?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
qom: Avoid leaking str and bool properties on failure
[qemu.git] / target-i386 / translate.c
blob707ebd5ca0e5f9f1da4d3d41e921069f50f08d3e
1 /*
2 * i386 translation
3 *
4 * Copyright (c) 2003 Fabrice Bellard
5 *
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2 of the License, or (at your option) any later version.
10 *
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, see <http://www.gnu.org/licenses/>.
18 */
19 #include <stdarg.h>
20 #include <stdlib.h>
21 #include <stdio.h>
22 #include <string.h>
23 #include <inttypes.h>
24 #include <signal.h>
25
26 #include "qemu/host-utils.h"
27 #include "cpu.h"
28 #include "disas/disas.h"
29 #include "tcg-op.h"
30
31 #include "helper.h"
32 #define GEN_HELPER 1
33 #include "helper.h"
34
35 #define PREFIX_REPZ 0x01
36 #define PREFIX_REPNZ 0x02
37 #define PREFIX_LOCK 0x04
38 #define PREFIX_DATA 0x08
39 #define PREFIX_ADR 0x10
40 #define PREFIX_VEX 0x20
41
42 #ifdef TARGET_X86_64
43 #define CODE64(s) ((s)->code64)
44 #define REX_X(s) ((s)->rex_x)
45 #define REX_B(s) ((s)->rex_b)
46 #else
47 #define CODE64(s) 0
48 #define REX_X(s) 0
49 #define REX_B(s) 0
50 #endif
51
52 #ifdef TARGET_X86_64
53 # define ctztl ctz64
54 # define clztl clz64
55 #else
56 # define ctztl ctz32
57 # define clztl clz32
58 #endif
59
60 //#define MACRO_TEST 1
61
62 /* global register indexes */
63 static TCGv_ptr cpu_env;
64 static TCGv cpu_A0;
65 static TCGv cpu_cc_dst, cpu_cc_src, cpu_cc_src2, cpu_cc_srcT;
66 static TCGv_i32 cpu_cc_op;
67 static TCGv cpu_regs[CPU_NB_REGS];
68 /* local temps */
69 static TCGv cpu_T[2];
70 /* local register indexes (only used inside old micro ops) */
71 static TCGv cpu_tmp0, cpu_tmp4;
72 static TCGv_ptr cpu_ptr0, cpu_ptr1;
73 static TCGv_i32 cpu_tmp2_i32, cpu_tmp3_i32;
74 static TCGv_i64 cpu_tmp1_i64;
75
76 static uint8_t gen_opc_cc_op[OPC_BUF_SIZE];
77
78 #include "exec/gen-icount.h"
79
80 #ifdef TARGET_X86_64
81 static int x86_64_hregs;
82 #endif
83
84 typedef struct DisasContext {
85 /* current insn context */
86 int override; /* -1 if no override */
87 int prefix;
88 TCGMemOp aflag;
89 TCGMemOp dflag;
90 target_ulong pc; /* pc = eip + cs_base */
91 int is_jmp; /* 1 = means jump (stop translation), 2 means CPU
92 static state change (stop translation) */
93 /* current block context */
94 target_ulong cs_base; /* base of CS segment */
95 int pe; /* protected mode */
96 int code32; /* 32 bit code segment */
97 #ifdef TARGET_X86_64
98 int lma; /* long mode active */
99 int code64; /* 64 bit code segment */
100 int rex_x, rex_b;
101 #endif
102 int vex_l; /* vex vector length */
103 int vex_v; /* vex vvvv register, without 1's compliment. */
104 int ss32; /* 32 bit stack segment */
105 CCOp cc_op; /* current CC operation */
106 bool cc_op_dirty;
107 int addseg; /* non zero if either DS/ES/SS have a non zero base */
108 int f_st; /* currently unused */
109 int vm86; /* vm86 mode */
110 int cpl;
111 int iopl;
112 int tf; /* TF cpu flag */
113 int singlestep_enabled; /* "hardware" single step enabled */
114 int jmp_opt; /* use direct block chaining for direct jumps */
115 int mem_index; /* select memory access functions */
116 uint64_t flags; /* all execution flags */
117 struct TranslationBlock *tb;
118 int popl_esp_hack; /* for correct popl with esp base handling */
119 int rip_offset; /* only used in x86_64, but left for simplicity */
120 int cpuid_features;
121 int cpuid_ext_features;
122 int cpuid_ext2_features;
123 int cpuid_ext3_features;
124 int cpuid_7_0_ebx_features;
125 } DisasContext;
126
127 static void gen_eob(DisasContext *s);
128 static void gen_jmp(DisasContext *s, target_ulong eip);
129 static void gen_jmp_tb(DisasContext *s, target_ulong eip, int tb_num);
130 static void gen_op(DisasContext *s1, int op, TCGMemOp ot, int d);
131
132 /* i386 arith/logic operations */
133 enum {
134 OP_ADDL,
135 OP_ORL,
136 OP_ADCL,
137 OP_SBBL,
138 OP_ANDL,
139 OP_SUBL,
140 OP_XORL,
141 OP_CMPL,
142 };
143
144 /* i386 shift ops */
145 enum {
146 OP_ROL,
147 OP_ROR,
148 OP_RCL,
149 OP_RCR,
150 OP_SHL,
151 OP_SHR,
152 OP_SHL1, /* undocumented */
153 OP_SAR = 7,
154 };
155
156 enum {
157 JCC_O,
158 JCC_B,
159 JCC_Z,
160 JCC_BE,
161 JCC_S,
162 JCC_P,
163 JCC_L,
164 JCC_LE,
165 };
166
167 enum {
168 /* I386 int registers */
169 OR_EAX, /* MUST be even numbered */
170 OR_ECX,
171 OR_EDX,
172 OR_EBX,
173 OR_ESP,
174 OR_EBP,
175 OR_ESI,
176 OR_EDI,
177
178 OR_TMP0 = 16, /* temporary operand register */
179 OR_TMP1,
180 OR_A0, /* temporary register used when doing address evaluation */
181 };
182
183 enum {
184 USES_CC_DST = 1,
185 USES_CC_SRC = 2,
186 USES_CC_SRC2 = 4,
187 USES_CC_SRCT = 8,
188 };
189
190 /* Bit set if the global variable is live after setting CC_OP to X. */
191 static const uint8_t cc_op_live[CC_OP_NB] = {
192 [CC_OP_DYNAMIC] = USES_CC_DST | USES_CC_SRC | USES_CC_SRC2,
193 [CC_OP_EFLAGS] = USES_CC_SRC,
194 [CC_OP_MULB ... CC_OP_MULQ] = USES_CC_DST | USES_CC_SRC,
195 [CC_OP_ADDB ... CC_OP_ADDQ] = USES_CC_DST | USES_CC_SRC,
196 [CC_OP_ADCB ... CC_OP_ADCQ] = USES_CC_DST | USES_CC_SRC | USES_CC_SRC2,
197 [CC_OP_SUBB ... CC_OP_SUBQ] = USES_CC_DST | USES_CC_SRC | USES_CC_SRCT,
198 [CC_OP_SBBB ... CC_OP_SBBQ] = USES_CC_DST | USES_CC_SRC | USES_CC_SRC2,
199 [CC_OP_LOGICB ... CC_OP_LOGICQ] = USES_CC_DST,
200 [CC_OP_INCB ... CC_OP_INCQ] = USES_CC_DST | USES_CC_SRC,
201 [CC_OP_DECB ... CC_OP_DECQ] = USES_CC_DST | USES_CC_SRC,
202 [CC_OP_SHLB ... CC_OP_SHLQ] = USES_CC_DST | USES_CC_SRC,
203 [CC_OP_SARB ... CC_OP_SARQ] = USES_CC_DST | USES_CC_SRC,
204 [CC_OP_BMILGB ... CC_OP_BMILGQ] = USES_CC_DST | USES_CC_SRC,
205 [CC_OP_ADCX] = USES_CC_DST | USES_CC_SRC,
206 [CC_OP_ADOX] = USES_CC_SRC | USES_CC_SRC2,
207 [CC_OP_ADCOX] = USES_CC_DST | USES_CC_SRC | USES_CC_SRC2,
208 [CC_OP_CLR] = 0,
209 };
210
211 static void set_cc_op(DisasContext *s, CCOp op)
212 {
213 int dead;
214
215 if (s->cc_op == op) {
216 return;
217 }
218
219 /* Discard CC computation that will no longer be used. */
220 dead = cc_op_live[s->cc_op] & ~cc_op_live[op];
221 if (dead & USES_CC_DST) {
222 tcg_gen_discard_tl(cpu_cc_dst);
223 }
224 if (dead & USES_CC_SRC) {
225 tcg_gen_discard_tl(cpu_cc_src);
226 }
227 if (dead & USES_CC_SRC2) {
228 tcg_gen_discard_tl(cpu_cc_src2);
229 }
230 if (dead & USES_CC_SRCT) {
231 tcg_gen_discard_tl(cpu_cc_srcT);
232 }
233
234 if (op == CC_OP_DYNAMIC) {
235 /* The DYNAMIC setting is translator only, and should never be
236 stored. Thus we always consider it clean. */
237 s->cc_op_dirty = false;
238 } else {
239 /* Discard any computed CC_OP value (see shifts). */
240 if (s->cc_op == CC_OP_DYNAMIC) {
241 tcg_gen_discard_i32(cpu_cc_op);
242 }
243 s->cc_op_dirty = true;
244 }
245 s->cc_op = op;
246 }
247
248 static void gen_update_cc_op(DisasContext *s)
249 {
250 if (s->cc_op_dirty) {
251 tcg_gen_movi_i32(cpu_cc_op, s->cc_op);
252 s->cc_op_dirty = false;
253 }
254 }
255
256 #ifdef TARGET_X86_64
257
258 #define NB_OP_SIZES 4
259
260 #else /* !TARGET_X86_64 */
261
262 #define NB_OP_SIZES 3
263
264 #endif /* !TARGET_X86_64 */
265
266 #if defined(HOST_WORDS_BIGENDIAN)
267 #define REG_B_OFFSET (sizeof(target_ulong) - 1)
268 #define REG_H_OFFSET (sizeof(target_ulong) - 2)
269 #define REG_W_OFFSET (sizeof(target_ulong) - 2)
270 #define REG_L_OFFSET (sizeof(target_ulong) - 4)
271 #define REG_LH_OFFSET (sizeof(target_ulong) - 8)
272 #else
273 #define REG_B_OFFSET 0
274 #define REG_H_OFFSET 1
275 #define REG_W_OFFSET 0
276 #define REG_L_OFFSET 0
277 #define REG_LH_OFFSET 4
278 #endif
279
280 /* In instruction encodings for byte register accesses the
281 * register number usually indicates "low 8 bits of register N";
282 * however there are some special cases where N 4..7 indicates
283 * [AH, CH, DH, BH], ie "bits 15..8 of register N-4". Return
284 * true for this special case, false otherwise.
285 */
286 static inline bool byte_reg_is_xH(int reg)
287 {
288 if (reg < 4) {
289 return false;
290 }
291 #ifdef TARGET_X86_64
292 if (reg >= 8 || x86_64_hregs) {
293 return false;
294 }
295 #endif
296 return true;
297 }
298
299 /* Select the size of a push/pop operation. */
300 static inline TCGMemOp mo_pushpop(DisasContext *s, TCGMemOp ot)
301 {
302 if (CODE64(s)) {
303 return ot == MO_16 ? MO_16 : MO_64;
304 } else {
305 return ot;
306 }
307 }
308
309 /* Select only size 64 else 32. Used for SSE operand sizes. */
310 static inline TCGMemOp mo_64_32(TCGMemOp ot)
311 {
312 #ifdef TARGET_X86_64
313 return ot == MO_64 ? MO_64 : MO_32;
314 #else
315 return MO_32;
316 #endif
317 }
318
319 /* Select size 8 if lsb of B is clear, else OT. Used for decoding
320 byte vs word opcodes. */
321 static inline TCGMemOp mo_b_d(int b, TCGMemOp ot)
322 {
323 return b & 1 ? ot : MO_8;
324 }
325
326 /* Select size 8 if lsb of B is clear, else OT capped at 32.
327 Used for decoding operand size of port opcodes. */
328 static inline TCGMemOp mo_b_d32(int b, TCGMemOp ot)
329 {
330 return b & 1 ? (ot == MO_16 ? MO_16 : MO_32) : MO_8;
331 }
332
333 static void gen_op_mov_reg_v(TCGMemOp ot, int reg, TCGv t0)
334 {
335 switch(ot) {
336 case MO_8:
337 if (!byte_reg_is_xH(reg)) {
338 tcg_gen_deposit_tl(cpu_regs[reg], cpu_regs[reg], t0, 0, 8);
339 } else {
340 tcg_gen_deposit_tl(cpu_regs[reg - 4], cpu_regs[reg - 4], t0, 8, 8);
341 }
342 break;
343 case MO_16:
344 tcg_gen_deposit_tl(cpu_regs[reg], cpu_regs[reg], t0, 0, 16);
345 break;
346 case MO_32:
347 /* For x86_64, this sets the higher half of register to zero.
348 For i386, this is equivalent to a mov. */
349 tcg_gen_ext32u_tl(cpu_regs[reg], t0);
350 break;
351 #ifdef TARGET_X86_64
352 case MO_64:
353 tcg_gen_mov_tl(cpu_regs[reg], t0);
354 break;
355 #endif
356 default:
357 tcg_abort();
358 }
359 }
360
361 static inline void gen_op_mov_v_reg(TCGMemOp ot, TCGv t0, int reg)
362 {
363 if (ot == MO_8 && byte_reg_is_xH(reg)) {
364 tcg_gen_shri_tl(t0, cpu_regs[reg - 4], 8);
365 tcg_gen_ext8u_tl(t0, t0);
366 } else {
367 tcg_gen_mov_tl(t0, cpu_regs[reg]);
368 }
369 }
370
371 static inline void gen_op_movl_A0_reg(int reg)
372 {
373 tcg_gen_mov_tl(cpu_A0, cpu_regs[reg]);
374 }
375
376 static inline void gen_op_addl_A0_im(int32_t val)
377 {
378 tcg_gen_addi_tl(cpu_A0, cpu_A0, val);
379 #ifdef TARGET_X86_64
380 tcg_gen_andi_tl(cpu_A0, cpu_A0, 0xffffffff);
381 #endif
382 }
383
384 #ifdef TARGET_X86_64
385 static inline void gen_op_addq_A0_im(int64_t val)
386 {
387 tcg_gen_addi_tl(cpu_A0, cpu_A0, val);
388 }
389 #endif
390
391 static void gen_add_A0_im(DisasContext *s, int val)
392 {
393 #ifdef TARGET_X86_64
394 if (CODE64(s))
395 gen_op_addq_A0_im(val);
396 else
397 #endif
398 gen_op_addl_A0_im(val);
399 }
400
401 static inline void gen_op_jmp_v(TCGv dest)
402 {
403 tcg_gen_st_tl(dest, cpu_env, offsetof(CPUX86State, eip));
404 }
405
406 static inline void gen_op_add_reg_im(TCGMemOp size, int reg, int32_t val)
407 {
408 tcg_gen_addi_tl(cpu_tmp0, cpu_regs[reg], val);
409 gen_op_mov_reg_v(size, reg, cpu_tmp0);
410 }
411
412 static inline void gen_op_add_reg_T0(TCGMemOp size, int reg)
413 {
414 tcg_gen_add_tl(cpu_tmp0, cpu_regs[reg], cpu_T[0]);
415 gen_op_mov_reg_v(size, reg, cpu_tmp0);
416 }
417
418 static inline void gen_op_addl_A0_reg_sN(int shift, int reg)
419 {
420 tcg_gen_mov_tl(cpu_tmp0, cpu_regs[reg]);
421 if (shift != 0)
422 tcg_gen_shli_tl(cpu_tmp0, cpu_tmp0, shift);
423 tcg_gen_add_tl(cpu_A0, cpu_A0, cpu_tmp0);
424 /* For x86_64, this sets the higher half of register to zero.
425 For i386, this is equivalent to a nop. */
426 tcg_gen_ext32u_tl(cpu_A0, cpu_A0);
427 }
428
429 static inline void gen_op_movl_A0_seg(int reg)
430 {
431 tcg_gen_ld32u_tl(cpu_A0, cpu_env, offsetof(CPUX86State, segs[reg].base) + REG_L_OFFSET);
432 }
433
434 static inline void gen_op_addl_A0_seg(DisasContext *s, int reg)
435 {
436 tcg_gen_ld_tl(cpu_tmp0, cpu_env, offsetof(CPUX86State, segs[reg].base));
437 #ifdef TARGET_X86_64
438 if (CODE64(s)) {
439 tcg_gen_andi_tl(cpu_A0, cpu_A0, 0xffffffff);
440 tcg_gen_add_tl(cpu_A0, cpu_A0, cpu_tmp0);
441 } else {
442 tcg_gen_add_tl(cpu_A0, cpu_A0, cpu_tmp0);
443 tcg_gen_andi_tl(cpu_A0, cpu_A0, 0xffffffff);
444 }
445 #else
446 tcg_gen_add_tl(cpu_A0, cpu_A0, cpu_tmp0);
447 #endif
448 }
449
450 #ifdef TARGET_X86_64
451 static inline void gen_op_movq_A0_seg(int reg)
452 {
453 tcg_gen_ld_tl(cpu_A0, cpu_env, offsetof(CPUX86State, segs[reg].base));
454 }
455
456 static inline void gen_op_addq_A0_seg(int reg)
457 {
458 tcg_gen_ld_tl(cpu_tmp0, cpu_env, offsetof(CPUX86State, segs[reg].base));
459 tcg_gen_add_tl(cpu_A0, cpu_A0, cpu_tmp0);
460 }
461
462 static inline void gen_op_movq_A0_reg(int reg)
463 {
464 tcg_gen_mov_tl(cpu_A0, cpu_regs[reg]);
465 }
466
467 static inline void gen_op_addq_A0_reg_sN(int shift, int reg)
468 {
469 tcg_gen_mov_tl(cpu_tmp0, cpu_regs[reg]);
470 if (shift != 0)
471 tcg_gen_shli_tl(cpu_tmp0, cpu_tmp0, shift);
472 tcg_gen_add_tl(cpu_A0, cpu_A0, cpu_tmp0);
473 }
474 #endif
475
476 static inline void gen_op_ld_v(DisasContext *s, int idx, TCGv t0, TCGv a0)
477 {
478 tcg_gen_qemu_ld_tl(t0, a0, s->mem_index, idx | MO_LE);
479 }
480
481 static inline void gen_op_st_v(DisasContext *s, int idx, TCGv t0, TCGv a0)
482 {
483 tcg_gen_qemu_st_tl(t0, a0, s->mem_index, idx | MO_LE);
484 }
485
486 static inline void gen_op_st_rm_T0_A0(DisasContext *s, int idx, int d)
487 {
488 if (d == OR_TMP0) {
489 gen_op_st_v(s, idx, cpu_T[0], cpu_A0);
490 } else {
491 gen_op_mov_reg_v(idx, d, cpu_T[0]);
492 }
493 }
494
495 static inline void gen_jmp_im(target_ulong pc)
496 {
497 tcg_gen_movi_tl(cpu_tmp0, pc);
498 gen_op_jmp_v(cpu_tmp0);
499 }
500
501 static inline void gen_string_movl_A0_ESI(DisasContext *s)
502 {
503 int override;
504
505 override = s->override;
506 switch (s->aflag) {
507 #ifdef TARGET_X86_64
508 case MO_64:
509 if (override >= 0) {
510 gen_op_movq_A0_seg(override);
511 gen_op_addq_A0_reg_sN(0, R_ESI);
512 } else {
513 gen_op_movq_A0_reg(R_ESI);
514 }
515 break;
516 #endif
517 case MO_32:
518 /* 32 bit address */
519 if (s->addseg && override < 0)
520 override = R_DS;
521 if (override >= 0) {
522 gen_op_movl_A0_seg(override);
523 gen_op_addl_A0_reg_sN(0, R_ESI);
524 } else {
525 gen_op_movl_A0_reg(R_ESI);
526 }
527 break;
528 case MO_16:
529 /* 16 address, always override */
530 if (override < 0)
531 override = R_DS;
532 tcg_gen_ext16u_tl(cpu_A0, cpu_regs[R_ESI]);
533 gen_op_addl_A0_seg(s, override);
534 break;
535 default:
536 tcg_abort();
537 }
538 }
539
540 static inline void gen_string_movl_A0_EDI(DisasContext *s)
541 {
542 switch (s->aflag) {
543 #ifdef TARGET_X86_64
544 case MO_64:
545 gen_op_movq_A0_reg(R_EDI);
546 break;
547 #endif
548 case MO_32:
549 if (s->addseg) {
550 gen_op_movl_A0_seg(R_ES);
551 gen_op_addl_A0_reg_sN(0, R_EDI);
552 } else {
553 gen_op_movl_A0_reg(R_EDI);
554 }
555 break;
556 case MO_16:
557 tcg_gen_ext16u_tl(cpu_A0, cpu_regs[R_EDI]);
558 gen_op_addl_A0_seg(s, R_ES);
559 break;
560 default:
561 tcg_abort();
562 }
563 }
564
565 static inline void gen_op_movl_T0_Dshift(TCGMemOp ot)
566 {
567 tcg_gen_ld32s_tl(cpu_T[0], cpu_env, offsetof(CPUX86State, df));
568 tcg_gen_shli_tl(cpu_T[0], cpu_T[0], ot);
569 };
570
571 static TCGv gen_ext_tl(TCGv dst, TCGv src, TCGMemOp size, bool sign)
572 {
573 switch (size) {
574 case MO_8:
575 if (sign) {
576 tcg_gen_ext8s_tl(dst, src);
577 } else {
578 tcg_gen_ext8u_tl(dst, src);
579 }
580 return dst;
581 case MO_16:
582 if (sign) {
583 tcg_gen_ext16s_tl(dst, src);
584 } else {
585 tcg_gen_ext16u_tl(dst, src);
586 }
587 return dst;
588 #ifdef TARGET_X86_64
589 case MO_32:
590 if (sign) {
591 tcg_gen_ext32s_tl(dst, src);
592 } else {
593 tcg_gen_ext32u_tl(dst, src);
594 }
595 return dst;
596 #endif
597 default:
598 return src;
599 }
600 }
601
602 static void gen_extu(TCGMemOp ot, TCGv reg)
603 {
604 gen_ext_tl(reg, reg, ot, false);
605 }
606
607 static void gen_exts(TCGMemOp ot, TCGv reg)
608 {
609 gen_ext_tl(reg, reg, ot, true);
610 }
611
612 static inline void gen_op_jnz_ecx(TCGMemOp size, int label1)
613 {
614 tcg_gen_mov_tl(cpu_tmp0, cpu_regs[R_ECX]);
615 gen_extu(size, cpu_tmp0);
616 tcg_gen_brcondi_tl(TCG_COND_NE, cpu_tmp0, 0, label1);
617 }
618
619 static inline void gen_op_jz_ecx(TCGMemOp size, int label1)
620 {
621 tcg_gen_mov_tl(cpu_tmp0, cpu_regs[R_ECX]);
622 gen_extu(size, cpu_tmp0);
623 tcg_gen_brcondi_tl(TCG_COND_EQ, cpu_tmp0, 0, label1);
624 }
625
626 static void gen_helper_in_func(TCGMemOp ot, TCGv v, TCGv_i32 n)
627 {
628 switch (ot) {
629 case MO_8:
630 gen_helper_inb(v, n);
631 break;
632 case MO_16:
633 gen_helper_inw(v, n);
634 break;
635 case MO_32:
636 gen_helper_inl(v, n);
637 break;
638 default:
639 tcg_abort();
640 }
641 }
642
643 static void gen_helper_out_func(TCGMemOp ot, TCGv_i32 v, TCGv_i32 n)
644 {
645 switch (ot) {
646 case MO_8:
647 gen_helper_outb(v, n);
648 break;
649 case MO_16:
650 gen_helper_outw(v, n);
651 break;
652 case MO_32:
653 gen_helper_outl(v, n);
654 break;
655 default:
656 tcg_abort();
657 }
658 }
659
660 static void gen_check_io(DisasContext *s, TCGMemOp ot, target_ulong cur_eip,
661 uint32_t svm_flags)
662 {
663 int state_saved;
664 target_ulong next_eip;
665
666 state_saved = 0;
667 if (s->pe && (s->cpl > s->iopl || s->vm86)) {
668 gen_update_cc_op(s);
669 gen_jmp_im(cur_eip);
670 state_saved = 1;
671 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
672 switch (ot) {
673 case MO_8:
674 gen_helper_check_iob(cpu_env, cpu_tmp2_i32);
675 break;
676 case MO_16:
677 gen_helper_check_iow(cpu_env, cpu_tmp2_i32);
678 break;
679 case MO_32:
680 gen_helper_check_iol(cpu_env, cpu_tmp2_i32);
681 break;
682 default:
683 tcg_abort();
684 }
685 }
686 if(s->flags & HF_SVMI_MASK) {
687 if (!state_saved) {
688 gen_update_cc_op(s);
689 gen_jmp_im(cur_eip);
690 }
691 svm_flags |= (1 << (4 + ot));
692 next_eip = s->pc - s->cs_base;
693 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
694 gen_helper_svm_check_io(cpu_env, cpu_tmp2_i32,
695 tcg_const_i32(svm_flags),
696 tcg_const_i32(next_eip - cur_eip));
697 }
698 }
699
700 static inline void gen_movs(DisasContext *s, TCGMemOp ot)
701 {
702 gen_string_movl_A0_ESI(s);
703 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
704 gen_string_movl_A0_EDI(s);
705 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
706 gen_op_movl_T0_Dshift(ot);
707 gen_op_add_reg_T0(s->aflag, R_ESI);
708 gen_op_add_reg_T0(s->aflag, R_EDI);
709 }
710
711 static void gen_op_update1_cc(void)
712 {
713 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
714 }
715
716 static void gen_op_update2_cc(void)
717 {
718 tcg_gen_mov_tl(cpu_cc_src, cpu_T[1]);
719 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
720 }
721
722 static void gen_op_update3_cc(TCGv reg)
723 {
724 tcg_gen_mov_tl(cpu_cc_src2, reg);
725 tcg_gen_mov_tl(cpu_cc_src, cpu_T[1]);
726 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
727 }
728
729 static inline void gen_op_testl_T0_T1_cc(void)
730 {
731 tcg_gen_and_tl(cpu_cc_dst, cpu_T[0], cpu_T[1]);
732 }
733
734 static void gen_op_update_neg_cc(void)
735 {
736 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
737 tcg_gen_neg_tl(cpu_cc_src, cpu_T[0]);
738 tcg_gen_movi_tl(cpu_cc_srcT, 0);
739 }
740
741 /* compute all eflags to cc_src */
742 static void gen_compute_eflags(DisasContext *s)
743 {
744 TCGv zero, dst, src1, src2;
745 int live, dead;
746
747 if (s->cc_op == CC_OP_EFLAGS) {
748 return;
749 }
750 if (s->cc_op == CC_OP_CLR) {
751 tcg_gen_movi_tl(cpu_cc_src, CC_Z | CC_P);
752 set_cc_op(s, CC_OP_EFLAGS);
753 return;
754 }
755
756 TCGV_UNUSED(zero);
757 dst = cpu_cc_dst;
758 src1 = cpu_cc_src;
759 src2 = cpu_cc_src2;
760
761 /* Take care to not read values that are not live. */
762 live = cc_op_live[s->cc_op] & ~USES_CC_SRCT;
763 dead = live ^ (USES_CC_DST | USES_CC_SRC | USES_CC_SRC2);
764 if (dead) {
765 zero = tcg_const_tl(0);
766 if (dead & USES_CC_DST) {
767 dst = zero;
768 }
769 if (dead & USES_CC_SRC) {
770 src1 = zero;
771 }
772 if (dead & USES_CC_SRC2) {
773 src2 = zero;
774 }
775 }
776
777 gen_update_cc_op(s);
778 gen_helper_cc_compute_all(cpu_cc_src, dst, src1, src2, cpu_cc_op);
779 set_cc_op(s, CC_OP_EFLAGS);
780
781 if (dead) {
782 tcg_temp_free(zero);
783 }
784 }
785
786 typedef struct CCPrepare {
787 TCGCond cond;
788 TCGv reg;
789 TCGv reg2;
790 target_ulong imm;
791 target_ulong mask;
792 bool use_reg2;
793 bool no_setcond;
794 } CCPrepare;
795
796 /* compute eflags.C to reg */
797 static CCPrepare gen_prepare_eflags_c(DisasContext *s, TCGv reg)
798 {
799 TCGv t0, t1;
800 int size, shift;
801
802 switch (s->cc_op) {
803 case CC_OP_SUBB ... CC_OP_SUBQ:
804 /* (DATA_TYPE)CC_SRCT < (DATA_TYPE)CC_SRC */
805 size = s->cc_op - CC_OP_SUBB;
806 t1 = gen_ext_tl(cpu_tmp0, cpu_cc_src, size, false);
807 /* If no temporary was used, be careful not to alias t1 and t0. */
808 t0 = TCGV_EQUAL(t1, cpu_cc_src) ? cpu_tmp0 : reg;
809 tcg_gen_mov_tl(t0, cpu_cc_srcT);
810 gen_extu(size, t0);
811 goto add_sub;
812
813 case CC_OP_ADDB ... CC_OP_ADDQ:
814 /* (DATA_TYPE)CC_DST < (DATA_TYPE)CC_SRC */
815 size = s->cc_op - CC_OP_ADDB;
816 t1 = gen_ext_tl(cpu_tmp0, cpu_cc_src, size, false);
817 t0 = gen_ext_tl(reg, cpu_cc_dst, size, false);
818 add_sub:
819 return (CCPrepare) { .cond = TCG_COND_LTU, .reg = t0,
820 .reg2 = t1, .mask = -1, .use_reg2 = true };
821
822 case CC_OP_LOGICB ... CC_OP_LOGICQ:
823 case CC_OP_CLR:
824 return (CCPrepare) { .cond = TCG_COND_NEVER, .mask = -1 };
825
826 case CC_OP_INCB ... CC_OP_INCQ:
827 case CC_OP_DECB ... CC_OP_DECQ:
828 return (CCPrepare) { .cond = TCG_COND_NE, .reg = cpu_cc_src,
829 .mask = -1, .no_setcond = true };
830
831 case CC_OP_SHLB ... CC_OP_SHLQ:
832 /* (CC_SRC >> (DATA_BITS - 1)) & 1 */
833 size = s->cc_op - CC_OP_SHLB;
834 shift = (8 << size) - 1;
835 return (CCPrepare) { .cond = TCG_COND_NE, .reg = cpu_cc_src,
836 .mask = (target_ulong)1 << shift };
837
838 case CC_OP_MULB ... CC_OP_MULQ:
839 return (CCPrepare) { .cond = TCG_COND_NE,
840 .reg = cpu_cc_src, .mask = -1 };
841
842 case CC_OP_BMILGB ... CC_OP_BMILGQ:
843 size = s->cc_op - CC_OP_BMILGB;
844 t0 = gen_ext_tl(reg, cpu_cc_src, size, false);
845 return (CCPrepare) { .cond = TCG_COND_EQ, .reg = t0, .mask = -1 };
846
847 case CC_OP_ADCX:
848 case CC_OP_ADCOX:
849 return (CCPrepare) { .cond = TCG_COND_NE, .reg = cpu_cc_dst,
850 .mask = -1, .no_setcond = true };
851
852 case CC_OP_EFLAGS:
853 case CC_OP_SARB ... CC_OP_SARQ:
854 /* CC_SRC & 1 */
855 return (CCPrepare) { .cond = TCG_COND_NE,
856 .reg = cpu_cc_src, .mask = CC_C };
857
858 default:
859 /* The need to compute only C from CC_OP_DYNAMIC is important
860 in efficiently implementing e.g. INC at the start of a TB. */
861 gen_update_cc_op(s);
862 gen_helper_cc_compute_c(reg, cpu_cc_dst, cpu_cc_src,
863 cpu_cc_src2, cpu_cc_op);
864 return (CCPrepare) { .cond = TCG_COND_NE, .reg = reg,
865 .mask = -1, .no_setcond = true };
866 }
867 }
868
869 /* compute eflags.P to reg */
870 static CCPrepare gen_prepare_eflags_p(DisasContext *s, TCGv reg)
871 {
872 gen_compute_eflags(s);
873 return (CCPrepare) { .cond = TCG_COND_NE, .reg = cpu_cc_src,
874 .mask = CC_P };
875 }
876
877 /* compute eflags.S to reg */
878 static CCPrepare gen_prepare_eflags_s(DisasContext *s, TCGv reg)
879 {
880 switch (s->cc_op) {
881 case CC_OP_DYNAMIC:
882 gen_compute_eflags(s);
883 /* FALLTHRU */
884 case CC_OP_EFLAGS:
885 case CC_OP_ADCX:
886 case CC_OP_ADOX:
887 case CC_OP_ADCOX:
888 return (CCPrepare) { .cond = TCG_COND_NE, .reg = cpu_cc_src,
889 .mask = CC_S };
890 case CC_OP_CLR:
891 return (CCPrepare) { .cond = TCG_COND_NEVER, .mask = -1 };
892 default:
893 {
894 TCGMemOp size = (s->cc_op - CC_OP_ADDB) & 3;
895 TCGv t0 = gen_ext_tl(reg, cpu_cc_dst, size, true);
896 return (CCPrepare) { .cond = TCG_COND_LT, .reg = t0, .mask = -1 };
897 }
898 }
899 }
900
901 /* compute eflags.O to reg */
902 static CCPrepare gen_prepare_eflags_o(DisasContext *s, TCGv reg)
903 {
904 switch (s->cc_op) {
905 case CC_OP_ADOX:
906 case CC_OP_ADCOX:
907 return (CCPrepare) { .cond = TCG_COND_NE, .reg = cpu_cc_src2,
908 .mask = -1, .no_setcond = true };
909 case CC_OP_CLR:
910 return (CCPrepare) { .cond = TCG_COND_NEVER, .mask = -1 };
911 default:
912 gen_compute_eflags(s);
913 return (CCPrepare) { .cond = TCG_COND_NE, .reg = cpu_cc_src,
914 .mask = CC_O };
915 }
916 }
917
918 /* compute eflags.Z to reg */
919 static CCPrepare gen_prepare_eflags_z(DisasContext *s, TCGv reg)
920 {
921 switch (s->cc_op) {
922 case CC_OP_DYNAMIC:
923 gen_compute_eflags(s);
924 /* FALLTHRU */
925 case CC_OP_EFLAGS:
926 case CC_OP_ADCX:
927 case CC_OP_ADOX:
928 case CC_OP_ADCOX:
929 return (CCPrepare) { .cond = TCG_COND_NE, .reg = cpu_cc_src,
930 .mask = CC_Z };
931 case CC_OP_CLR:
932 return (CCPrepare) { .cond = TCG_COND_ALWAYS, .mask = -1 };
933 default:
934 {
935 TCGMemOp size = (s->cc_op - CC_OP_ADDB) & 3;
936 TCGv t0 = gen_ext_tl(reg, cpu_cc_dst, size, false);
937 return (CCPrepare) { .cond = TCG_COND_EQ, .reg = t0, .mask = -1 };
938 }
939 }
940 }
941
942 /* perform a conditional store into register 'reg' according to jump opcode
943 value 'b'. In the fast case, T0 is guaranted not to be used. */
944 static CCPrepare gen_prepare_cc(DisasContext *s, int b, TCGv reg)
945 {
946 int inv, jcc_op, cond;
947 TCGMemOp size;
948 CCPrepare cc;
949 TCGv t0;
950
951 inv = b & 1;
952 jcc_op = (b >> 1) & 7;
953
954 switch (s->cc_op) {
955 case CC_OP_SUBB ... CC_OP_SUBQ:
956 /* We optimize relational operators for the cmp/jcc case. */
957 size = s->cc_op - CC_OP_SUBB;
958 switch (jcc_op) {
959 case JCC_BE:
960 tcg_gen_mov_tl(cpu_tmp4, cpu_cc_srcT);
961 gen_extu(size, cpu_tmp4);
962 t0 = gen_ext_tl(cpu_tmp0, cpu_cc_src, size, false);
963 cc = (CCPrepare) { .cond = TCG_COND_LEU, .reg = cpu_tmp4,
964 .reg2 = t0, .mask = -1, .use_reg2 = true };
965 break;
966
967 case JCC_L:
968 cond = TCG_COND_LT;
969 goto fast_jcc_l;
970 case JCC_LE:
971 cond = TCG_COND_LE;
972 fast_jcc_l:
973 tcg_gen_mov_tl(cpu_tmp4, cpu_cc_srcT);
974 gen_exts(size, cpu_tmp4);
975 t0 = gen_ext_tl(cpu_tmp0, cpu_cc_src, size, true);
976 cc = (CCPrepare) { .cond = cond, .reg = cpu_tmp4,
977 .reg2 = t0, .mask = -1, .use_reg2 = true };
978 break;
979
980 default:
981 goto slow_jcc;
982 }
983 break;
984
985 default:
986 slow_jcc:
987 /* This actually generates good code for JC, JZ and JS. */
988 switch (jcc_op) {
989 case JCC_O:
990 cc = gen_prepare_eflags_o(s, reg);
991 break;
992 case JCC_B:
993 cc = gen_prepare_eflags_c(s, reg);
994 break;
995 case JCC_Z:
996 cc = gen_prepare_eflags_z(s, reg);
997 break;
998 case JCC_BE:
999 gen_compute_eflags(s);
1000 cc = (CCPrepare) { .cond = TCG_COND_NE, .reg = cpu_cc_src,
1001 .mask = CC_Z | CC_C };
1002 break;
1003 case JCC_S:
1004 cc = gen_prepare_eflags_s(s, reg);
1005 break;
1006 case JCC_P:
1007 cc = gen_prepare_eflags_p(s, reg);
1008 break;
1009 case JCC_L:
1010 gen_compute_eflags(s);
1011 if (TCGV_EQUAL(reg, cpu_cc_src)) {
1012 reg = cpu_tmp0;
1013 }
1014 tcg_gen_shri_tl(reg, cpu_cc_src, 4); /* CC_O -> CC_S */
1015 tcg_gen_xor_tl(reg, reg, cpu_cc_src);
1016 cc = (CCPrepare) { .cond = TCG_COND_NE, .reg = reg,
1017 .mask = CC_S };
1018 break;
1019 default:
1020 case JCC_LE:
1021 gen_compute_eflags(s);
1022 if (TCGV_EQUAL(reg, cpu_cc_src)) {
1023 reg = cpu_tmp0;
1024 }
1025 tcg_gen_shri_tl(reg, cpu_cc_src, 4); /* CC_O -> CC_S */
1026 tcg_gen_xor_tl(reg, reg, cpu_cc_src);
1027 cc = (CCPrepare) { .cond = TCG_COND_NE, .reg = reg,
1028 .mask = CC_S | CC_Z };
1029 break;
1030 }
1031 break;
1032 }
1033
1034 if (inv) {
1035 cc.cond = tcg_invert_cond(cc.cond);
1036 }
1037 return cc;
1038 }
1039
1040 static void gen_setcc1(DisasContext *s, int b, TCGv reg)
1041 {
1042 CCPrepare cc = gen_prepare_cc(s, b, reg);
1043
1044 if (cc.no_setcond) {
1045 if (cc.cond == TCG_COND_EQ) {
1046 tcg_gen_xori_tl(reg, cc.reg, 1);
1047 } else {
1048 tcg_gen_mov_tl(reg, cc.reg);
1049 }
1050 return;
1051 }
1052
1053 if (cc.cond == TCG_COND_NE && !cc.use_reg2 && cc.imm == 0 &&
1054 cc.mask != 0 && (cc.mask & (cc.mask - 1)) == 0) {
1055 tcg_gen_shri_tl(reg, cc.reg, ctztl(cc.mask));
1056 tcg_gen_andi_tl(reg, reg, 1);
1057 return;
1058 }
1059 if (cc.mask != -1) {
1060 tcg_gen_andi_tl(reg, cc.reg, cc.mask);
1061 cc.reg = reg;
1062 }
1063 if (cc.use_reg2) {
1064 tcg_gen_setcond_tl(cc.cond, reg, cc.reg, cc.reg2);
1065 } else {
1066 tcg_gen_setcondi_tl(cc.cond, reg, cc.reg, cc.imm);
1067 }
1068 }
1069
1070 static inline void gen_compute_eflags_c(DisasContext *s, TCGv reg)
1071 {
1072 gen_setcc1(s, JCC_B << 1, reg);
1073 }
1074
1075 /* generate a conditional jump to label 'l1' according to jump opcode
1076 value 'b'. In the fast case, T0 is guaranted not to be used. */
1077 static inline void gen_jcc1_noeob(DisasContext *s, int b, int l1)
1078 {
1079 CCPrepare cc = gen_prepare_cc(s, b, cpu_T[0]);
1080
1081 if (cc.mask != -1) {
1082 tcg_gen_andi_tl(cpu_T[0], cc.reg, cc.mask);
1083 cc.reg = cpu_T[0];
1084 }
1085 if (cc.use_reg2) {
1086 tcg_gen_brcond_tl(cc.cond, cc.reg, cc.reg2, l1);
1087 } else {
1088 tcg_gen_brcondi_tl(cc.cond, cc.reg, cc.imm, l1);
1089 }
1090 }
1091
1092 /* Generate a conditional jump to label 'l1' according to jump opcode
1093 value 'b'. In the fast case, T0 is guaranted not to be used.
