t);
1619 } else {
1620 v = gv(RC_INT);
1621 orex(0,v,v,0x85);
1622 o(0xc0 + REG_VALUE(v) * 9);
1623 g(0x0f);
1624 t = psym(0x85 ^ inv, t);
1627 vtop--;
1628 return t;
1631 /* generate an integer binary operation */
1632 void gen_opi(int op)
1634 int r, fr, opc, fc, c, ll, uu, cc, tt2;
1636 fr = vtop[0].r;
1637 fc = vtop->c.ul;
1638 ll = is64_type(vtop[-1].type.t);
1639 cc = (fr & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST;
1640 tt2 = (fr & (VT_LVAL | VT_LVAL_TYPE)) == VT_LVAL;
1642 switch(op) {
1643 case '+':
1644 case TOK_ADDC1: /* add with carry generation */
1645 opc = 0;
1646 gen_op8:
1647 vswap();
1648 r = gv(RC_INT);
1649 vswap();
1650 if (cc && (!ll || (int)vtop->c.ll == vtop->c.ll)) {
1651 /* constant case */
1652 c = vtop->c.i;
1653 if (c == (char)c) {
1654 /* XXX: generate inc and dec for smaller code ? */
1655 orex(ll, r, 0, 0x83);
1656 o(0xc0 + REG_VALUE(r) + opc*8);
1657 g(c);
1658 } else {
1659 orex(ll, r, 0, 0x81);
1660 oad(0xc0 + REG_VALUE(r) + opc*8, c);
1662 } else {
1663 if(!tt2)
1664 fr = gv(RC_INT);
1665 orex(ll, fr, r, 0x03 + opc*8);
1666 if(fr >= VT_CONST)
1667 gen_modrm(r, fr, vtop->sym, fc);
1668 else
1669 o(0xc0 + REG_VALUE(fr) + REG_VALUE(r)*8);
1671 vtop--;
1672 if (op >= TOK_ULT && op <= TOK_GT) {
1673 vtop->r = VT_CMP;
1674 vtop->c.i = op;
1676 break;
1677 case '-':
1678 case TOK_SUBC1: /* sub with carry generation */
1679 opc = 5;
1680 goto gen_op8;
1681 case TOK_ADDC2: /* add with carry use */
1682 opc = 2;
1683 goto gen_op8;
1684 case TOK_SUBC2: /* sub with carry use */
1685 opc = 3;
1686 goto gen_op8;
1687 case '&':
1688 opc = 4;
1689 goto gen_op8;
1690 case '^':
1691 opc = 6;
1692 goto gen_op8;
1693 case '|':
1694 opc = 1;
1695 goto gen_op8;
1696 case '*':
1697 opc = 5;
1698 vswap();
1699 r = gv(RC_INT);
1700 vswap();
1701 if(!tt2)
1702 fr = gv(RC_INT);
1703 if(r == TREG_RAX){
1704 if(fr != TREG_RDX)
1705 save_reg(TREG_RDX);
1706 orex(ll, fr, r, 0xf7);
1707 if(fr >= VT_CONST)
1708 gen_modrm(opc, fr, vtop->sym, fc);
1709 else
1710 o(0xc0 + REG_VALUE(fr) + opc*8);
1711 }else{
1712 orex(ll, fr, r, 0xaf0f); /* imul fr, r */
1713 if(fr >= VT_CONST)
1714 gen_modrm(r, fr, vtop->sym, fc);
1715 else
1716 o(0xc0 + REG_VALUE(fr) + REG_VALUE(r)*8);
1718 vtop--;
1719 break;
1720 case TOK_SHL:
1721 opc = 4;
1722 goto gen_shift;
1723 case TOK_SHR:
1724 opc = 5;
1725 goto gen_shift;
1726 case TOK_SAR:
1727 opc = 7;
1728 gen_shift:
1729 if (cc) {
1730 /* constant case */
1731 vswap();
1732 r = gv(RC_INT);
1733 vswap();
1734 c = vtop->c.i;
1735 if(c == 1){
1736 orex(ll, r, 0, 0xd1);
1737 o(0xc0 + REG_VALUE(r) + opc*8);
1738 }else{
1739 orex(ll, r, 0, 0xc1); /* shl/shr/sar $xxx, r */
1740 o(0xc0 + REG_VALUE(r) + opc*8);
1741 g(c & (ll ? 0x3f : 0x1f));
1743 } else {
1744 /* we generate the shift in ecx */
1745 gv2(RC_INT, RC_RCX);
1746 r = vtop[-1].r;
1747 orex(ll, r, 0, 0xd3); /* shl/shr/sar %cl, r */
1748 o(0xc0 + REG_VALUE(r) + opc*8);
1750 vtop--;
1751 break;
1752 case TOK_UDIV:
1753 case TOK_UMOD:
1754 opc = 6;
1755 uu = 1;
1756 goto divmod;
1757 case '/':
1758 case '%':
1759 case TOK_PDIV:
