1 | /*
|
---|
2 | * x86-64 code generator for TCC
|
---|
3 | *
|
---|
4 | * Copyright (c) 2008 Shinichiro Hamaji
|
---|
5 | *
|
---|
6 | * Based on i386-gen.c by Fabrice Bellard
|
---|
7 | *
|
---|
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.
|
---|
12 | *
|
---|
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.
|
---|
17 | *
|
---|
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
|
---|
21 | */
|
---|
22 |
|
---|
23 | #ifdef TARGET_DEFS_ONLY
|
---|
24 |
|
---|
25 | /* number of available registers */
|
---|
26 | #define NB_REGS 5
|
---|
27 | #define NB_ASM_REGS 8
|
---|
28 |
|
---|
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_R8 0x0100
|
---|
38 | #define RC_R9 0x0200
|
---|
39 | #define RC_R10 0x0400
|
---|
40 | #define RC_R11 0x0800
|
---|
41 | #define RC_XMM0 0x0020
|
---|
42 | #define RC_ST0 0x0040 /* only for long double */
|
---|
43 | #define RC_IRET RC_RAX /* function return: integer register */
|
---|
44 | #define RC_LRET RC_RDX /* function return: second integer register */
|
---|
45 | #define RC_FRET RC_XMM0 /* function return: float register */
|
---|
46 |
|
---|
47 | /* pretty names for the registers */
|
---|
48 | enum {
|
---|
49 | TREG_RAX = 0,
|
---|
50 | TREG_RCX = 1,
|
---|
51 | TREG_RDX = 2,
|
---|
52 | TREG_XMM0 = 3,
|
---|
53 | TREG_ST0 = 4,
|
---|
54 |
|
---|
55 | TREG_RSI = 6,
|
---|
56 | TREG_RDI = 7,
|
---|
57 | TREG_R8 = 8,
|
---|
58 | TREG_R9 = 9,
|
---|
59 |
|
---|
60 | TREG_R10 = 10,
|
---|
61 | TREG_R11 = 11,
|
---|
62 |
|
---|
63 | TREG_MEM = 0x10,
|
---|
64 | };
|
---|
65 |
|
---|
66 | #define REX_BASE(reg) (((reg) >> 3) & 1)
|
---|
67 | #define REG_VALUE(reg) ((reg) & 7)
|
---|
68 |
|
---|
69 | /* return registers for function */
|
---|
70 | #define REG_IRET TREG_RAX /* single word int return register */
|
---|
71 | #define REG_LRET TREG_RDX /* second word return register (for long long) */
|
---|
72 | #define REG_FRET TREG_XMM0 /* float return register */
|
---|
73 |
|
---|
74 | /* defined if function parameters must be evaluated in reverse order */
|
---|
75 | #define INVERT_FUNC_PARAMS
|
---|
76 |
|
---|
77 | /* pointer size, in bytes */
|
---|
78 | #define PTR_SIZE 8
|
---|
79 |
|
---|
80 | /* long double size and alignment, in bytes */
|
---|
81 | #define LDOUBLE_SIZE 16
|
---|
82 | #define LDOUBLE_ALIGN 8
|
---|
83 | /* maximum alignment (for aligned attribute support) */
|
---|
84 | #define MAX_ALIGN 8
|
---|
85 |
|
---|
86 | /******************************************************/
|
---|
87 | /* ELF defines */
|
---|
88 |
|
---|
89 | #define EM_TCC_TARGET EM_X86_64
|
---|
90 |
|
---|
91 | /* relocation type for 32 bit data relocation */
|
---|
92 | #define R_DATA_32 R_X86_64_32
|
---|
93 | #define R_DATA_PTR R_X86_64_64
|
---|
94 | #define R_JMP_SLOT R_X86_64_JUMP_SLOT
|
---|
95 | #define R_COPY R_X86_64_COPY
|
---|
96 |
|
---|
97 | #define ELF_START_ADDR 0x08048000
|
---|
98 | #define ELF_PAGE_SIZE 0x1000
|
---|
99 |
|
---|
100 | /******************************************************/
|
---|
101 | #else /* ! TARGET_DEFS_ONLY */
|
---|
102 | /******************************************************/
|
---|
103 | #include "tcc.h"
|
---|
104 | #include <assert.h>
|
---|
105 |
|
---|
106 | ST_DATA const int reg_classes[NB_REGS+7] = {
|
---|
107 | /* eax */ RC_INT | RC_RAX,
|
---|
108 | /* ecx */ RC_INT | RC_RCX,
|
---|
109 | /* edx */ RC_INT | RC_RDX,
|
---|
110 | /* xmm0 */ RC_FLOAT | RC_XMM0,
|
---|
111 | /* st0 */ RC_ST0,
|
---|
112 | 0,
|
---|
113 | 0,
|
---|
114 | 0,
|
---|
115 | RC_INT | RC_R8,
|
---|
116 | RC_INT | RC_R9,
|
---|
117 | RC_INT | RC_R10,
|
---|
118 | RC_INT | RC_R11
|
---|
119 | };
|
---|
120 |
|
---|
121 | static unsigned long func_sub_sp_offset;
|
---|
122 | static int func_ret_sub;
|
---|
123 |
|
---|
124 | /* XXX: make it faster ? */
|
---|
125 | void g(int c)
|
---|
126 | {
|
---|
127 | int ind1;
|
---|
128 | ind1 = ind + 1;
|
---|
129 | if (ind1 > cur_text_section->data_allocated)
|
---|
130 | section_realloc(cur_text_section, ind1);
|
---|
131 | cur_text_section->data[ind] = c;
|
---|
132 | ind = ind1;
|
---|
133 | }
|
---|
134 |
|
---|
135 | void o(unsigned int c)
|
---|
136 | {
|
---|
137 | while (c) {
|
---|
138 | g(c);
|
---|
139 | c = c >> 8;
|
---|
140 | }
|
---|
141 | }
|
---|
142 |
|
---|
143 | void gen_le16(int v)
|
---|
144 | {
|
---|
145 | g(v);
|
---|
146 | g(v >> 8);
|
---|
147 | }
|
---|
148 |
|
---|
149 | void gen_le32(int c)
|
---|
150 | {
|
---|
151 | g(c);
|
---|
152 | g(c >> 8);
|
---|
153 | g(c >> 16);
|
---|
154 | g(c >> 24);
|
---|
155 | }
|
---|
156 |
|
---|
157 | void gen_le64(int64_t c)
|
---|
158 | {
|
---|
159 | g(c);
|
---|
160 | g(c >> 8);
|
---|
161 | g(c >> 16);
|
---|
162 | g(c >> 24);
|
---|
163 | g(c >> 32);
|
---|
164 | g(c >> 40);
|
---|
165 | g(c >> 48);
|
---|
166 | g(c >> 56);
|
---|
167 | }
|
---|
168 |
|
---|
169 | void orex(int ll, int r, int r2, int b)
|
---|
170 | {
|
---|
171 | if ((r & VT_VALMASK) >= VT_CONST)
|
---|
172 | r = 0;
|
---|
173 | if ((r2 & VT_VALMASK) >= VT_CONST)
|
---|
174 | r2 = 0;
|
---|
175 | if (ll || REX_BASE(r) || REX_BASE(r2))
|
---|
176 | o(0x40 | REX_BASE(r) | (REX_BASE(r2) << 2) | (ll << 3));
|
---|
177 | o(b);
|
---|
178 | }
|
---|
179 |
|
---|
180 | /* output a symbol and patch all calls to it */
|
---|
181 | void gsym_addr(int t, int a)
|
---|
182 | {
|
---|
183 | int n, *ptr;
|
---|
184 | while (t) {
|
---|
185 | ptr = (int *)(cur_text_section->data + t);
|
---|
186 | n = *ptr; /* next value */
|
---|
187 | *ptr = a - t - 4;
|
---|
188 | t = n;
|
---|
189 | }
|
---|
190 | }
|
---|
191 |
|
---|
192 | void gsym(int t)
|
---|
193 | {
|
---|
194 | gsym_addr(t, ind);
|
---|
195 | }
|
---|
196 |
|
---|
197 | /* psym is used to put an instruction with a data field which is a
|
---|
198 | reference to a symbol. It is in fact the same as oad ! */
|
---|
199 | #define psym oad
|
---|
200 |
|
---|
201 | static int is64_type(int t)
|
---|
202 | {
|
---|
203 | return ((t & VT_BTYPE) == VT_PTR ||
|
---|
204 | (t & VT_BTYPE) == VT_FUNC ||
|
---|
205 | (t & VT_BTYPE) == VT_LLONG);
|
---|
206 | }
|
---|
207 |
|
---|
208 | static int is_sse_float(int t) {
|
---|
209 | int bt;
|
---|
210 | bt = t & VT_BTYPE;
|
---|
211 | return bt == VT_DOUBLE || bt == VT_FLOAT;
|
---|
212 | }
|
---|
213 |
|
---|
214 |
|
---|
215 | /* instruction + 4 bytes data. Return the address of the data */
|
---|
216 | ST_FUNC int oad(int c, int s)
|
---|
217 | {
|
---|
218 | int ind1;
|
---|
219 |
|
---|
220 | o(c);
|
---|
221 | ind1 = ind + 4;
|
---|
222 | if (ind1 > cur_text_section->data_allocated)
|
---|
223 | section_realloc(cur_text_section, ind1);
|
---|
224 | *(int *)(cur_text_section->data + ind) = s;
|
---|
225 | s = ind;
|
---|
226 | ind = ind1;
|
---|
227 | return s;
|
---|
228 | }
|
---|
229 |
|
---|
230 | ST_FUNC void gen_addr32(int r, Sym *sym, int c)
|
---|
231 | {