1094 A translation block must end soon. */
1095 static inline void gen_jcc1(DisasContext *s, int b, int l1)
1096 {
1097 CCPrepare cc = gen_prepare_cc(s, b, cpu_T[0]);
1098
1099 gen_update_cc_op(s);
1100 if (cc.mask != -1) {
1101 tcg_gen_andi_tl(cpu_T[0], cc.reg, cc.mask);
1102 cc.reg = cpu_T[0];
1103 }
1104 set_cc_op(s, CC_OP_DYNAMIC);
1105 if (cc.use_reg2) {
1106 tcg_gen_brcond_tl(cc.cond, cc.reg, cc.reg2, l1);
1107 } else {
1108 tcg_gen_brcondi_tl(cc.cond, cc.reg, cc.imm, l1);
1109 }
1110 }
1111
1112 /* XXX: does not work with gdbstub "ice" single step - not a
1113 serious problem */
1114 static int gen_jz_ecx_string(DisasContext *s, target_ulong next_eip)
1115 {
1116 int l1, l2;
1117
1118 l1 = gen_new_label();
1119 l2 = gen_new_label();
1120 gen_op_jnz_ecx(s->aflag, l1);
1121 gen_set_label(l2);
1122 gen_jmp_tb(s, next_eip, 1);
1123 gen_set_label(l1);
1124 return l2;
1125 }
1126
1127 static inline void gen_stos(DisasContext *s, TCGMemOp ot)
1128 {
1129 gen_op_mov_v_reg(MO_32, cpu_T[0], R_EAX);
1130 gen_string_movl_A0_EDI(s);
1131 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
1132 gen_op_movl_T0_Dshift(ot);
1133 gen_op_add_reg_T0(s->aflag, R_EDI);
1134 }
1135
1136 static inline void gen_lods(DisasContext *s, TCGMemOp ot)
1137 {
1138 gen_string_movl_A0_ESI(s);
1139 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
1140 gen_op_mov_reg_v(ot, R_EAX, cpu_T[0]);
1141 gen_op_movl_T0_Dshift(ot);
1142 gen_op_add_reg_T0(s->aflag, R_ESI);
1143 }
1144
1145 static inline void gen_scas(DisasContext *s, TCGMemOp ot)
1146 {
1147 gen_string_movl_A0_EDI(s);
1148 gen_op_ld_v(s, ot, cpu_T[1], cpu_A0);
1149 gen_op(s, OP_CMPL, ot, R_EAX);
1150 gen_op_movl_T0_Dshift(ot);
1151 gen_op_add_reg_T0(s->aflag, R_EDI);
1152 }
1153
1154 static inline void gen_cmps(DisasContext *s, TCGMemOp ot)
1155 {
1156 gen_string_movl_A0_EDI(s);
1157 gen_op_ld_v(s, ot, cpu_T[1], cpu_A0);
1158 gen_string_movl_A0_ESI(s);
1159 gen_op(s, OP_CMPL, ot, OR_TMP0);
1160 gen_op_movl_T0_Dshift(ot);
1161 gen_op_add_reg_T0(s->aflag, R_ESI);
1162 gen_op_add_reg_T0(s->aflag, R_EDI);
1163 }
1164
1165 static inline void gen_ins(DisasContext *s, TCGMemOp ot)
1166 {
1167 if (use_icount)
1168 gen_io_start();
1169 gen_string_movl_A0_EDI(s);
1170 /* Note: we must do this dummy write first to be restartable in
1171 case of page fault. */
1172 tcg_gen_movi_tl(cpu_T[0], 0);
1173 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
1174 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_regs[R_EDX]);
1175 tcg_gen_andi_i32(cpu_tmp2_i32, cpu_tmp2_i32, 0xffff);
1176 gen_helper_in_func(ot, cpu_T[0], cpu_tmp2_i32);
1177 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
1178 gen_op_movl_T0_Dshift(ot);
1179 gen_op_add_reg_T0(s->aflag, R_EDI);
1180 if (use_icount)
1181 gen_io_end();
1182 }
1183
1184 static inline void gen_outs(DisasContext *s, TCGMemOp ot)
1185 {
1186 if (use_icount)
1187 gen_io_start();
1188 gen_string_movl_A0_ESI(s);
1189 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
1190
1191 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_regs[R_EDX]);
1192 tcg_gen_andi_i32(cpu_tmp2_i32, cpu_tmp2_i32, 0xffff);
1193 tcg_gen_trunc_tl_i32(cpu_tmp3_i32, cpu_T[0]);
1194 gen_helper_out_func(ot, cpu_tmp2_i32, cpu_tmp3_i32);
1195
1196 gen_op_movl_T0_Dshift(ot);
1197 gen_op_add_reg_T0(s->aflag, R_ESI);
1198 if (use_icount)
1199 gen_io_end();
1200 }
1201
1202 /* same method as Valgrind : we generate jumps to current or next
1203 instruction */
1204 #define GEN_REPZ(op) \
1205 static inline void gen_repz_ ## op(DisasContext *s, TCGMemOp ot, \
1206 target_ulong cur_eip, target_ulong next_eip) \
1207 { \
1208 int l2;\
1209 gen_update_cc_op(s); \
1210 l2 = gen_jz_ecx_string(s, next_eip); \
1211 gen_ ## op(s, ot); \
1212 gen_op_add_reg_im(s->aflag, R_ECX, -1); \
1213 /* a loop would cause two single step exceptions if ECX = 1 \
1214 before rep string_insn */ \
1215 if (!s->jmp_opt) \
1216 gen_op_jz_ecx(s->aflag, l2); \
1217 gen_jmp(s, cur_eip); \
1218 }
1219
1220 #define GEN_REPZ2(op) \
1221 static inline void gen_repz_ ## op(DisasContext *s, TCGMemOp ot, \
1222 target_ulong cur_eip, \
1223 target_ulong next_eip, \
1224 int nz) \
1225 { \
1226 int l2;\
1227 gen_update_cc_op(s); \
1228 l2 = gen_jz_ecx_string(s, next_eip); \
1229 gen_ ## op(s, ot); \
1230 gen_op_add_reg_im(s->aflag, R_ECX, -1); \
1231 gen_update_cc_op(s); \
1232 gen_jcc1(s, (JCC_Z << 1) | (nz ^ 1), l2); \
1233 if (!s->jmp_opt) \
1234 gen_op_jz_ecx(s->aflag, l2); \
1235 gen_jmp(s, cur_eip); \
1236 }
1237
1238 GEN_REPZ(movs)
1239 GEN_REPZ(stos)
1240 GEN_REPZ(lods)
1241 GEN_REPZ(ins)
1242 GEN_REPZ(outs)
1243 GEN_REPZ2(scas)
1244 GEN_REPZ2(cmps)
1245
1246 static void gen_helper_fp_arith_ST0_FT0(int op)
1247 {
1248 switch (op) {
1249 case 0:
1250 gen_helper_fadd_ST0_FT0(cpu_env);
1251 break;
1252 case 1:
1253 gen_helper_fmul_ST0_FT0(cpu_env);
1254 break;
1255 case 2:
1256 gen_helper_fcom_ST0_FT0(cpu_env);
1257 break;
1258 case 3:
1259 gen_helper_fcom_ST0_FT0(cpu_env);
1260 break;
1261 case 4:
1262 gen_helper_fsub_ST0_FT0(cpu_env);
1263 break;
1264 case 5:
1265 gen_helper_fsubr_ST0_FT0(cpu_env);
1266 break;
1267 case 6:
1268 gen_helper_fdiv_ST0_FT0(cpu_env);
1269 break;
1270 case 7:
1271 gen_helper_fdivr_ST0_FT0(cpu_env);
1272 break;
1273 }
1274 }
1275
1276 /* NOTE the exception in "r" op ordering */
1277 static void gen_helper_fp_arith_STN_ST0(int op, int opreg)
1278 {
1279 TCGv_i32 tmp = tcg_const_i32(opreg);
1280 switch (op) {
1281 case 0:
1282 gen_helper_fadd_STN_ST0(cpu_env, tmp);
1283 break;
1284 case 1:
1285 gen_helper_fmul_STN_ST0(cpu_env, tmp);
1286 break;
1287 case 4:
1288 gen_helper_fsubr_STN_ST0(cpu_env, tmp);
1289 break;
1290 case 5:
1291 gen_helper_fsub_STN_ST0(cpu_env, tmp);
1292 break;
1293 case 6:
1294 gen_helper_fdivr_STN_ST0(cpu_env, tmp);
1295 break;
1296 case 7:
1297 gen_helper_fdiv_STN_ST0(cpu_env, tmp);
1298 break;
1299 }
1300 }
1301
1302 /* if d == OR_TMP0, it means memory operand (address in A0) */
1303 static void gen_op(DisasContext *s1, int op, TCGMemOp ot, int d)
1304 {
1305 if (d != OR_TMP0) {
1306 gen_op_mov_v_reg(ot, cpu_T[0], d);
1307 } else {
1308 gen_op_ld_v(s1, ot, cpu_T[0], cpu_A0);
1309 }
1310 switch(op) {
1311 case OP_ADCL:
1312 gen_compute_eflags_c(s1, cpu_tmp4);
1313 tcg_gen_add_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
1314 tcg_gen_add_tl(cpu_T[0], cpu_T[0], cpu_tmp4);
1315 gen_op_st_rm_T0_A0(s1, ot, d);
1316 gen_op_update3_cc(cpu_tmp4);
1317 set_cc_op(s1, CC_OP_ADCB + ot);
1318 break;
1319 case OP_SBBL:
1320 gen_compute_eflags_c(s1, cpu_tmp4);
1321 tcg_gen_sub_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
1322 tcg_gen_sub_tl(cpu_T[0], cpu_T[0], cpu_tmp4);
1323 gen_op_st_rm_T0_A0(s1, ot, d);
1324 gen_op_update3_cc(cpu_tmp4);
1325 set_cc_op(s1, CC_OP_SBBB + ot);
1326 break;
1327 case OP_ADDL:
1328 tcg_gen_add_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
1329 gen_op_st_rm_T0_A0(s1, ot, d);
1330 gen_op_update2_cc();
1331 set_cc_op(s1, CC_OP_ADDB + ot);
1332 break;
1333 case OP_SUBL:
1334 tcg_gen_mov_tl(cpu_cc_srcT, cpu_T[0]);
1335 tcg_gen_sub_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
1336 gen_op_st_rm_T0_A0(s1, ot, d);
1337 gen_op_update2_cc();
1338 set_cc_op(s1, CC_OP_SUBB + ot);
1339 break;
1340 default:
1341 case OP_ANDL:
1342 tcg_gen_and_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
1343 gen_op_st_rm_T0_A0(s1, ot, d);
1344 gen_op_update1_cc();
1345 set_cc_op(s1, CC_OP_LOGICB + ot);
1346 break;
1347 case OP_ORL:
1348 tcg_gen_or_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
1349 gen_op_st_rm_T0_A0(s1, ot, d);
1350 gen_op_update1_cc();
1351 set_cc_op(s1, CC_OP_LOGICB + ot);
1352 break;
1353 case OP_XORL:
1354 tcg_gen_xor_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
1355 gen_op_st_rm_T0_A0(s1, ot, d);
1356 gen_op_update1_cc();
1357 set_cc_op(s1, CC_OP_LOGICB + ot);
1358 break;
1359 case OP_CMPL:
1360 tcg_gen_mov_tl(cpu_cc_src, cpu_T[1]);
1361 tcg_gen_mov_tl(cpu_cc_srcT, cpu_T[0]);
1362 tcg_gen_sub_tl(cpu_cc_dst, cpu_T[0], cpu_T[1]);
1363 set_cc_op(s1, CC_OP_SUBB + ot);
1364 break;
1365 }
1366 }
1367
1368 /* if d == OR_TMP0, it means memory operand (address in A0) */
1369 static void gen_inc(DisasContext *s1, TCGMemOp ot, int d, int c)
1370 {
1371 if (d != OR_TMP0) {
1372 gen_op_mov_v_reg(ot, cpu_T[0], d);
1373 } else {
1374 gen_op_ld_v(s1, ot, cpu_T[0], cpu_A0);
1375 }
1376 gen_compute_eflags_c(s1, cpu_cc_src);
1377 if (c > 0) {
1378 tcg_gen_addi_tl(cpu_T[0], cpu_T[0], 1);
1379 set_cc_op(s1, CC_OP_INCB + ot);
1380 } else {
1381 tcg_gen_addi_tl(cpu_T[0], cpu_T[0], -1);
1382 set_cc_op(s1, CC_OP_DECB + ot);
1383 }
1384 gen_op_st_rm_T0_A0(s1, ot, d);
1385 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
1386 }
1387
1388 static void gen_shift_flags(DisasContext *s, TCGMemOp ot, TCGv result,
1389 TCGv shm1, TCGv count, bool is_right)
1390 {
1391 TCGv_i32 z32, s32, oldop;
1392 TCGv z_tl;
1393
1394 /* Store the results into the CC variables. If we know that the
1395 variable must be dead, store unconditionally. Otherwise we'll
1396 need to not disrupt the current contents. */
1397 z_tl = tcg_const_tl(0);
1398 if (cc_op_live[s->cc_op] & USES_CC_DST) {
1399 tcg_gen_movcond_tl(TCG_COND_NE, cpu_cc_dst, count, z_tl,
1400 result, cpu_cc_dst);
1401 } else {
1402 tcg_gen_mov_tl(cpu_cc_dst, result);
1403 }
1404 if (cc_op_live[s->cc_op] & USES_CC_SRC) {
1405 tcg_gen_movcond_tl(TCG_COND_NE, cpu_cc_src, count, z_tl,
1406 shm1, cpu_cc_src);
1407 } else {
1408 tcg_gen_mov_tl(cpu_cc_src, shm1);
1409 }
1410 tcg_temp_free(z_tl);
1411
1412 /* Get the two potential CC_OP values into temporaries. */
1413 tcg_gen_movi_i32(cpu_tmp2_i32, (is_right ? CC_OP_SARB : CC_OP_SHLB) + ot);
1414 if (s->cc_op == CC_OP_DYNAMIC) {
1415 oldop = cpu_cc_op;
1416 } else {
1417 tcg_gen_movi_i32(cpu_tmp3_i32, s->cc_op);
1418 oldop = cpu_tmp3_i32;
1419 }
1420
1421 /* Conditionally store the CC_OP value. */
1422 z32 = tcg_const_i32(0);
1423 s32 = tcg_temp_new_i32();
1424 tcg_gen_trunc_tl_i32(s32, count);
1425 tcg_gen_movcond_i32(TCG_COND_NE, cpu_cc_op, s32, z32, cpu_tmp2_i32, oldop);
1426 tcg_temp_free_i32(z32);
1427 tcg_temp_free_i32(s32);
1428
1429 /* The CC_OP value is no longer predictable. */
1430 set_cc_op(s, CC_OP_DYNAMIC);
1431 }
1432
1433 static void gen_shift_rm_T1(DisasContext *s, TCGMemOp ot, int op1,
1434 int is_right, int is_arith)
1435 {
1436 target_ulong mask = (ot == MO_64 ? 0x3f : 0x1f);
1437
1438 /* load */
1439 if (op1 == OR_TMP0) {
1440 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
1441 } else {
1442 gen_op_mov_v_reg(ot, cpu_T[0], op1);
1443 }
1444
1445 tcg_gen_andi_tl(cpu_T[1], cpu_T[1], mask);
1446 tcg_gen_subi_tl(cpu_tmp0, cpu_T[1], 1);
1447
1448 if (is_right) {
1449 if (is_arith) {
1450 gen_exts(ot, cpu_T[0]);
1451 tcg_gen_sar_tl(cpu_tmp0, cpu_T[0], cpu_tmp0);
1452 tcg_gen_sar_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
1453 } else {
1454 gen_extu(ot, cpu_T[0]);
1455 tcg_gen_shr_tl(cpu_tmp0, cpu_T[0], cpu_tmp0);
1456 tcg_gen_shr_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
1457 }
1458 } else {
1459 tcg_gen_shl_tl(cpu_tmp0, cpu_T[0], cpu_tmp0);
1460 tcg_gen_shl_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
1461 }
1462
1463 /* store */
1464 gen_op_st_rm_T0_A0(s, ot, op1);
1465
1466 gen_shift_flags(s, ot, cpu_T[0], cpu_tmp0, cpu_T[1], is_right);
1467 }
1468
1469 static void gen_shift_rm_im(DisasContext *s, TCGMemOp ot, int op1, int op2,
1470 int is_right, int is_arith)
1471 {
1472 int mask = (ot == MO_64 ? 0x3f : 0x1f);
1473
1474 /* load */
1475 if (op1 == OR_TMP0)
1476 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
1477 else
1478 gen_op_mov_v_reg(ot, cpu_T[0], op1);
1479
1480 op2 &= mask;
1481 if (op2 != 0) {
1482 if (is_right) {
1483 if (is_arith) {
1484 gen_exts(ot, cpu_T[0]);
1485 tcg_gen_sari_tl(cpu_tmp4, cpu_T[0], op2 - 1);
1486 tcg_gen_sari_tl(cpu_T[0], cpu_T[0], op2);
1487 } else {
1488 gen_extu(ot, cpu_T[0]);
1489 tcg_gen_shri_tl(cpu_tmp4, cpu_T[0], op2 - 1);
1490 tcg_gen_shri_tl(cpu_T[0], cpu_T[0], op2);
1491 }
1492 } else {
1493 tcg_gen_shli_tl(cpu_tmp4, cpu_T[0], op2 - 1);
1494 tcg_gen_shli_tl(cpu_T[0], cpu_T[0], op2);
1495 }
1496 }
1497
1498 /* store */
1499 gen_op_st_rm_T0_A0(s, ot, op1);
1500
1501 /* update eflags if non zero shift */
1502 if (op2 != 0) {
1503 tcg_gen_mov_tl(cpu_cc_src, cpu_tmp4);
1504 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
1505 set_cc_op(s, (is_right ? CC_OP_SARB : CC_OP_SHLB) + ot);
1506 }
1507 }
1508
1509 static inline void tcg_gen_lshift(TCGv ret, TCGv arg1, target_long arg2)
1510 {
1511 if (arg2 >= 0)
1512 tcg_gen_shli_tl(ret, arg1, arg2);
1513 else
1514 tcg_gen_shri_tl(ret, arg1, -arg2);
1515 }
1516
1517 static void gen_rot_rm_T1(DisasContext *s, TCGMemOp ot, int op1, int is_right)
1518 {
1519 target_ulong mask = (ot == MO_64 ? 0x3f : 0x1f);
1520 TCGv_i32 t0, t1;
1521
1522 /* load */
1523 if (op1 == OR_TMP0) {
1524 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
1525 } else {
1526 gen_op_mov_v_reg(ot, cpu_T[0], op1);
1527 }
1528
1529 tcg_gen_andi_tl(cpu_T[1], cpu_T[1], mask);
1530
1531 switch (ot) {
1532 case MO_8:
1533 /* Replicate the 8-bit input so that a 32-bit rotate works. */
1534 tcg_gen_ext8u_tl(cpu_T[0], cpu_T[0]);
1535 tcg_gen_muli_tl(cpu_T[0], cpu_T[0], 0x01010101);
1536 goto do_long;
1537 case MO_16:
1538 /* Replicate the 16-bit input so that a 32-bit rotate works. */
1539 tcg_gen_deposit_tl(cpu_T[0], cpu_T[0], cpu_T[0], 16, 16);
1540 goto do_long;
1541 do_long:
1542 #ifdef TARGET_X86_64
1543 case MO_32:
1544 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
1545 tcg_gen_trunc_tl_i32(cpu_tmp3_i32, cpu_T[1]);
1546 if (is_right) {
1547 tcg_gen_rotr_i32(cpu_tmp2_i32, cpu_tmp2_i32, cpu_tmp3_i32);
1548 } else {
1549 tcg_gen_rotl_i32(cpu_tmp2_i32, cpu_tmp2_i32, cpu_tmp3_i32);
1550 }
1551 tcg_gen_extu_i32_tl(cpu_T[0], cpu_tmp2_i32);
1552 break;
1553 #endif
1554 default:
1555 if (is_right) {
1556 tcg_gen_rotr_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
1557 } else {
1558 tcg_gen_rotl_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
1559 }
1560 break;
1561 }
1562
1563 /* store */
1564 gen_op_st_rm_T0_A0(s, ot, op1);
1565
1566 /* We'll need the flags computed into CC_SRC. */
1567 gen_compute_eflags(s);
1568
1569 /* The value that was "rotated out" is now present at the other end
1570 of the word. Compute C into CC_DST and O into CC_SRC2. Note that
1571 since we've computed the flags into CC_SRC, these variables are
1572 currently dead. */
1573 if (is_right) {
1574 tcg_gen_shri_tl(cpu_cc_src2, cpu_T[0], mask - 1);
1575 tcg_gen_shri_tl(cpu_cc_dst, cpu_T[0], mask);
1576 tcg_gen_andi_tl(cpu_cc_dst, cpu_cc_dst, 1);
1577 } else {
1578 tcg_gen_shri_tl(cpu_cc_src2, cpu_T[0], mask);
1579 tcg_gen_andi_tl(cpu_cc_dst, cpu_T[0], 1);
1580 }
1581 tcg_gen_andi_tl(cpu_cc_src2, cpu_cc_src2, 1);
1582 tcg_gen_xor_tl(cpu_cc_src2, cpu_cc_src2, cpu_cc_dst);
1583
1584 /* Now conditionally store the new CC_OP value. If the shift count
1585 is 0 we keep the CC_OP_EFLAGS setting so that only CC_SRC is live.
1586 Otherwise reuse CC_OP_ADCOX which have the C and O flags split out
1587 exactly as we computed above. */
1588 t0 = tcg_const_i32(0);
1589 t1 = tcg_temp_new_i32();
1590 tcg_gen_trunc_tl_i32(t1, cpu_T[1]);
1591 tcg_gen_movi_i32(cpu_tmp2_i32, CC_OP_ADCOX);
1592 tcg_gen_movi_i32(cpu_tmp3_i32, CC_OP_EFLAGS);
1593 tcg_gen_movcond_i32(TCG_COND_NE, cpu_cc_op, t1, t0,
1594 cpu_tmp2_i32, cpu_tmp3_i32);
1595 tcg_temp_free_i32(t0);
1596 tcg_temp_free_i32(t1);
1597
1598 /* The CC_OP value is no longer predictable. */
1599 set_cc_op(s, CC_OP_DYNAMIC);
1600 }
1601
1602 static void gen_rot_rm_im(DisasContext *s, TCGMemOp ot, int op1, int op2,
1603 int is_right)
1604 {
1605 int mask = (ot == MO_64 ? 0x3f : 0x1f);
1606 int shift;
1607
1608 /* load */
1609 if (op1 == OR_TMP0) {
1610 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
1611 } else {
1612 gen_op_mov_v_reg(ot, cpu_T[0], op1);
1613 }
1614
1615 op2 &= mask;
1616 if (op2 != 0) {
1617 switch (ot) {
1618 #ifdef TARGET_X86_64
1619 case MO_32:
1620 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
1621 if (is_right) {
1622 tcg_gen_rotri_i32(cpu_tmp2_i32, cpu_tmp2_i32, op2);
1623 } else {
1624 tcg_gen_rotli_i32(cpu_tmp2_i32, cpu_tmp2_i32, op2);
1625 }
1626 tcg_gen_extu_i32_tl(cpu_T[0], cpu_tmp2_i32);
1627 break;
1628 #endif
1629 default:
1630 if (is_right) {
1631 tcg_gen_rotri_tl(cpu_T[0], cpu_T[0], op2);
1632 } else {
1633 tcg_gen_rotli_tl(cpu_T[0], cpu_T[0], op2);
1634 }
1635 break;
1636 case MO_8:
1637 mask = 7;
1638 goto do_shifts;
1639 case MO_16:
1640 mask = 15;
1641 do_shifts:
1642 shift = op2 & mask;
1643 if (is_right) {
1644 shift = mask + 1 - shift;
1645 }
1646 gen_extu(ot, cpu_T[0]);
1647 tcg_gen_shli_tl(cpu_tmp0, cpu_T[0], shift);
1648 tcg_gen_shri_tl(cpu_T[0], cpu_T[0], mask + 1 - shift);
1649 tcg_gen_or_tl(cpu_T[0], cpu_T[0], cpu_tmp0);
1650 break;
1651 }
1652 }
1653
1654 /* store */
1655 gen_op_st_rm_T0_A0(s, ot, op1);
1656
1657 if (op2 != 0) {
1658 /* Compute the flags into CC_SRC. */
1659 gen_compute_eflags(s);
1660
1661 /* The value that was "rotated out" is now present at the other end
1662 of the word. Compute C into CC_DST and O into CC_SRC2. Note that
1663 since we've computed the flags into CC_SRC, these variables are
1664 currently dead. */
1665 if (is_right) {
1666 tcg_gen_shri_tl(cpu_cc_src2, cpu_T[0], mask - 1);
1667 tcg_gen_shri_tl(cpu_cc_dst, cpu_T[0], mask);
1668 tcg_gen_andi_tl(cpu_cc_dst, cpu_cc_dst, 1);
1669 } else {
1670 tcg_gen_shri_tl(cpu_cc_src2, cpu_T[0], mask);
1671 tcg_gen_andi_tl(cpu_cc_dst, cpu_T[0], 1);
1672 }
1673 tcg_gen_andi_tl(cpu_cc_src2, cpu_cc_src2, 1);
1674 tcg_gen_xor_tl(cpu_cc_src2, cpu_cc_src2, cpu_cc_dst);
1675 set_cc_op(s, CC_OP_ADCOX);
1676 }
1677 }
1678
1679 /* XXX: add faster immediate = 1 case */
1680 static void gen_rotc_rm_T1(DisasContext *s, TCGMemOp ot, int op1,
1681 int is_right)
1682 {
1683 gen_compute_eflags(s);
1684 assert(s->cc_op == CC_OP_EFLAGS);
1685
1686 /* load */
1687 if (op1 == OR_TMP0)
1688 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
1689 else
1690 gen_op_mov_v_reg(ot, cpu_T[0], op1);
1691
1692 if (is_right) {
1693 switch (ot) {
1694 case MO_8:
1695 gen_helper_rcrb(cpu_T[0], cpu_env, cpu_T[0], cpu_T[1]);
1696 break;
1697 case MO_16:
1698 gen_helper_rcrw(cpu_T[0], cpu_env, cpu_T[0], cpu_T[1]);
1699 break;
1700 case MO_32:
1701 gen_helper_rcrl(cpu_T[0], cpu_env, cpu_T[0], cpu_T[1]);
1702 break;
1703 #ifdef TARGET_X86_64
1704 case MO_64:
1705 gen_helper_rcrq(cpu_T[0], cpu_env, cpu_T[0], cpu_T[1]);
1706 break;
1707 #endif
1708 default:
1709 tcg_abort();
1710 }
1711 } else {
1712 switch (ot) {
1713 case MO_8:
1714 gen_helper_rclb(cpu_T[0], cpu_env, cpu_T[0], cpu_T[1]);
1715 break;
1716 case MO_16:
1717 gen_helper_rclw(cpu_T[0], cpu_env, cpu_T[0], cpu_T[1]);
1718 break;
1719 case MO_32:
1720 gen_helper_rcll(cpu_T[0], cpu_env, cpu_T[0], cpu_T[1]);
1721 break;
1722 #ifdef TARGET_X86_64
1723 case MO_64:
1724 gen_helper_rclq(cpu_T[0], cpu_env, cpu_T[0], cpu_T[1]);
1725 break;
1726 #endif
1727 default:
1728 tcg_abort();
1729 }
1730 }
1731 /* store */
1732 gen_op_st_rm_T0_A0(s, ot, op1);
1733 }
1734
1735 /* XXX: add faster immediate case */
1736 static void gen_shiftd_rm_T1(DisasContext *s, TCGMemOp ot, int op1,
1737 bool is_right, TCGv count_in)
1738 {
1739 target_ulong mask = (ot == MO_64 ? 63 : 31);
1740 TCGv count;
1741
1742 /* load */
1743 if (op1 == OR_TMP0) {
1744 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
1745 } else {
1746 gen_op_mov_v_reg(ot, cpu_T[0], op1);
1747 }
1748
1749 count = tcg_temp_new();
1750 tcg_gen_andi_tl(count, count_in, mask);
1751
1752 switch (ot) {
1753 case MO_16:
1754 /* Note: we implement the Intel behaviour for shift count > 16.
