]> git.ipfire.org Git - thirdparty/qemu.git/blob - tcg/i386/tcg-target.inc.c
projects / thirdparty / qemu.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge remote-tracking branch 'remotes/dgibson/tags/ppc-for-2.11-20170908' into staging
[thirdparty/qemu.git] / tcg / i386 / tcg-target.inc.c
1 /*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Fabrice Bellard
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25 #include "tcg-pool.inc.c"
26
27 #ifdef CONFIG_DEBUG_TCG
28 static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
29 #if TCG_TARGET_REG_BITS == 64
30 "%rax", "%rcx", "%rdx", "%rbx", "%rsp", "%rbp", "%rsi", "%rdi",
31 "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15",
32 #else
33 "%eax", "%ecx", "%edx", "%ebx", "%esp", "%ebp", "%esi", "%edi",
34 #endif
35 };
36 #endif
37
38 static const int tcg_target_reg_alloc_order[] = {
39 #if TCG_TARGET_REG_BITS == 64
40 TCG_REG_RBP,
41 TCG_REG_RBX,
42 TCG_REG_R12,
43 TCG_REG_R13,
44 TCG_REG_R14,
45 TCG_REG_R15,
46 TCG_REG_R10,
47 TCG_REG_R11,
48 TCG_REG_R9,
49 TCG_REG_R8,
50 TCG_REG_RCX,
51 TCG_REG_RDX,
52 TCG_REG_RSI,
53 TCG_REG_RDI,
54 TCG_REG_RAX,
55 #else
56 TCG_REG_EBX,
57 TCG_REG_ESI,
58 TCG_REG_EDI,
59 TCG_REG_EBP,
60 TCG_REG_ECX,
61 TCG_REG_EDX,
62 TCG_REG_EAX,
63 #endif
64 };
65
66 static const int tcg_target_call_iarg_regs[] = {
67 #if TCG_TARGET_REG_BITS == 64
68 #if defined(_WIN64)
69 TCG_REG_RCX,
70 TCG_REG_RDX,
71 #else
72 TCG_REG_RDI,
73 TCG_REG_RSI,
74 TCG_REG_RDX,
75 TCG_REG_RCX,
76 #endif
77 TCG_REG_R8,
78 TCG_REG_R9,
79 #else
80 /* 32 bit mode uses stack based calling convention (GCC default). */
81 #endif
82 };
83
84 static const int tcg_target_call_oarg_regs[] = {
85 TCG_REG_EAX,
86 #if TCG_TARGET_REG_BITS == 32
87 TCG_REG_EDX
88 #endif
89 };
90
91 /* Constants we accept. */
92 #define TCG_CT_CONST_S32 0x100
93 #define TCG_CT_CONST_U32 0x200
94 #define TCG_CT_CONST_I32 0x400
95 #define TCG_CT_CONST_WSZ 0x800
96
97 /* Registers used with L constraint, which are the first argument
98 registers on x86_64, and two random call clobbered registers on
99 i386. */
100 #if TCG_TARGET_REG_BITS == 64
101 # define TCG_REG_L0 tcg_target_call_iarg_regs[0]
102 # define TCG_REG_L1 tcg_target_call_iarg_regs[1]
103 #else
104 # define TCG_REG_L0 TCG_REG_EAX
105 # define TCG_REG_L1 TCG_REG_EDX
106 #endif
107
108 /* The host compiler should supply <cpuid.h> to enable runtime features
109 detection, as we're not going to go so far as our own inline assembly.
110 If not available, default values will be assumed. */
111 #if defined(CONFIG_CPUID_H)
112 #include "qemu/cpuid.h"
113 #endif
114
115 /* For 64-bit, we always know that CMOV is available. */
116 #if TCG_TARGET_REG_BITS == 64
117 # define have_cmov 1
118 #elif defined(CONFIG_CPUID_H)
119 static bool have_cmov;
120 #else
121 # define have_cmov 0
122 #endif
123
124 /* We need these symbols in tcg-target.h, and we can't properly conditionalize
125 it there. Therefore we always define the variable. */
126 bool have_bmi1;
127 bool have_popcnt;
128
129 #ifdef CONFIG_CPUID_H
130 static bool have_movbe;
131 static bool have_bmi2;
132 static bool have_lzcnt;
133 #else
134 # define have_movbe 0
135 # define have_bmi2 0
136 # define have_lzcnt 0
137 #endif
138
139 static tcg_insn_unit *tb_ret_addr;
140
141 static void patch_reloc(tcg_insn_unit *code_ptr, int type,
142 intptr_t value, intptr_t addend)
143 {
144 value += addend;
145 switch(type) {
146 case R_386_PC32:
147 value -= (uintptr_t)code_ptr;
148 if (value != (int32_t)value) {
149 tcg_abort();
150 }
151 tcg_patch32(code_ptr, value);
152 break;
153 case R_386_PC8:
154 value -= (uintptr_t)code_ptr;
155 if (value != (int8_t)value) {
156 tcg_abort();
157 }
158 tcg_patch8(code_ptr, value);
159 break;
160 default:
161 tcg_abort();
162 }
163 }
164
165 /* parse target specific constraints */
166 static const char *target_parse_constraint(TCGArgConstraint *ct,
167 const char *ct_str, TCGType type)
168 {
169 switch(*ct_str++) {
170 case 'a':
171 ct->ct |= TCG_CT_REG;
172 tcg_regset_set_reg(ct->u.regs, TCG_REG_EAX);
173 break;
174 case 'b':
175 ct->ct |= TCG_CT_REG;
176 tcg_regset_set_reg(ct->u.regs, TCG_REG_EBX);
177 break;
178 case 'c':
179 ct->ct |= TCG_CT_REG;
180 tcg_regset_set_reg(ct->u.regs, TCG_REG_ECX);
181 break;
182 case 'd':
183 ct->ct |= TCG_CT_REG;
184 tcg_regset_set_reg(ct->u.regs, TCG_REG_EDX);
185 break;
186 case 'S':
187 ct->ct |= TCG_CT_REG;
188 tcg_regset_set_reg(ct->u.regs, TCG_REG_ESI);
189 break;
190 case 'D':
191 ct->ct |= TCG_CT_REG;
192 tcg_regset_set_reg(ct->u.regs, TCG_REG_EDI);
193 break;
194 case 'q':
195 ct->ct |= TCG_CT_REG;
196 if (TCG_TARGET_REG_BITS == 64) {
197 tcg_regset_set32(ct->u.regs, 0, 0xffff);
198 } else {
199 tcg_regset_set32(ct->u.regs, 0, 0xf);
200 }
201 break;
202 case 'Q':
203 ct->ct |= TCG_CT_REG;
204 tcg_regset_set32(ct->u.regs, 0, 0xf);
205 break;
206 case 'r':
207 ct->ct |= TCG_CT_REG;
208 if (TCG_TARGET_REG_BITS == 64) {
209 tcg_regset_set32(ct->u.regs, 0, 0xffff);
210 } else {
211 tcg_regset_set32(ct->u.regs, 0, 0xff);
212 }
213 break;
214 case 'W':
215 /* With TZCNT/LZCNT, we can have operand-size as an input. */
216 ct->ct |= TCG_CT_CONST_WSZ;
217 break;
218
219 /* qemu_ld/st address constraint */
220 case 'L':
221 ct->ct |= TCG_CT_REG;
222 if (TCG_TARGET_REG_BITS == 64) {
223 tcg_regset_set32(ct->u.regs, 0, 0xffff);
224 } else {
225 tcg_regset_set32(ct->u.regs, 0, 0xff);
226 }
227 tcg_regset_reset_reg(ct->u.regs, TCG_REG_L0);
228 tcg_regset_reset_reg(ct->u.regs, TCG_REG_L1);
229 break;
230
231 case 'e':
232 ct->ct |= (type == TCG_TYPE_I32 ? TCG_CT_CONST : TCG_CT_CONST_S32);
233 break;
234 case 'Z':
235 ct->ct |= (type == TCG_TYPE_I32 ? TCG_CT_CONST : TCG_CT_CONST_U32);
236 break;
237 case 'I':
238 ct->ct |= (type == TCG_TYPE_I32 ? TCG_CT_CONST : TCG_CT_CONST_I32);
239 break;
240
241 default:
242 return NULL;
243 }
244 return ct_str;
245 }
246
247 /* test if a constant matches the constraint */
248 static inline int tcg_target_const_match(tcg_target_long val, TCGType type,
249 const TCGArgConstraint *arg_ct)
250 {
251 int ct = arg_ct->ct;
252 if (ct & TCG_CT_CONST) {
253 return 1;
254 }
255 if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
256 return 1;
257 }
258 if ((ct & TCG_CT_CONST_U32) && val == (uint32_t)val) {
259 return 1;
260 }
261 if ((ct & TCG_CT_CONST_I32) && ~val == (int32_t)~val) {
262 return 1;
263 }
264 if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
265 return 1;
266 }
267 return 0;
268 }
269
270 #if TCG_TARGET_REG_BITS == 64
271 # define LOWREGMASK(x) ((x) & 7)
272 #else
273 # define LOWREGMASK(x) (x)
274 #endif
275
276 #define P_EXT 0x100 /* 0x0f opcode prefix */
277 #define P_EXT38 0x200 /* 0x0f 0x38 opcode prefix */
278 #define P_DATA16 0x400 /* 0x66 opcode prefix */
279 #if TCG_TARGET_REG_BITS == 64
280 # define P_ADDR32 0x800 /* 0x67 opcode prefix */
281 # define P_REXW 0x1000 /* Set REX.W = 1 */
282 # define P_REXB_R 0x2000 /* REG field as byte register */
283 # define P_REXB_RM 0x4000 /* R/M field as byte register */
284 # define P_GS 0x8000 /* gs segment override */
285 #else
286 # define P_ADDR32 0
287 # define P_REXW 0
288 # define P_REXB_R 0
289 # define P_REXB_RM 0
290 # define P_GS 0
291 #endif
292 #define P_SIMDF3 0x10000 /* 0xf3 opcode prefix */
293 #define P_SIMDF2 0x20000 /* 0xf2 opcode prefix */
294
295 #define OPC_ARITH_EvIz (0x81)
296 #define OPC_ARITH_EvIb (0x83)
297 #define OPC_ARITH_GvEv (0x03) /* ... plus (ARITH_FOO << 3) */
298 #define OPC_ANDN (0xf2 | P_EXT38)
299 #define OPC_ADD_GvEv (OPC_ARITH_GvEv | (ARITH_ADD << 3))
300 #define OPC_BSF (0xbc | P_EXT)
301 #define OPC_BSR (0xbd | P_EXT)
302 #define OPC_BSWAP (0xc8 | P_EXT)
303 #define OPC_CALL_Jz (0xe8)
304 #define OPC_CMOVCC (0x40 | P_EXT) /* ... plus condition code */
305 #define OPC_CMP_GvEv (OPC_ARITH_GvEv | (ARITH_CMP << 3))
306 #define OPC_DEC_r32 (0x48)
307 #define OPC_IMUL_GvEv (0xaf | P_EXT)
308 #define OPC_IMUL_GvEvIb (0x6b)
309 #define OPC_IMUL_GvEvIz (0x69)
310 #define OPC_INC_r32 (0x40)
311 #define OPC_JCC_long (0x80 | P_EXT) /* ... plus condition code */
312 #define OPC_JCC_short (0x70) /* ... plus condition code */
313 #define OPC_JMP_long (0xe9)
314 #define OPC_JMP_short (0xeb)
315 #define OPC_LEA (0x8d)
316 #define OPC_LZCNT (0xbd | P_EXT | P_SIMDF3)
317 #define OPC_MOVB_EvGv (0x88) /* stores, more or less */
318 #define OPC_MOVL_EvGv (0x89) /* stores, more or less */
319 #define OPC_MOVL_GvEv (0x8b) /* loads, more or less */
320 #define OPC_MOVB_EvIz (0xc6)
321 #define OPC_MOVL_EvIz (0xc7)
322 #define OPC_MOVL_Iv (0xb8)
323 #define OPC_MOVBE_GyMy (0xf0 | P_EXT38)
324 #define OPC_MOVBE_MyGy (0xf1 | P_EXT38)
325 #define OPC_MOVSBL (0xbe | P_EXT)
326 #define OPC_MOVSWL (0xbf | P_EXT)
327 #define OPC_MOVSLQ (0x63 | P_REXW)
328 #define OPC_MOVZBL (0xb6 | P_EXT)
329 #define OPC_MOVZWL (0xb7 | P_EXT)
330 #define OPC_POP_r32 (0x58)
331 #define OPC_POPCNT (0xb8 | P_EXT | P_SIMDF3)
332 #define OPC_PUSH_r32 (0x50)
333 #define OPC_PUSH_Iv (0x68)
334 #define OPC_PUSH_Ib (0x6a)
335 #define OPC_RET (0xc3)
336 #define OPC_SETCC (0x90 | P_EXT | P_REXB_RM) /* ... plus cc */
337 #define OPC_SHIFT_1 (0xd1)
338 #define OPC_SHIFT_Ib (0xc1)
339 #define OPC_SHIFT_cl (0xd3)
340 #define OPC_SARX (0xf7 | P_EXT38 | P_SIMDF3)
341 #define OPC_SHLX (0xf7 | P_EXT38 | P_DATA16)
342 #define OPC_SHRX (0xf7 | P_EXT38 | P_SIMDF2)
343 #define OPC_TESTL (0x85)
344 #define OPC_TZCNT (0xbc | P_EXT | P_SIMDF3)
345 #define OPC_XCHG_ax_r32 (0x90)
346
347 #define OPC_GRP3_Ev (0xf7)
348 #define OPC_GRP5 (0xff)
349
350 /* Group 1 opcode extensions for 0x80-0x83.