1760 opc = 7;
1761 uu = 0;
1762 divmod:
1763 /* first operand must be in eax */
1764 /* XXX: need better constraint for second operand */
1765 if(!tt2){
1766 gv2(RC_RAX, RC_INT2);
1767 fr = vtop[0].r;
1768 }else{
1769 vswap();
1770 gv(RC_RAX);
1771 vswap();
1773 save_reg(TREG_RDX);
1774 orex(ll, 0, 0, uu ? 0xd231 : 0x99); /* xor %edx,%edx : cdq RDX:RAX <- sign-extend of RAX. */
1775 orex(ll, fr, 0, 0xf7); /* div fr, %eax */
1776 if(fr >= VT_CONST)
1777 gen_modrm(opc, fr, vtop->sym, fc);
1778 else
1779 o(0xc0 + REG_VALUE(fr) + opc*8);
1780 if (op == '%' || op == TOK_UMOD)
1781 r = TREG_RDX;
1782 else
1783 r = TREG_RAX;
1784 vtop--;
1785 vtop->r = r;
1786 break;
1787 default:
1788 opc = 7;
1789 goto gen_op8;
1793 void gen_opl(int op)
1795 gen_opi(op);
1798 /* generate a floating point operation 'v = t1 op t2' instruction. The
1799 two operands are guaranted to have the same floating point type */
1800 /* XXX: need to use ST1 too */
1801 void gen_opf(int op)
1803 int a, ft, fc, swapped, fr, r;
1804 int float_type = (vtop->type.t & VT_BTYPE) == VT_LDOUBLE ? RC_ST0 : RC_FLOAT;
1806 /* convert constants to memory references */
1807 if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
1808 vswap();
1809 gv(float_type);
1810 vswap();
1812 if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
1813 gv(float_type);
1815 swapped = 0;
1816 fc = vtop->c.ul;
1817 ft = vtop->type.t;
1819 if ((ft & VT_BTYPE) == VT_LDOUBLE) {
1820 /* swap the stack if needed so that t1 is the register and t2 is
1821 the memory reference */
1822 /* must put at least one value in the floating point register */
1823 if ((vtop[-1].r & VT_LVAL) && (vtop[0].r & VT_LVAL)) {
1824 vswap();
1825 gv(float_type);
1826 vswap();
1828 if (vtop[-1].r & VT_LVAL) {
1829 vswap();
1830 swapped = 1;
1832 if (op >= TOK_ULT && op <= TOK_GT) {
1833 /* load on stack second operand */
1834 load(TREG_ST0, vtop);
1835 save_reg(TREG_RAX); /* eax is used by FP comparison code */
1836 if (op == TOK_GE || op == TOK_GT)
1837 swapped = !swapped;
1838 else if (op == TOK_EQ || op == TOK_NE)
1839 swapped = 0;
1840 if (swapped)
1841 o(0xc9d9); /* fxch %st(1) */
1842 if (op == TOK_EQ || op == TOK_NE)
1843 o(0xe9da); /* fucompp */
1844 else
1845 o(0xd9de); /* fcompp */
1846 o(0xe0df); /* fnstsw %ax */
1847 if (op == TOK_EQ) {
1848 o(0x45e480); /* and $0x45, %ah */
1849 o(0x40fC80); /* cmp $0x40, %ah */
1850 } else if (op == TOK_NE) {
1851 o(0x45e480); /* and $0x45, %ah */
1852 o(0x40f480); /* xor $0x40, %ah */
1853 op = TOK_NE;
1854 } else if (op == TOK_GE || op == TOK_LE) {
1855 o(0x05c4f6); /* test $0x05, %ah */
1856 op = TOK_EQ;
1857 } else {
1858 o(0x45c4f6); /* test $0x45, %ah */
1859 op = TOK_EQ;
1861 vtop--;
1862 vtop->r = VT_CMP;
1863 vtop->c.i = op;
1864 } else {
1865 /* no memory reference possible for long double operations */
1866 load(TREG_ST0, vtop);
1867 swapped = !swapped;
1868 switch(op) {
1869 default:
1870 case '+':
1871 a = 0;
1872 break;
1873 case '-':
1874 a = 4;
1875 if (swapped)
1876 a++;
1877 break;
1878 case '*':
1879 a = 1;
1880 break;
1881 case '/':
1882 a = 6;
1883 if (swapped)