1755 This means "shrdw C, B, A" shifts A:B:A >> C. Build the B:A
1756 portion by constructing it as a 32-bit value. */
1757 if (is_right) {
1758 tcg_gen_deposit_tl(cpu_tmp0, cpu_T[0], cpu_T[1], 16, 16);
1759 tcg_gen_mov_tl(cpu_T[1], cpu_T[0]);
1760 tcg_gen_mov_tl(cpu_T[0], cpu_tmp0);
1761 } else {
1762 tcg_gen_deposit_tl(cpu_T[1], cpu_T[0], cpu_T[1], 16, 16);
1763 }
1764 /* FALLTHRU */
1765 #ifdef TARGET_X86_64
1766 case MO_32:
1767 /* Concatenate the two 32-bit values and use a 64-bit shift. */
1768 tcg_gen_subi_tl(cpu_tmp0, count, 1);
1769 if (is_right) {
1770 tcg_gen_concat_tl_i64(cpu_T[0], cpu_T[0], cpu_T[1]);
1771 tcg_gen_shr_i64(cpu_tmp0, cpu_T[0], cpu_tmp0);
1772 tcg_gen_shr_i64(cpu_T[0], cpu_T[0], count);
1773 } else {
1774 tcg_gen_concat_tl_i64(cpu_T[0], cpu_T[1], cpu_T[0]);
1775 tcg_gen_shl_i64(cpu_tmp0, cpu_T[0], cpu_tmp0);
1776 tcg_gen_shl_i64(cpu_T[0], cpu_T[0], count);
1777 tcg_gen_shri_i64(cpu_tmp0, cpu_tmp0, 32);
1778 tcg_gen_shri_i64(cpu_T[0], cpu_T[0], 32);
1779 }
1780 break;
1781 #endif
1782 default:
1783 tcg_gen_subi_tl(cpu_tmp0, count, 1);
1784 if (is_right) {
1785 tcg_gen_shr_tl(cpu_tmp0, cpu_T[0], cpu_tmp0);
1786
1787 tcg_gen_subfi_tl(cpu_tmp4, mask + 1, count);
1788 tcg_gen_shr_tl(cpu_T[0], cpu_T[0], count);
1789 tcg_gen_shl_tl(cpu_T[1], cpu_T[1], cpu_tmp4);
1790 } else {
1791 tcg_gen_shl_tl(cpu_tmp0, cpu_T[0], cpu_tmp0);
1792 if (ot == MO_16) {
1793 /* Only needed if count > 16, for Intel behaviour. */
1794 tcg_gen_subfi_tl(cpu_tmp4, 33, count);
1795 tcg_gen_shr_tl(cpu_tmp4, cpu_T[1], cpu_tmp4);
1796 tcg_gen_or_tl(cpu_tmp0, cpu_tmp0, cpu_tmp4);
1797 }
1798
1799 tcg_gen_subfi_tl(cpu_tmp4, mask + 1, count);
1800 tcg_gen_shl_tl(cpu_T[0], cpu_T[0], count);
1801 tcg_gen_shr_tl(cpu_T[1], cpu_T[1], cpu_tmp4);
1802 }
1803 tcg_gen_movi_tl(cpu_tmp4, 0);
1804 tcg_gen_movcond_tl(TCG_COND_EQ, cpu_T[1], count, cpu_tmp4,
1805 cpu_tmp4, cpu_T[1]);
1806 tcg_gen_or_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
1807 break;
1808 }
1809
1810 /* store */
1811 gen_op_st_rm_T0_A0(s, ot, op1);
1812
1813 gen_shift_flags(s, ot, cpu_T[0], cpu_tmp0, count, is_right);
1814 tcg_temp_free(count);
1815 }
1816
1817 static void gen_shift(DisasContext *s1, int op, TCGMemOp ot, int d, int s)
1818 {
1819 if (s != OR_TMP1)
1820 gen_op_mov_v_reg(ot, cpu_T[1], s);
1821 switch(op) {
1822 case OP_ROL:
1823 gen_rot_rm_T1(s1, ot, d, 0);
1824 break;
1825 case OP_ROR:
1826 gen_rot_rm_T1(s1, ot, d, 1);
1827 break;
1828 case OP_SHL:
1829 case OP_SHL1:
1830 gen_shift_rm_T1(s1, ot, d, 0, 0);
1831 break;
1832 case OP_SHR:
1833 gen_shift_rm_T1(s1, ot, d, 1, 0);
1834 break;
1835 case OP_SAR:
1836 gen_shift_rm_T1(s1, ot, d, 1, 1);
1837 break;
1838 case OP_RCL:
1839 gen_rotc_rm_T1(s1, ot, d, 0);
1840 break;
1841 case OP_RCR:
1842 gen_rotc_rm_T1(s1, ot, d, 1);
1843 break;
1844 }
1845 }
1846
1847 static void gen_shifti(DisasContext *s1, int op, TCGMemOp ot, int d, int c)
1848 {
1849 switch(op) {
1850 case OP_ROL:
1851 gen_rot_rm_im(s1, ot, d, c, 0);
1852 break;
1853 case OP_ROR:
1854 gen_rot_rm_im(s1, ot, d, c, 1);
1855 break;
1856 case OP_SHL:
1857 case OP_SHL1:
1858 gen_shift_rm_im(s1, ot, d, c, 0, 0);
1859 break;
1860 case OP_SHR:
1861 gen_shift_rm_im(s1, ot, d, c, 1, 0);
1862 break;
1863 case OP_SAR:
1864 gen_shift_rm_im(s1, ot, d, c, 1, 1);
1865 break;
1866 default:
1867 /* currently not optimized */
1868 tcg_gen_movi_tl(cpu_T[1], c);
1869 gen_shift(s1, op, ot, d, OR_TMP1);
1870 break;
1871 }
1872 }
1873
1874 static void gen_lea_modrm(CPUX86State *env, DisasContext *s, int modrm)
1875 {
1876 target_long disp;
1877 int havesib;
1878 int base;
1879 int index;
1880 int scale;
1881 int mod, rm, code, override, must_add_seg;
1882 TCGv sum;
1883
1884 override = s->override;
1885 must_add_seg = s->addseg;
1886 if (override >= 0)
1887 must_add_seg = 1;
1888 mod = (modrm >> 6) & 3;
1889 rm = modrm & 7;
1890
1891 switch (s->aflag) {
1892 case MO_64:
1893 case MO_32:
1894 havesib = 0;
1895 base = rm;
1896 index = -1;
1897 scale = 0;
1898
1899 if (base == 4) {
1900 havesib = 1;
1901 code = cpu_ldub_code(env, s->pc++);
1902 scale = (code >> 6) & 3;
1903 index = ((code >> 3) & 7) | REX_X(s);
1904 if (index == 4) {
1905 index = -1; /* no index */
1906 }
1907 base = (code & 7);
1908 }
1909 base |= REX_B(s);
1910
1911 switch (mod) {
1912 case 0:
1913 if ((base & 7) == 5) {
1914 base = -1;
1915 disp = (int32_t)cpu_ldl_code(env, s->pc);
1916 s->pc += 4;
1917 if (CODE64(s) && !havesib) {
1918 disp += s->pc + s->rip_offset;
1919 }
1920 } else {
1921 disp = 0;
1922 }
1923 break;
1924 case 1:
1925 disp = (int8_t)cpu_ldub_code(env, s->pc++);
1926 break;
1927 default:
1928 case 2:
1929 disp = (int32_t)cpu_ldl_code(env, s->pc);
1930 s->pc += 4;
1931 break;
1932 }
1933
1934 /* For correct popl handling with esp. */
1935 if (base == R_ESP && s->popl_esp_hack) {
1936 disp += s->popl_esp_hack;
1937 }
1938
1939 /* Compute the address, with a minimum number of TCG ops. */
1940 TCGV_UNUSED(sum);
1941 if (index >= 0) {
1942 if (scale == 0) {
1943 sum = cpu_regs[index];
1944 } else {
1945 tcg_gen_shli_tl(cpu_A0, cpu_regs[index], scale);
1946 sum = cpu_A0;
1947 }
1948 if (base >= 0) {
1949 tcg_gen_add_tl(cpu_A0, sum, cpu_regs[base]);
1950 sum = cpu_A0;
1951 }
1952 } else if (base >= 0) {
1953 sum = cpu_regs[base];
1954 }
1955 if (TCGV_IS_UNUSED(sum)) {
1956 tcg_gen_movi_tl(cpu_A0, disp);
1957 } else {
1958 tcg_gen_addi_tl(cpu_A0, sum, disp);
1959 }
1960
1961 if (must_add_seg) {
1962 if (override < 0) {
1963 if (base == R_EBP || base == R_ESP) {
1964 override = R_SS;
1965 } else {
1966 override = R_DS;
1967 }
1968 }
1969
1970 tcg_gen_ld_tl(cpu_tmp0, cpu_env,
1971 offsetof(CPUX86State, segs[override].base));
1972 if (CODE64(s)) {
1973 if (s->aflag == MO_32) {
1974 tcg_gen_ext32u_tl(cpu_A0, cpu_A0);
1975 }
1976 tcg_gen_add_tl(cpu_A0, cpu_A0, cpu_tmp0);
1977 return;
1978 }
1979
1980 tcg_gen_add_tl(cpu_A0, cpu_A0, cpu_tmp0);
1981 }
1982
1983 if (s->aflag == MO_32) {
1984 tcg_gen_ext32u_tl(cpu_A0, cpu_A0);
1985 }
1986 break;
1987
1988 case MO_16:
1989 switch (mod) {
1990 case 0:
1991 if (rm == 6) {
1992 disp = cpu_lduw_code(env, s->pc);
1993 s->pc += 2;
1994 tcg_gen_movi_tl(cpu_A0, disp);
1995 rm = 0; /* avoid SS override */
1996 goto no_rm;
1997 } else {
1998 disp = 0;
1999 }
2000 break;
2001 case 1:
2002 disp = (int8_t)cpu_ldub_code(env, s->pc++);
2003 break;
2004 default:
2005 case 2:
2006 disp = (int16_t)cpu_lduw_code(env, s->pc);
2007 s->pc += 2;
2008 break;
2009 }
2010
2011 sum = cpu_A0;
2012 switch (rm) {
2013 case 0:
2014 tcg_gen_add_tl(cpu_A0, cpu_regs[R_EBX], cpu_regs[R_ESI]);
2015 break;
2016 case 1:
2017 tcg_gen_add_tl(cpu_A0, cpu_regs[R_EBX], cpu_regs[R_EDI]);
2018 break;
2019 case 2:
2020 tcg_gen_add_tl(cpu_A0, cpu_regs[R_EBP], cpu_regs[R_ESI]);
2021 break;
2022 case 3:
2023 tcg_gen_add_tl(cpu_A0, cpu_regs[R_EBP], cpu_regs[R_EDI]);
2024 break;
2025 case 4:
2026 sum = cpu_regs[R_ESI];
2027 break;
2028 case 5:
2029 sum = cpu_regs[R_EDI];
2030 break;
2031 case 6:
2032 sum = cpu_regs[R_EBP];
2033 break;
2034 default:
2035 case 7:
2036 sum = cpu_regs[R_EBX];
2037 break;
2038 }
2039 tcg_gen_addi_tl(cpu_A0, sum, disp);
2040 tcg_gen_ext16u_tl(cpu_A0, cpu_A0);
2041 no_rm:
2042 if (must_add_seg) {
2043 if (override < 0) {
2044 if (rm == 2 || rm == 3 || rm == 6) {
2045 override = R_SS;
2046 } else {
2047 override = R_DS;
2048 }
2049 }
2050 gen_op_addl_A0_seg(s, override);
2051 }
2052 break;
2053
2054 default:
2055 tcg_abort();
2056 }
2057 }
2058
2059 static void gen_nop_modrm(CPUX86State *env, DisasContext *s, int modrm)
2060 {
2061 int mod, rm, base, code;
2062
2063 mod = (modrm >> 6) & 3;
2064 if (mod == 3)
2065 return;
2066 rm = modrm & 7;
2067
2068 switch (s->aflag) {
2069 case MO_64:
2070 case MO_32:
2071 base = rm;
2072
2073 if (base == 4) {
2074 code = cpu_ldub_code(env, s->pc++);
2075 base = (code & 7);
2076 }
2077
2078 switch (mod) {
2079 case 0:
2080 if (base == 5) {
2081 s->pc += 4;
2082 }
2083 break;
2084 case 1:
2085 s->pc++;
2086 break;
2087 default:
2088 case 2:
2089 s->pc += 4;
2090 break;
2091 }
2092 break;
2093
2094 case MO_16:
2095 switch (mod) {
2096 case 0:
2097 if (rm == 6) {
2098 s->pc += 2;
2099 }
2100 break;
2101 case 1:
2102 s->pc++;
2103 break;
2104 default:
2105 case 2:
2106 s->pc += 2;
2107 break;
2108 }
2109 break;
2110
2111 default:
2112 tcg_abort();
2113 }
2114 }
2115
2116 /* used for LEA and MOV AX, mem */
2117 static void gen_add_A0_ds_seg(DisasContext *s)
2118 {
2119 int override, must_add_seg;
2120 must_add_seg = s->addseg;
2121 override = R_DS;
2122 if (s->override >= 0) {
2123 override = s->override;
2124 must_add_seg = 1;
2125 }
2126 if (must_add_seg) {
2127 #ifdef TARGET_X86_64
2128 if (CODE64(s)) {
2129 gen_op_addq_A0_seg(override);
2130 } else
2131 #endif
2132 {
2133 gen_op_addl_A0_seg(s, override);
2134 }
2135 }
2136 }
2137
2138 /* generate modrm memory load or store of 'reg'. TMP0 is used if reg ==
2139 OR_TMP0 */
2140 static void gen_ldst_modrm(CPUX86State *env, DisasContext *s, int modrm,
2141 TCGMemOp ot, int reg, int is_store)
2142 {
2143 int mod, rm;
2144
2145 mod = (modrm >> 6) & 3;
2146 rm = (modrm & 7) | REX_B(s);
2147 if (mod == 3) {
2148 if (is_store) {
2149 if (reg != OR_TMP0)
2150 gen_op_mov_v_reg(ot, cpu_T[0], reg);
2151 gen_op_mov_reg_v(ot, rm, cpu_T[0]);
2152 } else {
2153 gen_op_mov_v_reg(ot, cpu_T[0], rm);
2154 if (reg != OR_TMP0)
2155 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
2156 }
2157 } else {
2158 gen_lea_modrm(env, s, modrm);
2159 if (is_store) {
2160 if (reg != OR_TMP0)
2161 gen_op_mov_v_reg(ot, cpu_T[0], reg);
2162 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
2163 } else {
2164 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
2165 if (reg != OR_TMP0)
2166 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
2167 }
2168 }
2169 }
2170
2171 static inline uint32_t insn_get(CPUX86State *env, DisasContext *s, TCGMemOp ot)
2172 {
2173 uint32_t ret;
2174
2175 switch (ot) {
2176 case MO_8:
2177 ret = cpu_ldub_code(env, s->pc);
2178 s->pc++;
2179 break;
2180 case MO_16:
2181 ret = cpu_lduw_code(env, s->pc);
2182 s->pc += 2;
2183 break;
2184 case MO_32:
2185 #ifdef TARGET_X86_64
2186 case MO_64:
2187 #endif
2188 ret = cpu_ldl_code(env, s->pc);
2189 s->pc += 4;
2190 break;
2191 default:
2192 tcg_abort();
2193 }
2194 return ret;
2195 }
2196
2197 static inline int insn_const_size(TCGMemOp ot)
2198 {
2199 if (ot <= MO_32) {
2200 return 1 << ot;
2201 } else {
2202 return 4;
2203 }
2204 }
2205
2206 static inline void gen_goto_tb(DisasContext *s, int tb_num, target_ulong eip)
2207 {
2208 TranslationBlock *tb;
2209 target_ulong pc;
2210
2211 pc = s->cs_base + eip;
2212 tb = s->tb;
2213 /* NOTE: we handle the case where the TB spans two pages here */
2214 if ((pc & TARGET_PAGE_MASK) == (tb->pc & TARGET_PAGE_MASK) ||
2215 (pc & TARGET_PAGE_MASK) == ((s->pc - 1) & TARGET_PAGE_MASK)) {
2216 /* jump to same page: we can use a direct jump */
2217 tcg_gen_goto_tb(tb_num);
2218 gen_jmp_im(eip);
2219 tcg_gen_exit_tb((uintptr_t)tb + tb_num);
2220 } else {
2221 /* jump to another page: currently not optimized */
2222 gen_jmp_im(eip);
2223 gen_eob(s);
2224 }
2225 }
2226
2227 static inline void gen_jcc(DisasContext *s, int b,
2228 target_ulong val, target_ulong next_eip)
2229 {
2230 int l1, l2;
2231
2232 if (s->jmp_opt) {
2233 l1 = gen_new_label();
2234 gen_jcc1(s, b, l1);
2235
2236 gen_goto_tb(s, 0, next_eip);
2237
2238 gen_set_label(l1);
2239 gen_goto_tb(s, 1, val);
2240 s->is_jmp = DISAS_TB_JUMP;
2241 } else {
2242 l1 = gen_new_label();
2243 l2 = gen_new_label();
2244 gen_jcc1(s, b, l1);
2245
2246 gen_jmp_im(next_eip);
2247 tcg_gen_br(l2);
2248
2249 gen_set_label(l1);
2250 gen_jmp_im(val);
2251 gen_set_label(l2);
2252 gen_eob(s);
2253 }
2254 }
2255
2256 static void gen_cmovcc1(CPUX86State *env, DisasContext *s, TCGMemOp ot, int b,
2257 int modrm, int reg)
2258 {
2259 CCPrepare cc;
2260
2261 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
2262
2263 cc = gen_prepare_cc(s, b, cpu_T[1]);
2264 if (cc.mask != -1) {
2265 TCGv t0 = tcg_temp_new();
2266 tcg_gen_andi_tl(t0, cc.reg, cc.mask);
2267 cc.reg = t0;
2268 }
2269 if (!cc.use_reg2) {
2270 cc.reg2 = tcg_const_tl(cc.imm);
2271 }
2272
2273 tcg_gen_movcond_tl(cc.cond, cpu_T[0], cc.reg, cc.reg2,
2274 cpu_T[0], cpu_regs[reg]);
2275 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
2276
2277 if (cc.mask != -1) {
2278 tcg_temp_free(cc.reg);
2279 }
2280 if (!cc.use_reg2) {
2281 tcg_temp_free(cc.reg2);
2282 }
2283 }
2284
2285 static inline void gen_op_movl_T0_seg(int seg_reg)
2286 {
2287 tcg_gen_ld32u_tl(cpu_T[0], cpu_env,
2288 offsetof(CPUX86State,segs[seg_reg].selector));
2289 }
2290
2291 static inline void gen_op_movl_seg_T0_vm(int seg_reg)
2292 {
2293 tcg_gen_andi_tl(cpu_T[0], cpu_T[0], 0xffff);
2294 tcg_gen_st32_tl(cpu_T[0], cpu_env,
2295 offsetof(CPUX86State,segs[seg_reg].selector));
2296 tcg_gen_shli_tl(cpu_T[0], cpu_T[0], 4);
2297 tcg_gen_st_tl(cpu_T[0], cpu_env,
2298 offsetof(CPUX86State,segs[seg_reg].base));
2299 }
2300
2301 /* move T0 to seg_reg and compute if the CPU state may change. Never
2302 call this function with seg_reg == R_CS */
2303 static void gen_movl_seg_T0(DisasContext *s, int seg_reg, target_ulong cur_eip)
2304 {
2305 if (s->pe && !s->vm86) {
2306 /* XXX: optimize by finding processor state dynamically */
2307 gen_update_cc_op(s);
2308 gen_jmp_im(cur_eip);
2309 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
2310 gen_helper_load_seg(cpu_env, tcg_const_i32(seg_reg), cpu_tmp2_i32);
2311 /* abort translation because the addseg value may change or
2312 because ss32 may change. For R_SS, translation must always
2313 stop as a special handling must be done to disable hardware
2314 interrupts for the next instruction */
2315 if (seg_reg == R_SS || (s->code32 && seg_reg < R_FS))
2316 s->is_jmp = DISAS_TB_JUMP;
2317 } else {
2318 gen_op_movl_seg_T0_vm(seg_reg);
2319 if (seg_reg == R_SS)
2320 s->is_jmp = DISAS_TB_JUMP;
2321 }
2322 }
2323
2324 static inline int svm_is_rep(int prefixes)
2325 {
2326 return ((prefixes & (PREFIX_REPZ | PREFIX_REPNZ)) ? 8 : 0);
2327 }
2328
2329 static inline void
2330 gen_svm_check_intercept_param(DisasContext *s, target_ulong pc_start,
2331 uint32_t type, uint64_t param)
2332 {
2333 /* no SVM activated; fast case */
2334 if (likely(!(s->flags & HF_SVMI_MASK)))
2335 return;
2336 gen_update_cc_op(s);
2337 gen_jmp_im(pc_start - s->cs_base);
2338 gen_helper_svm_check_intercept_param(cpu_env, tcg_const_i32(type),
2339 tcg_const_i64(param));
2340 }
2341
2342 static inline void
2343 gen_svm_check_intercept(DisasContext *s, target_ulong pc_start, uint64_t type)
2344 {
2345 gen_svm_check_intercept_param(s, pc_start, type, 0);
2346 }
2347
2348 static inline void gen_stack_update(DisasContext *s, int addend)
2349 {
2350 #ifdef TARGET_X86_64
2351 if (CODE64(s)) {
2352 gen_op_add_reg_im(MO_64, R_ESP, addend);
2353 } else
2354 #endif
2355 if (s->ss32) {
2356 gen_op_add_reg_im(MO_32, R_ESP, addend);
2357 } else {
2358 gen_op_add_reg_im(MO_16, R_ESP, addend);
2359 }
2360 }
2361
2362 /* Generate a push. It depends on ss32, addseg and dflag. */
2363 static void gen_push_v(DisasContext *s, TCGv val)
2364 {
2365 TCGMemOp a_ot, d_ot = mo_pushpop(s, s->dflag);
2366 int size = 1 << d_ot;
2367 TCGv new_esp = cpu_A0;
2368
2369 tcg_gen_subi_tl(cpu_A0, cpu_regs[R_ESP], size);
2370
2371 if (CODE64(s)) {
2372 a_ot = MO_64;
2373 } else if (s->ss32) {
2374 a_ot = MO_32;
2375 if (s->addseg) {
2376 new_esp = cpu_tmp4;
2377 tcg_gen_mov_tl(new_esp, cpu_A0);
2378 gen_op_addl_A0_seg(s, R_SS);
2379 } else {
2380 tcg_gen_ext32u_tl(cpu_A0, cpu_A0);
2381 }
2382 } else {
2383 a_ot = MO_16;
2384 new_esp = cpu_tmp4;
2385 tcg_gen_ext16u_tl(cpu_A0, cpu_A0);
2386 tcg_gen_mov_tl(new_esp, cpu_A0);
2387 gen_op_addl_A0_seg(s, R_SS);
2388 }
2389
2390 gen_op_st_v(s, d_ot, val, cpu_A0);
2391 gen_op_mov_reg_v(a_ot, R_ESP, new_esp);
2392 }
2393
2394 /* two step pop is necessary for precise exceptions */
2395 static TCGMemOp gen_pop_T0(DisasContext *s)
2396 {
2397 TCGMemOp d_ot = mo_pushpop(s, s->dflag);
2398 TCGv addr = cpu_A0;
2399
2400 if (CODE64(s)) {
2401 addr = cpu_regs[R_ESP];
2402 } else if (!s->ss32) {
2403 tcg_gen_ext16u_tl(cpu_A0, cpu_regs[R_ESP]);
2404 gen_op_addl_A0_seg(s, R_SS);
2405 } else if (s->addseg) {
2406 tcg_gen_mov_tl(cpu_A0, cpu_regs[R_ESP]);
2407 gen_op_addl_A0_seg(s, R_SS);
2408 } else {
2409 tcg_gen_ext32u_tl(cpu_A0, cpu_regs[R_ESP]);
2410 }
2411
2412 gen_op_ld_v(s, d_ot, cpu_T[0], addr);
2413 return d_ot;
2414 }
2415
2416 static void gen_pop_update(DisasContext *s, TCGMemOp ot)
2417 {
2418 gen_stack_update(s, 1 << ot);
2419 }
2420
2421 static void gen_stack_A0(DisasContext *s)
2422 {
2423 gen_op_movl_A0_reg(R_ESP);
2424 if (!s->ss32)
2425 tcg_gen_ext16u_tl(cpu_A0, cpu_A0);
2426 tcg_gen_mov_tl(cpu_T[1], cpu_A0);
2427 if (s->addseg)
2428 gen_op_addl_A0_seg(s, R_SS);
2429 }
2430
2431 /* NOTE: wrap around in 16 bit not fully handled */
2432 static void gen_pusha(DisasContext *s)
2433 {
2434 int i;
2435 gen_op_movl_A0_reg(R_ESP);
2436 gen_op_addl_A0_im(-8 << s->dflag);
2437 if (!s->ss32)
2438 tcg_gen_ext16u_tl(cpu_A0, cpu_A0);
2439 tcg_gen_mov_tl(cpu_T[1], cpu_A0);
2440 if (s->addseg)
2441 gen_op_addl_A0_seg(s, R_SS);
2442 for(i = 0;i < 8; i++) {
2443 gen_op_mov_v_reg(MO_32, cpu_T[0], 7 - i);
2444 gen_op_st_v(s, s->dflag, cpu_T[0], cpu_A0);
2445 gen_op_addl_A0_im(1 << s->dflag);
2446 }
2447 gen_op_mov_reg_v(MO_16 + s->ss32, R_ESP, cpu_T[1]);
2448 }
2449
2450 /* NOTE: wrap around in 16 bit not fully handled */
2451 static void gen_popa(DisasContext *s)
2452 {
2453 int i;
2454 gen_op_movl_A0_reg(R_ESP);
2455 if (!s->ss32)
2456 tcg_gen_ext16u_tl(cpu_A0, cpu_A0);
2457 tcg_gen_mov_tl(cpu_T[1], cpu_A0);
2458 tcg_gen_addi_tl(cpu_T[1], cpu_T[1], 8 << s->dflag);
2459 if (s->addseg)
2460 gen_op_addl_A0_seg(s, R_SS);
2461 for(i = 0;i < 8; i++) {
2462 /* ESP is not reloaded */
2463 if (i != 3) {
2464 gen_op_ld_v(s, s->dflag, cpu_T[0], cpu_A0);
2465 gen_op_mov_reg_v(s->dflag, 7 - i, cpu_T[0]);
2466 }
2467 gen_op_addl_A0_im(1 << s->dflag);
2468 }
2469 gen_op_mov_reg_v(MO_16 + s->ss32, R_ESP, cpu_T[1]);
2470 }
2471
2472 static void gen_enter(DisasContext *s, int esp_addend, int level)
2473 {
2474 TCGMemOp ot = mo_pushpop(s, s->dflag);
2475 int opsize = 1 << ot;
2476
2477 level &= 0x1f;
2478 #ifdef TARGET_X86_64
2479 if (CODE64(s)) {
2480 gen_op_movl_A0_reg(R_ESP);
2481 gen_op_addq_A0_im(-opsize);
2482 tcg_gen_mov_tl(cpu_T[1], cpu_A0);
2483
2484 /* push bp */
2485 gen_op_mov_v_reg(MO_32, cpu_T[0], R_EBP);
2486 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
2487 if (level) {
2488 /* XXX: must save state */
2489 gen_helper_enter64_level(cpu_env, tcg_const_i32(level),
2490 tcg_const_i32((ot == MO_64)),
2491 cpu_T[1]);
2492 }
2493 gen_op_mov_reg_v(ot, R_EBP, cpu_T[1]);
2494 tcg_gen_addi_tl(cpu_T[1], cpu_T[1], -esp_addend + (-opsize * level));
2495 gen_op_mov_reg_v(MO_64, R_ESP, cpu_T[1]);
2496 } else
2497 #endif
2498 {
2499 gen_op_movl_A0_reg(R_ESP);
2500 gen_op_addl_A0_im(-opsize);
2501 if (!s->ss32)
2502 tcg_gen_ext16u_tl(cpu_A0, cpu_A0);
2503 tcg_gen_mov_tl(cpu_T[1], cpu_A0);
2504 if (s->addseg)
2505 gen_op_addl_A0_seg(s, R_SS);
2506 /* push bp */
2507 gen_op_mov_v_reg(MO_32, cpu_T[0], R_EBP);
2508 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
2509 if (level) {
2510 /* XXX: must save state */
2511 gen_helper_enter_level(cpu_env, tcg_const_i32(level),
2512 tcg_const_i32(s->dflag - 1),
2513 cpu_T[1]);
2514 }
2515 gen_op_mov_reg_v(ot, R_EBP, cpu_T[1]);
2516 tcg_gen_addi_tl(cpu_T[1], cpu_T[1], -esp_addend + (-opsize * level));
2517 gen_op_mov_reg_v(MO_16 + s->ss32, R_ESP, cpu_T[1]);
2518 }
2519 }
2520
2521 static void gen_exception(DisasContext *s, int trapno, target_ulong cur_eip)
2522 {
2523 gen_update_cc_op(s);
2524 gen_jmp_im(cur_eip);
2525 gen_helper_raise_exception(cpu_env, tcg_const_i32(trapno));
2526 s->is_jmp = DISAS_TB_JUMP;
2527 }
2528
2529 /* an interrupt is different from an exception because of the
2530 privilege checks */
2531 static void gen_interrupt(DisasContext *s, int intno,
2532 target_ulong cur_eip, target_ulong next_eip)
2533 {
2534 gen_update_cc_op(s);
2535 gen_jmp_im(cur_eip);
2536 gen_helper_raise_interrupt(cpu_env, tcg_const_i32(intno),
2537 tcg_const_i32(next_eip - cur_eip));
2538 s->is_jmp = DISAS_TB_JUMP;
2539 }
2540
2541 static void gen_debug(DisasContext *s, target_ulong cur_eip)
2542 {
2543 gen_update_cc_op(s);
2544 gen_jmp_im(cur_eip);
2545 gen_helper_debug(cpu_env);
2546 s->is_jmp = DISAS_TB_JUMP;
2547 }
2548
2549 /* generate a generic end of block. Trace exception is also generated
2550 if needed */
2551 static void gen_eob(DisasContext *s)
2552 {
2553 gen_update_cc_op(s);
2554 if (s->tb->flags & HF_INHIBIT_IRQ_MASK) {
2555 gen_helper_reset_inhibit_irq(cpu_env);
2556 }
2557 if (s->tb->flags & HF_RF_MASK) {
2558 gen_helper_reset_rf(cpu_env);
2559 }
2560 if (s->singlestep_enabled) {
2561 gen_helper_debug(cpu_env);
2562 } else if (s->tf) {