351 These are also used as modifiers for OPC_ARITH. */
352 #define ARITH_ADD 0
353 #define ARITH_OR 1
354 #define ARITH_ADC 2
355 #define ARITH_SBB 3
356 #define ARITH_AND 4
357 #define ARITH_SUB 5
358 #define ARITH_XOR 6
359 #define ARITH_CMP 7
360
361 /* Group 2 opcode extensions for 0xc0, 0xc1, 0xd0-0xd3. */
362 #define SHIFT_ROL 0
363 #define SHIFT_ROR 1
364 #define SHIFT_SHL 4
365 #define SHIFT_SHR 5
366 #define SHIFT_SAR 7
367
368 /* Group 3 opcode extensions for 0xf6, 0xf7. To be used with OPC_GRP3. */
369 #define EXT3_NOT 2
370 #define EXT3_NEG 3
371 #define EXT3_MUL 4
372 #define EXT3_IMUL 5
373 #define EXT3_DIV 6
374 #define EXT3_IDIV 7
375
376 /* Group 5 opcode extensions for 0xff. To be used with OPC_GRP5. */
377 #define EXT5_INC_Ev 0
378 #define EXT5_DEC_Ev 1
379 #define EXT5_CALLN_Ev 2
380 #define EXT5_JMPN_Ev 4
381
382 /* Condition codes to be added to OPC_JCC_{long,short}. */
383 #define JCC_JMP (-1)
384 #define JCC_JO 0x0
385 #define JCC_JNO 0x1
386 #define JCC_JB 0x2
387 #define JCC_JAE 0x3
388 #define JCC_JE 0x4
389 #define JCC_JNE 0x5
390 #define JCC_JBE 0x6
391 #define JCC_JA 0x7
392 #define JCC_JS 0x8
393 #define JCC_JNS 0x9
394 #define JCC_JP 0xa
395 #define JCC_JNP 0xb
396 #define JCC_JL 0xc
397 #define JCC_JGE 0xd
398 #define JCC_JLE 0xe
399 #define JCC_JG 0xf
400
401 static const uint8_t tcg_cond_to_jcc[] = {
402 [TCG_COND_EQ] = JCC_JE,
403 [TCG_COND_NE] = JCC_JNE,
404 [TCG_COND_LT] = JCC_JL,
405 [TCG_COND_GE] = JCC_JGE,
406 [TCG_COND_LE] = JCC_JLE,
407 [TCG_COND_GT] = JCC_JG,
408 [TCG_COND_LTU] = JCC_JB,
409 [TCG_COND_GEU] = JCC_JAE,
410 [TCG_COND_LEU] = JCC_JBE,
411 [TCG_COND_GTU] = JCC_JA,
412 };
413
414 #if TCG_TARGET_REG_BITS == 64
415 static void tcg_out_opc(TCGContext *s, int opc, int r, int rm, int x)
416 {
417 int rex;
418
419 if (opc & P_GS) {
420 tcg_out8(s, 0x65);
421 }
422 if (opc & P_DATA16) {
423 /* We should never be asking for both 16 and 64-bit operation. */
424 tcg_debug_assert((opc & P_REXW) == 0);
425 tcg_out8(s, 0x66);
426 }
427 if (opc & P_ADDR32) {
428 tcg_out8(s, 0x67);
429 }
430 if (opc & P_SIMDF3) {
431 tcg_out8(s, 0xf3);
432 } else if (opc & P_SIMDF2) {
433 tcg_out8(s, 0xf2);
434 }
435
436 rex = 0;
437 rex |= (opc & P_REXW) ? 0x8 : 0x0; /* REX.W */
438 rex |= (r & 8) >> 1; /* REX.R */
439 rex |= (x & 8) >> 2; /* REX.X */
440 rex |= (rm & 8) >> 3; /* REX.B */
441
442 /* P_REXB_{R,RM} indicates that the given register is the low byte.
443 For %[abcd]l we need no REX prefix, but for %{si,di,bp,sp}l we do,
444 as otherwise the encoding indicates %[abcd]h. Note that the values
445 that are ORed in merely indicate that the REX byte must be present;
446 those bits get discarded in output. */
447 rex |= opc & (r >= 4 ? P_REXB_R : 0);
448 rex |= opc & (rm >= 4 ? P_REXB_RM : 0);
449
450 if (rex) {
451 tcg_out8(s, (uint8_t)(rex | 0x40));
452 }
453
454 if (opc & (P_EXT | P_EXT38)) {
455 tcg_out8(s, 0x0f);
456 if (opc & P_EXT38) {
457 tcg_out8(s, 0x38);
458 }
459 }
460
461 tcg_out8(s, opc);
462 }
463 #else
464 static void tcg_out_opc(TCGContext *s, int opc)
465 {
466 if (opc & P_DATA16) {
467 tcg_out8(s, 0x66);
468 }
469 if (opc & P_SIMDF3) {
470 tcg_out8(s, 0xf3);
471 } else if (opc & P_SIMDF2) {
472 tcg_out8(s, 0xf2);
473 }
474 if (opc & (P_EXT | P_EXT38)) {
475 tcg_out8(s, 0x0f);
476 if (opc & P_EXT38) {
477 tcg_out8(s, 0x38);
478 }
479 }
480 tcg_out8(s, opc);
481 }
482 /* Discard the register arguments to tcg_out_opc early, so as not to penalize
483 the 32-bit compilation paths. This method works with all versions of gcc,
484 whereas relying on optimization may not be able to exclude them. */
485 #define tcg_out_opc(s, opc, r, rm, x) (tcg_out_opc)(s, opc)
486 #endif
487
488 static void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
489 {
490 tcg_out_opc(s, opc, r, rm, 0);
491 tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
492 }
493
494 static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm)
495 {
496 int tmp;
497
498 if ((opc & (P_REXW | P_EXT | P_EXT38)) || (rm & 8)) {
499 /* Three byte VEX prefix. */
500 tcg_out8(s, 0xc4);
501
502 /* VEX.m-mmmm */
503 if (opc & P_EXT38) {
504 tmp = 2;
505 } else if (opc & P_EXT) {
506 tmp = 1;
507 } else {
508 tcg_abort();
509 }
510 tmp |= 0x40; /* VEX.X */
511 tmp |= (r & 8 ? 0 : 0x80); /* VEX.R */
512 tmp |= (rm & 8 ? 0 : 0x20); /* VEX.B */
513 tcg_out8(s, tmp);
514
515 tmp = (opc & P_REXW ? 0x80 : 0); /* VEX.W */
516 } else {
517 /* Two byte VEX prefix. */
518 tcg_out8(s, 0xc5);
519
520 tmp = (r & 8 ? 0 : 0x80); /* VEX.R */
521 }
522 /* VEX.pp */
523 if (opc & P_DATA16) {
524 tmp |= 1; /* 0x66 */
525 } else if (opc & P_SIMDF3) {
526 tmp |= 2; /* 0xf3 */
527 } else if (opc & P_SIMDF2) {
528 tmp |= 3; /* 0xf2 */
529 }
530 tmp |= (~v & 15) << 3; /* VEX.vvvv */
531 tcg_out8(s, tmp);
532 tcg_out8(s, opc);
533 tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
534 }
535
536 /* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
537 We handle either RM and INDEX missing with a negative value. In 64-bit
538 mode for absolute addresses, ~RM is the size of the immediate operand
539 that will follow the instruction. */
540
541 static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
542 int index, int shift, intptr_t offset)
543 {
544 int mod, len;
545
546 if (index < 0 && rm < 0) {
547 if (TCG_TARGET_REG_BITS == 64) {
548 /* Try for a rip-relative addressing mode. This has replaced
549 the 32-bit-mode absolute addressing encoding. */
550 intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm;
551 intptr_t disp = offset - pc;
552 if (disp == (int32_t)disp) {
553 tcg_out_opc(s, opc, r, 0, 0);
554 tcg_out8(s, (LOWREGMASK(r) << 3) | 5);
555 tcg_out32(s, disp);
556 return;
557 }
558
559 /* Try for an absolute address encoding. This requires the
560 use of the MODRM+SIB encoding and is therefore larger than
561 rip-relative addressing. */
562 if (offset == (int32_t)offset) {
563 tcg_out_opc(s, opc, r, 0, 0);
564 tcg_out8(s, (LOWREGMASK(r) << 3) | 4);
565 tcg_out8(s, (4 << 3) | 5);
566 tcg_out32(s, offset);
567 return;
568 }
569
570 /* ??? The memory isn't directly addressable. */
571 tcg_abort();
572 } else {
573 /* Absolute address. */
574 tcg_out_opc(s, opc, r, 0, 0);
575 tcg_out8(s, (r << 3) | 5);
576 tcg_out32(s, offset);
577 return;
578 }
579 }
580
581 /* Find the length of the immediate addend. Note that the encoding
582 that would be used for (%ebp) indicates absolute addressing. */
583 if (rm < 0) {
584 mod = 0, len = 4, rm = 5;
585 } else if (offset == 0 && LOWREGMASK(rm) != TCG_REG_EBP) {
586 mod = 0, len = 0;
587 } else if (offset == (int8_t)offset) {
588 mod = 0x40, len = 1;
589 } else {
590 mod = 0x80, len = 4;
591 }
592
593 /* Use a single byte MODRM format if possible. Note that the encoding
594 that would be used for %esp is the escape to the two byte form. */
595 if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) {
596 /* Single byte MODRM format. */
597 tcg_out_opc(s, opc, r, rm, 0);
598 tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
599 } else {
600 /* Two byte MODRM+SIB format. */
601
602 /* Note that the encoding that would place %esp into the index
603 field indicates no index register. In 64-bit mode, the REX.X
604 bit counts, so %r12 can be used as the index. */
605 if (index < 0) {
606 index = 4;
607 } else {
608 tcg_debug_assert(index != TCG_REG_ESP);
609 }
610
611 tcg_out_opc(s, opc, r, rm, index);
612 tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4);
613 tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm));
614 }
615
616 if (len == 1) {
617 tcg_out8(s, offset);
618 } else if (len == 4) {
619 tcg_out32(s, offset);
620 }
621 }
622
623 /* A simplification of the above with no index or shift. */
624 static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r,
625 int rm, intptr_t offset)
626 {
627 tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
628 }
629
630 /* Generate dest op= src. Uses the same ARITH_* codes as tgen_arithi. */
631 static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
632 {
633 /* Propagate an opcode prefix, such as P_REXW. */
634 int ext = subop & ~0x7;
635 subop &= 0x7;
636
637 tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src);
638 }
639
640 static inline void tcg_out_mov(TCGContext *s, TCGType type,
641 TCGReg ret, TCGReg arg)
642 {
643 if (arg != ret) {
644 int opc = OPC_MOVL_GvEv + (type == TCG_TYPE_I64 ? P_REXW : 0);
645 tcg_out_modrm(s, opc, ret, arg);
646 }
647 }
648
649 static void tcg_out_movi(TCGContext *s, TCGType type,
650 TCGReg ret, tcg_target_long arg)
651 {
652 tcg_target_long diff;
653
654 if (arg == 0) {
655 tgen_arithr(s, ARITH_XOR, ret, ret);
656 return;
657 }
658 if (arg == (uint32_t)arg || type == TCG_TYPE_I32) {
659 tcg_out_opc(s, OPC_MOVL_Iv + LOWREGMASK(ret), 0, ret, 0);
660 tcg_out32(s, arg);
661 return;
662 }
663 if (arg == (int32_t)arg) {
664 tcg_out_modrm(s, OPC_MOVL_EvIz + P_REXW, 0, ret);
665 tcg_out32(s, arg);
666 return;
667 }
668
669 /* Try a 7 byte pc-relative lea before the 10 byte movq. */
670 diff = arg - ((uintptr_t)s->code_ptr + 7);
671 if (diff == (int32_t)diff) {
672 tcg_out_opc(s, OPC_LEA | P_REXW, ret, 0, 0);
673 tcg_out8(s, (LOWREGMASK(ret) << 3) | 5);
674 tcg_out32(s, diff);
675 return;
676 }
677
678 tcg_out_opc(s, OPC_MOVL_Iv + P_REXW + LOWREGMASK(ret), 0, ret, 0);
679 tcg_out64(s, arg);
680 }
681
682 static inline void tcg_out_pushi(TCGContext *s, tcg_target_long val)
683 {
684 if (val == (int8_t)val) {
685 tcg_out_opc(s, OPC_PUSH_Ib, 0, 0, 0);
686 tcg_out8(s, val);
687 } else if (val == (int32_t)val) {
688 tcg_out_opc(s, OPC_PUSH_Iv, 0, 0, 0);
689 tcg_out32(s, val);
690 } else {
691 tcg_abort();
692 }
693 }
694
695 static inline void tcg_out_mb(TCGContext *s, TCGArg a0)
696 {
697 /* Given the strength of x86 memory ordering, we only need care for
698 store-load ordering. Experimentally, "lock orl $0,0(%esp)" is
699 faster than "mfence", so don't bother with the sse insn. */
700 if (a0 & TCG_MO_ST_LD) {
701 tcg_out8(s, 0xf0);
702 tcg_out_modrm_offset(s, OPC_ARITH_EvIb, ARITH_OR, TCG_REG_ESP, 0);
703 tcg_out8(s, 0);
704 }
705 }
706
707 static inline void tcg_out_push(TCGContext *s, int reg)
708 {
709 tcg_out_opc(s, OPC_PUSH_r32 + LOWREGMASK(reg), 0, reg, 0);
710 }
711
712 static inline void tcg_out_pop(TCGContext *s, int reg)