|
---|
232 | if (r & VT_SYM)
|
---|
233 | greloc(cur_text_section, sym, ind, R_X86_64_32);
|
---|
234 | gen_le32(c);
|
---|
235 | }
|
---|
236 |
|
---|
237 | /* output constant with relocation if 'r & VT_SYM' is true */
|
---|
238 | ST_FUNC void gen_addr64(int r, Sym *sym, int64_t c)
|
---|
239 | {
|
---|
240 | if (r & VT_SYM)
|
---|
241 | greloc(cur_text_section, sym, ind, R_X86_64_64);
|
---|
242 | gen_le64(c);
|
---|
243 | }
|
---|
244 |
|
---|
245 | /* output constant with relocation if 'r & VT_SYM' is true */
|
---|
246 | ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
|
---|
247 | {
|
---|
248 | if (r & VT_SYM)
|
---|
249 | greloc(cur_text_section, sym, ind, R_X86_64_PC32);
|
---|
250 | gen_le32(c-4);
|
---|
251 | }
|
---|
252 |
|
---|
253 | /* output got address with relocation */
|
---|
254 | static void gen_gotpcrel(int r, Sym *sym, int c)
|
---|
255 | {
|
---|
256 | #ifndef TCC_TARGET_PE
|
---|
257 | Section *sr;
|
---|
258 | ElfW(Rela) *rel;
|
---|
259 | greloc(cur_text_section, sym, ind, R_X86_64_GOTPCREL);
|
---|
260 | sr = cur_text_section->reloc;
|
---|
261 | rel = (ElfW(Rela) *)(sr->data + sr->data_offset - sizeof(ElfW(Rela)));
|
---|
262 | rel->r_addend = -4;
|
---|
263 | #else
|
---|
264 | printf("picpic: %s %x %x | %02x %02x %02x\n", get_tok_str(sym->v, NULL), c, r,
|
---|
265 | cur_text_section->data[ind-3],
|
---|
266 | cur_text_section->data[ind-2],
|
---|
267 | cur_text_section->data[ind-1]
|
---|
268 | );
|
---|
269 | greloc(cur_text_section, sym, ind, R_X86_64_PC32);
|
---|
270 | #endif
|
---|
271 | gen_le32(0);
|
---|
272 | if (c) {
|
---|
273 | /* we use add c, %xxx for displacement */
|
---|
274 | orex(1, r, 0, 0x81);
|
---|
275 | o(0xc0 + REG_VALUE(r));
|
---|
276 | gen_le32(c);
|
---|
277 | }
|
---|
278 | }
|
---|
279 |
|
---|
280 | static void gen_modrm_impl(int op_reg, int r, Sym *sym, int c, int is_got)
|
---|
281 | {
|
---|
282 | op_reg = REG_VALUE(op_reg) << 3;
|
---|
283 | if ((r & VT_VALMASK) == VT_CONST) {
|
---|
284 | /* constant memory reference */
|
---|
285 | o(0x05 | op_reg);
|
---|
286 | if (is_got) {
|
---|
287 | gen_gotpcrel(r, sym, c);
|
---|
288 | } else {
|
---|
289 | gen_addrpc32(r, sym, c);
|
---|
290 | }
|
---|
291 | } else if ((r & VT_VALMASK) == VT_LOCAL) {
|
---|
292 | /* currently, we use only ebp as base */
|
---|
293 | if (c == (char)c) {
|
---|
294 | /* short reference */
|
---|
295 | o(0x45 | op_reg);
|
---|
296 | g(c);
|
---|
297 | } else {
|
---|
298 | oad(0x85 | op_reg, c);
|
---|
299 | }
|
---|
300 | } else if ((r & VT_VALMASK) >= TREG_MEM) {
|
---|
301 | if (c) {
|
---|
302 | g(0x80 | op_reg | REG_VALUE(r));
|
---|
303 | gen_le32(c);
|
---|
304 | } else {
|
---|
305 | g(0x00 | op_reg | REG_VALUE(r));
|
---|
306 | }
|
---|
307 | } else {
|
---|
308 | g(0x00 | op_reg | REG_VALUE(r));
|
---|
309 | }
|
---|
310 | }
|
---|
311 |
|
---|
312 | /* generate a modrm reference. 'op_reg' contains the addtionnal 3
|
---|
313 | opcode bits */
|
---|
314 | static void gen_modrm(int op_reg, int r, Sym *sym, int c)
|
---|
315 | {
|
---|
316 | gen_modrm_impl(op_reg, r, sym, c, 0);
|
---|
317 | }
|
---|
318 |
|
---|
319 | /* generate a modrm reference. 'op_reg' contains the addtionnal 3
|
---|
320 | opcode bits */
|
---|
321 | static void gen_modrm64(int opcode, int op_reg, int r, Sym *sym, int c)
|
---|
322 | {
|
---|
323 | int is_got;
|
---|
324 | is_got = (op_reg & TREG_MEM) && !(sym->type.t & VT_STATIC);
|
---|
325 | orex(1, r, op_reg, opcode);
|
---|
326 | gen_modrm_impl(op_reg, r, sym, c, is_got);
|
---|
327 | }
|
---|
328 |
|
---|
329 |
|
---|
330 | /* load 'r' from value 'sv' */
|
---|
331 | void load(int r, SValue *sv)
|
---|
332 | {
|
---|
333 | int v, t, ft, fc, fr;
|
---|
334 | SValue v1;
|
---|
335 |
|
---|
336 | #ifdef TCC_TARGET_PE
|
---|
337 | SValue v2;
|
---|
338 | sv = pe_getimport(sv, &v2);
|
---|
339 | #endif
|
---|
340 |
|
---|
341 | fr = sv->r;
|
---|
342 | ft = sv->type.t;
|
---|
343 | fc = sv->c.ul;
|
---|
344 |
|
---|
345 | #ifndef TCC_TARGET_PE
|
---|
346 | /* we use indirect access via got */
|
---|
347 | if ((fr & VT_VALMASK) == VT_CONST && (fr & VT_SYM) &&
|
---|
348 | (fr & VT_LVAL) && !(sv->sym->type.t & VT_STATIC)) {
|
---|
349 | /* use the result register as a temporal register */
|
---|
350 | int tr = r | TREG_MEM;
|
---|
351 | if (is_float(ft)) {
|
---|
352 | /* we cannot use float registers as a temporal register */
|
---|
353 | tr = get_reg(RC_INT) | TREG_MEM;
|
---|
354 | }
|
---|
355 | gen_modrm64(0x8b, tr, fr, sv->sym, 0);
|
---|
356 |
|
---|
357 | /* load from the temporal register */
|
---|
358 | fr = tr | VT_LVAL;
|
---|
359 | }
|
---|
360 | #endif
|
---|
361 |
|
---|
362 | v = fr & VT_VALMASK;
|
---|
363 | if (fr & VT_LVAL) {
|
---|
364 | int b, ll;
|
---|
365 | if (v == VT_LLOCAL) {
|
---|
366 | v1.type.t = VT_PTR;
|
---|
367 | v1.r = VT_LOCAL | VT_LVAL;
|
---|
368 | v1.c.ul = fc;
|
---|
369 | fr = r;
|
---|
370 | if (!(reg_classes[fr] & RC_INT))
|
---|
371 | fr = get_reg(RC_INT);
|
---|
372 | load(fr, &v1);
|
---|
373 | }
|
---|
374 | ll = 0;
|
---|
375 | if ((ft & VT_BTYPE) == VT_FLOAT) {
|
---|
376 | b = 0x6e0f66, r = 0; /* movd */
|
---|
377 | } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
|
---|
378 | b = 0x7e0ff3, r = 0; /* movq */
|
---|
379 | } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
|
---|
380 | b = 0xdb, r = 5; /* fldt */
|
---|
381 | } else if ((ft & VT_TYPE) == VT_BYTE) {
|
---|
382 | b = 0xbe0f; /* movsbl */
|
---|
383 | } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
|
---|
384 | b = 0xb60f; /* movzbl */
|
---|
385 | } else if ((ft & VT_TYPE) == VT_SHORT) {
|
---|
386 | b = 0xbf0f; /* movswl */
|
---|
387 | } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
|
---|
388 | b = 0xb70f; /* movzwl */
|
---|
389 | } else {
|
---|
390 | ll = is64_type(ft);
|
---|
391 | b = 0x8b;
|
---|
392 | }
|
---|
393 | if (ll) {
|
---|
394 | gen_modrm64(b, r, fr, sv->sym, fc);
|
---|
395 | } else {
|
---|
396 | orex(ll, fr, r, b);
|
---|
397 | gen_modrm(r, fr, sv->sym, fc);
|
---|
398 | }
|
---|
399 | } else {
|
---|
400 | if (v == VT_CONST) {
|
---|
401 | if (fr & VT_SYM) {
|
---|
402 | #ifdef TCC_TARGET_PE
|
---|
403 | orex(1,0,r,0x8d);
|
---|
404 | o(0x05 + REG_VALUE(r) * 8); /* lea xx(%rip), r */
|
---|
405 | gen_addrpc32(fr, sv->sym, fc);
|
---|
406 | #else
|
---|
407 | if (sv->sym->type.t & VT_STATIC) {
|
---|
408 | orex(1,0,r,0x8d);
|
---|
409 | o(0x05 + REG_VALUE(r) * 8); /* lea xx(%rip), r */
|
---|
410 | gen_addrpc32(fr, sv->sym, fc);
|
---|
411 | } else {
|
---|
412 | orex(1,0,r,0x8b);
|
---|
413 | o(0x05 + REG_VALUE(r) * 8); /* mov xx(%rip), r */
|
---|
414 | gen_gotpcrel(r, sv->sym, fc);
|
---|
415 | }
|
---|
416 | #endif
|
---|
417 | } else if (is64_type(ft)) {
|
---|
418 | orex(1,r,0, 0xb8 + REG_VALUE(r)); /* mov $xx, r */
|
---|
419 | gen_le64(sv->c.ull);
|