2563 gen_helper_single_step(cpu_env);
2564 } else {
2565 tcg_gen_exit_tb(0);
2566 }
2567 s->is_jmp = DISAS_TB_JUMP;
2568 }
2569
2570 /* generate a jump to eip. No segment change must happen before as a
2571 direct call to the next block may occur */
2572 static void gen_jmp_tb(DisasContext *s, target_ulong eip, int tb_num)
2573 {
2574 gen_update_cc_op(s);
2575 set_cc_op(s, CC_OP_DYNAMIC);
2576 if (s->jmp_opt) {
2577 gen_goto_tb(s, tb_num, eip);
2578 s->is_jmp = DISAS_TB_JUMP;
2579 } else {
2580 gen_jmp_im(eip);
2581 gen_eob(s);
2582 }
2583 }
2584
2585 static void gen_jmp(DisasContext *s, target_ulong eip)
2586 {
2587 gen_jmp_tb(s, eip, 0);
2588 }
2589
2590 static inline void gen_ldq_env_A0(DisasContext *s, int offset)
2591 {
2592 tcg_gen_qemu_ld_i64(cpu_tmp1_i64, cpu_A0, s->mem_index, MO_LEQ);
2593 tcg_gen_st_i64(cpu_tmp1_i64, cpu_env, offset);
2594 }
2595
2596 static inline void gen_stq_env_A0(DisasContext *s, int offset)
2597 {
2598 tcg_gen_ld_i64(cpu_tmp1_i64, cpu_env, offset);
2599 tcg_gen_qemu_st_i64(cpu_tmp1_i64, cpu_A0, s->mem_index, MO_LEQ);
2600 }
2601
2602 static inline void gen_ldo_env_A0(DisasContext *s, int offset)
2603 {
2604 int mem_index = s->mem_index;
2605 tcg_gen_qemu_ld_i64(cpu_tmp1_i64, cpu_A0, mem_index, MO_LEQ);
2606 tcg_gen_st_i64(cpu_tmp1_i64, cpu_env, offset + offsetof(XMMReg, XMM_Q(0)));
2607 tcg_gen_addi_tl(cpu_tmp0, cpu_A0, 8);
2608 tcg_gen_qemu_ld_i64(cpu_tmp1_i64, cpu_tmp0, mem_index, MO_LEQ);
2609 tcg_gen_st_i64(cpu_tmp1_i64, cpu_env, offset + offsetof(XMMReg, XMM_Q(1)));
2610 }
2611
2612 static inline void gen_sto_env_A0(DisasContext *s, int offset)
2613 {
2614 int mem_index = s->mem_index;
2615 tcg_gen_ld_i64(cpu_tmp1_i64, cpu_env, offset + offsetof(XMMReg, XMM_Q(0)));
2616 tcg_gen_qemu_st_i64(cpu_tmp1_i64, cpu_A0, mem_index, MO_LEQ);
2617 tcg_gen_addi_tl(cpu_tmp0, cpu_A0, 8);
2618 tcg_gen_ld_i64(cpu_tmp1_i64, cpu_env, offset + offsetof(XMMReg, XMM_Q(1)));
2619 tcg_gen_qemu_st_i64(cpu_tmp1_i64, cpu_tmp0, mem_index, MO_LEQ);
2620 }
2621
2622 static inline void gen_op_movo(int d_offset, int s_offset)
2623 {
2624 tcg_gen_ld_i64(cpu_tmp1_i64, cpu_env, s_offset);
2625 tcg_gen_st_i64(cpu_tmp1_i64, cpu_env, d_offset);
2626 tcg_gen_ld_i64(cpu_tmp1_i64, cpu_env, s_offset + 8);
2627 tcg_gen_st_i64(cpu_tmp1_i64, cpu_env, d_offset + 8);
2628 }
2629
2630 static inline void gen_op_movq(int d_offset, int s_offset)
2631 {
2632 tcg_gen_ld_i64(cpu_tmp1_i64, cpu_env, s_offset);
2633 tcg_gen_st_i64(cpu_tmp1_i64, cpu_env, d_offset);
2634 }
2635
2636 static inline void gen_op_movl(int d_offset, int s_offset)
2637 {
2638 tcg_gen_ld_i32(cpu_tmp2_i32, cpu_env, s_offset);
2639 tcg_gen_st_i32(cpu_tmp2_i32, cpu_env, d_offset);
2640 }
2641
2642 static inline void gen_op_movq_env_0(int d_offset)
2643 {
2644 tcg_gen_movi_i64(cpu_tmp1_i64, 0);
2645 tcg_gen_st_i64(cpu_tmp1_i64, cpu_env, d_offset);
2646 }
2647
2648 typedef void (*SSEFunc_i_ep)(TCGv_i32 val, TCGv_ptr env, TCGv_ptr reg);
2649 typedef void (*SSEFunc_l_ep)(TCGv_i64 val, TCGv_ptr env, TCGv_ptr reg);
2650 typedef void (*SSEFunc_0_epi)(TCGv_ptr env, TCGv_ptr reg, TCGv_i32 val);
2651 typedef void (*SSEFunc_0_epl)(TCGv_ptr env, TCGv_ptr reg, TCGv_i64 val);
2652 typedef void (*SSEFunc_0_epp)(TCGv_ptr env, TCGv_ptr reg_a, TCGv_ptr reg_b);
2653 typedef void (*SSEFunc_0_eppi)(TCGv_ptr env, TCGv_ptr reg_a, TCGv_ptr reg_b,
2654 TCGv_i32 val);
2655 typedef void (*SSEFunc_0_ppi)(TCGv_ptr reg_a, TCGv_ptr reg_b, TCGv_i32 val);
2656 typedef void (*SSEFunc_0_eppt)(TCGv_ptr env, TCGv_ptr reg_a, TCGv_ptr reg_b,
2657 TCGv val);
2658
2659 #define SSE_SPECIAL ((void *)1)
2660 #define SSE_DUMMY ((void *)2)
2661
2662 #define MMX_OP2(x) { gen_helper_ ## x ## _mmx, gen_helper_ ## x ## _xmm }
2663 #define SSE_FOP(x) { gen_helper_ ## x ## ps, gen_helper_ ## x ## pd, \
2664 gen_helper_ ## x ## ss, gen_helper_ ## x ## sd, }
2665
2666 static const SSEFunc_0_epp sse_op_table1[256][4] = {
2667 /* 3DNow! extensions */
2668 [0x0e] = { SSE_DUMMY }, /* femms */
2669 [0x0f] = { SSE_DUMMY }, /* pf... */
2670 /* pure SSE operations */
2671 [0x10] = { SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL }, /* movups, movupd, movss, movsd */
2672 [0x11] = { SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL }, /* movups, movupd, movss, movsd */
2673 [0x12] = { SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL }, /* movlps, movlpd, movsldup, movddup */
2674 [0x13] = { SSE_SPECIAL, SSE_SPECIAL }, /* movlps, movlpd */
2675 [0x14] = { gen_helper_punpckldq_xmm, gen_helper_punpcklqdq_xmm },
2676 [0x15] = { gen_helper_punpckhdq_xmm, gen_helper_punpckhqdq_xmm },
2677 [0x16] = { SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL }, /* movhps, movhpd, movshdup */
2678 [0x17] = { SSE_SPECIAL, SSE_SPECIAL }, /* movhps, movhpd */
2679
2680 [0x28] = { SSE_SPECIAL, SSE_SPECIAL }, /* movaps, movapd */
2681 [0x29] = { SSE_SPECIAL, SSE_SPECIAL }, /* movaps, movapd */
2682 [0x2a] = { SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL }, /* cvtpi2ps, cvtpi2pd, cvtsi2ss, cvtsi2sd */
2683 [0x2b] = { SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL }, /* movntps, movntpd, movntss, movntsd */
2684 [0x2c] = { SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL }, /* cvttps2pi, cvttpd2pi, cvttsd2si, cvttss2si */
2685 [0x2d] = { SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL }, /* cvtps2pi, cvtpd2pi, cvtsd2si, cvtss2si */
2686 [0x2e] = { gen_helper_ucomiss, gen_helper_ucomisd },
2687 [0x2f] = { gen_helper_comiss, gen_helper_comisd },
2688 [0x50] = { SSE_SPECIAL, SSE_SPECIAL }, /* movmskps, movmskpd */
2689 [0x51] = SSE_FOP(sqrt),
2690 [0x52] = { gen_helper_rsqrtps, NULL, gen_helper_rsqrtss, NULL },
2691 [0x53] = { gen_helper_rcpps, NULL, gen_helper_rcpss, NULL },
2692 [0x54] = { gen_helper_pand_xmm, gen_helper_pand_xmm }, /* andps, andpd */
2693 [0x55] = { gen_helper_pandn_xmm, gen_helper_pandn_xmm }, /* andnps, andnpd */
2694 [0x56] = { gen_helper_por_xmm, gen_helper_por_xmm }, /* orps, orpd */
2695 [0x57] = { gen_helper_pxor_xmm, gen_helper_pxor_xmm }, /* xorps, xorpd */
2696 [0x58] = SSE_FOP(add),
2697 [0x59] = SSE_FOP(mul),
2698 [0x5a] = { gen_helper_cvtps2pd, gen_helper_cvtpd2ps,
2699 gen_helper_cvtss2sd, gen_helper_cvtsd2ss },
2700 [0x5b] = { gen_helper_cvtdq2ps, gen_helper_cvtps2dq, gen_helper_cvttps2dq },
2701 [0x5c] = SSE_FOP(sub),
2702 [0x5d] = SSE_FOP(min),
2703 [0x5e] = SSE_FOP(div),
2704 [0x5f] = SSE_FOP(max),
2705
2706 [0xc2] = SSE_FOP(cmpeq),
2707 [0xc6] = { (SSEFunc_0_epp)gen_helper_shufps,
2708 (SSEFunc_0_epp)gen_helper_shufpd }, /* XXX: casts */
2709
2710 /* SSSE3, SSE4, MOVBE, CRC32, BMI1, BMI2, ADX. */
2711 [0x38] = { SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL },
2712 [0x3a] = { SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL },
2713
2714 /* MMX ops and their SSE extensions */
2715 [0x60] = MMX_OP2(punpcklbw),
2716 [0x61] = MMX_OP2(punpcklwd),
2717 [0x62] = MMX_OP2(punpckldq),
2718 [0x63] = MMX_OP2(packsswb),
2719 [0x64] = MMX_OP2(pcmpgtb),
2720 [0x65] = MMX_OP2(pcmpgtw),
2721 [0x66] = MMX_OP2(pcmpgtl),
2722 [0x67] = MMX_OP2(packuswb),
2723 [0x68] = MMX_OP2(punpckhbw),
2724 [0x69] = MMX_OP2(punpckhwd),
2725 [0x6a] = MMX_OP2(punpckhdq),
2726 [0x6b] = MMX_OP2(packssdw),
2727 [0x6c] = { NULL, gen_helper_punpcklqdq_xmm },
2728 [0x6d] = { NULL, gen_helper_punpckhqdq_xmm },
2729 [0x6e] = { SSE_SPECIAL, SSE_SPECIAL }, /* movd mm, ea */
2730 [0x6f] = { SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL }, /* movq, movdqa, , movqdu */
2731 [0x70] = { (SSEFunc_0_epp)gen_helper_pshufw_mmx,
2732 (SSEFunc_0_epp)gen_helper_pshufd_xmm,
2733 (SSEFunc_0_epp)gen_helper_pshufhw_xmm,
2734 (SSEFunc_0_epp)gen_helper_pshuflw_xmm }, /* XXX: casts */
2735 [0x71] = { SSE_SPECIAL, SSE_SPECIAL }, /* shiftw */
2736 [0x72] = { SSE_SPECIAL, SSE_SPECIAL }, /* shiftd */
2737 [0x73] = { SSE_SPECIAL, SSE_SPECIAL }, /* shiftq */
2738 [0x74] = MMX_OP2(pcmpeqb),
2739 [0x75] = MMX_OP2(pcmpeqw),
2740 [0x76] = MMX_OP2(pcmpeql),
2741 [0x77] = { SSE_DUMMY }, /* emms */
2742 [0x78] = { NULL, SSE_SPECIAL, NULL, SSE_SPECIAL }, /* extrq_i, insertq_i */
2743 [0x79] = { NULL, gen_helper_extrq_r, NULL, gen_helper_insertq_r },
2744 [0x7c] = { NULL, gen_helper_haddpd, NULL, gen_helper_haddps },
2745 [0x7d] = { NULL, gen_helper_hsubpd, NULL, gen_helper_hsubps },
2746 [0x7e] = { SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL }, /* movd, movd, , movq */
2747 [0x7f] = { SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL }, /* movq, movdqa, movdqu */
2748 [0xc4] = { SSE_SPECIAL, SSE_SPECIAL }, /* pinsrw */
2749 [0xc5] = { SSE_SPECIAL, SSE_SPECIAL }, /* pextrw */
2750 [0xd0] = { NULL, gen_helper_addsubpd, NULL, gen_helper_addsubps },
2751 [0xd1] = MMX_OP2(psrlw),
2752 [0xd2] = MMX_OP2(psrld),
2753 [0xd3] = MMX_OP2(psrlq),
2754 [0xd4] = MMX_OP2(paddq),
2755 [0xd5] = MMX_OP2(pmullw),
2756 [0xd6] = { NULL, SSE_SPECIAL, SSE_SPECIAL, SSE_SPECIAL },
2757 [0xd7] = { SSE_SPECIAL, SSE_SPECIAL }, /* pmovmskb */
2758 [0xd8] = MMX_OP2(psubusb),
2759 [0xd9] = MMX_OP2(psubusw),
2760 [0xda] = MMX_OP2(pminub),
2761 [0xdb] = MMX_OP2(pand),
2762 [0xdc] = MMX_OP2(paddusb),
2763 [0xdd] = MMX_OP2(paddusw),
2764 [0xde] = MMX_OP2(pmaxub),
2765 [0xdf] = MMX_OP2(pandn),
2766 [0xe0] = MMX_OP2(pavgb),
2767 [0xe1] = MMX_OP2(psraw),
2768 [0xe2] = MMX_OP2(psrad),
2769 [0xe3] = MMX_OP2(pavgw),
2770 [0xe4] = MMX_OP2(pmulhuw),
2771 [0xe5] = MMX_OP2(pmulhw),
2772 [0xe6] = { NULL, gen_helper_cvttpd2dq, gen_helper_cvtdq2pd, gen_helper_cvtpd2dq },
2773 [0xe7] = { SSE_SPECIAL , SSE_SPECIAL }, /* movntq, movntq */
2774 [0xe8] = MMX_OP2(psubsb),
2775 [0xe9] = MMX_OP2(psubsw),
2776 [0xea] = MMX_OP2(pminsw),
2777 [0xeb] = MMX_OP2(por),
2778 [0xec] = MMX_OP2(paddsb),
2779 [0xed] = MMX_OP2(paddsw),
2780 [0xee] = MMX_OP2(pmaxsw),
2781 [0xef] = MMX_OP2(pxor),
2782 [0xf0] = { NULL, NULL, NULL, SSE_SPECIAL }, /* lddqu */
2783 [0xf1] = MMX_OP2(psllw),
2784 [0xf2] = MMX_OP2(pslld),
2785 [0xf3] = MMX_OP2(psllq),
2786 [0xf4] = MMX_OP2(pmuludq),
2787 [0xf5] = MMX_OP2(pmaddwd),
2788 [0xf6] = MMX_OP2(psadbw),
2789 [0xf7] = { (SSEFunc_0_epp)gen_helper_maskmov_mmx,
2790 (SSEFunc_0_epp)gen_helper_maskmov_xmm }, /* XXX: casts */
2791 [0xf8] = MMX_OP2(psubb),
2792 [0xf9] = MMX_OP2(psubw),
2793 [0xfa] = MMX_OP2(psubl),
2794 [0xfb] = MMX_OP2(psubq),
2795 [0xfc] = MMX_OP2(paddb),
2796 [0xfd] = MMX_OP2(paddw),
2797 [0xfe] = MMX_OP2(paddl),
2798 };
2799
2800 static const SSEFunc_0_epp sse_op_table2[3 * 8][2] = {
2801 [0 + 2] = MMX_OP2(psrlw),
2802 [0 + 4] = MMX_OP2(psraw),
2803 [0 + 6] = MMX_OP2(psllw),
2804 [8 + 2] = MMX_OP2(psrld),
2805 [8 + 4] = MMX_OP2(psrad),
2806 [8 + 6] = MMX_OP2(pslld),
2807 [16 + 2] = MMX_OP2(psrlq),
2808 [16 + 3] = { NULL, gen_helper_psrldq_xmm },
2809 [16 + 6] = MMX_OP2(psllq),
2810 [16 + 7] = { NULL, gen_helper_pslldq_xmm },
2811 };
2812
2813 static const SSEFunc_0_epi sse_op_table3ai[] = {
2814 gen_helper_cvtsi2ss,
2815 gen_helper_cvtsi2sd
2816 };
2817
2818 #ifdef TARGET_X86_64
2819 static const SSEFunc_0_epl sse_op_table3aq[] = {
2820 gen_helper_cvtsq2ss,
2821 gen_helper_cvtsq2sd
2822 };
2823 #endif
2824
2825 static const SSEFunc_i_ep sse_op_table3bi[] = {
2826 gen_helper_cvttss2si,
2827 gen_helper_cvtss2si,
2828 gen_helper_cvttsd2si,
2829 gen_helper_cvtsd2si
2830 };
2831
2832 #ifdef TARGET_X86_64
2833 static const SSEFunc_l_ep sse_op_table3bq[] = {
2834 gen_helper_cvttss2sq,
2835 gen_helper_cvtss2sq,
2836 gen_helper_cvttsd2sq,
2837 gen_helper_cvtsd2sq
2838 };
2839 #endif
2840
2841 static const SSEFunc_0_epp sse_op_table4[8][4] = {
2842 SSE_FOP(cmpeq),
2843 SSE_FOP(cmplt),
2844 SSE_FOP(cmple),
2845 SSE_FOP(cmpunord),
2846 SSE_FOP(cmpneq),
2847 SSE_FOP(cmpnlt),
2848 SSE_FOP(cmpnle),
2849 SSE_FOP(cmpord),
2850 };
2851
2852 static const SSEFunc_0_epp sse_op_table5[256] = {
2853 [0x0c] = gen_helper_pi2fw,
2854 [0x0d] = gen_helper_pi2fd,
2855 [0x1c] = gen_helper_pf2iw,
2856 [0x1d] = gen_helper_pf2id,
2857 [0x8a] = gen_helper_pfnacc,
2858 [0x8e] = gen_helper_pfpnacc,
2859 [0x90] = gen_helper_pfcmpge,
2860 [0x94] = gen_helper_pfmin,
2861 [0x96] = gen_helper_pfrcp,
2862 [0x97] = gen_helper_pfrsqrt,
2863 [0x9a] = gen_helper_pfsub,
2864 [0x9e] = gen_helper_pfadd,
2865 [0xa0] = gen_helper_pfcmpgt,
2866 [0xa4] = gen_helper_pfmax,
2867 [0xa6] = gen_helper_movq, /* pfrcpit1; no need to actually increase precision */
2868 [0xa7] = gen_helper_movq, /* pfrsqit1 */
2869 [0xaa] = gen_helper_pfsubr,
2870 [0xae] = gen_helper_pfacc,
2871 [0xb0] = gen_helper_pfcmpeq,
2872 [0xb4] = gen_helper_pfmul,
2873 [0xb6] = gen_helper_movq, /* pfrcpit2 */
2874 [0xb7] = gen_helper_pmulhrw_mmx,
2875 [0xbb] = gen_helper_pswapd,
2876 [0xbf] = gen_helper_pavgb_mmx /* pavgusb */
2877 };
2878
2879 struct SSEOpHelper_epp {
2880 SSEFunc_0_epp op[2];
2881 uint32_t ext_mask;
2882 };
2883
2884 struct SSEOpHelper_eppi {
2885 SSEFunc_0_eppi op[2];
2886 uint32_t ext_mask;
2887 };
2888
2889 #define SSSE3_OP(x) { MMX_OP2(x), CPUID_EXT_SSSE3 }
2890 #define SSE41_OP(x) { { NULL, gen_helper_ ## x ## _xmm }, CPUID_EXT_SSE41 }
2891 #define SSE42_OP(x) { { NULL, gen_helper_ ## x ## _xmm }, CPUID_EXT_SSE42 }
2892 #define SSE41_SPECIAL { { NULL, SSE_SPECIAL }, CPUID_EXT_SSE41 }
2893 #define PCLMULQDQ_OP(x) { { NULL, gen_helper_ ## x ## _xmm }, \
2894 CPUID_EXT_PCLMULQDQ }
2895 #define AESNI_OP(x) { { NULL, gen_helper_ ## x ## _xmm }, CPUID_EXT_AES }
2896
2897 static const struct SSEOpHelper_epp sse_op_table6[256] = {
2898 [0x00] = SSSE3_OP(pshufb),
2899 [0x01] = SSSE3_OP(phaddw),
2900 [0x02] = SSSE3_OP(phaddd),
2901 [0x03] = SSSE3_OP(phaddsw),
2902 [0x04] = SSSE3_OP(pmaddubsw),
2903 [0x05] = SSSE3_OP(phsubw),
2904 [0x06] = SSSE3_OP(phsubd),
2905 [0x07] = SSSE3_OP(phsubsw),
2906 [0x08] = SSSE3_OP(psignb),
2907 [0x09] = SSSE3_OP(psignw),
2908 [0x0a] = SSSE3_OP(psignd),
2909 [0x0b] = SSSE3_OP(pmulhrsw),
2910 [0x10] = SSE41_OP(pblendvb),
2911 [0x14] = SSE41_OP(blendvps),
2912 [0x15] = SSE41_OP(blendvpd),
2913 [0x17] = SSE41_OP(ptest),
2914 [0x1c] = SSSE3_OP(pabsb),
2915 [0x1d] = SSSE3_OP(pabsw),
2916 [0x1e] = SSSE3_OP(pabsd),
2917 [0x20] = SSE41_OP(pmovsxbw),
2918 [0x21] = SSE41_OP(pmovsxbd),
2919 [0x22] = SSE41_OP(pmovsxbq),
2920 [0x23] = SSE41_OP(pmovsxwd),
2921 [0x24] = SSE41_OP(pmovsxwq),
2922 [0x25] = SSE41_OP(pmovsxdq),
2923 [0x28] = SSE41_OP(pmuldq),
2924 [0x29] = SSE41_OP(pcmpeqq),
2925 [0x2a] = SSE41_SPECIAL, /* movntqda */
2926 [0x2b] = SSE41_OP(packusdw),
2927 [0x30] = SSE41_OP(pmovzxbw),
2928 [0x31] = SSE41_OP(pmovzxbd),
2929 [0x32] = SSE41_OP(pmovzxbq),
2930 [0x33] = SSE41_OP(pmovzxwd),
2931 [0x34] = SSE41_OP(pmovzxwq),
2932 [0x35] = SSE41_OP(pmovzxdq),
2933 [0x37] = SSE42_OP(pcmpgtq),
2934 [0x38] = SSE41_OP(pminsb),
2935 [0x39] = SSE41_OP(pminsd),
2936 [0x3a] = SSE41_OP(pminuw),
2937 [0x3b] = SSE41_OP(pminud),
2938 [0x3c] = SSE41_OP(pmaxsb),
2939 [0x3d] = SSE41_OP(pmaxsd),
2940 [0x3e] = SSE41_OP(pmaxuw),
2941 [0x3f] = SSE41_OP(pmaxud),
2942 [0x40] = SSE41_OP(pmulld),
2943 [0x41] = SSE41_OP(phminposuw),
2944 [0xdb] = AESNI_OP(aesimc),
2945 [0xdc] = AESNI_OP(aesenc),
2946 [0xdd] = AESNI_OP(aesenclast),
2947 [0xde] = AESNI_OP(aesdec),
2948 [0xdf] = AESNI_OP(aesdeclast),
2949 };
2950
2951 static const struct SSEOpHelper_eppi sse_op_table7[256] = {
2952 [0x08] = SSE41_OP(roundps),
2953 [0x09] = SSE41_OP(roundpd),
2954 [0x0a] = SSE41_OP(roundss),
2955 [0x0b] = SSE41_OP(roundsd),
2956 [0x0c] = SSE41_OP(blendps),
2957 [0x0d] = SSE41_OP(blendpd),
2958 [0x0e] = SSE41_OP(pblendw),
2959 [0x0f] = SSSE3_OP(palignr),
2960 [0x14] = SSE41_SPECIAL, /* pextrb */
2961 [0x15] = SSE41_SPECIAL, /* pextrw */
2962 [0x16] = SSE41_SPECIAL, /* pextrd/pextrq */
2963 [0x17] = SSE41_SPECIAL, /* extractps */
2964 [0x20] = SSE41_SPECIAL, /* pinsrb */
2965 [0x21] = SSE41_SPECIAL, /* insertps */
2966 [0x22] = SSE41_SPECIAL, /* pinsrd/pinsrq */
2967 [0x40] = SSE41_OP(dpps),
2968 [0x41] = SSE41_OP(dppd),
2969 [0x42] = SSE41_OP(mpsadbw),
2970 [0x44] = PCLMULQDQ_OP(pclmulqdq),
2971 [0x60] = SSE42_OP(pcmpestrm),
2972 [0x61] = SSE42_OP(pcmpestri),
2973 [0x62] = SSE42_OP(pcmpistrm),
2974 [0x63] = SSE42_OP(pcmpistri),
2975 [0xdf] = AESNI_OP(aeskeygenassist),
2976 };
2977
2978 static void gen_sse(CPUX86State *env, DisasContext *s, int b,
2979 target_ulong pc_start, int rex_r)
2980 {
2981 int b1, op1_offset, op2_offset, is_xmm, val;
2982 int modrm, mod, rm, reg;
2983 SSEFunc_0_epp sse_fn_epp;
2984 SSEFunc_0_eppi sse_fn_eppi;
2985 SSEFunc_0_ppi sse_fn_ppi;
2986 SSEFunc_0_eppt sse_fn_eppt;
2987 TCGMemOp ot;
2988
2989 b &= 0xff;
2990 if (s->prefix & PREFIX_DATA)
2991 b1 = 1;
2992 else if (s->prefix & PREFIX_REPZ)
2993 b1 = 2;
2994 else if (s->prefix & PREFIX_REPNZ)
2995 b1 = 3;
2996 else
2997 b1 = 0;
2998 sse_fn_epp = sse_op_table1[b][b1];
2999 if (!sse_fn_epp) {
3000 goto illegal_op;
3001 }
3002 if ((b <= 0x5f && b >= 0x10) || b == 0xc6 || b == 0xc2) {
3003 is_xmm = 1;
3004 } else {
3005 if (b1 == 0) {
3006 /* MMX case */
3007 is_xmm = 0;
3008 } else {
3009 is_xmm = 1;
3010 }
3011 }
3012 /* simple MMX/SSE operation */
3013 if (s->flags & HF_TS_MASK) {
3014 gen_exception(s, EXCP07_PREX, pc_start - s->cs_base);
3015 return;
3016 }
3017 if (s->flags & HF_EM_MASK) {
3018 illegal_op:
3019 gen_exception(s, EXCP06_ILLOP, pc_start - s->cs_base);
3020 return;
3021 }
3022 if (is_xmm && !(s->flags & HF_OSFXSR_MASK))
3023 if ((b != 0x38 && b != 0x3a) || (s->prefix & PREFIX_DATA))
3024 goto illegal_op;
3025 if (b == 0x0e) {
3026 if (!(s->cpuid_ext2_features & CPUID_EXT2_3DNOW))
3027 goto illegal_op;
3028 /* femms */
3029 gen_helper_emms(cpu_env);
3030 return;
3031 }
3032 if (b == 0x77) {
3033 /* emms */
3034 gen_helper_emms(cpu_env);
3035 return;
3036 }
3037 /* prepare MMX state (XXX: optimize by storing fptt and fptags in
3038 the static cpu state) */
3039 if (!is_xmm) {
3040 gen_helper_enter_mmx(cpu_env);
3041 }
3042
3043 modrm = cpu_ldub_code(env, s->pc++);
3044 reg = ((modrm >> 3) & 7);
3045 if (is_xmm)
3046 reg |= rex_r;
3047 mod = (modrm >> 6) & 3;
3048 if (sse_fn_epp == SSE_SPECIAL) {
3049 b |= (b1 << 8);
3050 switch(b) {
3051 case 0x0e7: /* movntq */
3052 if (mod == 3)
3053 goto illegal_op;
3054 gen_lea_modrm(env, s, modrm);
3055 gen_stq_env_A0(s, offsetof(CPUX86State, fpregs[reg].mmx));
3056 break;
3057 case 0x1e7: /* movntdq */
3058 case 0x02b: /* movntps */
3059 case 0x12b: /* movntps */
3060 if (mod == 3)
3061 goto illegal_op;
3062 gen_lea_modrm(env, s, modrm);
3063 gen_sto_env_A0(s, offsetof(CPUX86State, xmm_regs[reg]));
3064 break;
3065 case 0x3f0: /* lddqu */
3066 if (mod == 3)
3067 goto illegal_op;
3068 gen_lea_modrm(env, s, modrm);
3069 gen_ldo_env_A0(s, offsetof(CPUX86State, xmm_regs[reg]));
3070 break;
3071 case 0x22b: /* movntss */
3072 case 0x32b: /* movntsd */
3073 if (mod == 3)
3074 goto illegal_op;
3075 gen_lea_modrm(env, s, modrm);
3076 if (b1 & 1) {
3077 gen_stq_env_A0(s, offsetof(CPUX86State, xmm_regs[reg]));
3078 } else {
3079 tcg_gen_ld32u_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,
3080 xmm_regs[reg].XMM_L(0)));
3081 gen_op_st_v(s, MO_32, cpu_T[0], cpu_A0);
3082 }
3083 break;
3084 case 0x6e: /* movd mm, ea */
3085 #ifdef TARGET_X86_64
3086 if (s->dflag == MO_64) {
3087 gen_ldst_modrm(env, s, modrm, MO_64, OR_TMP0, 0);
3088 tcg_gen_st_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,fpregs[reg].mmx));
3089 } else
3090 #endif
3091 {
3092 gen_ldst_modrm(env, s, modrm, MO_32, OR_TMP0, 0);
3093 tcg_gen_addi_ptr(cpu_ptr0, cpu_env,
3094 offsetof(CPUX86State,fpregs[reg].mmx));
3095 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
3096 gen_helper_movl_mm_T0_mmx(cpu_ptr0, cpu_tmp2_i32);
3097 }
3098 break;
3099 case 0x16e: /* movd xmm, ea */
3100 #ifdef TARGET_X86_64
3101 if (s->dflag == MO_64) {
3102 gen_ldst_modrm(env, s, modrm, MO_64, OR_TMP0, 0);
3103 tcg_gen_addi_ptr(cpu_ptr0, cpu_env,
3104 offsetof(CPUX86State,xmm_regs[reg]));
3105 gen_helper_movq_mm_T0_xmm(cpu_ptr0, cpu_T[0]);
3106 } else
3107 #endif
3108 {
3109 gen_ldst_modrm(env, s, modrm, MO_32, OR_TMP0, 0);
3110 tcg_gen_addi_ptr(cpu_ptr0, cpu_env,
3111 offsetof(CPUX86State,xmm_regs[reg]));
3112 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
3113 gen_helper_movl_mm_T0_xmm(cpu_ptr0, cpu_tmp2_i32);
3114 }
3115 break;
3116 case 0x6f: /* movq mm, ea */
3117 if (mod != 3) {
3118 gen_lea_modrm(env, s, modrm);
3119 gen_ldq_env_A0(s, offsetof(CPUX86State, fpregs[reg].mmx));
3120 } else {
3121 rm = (modrm & 7);
3122 tcg_gen_ld_i64(cpu_tmp1_i64, cpu_env,
3123 offsetof(CPUX86State,fpregs[rm].mmx));
3124 tcg_gen_st_i64(cpu_tmp1_i64, cpu_env,
3125 offsetof(CPUX86State,fpregs[reg].mmx));
3126 }
3127 break;
3128 case 0x010: /* movups */
3129 case 0x110: /* movupd */
3130 case 0x028: /* movaps */
3131 case 0x128: /* movapd */
3132 case 0x16f: /* movdqa xmm, ea */
3133 case 0x26f: /* movdqu xmm, ea */
3134 if (mod != 3) {
3135 gen_lea_modrm(env, s, modrm);
3136 gen_ldo_env_A0(s, offsetof(CPUX86State, xmm_regs[reg]));
3137 } else {
3138 rm = (modrm & 7) | REX_B(s);
3139 gen_op_movo(offsetof(CPUX86State,xmm_regs[reg]),
3140 offsetof(CPUX86State,xmm_regs[rm]));
3141 }