713 {
714 tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0);
715 }
716
717 static inline void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
718 TCGReg arg1, intptr_t arg2)
719 {
720 int opc = OPC_MOVL_GvEv + (type == TCG_TYPE_I64 ? P_REXW : 0);
721 tcg_out_modrm_offset(s, opc, ret, arg1, arg2);
722 }
723
724 static inline void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
725 TCGReg arg1, intptr_t arg2)
726 {
727 int opc = OPC_MOVL_EvGv + (type == TCG_TYPE_I64 ? P_REXW : 0);
728 tcg_out_modrm_offset(s, opc, arg, arg1, arg2);
729 }
730
731 static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
732 TCGReg base, intptr_t ofs)
733 {
734 int rexw = 0;
735 if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I64) {
736 if (val != (int32_t)val) {
737 return false;
738 }
739 rexw = P_REXW;
740 }
741 tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs);
742 tcg_out32(s, val);
743 return true;
744 }
745
746 static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
747 {
748 /* Propagate an opcode prefix, such as P_DATA16. */
749 int ext = subopc & ~0x7;
750 subopc &= 0x7;
751
752 if (count == 1) {
753 tcg_out_modrm(s, OPC_SHIFT_1 + ext, subopc, reg);
754 } else {
755 tcg_out_modrm(s, OPC_SHIFT_Ib + ext, subopc, reg);
756 tcg_out8(s, count);
757 }
758 }
759
760 static inline void tcg_out_bswap32(TCGContext *s, int reg)
761 {
762 tcg_out_opc(s, OPC_BSWAP + LOWREGMASK(reg), 0, reg, 0);
763 }
764
765 static inline void tcg_out_rolw_8(TCGContext *s, int reg)
766 {
767 tcg_out_shifti(s, SHIFT_ROL + P_DATA16, reg, 8);
768 }
769
770 static inline void tcg_out_ext8u(TCGContext *s, int dest, int src)
771 {
772 /* movzbl */
773 tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
774 tcg_out_modrm(s, OPC_MOVZBL + P_REXB_RM, dest, src);
775 }
776
777 static void tcg_out_ext8s(TCGContext *s, int dest, int src, int rexw)
778 {
779 /* movsbl */
780 tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
781 tcg_out_modrm(s, OPC_MOVSBL + P_REXB_RM + rexw, dest, src);
782 }
783
784 static inline void tcg_out_ext16u(TCGContext *s, int dest, int src)
785 {
786 /* movzwl */
787 tcg_out_modrm(s, OPC_MOVZWL, dest, src);
788 }
789
790 static inline void tcg_out_ext16s(TCGContext *s, int dest, int src, int rexw)
791 {
792 /* movsw[lq] */
793 tcg_out_modrm(s, OPC_MOVSWL + rexw, dest, src);
794 }
795
796 static inline void tcg_out_ext32u(TCGContext *s, int dest, int src)
797 {
798 /* 32-bit mov zero extends. */
799 tcg_out_modrm(s, OPC_MOVL_GvEv, dest, src);
800 }
801
802 static inline void tcg_out_ext32s(TCGContext *s, int dest, int src)
803 {
804 tcg_out_modrm(s, OPC_MOVSLQ, dest, src);
805 }
806
807 static inline void tcg_out_bswap64(TCGContext *s, int reg)
808 {
809 tcg_out_opc(s, OPC_BSWAP + P_REXW + LOWREGMASK(reg), 0, reg, 0);
810 }
811
812 static void tgen_arithi(TCGContext *s, int c, int r0,
813 tcg_target_long val, int cf)
814 {
815 int rexw = 0;
816
817 if (TCG_TARGET_REG_BITS == 64) {
818 rexw = c & -8;
819 c &= 7;
820 }
821
822 /* ??? While INC is 2 bytes shorter than ADDL $1, they also induce
823 partial flags update stalls on Pentium4 and are not recommended
824 by current Intel optimization manuals. */
825 if (!cf && (c == ARITH_ADD || c == ARITH_SUB) && (val == 1 || val == -1)) {
826 int is_inc = (c == ARITH_ADD) ^ (val < 0);
827 if (TCG_TARGET_REG_BITS == 64) {
828 /* The single-byte increment encodings are re-tasked as the
829 REX prefixes. Use the MODRM encoding. */
830 tcg_out_modrm(s, OPC_GRP5 + rexw,
831 (is_inc ? EXT5_INC_Ev : EXT5_DEC_Ev), r0);
832 } else {
833 tcg_out8(s, (is_inc ? OPC_INC_r32 : OPC_DEC_r32) + r0);
834 }
835 return;
836 }
837
838 if (c == ARITH_AND) {
839 if (TCG_TARGET_REG_BITS == 64) {
840 if (val == 0xffffffffu) {
841 tcg_out_ext32u(s, r0, r0);
842 return;
843 }
844 if (val == (uint32_t)val) {
845 /* AND with no high bits set can use a 32-bit operation. */
846 rexw = 0;
847 }
848 }
849 if (val == 0xffu && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
850 tcg_out_ext8u(s, r0, r0);
851 return;
852 }
853 if (val == 0xffffu) {
854 tcg_out_ext16u(s, r0, r0);
855 return;
856 }
857 }
858
859 if (val == (int8_t)val) {
860 tcg_out_modrm(s, OPC_ARITH_EvIb + rexw, c, r0);
861 tcg_out8(s, val);
862 return;
863 }
864 if (rexw == 0 || val == (int32_t)val) {
865 tcg_out_modrm(s, OPC_ARITH_EvIz + rexw, c, r0);
866 tcg_out32(s, val);
867 return;
868 }
869
870 tcg_abort();
871 }
872
873 static void tcg_out_addi(TCGContext *s, int reg, tcg_target_long val)
874 {
875 if (val != 0) {
876 tgen_arithi(s, ARITH_ADD + P_REXW, reg, val, 0);
877 }
878 }
879
880 /* Use SMALL != 0 to force a short forward branch. */
881 static void tcg_out_jxx(TCGContext *s, int opc, TCGLabel *l, int small)
882 {
883 int32_t val, val1;
884
885 if (l->has_value) {
886 val = tcg_pcrel_diff(s, l->u.value_ptr);
887 val1 = val - 2;
888 if ((int8_t)val1 == val1) {
889 if (opc == -1) {
890 tcg_out8(s, OPC_JMP_short);
891 } else {
892 tcg_out8(s, OPC_JCC_short + opc);
893 }
894 tcg_out8(s, val1);
895 } else {
896 if (small) {
897 tcg_abort();
898 }
899 if (opc == -1) {
900 tcg_out8(s, OPC_JMP_long);
901 tcg_out32(s, val - 5);
902 } else {
903 tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
904 tcg_out32(s, val - 6);
905 }
906 }
907 } else if (small) {
908 if (opc == -1) {
909 tcg_out8(s, OPC_JMP_short);
910 } else {
911 tcg_out8(s, OPC_JCC_short + opc);
912 }
913 tcg_out_reloc(s, s->code_ptr, R_386_PC8, l, -1);
914 s->code_ptr += 1;
915 } else {
916 if (opc == -1) {
917 tcg_out8(s, OPC_JMP_long);
918 } else {
919 tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
920 }
921 tcg_out_reloc(s, s->code_ptr, R_386_PC32, l, -4);
922 s->code_ptr += 4;
923 }
924 }
925
926 static void tcg_out_cmp(TCGContext *s, TCGArg arg1, TCGArg arg2,
927 int const_arg2, int rexw)
928 {
929 if (const_arg2) {
930 if (arg2 == 0) {
931 /* test r, r */
932 tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg1);
933 } else {
934 tgen_arithi(s, ARITH_CMP + rexw, arg1, arg2, 0);
935 }
936 } else {
937 tgen_arithr(s, ARITH_CMP + rexw, arg1, arg2);
938 }
939 }
940
941 static void tcg_out_brcond32(TCGContext *s, TCGCond cond,
942 TCGArg arg1, TCGArg arg2, int const_arg2,
943 TCGLabel *label, int small)
944 {
945 tcg_out_cmp(s, arg1, arg2, const_arg2, 0);
946 tcg_out_jxx(s, tcg_cond_to_jcc[cond], label, small);
947 }
948
949 #if TCG_TARGET_REG_BITS == 64
950 static void tcg_out_brcond64(TCGContext *s, TCGCond cond,
951 TCGArg arg1, TCGArg arg2, int const_arg2,
952 TCGLabel *label, int small)
953 {
954 tcg_out_cmp(s, arg1, arg2, const_arg2, P_REXW);
955 tcg_out_jxx(s, tcg_cond_to_jcc[cond], label, small);
956 }
957 #else
958 /* XXX: we implement it at the target level to avoid having to
959 handle cross basic blocks temporaries */
960 static void tcg_out_brcond2(TCGContext *s, const TCGArg *args,
961 const int *const_args, int small)
962 {
963 TCGLabel *label_next = gen_new_label();
964 TCGLabel *label_this = arg_label(args[5]);
965
966 switch(args[4]) {
967 case TCG_COND_EQ:
968 tcg_out_brcond32(s, TCG_COND_NE, args[0], args[2], const_args[2],
969 label_next, 1);
970 tcg_out_brcond32(s, TCG_COND_EQ, args[1], args[3], const_args[3],
971 label_this, small);
972 break;
973 case TCG_COND_NE:
974 tcg_out_brcond32(s, TCG_COND_NE, args[0], args[2], const_args[2],
975 label_this, small);
976 tcg_out_brcond32(s, TCG_COND_NE, args[1], args[3], const_args[3],
977 label_this, small);
978 break;
979 case TCG_COND_LT:
980 tcg_out_brcond32(s, TCG_COND_LT, args[1], args[3], const_args[3],
981 label_this, small);
982 tcg_out_jxx(s, JCC_JNE, label_next, 1);
983 tcg_out_brcond32(s, TCG_COND_LTU, args[0], args[2], const_args[2],
984 label_this, small);
985 break;
986 case TCG_COND_LE:
987 tcg_out_brcond32(s, TCG_COND_LT, args[1], args[3], const_args[3],
988 label_this, small);
989 tcg_out_jxx(s, JCC_JNE, label_next, 1);
990 tcg_out_brcond32(s, TCG_COND_LEU, args[0], args[2], const_args[2],
991 label_this, small);
992 break;
993 case TCG_COND_GT:
994 tcg_out_brcond32(s, TCG_COND_GT, args[1], args[3], const_args[3],
995 label_this, small);
996 tcg_out_jxx(s, JCC_JNE, label_next, 1);
997 tcg_out_brcond32(s, TCG_COND_GTU, args[0], args[2], const_args[2],
998 label_this, small);
999 break;
1000 case TCG_COND_GE:
1001 tcg_out_brcond32(s, TCG_COND_GT, args[1], args[3], const_args[3],
1002 label_this, small);
1003 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1004 tcg_out_brcond32(s, TCG_COND_GEU, args[0], args[2], const_args[2],
1005 label_this, small);
1006 break;
1007 case TCG_COND_LTU:
1008 tcg_out_brcond32(s, TCG_COND_LTU, args[1], args[3], const_args[3],
1009 label_this, small);
1010 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1011 tcg_out_brcond32(s, TCG_COND_LTU, args[0], args[2], const_args[2],
1012 label_this, small);
1013 break;
1014 case TCG_COND_LEU:
1015 tcg_out_brcond32(s, TCG_COND_LTU, args[1], args[3], const_args[3],
1016 label_this, small);
1017 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1018 tcg_out_brcond32(s, TCG_COND_LEU, args[0], args[2], const_args[2],
1019 label_this, small);
1020 break;
1021 case TCG_COND_GTU:
1022 tcg_out_brcond32(s, TCG_COND_GTU, args[1], args[3], const_args[3],
1023 label_this, small);
1024 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1025 tcg_out_brcond32(s, TCG_COND_GTU, args[0], args[2], const_args[2],
1026 label_this, small);
1027 break;
1028 case TCG_COND_GEU:
1029 tcg_out_brcond32(s, TCG_COND_GTU, args[1], args[3], const_args[3],
1030 label_this, small);
1031 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1032 tcg_out_brcond32(s, TCG_COND_GEU, args[0], args[2], const_args[2],
1033 label_this, small);
1034 break;
1035 default:
1036 tcg_abort();
1037 }
1038 tcg_out_label(s, label_next, s->code_ptr);
1039 }
1040 #endif
1041
1042 static void tcg_out_setcond32(TCGContext *s, TCGCond cond, TCGArg dest,
1043 TCGArg arg1, TCGArg arg2, int const_arg2)
1044 {
1045 tcg_out_cmp(s, arg1, arg2, const_arg2, 0);
1046 tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
1047 tcg_out_ext8u(s, dest, dest);
1048 }
1049
1050 #if TCG_TARGET_REG_BITS == 64
1051 static void tcg_out_setcond64(TCGContext *s, TCGCond cond, TCGArg dest,
1052 TCGArg arg1, TCGArg arg2, int const_arg2)
1053 {
1054 tcg_out_cmp(s, arg1, arg2, const_arg2, P_REXW);