---|
420 | } else {
|
---|
421 | orex(0,r,0, 0xb8 + REG_VALUE(r)); /* mov $xx, r */
|
---|
422 | gen_le32(fc);
|
---|
423 | }
|
---|
424 | } else if (v == VT_LOCAL) {
|
---|
425 | orex(1,0,r,0x8d); /* lea xxx(%ebp), r */
|
---|
426 | gen_modrm(r, VT_LOCAL, sv->sym, fc);
|
---|
427 | } else if (v == VT_CMP) {
|
---|
428 | orex(0,r,0,0);
|
---|
429 | if ((fc & ~0x100) != TOK_NE)
|
---|
430 | oad(0xb8 + REG_VALUE(r), 0); /* mov $0, r */
|
---|
431 | else
|
---|
432 | oad(0xb8 + REG_VALUE(r), 1); /* mov $1, r */
|
---|
433 | if (fc & 0x100)
|
---|
434 | {
|
---|
435 | /* This was a float compare. If the parity bit is
|
---|
436 | set the result was unordered, meaning false for everything
|
---|
437 | except TOK_NE, and true for TOK_NE. */
|
---|
438 | fc &= ~0x100;
|
---|
439 | o(0x037a + (REX_BASE(r) << 8));
|
---|
440 | }
|
---|
441 | orex(0,r,0, 0x0f); /* setxx %br */
|
---|
442 | o(fc);
|
---|
443 | o(0xc0 + REG_VALUE(r));
|
---|
444 | } else if (v == VT_JMP || v == VT_JMPI) {
|
---|
445 | t = v & 1;
|
---|
446 | orex(0,r,0,0);
|
---|
447 | oad(0xb8 + REG_VALUE(r), t); /* mov $1, r */
|
---|
448 | o(0x05eb + (REX_BASE(r) << 8)); /* jmp after */
|
---|
449 | gsym(fc);
|
---|
450 | orex(0,r,0,0);
|
---|
451 | oad(0xb8 + REG_VALUE(r), t ^ 1); /* mov $0, r */
|
---|
452 | } else if (v != r) {
|
---|
453 | if (r == TREG_XMM0) {
|
---|
454 | assert(v == TREG_ST0);
|
---|
455 | /* gen_cvt_ftof(VT_DOUBLE); */
|
---|
456 | o(0xf0245cdd); /* fstpl -0x10(%rsp) */
|
---|
457 | /* movsd -0x10(%rsp),%xmm0 */
|
---|
458 | o(0x44100ff2);
|
---|
459 | o(0xf024);
|
---|
460 | } else if (r == TREG_ST0) {
|
---|
461 | assert(v == TREG_XMM0);
|
---|
462 | /* gen_cvt_ftof(VT_LDOUBLE); */
|
---|
463 | /* movsd %xmm0,-0x10(%rsp) */
|
---|
464 | o(0x44110ff2);
|
---|
465 | o(0xf024);
|
---|
466 | o(0xf02444dd); /* fldl -0x10(%rsp) */
|
---|
467 | } else {
|
---|
468 | orex(1,r,v, 0x89);
|
---|
469 | o(0xc0 + REG_VALUE(r) + REG_VALUE(v) * 8); /* mov v, r */
|
---|
470 | }
|
---|
471 | }
|
---|
472 | }
|
---|
473 | }
|
---|
474 |
|
---|
475 | /* store register 'r' in lvalue 'v' */
|
---|
476 | void store(int r, SValue *v)
|
---|
477 | {
|
---|
478 | int fr, bt, ft, fc;
|
---|
479 | int op64 = 0;
|
---|
480 | /* store the REX prefix in this variable when PIC is enabled */
|
---|
481 | int pic = 0;
|
---|
482 |
|
---|
483 | #ifdef TCC_TARGET_PE
|
---|
484 | SValue v2;
|
---|
485 | v = pe_getimport(v, &v2);
|
---|
486 | #endif
|
---|
487 |
|
---|
488 | ft = v->type.t;
|
---|
489 | fc = v->c.ul;
|
---|
490 | fr = v->r & VT_VALMASK;
|
---|
491 | bt = ft & VT_BTYPE;
|
---|
492 |
|
---|
493 | #ifndef TCC_TARGET_PE
|
---|
494 | /* we need to access the variable via got */
|
---|
495 | if (fr == VT_CONST && (v->r & VT_SYM)) {
|
---|
496 | /* mov xx(%rip), %r11 */
|
---|
497 | o(0x1d8b4c);
|
---|
498 | gen_gotpcrel(TREG_R11, v->sym, v->c.ul);
|
---|
499 | pic = is64_type(bt) ? 0x49 : 0x41;
|
---|
500 | }
|
---|
501 | #endif
|
---|
502 |
|
---|
503 | /* XXX: incorrect if float reg to reg */
|
---|
504 | if (bt == VT_FLOAT) {
|
---|
505 | o(0x66);
|
---|
506 | o(pic);
|
---|
507 | o(0x7e0f); /* movd */
|
---|
508 | r = 0;
|
---|
509 | } else if (bt == VT_DOUBLE) {
|
---|
510 | o(0x66);
|
---|
511 | o(pic);
|
---|
512 | o(0xd60f); /* movq */
|
---|
513 | r = 0;
|
---|
514 | } else if (bt == VT_LDOUBLE) {
|
---|
515 | o(0xc0d9); /* fld %st(0) */
|
---|
516 | o(pic);
|
---|
517 | o(0xdb); /* fstpt */
|
---|
518 | r = 7;
|
---|
519 | } else {
|
---|
520 | if (bt == VT_SHORT)
|
---|
521 | o(0x66);
|
---|
522 | o(pic);
|
---|
523 | if (bt == VT_BYTE || bt == VT_BOOL)
|
---|
524 | orex(0, 0, r, 0x88);
|
---|
525 | else if (is64_type(bt))
|
---|
526 | op64 = 0x89;
|
---|
527 | else
|
---|
528 | orex(0, 0, r, 0x89);
|
---|
529 | }
|
---|
530 | if (pic) {
|
---|
531 | /* xxx r, (%r11) where xxx is mov, movq, fld, or etc */
|
---|
532 | if (op64)
|
---|
533 | o(op64);
|
---|
534 | o(3 + (r << 3));
|
---|
535 | } else if (op64) {
|
---|
536 | if (fr == VT_CONST || fr == VT_LOCAL || (v->r & VT_LVAL)) {
|
---|
537 | gen_modrm64(op64, r, v->r, v->sym, fc);
|
---|
538 | } else if (fr != r) {
|
---|
539 | /* XXX: don't we really come here? */
|
---|
540 | abort();
|
---|
541 | o(0xc0 + fr + r * 8); /* mov r, fr */
|
---|
542 | }
|
---|
543 | } else {
|
---|
544 | if (fr == VT_CONST || fr == VT_LOCAL || (v->r & VT_LVAL)) {
|
---|
545 | gen_modrm(r, v->r, v->sym, fc);
|
---|
546 | } else if (fr != r) {
|
---|
547 | /* XXX: don't we really come here? */
|
---|
548 | abort();
|
---|
549 | o(0xc0 + fr + r * 8); /* mov r, fr */
|
---|
550 | }
|
---|
551 | }
|
---|
552 | }
|
---|
553 |
|
---|
554 | /* 'is_jmp' is '1' if it is a jump */
|
---|
555 | static void gcall_or_jmp(int is_jmp)
|
---|
556 | {
|
---|
557 | int r;
|
---|
558 | if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
|
---|
559 | /* constant case */
|
---|
560 | if (vtop->r & VT_SYM) {
|
---|
561 | /* relocation case */
|
---|
562 | greloc(cur_text_section, vtop->sym,
|
---|
563 | ind + 1, R_X86_64_PC32);
|
---|
564 | } else {
|
---|
565 | /* put an empty PC32 relocation */
|
---|
566 | put_elf_reloc(symtab_section, cur_text_section,
|
---|
567 | ind + 1, R_X86_64_PC32, 0);
|
---|
568 | }
|
---|
569 | oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
|
---|
570 | } else {
|
---|
571 | /* otherwise, indirect call */
|
---|
572 | r = TREG_R11;
|
---|
573 | load(r, vtop);
|
---|
574 | o(0x41); /* REX */
|
---|
575 | o(0xff); /* call/jmp *r */
|
---|
576 | o(0xd0 + REG_VALUE(r) + (is_jmp << 4));
|
---|
577 | }
|
---|
578 | }
|
---|
579 |
|
---|
580 | #ifdef TCC_TARGET_PE
|
---|
581 |
|
---|
582 | #define REGN 4
|
---|
583 | static const uint8_t arg_regs[] = {
|
---|
584 | TREG_RCX, TREG_RDX, TREG_R8, TREG_R9
|
---|
585 | };
|
---|
586 |
|
---|
587 | static int func_scratch;
|
---|
588 |
|
---|
589 | /* Generate function call. The function address is pushed first, then
|
---|
590 | all the parameters in call order. This functions pops all the
|
---|
591 | parameters and the function address. */
|
---|
592 |
|
---|
593 | void gen_offs_sp(int b, int r, int d)
|
---|
594 | {
|
---|
595 | orex(1,0,r & 0x100 ? 0 : r, b);
|
---|
596 | if (d == (char)d) {
|
---|
597 | o(0x2444 | (REG_VALUE(r) << 3));
|
---|
598 | g(d);
|
---|
599 | } else {
|
---|
600 | o(0x2484 | (REG_VALUE(r) << 3));
|
---|
601 | gen_le32(d);
|
---|
602 | }
|
---|
603 | }
|
---|
604 |
|
---|
605 | void gfunc_call(int nb_args)
|
---|
606 | {
|
---|
607 | int size, align, r, args_size, i, d, j, bt, struct_size;
|
---|
608 | int nb_reg_args, gen_reg;
|
---|
609 |
|
---|
610 | nb_reg_args = nb_args;
|
---|
611 | args_size = (nb_reg_args < REGN ? REGN : nb_reg_args) * PTR_SIZE;
|
---|
612 |
|
---|
613 | /* for struct arguments, we need to call memcpy and the function
|
---|
614 | call breaks register passing arguments we are preparing.