3142 break;
3143 case 0x210: /* movss xmm, ea */
3144 if (mod != 3) {
3145 gen_lea_modrm(env, s, modrm);
3146 gen_op_ld_v(s, MO_32, cpu_T[0], cpu_A0);
3147 tcg_gen_st32_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,xmm_regs[reg].XMM_L(0)));
3148 tcg_gen_movi_tl(cpu_T[0], 0);
3149 tcg_gen_st32_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,xmm_regs[reg].XMM_L(1)));
3150 tcg_gen_st32_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,xmm_regs[reg].XMM_L(2)));
3151 tcg_gen_st32_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,xmm_regs[reg].XMM_L(3)));
3152 } else {
3153 rm = (modrm & 7) | REX_B(s);
3154 gen_op_movl(offsetof(CPUX86State,xmm_regs[reg].XMM_L(0)),
3155 offsetof(CPUX86State,xmm_regs[rm].XMM_L(0)));
3156 }
3157 break;
3158 case 0x310: /* movsd xmm, ea */
3159 if (mod != 3) {
3160 gen_lea_modrm(env, s, modrm);
3161 gen_ldq_env_A0(s, offsetof(CPUX86State,
3162 xmm_regs[reg].XMM_Q(0)));
3163 tcg_gen_movi_tl(cpu_T[0], 0);
3164 tcg_gen_st32_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,xmm_regs[reg].XMM_L(2)));
3165 tcg_gen_st32_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,xmm_regs[reg].XMM_L(3)));
3166 } else {
3167 rm = (modrm & 7) | REX_B(s);
3168 gen_op_movq(offsetof(CPUX86State,xmm_regs[reg].XMM_Q(0)),
3169 offsetof(CPUX86State,xmm_regs[rm].XMM_Q(0)));
3170 }
3171 break;
3172 case 0x012: /* movlps */
3173 case 0x112: /* movlpd */
3174 if (mod != 3) {
3175 gen_lea_modrm(env, s, modrm);
3176 gen_ldq_env_A0(s, offsetof(CPUX86State,
3177 xmm_regs[reg].XMM_Q(0)));
3178 } else {
3179 /* movhlps */
3180 rm = (modrm & 7) | REX_B(s);
3181 gen_op_movq(offsetof(CPUX86State,xmm_regs[reg].XMM_Q(0)),
3182 offsetof(CPUX86State,xmm_regs[rm].XMM_Q(1)));
3183 }
3184 break;
3185 case 0x212: /* movsldup */
3186 if (mod != 3) {
3187 gen_lea_modrm(env, s, modrm);
3188 gen_ldo_env_A0(s, offsetof(CPUX86State, xmm_regs[reg]));
3189 } else {
3190 rm = (modrm & 7) | REX_B(s);
3191 gen_op_movl(offsetof(CPUX86State,xmm_regs[reg].XMM_L(0)),
3192 offsetof(CPUX86State,xmm_regs[rm].XMM_L(0)));
3193 gen_op_movl(offsetof(CPUX86State,xmm_regs[reg].XMM_L(2)),
3194 offsetof(CPUX86State,xmm_regs[rm].XMM_L(2)));
3195 }
3196 gen_op_movl(offsetof(CPUX86State,xmm_regs[reg].XMM_L(1)),
3197 offsetof(CPUX86State,xmm_regs[reg].XMM_L(0)));
3198 gen_op_movl(offsetof(CPUX86State,xmm_regs[reg].XMM_L(3)),
3199 offsetof(CPUX86State,xmm_regs[reg].XMM_L(2)));
3200 break;
3201 case 0x312: /* movddup */
3202 if (mod != 3) {
3203 gen_lea_modrm(env, s, modrm);
3204 gen_ldq_env_A0(s, offsetof(CPUX86State,
3205 xmm_regs[reg].XMM_Q(0)));
3206 } else {
3207 rm = (modrm & 7) | REX_B(s);
3208 gen_op_movq(offsetof(CPUX86State,xmm_regs[reg].XMM_Q(0)),
3209 offsetof(CPUX86State,xmm_regs[rm].XMM_Q(0)));
3210 }
3211 gen_op_movq(offsetof(CPUX86State,xmm_regs[reg].XMM_Q(1)),
3212 offsetof(CPUX86State,xmm_regs[reg].XMM_Q(0)));
3213 break;
3214 case 0x016: /* movhps */
3215 case 0x116: /* movhpd */
3216 if (mod != 3) {
3217 gen_lea_modrm(env, s, modrm);
3218 gen_ldq_env_A0(s, offsetof(CPUX86State,
3219 xmm_regs[reg].XMM_Q(1)));
3220 } else {
3221 /* movlhps */
3222 rm = (modrm & 7) | REX_B(s);
3223 gen_op_movq(offsetof(CPUX86State,xmm_regs[reg].XMM_Q(1)),
3224 offsetof(CPUX86State,xmm_regs[rm].XMM_Q(0)));
3225 }
3226 break;
3227 case 0x216: /* movshdup */
3228 if (mod != 3) {
3229 gen_lea_modrm(env, s, modrm);
3230 gen_ldo_env_A0(s, offsetof(CPUX86State, xmm_regs[reg]));
3231 } else {
3232 rm = (modrm & 7) | REX_B(s);
3233 gen_op_movl(offsetof(CPUX86State,xmm_regs[reg].XMM_L(1)),
3234 offsetof(CPUX86State,xmm_regs[rm].XMM_L(1)));
3235 gen_op_movl(offsetof(CPUX86State,xmm_regs[reg].XMM_L(3)),
3236 offsetof(CPUX86State,xmm_regs[rm].XMM_L(3)));
3237 }
3238 gen_op_movl(offsetof(CPUX86State,xmm_regs[reg].XMM_L(0)),
3239 offsetof(CPUX86State,xmm_regs[reg].XMM_L(1)));
3240 gen_op_movl(offsetof(CPUX86State,xmm_regs[reg].XMM_L(2)),
3241 offsetof(CPUX86State,xmm_regs[reg].XMM_L(3)));
3242 break;
3243 case 0x178:
3244 case 0x378:
3245 {
3246 int bit_index, field_length;
3247
3248 if (b1 == 1 && reg != 0)
3249 goto illegal_op;
3250 field_length = cpu_ldub_code(env, s->pc++) & 0x3F;
3251 bit_index = cpu_ldub_code(env, s->pc++) & 0x3F;
3252 tcg_gen_addi_ptr(cpu_ptr0, cpu_env,
3253 offsetof(CPUX86State,xmm_regs[reg]));
3254 if (b1 == 1)
3255 gen_helper_extrq_i(cpu_env, cpu_ptr0,
3256 tcg_const_i32(bit_index),
3257 tcg_const_i32(field_length));
3258 else
3259 gen_helper_insertq_i(cpu_env, cpu_ptr0,
3260 tcg_const_i32(bit_index),
3261 tcg_const_i32(field_length));
3262 }
3263 break;
3264 case 0x7e: /* movd ea, mm */
3265 #ifdef TARGET_X86_64
3266 if (s->dflag == MO_64) {
3267 tcg_gen_ld_i64(cpu_T[0], cpu_env,
3268 offsetof(CPUX86State,fpregs[reg].mmx));
3269 gen_ldst_modrm(env, s, modrm, MO_64, OR_TMP0, 1);
3270 } else
3271 #endif
3272 {
3273 tcg_gen_ld32u_tl(cpu_T[0], cpu_env,
3274 offsetof(CPUX86State,fpregs[reg].mmx.MMX_L(0)));
3275 gen_ldst_modrm(env, s, modrm, MO_32, OR_TMP0, 1);
3276 }
3277 break;
3278 case 0x17e: /* movd ea, xmm */
3279 #ifdef TARGET_X86_64
3280 if (s->dflag == MO_64) {
3281 tcg_gen_ld_i64(cpu_T[0], cpu_env,
3282 offsetof(CPUX86State,xmm_regs[reg].XMM_Q(0)));
3283 gen_ldst_modrm(env, s, modrm, MO_64, OR_TMP0, 1);
3284 } else
3285 #endif
3286 {
3287 tcg_gen_ld32u_tl(cpu_T[0], cpu_env,
3288 offsetof(CPUX86State,xmm_regs[reg].XMM_L(0)));
3289 gen_ldst_modrm(env, s, modrm, MO_32, OR_TMP0, 1);
3290 }
3291 break;
3292 case 0x27e: /* movq xmm, ea */
3293 if (mod != 3) {
3294 gen_lea_modrm(env, s, modrm);
3295 gen_ldq_env_A0(s, offsetof(CPUX86State,
3296 xmm_regs[reg].XMM_Q(0)));
3297 } else {
3298 rm = (modrm & 7) | REX_B(s);
3299 gen_op_movq(offsetof(CPUX86State,xmm_regs[reg].XMM_Q(0)),
3300 offsetof(CPUX86State,xmm_regs[rm].XMM_Q(0)));
3301 }
3302 gen_op_movq_env_0(offsetof(CPUX86State,xmm_regs[reg].XMM_Q(1)));
3303 break;
3304 case 0x7f: /* movq ea, mm */
3305 if (mod != 3) {
3306 gen_lea_modrm(env, s, modrm);
3307 gen_stq_env_A0(s, offsetof(CPUX86State, fpregs[reg].mmx));
3308 } else {
3309 rm = (modrm & 7);
3310 gen_op_movq(offsetof(CPUX86State,fpregs[rm].mmx),
3311 offsetof(CPUX86State,fpregs[reg].mmx));
3312 }
3313 break;
3314 case 0x011: /* movups */
3315 case 0x111: /* movupd */
3316 case 0x029: /* movaps */
3317 case 0x129: /* movapd */
3318 case 0x17f: /* movdqa ea, xmm */
3319 case 0x27f: /* movdqu ea, xmm */
3320 if (mod != 3) {
3321 gen_lea_modrm(env, s, modrm);
3322 gen_sto_env_A0(s, offsetof(CPUX86State, xmm_regs[reg]));
3323 } else {
3324 rm = (modrm & 7) | REX_B(s);
3325 gen_op_movo(offsetof(CPUX86State,xmm_regs[rm]),
3326 offsetof(CPUX86State,xmm_regs[reg]));
3327 }
3328 break;
3329 case 0x211: /* movss ea, xmm */
3330 if (mod != 3) {
3331 gen_lea_modrm(env, s, modrm);
3332 tcg_gen_ld32u_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,xmm_regs[reg].XMM_L(0)));
3333 gen_op_st_v(s, MO_32, cpu_T[0], cpu_A0);
3334 } else {
3335 rm = (modrm & 7) | REX_B(s);
3336 gen_op_movl(offsetof(CPUX86State,xmm_regs[rm].XMM_L(0)),
3337 offsetof(CPUX86State,xmm_regs[reg].XMM_L(0)));
3338 }
3339 break;
3340 case 0x311: /* movsd ea, xmm */
3341 if (mod != 3) {
3342 gen_lea_modrm(env, s, modrm);
3343 gen_stq_env_A0(s, offsetof(CPUX86State,
3344 xmm_regs[reg].XMM_Q(0)));
3345 } else {
3346 rm = (modrm & 7) | REX_B(s);
3347 gen_op_movq(offsetof(CPUX86State,xmm_regs[rm].XMM_Q(0)),
3348 offsetof(CPUX86State,xmm_regs[reg].XMM_Q(0)));
3349 }
3350 break;
3351 case 0x013: /* movlps */
3352 case 0x113: /* movlpd */
3353 if (mod != 3) {
3354 gen_lea_modrm(env, s, modrm);
3355 gen_stq_env_A0(s, offsetof(CPUX86State,
3356 xmm_regs[reg].XMM_Q(0)));
3357 } else {
3358 goto illegal_op;
3359 }
3360 break;
3361 case 0x017: /* movhps */
3362 case 0x117: /* movhpd */
3363 if (mod != 3) {
3364 gen_lea_modrm(env, s, modrm);
3365 gen_stq_env_A0(s, offsetof(CPUX86State,
3366 xmm_regs[reg].XMM_Q(1)));
3367 } else {
3368 goto illegal_op;
3369 }
3370 break;
3371 case 0x71: /* shift mm, im */
3372 case 0x72:
3373 case 0x73:
3374 case 0x171: /* shift xmm, im */
3375 case 0x172:
3376 case 0x173:
3377 if (b1 >= 2) {
3378 goto illegal_op;
3379 }
3380 val = cpu_ldub_code(env, s->pc++);
3381 if (is_xmm) {
3382 tcg_gen_movi_tl(cpu_T[0], val);
3383 tcg_gen_st32_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,xmm_t0.XMM_L(0)));
3384 tcg_gen_movi_tl(cpu_T[0], 0);
3385 tcg_gen_st32_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,xmm_t0.XMM_L(1)));
3386 op1_offset = offsetof(CPUX86State,xmm_t0);
3387 } else {
3388 tcg_gen_movi_tl(cpu_T[0], val);
3389 tcg_gen_st32_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,mmx_t0.MMX_L(0)));
3390 tcg_gen_movi_tl(cpu_T[0], 0);
3391 tcg_gen_st32_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,mmx_t0.MMX_L(1)));
3392 op1_offset = offsetof(CPUX86State,mmx_t0);
3393 }
3394 sse_fn_epp = sse_op_table2[((b - 1) & 3) * 8 +
3395 (((modrm >> 3)) & 7)][b1];
3396 if (!sse_fn_epp) {
3397 goto illegal_op;
3398 }
3399 if (is_xmm) {
3400 rm = (modrm & 7) | REX_B(s);
3401 op2_offset = offsetof(CPUX86State,xmm_regs[rm]);
3402 } else {
3403 rm = (modrm & 7);
3404 op2_offset = offsetof(CPUX86State,fpregs[rm].mmx);
3405 }
3406 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, op2_offset);
3407 tcg_gen_addi_ptr(cpu_ptr1, cpu_env, op1_offset);
3408 sse_fn_epp(cpu_env, cpu_ptr0, cpu_ptr1);
3409 break;
3410 case 0x050: /* movmskps */
3411 rm = (modrm & 7) | REX_B(s);
3412 tcg_gen_addi_ptr(cpu_ptr0, cpu_env,
3413 offsetof(CPUX86State,xmm_regs[rm]));
3414 gen_helper_movmskps(cpu_tmp2_i32, cpu_env, cpu_ptr0);
3415 tcg_gen_extu_i32_tl(cpu_regs[reg], cpu_tmp2_i32);
3416 break;
3417 case 0x150: /* movmskpd */
3418 rm = (modrm & 7) | REX_B(s);
3419 tcg_gen_addi_ptr(cpu_ptr0, cpu_env,
3420 offsetof(CPUX86State,xmm_regs[rm]));
3421 gen_helper_movmskpd(cpu_tmp2_i32, cpu_env, cpu_ptr0);
3422 tcg_gen_extu_i32_tl(cpu_regs[reg], cpu_tmp2_i32);
3423 break;
3424 case 0x02a: /* cvtpi2ps */
3425 case 0x12a: /* cvtpi2pd */
3426 gen_helper_enter_mmx(cpu_env);
3427 if (mod != 3) {
3428 gen_lea_modrm(env, s, modrm);
3429 op2_offset = offsetof(CPUX86State,mmx_t0);
3430 gen_ldq_env_A0(s, op2_offset);
3431 } else {
3432 rm = (modrm & 7);
3433 op2_offset = offsetof(CPUX86State,fpregs[rm].mmx);
3434 }
3435 op1_offset = offsetof(CPUX86State,xmm_regs[reg]);
3436 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, op1_offset);
3437 tcg_gen_addi_ptr(cpu_ptr1, cpu_env, op2_offset);
3438 switch(b >> 8) {
3439 case 0x0:
3440 gen_helper_cvtpi2ps(cpu_env, cpu_ptr0, cpu_ptr1);
3441 break;
3442 default:
3443 case 0x1:
3444 gen_helper_cvtpi2pd(cpu_env, cpu_ptr0, cpu_ptr1);
3445 break;
3446 }
3447 break;
3448 case 0x22a: /* cvtsi2ss */
3449 case 0x32a: /* cvtsi2sd */
3450 ot = mo_64_32(s->dflag);
3451 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
3452 op1_offset = offsetof(CPUX86State,xmm_regs[reg]);
3453 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, op1_offset);
3454 if (ot == MO_32) {
3455 SSEFunc_0_epi sse_fn_epi = sse_op_table3ai[(b >> 8) & 1];
3456 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
3457 sse_fn_epi(cpu_env, cpu_ptr0, cpu_tmp2_i32);
3458 } else {
3459 #ifdef TARGET_X86_64
3460 SSEFunc_0_epl sse_fn_epl = sse_op_table3aq[(b >> 8) & 1];
3461 sse_fn_epl(cpu_env, cpu_ptr0, cpu_T[0]);
3462 #else
3463 goto illegal_op;
3464 #endif
3465 }
3466 break;
3467 case 0x02c: /* cvttps2pi */
3468 case 0x12c: /* cvttpd2pi */
3469 case 0x02d: /* cvtps2pi */
3470 case 0x12d: /* cvtpd2pi */
3471 gen_helper_enter_mmx(cpu_env);
3472 if (mod != 3) {
3473 gen_lea_modrm(env, s, modrm);
3474 op2_offset = offsetof(CPUX86State,xmm_t0);
3475 gen_ldo_env_A0(s, op2_offset);
3476 } else {
3477 rm = (modrm & 7) | REX_B(s);
3478 op2_offset = offsetof(CPUX86State,xmm_regs[rm]);
3479 }
3480 op1_offset = offsetof(CPUX86State,fpregs[reg & 7].mmx);
3481 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, op1_offset);
3482 tcg_gen_addi_ptr(cpu_ptr1, cpu_env, op2_offset);
3483 switch(b) {
3484 case 0x02c:
3485 gen_helper_cvttps2pi(cpu_env, cpu_ptr0, cpu_ptr1);
3486 break;
3487 case 0x12c:
3488 gen_helper_cvttpd2pi(cpu_env, cpu_ptr0, cpu_ptr1);
3489 break;
3490 case 0x02d:
3491 gen_helper_cvtps2pi(cpu_env, cpu_ptr0, cpu_ptr1);
3492 break;
3493 case 0x12d:
3494 gen_helper_cvtpd2pi(cpu_env, cpu_ptr0, cpu_ptr1);
3495 break;
3496 }
3497 break;
3498 case 0x22c: /* cvttss2si */
3499 case 0x32c: /* cvttsd2si */
3500 case 0x22d: /* cvtss2si */
3501 case 0x32d: /* cvtsd2si */
3502 ot = mo_64_32(s->dflag);
3503 if (mod != 3) {
3504 gen_lea_modrm(env, s, modrm);
3505 if ((b >> 8) & 1) {
3506 gen_ldq_env_A0(s, offsetof(CPUX86State, xmm_t0.XMM_Q(0)));
3507 } else {
3508 gen_op_ld_v(s, MO_32, cpu_T[0], cpu_A0);
3509 tcg_gen_st32_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,xmm_t0.XMM_L(0)));
3510 }
3511 op2_offset = offsetof(CPUX86State,xmm_t0);
3512 } else {
3513 rm = (modrm & 7) | REX_B(s);
3514 op2_offset = offsetof(CPUX86State,xmm_regs[rm]);
3515 }
3516 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, op2_offset);
3517 if (ot == MO_32) {
3518 SSEFunc_i_ep sse_fn_i_ep =
3519 sse_op_table3bi[((b >> 7) & 2) | (b & 1)];
3520 sse_fn_i_ep(cpu_tmp2_i32, cpu_env, cpu_ptr0);
3521 tcg_gen_extu_i32_tl(cpu_T[0], cpu_tmp2_i32);
3522 } else {
3523 #ifdef TARGET_X86_64
3524 SSEFunc_l_ep sse_fn_l_ep =
3525 sse_op_table3bq[((b >> 7) & 2) | (b & 1)];
3526 sse_fn_l_ep(cpu_T[0], cpu_env, cpu_ptr0);
3527 #else
3528 goto illegal_op;
3529 #endif
3530 }
3531 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
3532 break;
3533 case 0xc4: /* pinsrw */
3534 case 0x1c4:
3535 s->rip_offset = 1;
3536 gen_ldst_modrm(env, s, modrm, MO_16, OR_TMP0, 0);
3537 val = cpu_ldub_code(env, s->pc++);
3538 if (b1) {
3539 val &= 7;
3540 tcg_gen_st16_tl(cpu_T[0], cpu_env,
3541 offsetof(CPUX86State,xmm_regs[reg].XMM_W(val)));
3542 } else {
3543 val &= 3;
3544 tcg_gen_st16_tl(cpu_T[0], cpu_env,
3545 offsetof(CPUX86State,fpregs[reg].mmx.MMX_W(val)));
3546 }
3547 break;
3548 case 0xc5: /* pextrw */
3549 case 0x1c5:
3550 if (mod != 3)
3551 goto illegal_op;
3552 ot = mo_64_32(s->dflag);
3553 val = cpu_ldub_code(env, s->pc++);
3554 if (b1) {
3555 val &= 7;
3556 rm = (modrm & 7) | REX_B(s);
3557 tcg_gen_ld16u_tl(cpu_T[0], cpu_env,
3558 offsetof(CPUX86State,xmm_regs[rm].XMM_W(val)));
3559 } else {
3560 val &= 3;
3561 rm = (modrm & 7);
3562 tcg_gen_ld16u_tl(cpu_T[0], cpu_env,
3563 offsetof(CPUX86State,fpregs[rm].mmx.MMX_W(val)));
3564 }
3565 reg = ((modrm >> 3) & 7) | rex_r;
3566 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
3567 break;
3568 case 0x1d6: /* movq ea, xmm */
3569 if (mod != 3) {
3570 gen_lea_modrm(env, s, modrm);
3571 gen_stq_env_A0(s, offsetof(CPUX86State,
3572 xmm_regs[reg].XMM_Q(0)));
3573 } else {
3574 rm = (modrm & 7) | REX_B(s);
3575 gen_op_movq(offsetof(CPUX86State,xmm_regs[rm].XMM_Q(0)),
3576 offsetof(CPUX86State,xmm_regs[reg].XMM_Q(0)));
3577 gen_op_movq_env_0(offsetof(CPUX86State,xmm_regs[rm].XMM_Q(1)));
3578 }
3579 break;
3580 case 0x2d6: /* movq2dq */
3581 gen_helper_enter_mmx(cpu_env);
3582 rm = (modrm & 7);
3583 gen_op_movq(offsetof(CPUX86State,xmm_regs[reg].XMM_Q(0)),
3584 offsetof(CPUX86State,fpregs[rm].mmx));
3585 gen_op_movq_env_0(offsetof(CPUX86State,xmm_regs[reg].XMM_Q(1)));
3586 break;
3587 case 0x3d6: /* movdq2q */
3588 gen_helper_enter_mmx(cpu_env);
3589 rm = (modrm & 7) | REX_B(s);
3590 gen_op_movq(offsetof(CPUX86State,fpregs[reg & 7].mmx),
3591 offsetof(CPUX86State,xmm_regs[rm].XMM_Q(0)));
3592 break;
3593 case 0xd7: /* pmovmskb */
3594 case 0x1d7:
3595 if (mod != 3)
3596 goto illegal_op;
3597 if (b1) {
3598 rm = (modrm & 7) | REX_B(s);
3599 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, offsetof(CPUX86State,xmm_regs[rm]));
3600 gen_helper_pmovmskb_xmm(cpu_tmp2_i32, cpu_env, cpu_ptr0);
3601 } else {
3602 rm = (modrm & 7);
3603 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, offsetof(CPUX86State,fpregs[rm].mmx));
3604 gen_helper_pmovmskb_mmx(cpu_tmp2_i32, cpu_env, cpu_ptr0);
3605 }
3606 reg = ((modrm >> 3) & 7) | rex_r;
3607 tcg_gen_extu_i32_tl(cpu_regs[reg], cpu_tmp2_i32);
3608 break;
3609
3610 case 0x138:
3611 case 0x038:
3612 b = modrm;
3613 if ((b & 0xf0) == 0xf0) {
3614 goto do_0f_38_fx;
3615 }
3616 modrm = cpu_ldub_code(env, s->pc++);
3617 rm = modrm & 7;
3618 reg = ((modrm >> 3) & 7) | rex_r;
3619 mod = (modrm >> 6) & 3;
3620 if (b1 >= 2) {
3621 goto illegal_op;
3622 }
3623
3624 sse_fn_epp = sse_op_table6[b].op[b1];
3625 if (!sse_fn_epp) {
3626 goto illegal_op;
3627 }
3628 if (!(s->cpuid_ext_features & sse_op_table6[b].ext_mask))
3629 goto illegal_op;
3630
3631 if (b1) {
3632 op1_offset = offsetof(CPUX86State,xmm_regs[reg]);
3633 if (mod == 3) {
3634 op2_offset = offsetof(CPUX86State,xmm_regs[rm | REX_B(s)]);
3635 } else {
3636 op2_offset = offsetof(CPUX86State,xmm_t0);
3637 gen_lea_modrm(env, s, modrm);
3638 switch (b) {
3639 case 0x20: case 0x30: /* pmovsxbw, pmovzxbw */
3640 case 0x23: case 0x33: /* pmovsxwd, pmovzxwd */
3641 case 0x25: case 0x35: /* pmovsxdq, pmovzxdq */
3642 gen_ldq_env_A0(s, op2_offset +
3643 offsetof(XMMReg, XMM_Q(0)));
3644 break;
3645 case 0x21: case 0x31: /* pmovsxbd, pmovzxbd */
3646 case 0x24: case 0x34: /* pmovsxwq, pmovzxwq */
3647 tcg_gen_qemu_ld_i32(cpu_tmp2_i32, cpu_A0,
3648 s->mem_index, MO_LEUL);
3649 tcg_gen_st_i32(cpu_tmp2_i32, cpu_env, op2_offset +
3650 offsetof(XMMReg, XMM_L(0)));
3651 break;
3652 case 0x22: case 0x32: /* pmovsxbq, pmovzxbq */
3653 tcg_gen_qemu_ld_tl(cpu_tmp0, cpu_A0,
3654 s->mem_index, MO_LEUW);
3655 tcg_gen_st16_tl(cpu_tmp0, cpu_env, op2_offset +
3656 offsetof(XMMReg, XMM_W(0)));
3657 break;
3658 case 0x2a: /* movntqda */
3659 gen_ldo_env_A0(s, op1_offset);
3660 return;
3661 default:
3662 gen_ldo_env_A0(s, op2_offset);
3663 }
3664 }
3665 } else {
3666 op1_offset = offsetof(CPUX86State,fpregs[reg].mmx);
3667 if (mod == 3) {
3668 op2_offset = offsetof(CPUX86State,fpregs[rm].mmx);
3669 } else {
3670 op2_offset = offsetof(CPUX86State,mmx_t0);
3671 gen_lea_modrm(env, s, modrm);
3672 gen_ldq_env_A0(s, op2_offset);
3673 }
3674 }
3675 if (sse_fn_epp == SSE_SPECIAL) {
3676 goto illegal_op;
3677 }
3678
3679 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, op1_offset);
3680 tcg_gen_addi_ptr(cpu_ptr1, cpu_env, op2_offset);
3681 sse_fn_epp(cpu_env, cpu_ptr0, cpu_ptr1);
3682
3683 if (b == 0x17) {
3684 set_cc_op(s, CC_OP_EFLAGS);
3685 }
3686 break;
3687
3688 case 0x238:
3689 case 0x338:
3690 do_0f_38_fx:
3691 /* Various integer extensions at 0f 38 f[0-f]. */
3692 b = modrm | (b1 << 8);
3693 modrm = cpu_ldub_code(env, s->pc++);
3694 reg = ((modrm >> 3) & 7) | rex_r;
3695
3696 switch (b) {
3697 case 0x3f0: /* crc32 Gd,Eb */
3698 case 0x3f1: /* crc32 Gd,Ey */
3699 do_crc32:
3700 if (!(s->cpuid_ext_features & CPUID_EXT_SSE42)) {
3701 goto illegal_op;
3702 }
3703 if ((b & 0xff) == 0xf0) {
3704 ot = MO_8;
3705 } else if (s->dflag != MO_64) {
3706 ot = (s->prefix & PREFIX_DATA ? MO_16 : MO_32);
3707 } else {
3708 ot = MO_64;
3709 }
3710
3711 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_regs[reg]);
3712 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
3713 gen_helper_crc32(cpu_T[0], cpu_tmp2_i32,
3714 cpu_T[0], tcg_const_i32(8 << ot));
3715
3716 ot = mo_64_32(s->dflag);
3717 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
3718 break;
3719
3720 case 0x1f0: /* crc32 or movbe */
3721 case 0x1f1:
3722 /* For these insns, the f3 prefix is supposed to have priority
3723 over the 66 prefix, but that's not what we implement above
3724 setting b1. */
3725 if (s->prefix & PREFIX_REPNZ) {
3726 goto do_crc32;
3727 }
3728 /* FALLTHRU */
3729 case 0x0f0: /* movbe Gy,My */
3730 case 0x0f1: /* movbe My,Gy */
3731 if (!(s->cpuid_ext_features & CPUID_EXT_MOVBE)) {
3732 goto illegal_op;
3733 }
3734 if (s->dflag != MO_64) {
3735 ot = (s->prefix & PREFIX_DATA ? MO_16 : MO_32);
3736 } else {
3737 ot = MO_64;
3738 }
3739
3740 gen_lea_modrm(env, s, modrm);
3741 if ((b & 1) == 0) {
3742 tcg_gen_qemu_ld_tl(cpu_T[0], cpu_A0,
3743 s->mem_index, ot | MO_BE);
3744 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
3745 } else {
3746 tcg_gen_qemu_st_tl(cpu_regs[reg], cpu_A0,
3747 s->mem_index, ot | MO_BE);
3748 }
3749 break;
3750
3751 case 0x0f2: /* andn Gy, By, Ey */
3752 if (!(s->cpuid_7_0_ebx_features & CPUID_7_0_EBX_BMI1)
3753 || !(s->prefix & PREFIX_VEX)
3754 || s->vex_l != 0) {
3755 goto illegal_op;
3756 }
3757 ot = mo_64_32(s->dflag);
3758 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
3759 tcg_gen_andc_tl(cpu_T[0], cpu_regs[s->vex_v], cpu_T[0]);
3760 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
3761 gen_op_update1_cc();
3762 set_cc_op(s, CC_OP_LOGICB + ot);
3763 break;
3764
3765 case 0x0f7: /* bextr Gy, Ey, By */
3766 if (!(s->cpuid_7_0_ebx_features & CPUID_7_0_EBX_BMI1)
3767 || !(s->prefix & PREFIX_VEX)
3768 || s->vex_l != 0) {
3769 goto illegal_op;
3770 }
3771 ot = mo_64_32(s->dflag);
3772 {
3773 TCGv bound, zero;
3774
3775 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
3776 /* Extract START, and shift the operand.