1055 tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
1056 tcg_out_ext8u(s, dest, dest);
1057 }
1058 #else
1059 static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
1060 const int *const_args)
1061 {
1062 TCGArg new_args[6];
1063 TCGLabel *label_true, *label_over;
1064
1065 memcpy(new_args, args+1, 5*sizeof(TCGArg));
1066
1067 if (args[0] == args[1] || args[0] == args[2]
1068 || (!const_args[3] && args[0] == args[3])
1069 || (!const_args[4] && args[0] == args[4])) {
1070 /* When the destination overlaps with one of the argument
1071 registers, don't do anything tricky. */
1072 label_true = gen_new_label();
1073 label_over = gen_new_label();
1074
1075 new_args[5] = label_arg(label_true);
1076 tcg_out_brcond2(s, new_args, const_args+1, 1);
1077
1078 tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1079 tcg_out_jxx(s, JCC_JMP, label_over, 1);
1080 tcg_out_label(s, label_true, s->code_ptr);
1081
1082 tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
1083 tcg_out_label(s, label_over, s->code_ptr);
1084 } else {
1085 /* When the destination does not overlap one of the arguments,
1086 clear the destination first, jump if cond false, and emit an
1087 increment in the true case. This results in smaller code. */
1088
1089 tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1090
1091 label_over = gen_new_label();
1092 new_args[4] = tcg_invert_cond(new_args[4]);
1093 new_args[5] = label_arg(label_over);
1094 tcg_out_brcond2(s, new_args, const_args+1, 1);
1095
1096 tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
1097 tcg_out_label(s, label_over, s->code_ptr);
1098 }
1099 }
1100 #endif
1101
1102 static void tcg_out_cmov(TCGContext *s, TCGCond cond, int rexw,
1103 TCGReg dest, TCGReg v1)
1104 {
1105 if (have_cmov) {
1106 tcg_out_modrm(s, OPC_CMOVCC | tcg_cond_to_jcc[cond] | rexw, dest, v1);
1107 } else {
1108 TCGLabel *over = gen_new_label();
1109 tcg_out_jxx(s, tcg_cond_to_jcc[tcg_invert_cond(cond)], over, 1);
1110 tcg_out_mov(s, TCG_TYPE_I32, dest, v1);
1111 tcg_out_label(s, over, s->code_ptr);
1112 }
1113 }
1114
1115 static void tcg_out_movcond32(TCGContext *s, TCGCond cond, TCGReg dest,
1116 TCGReg c1, TCGArg c2, int const_c2,
1117 TCGReg v1)
1118 {
1119 tcg_out_cmp(s, c1, c2, const_c2, 0);
1120 tcg_out_cmov(s, cond, 0, dest, v1);
1121 }
1122
1123 #if TCG_TARGET_REG_BITS == 64
1124 static void tcg_out_movcond64(TCGContext *s, TCGCond cond, TCGReg dest,
1125 TCGReg c1, TCGArg c2, int const_c2,
1126 TCGReg v1)
1127 {
1128 tcg_out_cmp(s, c1, c2, const_c2, P_REXW);
1129 tcg_out_cmov(s, cond, P_REXW, dest, v1);
1130 }
1131 #endif
1132
1133 static void tcg_out_ctz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1134 TCGArg arg2, bool const_a2)
1135 {
1136 if (have_bmi1) {
1137 tcg_out_modrm(s, OPC_TZCNT + rexw, dest, arg1);
1138 if (const_a2) {
1139 tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1140 } else {
1141 tcg_debug_assert(dest != arg2);
1142 tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1143 }
1144 } else {
1145 tcg_debug_assert(dest != arg2);
1146 tcg_out_modrm(s, OPC_BSF + rexw, dest, arg1);
1147 tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1148 }
1149 }
1150
1151 static void tcg_out_clz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1152 TCGArg arg2, bool const_a2)
1153 {
1154 if (have_lzcnt) {
1155 tcg_out_modrm(s, OPC_LZCNT + rexw, dest, arg1);
1156 if (const_a2) {
1157 tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1158 } else {
1159 tcg_debug_assert(dest != arg2);
1160 tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1161 }
1162 } else {
1163 tcg_debug_assert(!const_a2);
1164 tcg_debug_assert(dest != arg1);
1165 tcg_debug_assert(dest != arg2);
1166
1167 /* Recall that the output of BSR is the index not the count. */
1168 tcg_out_modrm(s, OPC_BSR + rexw, dest, arg1);
1169 tgen_arithi(s, ARITH_XOR + rexw, dest, rexw ? 63 : 31, 0);
1170
1171 /* Since we have destroyed the flags from BSR, we have to re-test. */
1172 tcg_out_cmp(s, arg1, 0, 1, rexw);
1173 tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1174 }
1175 }
1176
1177 static void tcg_out_branch(TCGContext *s, int call, tcg_insn_unit *dest)
1178 {
1179 intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1180
1181 if (disp == (int32_t)disp) {
1182 tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1183 tcg_out32(s, disp);
1184 } else {
1185 /* rip-relative addressing into the constant pool.
1186 This is 6 + 8 = 14 bytes, as compared to using an
1187 an immediate load 10 + 6 = 16 bytes, plus we may
1188 be able to re-use the pool constant for more calls. */
1189 tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1190 tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1191 new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1192 tcg_out32(s, 0);
1193 }
1194 }
1195
1196 static inline void tcg_out_call(TCGContext *s, tcg_insn_unit *dest)
1197 {
1198 tcg_out_branch(s, 1, dest);
1199 }
1200
1201 static void tcg_out_jmp(TCGContext *s, tcg_insn_unit *dest)
1202 {
1203 tcg_out_branch(s, 0, dest);
1204 }
1205
1206 static void tcg_out_nopn(TCGContext *s, int n)
1207 {
1208 int i;
1209 /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1210 * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1211 * duplicate prefix, and all of the interesting recent cores can
1212 * decode and discard the duplicates in a single cycle.
1213 */
1214 tcg_debug_assert(n >= 1);
1215 for (i = 1; i < n; ++i) {
1216 tcg_out8(s, 0x66);
1217 }
1218 tcg_out8(s, 0x90);
1219 }
1220
1221 #if defined(CONFIG_SOFTMMU)
1222 #include "tcg-ldst.inc.c"
1223
1224 /* helper signature: helper_ret_ld_mmu(CPUState *env, target_ulong addr,
1225 * int mmu_idx, uintptr_t ra)
1226 */
1227 static void * const qemu_ld_helpers[16] = {
1228 [MO_UB] = helper_ret_ldub_mmu,
1229 [MO_LEUW] = helper_le_lduw_mmu,
1230 [MO_LEUL] = helper_le_ldul_mmu,
1231 [MO_LEQ] = helper_le_ldq_mmu,
1232 [MO_BEUW] = helper_be_lduw_mmu,
1233 [MO_BEUL] = helper_be_ldul_mmu,
1234 [MO_BEQ] = helper_be_ldq_mmu,
1235 };
1236
1237 /* helper signature: helper_ret_st_mmu(CPUState *env, target_ulong addr,
1238 * uintxx_t val, int mmu_idx, uintptr_t ra)
1239 */
1240 static void * const qemu_st_helpers[16] = {
1241 [MO_UB] = helper_ret_stb_mmu,
1242 [MO_LEUW] = helper_le_stw_mmu,
1243 [MO_LEUL] = helper_le_stl_mmu,
1244 [MO_LEQ] = helper_le_stq_mmu,
1245 [MO_BEUW] = helper_be_stw_mmu,
1246 [MO_BEUL] = helper_be_stl_mmu,
1247 [MO_BEQ] = helper_be_stq_mmu,
1248 };
1249
1250 /* Perform the TLB load and compare.
1251
1252 Inputs:
1253 ADDRLO and ADDRHI contain the low and high part of the address.
1254
1255 MEM_INDEX and S_BITS are the memory context and log2 size of the load.
1256
1257 WHICH is the offset into the CPUTLBEntry structure of the slot to read.
1258 This should be offsetof addr_read or addr_write.
1259
1260 Outputs:
1261 LABEL_PTRS is filled with 1 (32-bit addresses) or 2 (64-bit addresses)
1262 positions of the displacements of forward jumps to the TLB miss case.
1263
1264 Second argument register is loaded with the low part of the address.
1265 In the TLB hit case, it has been adjusted as indicated by the TLB
1266 and so is a host address. In the TLB miss case, it continues to
1267 hold a guest address.
1268
1269 First argument register is clobbered. */
1270
1271 static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
1272 int mem_index, TCGMemOp opc,
1273 tcg_insn_unit **label_ptr, int which)
1274 {
1275 const TCGReg r0 = TCG_REG_L0;
1276 const TCGReg r1 = TCG_REG_L1;
1277 TCGType ttype = TCG_TYPE_I32;
1278 TCGType tlbtype = TCG_TYPE_I32;
1279 int trexw = 0, hrexw = 0, tlbrexw = 0;
1280 unsigned a_bits = get_alignment_bits(opc);
1281 unsigned s_bits = opc & MO_SIZE;
1282 unsigned a_mask = (1 << a_bits) - 1;
1283 unsigned s_mask = (1 << s_bits) - 1;
1284 target_ulong tlb_mask;
1285
1286 if (TCG_TARGET_REG_BITS == 64) {
1287 if (TARGET_LONG_BITS == 64) {
1288 ttype = TCG_TYPE_I64;
1289 trexw = P_REXW;
1290 }
1291 if (TCG_TYPE_PTR == TCG_TYPE_I64) {
1292 hrexw = P_REXW;
1293 if (TARGET_PAGE_BITS + CPU_TLB_BITS > 32) {
1294 tlbtype = TCG_TYPE_I64;
1295 tlbrexw = P_REXW;
1296 }
1297 }
1298 }
1299
1300 tcg_out_mov(s, tlbtype, r0, addrlo);
1301 /* If the required alignment is at least as large as the access, simply
1302 copy the address and mask. For lesser alignments, check that we don't
1303 cross pages for the complete access. */
1304 if (a_bits >= s_bits) {
1305 tcg_out_mov(s, ttype, r1, addrlo);
1306 } else {
1307 tcg_out_modrm_offset(s, OPC_LEA + trexw, r1, addrlo, s_mask - a_mask);
1308 }
1309 tlb_mask = (target_ulong)TARGET_PAGE_MASK | a_mask;
1310
1311 tcg_out_shifti(s, SHIFT_SHR + tlbrexw, r0,
1312 TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
1313
1314 tgen_arithi(s, ARITH_AND + trexw, r1, tlb_mask, 0);
1315 tgen_arithi(s, ARITH_AND + tlbrexw, r0,
1316 (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS, 0);
1317
1318 tcg_out_modrm_sib_offset(s, OPC_LEA + hrexw, r0, TCG_AREG0, r0, 0,
1319 offsetof(CPUArchState, tlb_table[mem_index][0])
1320 + which);
1321
1322 /* cmp 0(r0), r1 */
1323 tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw, r1, r0, 0);
1324
1325 /* Prepare for both the fast path add of the tlb addend, and the slow
1326 path function argument setup. There are two cases worth note:
1327 For 32-bit guest and x86_64 host, MOVL zero-extends the guest address
1328 before the fastpath ADDQ below. For 64-bit guest and x32 host, MOVQ
1329 copies the entire guest address for the slow path, while truncation
1330 for the 32-bit host happens with the fastpath ADDL below. */
1331 tcg_out_mov(s, ttype, r1, addrlo);
1332
1333 /* jne slow_path */
1334 tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
1335 label_ptr[0] = s->code_ptr;
1336 s->code_ptr += 4;
1337
1338 if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1339 /* cmp 4(r0), addrhi */
1340 tcg_out_modrm_offset(s, OPC_CMP_GvEv, addrhi, r0, 4);
1341
1342 /* jne slow_path */
1343 tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
1344 label_ptr[1] = s->code_ptr;
1345 s->code_ptr += 4;
1346 }
1347
1348 /* TLB Hit. */
1349
1350 /* add addend(r0), r1 */
1351 tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r1, r0,
1352 offsetof(CPUTLBEntry, addend) - which);
1353 }
1354