|
---|
615 | So, we process arguments which will be passed by stack first. */
|
---|
616 | struct_size = args_size;
|
---|
617 | for(i = 0; i < nb_args; i++) {
|
---|
618 | SValue *sv = &vtop[-i];
|
---|
619 | bt = (sv->type.t & VT_BTYPE);
|
---|
620 | if (bt == VT_STRUCT) {
|
---|
621 | size = type_size(&sv->type, &align);
|
---|
622 | /* align to stack align size */
|
---|
623 | size = (size + 15) & ~15;
|
---|
624 | /* generate structure store */
|
---|
625 | r = get_reg(RC_INT);
|
---|
626 | gen_offs_sp(0x8d, r, struct_size);
|
---|
627 | struct_size += size;
|
---|
628 |
|
---|
629 | /* generate memcpy call */
|
---|
630 | vset(&sv->type, r | VT_LVAL, 0);
|
---|
631 | vpushv(sv);
|
---|
632 | vstore();
|
---|
633 | --vtop;
|
---|
634 |
|
---|
635 | } else if (bt == VT_LDOUBLE) {
|
---|
636 |
|
---|
637 | gv(RC_ST0);
|
---|
638 | gen_offs_sp(0xdb, 0x107, struct_size);
|
---|
639 | struct_size += 16;
|
---|
640 |
|
---|
641 | }
|
---|
642 | }
|
---|
643 |
|
---|
644 | if (func_scratch < struct_size)
|
---|
645 | func_scratch = struct_size;
|
---|
646 | #if 1
|
---|
647 | for (i = 0; i < REGN; ++i)
|
---|
648 | save_reg(arg_regs[i]);
|
---|
649 | save_reg(TREG_RAX);
|
---|
650 | #endif
|
---|
651 | gen_reg = nb_reg_args;
|
---|
652 | struct_size = args_size;
|
---|
653 |
|
---|
654 | for(i = 0; i < nb_args; i++) {
|
---|
655 | bt = (vtop->type.t & VT_BTYPE);
|
---|
656 |
|
---|
657 | if (bt == VT_STRUCT || bt == VT_LDOUBLE) {
|
---|
658 | if (bt == VT_LDOUBLE)
|
---|
659 | size = 16;
|
---|
660 | else
|
---|
661 | size = type_size(&vtop->type, &align);
|
---|
662 | /* align to stack align size */
|
---|
663 | size = (size + 15) & ~15;
|
---|
664 | j = --gen_reg;
|
---|
665 | if (j >= REGN) {
|
---|
666 | d = TREG_RAX;
|
---|
667 | gen_offs_sp(0x8d, d, struct_size);
|
---|
668 | gen_offs_sp(0x89, d, j*8);
|
---|
669 | } else {
|
---|
670 | d = arg_regs[j];
|
---|
671 | gen_offs_sp(0x8d, d, struct_size);
|
---|
672 | }
|
---|
673 | struct_size += size;
|
---|
674 |
|
---|
675 | } else if (is_sse_float(vtop->type.t)) {
|
---|
676 | gv(RC_FLOAT); /* only one float register */
|
---|
677 | j = --gen_reg;
|
---|
678 | if (j >= REGN) {
|
---|
679 | /* movq %xmm0, j*8(%rsp) */
|
---|
680 | gen_offs_sp(0xd60f66, 0x100, j*8);
|
---|
681 | } else {
|
---|
682 | /* movaps %xmm0, %xmmN */
|
---|
683 | o(0x280f);
|
---|
684 | o(0xc0 + (j << 3));
|
---|
685 | d = arg_regs[j];
|
---|
686 | /* mov %xmm0, %rxx */
|
---|
687 | o(0x66);
|
---|
688 | orex(1,d,0, 0x7e0f);
|
---|
689 | o(0xc0 + REG_VALUE(d));
|
---|
690 | }
|
---|
691 | } else {
|
---|
692 | j = --gen_reg;
|
---|
693 | if (j >= REGN) {
|
---|
694 | r = gv(RC_INT);
|
---|
695 | gen_offs_sp(0x89, r, j*8);
|
---|
696 | } else {
|
---|
697 | d = arg_regs[j];
|
---|
698 | if (d < NB_REGS) {
|
---|
699 | gv(reg_classes[d] & ~RC_INT);
|
---|
700 | } else {
|
---|
701 | r = gv(RC_INT);
|
---|
702 | if (d != r) {
|
---|
703 | orex(1,d,r, 0x89);
|
---|
704 | o(0xc0 + REG_VALUE(d) + REG_VALUE(r) * 8);
|
---|
705 | }
|
---|
706 | }
|
---|
707 |
|
---|
708 | }
|
---|
709 | }
|
---|
710 | vtop--;
|
---|
711 | }
|
---|
712 | save_regs(0);
|
---|
713 | gcall_or_jmp(0);
|
---|
714 | vtop--;
|
---|
715 | }
|
---|
716 |
|
---|
717 |
|
---|
718 | #define FUNC_PROLOG_SIZE 11
|
---|
719 |
|
---|
720 | /* generate function prolog of type 't' */
|
---|
721 | void gfunc_prolog(CType *func_type)
|
---|
722 | {
|
---|
723 | int addr, reg_param_index, bt;
|
---|
724 | Sym *sym;
|
---|
725 | CType *type;
|
---|
726 |
|
---|
727 | func_ret_sub = 0;
|
---|
728 | func_scratch = 0;
|
---|
729 | loc = 0;
|
---|
730 |
|
---|
731 | addr = PTR_SIZE * 2;
|
---|
732 | ind += FUNC_PROLOG_SIZE;
|
---|
733 | func_sub_sp_offset = ind;
|
---|
734 | reg_param_index = 0;
|
---|
735 |
|
---|
736 | sym = func_type->ref;
|
---|
737 |
|
---|
738 | /* if the function returns a structure, then add an
|
---|
739 | implicit pointer parameter */
|
---|
740 | func_vt = sym->type;
|
---|
741 | if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
|
---|
742 | gen_modrm64(0x89, arg_regs[reg_param_index], VT_LOCAL, NULL, addr);
|
---|
743 | reg_param_index++;
|
---|
744 | addr += PTR_SIZE;
|
---|
745 | }
|
---|
746 |
|
---|
747 | /* define parameters */
|
---|
748 | while ((sym = sym->next) != NULL) {
|
---|
749 | type = &sym->type;
|
---|
750 | bt = type->t & VT_BTYPE;
|
---|
751 | if (reg_param_index < REGN) {
|
---|
752 | /* save arguments passed by register */
|
---|
753 | gen_modrm64(0x89, arg_regs[reg_param_index], VT_LOCAL, NULL, addr);
|
---|
754 | }
|
---|
755 | if (bt == VT_STRUCT || bt == VT_LDOUBLE) {
|
---|
756 | sym_push(sym->v & ~SYM_FIELD, type, VT_LOCAL | VT_LVAL | VT_REF, addr);
|
---|
757 | } else {
|
---|
758 | sym_push(sym->v & ~SYM_FIELD, type, VT_LOCAL | VT_LVAL, addr);
|
---|
759 | }
|
---|
760 | reg_param_index++;
|
---|
761 | addr += PTR_SIZE;
|
---|
762 | }
|
---|
763 |
|
---|
764 | while (reg_param_index < REGN) {
|
---|
765 | if (func_type->ref->c == FUNC_ELLIPSIS)
|
---|
766 | gen_modrm64(0x89, arg_regs[reg_param_index], VT_LOCAL, NULL, addr);
|
---|
767 | reg_param_index++;
|
---|
768 | addr += PTR_SIZE;
|
---|
769 | }
|
---|
770 | }
|
---|
771 |
|
---|
772 | /* generate function epilog */
|
---|
773 | void gfunc_epilog(void)
|
---|
774 | {
|
---|
775 | int v, saved_ind;
|
---|
776 |
|
---|
777 | o(0xc9); /* leave */
|
---|
778 | if (func_ret_sub == 0) {
|
---|
779 | o(0xc3); /* ret */
|
---|
780 | } else {
|
---|
781 | o(0xc2); /* ret n */
|
---|
782 | g(func_ret_sub);
|
---|
783 | g(func_ret_sub >> 8);
|
---|
784 | }
|
---|
785 |
|
---|
786 | saved_ind = ind;
|
---|
787 | ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
|
---|
788 | /* align local size to word & save local variables */
|
---|
789 | v = (func_scratch + -loc + 15) & -16;
|
---|
790 |
|
---|
791 | if (v >= 4096) {
|
---|
792 | Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
|
---|
793 | oad(0xb8, v); /* mov stacksize, %eax */
|
---|
794 | oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
|
---|
795 | greloc(cur_text_section, sym, ind-4, R_X86_64_PC32);
|
---|
796 | o(0x90); /* fill for FUNC_PROLOG_SIZE = 11 bytes */
|
---|
797 | } else {
|
---|
798 | o(0xe5894855); /* push %rbp, mov %rsp, %rbp */
|
---|
799 | o(0xec8148); /* sub rsp, stacksize */
|
---|
800 | gen_le32(v);
|
---|
801 | }
|
---|
802 |
|
---|
803 | cur_text_section->data_offset = saved_ind;
|
---|
804 | pe_add_unwind_data(ind, saved_ind, v);
|
---|
805 | ind = cur_text_section->data_offset;
|
---|
806 | }
|
---|
807 |
|
---|
808 | #else
|
---|
809 |
|
---|
810 | static void gadd_sp(int val)
|
---|
811 | {
|
---|
812 | if (val == (char)val) {
|
---|
813 | o(0xc48348);
|
---|
814 | g(val);
|
---|
815 | } else {
|
---|
816 | oad(0xc48148, val); /* add $xxx, %rsp */
|
---|
817 | }
|
---|
818 | }
|
---|
819 |
|
---|
820 | #define REGN 6
|
---|
821 | static const uint8_t arg_regs[REGN] = {
|
---|
822 | TREG_RDI, TREG_RSI, TREG_RDX, TREG_RCX, TREG_R8, TREG_R9
|
---|
823 | };
|
---|
824 |
|
---|
825 | /* Generate function call. The function address is pushed first, then
|
---|
826 | all the parameters in call order. This functions pops all the
|
---|
827 | parameters and the function address. */
|
---|
828 | void gfunc_call(int nb_args)
|
---|
829 | {
|
---|
830 | int size, align, r, args_size, i;
|
---|
831 | int nb_reg_args = 0;
|
---|
832 | int nb_sse_args = 0;
|
---|
833 | int sse_reg, gen_reg;
|
---|
834 |
|
---|
835 | /* calculate the number of integer/float arguments */
|
---|
836 | args_size = 0;
|
---|
837 | for(i = 0; i < nb_args; i++) {
|
---|
838 | if ((vtop[-i].type.t & VT_BTYPE) == VT_STRUCT) {
|
---|
839 | args_size += type_size(&vtop[-i].type, &align);
|
---|
840 | args_size = (args_size + 7) & ~7;
|
---|
841 | } else if ((vtop[-i].type.t & VT_BTYPE) == VT_LDOUBLE) {
|
---|
842 | args_size += 16;
|
---|
843 | } else if (is_sse_float(vtop[-i].type.t)) {
|
---|
844 | nb_sse_args++;
|
---|
845 | if (nb_sse_args > 8) args_size += 8;
|
---|
846 | } else {
|
---|
847 | nb_reg_args++;
|
---|
848 | if (nb_reg_args > REGN) args_size += 8;
|
---|
849 | }
|
---|
850 | }
|
---|
851 |
|
---|
852 | /* for struct arguments, we need to call memcpy and the function
|
---|
853 | call breaks register passing arguments we are preparing.