3777 Shifts larger than operand size get zeros. */
3778 tcg_gen_ext8u_tl(cpu_A0, cpu_regs[s->vex_v]);
3779 tcg_gen_shr_tl(cpu_T[0], cpu_T[0], cpu_A0);
3780
3781 bound = tcg_const_tl(ot == MO_64 ? 63 : 31);
3782 zero = tcg_const_tl(0);
3783 tcg_gen_movcond_tl(TCG_COND_LEU, cpu_T[0], cpu_A0, bound,
3784 cpu_T[0], zero);
3785 tcg_temp_free(zero);
3786
3787 /* Extract the LEN into a mask. Lengths larger than
3788 operand size get all ones. */
3789 tcg_gen_shri_tl(cpu_A0, cpu_regs[s->vex_v], 8);
3790 tcg_gen_ext8u_tl(cpu_A0, cpu_A0);
3791 tcg_gen_movcond_tl(TCG_COND_LEU, cpu_A0, cpu_A0, bound,
3792 cpu_A0, bound);
3793 tcg_temp_free(bound);
3794 tcg_gen_movi_tl(cpu_T[1], 1);
3795 tcg_gen_shl_tl(cpu_T[1], cpu_T[1], cpu_A0);
3796 tcg_gen_subi_tl(cpu_T[1], cpu_T[1], 1);
3797 tcg_gen_and_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
3798
3799 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
3800 gen_op_update1_cc();
3801 set_cc_op(s, CC_OP_LOGICB + ot);
3802 }
3803 break;
3804
3805 case 0x0f5: /* bzhi Gy, Ey, By */
3806 if (!(s->cpuid_7_0_ebx_features & CPUID_7_0_EBX_BMI2)
3807 || !(s->prefix & PREFIX_VEX)
3808 || s->vex_l != 0) {
3809 goto illegal_op;
3810 }
3811 ot = mo_64_32(s->dflag);
3812 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
3813 tcg_gen_ext8u_tl(cpu_T[1], cpu_regs[s->vex_v]);
3814 {
3815 TCGv bound = tcg_const_tl(ot == MO_64 ? 63 : 31);
3816 /* Note that since we're using BMILG (in order to get O
3817 cleared) we need to store the inverse into C. */
3818 tcg_gen_setcond_tl(TCG_COND_LT, cpu_cc_src,
3819 cpu_T[1], bound);
3820 tcg_gen_movcond_tl(TCG_COND_GT, cpu_T[1], cpu_T[1],
3821 bound, bound, cpu_T[1]);
3822 tcg_temp_free(bound);
3823 }
3824 tcg_gen_movi_tl(cpu_A0, -1);
3825 tcg_gen_shl_tl(cpu_A0, cpu_A0, cpu_T[1]);
3826 tcg_gen_andc_tl(cpu_T[0], cpu_T[0], cpu_A0);
3827 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
3828 gen_op_update1_cc();
3829 set_cc_op(s, CC_OP_BMILGB + ot);
3830 break;
3831
3832 case 0x3f6: /* mulx By, Gy, rdx, Ey */
3833 if (!(s->cpuid_7_0_ebx_features & CPUID_7_0_EBX_BMI2)
3834 || !(s->prefix & PREFIX_VEX)
3835 || s->vex_l != 0) {
3836 goto illegal_op;
3837 }
3838 ot = mo_64_32(s->dflag);
3839 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
3840 switch (ot) {
3841 default:
3842 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
3843 tcg_gen_trunc_tl_i32(cpu_tmp3_i32, cpu_regs[R_EDX]);
3844 tcg_gen_mulu2_i32(cpu_tmp2_i32, cpu_tmp3_i32,
3845 cpu_tmp2_i32, cpu_tmp3_i32);
3846 tcg_gen_extu_i32_tl(cpu_regs[s->vex_v], cpu_tmp2_i32);
3847 tcg_gen_extu_i32_tl(cpu_regs[reg], cpu_tmp3_i32);
3848 break;
3849 #ifdef TARGET_X86_64
3850 case MO_64:
3851 tcg_gen_mulu2_i64(cpu_regs[s->vex_v], cpu_regs[reg],
3852 cpu_T[0], cpu_regs[R_EDX]);
3853 break;
3854 #endif
3855 }
3856 break;
3857
3858 case 0x3f5: /* pdep Gy, By, Ey */
3859 if (!(s->cpuid_7_0_ebx_features & CPUID_7_0_EBX_BMI2)
3860 || !(s->prefix & PREFIX_VEX)
3861 || s->vex_l != 0) {
3862 goto illegal_op;
3863 }
3864 ot = mo_64_32(s->dflag);
3865 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
3866 /* Note that by zero-extending the mask operand, we
3867 automatically handle zero-extending the result. */
3868 if (ot == MO_64) {
3869 tcg_gen_mov_tl(cpu_T[1], cpu_regs[s->vex_v]);
3870 } else {
3871 tcg_gen_ext32u_tl(cpu_T[1], cpu_regs[s->vex_v]);
3872 }
3873 gen_helper_pdep(cpu_regs[reg], cpu_T[0], cpu_T[1]);
3874 break;
3875
3876 case 0x2f5: /* pext Gy, By, Ey */
3877 if (!(s->cpuid_7_0_ebx_features & CPUID_7_0_EBX_BMI2)
3878 || !(s->prefix & PREFIX_VEX)
3879 || s->vex_l != 0) {
3880 goto illegal_op;
3881 }
3882 ot = mo_64_32(s->dflag);
3883 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
3884 /* Note that by zero-extending the mask operand, we
3885 automatically handle zero-extending the result. */
3886 if (ot == MO_64) {
3887 tcg_gen_mov_tl(cpu_T[1], cpu_regs[s->vex_v]);
3888 } else {
3889 tcg_gen_ext32u_tl(cpu_T[1], cpu_regs[s->vex_v]);
3890 }
3891 gen_helper_pext(cpu_regs[reg], cpu_T[0], cpu_T[1]);
3892 break;
3893
3894 case 0x1f6: /* adcx Gy, Ey */
3895 case 0x2f6: /* adox Gy, Ey */
3896 if (!(s->cpuid_7_0_ebx_features & CPUID_7_0_EBX_ADX)) {
3897 goto illegal_op;
3898 } else {
3899 TCGv carry_in, carry_out, zero;
3900 int end_op;
3901
3902 ot = mo_64_32(s->dflag);
3903 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
3904
3905 /* Re-use the carry-out from a previous round. */
3906 TCGV_UNUSED(carry_in);
3907 carry_out = (b == 0x1f6 ? cpu_cc_dst : cpu_cc_src2);
3908 switch (s->cc_op) {
3909 case CC_OP_ADCX:
3910 if (b == 0x1f6) {
3911 carry_in = cpu_cc_dst;
3912 end_op = CC_OP_ADCX;
3913 } else {
3914 end_op = CC_OP_ADCOX;
3915 }
3916 break;
3917 case CC_OP_ADOX:
3918 if (b == 0x1f6) {
3919 end_op = CC_OP_ADCOX;
3920 } else {
3921 carry_in = cpu_cc_src2;
3922 end_op = CC_OP_ADOX;
3923 }
3924 break;
3925 case CC_OP_ADCOX:
3926 end_op = CC_OP_ADCOX;
3927 carry_in = carry_out;
3928 break;
3929 default:
3930 end_op = (b == 0x1f6 ? CC_OP_ADCX : CC_OP_ADOX);
3931 break;
3932 }
3933 /* If we can't reuse carry-out, get it out of EFLAGS. */
3934 if (TCGV_IS_UNUSED(carry_in)) {
3935 if (s->cc_op != CC_OP_ADCX && s->cc_op != CC_OP_ADOX) {
3936 gen_compute_eflags(s);
3937 }
3938 carry_in = cpu_tmp0;
3939 tcg_gen_shri_tl(carry_in, cpu_cc_src,
3940 ctz32(b == 0x1f6 ? CC_C : CC_O));
3941 tcg_gen_andi_tl(carry_in, carry_in, 1);
3942 }
3943
3944 switch (ot) {
3945 #ifdef TARGET_X86_64
3946 case MO_32:
3947 /* If we know TL is 64-bit, and we want a 32-bit
3948 result, just do everything in 64-bit arithmetic. */
3949 tcg_gen_ext32u_i64(cpu_regs[reg], cpu_regs[reg]);
3950 tcg_gen_ext32u_i64(cpu_T[0], cpu_T[0]);
3951 tcg_gen_add_i64(cpu_T[0], cpu_T[0], cpu_regs[reg]);
3952 tcg_gen_add_i64(cpu_T[0], cpu_T[0], carry_in);
3953 tcg_gen_ext32u_i64(cpu_regs[reg], cpu_T[0]);
3954 tcg_gen_shri_i64(carry_out, cpu_T[0], 32);
3955 break;
3956 #endif
3957 default:
3958 /* Otherwise compute the carry-out in two steps. */
3959 zero = tcg_const_tl(0);
3960 tcg_gen_add2_tl(cpu_T[0], carry_out,
3961 cpu_T[0], zero,
3962 carry_in, zero);
3963 tcg_gen_add2_tl(cpu_regs[reg], carry_out,
3964 cpu_regs[reg], carry_out,
3965 cpu_T[0], zero);
3966 tcg_temp_free(zero);
3967 break;
3968 }
3969 set_cc_op(s, end_op);
3970 }
3971 break;
3972
3973 case 0x1f7: /* shlx Gy, Ey, By */
3974 case 0x2f7: /* sarx Gy, Ey, By */
3975 case 0x3f7: /* shrx Gy, Ey, By */
3976 if (!(s->cpuid_7_0_ebx_features & CPUID_7_0_EBX_BMI2)
3977 || !(s->prefix & PREFIX_VEX)
3978 || s->vex_l != 0) {
3979 goto illegal_op;
3980 }
3981 ot = mo_64_32(s->dflag);
3982 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
3983 if (ot == MO_64) {
3984 tcg_gen_andi_tl(cpu_T[1], cpu_regs[s->vex_v], 63);
3985 } else {
3986 tcg_gen_andi_tl(cpu_T[1], cpu_regs[s->vex_v], 31);
3987 }
3988 if (b == 0x1f7) {
3989 tcg_gen_shl_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
3990 } else if (b == 0x2f7) {
3991 if (ot != MO_64) {
3992 tcg_gen_ext32s_tl(cpu_T[0], cpu_T[0]);
3993 }
3994 tcg_gen_sar_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
3995 } else {
3996 if (ot != MO_64) {
3997 tcg_gen_ext32u_tl(cpu_T[0], cpu_T[0]);
3998 }
3999 tcg_gen_shr_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
4000 }
4001 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
4002 break;
4003
4004 case 0x0f3:
4005 case 0x1f3:
4006 case 0x2f3:
4007 case 0x3f3: /* Group 17 */
4008 if (!(s->cpuid_7_0_ebx_features & CPUID_7_0_EBX_BMI1)
4009 || !(s->prefix & PREFIX_VEX)
4010 || s->vex_l != 0) {
4011 goto illegal_op;
4012 }
4013 ot = mo_64_32(s->dflag);
4014 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
4015
4016 switch (reg & 7) {
4017 case 1: /* blsr By,Ey */
4018 tcg_gen_neg_tl(cpu_T[1], cpu_T[0]);
4019 tcg_gen_and_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
4020 gen_op_mov_reg_v(ot, s->vex_v, cpu_T[0]);
4021 gen_op_update2_cc();
4022 set_cc_op(s, CC_OP_BMILGB + ot);
4023 break;
4024
4025 case 2: /* blsmsk By,Ey */
4026 tcg_gen_mov_tl(cpu_cc_src, cpu_T[0]);
4027 tcg_gen_subi_tl(cpu_T[0], cpu_T[0], 1);
4028 tcg_gen_xor_tl(cpu_T[0], cpu_T[0], cpu_cc_src);
4029 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
4030 set_cc_op(s, CC_OP_BMILGB + ot);
4031 break;
4032
4033 case 3: /* blsi By, Ey */
4034 tcg_gen_mov_tl(cpu_cc_src, cpu_T[0]);
4035 tcg_gen_subi_tl(cpu_T[0], cpu_T[0], 1);
4036 tcg_gen_and_tl(cpu_T[0], cpu_T[0], cpu_cc_src);
4037 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
4038 set_cc_op(s, CC_OP_BMILGB + ot);
4039 break;
4040
4041 default:
4042 goto illegal_op;
4043 }
4044 break;
4045
4046 default:
4047 goto illegal_op;
4048 }
4049 break;
4050
4051 case 0x03a:
4052 case 0x13a:
4053 b = modrm;
4054 modrm = cpu_ldub_code(env, s->pc++);
4055 rm = modrm & 7;
4056 reg = ((modrm >> 3) & 7) | rex_r;
4057 mod = (modrm >> 6) & 3;
4058 if (b1 >= 2) {
4059 goto illegal_op;
4060 }
4061
4062 sse_fn_eppi = sse_op_table7[b].op[b1];
4063 if (!sse_fn_eppi) {
4064 goto illegal_op;
4065 }
4066 if (!(s->cpuid_ext_features & sse_op_table7[b].ext_mask))
4067 goto illegal_op;
4068
4069 if (sse_fn_eppi == SSE_SPECIAL) {
4070 ot = mo_64_32(s->dflag);
4071 rm = (modrm & 7) | REX_B(s);
4072 if (mod != 3)
4073 gen_lea_modrm(env, s, modrm);
4074 reg = ((modrm >> 3) & 7) | rex_r;
4075 val = cpu_ldub_code(env, s->pc++);
4076 switch (b) {
4077 case 0x14: /* pextrb */
4078 tcg_gen_ld8u_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,
4079 xmm_regs[reg].XMM_B(val & 15)));
4080 if (mod == 3) {
4081 gen_op_mov_reg_v(ot, rm, cpu_T[0]);
4082 } else {
4083 tcg_gen_qemu_st_tl(cpu_T[0], cpu_A0,
4084 s->mem_index, MO_UB);
4085 }
4086 break;
4087 case 0x15: /* pextrw */
4088 tcg_gen_ld16u_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,
4089 xmm_regs[reg].XMM_W(val & 7)));
4090 if (mod == 3) {
4091 gen_op_mov_reg_v(ot, rm, cpu_T[0]);
4092 } else {
4093 tcg_gen_qemu_st_tl(cpu_T[0], cpu_A0,
4094 s->mem_index, MO_LEUW);
4095 }
4096 break;
4097 case 0x16:
4098 if (ot == MO_32) { /* pextrd */
4099 tcg_gen_ld_i32(cpu_tmp2_i32, cpu_env,
4100 offsetof(CPUX86State,
4101 xmm_regs[reg].XMM_L(val & 3)));
4102 if (mod == 3) {
4103 tcg_gen_extu_i32_tl(cpu_regs[rm], cpu_tmp2_i32);
4104 } else {
4105 tcg_gen_qemu_st_i32(cpu_tmp2_i32, cpu_A0,
4106 s->mem_index, MO_LEUL);
4107 }
4108 } else { /* pextrq */
4109 #ifdef TARGET_X86_64
4110 tcg_gen_ld_i64(cpu_tmp1_i64, cpu_env,
4111 offsetof(CPUX86State,
4112 xmm_regs[reg].XMM_Q(val & 1)));
4113 if (mod == 3) {
4114 tcg_gen_mov_i64(cpu_regs[rm], cpu_tmp1_i64);
4115 } else {
4116 tcg_gen_qemu_st_i64(cpu_tmp1_i64, cpu_A0,
4117 s->mem_index, MO_LEQ);
4118 }
4119 #else
4120 goto illegal_op;
4121 #endif
4122 }
4123 break;
4124 case 0x17: /* extractps */
4125 tcg_gen_ld32u_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,
4126 xmm_regs[reg].XMM_L(val & 3)));
4127 if (mod == 3) {
4128 gen_op_mov_reg_v(ot, rm, cpu_T[0]);
4129 } else {
4130 tcg_gen_qemu_st_tl(cpu_T[0], cpu_A0,
4131 s->mem_index, MO_LEUL);
4132 }
4133 break;
4134 case 0x20: /* pinsrb */
4135 if (mod == 3) {
4136 gen_op_mov_v_reg(MO_32, cpu_T[0], rm);
4137 } else {
4138 tcg_gen_qemu_ld_tl(cpu_T[0], cpu_A0,
4139 s->mem_index, MO_UB);
4140 }
4141 tcg_gen_st8_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,
4142 xmm_regs[reg].XMM_B(val & 15)));
4143 break;
4144 case 0x21: /* insertps */
4145 if (mod == 3) {
4146 tcg_gen_ld_i32(cpu_tmp2_i32, cpu_env,
4147 offsetof(CPUX86State,xmm_regs[rm]
4148 .XMM_L((val >> 6) & 3)));
4149 } else {
4150 tcg_gen_qemu_ld_i32(cpu_tmp2_i32, cpu_A0,
4151 s->mem_index, MO_LEUL);
4152 }
4153 tcg_gen_st_i32(cpu_tmp2_i32, cpu_env,
4154 offsetof(CPUX86State,xmm_regs[reg]
4155 .XMM_L((val >> 4) & 3)));
4156 if ((val >> 0) & 1)
4157 tcg_gen_st_i32(tcg_const_i32(0 /*float32_zero*/),
4158 cpu_env, offsetof(CPUX86State,
4159 xmm_regs[reg].XMM_L(0)));
4160 if ((val >> 1) & 1)
4161 tcg_gen_st_i32(tcg_const_i32(0 /*float32_zero*/),
4162 cpu_env, offsetof(CPUX86State,
4163 xmm_regs[reg].XMM_L(1)));
4164 if ((val >> 2) & 1)
4165 tcg_gen_st_i32(tcg_const_i32(0 /*float32_zero*/),
4166 cpu_env, offsetof(CPUX86State,
4167 xmm_regs[reg].XMM_L(2)));
4168 if ((val >> 3) & 1)
4169 tcg_gen_st_i32(tcg_const_i32(0 /*float32_zero*/),
4170 cpu_env, offsetof(CPUX86State,
4171 xmm_regs[reg].XMM_L(3)));
4172 break;
4173 case 0x22:
4174 if (ot == MO_32) { /* pinsrd */
4175 if (mod == 3) {
4176 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_regs[rm]);
4177 } else {
4178 tcg_gen_qemu_ld_i32(cpu_tmp2_i32, cpu_A0,
4179 s->mem_index, MO_LEUL);
4180 }
4181 tcg_gen_st_i32(cpu_tmp2_i32, cpu_env,
4182 offsetof(CPUX86State,
4183 xmm_regs[reg].XMM_L(val & 3)));
4184 } else { /* pinsrq */
4185 #ifdef TARGET_X86_64
4186 if (mod == 3) {
4187 gen_op_mov_v_reg(ot, cpu_tmp1_i64, rm);
4188 } else {
4189 tcg_gen_qemu_ld_i64(cpu_tmp1_i64, cpu_A0,
4190 s->mem_index, MO_LEQ);
4191 }
4192 tcg_gen_st_i64(cpu_tmp1_i64, cpu_env,
4193 offsetof(CPUX86State,
4194 xmm_regs[reg].XMM_Q(val & 1)));
4195 #else
4196 goto illegal_op;
4197 #endif
4198 }
4199 break;
4200 }
4201 return;
4202 }
4203
4204 if (b1) {
4205 op1_offset = offsetof(CPUX86State,xmm_regs[reg]);
4206 if (mod == 3) {
4207 op2_offset = offsetof(CPUX86State,xmm_regs[rm | REX_B(s)]);
4208 } else {
4209 op2_offset = offsetof(CPUX86State,xmm_t0);
4210 gen_lea_modrm(env, s, modrm);
4211 gen_ldo_env_A0(s, op2_offset);
4212 }
4213 } else {
4214 op1_offset = offsetof(CPUX86State,fpregs[reg].mmx);
4215 if (mod == 3) {
4216 op2_offset = offsetof(CPUX86State,fpregs[rm].mmx);
4217 } else {
4218 op2_offset = offsetof(CPUX86State,mmx_t0);
4219 gen_lea_modrm(env, s, modrm);
4220 gen_ldq_env_A0(s, op2_offset);
4221 }
4222 }
4223 val = cpu_ldub_code(env, s->pc++);
4224
4225 if ((b & 0xfc) == 0x60) { /* pcmpXstrX */
4226 set_cc_op(s, CC_OP_EFLAGS);
4227
4228 if (s->dflag == MO_64) {
4229 /* The helper must use entire 64-bit gp registers */
4230 val |= 1 << 8;
4231 }
4232 }
4233
4234 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, op1_offset);
4235 tcg_gen_addi_ptr(cpu_ptr1, cpu_env, op2_offset);
4236 sse_fn_eppi(cpu_env, cpu_ptr0, cpu_ptr1, tcg_const_i32(val));
4237 break;
4238
4239 case 0x33a:
4240 /* Various integer extensions at 0f 3a f[0-f]. */
4241 b = modrm | (b1 << 8);
4242 modrm = cpu_ldub_code(env, s->pc++);
4243 reg = ((modrm >> 3) & 7) | rex_r;
4244
4245 switch (b) {
4246 case 0x3f0: /* rorx Gy,Ey, Ib */
4247 if (!(s->cpuid_7_0_ebx_features & CPUID_7_0_EBX_BMI2)
4248 || !(s->prefix & PREFIX_VEX)
4249 || s->vex_l != 0) {
4250 goto illegal_op;
4251 }
4252 ot = mo_64_32(s->dflag);
4253 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
4254 b = cpu_ldub_code(env, s->pc++);
4255 if (ot == MO_64) {
4256 tcg_gen_rotri_tl(cpu_T[0], cpu_T[0], b & 63);
4257 } else {
4258 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
4259 tcg_gen_rotri_i32(cpu_tmp2_i32, cpu_tmp2_i32, b & 31);
4260 tcg_gen_extu_i32_tl(cpu_T[0], cpu_tmp2_i32);
4261 }
4262 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
4263 break;
4264
4265 default:
4266 goto illegal_op;
4267 }
4268 break;
4269
4270 default:
4271 goto illegal_op;
4272 }
4273 } else {
4274 /* generic MMX or SSE operation */
4275 switch(b) {
4276 case 0x70: /* pshufx insn */
4277 case 0xc6: /* pshufx insn */
4278 case 0xc2: /* compare insns */
4279 s->rip_offset = 1;
4280 break;
4281 default:
4282 break;
4283 }
4284 if (is_xmm) {
4285 op1_offset = offsetof(CPUX86State,xmm_regs[reg]);
4286 if (mod != 3) {
4287 int sz = 4;
4288
4289 gen_lea_modrm(env, s, modrm);
4290 op2_offset = offsetof(CPUX86State,xmm_t0);
4291
4292 switch (b) {
4293 case 0x50 ... 0x5a:
4294 case 0x5c ... 0x5f:
4295 case 0xc2:
4296 /* Most sse scalar operations. */
4297 if (b1 == 2) {
4298 sz = 2;
4299 } else if (b1 == 3) {
4300 sz = 3;
4301 }
4302 break;
4303
4304 case 0x2e: /* ucomis[sd] */
4305 case 0x2f: /* comis[sd] */
4306 if (b1 == 0) {
4307 sz = 2;
4308 } else {
4309 sz = 3;
4310 }
4311 break;
4312 }
4313
4314 switch (sz) {
4315 case 2:
4316 /* 32 bit access */
4317 gen_op_ld_v(s, MO_32, cpu_T[0], cpu_A0);
4318 tcg_gen_st32_tl(cpu_T[0], cpu_env,
4319 offsetof(CPUX86State,xmm_t0.XMM_L(0)));
4320 break;
4321 case 3:
4322 /* 64 bit access */
4323 gen_ldq_env_A0(s, offsetof(CPUX86State, xmm_t0.XMM_D(0)));
4324 break;
4325 default:
4326 /* 128 bit access */
4327 gen_ldo_env_A0(s, op2_offset);
4328 break;
4329 }
4330 } else {
4331 rm = (modrm & 7) | REX_B(s);
4332 op2_offset = offsetof(CPUX86State,xmm_regs[rm]);
4333 }
4334 } else {
4335 op1_offset = offsetof(CPUX86State,fpregs[reg].mmx);
4336 if (mod != 3) {
4337 gen_lea_modrm(env, s, modrm);
4338 op2_offset = offsetof(CPUX86State,mmx_t0);
4339 gen_ldq_env_A0(s, op2_offset);
4340 } else {
4341 rm = (modrm & 7);
4342 op2_offset = offsetof(CPUX86State,fpregs[rm].mmx);
4343 }
4344 }
4345 switch(b) {
4346 case 0x0f: /* 3DNow! data insns */
4347 if (!(s->cpuid_ext2_features & CPUID_EXT2_3DNOW))
4348 goto illegal_op;
4349 val = cpu_ldub_code(env, s->pc++);
4350 sse_fn_epp = sse_op_table5[val];
4351 if (!sse_fn_epp) {
4352 goto illegal_op;
4353 }
4354 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, op1_offset);
4355 tcg_gen_addi_ptr(cpu_ptr1, cpu_env, op2_offset);
4356 sse_fn_epp(cpu_env, cpu_ptr0, cpu_ptr1);
4357 break;
4358 case 0x70: /* pshufx insn */
4359 case 0xc6: /* pshufx insn */
4360 val = cpu_ldub_code(env, s->pc++);
4361 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, op1_offset);
4362 tcg_gen_addi_ptr(cpu_ptr1, cpu_env, op2_offset);
4363 /* XXX: introduce a new table? */
4364 sse_fn_ppi = (SSEFunc_0_ppi)sse_fn_epp;
4365 sse_fn_ppi(cpu_ptr0, cpu_ptr1, tcg_const_i32(val));
4366 break;
4367 case 0xc2:
4368 /* compare insns */
4369 val = cpu_ldub_code(env, s->pc++);
4370 if (val >= 8)
4371 goto illegal_op;
4372 sse_fn_epp = sse_op_table4[val][b1];
4373
4374 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, op1_offset);
4375 tcg_gen_addi_ptr(cpu_ptr1, cpu_env, op2_offset);
4376 sse_fn_epp(cpu_env, cpu_ptr0, cpu_ptr1);
4377 break;
4378 case 0xf7:
4379 /* maskmov : we must prepare A0 */
4380 if (mod != 3)
4381 goto illegal_op;
4382 tcg_gen_mov_tl(cpu_A0, cpu_regs[R_EDI]);
4383 gen_extu(s->aflag, cpu_A0);
4384 gen_add_A0_ds_seg(s);
4385
4386 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, op1_offset);
4387 tcg_gen_addi_ptr(cpu_ptr1, cpu_env, op2_offset);
4388 /* XXX: introduce a new table? */
4389 sse_fn_eppt = (SSEFunc_0_eppt)sse_fn_epp;
4390 sse_fn_eppt(cpu_env, cpu_ptr0, cpu_ptr1, cpu_A0);
4391 break;
4392 default:
4393 tcg_gen_addi_ptr(cpu_ptr0, cpu_env, op1_offset);
4394 tcg_gen_addi_ptr(cpu_ptr1, cpu_env, op2_offset);
4395 sse_fn_epp(cpu_env, cpu_ptr0, cpu_ptr1);
4396 break;
4397 }
4398 if (b == 0x2e || b == 0x2f) {
4399 set_cc_op(s, CC_OP_EFLAGS);
4400 }
4401 }
4402 }
4403
4404 /* convert one instruction. s->is_jmp is set if the translation must
4405 be stopped. Return the next pc value */
4406 static target_ulong disas_insn(CPUX86State *env, DisasContext *s,
4407 target_ulong pc_start)
4408 {
4409 int b, prefixes;
4410 int shift;
4411 TCGMemOp ot, aflag, dflag;
4412 int modrm, reg, rm, mod, op, opreg, val;
4413 target_ulong next_eip, tval;
4414 int rex_w, rex_r;
4415
4416 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP | CPU_LOG_TB_OP_OPT))) {
4417 tcg_gen_debug_insn_start(pc_start);
4418 }
4419 s->pc = pc_start;
4420 prefixes = 0;
4421 s->override = -1;
4422 rex_w = -1;
4423 rex_r = 0;
4424 #ifdef TARGET_X86_64
4425 s->rex_x = 0;
4426 s->rex_b = 0;
4427 x86_64_hregs = 0;
4428 #endif
4429 s->rip_offset = 0; /* for relative ip address */
4430 s->vex_l = 0;
4431 s->vex_v = 0;
4432 next_byte:
4433 b = cpu_ldub_code(env, s->pc);
4434 s->pc++;
4435 /* Collect prefixes. */
4436 switch (b) {
4437 case 0xf3:
4438 prefixes |= PREFIX_REPZ;
4439 goto next_byte;
4440 case 0xf2:
4441 prefixes |= PREFIX_REPNZ;
4442 goto next_byte;
4443 case 0xf0:
4444 prefixes |= PREFIX_LOCK;
4445 goto next_byte;
4446 case 0x2e:
4447 s->override = R_CS;
4448 goto next_byte;
4449 case 0x36:
4450 s->override = R_SS;
4451 goto next_byte;
4452 case 0x3e:
4453 s->override = R_DS;
4454 goto next_byte;
4455 case 0x26:
4456 s->override = R_ES;
4457 goto next_byte;
4458 case 0x64:
4459 s->override = R_FS;
4460 goto next_byte;
4461 case 0x65:
4462 s->override = R_GS;
4463 goto next_byte;
4464 case 0x66:
4465 prefixes |= PREFIX_DATA;
4466 goto next_byte;
4467 case 0x67:
4468 prefixes |= PREFIX_ADR;
4469 goto next_byte;
4470 #ifdef TARGET_X86_64
4471 case 0x40 ... 0x4f:
4472 if (CODE64(s)) {
4473 /* REX prefix */
4474 rex_w = (b >> 3) & 1;
4475 rex_r = (b & 0x4) << 1;
4476 s->rex_x = (b & 0x2) << 2;
4477 REX_B(s) = (b & 0x1) << 3;
4478 x86_64_hregs = 1; /* select uniform byte register addressing */
4479 goto next_byte;
4480 }
4481 break;
4482 #endif
4483 case 0xc5: /* 2-byte VEX */
4484 case 0xc4: /* 3-byte VEX */
4485 /* VEX prefixes cannot be used except in 32-bit mode.
4486 Otherwise the instruction is LES or LDS. */
4487 if (s->code32 && !s->vm86) {
4488 static const int pp_prefix[4] = {
4489 0, PREFIX_DATA, PREFIX_REPZ, PREFIX_REPNZ
4490 };
4491 int vex3, vex2 = cpu_ldub_code(env, s->pc);
4492
4493 if (!CODE64(s) && (vex2 & 0xc0) != 0xc0) {
4494 /* 4.1.4.6: In 32-bit mode, bits [7:6] must be 11b,
4495 otherwise the instruction is LES or LDS. */
4496 break;
4497 }
4498 s->pc++;
4499
4500 /* 4.1.1-4.1.3: No preceding lock, 66, f2, f3, or rex prefixes. */
4501 if (prefixes & (PREFIX_REPZ | PREFIX_REPNZ
4502 | PREFIX_LOCK | PREFIX_DATA)) {
4503 goto illegal_op;
4504 }
4505 #ifdef TARGET_X86_64
4506 if (x86_64_hregs) {
4507 goto illegal_op;
4508 }
4509 #endif
4510 rex_r = (~vex2 >> 4) & 8;
4511 if (b == 0xc5) {
4512 vex3 = vex2;
4513 b = cpu_ldub_code(env, s->pc++);
4514 } else {
4515 #ifdef TARGET_X86_64
4516 s->rex_x = (~vex2 >> 3) & 8;
4517 s->rex_b = (~vex2 >> 2) & 8;
4518 #endif
4519 vex3 = cpu_ldub_code(env, s->pc++);
4520 rex_w = (vex3 >> 7) & 1;
4521 switch (vex2 & 0x1f) {
4522 case 0x01: /* Implied 0f leading opcode bytes. */
4523 b = cpu_ldub_code(env, s->pc++) | 0x100;
4524 break;
4525 case 0x02: /* Implied 0f 38 leading opcode bytes. */
4526 b = 0x138;
4527 break;
4528 case 0x03: /* Implied 0f 3a leading opcode bytes. */
4529 b = 0x13a;
4530 break;
4531 default: /* Reserved for future use. */
4532 goto illegal_op;
4533 }
4534 }
4535 s->vex_v = (~vex3 >> 3) & 0xf;
4536 s->vex_l = (vex3 >> 2) & 1;
4537 prefixes |= pp_prefix[vex3 & 3] | PREFIX_VEX;
4538 }
4539 break;
4540 }
4541
4542 /* Post-process prefixes. */
4543 if (CODE64(s)) {
4544 /* In 64-bit mode, the default data size is 32-bit. Select 64-bit
4545 data with rex_w, and 16-bit data with 0x66; rex_w takes precedence
4546 over 0x66 if both are present. */
4547 dflag = (rex_w > 0 ? MO_64 : prefixes & PREFIX_DATA ? MO_16 : MO_32);
4548 /* In 64-bit mode, 0x67 selects 32-bit addressing. */
4549 aflag = (prefixes & PREFIX_ADR ? MO_32 : MO_64);
4550 } else {
4551 /* In 16/32-bit mode, 0x66 selects the opposite data size. */
4552 if (s->code32 ^ ((prefixes & PREFIX_DATA) != 0)) {
4553 dflag = MO_32;
4554 } else {
4555 dflag = MO_16;
4556 }
4557 /* In 16/32-bit mode, 0x67 selects the opposite addressing. */
4558 if (s->code32 ^ ((prefixes & PREFIX_ADR) != 0)) {
4559 aflag = MO_32;
4560 } else {
4561 aflag = MO_16;
4562 }
4563 }
4564
4565 s->prefix = prefixes;
4566 s->aflag = aflag;
4567 s->dflag = dflag;
4568
4569 /* lock generation */
4570 if (prefixes & PREFIX_LOCK)
4571 gen_helper_lock();
4572
4573 /* now check op code */
4574 reswitch:
4575 switch(b) {
4576 case 0x0f:
4577 /**************************/
4578 /* extended op code */
4579 b = cpu_ldub_code(env, s->pc++) | 0x100;
4580 goto reswitch;
4581
4582 /**************************/
4583 /* arith & logic */
4584 case 0x00 ... 0x05:
4585 case 0x08 ... 0x0d:
4586 case 0x10 ... 0x15:
4587 case 0x18 ... 0x1d:
4588 case 0x20 ... 0x25:
4589 case 0x28 ... 0x2d:
4590 case 0x30 ... 0x35:
4591 case 0x38 ... 0x3d:
4592 {
4593 int op, f, val;
4594 op = (b >> 3) & 7;
4595 f = (b >> 1) & 3;
4596
4597 ot = mo_b_d(b, dflag);
4598
4599 switch(f) {
4600 case 0: /* OP Ev, Gv */
4601 modrm = cpu_ldub_code(env, s->pc++);
4602 reg = ((modrm >> 3) & 7) | rex_r;
4603 mod = (modrm >> 6) & 3;
4604 rm = (modrm & 7) | REX_B(s);
4605 if (mod != 3) {
4606 gen_lea_modrm(env, s, modrm);
4607 opreg = OR_TMP0;
4608 } else if (op == OP_XORL && rm == reg) {
4609 xor_zero:
4610 /* xor reg, reg optimisation */
4611 set_cc_op(s, CC_OP_CLR);
4612 tcg_gen_movi_tl(cpu_T[0], 0);
4613 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
4614 break;
4615 } else {
4616 opreg = rm;
4617 }
4618 gen_op_mov_v_reg(ot, cpu_T[1], reg);
4619 gen_op(s, op, ot, opreg);
4620 break;
4621 case 1: /* OP Gv, Ev */
4622 modrm = cpu_ldub_code(env, s->pc++);
4623 mod = (modrm >> 6) & 3;
4624 reg = ((modrm >> 3) & 7) | rex_r;
4625 rm = (modrm & 7) | REX_B(s);
4626 if (mod != 3) {
4627 gen_lea_modrm(env, s, modrm);
4628 gen_op_ld_v(s, ot, cpu_T[1], cpu_A0);
4629 } else if (op == OP_XORL && rm == reg) {
4630 goto xor_zero;
4631 } else {
4632 gen_op_mov_v_reg(ot, cpu_T[1], rm);
4633 }
4634 gen_op(s, op, ot, reg);
4635 break;
4636 case 2: /* OP A, Iv */
4637 val = insn_get(env, s, ot);
4638 tcg_gen_movi_tl(cpu_T[1], val);
4639 gen_op(s, op, ot, OR_EAX);
4640 break;
4641 }
4642 }
4643 break;
4644
4645 case 0x82:
4646 if (CODE64(s))