1355 /*
1356 * Record the context of a call to the out of line helper code for the slow path
1357 * for a load or store, so that we can later generate the correct helper code
1358 */
1359 static void add_qemu_ldst_label(TCGContext *s, bool is_ld, TCGMemOpIdx oi,
1360 TCGReg datalo, TCGReg datahi,
1361 TCGReg addrlo, TCGReg addrhi,
1362 tcg_insn_unit *raddr,
1363 tcg_insn_unit **label_ptr)
1364 {
1365 TCGLabelQemuLdst *label = new_ldst_label(s);
1366
1367 label->is_ld = is_ld;
1368 label->oi = oi;
1369 label->datalo_reg = datalo;
1370 label->datahi_reg = datahi;
1371 label->addrlo_reg = addrlo;
1372 label->addrhi_reg = addrhi;
1373 label->raddr = raddr;
1374 label->label_ptr[0] = label_ptr[0];
1375 if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1376 label->label_ptr[1] = label_ptr[1];
1377 }
1378 }
1379
1380 /*
1381 * Generate code for the slow path for a load at the end of block
1382 */
1383 static void tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1384 {
1385 TCGMemOpIdx oi = l->oi;
1386 TCGMemOp opc = get_memop(oi);
1387 TCGReg data_reg;
1388 tcg_insn_unit **label_ptr = &l->label_ptr[0];
1389
1390 /* resolve label address */
1391 tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1392 if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1393 tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1394 }
1395
1396 if (TCG_TARGET_REG_BITS == 32) {
1397 int ofs = 0;
1398
1399 tcg_out_st(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP, ofs);
1400 ofs += 4;
1401
1402 tcg_out_st(s, TCG_TYPE_I32, l->addrlo_reg, TCG_REG_ESP, ofs);
1403 ofs += 4;
1404
1405 if (TARGET_LONG_BITS == 64) {
1406 tcg_out_st(s, TCG_TYPE_I32, l->addrhi_reg, TCG_REG_ESP, ofs);
1407 ofs += 4;
1408 }
1409
1410 tcg_out_sti(s, TCG_TYPE_I32, oi, TCG_REG_ESP, ofs);
1411 ofs += 4;
1412
1413 tcg_out_sti(s, TCG_TYPE_PTR, (uintptr_t)l->raddr, TCG_REG_ESP, ofs);
1414 } else {
1415 tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
1416 /* The second argument is already loaded with addrlo. */
1417 tcg_out_movi(s, TCG_TYPE_I32, tcg_target_call_iarg_regs[2], oi);
1418 tcg_out_movi(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[3],
1419 (uintptr_t)l->raddr);
1420 }
1421
1422 tcg_out_call(s, qemu_ld_helpers[opc & (MO_BSWAP | MO_SIZE)]);
1423
1424 data_reg = l->datalo_reg;
1425 switch (opc & MO_SSIZE) {
1426 case MO_SB:
1427 tcg_out_ext8s(s, data_reg, TCG_REG_EAX, P_REXW);
1428 break;
1429 case MO_SW:
1430 tcg_out_ext16s(s, data_reg, TCG_REG_EAX, P_REXW);
1431 break;
1432 #if TCG_TARGET_REG_BITS == 64
1433 case MO_SL:
1434 tcg_out_ext32s(s, data_reg, TCG_REG_EAX);
1435 break;
1436 #endif
1437 case MO_UB:
1438 case MO_UW:
1439 /* Note that the helpers have zero-extended to tcg_target_long. */
1440 case MO_UL:
1441 tcg_out_mov(s, TCG_TYPE_I32, data_reg, TCG_REG_EAX);
1442 break;
1443 case MO_Q:
1444 if (TCG_TARGET_REG_BITS == 64) {
1445 tcg_out_mov(s, TCG_TYPE_I64, data_reg, TCG_REG_RAX);
1446 } else if (data_reg == TCG_REG_EDX) {
1447 /* xchg %edx, %eax */
1448 tcg_out_opc(s, OPC_XCHG_ax_r32 + TCG_REG_EDX, 0, 0, 0);
1449 tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EAX);
1450 } else {
1451 tcg_out_mov(s, TCG_TYPE_I32, data_reg, TCG_REG_EAX);
1452 tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EDX);
1453 }
1454 break;
1455 default:
1456 tcg_abort();
1457 }
1458
1459 /* Jump to the code corresponding to next IR of qemu_st */
1460 tcg_out_jmp(s, l->raddr);
1461 }
1462
1463 /*
1464 * Generate code for the slow path for a store at the end of block
1465 */
1466 static void tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1467 {
1468 TCGMemOpIdx oi = l->oi;
1469 TCGMemOp opc = get_memop(oi);
1470 TCGMemOp s_bits = opc & MO_SIZE;
1471 tcg_insn_unit **label_ptr = &l->label_ptr[0];
1472 TCGReg retaddr;
1473
1474 /* resolve label address */
1475 tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1476 if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1477 tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1478 }
1479
1480 if (TCG_TARGET_REG_BITS == 32) {
1481 int ofs = 0;
1482
1483 tcg_out_st(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP, ofs);
1484 ofs += 4;
1485
1486 tcg_out_st(s, TCG_TYPE_I32, l->addrlo_reg, TCG_REG_ESP, ofs);
1487 ofs += 4;
1488
1489 if (TARGET_LONG_BITS == 64) {
1490 tcg_out_st(s, TCG_TYPE_I32, l->addrhi_reg, TCG_REG_ESP, ofs);
1491 ofs += 4;
1492 }
1493
1494 tcg_out_st(s, TCG_TYPE_I32, l->datalo_reg, TCG_REG_ESP, ofs);
1495 ofs += 4;
1496
1497 if (s_bits == MO_64) {
1498 tcg_out_st(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_ESP, ofs);
1499 ofs += 4;
1500 }
1501
1502 tcg_out_sti(s, TCG_TYPE_I32, oi, TCG_REG_ESP, ofs);
1503 ofs += 4;
1504
1505 retaddr = TCG_REG_EAX;
1506 tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
1507 tcg_out_st(s, TCG_TYPE_PTR, retaddr, TCG_REG_ESP, ofs);
1508 } else {
1509 tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
1510 /* The second argument is already loaded with addrlo. */
1511 tcg_out_mov(s, (s_bits == MO_64 ? TCG_TYPE_I64 : TCG_TYPE_I32),
1512 tcg_target_call_iarg_regs[2], l->datalo_reg);
1513 tcg_out_movi(s, TCG_TYPE_I32, tcg_target_call_iarg_regs[3], oi);
1514
1515 if (ARRAY_SIZE(tcg_target_call_iarg_regs) > 4) {
1516 retaddr = tcg_target_call_iarg_regs[4];
1517 tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
1518 } else {
1519 retaddr = TCG_REG_RAX;
1520 tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
1521 tcg_out_st(s, TCG_TYPE_PTR, retaddr, TCG_REG_ESP,
1522 TCG_TARGET_CALL_STACK_OFFSET);
1523 }
1524 }
1525
1526 /* "Tail call" to the helper, with the return address back inline. */
1527 tcg_out_push(s, retaddr);
1528 tcg_out_jmp(s, qemu_st_helpers[opc & (MO_BSWAP | MO_SIZE)]);
1529 }
1530 #elif defined(__x86_64__) && defined(__linux__)
1531 # include <asm/prctl.h>
1532 # include <sys/prctl.h>
1533
1534 int arch_prctl(int code, unsigned long addr);
1535
1536 static int guest_base_flags;
1537 static inline void setup_guest_base_seg(void)
1538 {
1539 if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
1540 guest_base_flags = P_GS;
1541 }
1542 }
1543 #else
1544 # define guest_base_flags 0
1545 static inline void setup_guest_base_seg(void) { }
1546 #endif /* SOFTMMU */
1547
1548 static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
1549 TCGReg base, int index, intptr_t ofs,
1550 int seg, TCGMemOp memop)
1551 {
1552 const TCGMemOp real_bswap = memop & MO_BSWAP;
1553 TCGMemOp bswap = real_bswap;
1554 int movop = OPC_MOVL_GvEv;
1555
1556 if (have_movbe && real_bswap) {
1557 bswap = 0;
1558 movop = OPC_MOVBE_GyMy;
1559 }
1560
1561 switch (memop & MO_SSIZE) {
1562 case MO_UB:
1563 tcg_out_modrm_sib_offset(s, OPC_MOVZBL + seg, datalo,
1564 base, index, 0, ofs);
1565 break;
1566 case MO_SB:
1567 tcg_out_modrm_sib_offset(s, OPC_MOVSBL + P_REXW + seg, datalo,
1568 base, index, 0, ofs);
1569 break;
1570 case MO_UW:
1571 tcg_out_modrm_sib_offset(s, OPC_MOVZWL + seg, datalo,
1572 base, index, 0, ofs);
1573 if (real_bswap) {
1574 tcg_out_rolw_8(s, datalo);
1575 }
1576 break;
1577 case MO_SW:
1578 if (real_bswap) {
1579 if (have_movbe) {
1580 tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + seg,
1581 datalo, base, index, 0, ofs);
1582 } else {
1583 tcg_out_modrm_sib_offset(s, OPC_MOVZWL + seg, datalo,
1584 base, index, 0, ofs);
1585 tcg_out_rolw_8(s, datalo);
1586 }
1587 tcg_out_modrm(s, OPC_MOVSWL + P_REXW, datalo, datalo);
1588 } else {
1589 tcg_out_modrm_sib_offset(s, OPC_MOVSWL + P_REXW + seg,
1590 datalo, base, index, 0, ofs);
1591 }
1592 break;
1593 case MO_UL:
1594 tcg_out_modrm_sib_offset(s, movop + seg, datalo, base, index, 0, ofs);
1595 if (bswap) {
1596 tcg_out_bswap32(s, datalo);
1597 }
1598 break;
1599 #if TCG_TARGET_REG_BITS == 64
1600 case MO_SL:
1601 if (real_bswap) {
1602 tcg_out_modrm_sib_offset(s, movop + seg, datalo,
1603 base, index, 0, ofs);
1604 if (bswap) {
1605 tcg_out_bswap32(s, datalo);
1606 }
1607 tcg_out_ext32s(s, datalo, datalo);
1608 } else {
1609 tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + seg, datalo,
1610 base, index, 0, ofs);
1611 }
1612 break;
1613 #endif
1614 case MO_Q:
1615 if (TCG_TARGET_REG_BITS == 64) {
1616 tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datalo,
1617 base, index, 0, ofs);
1618 if (bswap) {
1619 tcg_out_bswap64(s, datalo);
1620 }
1621 } else {
1622 if (real_bswap) {
1623 int t = datalo;
1624 datalo = datahi;
1625 datahi = t;
1626 }
1627 if (base != datalo) {
1628 tcg_out_modrm_sib_offset(s, movop + seg, datalo,
1629 base, index, 0, ofs);
1630 tcg_out_modrm_sib_offset(s, movop + seg, datahi,
1631 base, index, 0, ofs + 4);
1632 } else {
1633 tcg_out_modrm_sib_offset(s, movop + seg, datahi,
1634 base, index, 0, ofs + 4);
1635 tcg_out_modrm_sib_offset(s, movop + seg, datalo,
1636 base, index, 0, ofs);
1637 }
1638 if (bswap) {
1639 tcg_out_bswap32(s, datalo);
1640 tcg_out_bswap32(s, datahi);
1641 }
1642 }
1643 break;
1644 default:
1645 tcg_abort();
1646 }
1647 }
1648
1649 /* XXX: qemu_ld and qemu_st could be modified to clobber only EDX and
1650 EAX. It will be useful once fixed registers globals are less
1651 common. */
1652 static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is64)
1653 {
1654 TCGReg datalo, datahi, addrlo;
1655 TCGReg addrhi __attribute__((unused));
1656 TCGMemOpIdx oi;
1657 TCGMemOp opc;
1658 #if defined(CONFIG_SOFTMMU)
1659 int mem_index;
1660 tcg_insn_unit *label_ptr[2];
1661 #endif
1662
1663 datalo = *args++;
1664 datahi = (TCG_TARGET_REG_BITS == 32 && is64 ? *args++ : 0);
1665 addrlo = *args++;
1666 addrhi = (TARGET_LONG_BITS > TCG_TARGET_REG_BITS ? *args++ : 0);
1667 oi = *args++;
1668 opc = get_memop(oi);
1669
1670 #if defined(CONFIG_SOFTMMU)
1671 mem_index = get_mmuidx(oi);
1672
1673 tcg_out_tlb_load(s, addrlo, addrhi, mem_index, opc,
1674 label_ptr, offsetof(CPUTLBEntry, addr_read));
1675
1676 /* TLB Hit. */
1677 tcg_out_qemu_ld_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, opc);
1678
1679 /* Record the current context of a load into ldst label */
1680 add_qemu_ldst_label(s, true, oi, datalo, datahi, addrlo, addrhi,
1681 s->code_ptr, label_ptr);
1682 #else
1683 {
1684 int32_t offset = guest_base;
1685 TCGReg base = addrlo;
1686 int index = -1;
1687 int seg = 0;
1688
1689 /* For a 32-bit guest, the high 32 bits may contain garbage.