|
---|
854 | So, we process arguments which will be passed by stack first. */
|
---|
855 | gen_reg = nb_reg_args;
|
---|
856 | sse_reg = nb_sse_args;
|
---|
857 |
|
---|
858 | /* adjust stack to align SSE boundary */
|
---|
859 | if (args_size &= 15) {
|
---|
860 | /* fetch cpu flag before the following sub will change the value */
|
---|
861 | if (vtop >= vstack && (vtop->r & VT_VALMASK) == VT_CMP)
|
---|
862 | gv(RC_INT);
|
---|
863 |
|
---|
864 | args_size = 16 - args_size;
|
---|
865 | o(0x48);
|
---|
866 | oad(0xec81, args_size); /* sub $xxx, %rsp */
|
---|
867 | }
|
---|
868 |
|
---|
869 | for(i = 0; i < nb_args; i++) {
|
---|
870 | /* Swap argument to top, it will possibly be changed here,
|
---|
871 | and might use more temps. All arguments must remain on the
|
---|
872 | stack, so that get_reg can correctly evict some of them onto
|
---|
873 | stack. We could use also use a vrott(nb_args) at the end
|
---|
874 | of this loop, but this seems faster. */
|
---|
875 | SValue tmp = vtop[0];
|
---|
876 | vtop[0] = vtop[-i];
|
---|
877 | vtop[-i] = tmp;
|
---|
878 | if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
|
---|
879 | size = type_size(&vtop->type, &align);
|
---|
880 | /* align to stack align size */
|
---|
881 | size = (size + 7) & ~7;
|
---|
882 | /* allocate the necessary size on stack */
|
---|
883 | o(0x48);
|
---|
884 | oad(0xec81, size); /* sub $xxx, %rsp */
|
---|
885 | /* generate structure store */
|
---|
886 | r = get_reg(RC_INT);
|
---|
887 | orex(1, r, 0, 0x89); /* mov %rsp, r */
|
---|
888 | o(0xe0 + REG_VALUE(r));
|
---|
889 | vset(&vtop->type, r | VT_LVAL, 0);
|
---|
890 | vswap();
|
---|
891 | vstore();
|
---|
892 | args_size += size;
|
---|
893 | } else if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
|
---|
894 | gv(RC_ST0);
|
---|
895 | size = LDOUBLE_SIZE;
|
---|
896 | oad(0xec8148, size); /* sub $xxx, %rsp */
|
---|
897 | o(0x7cdb); /* fstpt 0(%rsp) */
|
---|
898 | g(0x24);
|
---|
899 | g(0x00);
|
---|
900 | args_size += size;
|
---|
901 | } else if (is_sse_float(vtop->type.t)) {
|
---|
902 | int j = --sse_reg;
|
---|
903 | if (j >= 8) {
|
---|
904 | gv(RC_FLOAT);
|
---|
905 | o(0x50); /* push $rax */
|
---|
906 | /* movq %xmm0, (%rsp) */
|
---|
907 | o(0x04d60f66);
|
---|
908 | o(0x24);
|
---|
909 | args_size += 8;
|
---|
910 | }
|
---|
911 | } else {
|
---|
912 | int j = --gen_reg;
|
---|
913 | /* simple type */
|
---|
914 | /* XXX: implicit cast ? */
|
---|
915 | if (j >= REGN) {
|
---|
916 | r = gv(RC_INT);
|
---|
917 | orex(0,r,0,0x50 + REG_VALUE(r)); /* push r */
|
---|
918 | args_size += 8;
|
---|
919 | }
|
---|
920 | }
|
---|
921 |
|
---|
922 | /* And swap the argument back to it's original position. */
|
---|
923 | tmp = vtop[0];
|
---|
924 | vtop[0] = vtop[-i];
|
---|
925 | vtop[-i] = tmp;
|
---|
926 | }
|
---|
927 |
|
---|
928 | /* XXX This should be superfluous. */
|
---|
929 | save_regs(0); /* save used temporary registers */
|
---|
930 |
|
---|
931 | /* then, we prepare register passing arguments.
|
---|
932 | Note that we cannot set RDX and RCX in this loop because gv()
|
---|
933 | may break these temporary registers. Let's use R10 and R11
|
---|
934 | instead of them */
|
---|
935 | gen_reg = nb_reg_args;
|
---|
936 | sse_reg = nb_sse_args;
|
---|
937 | for(i = 0; i < nb_args; i++) {
|
---|
938 | if ((vtop->type.t & VT_BTYPE) == VT_STRUCT ||
|
---|
939 | (vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
|
---|
940 | } else if (is_sse_float(vtop->type.t)) {
|
---|
941 | int j = --sse_reg;
|
---|
942 | if (j < 8) {
|
---|
943 | gv(RC_FLOAT); /* only one float register */
|
---|
944 | /* movaps %xmm0, %xmmN */
|
---|
945 | o(0x280f);
|
---|
946 | o(0xc0 + (sse_reg << 3));
|
---|
947 | }
|
---|
948 | } else {
|
---|
949 | int j = --gen_reg;
|
---|
950 | /* simple type */
|
---|
951 | /* XXX: implicit cast ? */
|
---|
952 | if (j < REGN) {
|
---|
953 | int d = arg_regs[j];
|
---|
954 | r = gv(RC_INT);
|
---|
955 | if (j == 2 || j == 3)
|
---|
956 | /* j=2: r10, j=3: r11 */
|
---|
957 | d = j + 8;
|
---|
958 | orex(1,d,r,0x89); /* mov */
|
---|
959 | o(0xc0 + REG_VALUE(r) * 8 + REG_VALUE(d));
|
---|
960 | }
|
---|
961 | }
|
---|
962 | vtop--;
|
---|
963 | }
|
---|
964 |
|
---|
965 | /* We shouldn't have many operands on the stack anymore, but the
|
---|
966 | call address itself is still there, and it might be in %eax
|
---|
967 | (or edx/ecx) currently, which the below writes would clobber.
|
---|
968 | So evict all remaining operands here. */
|
---|
969 | save_regs(0);
|
---|
970 |
|
---|
971 | /* Copy R10 and R11 into RDX and RCX, respectively */
|
---|
972 | if (nb_reg_args > 2) {
|
---|
973 | o(0xd2894c); /* mov %r10, %rdx */
|
---|
974 | if (nb_reg_args > 3) {
|
---|
975 | o(0xd9894c); /* mov %r11, %rcx */
|
---|
976 | }
|
---|
977 | }
|
---|
978 |
|
---|
979 | oad(0xb8, nb_sse_args < 8 ? nb_sse_args : 8); /* mov nb_sse_args, %eax */
|
---|
980 | gcall_or_jmp(0);
|
---|
981 | if (args_size)
|
---|
982 | gadd_sp(args_size);
|
---|
983 | vtop--;
|
---|
984 | }
|
---|
985 |
|
---|
986 |
|
---|
987 | #define FUNC_PROLOG_SIZE 11
|
---|
988 |
|
---|
989 | static void push_arg_reg(int i) {
|
---|
990 | loc -= 8;
|
---|
991 | gen_modrm64(0x89, arg_regs[i], VT_LOCAL, NULL, loc);
|
---|
992 | }
|
---|
993 |
|
---|
994 | /* generate function prolog of type 't' */
|
---|
995 | void gfunc_prolog(CType *func_type)
|
---|
996 | {
|
---|
997 | int i, addr, align, size;
|
---|
998 | int param_index, param_addr, reg_param_index, sse_param_index;
|
---|
999 | Sym *sym;
|
---|
1000 | CType *type;
|
---|
1001 |
|
---|
1002 | sym = func_type->ref;
|
---|
1003 | addr = PTR_SIZE * 2;
|
---|
1004 | loc = 0;
|
---|
1005 | ind += FUNC_PROLOG_SIZE;
|
---|
1006 | func_sub_sp_offset = ind;
|
---|
1007 | func_ret_sub = 0;
|
---|
1008 |
|
---|
1009 | if (func_type->ref->c == FUNC_ELLIPSIS) {
|
---|
1010 | int seen_reg_num, seen_sse_num, seen_stack_size;
|
---|
1011 | seen_reg_num = seen_sse_num = 0;
|
---|
1012 | /* frame pointer and return address */
|
---|
1013 | seen_stack_size = PTR_SIZE * 2;
|
---|
1014 | /* count the number of seen parameters */
|
---|
1015 | sym = func_type->ref;
|
---|
1016 | while ((sym = sym->next) != NULL) {
|
---|
1017 | type = &sym->type;
|
---|
1018 | if (is_sse_float(type->t)) {
|
---|
1019 | if (seen_sse_num < 8) {
|
---|
1020 | seen_sse_num++;
|
---|
1021 | } else {
|
---|
1022 | seen_stack_size += 8;
|
---|
1023 | }
|
---|
1024 | } else if ((type->t & VT_BTYPE) == VT_STRUCT) {
|
---|
1025 | size = type_size(type, &align);
|
---|
1026 | size = (size + 7) & ~7;
|
---|
1027 | seen_stack_size += size;
|
---|
1028 | } else if ((type->t & VT_BTYPE) == VT_LDOUBLE) {
|
---|
1029 | seen_stack_size += LDOUBLE_SIZE;
|
---|
1030 | } else {
|
---|
1031 | if (seen_reg_num < REGN) {
|
---|
1032 | seen_reg_num++;
|
---|
1033 | } else {
|
---|
1034 | seen_stack_size += 8;
|
---|
1035 | }
|
---|
1036 | }
|
---|
1037 | }
|
---|
1038 |
|
---|
1039 | loc -= 16;
|
---|
1040 | /* movl $0x????????, -0x10(%rbp) */
|
---|
1041 | o(0xf045c7);
|
---|
1042 | gen_le32(seen_reg_num * 8);
|
---|
1043 | /* movl $0x????????, -0xc(%rbp) */
|
---|
1044 | o(0xf445c7);
|
---|
1045 | gen_le32(seen_sse_num * 16 + 48);
|
---|
1046 | /* movl $0x????????, -0x8(%rbp) */
|
---|
1047 | o(0xf845c7);
|
---|
1048 | gen_le32(seen_stack_size);