1884 a++;
1885 break;
1887 o(0xde); /* fxxxp %st, %st(1) */
1888 o(0xc1 + (a << 3));
1889 vtop--;
1891 } else {
1892 vswap();
1893 gv(float_type);
1894 vswap();
1895 fr = vtop->r;
1896 r = vtop[-1].r;
1897 if (op >= TOK_ULT && op <= TOK_GT) {
1898 switch(op){
1899 case TOK_LE:
1900 op = TOK_ULE; /* setae */
1901 break;
1902 case TOK_LT:
1903 op = TOK_ULT;
1904 break;
1905 case TOK_GE:
1906 op = TOK_UGE;
1907 break;
1908 case TOK_GT:
1909 op = TOK_UGT; /* seta */
1910 break;
1912 assert(!(vtop[-1].r & VT_LVAL));
1913 if ((ft & VT_BTYPE) == VT_DOUBLE)
1914 o(0x66);
1915 o(0x2e0f); /* ucomisd */
1916 if(fr >= VT_CONST)
1917 gen_modrm(r, fr, vtop->sym, fc);
1918 else
1919 o(0xc0 + REG_VALUE(fr) + REG_VALUE(r)*8);
1920 vtop--;
1921 vtop->r = VT_CMP;
1922 vtop->c.i = op | 0x100;
1923 } else {
1924 assert((vtop->type.t & VT_BTYPE) != VT_LDOUBLE);
1925 /* no memory reference possible for long double operations */
1926 switch(op) {
1927 default:
1928 case '+':
1929 a = 0;
1930 break;
1931 case '-':
1932 a = 4;
1933 break;
1934 case '*':
1935 a = 1;
1936 break;
1937 case '/':
1938 a = 6;
1939 break;
1941 assert((ft & VT_BTYPE) != VT_LDOUBLE);
1942 assert(!(vtop[-1].r & VT_LVAL));
1943 if ((ft & VT_BTYPE) == VT_DOUBLE) {
1944 o(0xf2);
1945 } else {
1946 o(0xf3);
1948 o(0x0f);
1949 o(0x58 + a);
1950 if(fr >= VT_CONST)
1951 gen_modrm(r, fr, vtop->sym, fc);
1952 else
1953 o(0xc0 + REG_VALUE(fr) + REG_VALUE(r)*8);
1954 vtop--;
1959 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1960 and 'long long' cases. */
1961 void gen_cvt_itof(int t)
1963 int ft, bt, tbt, r;
1965 ft = vtop->type.t;
1966 bt = ft & VT_BTYPE;
1967 tbt = t & VT_BTYPE;
1968 r = gv(RC_INT);
1970 if (tbt == VT_LDOUBLE) {
1971 save_reg(TREG_ST0);
1972 if ((ft & VT_BTYPE) == VT_LLONG) {
1973 /* signed long long to float/double/long double (unsigned case
1974 is handled generically) */
1975 o(0x50 + REG_VALUE(r)); /* push r */
1976 o(0x242cdf); /* fildll (%rsp) */
1977 o(0x08c48348); /* add $8, %rsp */
1978 } else if ((ft & (VT_BTYPE | VT_UNSIGNED)) == (VT_INT | VT_UNSIGNED)) {
1979 /* unsigned int to float/double/long double */
1980 o(0x6a); /* push $0 */
1981 g(0x00);
1982 o(0x50 + REG_VALUE(r)); /* push r */
1983 o(0x242cdf); /* fildll (%rsp) */
1984 o(0x10c48348); /* add $16, %rsp */
1985 } else {
1986 /* int to float/double/long double */
1987 o(0x50 + REG_VALUE(r)); /* push r */
1988 o(0x2404db); /* fildl (%rsp) */
1989 o(0x08c48348); /* add $8, %rsp */
1991 vtop->r = TREG_ST0;
1992 } else {
1993 int r_xmm;
1994 r_xmm = get_reg(RC_FLOAT);
1995 o(0xf2 + (tbt == VT_FLOAT));
1996 if ((ft & (VT_BTYPE | VT_UNSIGNED)) == (VT_INT | VT_UNSIGNED) || bt == VT_LLONG) {
1997 o(0x48); /* REX */
1999 o(0x2a0f);
2000 o(0xc0 + REG_VALUE(r) + REG_VALUE(r_xmm)*8); /* cvtsi2sd or cvtsi2ss */
2001 vtop->r = r_xmm;
2005 /* convert from one floating point type to another */
2006 void gen_cvt_ftof(int t)
2008 int ft, bt, tbt, r;
2010 ft = vtop->type.t;
2011 bt = ft & VT_BTYPE;
2012 tbt = t & VT_BTYPE;
2014 if(bt == VT_LDOUBLE)
2015 r = get_reg(RC_FLOAT);
2016 else
2017 r = gv(RC_FLOAT);