1690 We can do this with the ADDR32 prefix if we're not using
1691 a guest base, or when using segmentation. Otherwise we
1692 need to zero-extend manually. */
1693 if (guest_base == 0 || guest_base_flags) {
1694 seg = guest_base_flags;
1695 offset = 0;
1696 if (TCG_TARGET_REG_BITS > TARGET_LONG_BITS) {
1697 seg |= P_ADDR32;
1698 }
1699 } else if (TCG_TARGET_REG_BITS == 64) {
1700 if (TARGET_LONG_BITS == 32) {
1701 tcg_out_ext32u(s, TCG_REG_L0, base);
1702 base = TCG_REG_L0;
1703 }
1704 if (offset != guest_base) {
1705 tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_L1, guest_base);
1706 index = TCG_REG_L1;
1707 offset = 0;
1708 }
1709 }
1710
1711 tcg_out_qemu_ld_direct(s, datalo, datahi,
1712 base, index, offset, seg, opc);
1713 }
1714 #endif
1715 }
1716
1717 static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
1718 TCGReg base, intptr_t ofs, int seg,
1719 TCGMemOp memop)
1720 {
1721 /* ??? Ideally we wouldn't need a scratch register. For user-only,
1722 we could perform the bswap twice to restore the original value
1723 instead of moving to the scratch. But as it is, the L constraint
1724 means that TCG_REG_L0 is definitely free here. */
1725 const TCGReg scratch = TCG_REG_L0;
1726 const TCGMemOp real_bswap = memop & MO_BSWAP;
1727 TCGMemOp bswap = real_bswap;
1728 int movop = OPC_MOVL_EvGv;
1729
1730 if (have_movbe && real_bswap) {
1731 bswap = 0;
1732 movop = OPC_MOVBE_MyGy;
1733 }
1734
1735 switch (memop & MO_SIZE) {
1736 case MO_8:
1737 /* In 32-bit mode, 8-bit stores can only happen from [abcd]x.
1738 Use the scratch register if necessary. */
1739 if (TCG_TARGET_REG_BITS == 32 && datalo >= 4) {
1740 tcg_out_mov(s, TCG_TYPE_I32, scratch, datalo);
1741 datalo = scratch;
1742 }
1743 tcg_out_modrm_offset(s, OPC_MOVB_EvGv + P_REXB_R + seg,
1744 datalo, base, ofs);
1745 break;
1746 case MO_16:
1747 if (bswap) {
1748 tcg_out_mov(s, TCG_TYPE_I32, scratch, datalo);
1749 tcg_out_rolw_8(s, scratch);
1750 datalo = scratch;
1751 }
1752 tcg_out_modrm_offset(s, movop + P_DATA16 + seg, datalo, base, ofs);
1753 break;
1754 case MO_32:
1755 if (bswap) {
1756 tcg_out_mov(s, TCG_TYPE_I32, scratch, datalo);
1757 tcg_out_bswap32(s, scratch);
1758 datalo = scratch;
1759 }
1760 tcg_out_modrm_offset(s, movop + seg, datalo, base, ofs);
1761 break;
1762 case MO_64:
1763 if (TCG_TARGET_REG_BITS == 64) {
1764 if (bswap) {
1765 tcg_out_mov(s, TCG_TYPE_I64, scratch, datalo);
1766 tcg_out_bswap64(s, scratch);
1767 datalo = scratch;
1768 }
1769 tcg_out_modrm_offset(s, movop + P_REXW + seg, datalo, base, ofs);
1770 } else if (bswap) {
1771 tcg_out_mov(s, TCG_TYPE_I32, scratch, datahi);
1772 tcg_out_bswap32(s, scratch);
1773 tcg_out_modrm_offset(s, OPC_MOVL_EvGv + seg, scratch, base, ofs);
1774 tcg_out_mov(s, TCG_TYPE_I32, scratch, datalo);
1775 tcg_out_bswap32(s, scratch);
1776 tcg_out_modrm_offset(s, OPC_MOVL_EvGv + seg, scratch, base, ofs+4);
1777 } else {
1778 if (real_bswap) {
1779 int t = datalo;
1780 datalo = datahi;
1781 datahi = t;
1782 }
1783 tcg_out_modrm_offset(s, movop + seg, datalo, base, ofs);
1784 tcg_out_modrm_offset(s, movop + seg, datahi, base, ofs+4);
1785 }
1786 break;
1787 default:
1788 tcg_abort();
1789 }
1790 }
1791
1792 static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
1793 {
1794 TCGReg datalo, datahi, addrlo;
1795 TCGReg addrhi __attribute__((unused));
1796 TCGMemOpIdx oi;
1797 TCGMemOp opc;
1798 #if defined(CONFIG_SOFTMMU)
1799 int mem_index;
1800 tcg_insn_unit *label_ptr[2];
1801 #endif
1802
1803 datalo = *args++;
1804 datahi = (TCG_TARGET_REG_BITS == 32 && is64 ? *args++ : 0);
1805 addrlo = *args++;
1806 addrhi = (TARGET_LONG_BITS > TCG_TARGET_REG_BITS ? *args++ : 0);
1807 oi = *args++;
1808 opc = get_memop(oi);
1809
1810 #if defined(CONFIG_SOFTMMU)
1811 mem_index = get_mmuidx(oi);
1812
1813 tcg_out_tlb_load(s, addrlo, addrhi, mem_index, opc,
1814 label_ptr, offsetof(CPUTLBEntry, addr_write));
1815
1816 /* TLB Hit. */
1817 tcg_out_qemu_st_direct(s, datalo, datahi, TCG_REG_L1, 0, 0, opc);
1818
1819 /* Record the current context of a store into ldst label */
1820 add_qemu_ldst_label(s, false, oi, datalo, datahi, addrlo, addrhi,
1821 s->code_ptr, label_ptr);
1822 #else
1823 {
1824 int32_t offset = guest_base;
1825 TCGReg base = addrlo;
1826 int seg = 0;
1827
1828 /* See comment in tcg_out_qemu_ld re zero-extension of addrlo. */
1829 if (guest_base == 0 || guest_base_flags) {
1830 seg = guest_base_flags;
1831 offset = 0;
1832 if (TCG_TARGET_REG_BITS > TARGET_LONG_BITS) {
1833 seg |= P_ADDR32;
1834 }
1835 } else if (TCG_TARGET_REG_BITS == 64) {
1836 /* ??? Note that we can't use the same SIB addressing scheme
1837 as for loads, since we require L0 free for bswap. */
1838 if (offset != guest_base) {
1839 if (TARGET_LONG_BITS == 32) {
1840 tcg_out_ext32u(s, TCG_REG_L0, base);
1841 base = TCG_REG_L0;
1842 }
1843 tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_L1, guest_base);
1844 tgen_arithr(s, ARITH_ADD + P_REXW, TCG_REG_L1, base);
1845 base = TCG_REG_L1;
1846 offset = 0;
1847 } else if (TARGET_LONG_BITS == 32) {
1848 tcg_out_ext32u(s, TCG_REG_L1, base);
1849 base = TCG_REG_L1;
1850 }
1851 }
1852
1853 tcg_out_qemu_st_direct(s, datalo, datahi, base, offset, seg, opc);
1854 }
1855 #endif
1856 }
1857
1858 static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
1859 const TCGArg *args, const int *const_args)
1860 {
1861 TCGArg a0, a1, a2;
1862 int c, const_a2, vexop, rexw = 0;
1863
1864 #if TCG_TARGET_REG_BITS == 64
1865 # define OP_32_64(x) \
1866 case glue(glue(INDEX_op_, x), _i64): \
1867 rexw = P_REXW; /* FALLTHRU */ \
1868 case glue(glue(INDEX_op_, x), _i32)
1869 #else
1870 # define OP_32_64(x) \
1871 case glue(glue(INDEX_op_, x), _i32)
1872 #endif
1873
1874 /* Hoist the loads of the most common arguments. */
1875 a0 = args[0];
1876 a1 = args[1];
1877 a2 = args[2];
1878 const_a2 = const_args[2];
1879
1880 switch (opc) {
1881 case INDEX_op_exit_tb:
1882 /* Reuse the zeroing that exists for goto_ptr. */
1883 if (a0 == 0) {
1884 tcg_out_jmp(s, s->code_gen_epilogue);
1885 } else {
1886 tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
1887 tcg_out_jmp(s, tb_ret_addr);
1888 }
1889 break;
1890 case INDEX_op_goto_tb:
1891 if (s->tb_jmp_insn_offset) {
1892 /* direct jump method */
1893 int gap;
1894 /* jump displacement must be aligned for atomic patching;
1895 * see if we need to add extra nops before jump
1896 */
1897 gap = tcg_pcrel_diff(s, QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4));
1898 if (gap != 1) {
1899 tcg_out_nopn(s, gap - 1);
1900 }
1901 tcg_out8(s, OPC_JMP_long); /* jmp im */
1902 s->tb_jmp_insn_offset[a0] = tcg_current_code_size(s);
1903 tcg_out32(s, 0);
1904 } else {
1905 /* indirect jump method */
1906 tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, -1,
1907 (intptr_t)(s->tb_jmp_target_addr + a0));
1908 }
1909 s->tb_jmp_reset_offset[a0] = tcg_current_code_size(s);
1910 break;
1911 case INDEX_op_goto_ptr:
1912 /* jmp to the given host address (could be epilogue) */
1913 tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
1914 break;
1915 case INDEX_op_br:
1916 tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
1917 break;
1918 OP_32_64(ld8u):
1919 /* Note that we can ignore REXW for the zero-extend to 64-bit. */
1920 tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
1921 break;
1922 OP_32_64(ld8s):
1923 tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
1924 break;
1925 OP_32_64(ld16u):
1926 /* Note that we can ignore REXW for the zero-extend to 64-bit. */
1927 tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
1928 break;
1929 OP_32_64(ld16s):
1930 tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
1931 break;
1932 #if TCG_TARGET_REG_BITS == 64
1933 case INDEX_op_ld32u_i64:
1934 #endif
1935 case INDEX_op_ld_i32:
1936 tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
1937 break;
1938
1939 OP_32_64(st8):
1940 if (const_args[0]) {
1941 tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
1942 tcg_out8(s, a0);
1943 } else {
1944 tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
1945 }
1946 break;
1947 OP_32_64(st16):
1948 if (const_args[0]) {
1949 tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
1950 tcg_out16(s, a0);
1951 } else {
1952 tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
1953 }
1954 break;
1955 #if TCG_TARGET_REG_BITS == 64
1956 case INDEX_op_st32_i64:
1957 #endif
1958 case INDEX_op_st_i32:
1959 if (const_args[0]) {
1960 tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
1961 tcg_out32(s, a0);
1962 } else {
1963 tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
1964 }
1965 break;
1966
1967 OP_32_64(add):
1968 /* For 3-operand addition, use LEA. */
1969 if (a0 != a1) {
1970 TCGArg c3 = 0;
1971 if (const_a2) {
1972 c3 = a2, a2 = -1;
1973 } else if (a0 == a2) {
1974 /* Watch out for dest = src + dest, since we've removed
1975 the matching constraint on the add. */
1976 tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
1977 break;
1978 }
1979
1980 tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, c3);
1981 break;
1982 }
1983 c = ARITH_ADD;
1984 goto gen_arith;
1985 OP_32_64(sub):
1986 c = ARITH_SUB;
1987 goto gen_arith;
1988 OP_32_64(and):
1989 c = ARITH_AND;
1990 goto gen_arith;
1991 OP_32_64(or):
1992 c = ARITH_OR;
1993 goto gen_arith;
1994 OP_32_64(xor):
1995 c = ARITH_XOR;
1996 goto gen_arith;
1997 gen_arith:
1998 if (const_a2) {
1999 tgen_arithi(s, c + rexw, a0, a2, 0);
2000 } else {
2001 tgen_arithr(s, c + rexw, a0, a2);
2002 }
2003 break;
2004
2005 OP_32_64(andc):
2006 if (const_a2) {
2007 tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2008 tgen_arithi(s, ARITH_AND + rexw, a0, ~a2, 0);
2009 } else {
2010 tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2011 }
2012 break;
2013
2014 OP_32_64(mul):
2015 if (const_a2) {
2016 int32_t val;
2017 val = a2;
2018 if (val == (int8_t)val) {
2019 tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2020 tcg_out8(s, val);
2021 } else {
2022 tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2023 tcg_out32(s, val);
2024 }
2025 } else {
2026 tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2027 }
2028 break;
2029
2030 OP_32_64(div2):
2031 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, args[4]);
2032 break;
2033 OP_32_64(divu2):
2034 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, args[4]);
2035 break;
2036
2037 OP_32_64(shl):
2038 /* For small constant 3-operand shift, use LEA. */
2039 if (const_a2 && a0 != a1 && (a2 - 1) < 3) {
2040 if (a2 - 1 == 0) {
2041 /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2042 tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2043 } else {