|
---|
1049 |
|
---|
1050 | /* save all register passing arguments */
|
---|
1051 | for (i = 0; i < 8; i++) {
|
---|
1052 | loc -= 16;
|
---|
1053 | o(0xd60f66); /* movq */
|
---|
1054 | gen_modrm(7 - i, VT_LOCAL, NULL, loc);
|
---|
1055 | /* movq $0, loc+8(%rbp) */
|
---|
1056 | o(0x85c748);
|
---|
1057 | gen_le32(loc + 8);
|
---|
1058 | gen_le32(0);
|
---|
1059 | }
|
---|
1060 | for (i = 0; i < REGN; i++) {
|
---|
1061 | push_arg_reg(REGN-1-i);
|
---|
1062 | }
|
---|
1063 | }
|
---|
1064 |
|
---|
1065 | sym = func_type->ref;
|
---|
1066 | param_index = 0;
|
---|
1067 | reg_param_index = 0;
|
---|
1068 | sse_param_index = 0;
|
---|
1069 |
|
---|
1070 | /* if the function returns a structure, then add an
|
---|
1071 | implicit pointer parameter */
|
---|
1072 | func_vt = sym->type;
|
---|
1073 | if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
|
---|
1074 | push_arg_reg(reg_param_index);
|
---|
1075 | param_addr = loc;
|
---|
1076 |
|
---|
1077 | func_vc = loc;
|
---|
1078 | param_index++;
|
---|
1079 | reg_param_index++;
|
---|
1080 | }
|
---|
1081 | /* define parameters */
|
---|
1082 | while ((sym = sym->next) != NULL) {
|
---|
1083 | type = &sym->type;
|
---|
1084 | size = type_size(type, &align);
|
---|
1085 | size = (size + 7) & ~7;
|
---|
1086 | if (is_sse_float(type->t)) {
|
---|
1087 | if (sse_param_index < 8) {
|
---|
1088 | /* save arguments passed by register */
|
---|
1089 | loc -= 8;
|
---|
1090 | o(0xd60f66); /* movq */
|
---|
1091 | gen_modrm(sse_param_index, VT_LOCAL, NULL, loc);
|
---|
1092 | param_addr = loc;
|
---|
1093 | } else {
|
---|
1094 | param_addr = addr;
|
---|
1095 | addr += size;
|
---|
1096 | }
|
---|
1097 | sse_param_index++;
|
---|
1098 |
|
---|
1099 | } else if ((type->t & VT_BTYPE) == VT_STRUCT ||
|
---|
1100 | (type->t & VT_BTYPE) == VT_LDOUBLE) {
|
---|
1101 | param_addr = addr;
|
---|
1102 | addr += size;
|
---|
1103 | } else {
|
---|
1104 | if (reg_param_index < REGN) {
|
---|
1105 | /* save arguments passed by register */
|
---|
1106 | push_arg_reg(reg_param_index);
|
---|
1107 | param_addr = loc;
|
---|
1108 | } else {
|
---|
1109 | param_addr = addr;
|
---|
1110 | addr += 8;
|
---|
1111 | }
|
---|
1112 | reg_param_index++;
|
---|
1113 | }
|
---|
1114 | sym_push(sym->v & ~SYM_FIELD, type,
|
---|
1115 | VT_LOCAL | VT_LVAL, param_addr);
|
---|
1116 | param_index++;
|
---|
1117 | }
|
---|
1118 | }
|
---|
1119 |
|
---|
1120 | /* generate function epilog */
|
---|
1121 | void gfunc_epilog(void)
|
---|
1122 | {
|
---|
1123 | int v, saved_ind;
|
---|
1124 |
|
---|
1125 | o(0xc9); /* leave */
|
---|
1126 | if (func_ret_sub == 0) {
|
---|
1127 | o(0xc3); /* ret */
|
---|
1128 | } else {
|
---|
1129 | o(0xc2); /* ret n */
|
---|
1130 | g(func_ret_sub);
|
---|
1131 | g(func_ret_sub >> 8);
|
---|
1132 | }
|
---|
1133 | /* align local size to word & save local variables */
|
---|
1134 | v = (-loc + 15) & -16;
|
---|
1135 | saved_ind = ind;
|
---|
1136 | ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
|
---|
1137 | o(0xe5894855); /* push %rbp, mov %rsp, %rbp */
|
---|
1138 | o(0xec8148); /* sub rsp, stacksize */
|
---|
1139 | gen_le32(v);
|
---|
1140 | ind = saved_ind;
|
---|
1141 | }
|
---|
1142 |
|
---|
1143 | #endif /* not PE */
|
---|
1144 |
|
---|
1145 | /* generate a jump to a label */
|
---|
1146 | int gjmp(int t)
|
---|
1147 | {
|
---|
1148 | return psym(0xe9, t);
|
---|
1149 | }
|
---|
1150 |
|
---|
1151 | /* generate a jump to a fixed address */
|
---|
1152 | void gjmp_addr(int a)
|
---|
1153 | {
|
---|
1154 | int r;
|
---|
1155 | r = a - ind - 2;
|
---|
1156 | if (r == (char)r) {
|
---|
1157 | g(0xeb);
|
---|
1158 | g(r);
|
---|
1159 | } else {
|
---|
1160 | oad(0xe9, a - ind - 5);
|
---|
1161 | }
|
---|
1162 | }
|
---|
1163 |
|
---|
1164 | /* generate a test. set 'inv' to invert test. Stack entry is popped */
|
---|
1165 | int gtst(int inv, int t)
|
---|
1166 | {
|
---|
1167 | int v, *p;
|
---|
1168 |
|
---|
1169 | v = vtop->r & VT_VALMASK;
|
---|
1170 | if (v == VT_CMP) {
|
---|
1171 | /* fast case : can jump directly since flags are set */
|
---|
1172 | if (vtop->c.i & 0x100)
|
---|
1173 | {
|
---|
1174 | /* This was a float compare. If the parity flag is set
|
---|
1175 | the result was unordered. For anything except != this
|
---|
1176 | means false and we don't jump (anding both conditions).
|
---|
1177 | For != this means true (oring both).
|
---|
1178 | Take care about inverting the test. We need to jump
|
---|
1179 | to our target if the result was unordered and test wasn't NE,
|
---|
1180 | otherwise if unordered we don't want to jump. */
|
---|
1181 | vtop->c.i &= ~0x100;
|
---|
1182 | if (!inv == (vtop->c.i != TOK_NE))
|
---|
1183 | o(0x067a); /* jp +6 */
|
---|
1184 | else
|
---|
1185 | {
|
---|
1186 | g(0x0f);
|
---|
1187 | t = psym(0x8a, t); /* jp t */
|
---|
1188 | }
|
---|
1189 | }
|
---|
1190 | g(0x0f);
|
---|
1191 | t = psym((vtop->c.i - 16) ^ inv, t);
|
---|
1192 | } else if (v == VT_JMP || v == VT_JMPI) {
|
---|
1193 | /* && or || optimization */
|
---|
1194 | if ((v & 1) == inv) {
|
---|
1195 | /* insert vtop->c jump list in t */
|
---|
1196 | p = &vtop->c.i;
|
---|
1197 | while (*p != 0)
|
---|
1198 | p = (int *)(cur_text_section->data + *p);
|
---|
1199 | *p = t;
|
---|
1200 | t = vtop->c.i;
|
---|
1201 | } else {
|
---|
1202 | t = gjmp(t);
|
---|
1203 | gsym(vtop->c.i);
|
---|
1204 | }
|
---|
1205 | } else {
|
---|
1206 | if (is_float(vtop->type.t) ||
|
---|
1207 | (vtop->type.t & VT_BTYPE) == VT_LLONG) {
|
---|
1208 | vpushi(0);
|
---|
1209 | gen_op(TOK_NE);
|
---|
1210 | }
|
---|
1211 | if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
|
---|
1212 | /* constant jmp optimization */
|
---|
1213 | if ((vtop->c.i != 0) != inv)
|
---|
1214 | t = gjmp(t);
|
---|
1215 | } else {
|
---|
1216 | v = gv(RC_INT);
|
---|
1217 | orex(0,v,v,0x85);
|
---|
1218 | o(0xc0 + REG_VALUE(v) * 9);
|
---|
1219 | g(0x0f);
|
---|
1220 | t = psym(0x85 ^ inv, t);
|
---|
1221 | }
|
---|
1222 | }
|
---|
1223 | vtop--;
|
---|
1224 | return t;
|
---|
1225 | }
|
---|
1226 |
|
---|
1227 | /* generate an integer binary operation */
|
---|
1228 | void gen_opi(int op)
|
---|
1229 | {
|
---|
1230 | int r, fr, opc, c;
|
---|
1231 | int ll, uu, cc;
|
---|
1232 |
|
---|
1233 | ll = is64_type(vtop[-1].type.t);
|
---|
1234 | uu = (vtop[-1].type.t & VT_UNSIGNED) != 0;
|
---|
1235 | cc = (vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST;
|
---|
1236 |
|
---|
1237 | switch(op) {
|
---|
1238 | case '+':
|
---|
1239 | case TOK_ADDC1: /* add with carry generation */
|
---|
1240 | opc = 0;
|
---|
1241 | gen_op8:
|
---|
1242 | if (cc && (!ll || (int)vtop->c.ll == vtop->c.ll)) {
|
---|
1243 | /* constant case */
|
---|
1244 | vswap();
|
---|
1245 | r = gv(RC_INT);
|
---|
1246 | vswap();
|
---|
1247 | c = vtop->c.i;
|
---|
1248 | if (c == (char)c) {
|
---|
1249 | /* XXX: generate inc and dec for smaller code ? */
|
---|
1250 | orex(ll, r, 0, 0x83);
|
---|
1251 | o(0xc0 | (opc << 3) | REG_VALUE(r));
|
---|
1252 | g(c);
|
---|
1253 | } else {
|
---|
1254 | orex(ll, r, 0, 0x81);
|
---|
1255 | oad(0xc0 | (opc << 3) | REG_VALUE(r), c);
|
---|
1256 | }
|
---|
1257 | } else {
|
---|
1258 | gv2(RC_INT, RC_INT);
|
---|
1259 | r = vtop[-1].r;
|
---|
1260 | fr = vtop[0].r;
|
---|
1261 | orex(ll, r, fr, (opc << 3) | 0x01);
|
---|