4647 goto illegal_op;
4648 case 0x80: /* GRP1 */
4649 case 0x81:
4650 case 0x83:
4651 {
4652 int val;
4653
4654 ot = mo_b_d(b, dflag);
4655
4656 modrm = cpu_ldub_code(env, s->pc++);
4657 mod = (modrm >> 6) & 3;
4658 rm = (modrm & 7) | REX_B(s);
4659 op = (modrm >> 3) & 7;
4660
4661 if (mod != 3) {
4662 if (b == 0x83)
4663 s->rip_offset = 1;
4664 else
4665 s->rip_offset = insn_const_size(ot);
4666 gen_lea_modrm(env, s, modrm);
4667 opreg = OR_TMP0;
4668 } else {
4669 opreg = rm;
4670 }
4671
4672 switch(b) {
4673 default:
4674 case 0x80:
4675 case 0x81:
4676 case 0x82:
4677 val = insn_get(env, s, ot);
4678 break;
4679 case 0x83:
4680 val = (int8_t)insn_get(env, s, MO_8);
4681 break;
4682 }
4683 tcg_gen_movi_tl(cpu_T[1], val);
4684 gen_op(s, op, ot, opreg);
4685 }
4686 break;
4687
4688 /**************************/
4689 /* inc, dec, and other misc arith */
4690 case 0x40 ... 0x47: /* inc Gv */
4691 ot = dflag;
4692 gen_inc(s, ot, OR_EAX + (b & 7), 1);
4693 break;
4694 case 0x48 ... 0x4f: /* dec Gv */
4695 ot = dflag;
4696 gen_inc(s, ot, OR_EAX + (b & 7), -1);
4697 break;
4698 case 0xf6: /* GRP3 */
4699 case 0xf7:
4700 ot = mo_b_d(b, dflag);
4701
4702 modrm = cpu_ldub_code(env, s->pc++);
4703 mod = (modrm >> 6) & 3;
4704 rm = (modrm & 7) | REX_B(s);
4705 op = (modrm >> 3) & 7;
4706 if (mod != 3) {
4707 if (op == 0)
4708 s->rip_offset = insn_const_size(ot);
4709 gen_lea_modrm(env, s, modrm);
4710 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
4711 } else {
4712 gen_op_mov_v_reg(ot, cpu_T[0], rm);
4713 }
4714
4715 switch(op) {
4716 case 0: /* test */
4717 val = insn_get(env, s, ot);
4718 tcg_gen_movi_tl(cpu_T[1], val);
4719 gen_op_testl_T0_T1_cc();
4720 set_cc_op(s, CC_OP_LOGICB + ot);
4721 break;
4722 case 2: /* not */
4723 tcg_gen_not_tl(cpu_T[0], cpu_T[0]);
4724 if (mod != 3) {
4725 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
4726 } else {
4727 gen_op_mov_reg_v(ot, rm, cpu_T[0]);
4728 }
4729 break;
4730 case 3: /* neg */
4731 tcg_gen_neg_tl(cpu_T[0], cpu_T[0]);
4732 if (mod != 3) {
4733 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
4734 } else {
4735 gen_op_mov_reg_v(ot, rm, cpu_T[0]);
4736 }
4737 gen_op_update_neg_cc();
4738 set_cc_op(s, CC_OP_SUBB + ot);
4739 break;
4740 case 4: /* mul */
4741 switch(ot) {
4742 case MO_8:
4743 gen_op_mov_v_reg(MO_8, cpu_T[1], R_EAX);
4744 tcg_gen_ext8u_tl(cpu_T[0], cpu_T[0]);
4745 tcg_gen_ext8u_tl(cpu_T[1], cpu_T[1]);
4746 /* XXX: use 32 bit mul which could be faster */
4747 tcg_gen_mul_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
4748 gen_op_mov_reg_v(MO_16, R_EAX, cpu_T[0]);
4749 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
4750 tcg_gen_andi_tl(cpu_cc_src, cpu_T[0], 0xff00);
4751 set_cc_op(s, CC_OP_MULB);
4752 break;
4753 case MO_16:
4754 gen_op_mov_v_reg(MO_16, cpu_T[1], R_EAX);
4755 tcg_gen_ext16u_tl(cpu_T[0], cpu_T[0]);
4756 tcg_gen_ext16u_tl(cpu_T[1], cpu_T[1]);
4757 /* XXX: use 32 bit mul which could be faster */
4758 tcg_gen_mul_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
4759 gen_op_mov_reg_v(MO_16, R_EAX, cpu_T[0]);
4760 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
4761 tcg_gen_shri_tl(cpu_T[0], cpu_T[0], 16);
4762 gen_op_mov_reg_v(MO_16, R_EDX, cpu_T[0]);
4763 tcg_gen_mov_tl(cpu_cc_src, cpu_T[0]);
4764 set_cc_op(s, CC_OP_MULW);
4765 break;
4766 default:
4767 case MO_32:
4768 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
4769 tcg_gen_trunc_tl_i32(cpu_tmp3_i32, cpu_regs[R_EAX]);
4770 tcg_gen_mulu2_i32(cpu_tmp2_i32, cpu_tmp3_i32,
4771 cpu_tmp2_i32, cpu_tmp3_i32);
4772 tcg_gen_extu_i32_tl(cpu_regs[R_EAX], cpu_tmp2_i32);
4773 tcg_gen_extu_i32_tl(cpu_regs[R_EDX], cpu_tmp3_i32);
4774 tcg_gen_mov_tl(cpu_cc_dst, cpu_regs[R_EAX]);
4775 tcg_gen_mov_tl(cpu_cc_src, cpu_regs[R_EDX]);
4776 set_cc_op(s, CC_OP_MULL);
4777 break;
4778 #ifdef TARGET_X86_64
4779 case MO_64:
4780 tcg_gen_mulu2_i64(cpu_regs[R_EAX], cpu_regs[R_EDX],
4781 cpu_T[0], cpu_regs[R_EAX]);
4782 tcg_gen_mov_tl(cpu_cc_dst, cpu_regs[R_EAX]);
4783 tcg_gen_mov_tl(cpu_cc_src, cpu_regs[R_EDX]);
4784 set_cc_op(s, CC_OP_MULQ);
4785 break;
4786 #endif
4787 }
4788 break;
4789 case 5: /* imul */
4790 switch(ot) {
4791 case MO_8:
4792 gen_op_mov_v_reg(MO_8, cpu_T[1], R_EAX);
4793 tcg_gen_ext8s_tl(cpu_T[0], cpu_T[0]);
4794 tcg_gen_ext8s_tl(cpu_T[1], cpu_T[1]);
4795 /* XXX: use 32 bit mul which could be faster */
4796 tcg_gen_mul_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
4797 gen_op_mov_reg_v(MO_16, R_EAX, cpu_T[0]);
4798 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
4799 tcg_gen_ext8s_tl(cpu_tmp0, cpu_T[0]);
4800 tcg_gen_sub_tl(cpu_cc_src, cpu_T[0], cpu_tmp0);
4801 set_cc_op(s, CC_OP_MULB);
4802 break;
4803 case MO_16:
4804 gen_op_mov_v_reg(MO_16, cpu_T[1], R_EAX);
4805 tcg_gen_ext16s_tl(cpu_T[0], cpu_T[0]);
4806 tcg_gen_ext16s_tl(cpu_T[1], cpu_T[1]);
4807 /* XXX: use 32 bit mul which could be faster */
4808 tcg_gen_mul_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
4809 gen_op_mov_reg_v(MO_16, R_EAX, cpu_T[0]);
4810 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
4811 tcg_gen_ext16s_tl(cpu_tmp0, cpu_T[0]);
4812 tcg_gen_sub_tl(cpu_cc_src, cpu_T[0], cpu_tmp0);
4813 tcg_gen_shri_tl(cpu_T[0], cpu_T[0], 16);
4814 gen_op_mov_reg_v(MO_16, R_EDX, cpu_T[0]);
4815 set_cc_op(s, CC_OP_MULW);
4816 break;
4817 default:
4818 case MO_32:
4819 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
4820 tcg_gen_trunc_tl_i32(cpu_tmp3_i32, cpu_regs[R_EAX]);
4821 tcg_gen_muls2_i32(cpu_tmp2_i32, cpu_tmp3_i32,
4822 cpu_tmp2_i32, cpu_tmp3_i32);
4823 tcg_gen_extu_i32_tl(cpu_regs[R_EAX], cpu_tmp2_i32);
4824 tcg_gen_extu_i32_tl(cpu_regs[R_EDX], cpu_tmp3_i32);
4825 tcg_gen_sari_i32(cpu_tmp2_i32, cpu_tmp2_i32, 31);
4826 tcg_gen_mov_tl(cpu_cc_dst, cpu_regs[R_EAX]);
4827 tcg_gen_sub_i32(cpu_tmp2_i32, cpu_tmp2_i32, cpu_tmp3_i32);
4828 tcg_gen_extu_i32_tl(cpu_cc_src, cpu_tmp2_i32);
4829 set_cc_op(s, CC_OP_MULL);
4830 break;
4831 #ifdef TARGET_X86_64
4832 case MO_64:
4833 tcg_gen_muls2_i64(cpu_regs[R_EAX], cpu_regs[R_EDX],
4834 cpu_T[0], cpu_regs[R_EAX]);
4835 tcg_gen_mov_tl(cpu_cc_dst, cpu_regs[R_EAX]);
4836 tcg_gen_sari_tl(cpu_cc_src, cpu_regs[R_EAX], 63);
4837 tcg_gen_sub_tl(cpu_cc_src, cpu_cc_src, cpu_regs[R_EDX]);
4838 set_cc_op(s, CC_OP_MULQ);
4839 break;
4840 #endif
4841 }
4842 break;
4843 case 6: /* div */
4844 switch(ot) {
4845 case MO_8:
4846 gen_jmp_im(pc_start - s->cs_base);
4847 gen_helper_divb_AL(cpu_env, cpu_T[0]);
4848 break;
4849 case MO_16:
4850 gen_jmp_im(pc_start - s->cs_base);
4851 gen_helper_divw_AX(cpu_env, cpu_T[0]);
4852 break;
4853 default:
4854 case MO_32:
4855 gen_jmp_im(pc_start - s->cs_base);
4856 gen_helper_divl_EAX(cpu_env, cpu_T[0]);
4857 break;
4858 #ifdef TARGET_X86_64
4859 case MO_64:
4860 gen_jmp_im(pc_start - s->cs_base);
4861 gen_helper_divq_EAX(cpu_env, cpu_T[0]);
4862 break;
4863 #endif
4864 }
4865 break;
4866 case 7: /* idiv */
4867 switch(ot) {
4868 case MO_8:
4869 gen_jmp_im(pc_start - s->cs_base);
4870 gen_helper_idivb_AL(cpu_env, cpu_T[0]);
4871 break;
4872 case MO_16:
4873 gen_jmp_im(pc_start - s->cs_base);
4874 gen_helper_idivw_AX(cpu_env, cpu_T[0]);
4875 break;
4876 default:
4877 case MO_32:
4878 gen_jmp_im(pc_start - s->cs_base);
4879 gen_helper_idivl_EAX(cpu_env, cpu_T[0]);
4880 break;
4881 #ifdef TARGET_X86_64
4882 case MO_64:
4883 gen_jmp_im(pc_start - s->cs_base);
4884 gen_helper_idivq_EAX(cpu_env, cpu_T[0]);
4885 break;
4886 #endif
4887 }
4888 break;
4889 default:
4890 goto illegal_op;
4891 }
4892 break;
4893
4894 case 0xfe: /* GRP4 */
4895 case 0xff: /* GRP5 */
4896 ot = mo_b_d(b, dflag);
4897
4898 modrm = cpu_ldub_code(env, s->pc++);
4899 mod = (modrm >> 6) & 3;
4900 rm = (modrm & 7) | REX_B(s);
4901 op = (modrm >> 3) & 7;
4902 if (op >= 2 && b == 0xfe) {
4903 goto illegal_op;
4904 }
4905 if (CODE64(s)) {
4906 if (op == 2 || op == 4) {
4907 /* operand size for jumps is 64 bit */
4908 ot = MO_64;
4909 } else if (op == 3 || op == 5) {
4910 ot = dflag != MO_16 ? MO_32 + (rex_w == 1) : MO_16;
4911 } else if (op == 6) {
4912 /* default push size is 64 bit */
4913 ot = mo_pushpop(s, dflag);
4914 }
4915 }
4916 if (mod != 3) {
4917 gen_lea_modrm(env, s, modrm);
4918 if (op >= 2 && op != 3 && op != 5)
4919 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
4920 } else {
4921 gen_op_mov_v_reg(ot, cpu_T[0], rm);
4922 }
4923
4924 switch(op) {
4925 case 0: /* inc Ev */
4926 if (mod != 3)
4927 opreg = OR_TMP0;
4928 else
4929 opreg = rm;
4930 gen_inc(s, ot, opreg, 1);
4931 break;
4932 case 1: /* dec Ev */
4933 if (mod != 3)
4934 opreg = OR_TMP0;
4935 else
4936 opreg = rm;
4937 gen_inc(s, ot, opreg, -1);
4938 break;
4939 case 2: /* call Ev */
4940 /* XXX: optimize if memory (no 'and' is necessary) */
4941 if (dflag == MO_16) {
4942 tcg_gen_ext16u_tl(cpu_T[0], cpu_T[0]);
4943 }
4944 next_eip = s->pc - s->cs_base;
4945 tcg_gen_movi_tl(cpu_T[1], next_eip);
4946 gen_push_v(s, cpu_T[1]);
4947 gen_op_jmp_v(cpu_T[0]);
4948 gen_eob(s);
4949 break;
4950 case 3: /* lcall Ev */
4951 gen_op_ld_v(s, ot, cpu_T[1], cpu_A0);
4952 gen_add_A0_im(s, 1 << ot);
4953 gen_op_ld_v(s, MO_16, cpu_T[0], cpu_A0);
4954 do_lcall:
4955 if (s->pe && !s->vm86) {
4956 gen_update_cc_op(s);
4957 gen_jmp_im(pc_start - s->cs_base);
4958 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
4959 gen_helper_lcall_protected(cpu_env, cpu_tmp2_i32, cpu_T[1],
4960 tcg_const_i32(dflag - 1),
4961 tcg_const_i32(s->pc - pc_start));
4962 } else {
4963 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
4964 gen_helper_lcall_real(cpu_env, cpu_tmp2_i32, cpu_T[1],
4965 tcg_const_i32(dflag - 1),
4966 tcg_const_i32(s->pc - s->cs_base));
4967 }
4968 gen_eob(s);
4969 break;
4970 case 4: /* jmp Ev */
4971 if (dflag == MO_16) {
4972 tcg_gen_ext16u_tl(cpu_T[0], cpu_T[0]);
4973 }
4974 gen_op_jmp_v(cpu_T[0]);
4975 gen_eob(s);
4976 break;
4977 case 5: /* ljmp Ev */
4978 gen_op_ld_v(s, ot, cpu_T[1], cpu_A0);
4979 gen_add_A0_im(s, 1 << ot);
4980 gen_op_ld_v(s, MO_16, cpu_T[0], cpu_A0);
4981 do_ljmp:
4982 if (s->pe && !s->vm86) {
4983 gen_update_cc_op(s);
4984 gen_jmp_im(pc_start - s->cs_base);
4985 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
4986 gen_helper_ljmp_protected(cpu_env, cpu_tmp2_i32, cpu_T[1],
4987 tcg_const_i32(s->pc - pc_start));
4988 } else {
4989 gen_op_movl_seg_T0_vm(R_CS);
4990 gen_op_jmp_v(cpu_T[1]);
4991 }
4992 gen_eob(s);
4993 break;
4994 case 6: /* push Ev */
4995 gen_push_v(s, cpu_T[0]);
4996 break;
4997 default:
4998 goto illegal_op;
4999 }
5000 break;
5001
5002 case 0x84: /* test Ev, Gv */
5003 case 0x85:
5004 ot = mo_b_d(b, dflag);
5005
5006 modrm = cpu_ldub_code(env, s->pc++);
5007 reg = ((modrm >> 3) & 7) | rex_r;
5008
5009 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
5010 gen_op_mov_v_reg(ot, cpu_T[1], reg);
5011 gen_op_testl_T0_T1_cc();
5012 set_cc_op(s, CC_OP_LOGICB + ot);
5013 break;
5014
5015 case 0xa8: /* test eAX, Iv */
5016 case 0xa9:
5017 ot = mo_b_d(b, dflag);
5018 val = insn_get(env, s, ot);
5019
5020 gen_op_mov_v_reg(ot, cpu_T[0], OR_EAX);
5021 tcg_gen_movi_tl(cpu_T[1], val);
5022 gen_op_testl_T0_T1_cc();
5023 set_cc_op(s, CC_OP_LOGICB + ot);
5024 break;
5025
5026 case 0x98: /* CWDE/CBW */
5027 switch (dflag) {
5028 #ifdef TARGET_X86_64
5029 case MO_64:
5030 gen_op_mov_v_reg(MO_32, cpu_T[0], R_EAX);
5031 tcg_gen_ext32s_tl(cpu_T[0], cpu_T[0]);
5032 gen_op_mov_reg_v(MO_64, R_EAX, cpu_T[0]);
5033 break;
5034 #endif
5035 case MO_32:
5036 gen_op_mov_v_reg(MO_16, cpu_T[0], R_EAX);
5037 tcg_gen_ext16s_tl(cpu_T[0], cpu_T[0]);
5038 gen_op_mov_reg_v(MO_32, R_EAX, cpu_T[0]);
5039 break;
5040 case MO_16:
5041 gen_op_mov_v_reg(MO_8, cpu_T[0], R_EAX);
5042 tcg_gen_ext8s_tl(cpu_T[0], cpu_T[0]);
5043 gen_op_mov_reg_v(MO_16, R_EAX, cpu_T[0]);
5044 break;
5045 default:
5046 tcg_abort();
5047 }
5048 break;
5049 case 0x99: /* CDQ/CWD */
5050 switch (dflag) {
5051 #ifdef TARGET_X86_64
5052 case MO_64:
5053 gen_op_mov_v_reg(MO_64, cpu_T[0], R_EAX);
5054 tcg_gen_sari_tl(cpu_T[0], cpu_T[0], 63);
5055 gen_op_mov_reg_v(MO_64, R_EDX, cpu_T[0]);
5056 break;
5057 #endif
5058 case MO_32:
5059 gen_op_mov_v_reg(MO_32, cpu_T[0], R_EAX);
5060 tcg_gen_ext32s_tl(cpu_T[0], cpu_T[0]);
5061 tcg_gen_sari_tl(cpu_T[0], cpu_T[0], 31);
5062 gen_op_mov_reg_v(MO_32, R_EDX, cpu_T[0]);
5063 break;
5064 case MO_16:
5065 gen_op_mov_v_reg(MO_16, cpu_T[0], R_EAX);
5066 tcg_gen_ext16s_tl(cpu_T[0], cpu_T[0]);
5067 tcg_gen_sari_tl(cpu_T[0], cpu_T[0], 15);
5068 gen_op_mov_reg_v(MO_16, R_EDX, cpu_T[0]);
5069 break;
5070 default:
5071 tcg_abort();
5072 }
5073 break;
5074 case 0x1af: /* imul Gv, Ev */
5075 case 0x69: /* imul Gv, Ev, I */
5076 case 0x6b:
5077 ot = dflag;
5078 modrm = cpu_ldub_code(env, s->pc++);
5079 reg = ((modrm >> 3) & 7) | rex_r;
5080 if (b == 0x69)
5081 s->rip_offset = insn_const_size(ot);
5082 else if (b == 0x6b)
5083 s->rip_offset = 1;
5084 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
5085 if (b == 0x69) {
5086 val = insn_get(env, s, ot);
5087 tcg_gen_movi_tl(cpu_T[1], val);
5088 } else if (b == 0x6b) {
5089 val = (int8_t)insn_get(env, s, MO_8);
5090 tcg_gen_movi_tl(cpu_T[1], val);
5091 } else {
5092 gen_op_mov_v_reg(ot, cpu_T[1], reg);
5093 }
5094 switch (ot) {
5095 #ifdef TARGET_X86_64
5096 case MO_64:
5097 tcg_gen_muls2_i64(cpu_regs[reg], cpu_T[1], cpu_T[0], cpu_T[1]);
5098 tcg_gen_mov_tl(cpu_cc_dst, cpu_regs[reg]);
5099 tcg_gen_sari_tl(cpu_cc_src, cpu_cc_dst, 63);
5100 tcg_gen_sub_tl(cpu_cc_src, cpu_cc_src, cpu_T[1]);
5101 break;
5102 #endif
5103 case MO_32:
5104 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
5105 tcg_gen_trunc_tl_i32(cpu_tmp3_i32, cpu_T[1]);
5106 tcg_gen_muls2_i32(cpu_tmp2_i32, cpu_tmp3_i32,
5107 cpu_tmp2_i32, cpu_tmp3_i32);
5108 tcg_gen_extu_i32_tl(cpu_regs[reg], cpu_tmp2_i32);
5109 tcg_gen_sari_i32(cpu_tmp2_i32, cpu_tmp2_i32, 31);
5110 tcg_gen_mov_tl(cpu_cc_dst, cpu_regs[reg]);
5111 tcg_gen_sub_i32(cpu_tmp2_i32, cpu_tmp2_i32, cpu_tmp3_i32);
5112 tcg_gen_extu_i32_tl(cpu_cc_src, cpu_tmp2_i32);
5113 break;
5114 default:
5115 tcg_gen_ext16s_tl(cpu_T[0], cpu_T[0]);
5116 tcg_gen_ext16s_tl(cpu_T[1], cpu_T[1]);
5117 /* XXX: use 32 bit mul which could be faster */
5118 tcg_gen_mul_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
5119 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
5120 tcg_gen_ext16s_tl(cpu_tmp0, cpu_T[0]);
5121 tcg_gen_sub_tl(cpu_cc_src, cpu_T[0], cpu_tmp0);
5122 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
5123 break;
5124 }
5125 set_cc_op(s, CC_OP_MULB + ot);
5126 break;
5127 case 0x1c0:
5128 case 0x1c1: /* xadd Ev, Gv */
5129 ot = mo_b_d(b, dflag);
5130 modrm = cpu_ldub_code(env, s->pc++);
5131 reg = ((modrm >> 3) & 7) | rex_r;
5132 mod = (modrm >> 6) & 3;
5133 if (mod == 3) {
5134 rm = (modrm & 7) | REX_B(s);
5135 gen_op_mov_v_reg(ot, cpu_T[0], reg);
5136 gen_op_mov_v_reg(ot, cpu_T[1], rm);
5137 tcg_gen_add_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
5138 gen_op_mov_reg_v(ot, reg, cpu_T[1]);
5139 gen_op_mov_reg_v(ot, rm, cpu_T[0]);
5140 } else {
5141 gen_lea_modrm(env, s, modrm);
5142 gen_op_mov_v_reg(ot, cpu_T[0], reg);
5143 gen_op_ld_v(s, ot, cpu_T[1], cpu_A0);
5144 tcg_gen_add_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
5145 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
5146 gen_op_mov_reg_v(ot, reg, cpu_T[1]);
5147 }
5148 gen_op_update2_cc();
5149 set_cc_op(s, CC_OP_ADDB + ot);
5150 break;
5151 case 0x1b0:
5152 case 0x1b1: /* cmpxchg Ev, Gv */
5153 {
5154 int label1, label2;
5155 TCGv t0, t1, t2, a0;
5156
5157 ot = mo_b_d(b, dflag);
5158 modrm = cpu_ldub_code(env, s->pc++);
5159 reg = ((modrm >> 3) & 7) | rex_r;
5160 mod = (modrm >> 6) & 3;
5161 t0 = tcg_temp_local_new();
5162 t1 = tcg_temp_local_new();
5163 t2 = tcg_temp_local_new();
5164 a0 = tcg_temp_local_new();
5165 gen_op_mov_v_reg(ot, t1, reg);
5166 if (mod == 3) {
5167 rm = (modrm & 7) | REX_B(s);
5168 gen_op_mov_v_reg(ot, t0, rm);
5169 } else {
5170 gen_lea_modrm(env, s, modrm);
5171 tcg_gen_mov_tl(a0, cpu_A0);
5172 gen_op_ld_v(s, ot, t0, a0);
5173 rm = 0; /* avoid warning */
5174 }
5175 label1 = gen_new_label();
5176 tcg_gen_mov_tl(t2, cpu_regs[R_EAX]);
5177 gen_extu(ot, t0);
5178 gen_extu(ot, t2);
5179 tcg_gen_brcond_tl(TCG_COND_EQ, t2, t0, label1);
5180 label2 = gen_new_label();
5181 if (mod == 3) {
5182 gen_op_mov_reg_v(ot, R_EAX, t0);
5183 tcg_gen_br(label2);
5184 gen_set_label(label1);
5185 gen_op_mov_reg_v(ot, rm, t1);
5186 } else {
5187 /* perform no-op store cycle like physical cpu; must be
5188 before changing accumulator to ensure idempotency if
5189 the store faults and the instruction is restarted */
5190 gen_op_st_v(s, ot, t0, a0);
5191 gen_op_mov_reg_v(ot, R_EAX, t0);
5192 tcg_gen_br(label2);
5193 gen_set_label(label1);
5194 gen_op_st_v(s, ot, t1, a0);
5195 }
5196 gen_set_label(label2);
5197 tcg_gen_mov_tl(cpu_cc_src, t0);
5198 tcg_gen_mov_tl(cpu_cc_srcT, t2);
5199 tcg_gen_sub_tl(cpu_cc_dst, t2, t0);
5200 set_cc_op(s, CC_OP_SUBB + ot);
5201 tcg_temp_free(t0);
5202 tcg_temp_free(t1);
5203 tcg_temp_free(t2);
5204 tcg_temp_free(a0);
5205 }
5206 break;
5207 case 0x1c7: /* cmpxchg8b */
5208 modrm = cpu_ldub_code(env, s->pc++);
5209 mod = (modrm >> 6) & 3;
5210 if ((mod == 3) || ((modrm & 0x38) != 0x8))
5211 goto illegal_op;
5212 #ifdef TARGET_X86_64
5213 if (dflag == MO_64) {
5214 if (!(s->cpuid_ext_features & CPUID_EXT_CX16))
5215 goto illegal_op;
5216 gen_jmp_im(pc_start - s->cs_base);
5217 gen_update_cc_op(s);
5218 gen_lea_modrm(env, s, modrm);
5219 gen_helper_cmpxchg16b(cpu_env, cpu_A0);
5220 } else
5221 #endif
5222 {
5223 if (!(s->cpuid_features & CPUID_CX8))
5224 goto illegal_op;
5225 gen_jmp_im(pc_start - s->cs_base);
5226 gen_update_cc_op(s);
5227 gen_lea_modrm(env, s, modrm);
5228 gen_helper_cmpxchg8b(cpu_env, cpu_A0);
5229 }
5230 set_cc_op(s, CC_OP_EFLAGS);
5231 break;
5232
5233 /**************************/
5234 /* push/pop */
5235 case 0x50 ... 0x57: /* push */
5236 gen_op_mov_v_reg(MO_32, cpu_T[0], (b & 7) | REX_B(s));
5237 gen_push_v(s, cpu_T[0]);
5238 break;
5239 case 0x58 ... 0x5f: /* pop */
5240 ot = gen_pop_T0(s);
5241 /* NOTE: order is important for pop %sp */
5242 gen_pop_update(s, ot);
5243 gen_op_mov_reg_v(ot, (b & 7) | REX_B(s), cpu_T[0]);
5244 break;
5245 case 0x60: /* pusha */
5246 if (CODE64(s))
5247 goto illegal_op;
5248 gen_pusha(s);
5249 break;
5250 case 0x61: /* popa */
5251 if (CODE64(s))
5252 goto illegal_op;
5253 gen_popa(s);
5254 break;
5255 case 0x68: /* push Iv */
5256 case 0x6a:
5257 ot = mo_pushpop(s, dflag);
5258 if (b == 0x68)
5259 val = insn_get(env, s, ot);
5260 else
5261 val = (int8_t)insn_get(env, s, MO_8);
5262 tcg_gen_movi_tl(cpu_T[0], val);
5263 gen_push_v(s, cpu_T[0]);
5264 break;
5265 case 0x8f: /* pop Ev */
5266 modrm = cpu_ldub_code(env, s->pc++);
5267 mod = (modrm >> 6) & 3;
5268 ot = gen_pop_T0(s);
5269 if (mod == 3) {
5270 /* NOTE: order is important for pop %sp */
5271 gen_pop_update(s, ot);
5272 rm = (modrm & 7) | REX_B(s);
5273 gen_op_mov_reg_v(ot, rm, cpu_T[0]);
5274 } else {
5275 /* NOTE: order is important too for MMU exceptions */
5276 s->popl_esp_hack = 1 << ot;
5277 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 1);
5278 s->popl_esp_hack = 0;
5279 gen_pop_update(s, ot);
5280 }
5281 break;
5282 case 0xc8: /* enter */
5283 {
5284 int level;
5285 val = cpu_lduw_code(env, s->pc);
5286 s->pc += 2;
5287 level = cpu_ldub_code(env, s->pc++);
5288 gen_enter(s, val, level);
5289 }
5290 break;
5291 case 0xc9: /* leave */
5292 /* XXX: exception not precise (ESP is updated before potential exception) */
5293 if (CODE64(s)) {
5294 gen_op_mov_v_reg(MO_64, cpu_T[0], R_EBP);
5295 gen_op_mov_reg_v(MO_64, R_ESP, cpu_T[0]);
5296 } else if (s->ss32) {
5297 gen_op_mov_v_reg(MO_32, cpu_T[0], R_EBP);
5298 gen_op_mov_reg_v(MO_32, R_ESP, cpu_T[0]);
5299 } else {
5300 gen_op_mov_v_reg(MO_16, cpu_T[0], R_EBP);
5301 gen_op_mov_reg_v(MO_16, R_ESP, cpu_T[0]);
5302 }
5303 ot = gen_pop_T0(s);
5304 gen_op_mov_reg_v(ot, R_EBP, cpu_T[0]);
5305 gen_pop_update(s, ot);
5306 break;
5307 case 0x06: /* push es */
5308 case 0x0e: /* push cs */
5309 case 0x16: /* push ss */
5310 case 0x1e: /* push ds */
5311 if (CODE64(s))
5312 goto illegal_op;
5313 gen_op_movl_T0_seg(b >> 3);
5314 gen_push_v(s, cpu_T[0]);
5315 break;
5316 case 0x1a0: /* push fs */
5317 case 0x1a8: /* push gs */
5318 gen_op_movl_T0_seg((b >> 3) & 7);
5319 gen_push_v(s, cpu_T[0]);
5320 break;
5321 case 0x07: /* pop es */
5322 case 0x17: /* pop ss */
5323 case 0x1f: /* pop ds */
5324 if (CODE64(s))
5325 goto illegal_op;
5326 reg = b >> 3;
5327 ot = gen_pop_T0(s);
5328 gen_movl_seg_T0(s, reg, pc_start - s->cs_base);
5329 gen_pop_update(s, ot);
5330 if (reg == R_SS) {
5331 /* if reg == SS, inhibit interrupts/trace. */
5332 /* If several instructions disable interrupts, only the
5333 _first_ does it */
5334 if (!(s->tb->flags & HF_INHIBIT_IRQ_MASK))
5335 gen_helper_set_inhibit_irq(cpu_env);
5336 s->tf = 0;
5337 }
5338 if (s->is_jmp) {
5339 gen_jmp_im(s->pc - s->cs_base);
5340 gen_eob(s);
5341 }
5342 break;
5343 case 0x1a1: /* pop fs */
5344 case 0x1a9: /* pop gs */
5345 ot = gen_pop_T0(s);
5346 gen_movl_seg_T0(s, (b >> 3) & 7, pc_start - s->cs_base);
5347 gen_pop_update(s, ot);
5348 if (s->is_jmp) {
5349 gen_jmp_im(s->pc - s->cs_base);
5350 gen_eob(s);
5351 }
5352 break;
5353
5354 /**************************/
5355 /* mov */
5356 case 0x88:
5357 case 0x89: /* mov Gv, Ev */
5358 ot = mo_b_d(b, dflag);
5359 modrm = cpu_ldub_code(env, s->pc++);
5360 reg = ((modrm >> 3) & 7) | rex_r;
5361
5362 /* generate a generic store */
5363 gen_ldst_modrm(env, s, modrm, ot, reg, 1);
5364 break;
5365 case 0xc6:
5366 case 0xc7: /* mov Ev, Iv */
5367 ot = mo_b_d(b, dflag);
5368 modrm = cpu_ldub_code(env, s->pc++);
5369 mod = (modrm >> 6) & 3;
5370 if (mod != 3) {
5371 s->rip_offset = insn_const_size(ot);
5372 gen_lea_modrm(env, s, modrm);
5373 }
5374 val = insn_get(env, s, ot);
5375 tcg_gen_movi_tl(cpu_T[0], val);
5376 if (mod != 3) {
5377 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
5378 } else {
5379 gen_op_mov_reg_v(ot, (modrm & 7) | REX_B(s), cpu_T[0]);
5380 }
5381 break;
5382 case 0x8a:
5383 case 0x8b: /* mov Ev, Gv */
5384 ot = mo_b_d(b, dflag);
5385 modrm = cpu_ldub_code(env, s->pc++);
5386 reg = ((modrm >> 3) & 7) | rex_r;
5387
5388 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
5389 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
5390 break;
5391 case 0x8e: /* mov seg, Gv */
5392 modrm = cpu_ldub_code(env, s->pc++);
5393 reg = (modrm >> 3) & 7;
5394 if (reg >= 6 || reg == R_CS)
5395 goto illegal_op;
5396 gen_ldst_modrm(env, s, modrm, MO_16, OR_TMP0, 0);
5397 gen_movl_seg_T0(s, reg, pc_start - s->cs_base);
5398 if (reg == R_SS) {
5399 /* if reg == SS, inhibit interrupts/trace */
5400 /* If several instructions disable interrupts, only the
5401 _first_ does it */
5402 if (!(s->tb->flags & HF_INHIBIT_IRQ_MASK))
5403 gen_helper_set_inhibit_irq(cpu_env);
5404 s->tf = 0;
5405 }
5406 if (s->is_jmp) {
5407 gen_jmp_im(s->pc - s->cs_base);
5408 gen_eob(s);
5409 }
5410 break;
5411 case 0x8c: /* mov Gv, seg */
5412 modrm = cpu_ldub_code(env, s->pc++);
5413 reg = (modrm >> 3) & 7;
5414 mod = (modrm >> 6) & 3;
5415 if (reg >= 6)
5416 goto illegal_op;
5417 gen_op_movl_T0_seg(reg);
5418 ot = mod == 3 ? dflag : MO_16;
5419 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 1);
5420 break;
5421
5422 case 0x1b6: /* movzbS Gv, Eb */
5423 case 0x1b7: /* movzwS Gv, Eb */
5424 case 0x1be: /* movsbS Gv, Eb */
5425 case 0x1bf: /* movswS Gv, Eb */
5426 {
5427 TCGMemOp d_ot;
5428 TCGMemOp s_ot;
5429
5430 /* d_ot is the size of destination */
5431 d_ot = dflag;
5432 /* ot is the size of source */
5433 ot = (b & 1) + MO_8;
5434 /* s_ot is the sign+size of source */
5435 s_ot = b & 8 ? MO_SIGN | ot : ot;
5436
5437 modrm = cpu_ldub_code(env, s->pc++);
5438 reg = ((modrm >> 3) & 7) | rex_r;
5439 mod = (modrm >> 6) & 3;
5440 rm = (modrm & 7) | REX_B(s);
5441
5442 if (mod == 3) {
5443 gen_op_mov_v_reg(ot, cpu_T[0], rm);
5444 switch (s_ot) {
5445 case MO_UB:
5446 tcg_gen_ext8u_tl(cpu_T[0], cpu_T[0]);
5447 break;
5448 case MO_SB:
5449 tcg_gen_ext8s_tl(cpu_T[0], cpu_T[0]);
5450 break;
5451 case MO_UW:
5452 tcg_gen_ext16u_tl(cpu_T[0], cpu_T[0]);
5453 break;
5454 default:
5455 case MO_SW:
5456 tcg_gen_ext16s_tl(cpu_T[0], cpu_T[0]);
5457 break;
5458 }
5459 gen_op_mov_reg_v(d_ot, reg, cpu_T[0]);
5460 } else {
5461 gen_lea_modrm(env, s, modrm);
5462 gen_op_ld_v(s, s_ot, cpu_T[0], cpu_A0);
5463 gen_op_mov_reg_v(d_ot, reg, cpu_T[0]);
5464 }
5465 }
5466 break;
5467
5468 case 0x8d: /* lea */
5469 ot = dflag;
5470 modrm = cpu_ldub_code(env, s->pc++);
5471 mod = (modrm >> 6) & 3;
5472 if (mod == 3)
5473 goto illegal_op;
5474 reg = ((modrm >> 3) & 7) | rex_r;
5475 /* we must ensure that no segment is added */
5476 s->override = -1;
5477 val = s->addseg;
5478 s->addseg = 0;
5479 gen_lea_modrm(env, s, modrm);
5480 s->addseg = val;
5481 gen_op_mov_reg_v(ot, reg, cpu_A0);
5482 break;
5483
5484 case 0xa0: /* mov EAX, Ov */
5485 case 0xa1:
5486 case 0xa2: /* mov Ov, EAX */
5487 case 0xa3:
5488 {
5489 target_ulong offset_addr;
5490
5491 ot = mo_b_d(b, dflag);
5492 switch (s->aflag) {
5493 #ifdef TARGET_X86_64
5494 case MO_64:
5495 offset_addr = cpu_ldq_code(env, s->pc);
5496 s->pc += 8;
5497 break;
5498 #endif