2018 if (bt == VT_FLOAT) {
2019 if (tbt == VT_DOUBLE) {
2020 o(0x5a0f); /* cvtps2pd */
2021 o(0xc0 + REG_VALUE(r) + REG_VALUE(r) * 8);
2022 } else if (tbt == VT_LDOUBLE) {
2023 /* movss %xmm0-7,-0x10(%rsp) */
2024 o(0x110ff3);
2025 o(0xf02444 + REG_VALUE(r)*8);
2026 o(0xf02444d9); /* flds -0x10(%rsp) */
2027 vtop->r = TREG_ST0;
2029 } else if (bt == VT_DOUBLE) {
2030 if (tbt == VT_FLOAT) {
2031 o(0x5a0f66); /* cvtpd2ps */
2032 o(0xc0 + REG_VALUE(r) + REG_VALUE(r) * 8);
2033 } else if (tbt == VT_LDOUBLE) {
2034 /* movsd %xmm0-7,-0x10(%rsp) */
2035 o(0x110ff2);
2036 o(0xf02444 + REG_VALUE(r)*8);
2037 o(0xf02444dd); /* fldl -0x10(%rsp) */
2038 vtop->r = TREG_ST0;
2040 } else {
2041 gv(RC_ST0);
2042 if (tbt == VT_DOUBLE) {
2043 o(0xf0245cdd); /* fstpl -0x10(%rsp) */
2044 /* movsd -0x10(%rsp),%xmm0-7 */
2045 o(0x100ff2);
2046 o(0xf02444 + REG_VALUE(r)*8);
2047 vtop->r = r;
2048 } else if (tbt == VT_FLOAT) {
2049 o(0xf0245cd9); /* fstps -0x10(%rsp) */
2050 /* movss -0x10(%rsp),%xmm0-7 */
2051 o(0x100ff3);
2052 o(0xf02444 + REG_VALUE(r)*8);
2053 vtop->r = r;
2058 /* convert fp to int 't' type */
2059 void gen_cvt_ftoi(int t)
2061 int ft, bt, ll, r, r_xmm;
2063 ft = vtop->type.t;
2064 bt = ft & VT_BTYPE;
2066 if (bt == VT_LDOUBLE) {
2067 gen_cvt_ftof(VT_DOUBLE);
2068 bt = VT_DOUBLE;
2070 r_xmm = gv(RC_FLOAT);
2071 if ((t & VT_BTYPE) == VT_INT)
2072 ll = 0;
2073 else
2074 ll = 1;
2075 r = get_reg(RC_INT);
2076 if (bt == VT_FLOAT) {
2077 o(0xf3);
2078 } else if (bt == VT_DOUBLE) {
2079 o(0xf2);
2080 } else {
2081 assert(0);
2083 orex(ll, r, r_xmm, 0x2c0f); /* cvttss2si or cvttsd2si */
2084 o(0xc0 + REG_VALUE(r_xmm) + (REG_VALUE(r) << 3));
2085 vtop->r = r;
2088 /* computed goto support */
2089 void ggoto(void)
2091 gcall_or_jmp(1);
2092 vtop--;
2095 /* Save the stack pointer onto the stack and return the location of its address */
2096 ST_FUNC void gen_vla_sp_save(int addr) {
2097 /* mov %rsp,addr(%rbp)*/
2098 gen_modrm64(0x89, TREG_RSP, VT_LOCAL, NULL, addr);
2101 /* Restore the SP from a location on the stack */
2102 ST_FUNC void gen_vla_sp_restore(int addr) {
2103 gen_modrm64(0x8b, TREG_RSP, VT_LOCAL, NULL, addr);
2106 /* Subtract from the stack pointer, and push the resulting value onto the stack */
2107 ST_FUNC void gen_vla_alloc(CType *type, int align) {
2108 #ifdef TCC_TARGET_PE
2109 /* alloca does more than just adjust %rsp on Windows */
2110 vpush_global_sym(&func_old_type, TOK_alloca);
2111 vswap(); /* Move alloca ref past allocation size */
2112 gfunc_call(1);
2113 vset(type, REG_IRET, 0);
2114 #else
2115 int r;
2116 r = gv(RC_INT); /* allocation size */
2117 /* sub r,%rsp */
2118 o(0x2b48);
2119 o(0xe0 | REG_VALUE(r));
2120 /* We align to 16 bytes rather than align */
2121 /* and ~15, %rsp */
2122 o(0xf0e48348);
2123 /* mov %rsp, r */
2124 o(0x8948);
2125 o(0xe0 | REG_VALUE(r));
2126 vpop();
2127 vset(type, r, 0);
2128 #endif
2132 /* end of x86-64 code generator */
2133 /*************************************************************/
2134 #endif /* ! TARGET_DEFS_ONLY */
2135 /******************************************************/