1262 | o(0xc0 + REG_VALUE(r) + REG_VALUE(fr) * 8);
|
---|
1263 | }
|
---|
1264 | vtop--;
|
---|
1265 | if (op >= TOK_ULT && op <= TOK_GT) {
|
---|
1266 | vtop->r = VT_CMP;
|
---|
1267 | vtop->c.i = op;
|
---|
1268 | }
|
---|
1269 | break;
|
---|
1270 | case '-':
|
---|
1271 | case TOK_SUBC1: /* sub with carry generation */
|
---|
1272 | opc = 5;
|
---|
1273 | goto gen_op8;
|
---|
1274 | case TOK_ADDC2: /* add with carry use */
|
---|
1275 | opc = 2;
|
---|
1276 | goto gen_op8;
|
---|
1277 | case TOK_SUBC2: /* sub with carry use */
|
---|
1278 | opc = 3;
|
---|
1279 | goto gen_op8;
|
---|
1280 | case '&':
|
---|
1281 | opc = 4;
|
---|
1282 | goto gen_op8;
|
---|
1283 | case '^':
|
---|
1284 | opc = 6;
|
---|
1285 | goto gen_op8;
|
---|
1286 | case '|':
|
---|
1287 | opc = 1;
|
---|
1288 | goto gen_op8;
|
---|
1289 | case '*':
|
---|
1290 | gv2(RC_INT, RC_INT);
|
---|
1291 | r = vtop[-1].r;
|
---|
1292 | fr = vtop[0].r;
|
---|
1293 | orex(ll, fr, r, 0xaf0f); /* imul fr, r */
|
---|
1294 | o(0xc0 + REG_VALUE(fr) + REG_VALUE(r) * 8);
|
---|
1295 | vtop--;
|
---|
1296 | break;
|
---|
1297 | case TOK_SHL:
|
---|
1298 | opc = 4;
|
---|
1299 | goto gen_shift;
|
---|
1300 | case TOK_SHR:
|
---|
1301 | opc = 5;
|
---|
1302 | goto gen_shift;
|
---|
1303 | case TOK_SAR:
|
---|
1304 | opc = 7;
|
---|
1305 | gen_shift:
|
---|
1306 | opc = 0xc0 | (opc << 3);
|
---|
1307 | if (cc) {
|
---|
1308 | /* constant case */
|
---|
1309 | vswap();
|
---|
1310 | r = gv(RC_INT);
|
---|
1311 | vswap();
|
---|
1312 | orex(ll, r, 0, 0xc1); /* shl/shr/sar $xxx, r */
|
---|
1313 | o(opc | REG_VALUE(r));
|
---|
1314 | g(vtop->c.i & (ll ? 63 : 31));
|
---|
1315 | } else {
|
---|
1316 | /* we generate the shift in ecx */
|
---|
1317 | gv2(RC_INT, RC_RCX);
|
---|
1318 | r = vtop[-1].r;
|
---|
1319 | orex(ll, r, 0, 0xd3); /* shl/shr/sar %cl, r */
|
---|
1320 | o(opc | REG_VALUE(r));
|
---|
1321 | }
|
---|
1322 | vtop--;
|
---|
1323 | break;
|
---|
1324 | case TOK_UDIV:
|
---|
1325 | case TOK_UMOD:
|
---|
1326 | uu = 1;
|
---|
1327 | goto divmod;
|
---|
1328 | case '/':
|
---|
1329 | case '%':
|
---|
1330 | case TOK_PDIV:
|
---|
1331 | uu = 0;
|
---|
1332 | divmod:
|
---|
1333 | /* first operand must be in eax */
|
---|
1334 | /* XXX: need better constraint for second operand */
|
---|
1335 | gv2(RC_RAX, RC_RCX);
|
---|
1336 | r = vtop[-1].r;
|
---|
1337 | fr = vtop[0].r;
|
---|
1338 | vtop--;
|
---|
1339 | save_reg(TREG_RDX);
|
---|
1340 | orex(ll, 0, 0, uu ? 0xd231 : 0x99); /* xor %edx,%edx : cqto */
|
---|
1341 | orex(ll, fr, 0, 0xf7); /* div fr, %eax */
|
---|
1342 | o((uu ? 0xf0 : 0xf8) + REG_VALUE(fr));
|
---|
1343 | if (op == '%' || op == TOK_UMOD)
|
---|
1344 | r = TREG_RDX;
|
---|
1345 | else
|
---|
1346 | r = TREG_RAX;
|
---|
1347 | vtop->r = r;
|
---|
1348 | break;
|
---|
1349 | default:
|
---|
1350 | opc = 7;
|
---|
1351 | goto gen_op8;
|
---|
1352 | }
|
---|
1353 | }
|
---|
1354 |
|
---|
1355 | void gen_opl(int op)
|
---|
1356 | {
|
---|
1357 | gen_opi(op);
|
---|
1358 | }
|
---|
1359 |
|
---|
1360 | /* generate a floating point operation 'v = t1 op t2' instruction. The
|
---|
1361 | two operands are guaranted to have the same floating point type */
|
---|
1362 | /* XXX: need to use ST1 too */
|
---|
1363 | void gen_opf(int op)
|
---|
1364 | {
|
---|
1365 | int a, ft, fc, swapped, r;
|
---|
1366 | int float_type =
|
---|
1367 | (vtop->type.t & VT_BTYPE) == VT_LDOUBLE ? RC_ST0 : RC_FLOAT;
|
---|
1368 |
|
---|
1369 | /* convert constants to memory references */
|
---|
1370 | if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
|
---|
1371 | vswap();
|
---|
1372 | gv(float_type);
|
---|
1373 | vswap();
|
---|
1374 | }
|
---|
1375 | if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
|
---|
1376 | gv(float_type);
|
---|
1377 |
|
---|
1378 | /* must put at least one value in the floating point register */
|
---|
1379 | if ((vtop[-1].r & VT_LVAL) &&
|
---|
1380 | (vtop[0].r & VT_LVAL)) {
|
---|
1381 | vswap();
|
---|
1382 | gv(float_type);
|
---|
1383 | vswap();
|
---|
1384 | }
|
---|
1385 | swapped = 0;
|
---|
1386 | /* swap the stack if needed so that t1 is the register and t2 is
|
---|
1387 | the memory reference */
|
---|
1388 | if (vtop[-1].r & VT_LVAL) {
|
---|
1389 | vswap();
|
---|
1390 | swapped = 1;
|
---|
1391 | }
|
---|
1392 | if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
|
---|
1393 | if (op >= TOK_ULT && op <= TOK_GT) {
|
---|
1394 | /* load on stack second operand */
|
---|
1395 | load(TREG_ST0, vtop);
|
---|
1396 | save_reg(TREG_RAX); /* eax is used by FP comparison code */
|
---|
1397 | if (op == TOK_GE || op == TOK_GT)
|
---|
1398 | swapped = !swapped;
|
---|
1399 | else if (op == TOK_EQ || op == TOK_NE)
|
---|
1400 | swapped = 0;
|
---|
1401 | if (swapped)
|
---|
1402 | o(0xc9d9); /* fxch %st(1) */
|
---|
1403 | o(0xe9da); /* fucompp */
|
---|
1404 | o(0xe0df); /* fnstsw %ax */
|
---|
1405 | if (op == TOK_EQ) {
|
---|
1406 | o(0x45e480); /* and $0x45, %ah */
|
---|
1407 | o(0x40fC80); /* cmp $0x40, %ah */
|
---|
1408 | } else if (op == TOK_NE) {
|
---|
1409 | o(0x45e480); /* and $0x45, %ah */
|
---|
1410 | o(0x40f480); /* xor $0x40, %ah */
|
---|
1411 | op = TOK_NE;
|
---|
1412 | } else if (op == TOK_GE || op == TOK_LE) {
|
---|
1413 | o(0x05c4f6); /* test $0x05, %ah */
|
---|
1414 | op = TOK_EQ;
|
---|
1415 | } else {
|
---|
1416 | o(0x45c4f6); /* test $0x45, %ah */
|
---|
1417 | op = TOK_EQ;
|
---|
1418 | }
|
---|
1419 | vtop--;
|
---|
1420 | vtop->r = VT_CMP;
|
---|
1421 | vtop->c.i = op;
|
---|
1422 | } else {
|
---|
1423 | /* no memory reference possible for long double operations */
|
---|
1424 | load(TREG_ST0, vtop);
|
---|
1425 | swapped = !swapped;
|
---|
1426 |
|
---|
1427 | switch(op) {
|
---|
1428 | default:
|
---|
1429 | case '+':
|
---|
1430 | a = 0;
|
---|
1431 | break;
|
---|
1432 | case '-':
|
---|
1433 | a = 4;
|
---|
1434 | if (swapped)
|
---|
1435 | a++;
|
---|
1436 | break;
|
---|
1437 | case '*':
|
---|
1438 | a = 1;
|
---|
1439 | break;
|
---|
1440 | case '/':
|
---|
1441 | a = 6;
|
---|
1442 | if (swapped)
|
---|
1443 | a++;
|
---|
1444 | break;
|
---|
1445 | }
|
---|
1446 | ft = vtop->type.t;
|
---|
1447 | fc = vtop->c.ul;
|
---|
1448 | o(0xde); /* fxxxp %st, %st(1) */
|
---|
1449 | o(0xc1 + (a << 3));
|
---|
1450 | vtop--;
|
---|
1451 | }
|
---|
1452 | } else {
|
---|
1453 | if (op >= TOK_ULT && op <= TOK_GT) {
|
---|
1454 | /* if saved lvalue, then we must reload it */
|
---|
1455 | r = vtop->r;
|
---|
1456 | fc = vtop->c.ul;
|
---|
1457 | if ((r & VT_VALMASK) == VT_LLOCAL) {
|
---|
1458 | SValue v1;
|
---|
1459 | r = get_reg(RC_INT);
|
---|
1460 | v1.type.t = VT_PTR;
|
---|
1461 | v1.r = VT_LOCAL | VT_LVAL;
|
---|
1462 | v1.c.ul = fc;
|
---|
1463 | load(r, &v1);
|
---|
1464 | fc = 0;
|
---|
1465 | }
|
---|
1466 |
|
---|
1467 | if (op == TOK_EQ || op == TOK_NE) {
|
---|
1468 | swapped = 0;
|
---|
1469 | } else {
|
---|
1470 | if (op == TOK_LE || op == TOK_LT)
|
---|
1471 | swapped = !swapped;
|
---|
1472 | if (op == TOK_LE || op == TOK_GE) {
|
---|
1473 | op = 0x93; /* setae */
|
---|
1474 | } else {
|
---|
1475 | op = 0x97; /* seta */
|
---|
1476 | }
|
---|
1477 | }
|
---|
1478 |
|
---|
1479 | if (swapped) {
|
---|
1480 | o(0x7e0ff3); /* movq */
|
---|
1481 | gen_modrm(1, r, vtop->sym, fc);
|
---|
1482 |
|
---|