2044 /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2045 tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2046 }
2047 break;
2048 }
2049 c = SHIFT_SHL;
2050 vexop = OPC_SHLX;
2051 goto gen_shift_maybe_vex;
2052 OP_32_64(shr):
2053 c = SHIFT_SHR;
2054 vexop = OPC_SHRX;
2055 goto gen_shift_maybe_vex;
2056 OP_32_64(sar):
2057 c = SHIFT_SAR;
2058 vexop = OPC_SARX;
2059 goto gen_shift_maybe_vex;
2060 OP_32_64(rotl):
2061 c = SHIFT_ROL;
2062 goto gen_shift;
2063 OP_32_64(rotr):
2064 c = SHIFT_ROR;
2065 goto gen_shift;
2066 gen_shift_maybe_vex:
2067 if (have_bmi2) {
2068 if (!const_a2) {
2069 tcg_out_vex_modrm(s, vexop + rexw, a0, a2, a1);
2070 break;
2071 }
2072 tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2073 }
2074 /* FALLTHRU */
2075 gen_shift:
2076 if (const_a2) {
2077 tcg_out_shifti(s, c + rexw, a0, a2);
2078 } else {
2079 tcg_out_modrm(s, OPC_SHIFT_cl + rexw, c, a0);
2080 }
2081 break;
2082
2083 OP_32_64(ctz):
2084 tcg_out_ctz(s, rexw, args[0], args[1], args[2], const_args[2]);
2085 break;
2086 OP_32_64(clz):
2087 tcg_out_clz(s, rexw, args[0], args[1], args[2], const_args[2]);
2088 break;
2089 OP_32_64(ctpop):
2090 tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2091 break;
2092
2093 case INDEX_op_brcond_i32:
2094 tcg_out_brcond32(s, a2, a0, a1, const_args[1], arg_label(args[3]), 0);
2095 break;
2096 case INDEX_op_setcond_i32:
2097 tcg_out_setcond32(s, args[3], a0, a1, a2, const_a2);
2098 break;
2099 case INDEX_op_movcond_i32:
2100 tcg_out_movcond32(s, args[5], a0, a1, a2, const_a2, args[3]);
2101 break;
2102
2103 OP_32_64(bswap16):
2104 tcg_out_rolw_8(s, a0);
2105 break;
2106 OP_32_64(bswap32):
2107 tcg_out_bswap32(s, a0);
2108 break;
2109
2110 OP_32_64(neg):
2111 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
2112 break;
2113 OP_32_64(not):
2114 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
2115 break;
2116
2117 OP_32_64(ext8s):
2118 tcg_out_ext8s(s, a0, a1, rexw);
2119 break;
2120 OP_32_64(ext16s):
2121 tcg_out_ext16s(s, a0, a1, rexw);
2122 break;
2123 OP_32_64(ext8u):
2124 tcg_out_ext8u(s, a0, a1);
2125 break;
2126 OP_32_64(ext16u):
2127 tcg_out_ext16u(s, a0, a1);
2128 break;
2129
2130 case INDEX_op_qemu_ld_i32:
2131 tcg_out_qemu_ld(s, args, 0);
2132 break;
2133 case INDEX_op_qemu_ld_i64:
2134 tcg_out_qemu_ld(s, args, 1);
2135 break;
2136 case INDEX_op_qemu_st_i32:
2137 tcg_out_qemu_st(s, args, 0);
2138 break;
2139 case INDEX_op_qemu_st_i64:
2140 tcg_out_qemu_st(s, args, 1);
2141 break;
2142
2143 OP_32_64(mulu2):
2144 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
2145 break;
2146 OP_32_64(muls2):
2147 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, args[3]);
2148 break;
2149 OP_32_64(add2):
2150 if (const_args[4]) {
2151 tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
2152 } else {
2153 tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
2154 }
2155 if (const_args[5]) {
2156 tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
2157 } else {
2158 tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
2159 }
2160 break;
2161 OP_32_64(sub2):
2162 if (const_args[4]) {
2163 tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
2164 } else {
2165 tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
2166 }
2167 if (const_args[5]) {
2168 tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
2169 } else {
2170 tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
2171 }
2172 break;
2173
2174 #if TCG_TARGET_REG_BITS == 32
2175 case INDEX_op_brcond2_i32:
2176 tcg_out_brcond2(s, args, const_args, 0);
2177 break;
2178 case INDEX_op_setcond2_i32:
2179 tcg_out_setcond2(s, args, const_args);
2180 break;
2181 #else /* TCG_TARGET_REG_BITS == 64 */
2182 case INDEX_op_ld32s_i64:
2183 tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
2184 break;
2185 case INDEX_op_ld_i64:
2186 tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
2187 break;
2188 case INDEX_op_st_i64:
2189 if (const_args[0]) {
2190 tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
2191 tcg_out32(s, a0);
2192 } else {
2193 tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
2194 }
2195 break;
2196
2197 case INDEX_op_brcond_i64:
2198 tcg_out_brcond64(s, a2, a0, a1, const_args[1], arg_label(args[3]), 0);
2199 break;
2200 case INDEX_op_setcond_i64:
2201 tcg_out_setcond64(s, args[3], a0, a1, a2, const_a2);
2202 break;
2203 case INDEX_op_movcond_i64:
2204 tcg_out_movcond64(s, args[5], a0, a1, a2, const_a2, args[3]);
2205 break;
2206
2207 case INDEX_op_bswap64_i64:
2208 tcg_out_bswap64(s, a0);
2209 break;
2210 case INDEX_op_extu_i32_i64:
2211 case INDEX_op_ext32u_i64:
2212 tcg_out_ext32u(s, a0, a1);
2213 break;
2214 case INDEX_op_ext_i32_i64:
2215 case INDEX_op_ext32s_i64:
2216 tcg_out_ext32s(s, a0, a1);
2217 break;
2218 #endif
2219
2220 OP_32_64(deposit):
2221 if (args[3] == 0 && args[4] == 8) {
2222 /* load bits 0..7 */
2223 tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
2224 } else if (args[3] == 8 && args[4] == 8) {
2225 /* load bits 8..15 */
2226 tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
2227 } else if (args[3] == 0 && args[4] == 16) {
2228 /* load bits 0..15 */
2229 tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
2230 } else {
2231 tcg_abort();
2232 }
2233 break;
2234
2235 case INDEX_op_extract_i64:
2236 if (a2 + args[3] == 32) {
2237 /* This is a 32-bit zero-extending right shift. */
2238 tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
2239 tcg_out_shifti(s, SHIFT_SHR, a0, a2);
2240 break;
2241 }
2242 /* FALLTHRU */
2243 case INDEX_op_extract_i32:
2244 /* On the off-chance that we can use the high-byte registers.
2245 Otherwise we emit the same ext16 + shift pattern that we
2246 would have gotten from the normal tcg-op.c expansion. */
2247 tcg_debug_assert(a2 == 8 && args[3] == 8);
2248 if (a1 < 4 && a0 < 8) {
2249 tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
2250 } else {
2251 tcg_out_ext16u(s, a0, a1);
2252 tcg_out_shifti(s, SHIFT_SHR, a0, 8);
2253 }
2254 break;
2255
2256 case INDEX_op_sextract_i32:
2257 /* We don't implement sextract_i64, as we cannot sign-extend to
2258 64-bits without using the REX prefix that explicitly excludes
2259 access to the high-byte registers. */
2260 tcg_debug_assert(a2 == 8 && args[3] == 8);
2261 if (a1 < 4 && a0 < 8) {
2262 tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
2263 } else {
2264 tcg_out_ext16s(s, a0, a1, 0);
2265 tcg_out_shifti(s, SHIFT_SAR, a0, 8);
2266 }
2267 break;
2268
2269 case INDEX_op_mb:
2270 tcg_out_mb(s, a0);
2271 break;
2272 case INDEX_op_mov_i32: /* Always emitted via tcg_out_mov. */
2273 case INDEX_op_mov_i64:
2274 case INDEX_op_movi_i32: /* Always emitted via tcg_out_movi. */
2275 case INDEX_op_movi_i64:
2276 case INDEX_op_call: /* Always emitted via tcg_out_call. */
2277 default:
2278 tcg_abort();
2279 }
2280
2281 #undef OP_32_64
2282 }
2283
2284 static const TCGTargetOpDef *tcg_target_op_def(TCGOpcode op)
2285 {
2286 static const TCGTargetOpDef r = { .args_ct_str = { "r" } };
2287 static const TCGTargetOpDef ri_r = { .args_ct_str = { "ri", "r" } };
2288 static const TCGTargetOpDef re_r = { .args_ct_str = { "re", "r" } };
2289 static const TCGTargetOpDef qi_r = { .args_ct_str = { "qi", "r" } };
2290 static const TCGTargetOpDef r_r = { .args_ct_str = { "r", "r" } };
2291 static const TCGTargetOpDef r_q = { .args_ct_str = { "r", "q" } };
2292 static const TCGTargetOpDef r_re = { .args_ct_str = { "r", "re" } };
2293 static const TCGTargetOpDef r_0 = { .args_ct_str = { "r", "0" } };
2294 static const TCGTargetOpDef r_r_ri = { .args_ct_str = { "r", "r", "ri" } };
2295 static const TCGTargetOpDef r_r_re = { .args_ct_str = { "r", "r", "re" } };
2296 static const TCGTargetOpDef r_0_re = { .args_ct_str = { "r", "0", "re" } };
2297 static const TCGTargetOpDef r_0_ci = { .args_ct_str = { "r", "0", "ci" } };
2298 static const TCGTargetOpDef r_L = { .args_ct_str = { "r", "L" } };
2299 static const TCGTargetOpDef L_L = { .args_ct_str = { "L", "L" } };
2300 static const TCGTargetOpDef r_L_L = { .args_ct_str = { "r", "L", "L" } };
2301 static const TCGTargetOpDef r_r_L = { .args_ct_str = { "r", "r", "L" } };
2302 static const TCGTargetOpDef L_L_L = { .args_ct_str = { "L", "L", "L" } };
2303 static const TCGTargetOpDef r_r_L_L
2304 = { .args_ct_str = { "r", "r", "L", "L" } };
2305 static const TCGTargetOpDef L_L_L_L
2306 = { .args_ct_str = { "L", "L", "L", "L" } };
2307
2308 switch (op) {
2309 case INDEX_op_goto_ptr:
2310 return &r;
2311
2312 case INDEX_op_ld8u_i32:
2313 case INDEX_op_ld8u_i64:
2314 case INDEX_op_ld8s_i32:
2315 case INDEX_op_ld8s_i64:
2316 case INDEX_op_ld16u_i32:
2317 case INDEX_op_ld16u_i64:
2318 case INDEX_op_ld16s_i32:
2319 case INDEX_op_ld16s_i64:
2320 case INDEX_op_ld_i32:
2321 case INDEX_op_ld32u_i64:
2322 case INDEX_op_ld32s_i64:
2323 case INDEX_op_ld_i64:
2324 return &r_r;
2325
2326 case INDEX_op_st8_i32:
2327 case INDEX_op_st8_i64:
2328 return &qi_r;
2329 case INDEX_op_st16_i32:
2330 case INDEX_op_st16_i64:
2331 case INDEX_op_st_i32:
2332 case INDEX_op_st32_i64:
2333 return &ri_r;
2334 case INDEX_op_st_i64:
2335 return &re_r;
2336
2337 case INDEX_op_add_i32:
2338 case INDEX_op_add_i64:
2339 return &r_r_re;
2340 case INDEX_op_sub_i32:
2341 case INDEX_op_sub_i64:
2342 case INDEX_op_mul_i32:
2343 case INDEX_op_mul_i64:
2344 case INDEX_op_or_i32:
2345 case INDEX_op_or_i64:
2346 case INDEX_op_xor_i32:
2347 case INDEX_op_xor_i64:
2348 return &r_0_re;
2349
2350 case INDEX_op_and_i32:
2351 case INDEX_op_and_i64:
2352 {
2353 static const TCGTargetOpDef and
2354 = { .args_ct_str = { "r", "0", "reZ" } };
2355 return &and;
2356 }
2357 break;
2358 case INDEX_op_andc_i32:
2359 case INDEX_op_andc_i64:
2360 {
2361 static const TCGTargetOpDef andc
2362 = { .args_ct_str = { "r", "r", "rI" } };
2363 return &andc;
2364 }
2365 break;
2366
2367 case INDEX_op_shl_i32:
2368 case INDEX_op_shl_i64:
2369 case INDEX_op_shr_i32:
2370 case INDEX_op_shr_i64:
2371 case INDEX_op_sar_i32:
2372 case INDEX_op_sar_i64:
2373 return have_bmi2 ? &r_r_ri : &r_0_ci;
2374 case INDEX_op_rotl_i32:
2375 case INDEX_op_rotl_i64:
2376 case INDEX_op_rotr_i32:
2377 case INDEX_op_rotr_i64:
2378 return &r_0_ci;
2379
2380 case INDEX_op_brcond_i32:
2381 case INDEX_op_brcond_i64:
2382 return &r_re;
2383
2384 case INDEX_op_bswap16_i32:
2385 case INDEX_op_bswap16_i64:
2386 case INDEX_op_bswap32_i32:
2387 case INDEX_op_bswap32_i64:
2388 case INDEX_op_bswap64_i64:
2389 case INDEX_op_neg_i32:
2390 case INDEX_op_neg_i64:
2391 case INDEX_op_not_i32:
2392 case INDEX_op_not_i64:
2393 return &r_0;
2394
2395 case INDEX_op_ext8s_i32:
2396 case INDEX_op_ext8s_i64:
2397 case INDEX_op_ext8u_i32:
2398 case INDEX_op_ext8u_i64:
2399 return &r_q;
2400 case INDEX_op_ext16s_i32:
2401 case INDEX_op_ext16s_i64:
2402 case INDEX_op_ext16u_i32:
2403 case INDEX_op_ext16u_i64:
2404 case INDEX_op_ext32s_i64:
2405 case INDEX_op_ext32u_i64:
2406 case INDEX_op_ext_i32_i64:
2407 case INDEX_op_extu_i32_i64:
2408 case INDEX_op_extract_i32:
2409 case INDEX_op_extract_i64:
2410 case INDEX_op_sextract_i32:
2411 case INDEX_op_ctpop_i32:
2412 case INDEX_op_ctpop_i64:
2413 return &r_r;
2414
2415 case INDEX_op_deposit_i32:
2416 case INDEX_op_deposit_i64:
2417 {
2418 static const TCGTargetOpDef dep
2419 = { .args_ct_str = { "Q", "0", "Q" } };
2420 return &dep;
2421 }
2422 case INDEX_op_setcond_i32:
2423 case INDEX_op_setcond_i64:
2424 {
2425 static const TCGTargetOpDef setc
2426 = { .args_ct_str = { "q", "r", "re" } };
2427 return &setc;
2428 }
2429 case INDEX_op_movcond_i32:
2430 case INDEX_op_movcond_i64:
2431 {
2432 static const TCGTargetOpDef movc
2433 = { .args_ct_str = { "r", "r", "re", "r", "0" } };
2434 return &movc;
2435 }
2436 case INDEX_op_div2_i32:
2437 case INDEX_op_div2_i64:
2438 case INDEX_op_divu2_i32:
2439 case INDEX_op_divu2_i64:
2440 {
2441 static const TCGTargetOpDef div2
2442 = { .args_ct_str = { "a", "d", "0", "1", "r" } };
2443 return &div2;
2444 }
2445 case INDEX_op_mulu2_i32:
2446 case INDEX_op_mulu2_i64:
2447 case INDEX_op_muls2_i32:
2448 case INDEX_op_muls2_i64:
2449 {
2450 static const TCGTargetOpDef mul2
2451 = { .args_ct_str = { "a", "d", "a", "r" } };
2452 return &mul2;
2453 }
2454 case INDEX_op_add2_i32:
2455 case INDEX_op_add2_i64:
2456 case INDEX_op_sub2_i32:
2457 case INDEX_op_sub2_i64:
2458 {
2459 static const TCGTargetOpDef arith2
2460 = { .args_ct_str = { "r", "r", "0", "1", "re", "re" } };
2461 return &arith2;
2462 }
2463 case INDEX_op_ctz_i32:
2464 case INDEX_op_ctz_i64:
2465 {
2466 static const TCGTargetOpDef ctz[2] = {
2467 { .args_ct_str = { "&r", "r", "r" } },
2468 { .args_ct_str = { "&r", "r", "rW" } },
2469 };
2470 return &ctz[have_bmi1];
2471 }
2472 case INDEX_op_clz_i32:
2473 case INDEX_op_clz_i64:
2474 {
2475 static const TCGTargetOpDef clz[2] = {
2476 { .args_ct_str = { "&r", "r", "r" } },
2477 { .args_ct_str = { "&r", "r", "rW" } },
2478 };
2479 return &clz[have_lzcnt];
2480 }
2481
2482 case INDEX_op_qemu_ld_i32:
2483 return TARGET_LONG_BITS <= TCG_TARGET_REG_BITS ? &r_L : &r_L_L;
2484 case INDEX_op_qemu_st_i32:
2485 return TARGET_LONG_BITS <= TCG_TARGET_REG_BITS ? &L_L : &L_L_L;
2486 case INDEX_op_qemu_ld_i64:
2487 return (TCG_TARGET_REG_BITS == 64 ? &r_L
2488 : TARGET_LONG_BITS <= TCG_TARGET_REG_BITS ? &r_r_L
2489 : &r_r_L_L);
2490 case INDEX_op_qemu_st_i64:
2491 return (TCG_TARGET_REG_BITS == 64 ? &L_L
2492 : TARGET_LONG_BITS <= TCG_TARGET_REG_BITS ? &L_L_L
2493 : &L_L_L_L);
2494
2495 case INDEX_op_brcond2_i32:
2496 {
2497 static const TCGTargetOpDef b2
2498 = { .args_ct_str = { "r", "r", "ri", "ri" } };
2499 return &b2;
2500 }
2501 case INDEX_op_setcond2_i32:
2502 {
2503 static const TCGTargetOpDef s2
2504 = { .args_ct_str = { "r", "r", "r", "ri", "ri" } };
2505 return &s2;
2506 }
2507
2508 default:
2509 break;
2510 }
2511 return NULL;
2512 }
2513
2514 static int tcg_target_callee_save_regs[] = {
2515 #if TCG_TARGET_REG_BITS == 64
2516 TCG_REG_RBP,
2517 TCG_REG_RBX,
2518 #if defined(_WIN64)
2519 TCG_REG_RDI,
2520 TCG_REG_RSI,
2521 #endif
2522 TCG_REG_R12,
2523 TCG_REG_R13,
2524 TCG_REG_R14, /* Currently used for the global env. */
2525 TCG_REG_R15,
2526 #else
2527 TCG_REG_EBP, /* Currently used for the global env. */
2528 TCG_REG_EBX,
2529 TCG_REG_ESI,
2530 TCG_REG_EDI,
2531 #endif
2532 };
2533
2534 /* Compute frame size via macros, to share between tcg_target_qemu_prologue
2535 and tcg_register_jit. */
2536
2537 #define PUSH_SIZE \
2538 ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
2539 * (TCG_TARGET_REG_BITS / 8))
2540
2541 #define FRAME_SIZE \
2542 ((PUSH_SIZE \
2543 + TCG_STATIC_CALL_ARGS_SIZE \
2544 + CPU_TEMP_BUF_NLONGS * sizeof(long) \
2545 + TCG_TARGET_STACK_ALIGN - 1) \
2546 & ~(TCG_TARGET_STACK_ALIGN - 1))
2547
2548 /* Generate global QEMU prologue and epilogue code */
2549 static void tcg_target_qemu_prologue(TCGContext *s)
2550 {
2551 int i, stack_addend;
2552
2553 /* TB prologue */
2554
2555 /* Reserve some stack space, also for TCG temps. */
2556 stack_addend = FRAME_SIZE - PUSH_SIZE;
2557 tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
2558 CPU_TEMP_BUF_NLONGS * sizeof(long));
2559
2560 /* Save all callee saved registers. */
2561 for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
2562 tcg_out_push(s, tcg_target_callee_save_regs[i]);
2563 }
2564
2565 #if TCG_TARGET_REG_BITS == 32
2566 tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
2567 (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
2568 tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
2569 /* jmp *tb. */
2570 tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
2571 (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
2572 + stack_addend);
2573 #else
2574 tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
2575 tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
2576 /* jmp *tb. */
2577 tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
2578 #endif
2579
2580 /*
2581 * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
2582 * and fall through to the rest of the epilogue.
2583 */
2584 s->code_gen_epilogue = s->code_ptr;
2585 tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
2586
2587 /* TB epilogue */
2588 tb_ret_addr = s->code_ptr;
2589
2590 tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
2591
2592 for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
2593 tcg_out_pop(s, tcg_target_callee_save_regs[i]);
2594 }
2595 tcg_out_opc(s, OPC_RET, 0, 0, 0);
2596
2597 #if !defined(CONFIG_SOFTMMU)
2598 /* Try to set up a segment register to point to guest_base. */
2599 if (guest_base) {
2600 setup_guest_base_seg();
2601 }
2602 #endif
2603 }
2604
2605 static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
2606 {
2607 memset(p, 0x90, count);
2608 }
2609
2610 static void tcg_target_init(TCGContext *s)
2611 {
2612 #ifdef CONFIG_CPUID_H
2613 unsigned a, b, c, d;
2614 int max = __get_cpuid_max(0, 0);
2615
2616 if (max >= 1) {
2617 __cpuid(1, a, b, c, d);
2618 #ifndef have_cmov
2619 /* For 32-bit, 99% certainty that we're running on hardware that
2620 supports cmov, but we still need to check. In case cmov is not
2621 available, we'll use a small forward branch. */
2622 have_cmov = (d & bit_CMOV) != 0;
2623 #endif
2624 /* MOVBE is only available on Intel Atom and Haswell CPUs, so we
2625 need to probe for it. */
2626 have_movbe = (c & bit_MOVBE) != 0;
2627 have_popcnt = (c & bit_POPCNT) != 0;
2628 }
2629
2630 if (max >= 7) {
2631 /* BMI1 is available on AMD Piledriver and Intel Haswell CPUs. */
2632 __cpuid_count(7, 0, a, b, c, d);
2633 have_bmi1 = (b & bit_BMI) != 0;
2634 have_bmi2 = (b & bit_BMI2) != 0;
2635 }
2636
2637 max = __get_cpuid_max(0x8000000, 0);
2638 if (max >= 1) {
2639 __cpuid(0x80000001, a, b, c, d);
2640 /* LZCNT was introduced with AMD Barcelona and Intel Haswell CPUs. */
2641 have_lzcnt = (c & bit_LZCNT) != 0;
2642 }
2643 #endif /* CONFIG_CPUID_H */
2644
2645 if (TCG_TARGET_REG_BITS == 64) {
2646 tcg_regset_set32(tcg_target_available_regs[TCG_TYPE_I32], 0, 0xffff);
2647 tcg_regset_set32(tcg_target_available_regs[TCG_TYPE_I64], 0, 0xffff);
2648 } else {
2649 tcg_regset_set32(tcg_target_available_regs[TCG_TYPE_I32], 0, 0xff);
2650 }
2651
2652 tcg_regset_clear(tcg_target_call_clobber_regs);
2653 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
2654 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
2655 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
2656 if (TCG_TARGET_REG_BITS == 64) {
2657 #if !defined(_WIN64)
2658 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
2659 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
2660 #endif
2661 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
2662 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
2663 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
2664 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
2665 }
2666
2667 tcg_regset_clear(s->reserved_regs);
2668 tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2669 }
2670
2671 typedef struct {
2672 DebugFrameHeader h;
2673 uint8_t fde_def_cfa[4];
2674 uint8_t fde_reg_ofs[14];
2675 } DebugFrame;
2676
2677 /* We're expecting a 2 byte uleb128 encoded value. */
2678 QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
2679
2680 #if !defined(__ELF__)
2681 /* Host machine without ELF. */
2682 #elif TCG_TARGET_REG_BITS == 64
2683 #define ELF_HOST_MACHINE EM_X86_64
2684 static const DebugFrame debug_frame = {
2685 .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
2686 .h.cie.id = -1,
2687 .h.cie.version = 1,
2688 .h.cie.code_align = 1,
2689 .h.cie.data_align = 0x78, /* sleb128 -8 */
2690 .h.cie.return_column = 16,
2691
2692 /* Total FDE size does not include the "len" member. */
2693 .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
2694
2695 .fde_def_cfa = {
2696 12, 7, /* DW_CFA_def_cfa %rsp, ... */
2697 (FRAME_SIZE & 0x7f) | 0x80, /* ... uleb128 FRAME_SIZE */
2698 (FRAME_SIZE >> 7)
2699 },
2700 .fde_reg_ofs = {
2701 0x90, 1, /* DW_CFA_offset, %rip, -8 */
2702 /* The following ordering must match tcg_target_callee_save_regs. */
2703 0x86, 2, /* DW_CFA_offset, %rbp, -16 */
2704 0x83, 3, /* DW_CFA_offset, %rbx, -24 */
2705 0x8c, 4, /* DW_CFA_offset, %r12, -32 */
2706 0x8d, 5, /* DW_CFA_offset, %r13, -40 */
2707 0x8e, 6, /* DW_CFA_offset, %r14, -48 */
2708 0x8f, 7, /* DW_CFA_offset, %r15, -56 */
2709 }
2710 };
2711 #else
2712 #define ELF_HOST_MACHINE EM_386
2713 static const DebugFrame debug_frame = {
2714 .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
2715 .h.cie.id = -1,
2716 .h.cie.version = 1,
2717 .h.cie.code_align = 1,
2718 .h.cie.data_align = 0x7c, /* sleb128 -4 */
2719 .h.cie.return_column = 8,
2720
2721 /* Total FDE size does not include the "len" member. */
2722 .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
2723
2724 .fde_def_cfa = {
2725 12, 4, /* DW_CFA_def_cfa %esp, ... */
2726 (FRAME_SIZE & 0x7f) | 0x80, /* ... uleb128 FRAME_SIZE */
2727 (FRAME_SIZE >> 7)
2728 },
2729 .fde_reg_ofs = {
2730 0x88, 1, /* DW_CFA_offset, %eip, -4 */
2731 /* The following ordering must match tcg_target_callee_save_regs. */
2732 0x85, 2, /* DW_CFA_offset, %ebp, -8 */
2733 0x83, 3, /* DW_CFA_offset, %ebx, -12 */
2734 0x86, 4, /* DW_CFA_offset, %esi, -16 */
2735 0x87, 5, /* DW_CFA_offset, %edi, -20 */
2736 }
2737 };
2738 #endif
2739
2740 #if defined(ELF_HOST_MACHINE)
2741 void tcg_register_jit(void *buf, size_t buf_size)
2742 {
2743 tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
2744 }
2745 #endif
Mirror of https://git.qemu.org/git/qemu.git