1483 | if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE) {
|
---|
1484 | o(0x66);
|
---|
1485 | }
|
---|
1486 | o(0x2e0f); /* ucomisd %xmm0, %xmm1 */
|
---|
1487 | o(0xc8);
|
---|
1488 | } else {
|
---|
1489 | if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE) {
|
---|
1490 | o(0x66);
|
---|
1491 | }
|
---|
1492 | o(0x2e0f); /* ucomisd */
|
---|
1493 | gen_modrm(0, r, vtop->sym, fc);
|
---|
1494 | }
|
---|
1495 |
|
---|
1496 | vtop--;
|
---|
1497 | vtop->r = VT_CMP;
|
---|
1498 | vtop->c.i = op | 0x100;
|
---|
1499 | } else {
|
---|
1500 | /* no memory reference possible for long double operations */
|
---|
1501 | if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
|
---|
1502 | load(TREG_XMM0, vtop);
|
---|
1503 | swapped = !swapped;
|
---|
1504 | }
|
---|
1505 | switch(op) {
|
---|
1506 | default:
|
---|
1507 | case '+':
|
---|
1508 | a = 0;
|
---|
1509 | break;
|
---|
1510 | case '-':
|
---|
1511 | a = 4;
|
---|
1512 | break;
|
---|
1513 | case '*':
|
---|
1514 | a = 1;
|
---|
1515 | break;
|
---|
1516 | case '/':
|
---|
1517 | a = 6;
|
---|
1518 | break;
|
---|
1519 | }
|
---|
1520 | ft = vtop->type.t;
|
---|
1521 | fc = vtop->c.ul;
|
---|
1522 | if ((ft & VT_BTYPE) == VT_LDOUBLE) {
|
---|
1523 | o(0xde); /* fxxxp %st, %st(1) */
|
---|
1524 | o(0xc1 + (a << 3));
|
---|
1525 | } else {
|
---|
1526 | /* if saved lvalue, then we must reload it */
|
---|
1527 | r = vtop->r;
|
---|
1528 | if ((r & VT_VALMASK) == VT_LLOCAL) {
|
---|
1529 | SValue v1;
|
---|
1530 | r = get_reg(RC_INT);
|
---|
1531 | v1.type.t = VT_PTR;
|
---|
1532 | v1.r = VT_LOCAL | VT_LVAL;
|
---|
1533 | v1.c.ul = fc;
|
---|
1534 | load(r, &v1);
|
---|
1535 | fc = 0;
|
---|
1536 | }
|
---|
1537 | if (swapped) {
|
---|
1538 | /* movq %xmm0,%xmm1 */
|
---|
1539 | o(0x7e0ff3);
|
---|
1540 | o(0xc8);
|
---|
1541 | load(TREG_XMM0, vtop);
|
---|
1542 | /* subsd %xmm1,%xmm0 (f2 0f 5c c1) */
|
---|
1543 | if ((ft & VT_BTYPE) == VT_DOUBLE) {
|
---|
1544 | o(0xf2);
|
---|
1545 | } else {
|
---|
1546 | o(0xf3);
|
---|
1547 | }
|
---|
1548 | o(0x0f);
|
---|
1549 | o(0x58 + a);
|
---|
1550 | o(0xc1);
|
---|
1551 | } else {
|
---|
1552 | if ((ft & VT_BTYPE) == VT_DOUBLE) {
|
---|
1553 | o(0xf2);
|
---|
1554 | } else {
|
---|
1555 | o(0xf3);
|
---|
1556 | }
|
---|
1557 | o(0x0f);
|
---|
1558 | o(0x58 + a);
|
---|
1559 | gen_modrm(0, r, vtop->sym, fc);
|
---|
1560 | }
|
---|
1561 | }
|
---|
1562 | vtop--;
|
---|
1563 | }
|
---|
1564 | }
|
---|
1565 | }
|
---|
1566 |
|
---|
1567 | /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
|
---|
1568 | and 'long long' cases. */
|
---|
1569 | void gen_cvt_itof(int t)
|
---|
1570 | {
|
---|
1571 | if ((t & VT_BTYPE) == VT_LDOUBLE) {
|
---|
1572 | save_reg(TREG_ST0);
|
---|
1573 | gv(RC_INT);
|
---|
1574 | if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
|
---|
1575 | /* signed long long to float/double/long double (unsigned case
|
---|
1576 | is handled generically) */
|
---|
1577 | o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
|
---|
1578 | o(0x242cdf); /* fildll (%rsp) */
|
---|
1579 | o(0x08c48348); /* add $8, %rsp */
|
---|
1580 | } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
|
---|
1581 | (VT_INT | VT_UNSIGNED)) {
|
---|
1582 | /* unsigned int to float/double/long double */
|
---|
1583 | o(0x6a); /* push $0 */
|
---|
1584 | g(0x00);
|
---|
1585 | o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
|
---|
1586 | o(0x242cdf); /* fildll (%rsp) */
|
---|
1587 | o(0x10c48348); /* add $16, %rsp */
|
---|
1588 | } else {
|
---|
1589 | /* int to float/double/long double */
|
---|
1590 | o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
|
---|
1591 | o(0x2404db); /* fildl (%rsp) */
|
---|
1592 | o(0x08c48348); /* add $8, %rsp */
|
---|
1593 | }
|
---|
1594 | vtop->r = TREG_ST0;
|
---|
1595 | } else {
|
---|
1596 | save_reg(TREG_XMM0);
|
---|
1597 | gv(RC_INT);
|
---|
1598 | o(0xf2 + ((t & VT_BTYPE) == VT_FLOAT));
|
---|
1599 | if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
|
---|
1600 | (VT_INT | VT_UNSIGNED) ||
|
---|
1601 | (vtop->type.t & VT_BTYPE) == VT_LLONG) {
|
---|
1602 | o(0x48); /* REX */
|
---|
1603 | }
|
---|
1604 | o(0x2a0f);
|
---|
1605 | o(0xc0 + (vtop->r & VT_VALMASK)); /* cvtsi2sd */
|
---|
1606 | vtop->r = TREG_XMM0;
|
---|
1607 | }
|
---|
1608 | }
|
---|
1609 |
|
---|
1610 | /* convert from one floating point type to another */
|
---|
1611 | void gen_cvt_ftof(int t)
|
---|
1612 | {
|
---|
1613 | int ft, bt, tbt;
|
---|
1614 |
|
---|
1615 | ft = vtop->type.t;
|
---|
1616 | bt = ft & VT_BTYPE;
|
---|
1617 | tbt = t & VT_BTYPE;
|
---|
1618 |
|
---|
1619 | if (bt == VT_FLOAT) {
|
---|
1620 | gv(RC_FLOAT);
|
---|
1621 | if (tbt == VT_DOUBLE) {
|
---|
1622 | o(0xc0140f); /* unpcklps */
|
---|
1623 | o(0xc05a0f); /* cvtps2pd */
|
---|
1624 | } else if (tbt == VT_LDOUBLE) {
|
---|
1625 | /* movss %xmm0,-0x10(%rsp) */
|
---|
1626 | o(0x44110ff3);
|
---|
1627 | o(0xf024);
|
---|
1628 | o(0xf02444d9); /* flds -0x10(%rsp) */
|
---|
1629 | vtop->r = TREG_ST0;
|
---|
1630 | }
|
---|
1631 | } else if (bt == VT_DOUBLE) {
|
---|
1632 | gv(RC_FLOAT);
|
---|
1633 | if (tbt == VT_FLOAT) {
|
---|
1634 | o(0xc0140f66); /* unpcklpd */
|
---|
1635 | o(0xc05a0f66); /* cvtpd2ps */
|
---|
1636 | } else if (tbt == VT_LDOUBLE) {
|
---|
1637 | /* movsd %xmm0,-0x10(%rsp) */
|
---|
1638 | o(0x44110ff2);
|
---|
1639 | o(0xf024);
|
---|
1640 | o(0xf02444dd); /* fldl -0x10(%rsp) */
|
---|
1641 | vtop->r = TREG_ST0;
|
---|
1642 | }
|
---|
1643 | } else {
|
---|
1644 | gv(RC_ST0);
|
---|
1645 | if (tbt == VT_DOUBLE) {
|
---|
1646 | o(0xf0245cdd); /* fstpl -0x10(%rsp) */
|
---|
1647 | /* movsd -0x10(%rsp),%xmm0 */
|
---|
1648 | o(0x44100ff2);
|
---|
1649 | o(0xf024);
|
---|
1650 | vtop->r = TREG_XMM0;
|
---|
1651 | } else if (tbt == VT_FLOAT) {
|
---|
1652 | o(0xf0245cd9); /* fstps -0x10(%rsp) */
|
---|
1653 | /* movss -0x10(%rsp),%xmm0 */
|
---|
1654 | o(0x44100ff3);
|
---|
1655 | o(0xf024);
|
---|
1656 | vtop->r = TREG_XMM0;
|
---|
1657 | }
|
---|
1658 | }
|
---|
1659 | }
|
---|
1660 |
|
---|
1661 | /* convert fp to int 't' type */
|
---|
1662 | void gen_cvt_ftoi(int t)
|
---|
1663 | {
|
---|
1664 | int ft, bt, size, r;
|
---|
1665 | ft = vtop->type.t;
|
---|
1666 | bt = ft & VT_BTYPE;
|
---|
1667 | if (bt == VT_LDOUBLE) {
|
---|
1668 | gen_cvt_ftof(VT_DOUBLE);
|
---|
1669 | bt = VT_DOUBLE;
|
---|
1670 | }
|
---|
1671 |
|
---|
1672 | gv(RC_FLOAT);
|
---|
1673 | if (t != VT_INT)
|
---|
1674 | size = 8;
|
---|
1675 | else
|
---|
1676 | size = 4;
|
---|
1677 |
|
---|
1678 | r = get_reg(RC_INT);
|
---|
1679 | if (bt == VT_FLOAT) {
|
---|
1680 | o(0xf3);
|
---|
1681 | } else if (bt == VT_DOUBLE) {
|
---|
1682 | o(0xf2);
|
---|
1683 | } else {
|
---|
1684 | assert(0);
|
---|
1685 | }
|
---|
1686 | orex(size == 8, r, 0, 0x2c0f); /* cvttss2si or cvttsd2si */
|
---|
1687 | o(0xc0 + (REG_VALUE(r) << 3));
|
---|
1688 | vtop->r = r;
|
---|
1689 | }
|
---|
1690 |
|
---|
1691 | /* computed goto support */
|
---|
1692 | void ggoto(void)
|
---|
1693 | {
|
---|
1694 | gcall_or_jmp(1);
|
---|
1695 | vtop--;
|
---|
1696 | }
|
---|
1697 |
|
---|
1698 | /* end of x86-64 code generator */
|
---|
1699 | /*************************************************************/
|
---|
1700 | #endif /* ! TARGET_DEFS_ONLY */
|
---|
1701 | /******************************************************/
|
---|