5499 default:
5500 offset_addr = insn_get(env, s, s->aflag);
5501 break;
5502 }
5503 tcg_gen_movi_tl(cpu_A0, offset_addr);
5504 gen_add_A0_ds_seg(s);
5505 if ((b & 2) == 0) {
5506 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
5507 gen_op_mov_reg_v(ot, R_EAX, cpu_T[0]);
5508 } else {
5509 gen_op_mov_v_reg(ot, cpu_T[0], R_EAX);
5510 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
5511 }
5512 }
5513 break;
5514 case 0xd7: /* xlat */
5515 tcg_gen_mov_tl(cpu_A0, cpu_regs[R_EBX]);
5516 tcg_gen_ext8u_tl(cpu_T[0], cpu_regs[R_EAX]);
5517 tcg_gen_add_tl(cpu_A0, cpu_A0, cpu_T[0]);
5518 gen_extu(s->aflag, cpu_A0);
5519 gen_add_A0_ds_seg(s);
5520 gen_op_ld_v(s, MO_8, cpu_T[0], cpu_A0);
5521 gen_op_mov_reg_v(MO_8, R_EAX, cpu_T[0]);
5522 break;
5523 case 0xb0 ... 0xb7: /* mov R, Ib */
5524 val = insn_get(env, s, MO_8);
5525 tcg_gen_movi_tl(cpu_T[0], val);
5526 gen_op_mov_reg_v(MO_8, (b & 7) | REX_B(s), cpu_T[0]);
5527 break;
5528 case 0xb8 ... 0xbf: /* mov R, Iv */
5529 #ifdef TARGET_X86_64
5530 if (dflag == MO_64) {
5531 uint64_t tmp;
5532 /* 64 bit case */
5533 tmp = cpu_ldq_code(env, s->pc);
5534 s->pc += 8;
5535 reg = (b & 7) | REX_B(s);
5536 tcg_gen_movi_tl(cpu_T[0], tmp);
5537 gen_op_mov_reg_v(MO_64, reg, cpu_T[0]);
5538 } else
5539 #endif
5540 {
5541 ot = dflag;
5542 val = insn_get(env, s, ot);
5543 reg = (b & 7) | REX_B(s);
5544 tcg_gen_movi_tl(cpu_T[0], val);
5545 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
5546 }
5547 break;
5548
5549 case 0x91 ... 0x97: /* xchg R, EAX */
5550 do_xchg_reg_eax:
5551 ot = dflag;
5552 reg = (b & 7) | REX_B(s);
5553 rm = R_EAX;
5554 goto do_xchg_reg;
5555 case 0x86:
5556 case 0x87: /* xchg Ev, Gv */
5557 ot = mo_b_d(b, dflag);
5558 modrm = cpu_ldub_code(env, s->pc++);
5559 reg = ((modrm >> 3) & 7) | rex_r;
5560 mod = (modrm >> 6) & 3;
5561 if (mod == 3) {
5562 rm = (modrm & 7) | REX_B(s);
5563 do_xchg_reg:
5564 gen_op_mov_v_reg(ot, cpu_T[0], reg);
5565 gen_op_mov_v_reg(ot, cpu_T[1], rm);
5566 gen_op_mov_reg_v(ot, rm, cpu_T[0]);
5567 gen_op_mov_reg_v(ot, reg, cpu_T[1]);
5568 } else {
5569 gen_lea_modrm(env, s, modrm);
5570 gen_op_mov_v_reg(ot, cpu_T[0], reg);
5571 /* for xchg, lock is implicit */
5572 if (!(prefixes & PREFIX_LOCK))
5573 gen_helper_lock();
5574 gen_op_ld_v(s, ot, cpu_T[1], cpu_A0);
5575 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
5576 if (!(prefixes & PREFIX_LOCK))
5577 gen_helper_unlock();
5578 gen_op_mov_reg_v(ot, reg, cpu_T[1]);
5579 }
5580 break;
5581 case 0xc4: /* les Gv */
5582 /* In CODE64 this is VEX3; see above. */
5583 op = R_ES;
5584 goto do_lxx;
5585 case 0xc5: /* lds Gv */
5586 /* In CODE64 this is VEX2; see above. */
5587 op = R_DS;
5588 goto do_lxx;
5589 case 0x1b2: /* lss Gv */
5590 op = R_SS;
5591 goto do_lxx;
5592 case 0x1b4: /* lfs Gv */
5593 op = R_FS;
5594 goto do_lxx;
5595 case 0x1b5: /* lgs Gv */
5596 op = R_GS;
5597 do_lxx:
5598 ot = dflag != MO_16 ? MO_32 : MO_16;
5599 modrm = cpu_ldub_code(env, s->pc++);
5600 reg = ((modrm >> 3) & 7) | rex_r;
5601 mod = (modrm >> 6) & 3;
5602 if (mod == 3)
5603 goto illegal_op;
5604 gen_lea_modrm(env, s, modrm);
5605 gen_op_ld_v(s, ot, cpu_T[1], cpu_A0);
5606 gen_add_A0_im(s, 1 << ot);
5607 /* load the segment first to handle exceptions properly */
5608 gen_op_ld_v(s, MO_16, cpu_T[0], cpu_A0);
5609 gen_movl_seg_T0(s, op, pc_start - s->cs_base);
5610 /* then put the data */
5611 gen_op_mov_reg_v(ot, reg, cpu_T[1]);
5612 if (s->is_jmp) {
5613 gen_jmp_im(s->pc - s->cs_base);
5614 gen_eob(s);
5615 }
5616 break;
5617
5618 /************************/
5619 /* shifts */
5620 case 0xc0:
5621 case 0xc1:
5622 /* shift Ev,Ib */
5623 shift = 2;
5624 grp2:
5625 {
5626 ot = mo_b_d(b, dflag);
5627 modrm = cpu_ldub_code(env, s->pc++);
5628 mod = (modrm >> 6) & 3;
5629 op = (modrm >> 3) & 7;
5630
5631 if (mod != 3) {
5632 if (shift == 2) {
5633 s->rip_offset = 1;
5634 }
5635 gen_lea_modrm(env, s, modrm);
5636 opreg = OR_TMP0;
5637 } else {
5638 opreg = (modrm & 7) | REX_B(s);
5639 }
5640
5641 /* simpler op */
5642 if (shift == 0) {
5643 gen_shift(s, op, ot, opreg, OR_ECX);
5644 } else {
5645 if (shift == 2) {
5646 shift = cpu_ldub_code(env, s->pc++);
5647 }
5648 gen_shifti(s, op, ot, opreg, shift);
5649 }
5650 }
5651 break;
5652 case 0xd0:
5653 case 0xd1:
5654 /* shift Ev,1 */
5655 shift = 1;
5656 goto grp2;
5657 case 0xd2:
5658 case 0xd3:
5659 /* shift Ev,cl */
5660 shift = 0;
5661 goto grp2;
5662
5663 case 0x1a4: /* shld imm */
5664 op = 0;
5665 shift = 1;
5666 goto do_shiftd;
5667 case 0x1a5: /* shld cl */
5668 op = 0;
5669 shift = 0;
5670 goto do_shiftd;
5671 case 0x1ac: /* shrd imm */
5672 op = 1;
5673 shift = 1;
5674 goto do_shiftd;
5675 case 0x1ad: /* shrd cl */
5676 op = 1;
5677 shift = 0;
5678 do_shiftd:
5679 ot = dflag;
5680 modrm = cpu_ldub_code(env, s->pc++);
5681 mod = (modrm >> 6) & 3;
5682 rm = (modrm & 7) | REX_B(s);
5683 reg = ((modrm >> 3) & 7) | rex_r;
5684 if (mod != 3) {
5685 gen_lea_modrm(env, s, modrm);
5686 opreg = OR_TMP0;
5687 } else {
5688 opreg = rm;
5689 }
5690 gen_op_mov_v_reg(ot, cpu_T[1], reg);
5691
5692 if (shift) {
5693 TCGv imm = tcg_const_tl(cpu_ldub_code(env, s->pc++));
5694 gen_shiftd_rm_T1(s, ot, opreg, op, imm);
5695 tcg_temp_free(imm);
5696 } else {
5697 gen_shiftd_rm_T1(s, ot, opreg, op, cpu_regs[R_ECX]);
5698 }
5699 break;
5700
5701 /************************/
5702 /* floats */
5703 case 0xd8 ... 0xdf:
5704 if (s->flags & (HF_EM_MASK | HF_TS_MASK)) {
5705 /* if CR0.EM or CR0.TS are set, generate an FPU exception */
5706 /* XXX: what to do if illegal op ? */
5707 gen_exception(s, EXCP07_PREX, pc_start - s->cs_base);
5708 break;
5709 }
5710 modrm = cpu_ldub_code(env, s->pc++);
5711 mod = (modrm >> 6) & 3;
5712 rm = modrm & 7;
5713 op = ((b & 7) << 3) | ((modrm >> 3) & 7);
5714 if (mod != 3) {
5715 /* memory op */
5716 gen_lea_modrm(env, s, modrm);
5717 switch(op) {
5718 case 0x00 ... 0x07: /* fxxxs */
5719 case 0x10 ... 0x17: /* fixxxl */
5720 case 0x20 ... 0x27: /* fxxxl */
5721 case 0x30 ... 0x37: /* fixxx */
5722 {
5723 int op1;
5724 op1 = op & 7;
5725
5726 switch(op >> 4) {
5727 case 0:
5728 tcg_gen_qemu_ld_i32(cpu_tmp2_i32, cpu_A0,
5729 s->mem_index, MO_LEUL);
5730 gen_helper_flds_FT0(cpu_env, cpu_tmp2_i32);
5731 break;
5732 case 1:
5733 tcg_gen_qemu_ld_i32(cpu_tmp2_i32, cpu_A0,
5734 s->mem_index, MO_LEUL);
5735 gen_helper_fildl_FT0(cpu_env, cpu_tmp2_i32);
5736 break;
5737 case 2:
5738 tcg_gen_qemu_ld_i64(cpu_tmp1_i64, cpu_A0,
5739 s->mem_index, MO_LEQ);
5740 gen_helper_fldl_FT0(cpu_env, cpu_tmp1_i64);
5741 break;
5742 case 3:
5743 default:
5744 tcg_gen_qemu_ld_i32(cpu_tmp2_i32, cpu_A0,
5745 s->mem_index, MO_LESW);
5746 gen_helper_fildl_FT0(cpu_env, cpu_tmp2_i32);
5747 break;
5748 }
5749
5750 gen_helper_fp_arith_ST0_FT0(op1);
5751 if (op1 == 3) {
5752 /* fcomp needs pop */
5753 gen_helper_fpop(cpu_env);
5754 }
5755 }
5756 break;
5757 case 0x08: /* flds */
5758 case 0x0a: /* fsts */
5759 case 0x0b: /* fstps */
5760 case 0x18 ... 0x1b: /* fildl, fisttpl, fistl, fistpl */
5761 case 0x28 ... 0x2b: /* fldl, fisttpll, fstl, fstpl */
5762 case 0x38 ... 0x3b: /* filds, fisttps, fists, fistps */
5763 switch(op & 7) {
5764 case 0:
5765 switch(op >> 4) {
5766 case 0:
5767 tcg_gen_qemu_ld_i32(cpu_tmp2_i32, cpu_A0,
5768 s->mem_index, MO_LEUL);
5769 gen_helper_flds_ST0(cpu_env, cpu_tmp2_i32);
5770 break;
5771 case 1:
5772 tcg_gen_qemu_ld_i32(cpu_tmp2_i32, cpu_A0,
5773 s->mem_index, MO_LEUL);
5774 gen_helper_fildl_ST0(cpu_env, cpu_tmp2_i32);
5775 break;
5776 case 2:
5777 tcg_gen_qemu_ld_i64(cpu_tmp1_i64, cpu_A0,
5778 s->mem_index, MO_LEQ);
5779 gen_helper_fldl_ST0(cpu_env, cpu_tmp1_i64);
5780 break;
5781 case 3:
5782 default:
5783 tcg_gen_qemu_ld_i32(cpu_tmp2_i32, cpu_A0,
5784 s->mem_index, MO_LESW);
5785 gen_helper_fildl_ST0(cpu_env, cpu_tmp2_i32);
5786 break;
5787 }
5788 break;
5789 case 1:
5790 /* XXX: the corresponding CPUID bit must be tested ! */
5791 switch(op >> 4) {
5792 case 1:
5793 gen_helper_fisttl_ST0(cpu_tmp2_i32, cpu_env);
5794 tcg_gen_qemu_st_i32(cpu_tmp2_i32, cpu_A0,
5795 s->mem_index, MO_LEUL);
5796 break;
5797 case 2:
5798 gen_helper_fisttll_ST0(cpu_tmp1_i64, cpu_env);
5799 tcg_gen_qemu_st_i64(cpu_tmp1_i64, cpu_A0,
5800 s->mem_index, MO_LEQ);
5801 break;
5802 case 3:
5803 default:
5804 gen_helper_fistt_ST0(cpu_tmp2_i32, cpu_env);
5805 tcg_gen_qemu_st_i32(cpu_tmp2_i32, cpu_A0,
5806 s->mem_index, MO_LEUW);
5807 break;
5808 }
5809 gen_helper_fpop(cpu_env);
5810 break;
5811 default:
5812 switch(op >> 4) {
5813 case 0:
5814 gen_helper_fsts_ST0(cpu_tmp2_i32, cpu_env);
5815 tcg_gen_qemu_st_i32(cpu_tmp2_i32, cpu_A0,
5816 s->mem_index, MO_LEUL);
5817 break;
5818 case 1:
5819 gen_helper_fistl_ST0(cpu_tmp2_i32, cpu_env);
5820 tcg_gen_qemu_st_i32(cpu_tmp2_i32, cpu_A0,
5821 s->mem_index, MO_LEUL);
5822 break;
5823 case 2:
5824 gen_helper_fstl_ST0(cpu_tmp1_i64, cpu_env);
5825 tcg_gen_qemu_st_i64(cpu_tmp1_i64, cpu_A0,
5826 s->mem_index, MO_LEQ);
5827 break;
5828 case 3:
5829 default:
5830 gen_helper_fist_ST0(cpu_tmp2_i32, cpu_env);
5831 tcg_gen_qemu_st_i32(cpu_tmp2_i32, cpu_A0,
5832 s->mem_index, MO_LEUW);
5833 break;
5834 }
5835 if ((op & 7) == 3)
5836 gen_helper_fpop(cpu_env);
5837 break;
5838 }
5839 break;
5840 case 0x0c: /* fldenv mem */
5841 gen_update_cc_op(s);
5842 gen_jmp_im(pc_start - s->cs_base);
5843 gen_helper_fldenv(cpu_env, cpu_A0, tcg_const_i32(dflag - 1));
5844 break;
5845 case 0x0d: /* fldcw mem */
5846 tcg_gen_qemu_ld_i32(cpu_tmp2_i32, cpu_A0,
5847 s->mem_index, MO_LEUW);
5848 gen_helper_fldcw(cpu_env, cpu_tmp2_i32);
5849 break;
5850 case 0x0e: /* fnstenv mem */
5851 gen_update_cc_op(s);
5852 gen_jmp_im(pc_start - s->cs_base);
5853 gen_helper_fstenv(cpu_env, cpu_A0, tcg_const_i32(dflag - 1));
5854 break;
5855 case 0x0f: /* fnstcw mem */
5856 gen_helper_fnstcw(cpu_tmp2_i32, cpu_env);
5857 tcg_gen_qemu_st_i32(cpu_tmp2_i32, cpu_A0,
5858 s->mem_index, MO_LEUW);
5859 break;
5860 case 0x1d: /* fldt mem */
5861 gen_update_cc_op(s);
5862 gen_jmp_im(pc_start - s->cs_base);
5863 gen_helper_fldt_ST0(cpu_env, cpu_A0);
5864 break;
5865 case 0x1f: /* fstpt mem */
5866 gen_update_cc_op(s);
5867 gen_jmp_im(pc_start - s->cs_base);
5868 gen_helper_fstt_ST0(cpu_env, cpu_A0);
5869 gen_helper_fpop(cpu_env);
5870 break;
5871 case 0x2c: /* frstor mem */
5872 gen_update_cc_op(s);
5873 gen_jmp_im(pc_start - s->cs_base);
5874 gen_helper_frstor(cpu_env, cpu_A0, tcg_const_i32(dflag - 1));
5875 break;
5876 case 0x2e: /* fnsave mem */
5877 gen_update_cc_op(s);
5878 gen_jmp_im(pc_start - s->cs_base);
5879 gen_helper_fsave(cpu_env, cpu_A0, tcg_const_i32(dflag - 1));
5880 break;
5881 case 0x2f: /* fnstsw mem */
5882 gen_helper_fnstsw(cpu_tmp2_i32, cpu_env);
5883 tcg_gen_qemu_st_i32(cpu_tmp2_i32, cpu_A0,
5884 s->mem_index, MO_LEUW);
5885 break;
5886 case 0x3c: /* fbld */
5887 gen_update_cc_op(s);
5888 gen_jmp_im(pc_start - s->cs_base);
5889 gen_helper_fbld_ST0(cpu_env, cpu_A0);
5890 break;
5891 case 0x3e: /* fbstp */
5892 gen_update_cc_op(s);
5893 gen_jmp_im(pc_start - s->cs_base);
5894 gen_helper_fbst_ST0(cpu_env, cpu_A0);
5895 gen_helper_fpop(cpu_env);
5896 break;
5897 case 0x3d: /* fildll */
5898 tcg_gen_qemu_ld_i64(cpu_tmp1_i64, cpu_A0, s->mem_index, MO_LEQ);
5899 gen_helper_fildll_ST0(cpu_env, cpu_tmp1_i64);
5900 break;
5901 case 0x3f: /* fistpll */
5902 gen_helper_fistll_ST0(cpu_tmp1_i64, cpu_env);
5903 tcg_gen_qemu_st_i64(cpu_tmp1_i64, cpu_A0, s->mem_index, MO_LEQ);
5904 gen_helper_fpop(cpu_env);
5905 break;
5906 default:
5907 goto illegal_op;
5908 }
5909 } else {
5910 /* register float ops */
5911 opreg = rm;
5912
5913 switch(op) {
5914 case 0x08: /* fld sti */
5915 gen_helper_fpush(cpu_env);
5916 gen_helper_fmov_ST0_STN(cpu_env,
5917 tcg_const_i32((opreg + 1) & 7));
5918 break;
5919 case 0x09: /* fxchg sti */
5920 case 0x29: /* fxchg4 sti, undocumented op */
5921 case 0x39: /* fxchg7 sti, undocumented op */
5922 gen_helper_fxchg_ST0_STN(cpu_env, tcg_const_i32(opreg));
5923 break;
5924 case 0x0a: /* grp d9/2 */
5925 switch(rm) {
5926 case 0: /* fnop */
5927 /* check exceptions (FreeBSD FPU probe) */
5928 gen_update_cc_op(s);
5929 gen_jmp_im(pc_start - s->cs_base);
5930 gen_helper_fwait(cpu_env);
5931 break;
5932 default:
5933 goto illegal_op;
5934 }
5935 break;
5936 case 0x0c: /* grp d9/4 */
5937 switch(rm) {
5938 case 0: /* fchs */
5939 gen_helper_fchs_ST0(cpu_env);
5940 break;
5941 case 1: /* fabs */
5942 gen_helper_fabs_ST0(cpu_env);
5943 break;
5944 case 4: /* ftst */
5945 gen_helper_fldz_FT0(cpu_env);
5946 gen_helper_fcom_ST0_FT0(cpu_env);
5947 break;
5948 case 5: /* fxam */
5949 gen_helper_fxam_ST0(cpu_env);
5950 break;
5951 default:
5952 goto illegal_op;
5953 }
5954 break;
5955 case 0x0d: /* grp d9/5 */
5956 {
5957 switch(rm) {
5958 case 0:
5959 gen_helper_fpush(cpu_env);
5960 gen_helper_fld1_ST0(cpu_env);
5961 break;
5962 case 1:
5963 gen_helper_fpush(cpu_env);
5964 gen_helper_fldl2t_ST0(cpu_env);
5965 break;
5966 case 2:
5967 gen_helper_fpush(cpu_env);
5968 gen_helper_fldl2e_ST0(cpu_env);
5969 break;
5970 case 3:
5971 gen_helper_fpush(cpu_env);
5972 gen_helper_fldpi_ST0(cpu_env);
5973 break;
5974 case 4:
5975 gen_helper_fpush(cpu_env);
5976 gen_helper_fldlg2_ST0(cpu_env);
5977 break;
5978 case 5:
5979 gen_helper_fpush(cpu_env);
5980 gen_helper_fldln2_ST0(cpu_env);
5981 break;
5982 case 6:
5983 gen_helper_fpush(cpu_env);
5984 gen_helper_fldz_ST0(cpu_env);
5985 break;
5986 default:
5987 goto illegal_op;
5988 }
5989 }
5990 break;
5991 case 0x0e: /* grp d9/6 */
5992 switch(rm) {
5993 case 0: /* f2xm1 */
5994 gen_helper_f2xm1(cpu_env);
5995 break;
5996 case 1: /* fyl2x */
5997 gen_helper_fyl2x(cpu_env);
5998 break;
5999 case 2: /* fptan */
6000 gen_helper_fptan(cpu_env);
6001 break;
6002 case 3: /* fpatan */
6003 gen_helper_fpatan(cpu_env);
6004 break;
6005 case 4: /* fxtract */
6006 gen_helper_fxtract(cpu_env);
6007 break;
6008 case 5: /* fprem1 */
6009 gen_helper_fprem1(cpu_env);
6010 break;
6011 case 6: /* fdecstp */
6012 gen_helper_fdecstp(cpu_env);
6013 break;
6014 default:
6015 case 7: /* fincstp */
6016 gen_helper_fincstp(cpu_env);
6017 break;
6018 }
6019 break;
6020 case 0x0f: /* grp d9/7 */
6021 switch(rm) {
6022 case 0: /* fprem */
6023 gen_helper_fprem(cpu_env);
6024 break;
6025 case 1: /* fyl2xp1 */
6026 gen_helper_fyl2xp1(cpu_env);
6027 break;
6028 case 2: /* fsqrt */
6029 gen_helper_fsqrt(cpu_env);
6030 break;
6031 case 3: /* fsincos */
6032 gen_helper_fsincos(cpu_env);
6033 break;
6034 case 5: /* fscale */
6035 gen_helper_fscale(cpu_env);
6036 break;
6037 case 4: /* frndint */
6038 gen_helper_frndint(cpu_env);
6039 break;
6040 case 6: /* fsin */
6041 gen_helper_fsin(cpu_env);
6042 break;
6043 default:
6044 case 7: /* fcos */
6045 gen_helper_fcos(cpu_env);
6046 break;
6047 }
6048 break;
6049 case 0x00: case 0x01: case 0x04 ... 0x07: /* fxxx st, sti */
6050 case 0x20: case 0x21: case 0x24 ... 0x27: /* fxxx sti, st */
6051 case 0x30: case 0x31: case 0x34 ... 0x37: /* fxxxp sti, st */
6052 {
6053 int op1;
6054
6055 op1 = op & 7;
6056 if (op >= 0x20) {
6057 gen_helper_fp_arith_STN_ST0(op1, opreg);
6058 if (op >= 0x30)
6059 gen_helper_fpop(cpu_env);
6060 } else {
6061 gen_helper_fmov_FT0_STN(cpu_env, tcg_const_i32(opreg));
6062 gen_helper_fp_arith_ST0_FT0(op1);
6063 }
6064 }
6065 break;
6066 case 0x02: /* fcom */
6067 case 0x22: /* fcom2, undocumented op */
6068 gen_helper_fmov_FT0_STN(cpu_env, tcg_const_i32(opreg));
6069 gen_helper_fcom_ST0_FT0(cpu_env);
6070 break;
6071 case 0x03: /* fcomp */
6072 case 0x23: /* fcomp3, undocumented op */
6073 case 0x32: /* fcomp5, undocumented op */
6074 gen_helper_fmov_FT0_STN(cpu_env, tcg_const_i32(opreg));
6075 gen_helper_fcom_ST0_FT0(cpu_env);
6076 gen_helper_fpop(cpu_env);
6077 break;
6078 case 0x15: /* da/5 */
6079 switch(rm) {
6080 case 1: /* fucompp */
6081 gen_helper_fmov_FT0_STN(cpu_env, tcg_const_i32(1));
6082 gen_helper_fucom_ST0_FT0(cpu_env);
6083 gen_helper_fpop(cpu_env);
6084 gen_helper_fpop(cpu_env);
6085 break;
6086 default:
6087 goto illegal_op;
6088 }
6089 break;
6090 case 0x1c:
6091 switch(rm) {
6092 case 0: /* feni (287 only, just do nop here) */
6093 break;
6094 case 1: /* fdisi (287 only, just do nop here) */
6095 break;
6096 case 2: /* fclex */
6097 gen_helper_fclex(cpu_env);
6098 break;
6099 case 3: /* fninit */
6100 gen_helper_fninit(cpu_env);
6101 break;
6102 case 4: /* fsetpm (287 only, just do nop here) */
6103 break;
6104 default:
6105 goto illegal_op;
6106 }
6107 break;
6108 case 0x1d: /* fucomi */
6109 if (!(s->cpuid_features & CPUID_CMOV)) {
6110 goto illegal_op;
6111 }
6112 gen_update_cc_op(s);
6113 gen_helper_fmov_FT0_STN(cpu_env, tcg_const_i32(opreg));
6114 gen_helper_fucomi_ST0_FT0(cpu_env);
6115 set_cc_op(s, CC_OP_EFLAGS);
6116 break;
6117 case 0x1e: /* fcomi */
6118 if (!(s->cpuid_features & CPUID_CMOV)) {
6119 goto illegal_op;
6120 }
6121 gen_update_cc_op(s);
6122 gen_helper_fmov_FT0_STN(cpu_env, tcg_const_i32(opreg));
6123 gen_helper_fcomi_ST0_FT0(cpu_env);
6124 set_cc_op(s, CC_OP_EFLAGS);
6125 break;
6126 case 0x28: /* ffree sti */
6127 gen_helper_ffree_STN(cpu_env, tcg_const_i32(opreg));
6128 break;
6129 case 0x2a: /* fst sti */
6130 gen_helper_fmov_STN_ST0(cpu_env, tcg_const_i32(opreg));
6131 break;
6132 case 0x2b: /* fstp sti */
6133 case 0x0b: /* fstp1 sti, undocumented op */
6134 case 0x3a: /* fstp8 sti, undocumented op */
6135 case 0x3b: /* fstp9 sti, undocumented op */
6136 gen_helper_fmov_STN_ST0(cpu_env, tcg_const_i32(opreg));
6137 gen_helper_fpop(cpu_env);
6138 break;
6139 case 0x2c: /* fucom st(i) */
6140 gen_helper_fmov_FT0_STN(cpu_env, tcg_const_i32(opreg));
6141 gen_helper_fucom_ST0_FT0(cpu_env);
6142 break;
6143 case 0x2d: /* fucomp st(i) */
6144 gen_helper_fmov_FT0_STN(cpu_env, tcg_const_i32(opreg));
6145 gen_helper_fucom_ST0_FT0(cpu_env);
6146 gen_helper_fpop(cpu_env);
6147 break;
6148 case 0x33: /* de/3 */
6149 switch(rm) {
6150 case 1: /* fcompp */
6151 gen_helper_fmov_FT0_STN(cpu_env, tcg_const_i32(1));
6152 gen_helper_fcom_ST0_FT0(cpu_env);
6153 gen_helper_fpop(cpu_env);
6154 gen_helper_fpop(cpu_env);
6155 break;
6156 default:
6157 goto illegal_op;
6158 }
6159 break;
6160 case 0x38: /* ffreep sti, undocumented op */
6161 gen_helper_ffree_STN(cpu_env, tcg_const_i32(opreg));
6162 gen_helper_fpop(cpu_env);
6163 break;
6164 case 0x3c: /* df/4 */
6165 switch(rm) {
6166 case 0:
6167 gen_helper_fnstsw(cpu_tmp2_i32, cpu_env);
6168 tcg_gen_extu_i32_tl(cpu_T[0], cpu_tmp2_i32);
6169 gen_op_mov_reg_v(MO_16, R_EAX, cpu_T[0]);
6170 break;
6171 default:
6172 goto illegal_op;
6173 }
6174 break;
6175 case 0x3d: /* fucomip */
6176 if (!(s->cpuid_features & CPUID_CMOV)) {
6177 goto illegal_op;
6178 }
6179 gen_update_cc_op(s);
6180 gen_helper_fmov_FT0_STN(cpu_env, tcg_const_i32(opreg));
6181 gen_helper_fucomi_ST0_FT0(cpu_env);
6182 gen_helper_fpop(cpu_env);
6183 set_cc_op(s, CC_OP_EFLAGS);
6184 break;
6185 case 0x3e: /* fcomip */
6186 if (!(s->cpuid_features & CPUID_CMOV)) {
6187 goto illegal_op;
6188 }
6189 gen_update_cc_op(s);
6190 gen_helper_fmov_FT0_STN(cpu_env, tcg_const_i32(opreg));
6191 gen_helper_fcomi_ST0_FT0(cpu_env);
6192 gen_helper_fpop(cpu_env);
6193 set_cc_op(s, CC_OP_EFLAGS);
6194 break;
6195 case 0x10 ... 0x13: /* fcmovxx */
6196 case 0x18 ... 0x1b:
6197 {
6198 int op1, l1;
6199 static const uint8_t fcmov_cc[8] = {
6200 (JCC_B << 1),
6201 (JCC_Z << 1),
6202 (JCC_BE << 1),
6203 (JCC_P << 1),
6204 };
6205
6206 if (!(s->cpuid_features & CPUID_CMOV)) {
6207 goto illegal_op;
6208 }
6209 op1 = fcmov_cc[op & 3] | (((op >> 3) & 1) ^ 1);
6210 l1 = gen_new_label();
6211 gen_jcc1_noeob(s, op1, l1);
6212 gen_helper_fmov_ST0_STN(cpu_env, tcg_const_i32(opreg));
6213 gen_set_label(l1);
6214 }
6215 break;
6216 default:
6217 goto illegal_op;
6218 }
6219 }
6220 break;
6221 /************************/
6222 /* string ops */
6223
6224 case 0xa4: /* movsS */
6225 case 0xa5:
6226 ot = mo_b_d(b, dflag);
6227 if (prefixes & (PREFIX_REPZ | PREFIX_REPNZ)) {
6228 gen_repz_movs(s, ot, pc_start - s->cs_base, s->pc - s->cs_base);
6229 } else {
6230 gen_movs(s, ot);
6231 }
6232 break;
6233
6234 case 0xaa: /* stosS */
6235 case 0xab:
6236 ot = mo_b_d(b, dflag);
6237 if (prefixes & (PREFIX_REPZ | PREFIX_REPNZ)) {
6238 gen_repz_stos(s, ot, pc_start - s->cs_base, s->pc - s->cs_base);
6239 } else {
6240 gen_stos(s, ot);
6241 }
6242 break;
6243 case 0xac: /* lodsS */
6244 case 0xad:
6245 ot = mo_b_d(b, dflag);
6246 if (prefixes & (PREFIX_REPZ | PREFIX_REPNZ)) {
6247 gen_repz_lods(s, ot, pc_start - s->cs_base, s->pc - s->cs_base);
6248 } else {
6249 gen_lods(s, ot);
6250 }
6251 break;
6252 case 0xae: /* scasS */
6253 case 0xaf:
6254 ot = mo_b_d(b, dflag);
6255 if (prefixes & PREFIX_REPNZ) {
6256 gen_repz_scas(s, ot, pc_start - s->cs_base, s->pc - s->cs_base, 1);
6257 } else if (prefixes & PREFIX_REPZ) {
6258 gen_repz_scas(s, ot, pc_start - s->cs_base, s->pc - s->cs_base, 0);
6259 } else {
6260 gen_scas(s, ot);
6261 }
6262 break;
6263
6264 case 0xa6: /* cmpsS */
6265 case 0xa7:
6266 ot = mo_b_d(b, dflag);
6267 if (prefixes & PREFIX_REPNZ) {
6268 gen_repz_cmps(s, ot, pc_start - s->cs_base, s->pc - s->cs_base, 1);
6269 } else if (prefixes & PREFIX_REPZ) {
6270 gen_repz_cmps(s, ot, pc_start - s->cs_base, s->pc - s->cs_base, 0);
6271 } else {
6272 gen_cmps(s, ot);
6273 }
6274 break;
6275 case 0x6c: /* insS */
6276 case 0x6d:
6277 ot = mo_b_d32(b, dflag);
6278 tcg_gen_ext16u_tl(cpu_T[0], cpu_regs[R_EDX]);
6279 gen_check_io(s, ot, pc_start - s->cs_base,
6280 SVM_IOIO_TYPE_MASK | svm_is_rep(prefixes) | 4);
6281 if (prefixes & (PREFIX_REPZ | PREFIX_REPNZ)) {
6282 gen_repz_ins(s, ot, pc_start - s->cs_base, s->pc - s->cs_base);
6283 } else {
6284 gen_ins(s, ot);
6285 if (use_icount) {
6286 gen_jmp(s, s->pc - s->cs_base);
6287 }
6288 }
6289 break;
6290 case 0x6e: /* outsS */
6291 case 0x6f:
6292 ot = mo_b_d32(b, dflag);
6293 tcg_gen_ext16u_tl(cpu_T[0], cpu_regs[R_EDX]);
6294 gen_check_io(s, ot, pc_start - s->cs_base,
6295 svm_is_rep(prefixes) | 4);
6296 if (prefixes & (PREFIX_REPZ | PREFIX_REPNZ)) {
6297 gen_repz_outs(s, ot, pc_start - s->cs_base, s->pc - s->cs_base);
6298 } else {
6299 gen_outs(s, ot);
6300 if (use_icount) {
6301 gen_jmp(s, s->pc - s->cs_base);
6302 }
6303 }
6304 break;
6305
6306 /************************/
6307 /* port I/O */
6308
6309 case 0xe4:
6310 case 0xe5:
6311 ot = mo_b_d32(b, dflag);
6312 val = cpu_ldub_code(env, s->pc++);
6313 tcg_gen_movi_tl(cpu_T[0], val);
6314 gen_check_io(s, ot, pc_start - s->cs_base,
6315 SVM_IOIO_TYPE_MASK | svm_is_rep(prefixes));
6316 if (use_icount)
6317 gen_io_start();
6318 tcg_gen_movi_i32(cpu_tmp2_i32, val);
6319 gen_helper_in_func(ot, cpu_T[1], cpu_tmp2_i32);
6320 gen_op_mov_reg_v(ot, R_EAX, cpu_T[1]);
6321 if (use_icount) {
6322 gen_io_end();
6323 gen_jmp(s, s->pc - s->cs_base);
6324 }
6325 break;
6326 case 0xe6:
6327 case 0xe7:
6328 ot = mo_b_d32(b, dflag);
6329 val = cpu_ldub_code(env, s->pc++);
6330 tcg_gen_movi_tl(cpu_T[0], val);
6331 gen_check_io(s, ot, pc_start - s->cs_base,
6332 svm_is_rep(prefixes));
6333 gen_op_mov_v_reg(ot, cpu_T[1], R_EAX);
6334
6335 if (use_icount)
6336 gen_io_start();
6337 tcg_gen_movi_i32(cpu_tmp2_i32, val);
6338 tcg_gen_trunc_tl_i32(cpu_tmp3_i32, cpu_T[1]);
6339 gen_helper_out_func(ot, cpu_tmp2_i32, cpu_tmp3_i32);
6340 if (use_icount) {
6341 gen_io_end();
6342 gen_jmp(s, s->pc - s->cs_base);
6343 }
6344 break;
6345 case 0xec:
6346 case 0xed:
6347 ot = mo_b_d32(b, dflag);
6348 tcg_gen_ext16u_tl(cpu_T[0], cpu_regs[R_EDX]);
6349 gen_check_io(s, ot, pc_start - s->cs_base,
6350 SVM_IOIO_TYPE_MASK | svm_is_rep(prefixes));
6351 if (use_icount)
6352 gen_io_start();
6353 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
6354 gen_helper_in_func(ot, cpu_T[1], cpu_tmp2_i32);
6355 gen_op_mov_reg_v(ot, R_EAX, cpu_T[1]);
6356 if (use_icount) {
6357 gen_io_end();
6358 gen_jmp(s, s->pc - s->cs_base);
6359 }
6360 break;
6361 case 0xee:
6362 case 0xef:
6363 ot = mo_b_d32(b, dflag);
6364 tcg_gen_ext16u_tl(cpu_T[0], cpu_regs[R_EDX]);
6365 gen_check_io(s, ot, pc_start - s->cs_base,
6366 svm_is_rep(prefixes));
6367 gen_op_mov_v_reg(ot, cpu_T[1], R_EAX);
6368
6369 if (use_icount)
6370 gen_io_start();
6371 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
6372 tcg_gen_trunc_tl_i32(cpu_tmp3_i32, cpu_T[1]);
6373 gen_helper_out_func(ot, cpu_tmp2_i32, cpu_tmp3_i32);
6374 if (use_icount) {
6375 gen_io_end();
6376 gen_jmp(s, s->pc - s->cs_base);
6377 }
6378 break;
6379
6380 /************************/
6381 /* control */
6382 case 0xc2: /* ret im */
6383 val = cpu_ldsw_code(env, s->pc);
6384 s->pc += 2;
6385 ot = gen_pop_T0(s);
6386 gen_stack_update(s, val + (1 << ot));
6387 /* Note that gen_pop_T0 uses a zero-extending load. */
6388 gen_op_jmp_v(cpu_T[0]);
6389 gen_eob(s);
6390 break;
6391 case 0xc3: /* ret */
6392 ot = gen_pop_T0(s);
6393 gen_pop_update(s, ot);
6394 /* Note that gen_pop_T0 uses a zero-extending load. */
6395 gen_op_jmp_v(cpu_T[0]);
6396 gen_eob(s);
6397 break;
6398 case 0xca: /* lret im */
6399 val = cpu_ldsw_code(env, s->pc);
6400 s->pc += 2;
6401 do_lret:
6402 if (s->pe && !s->vm86) {
6403 gen_update_cc_op(s);
6404 gen_jmp_im(pc_start - s->cs_base);
6405 gen_helper_lret_protected(cpu_env, tcg_const_i32(dflag - 1),
6406 tcg_const_i32(val));
6407 } else {
6408 gen_stack_A0(s);
6409 /* pop offset */
6410 gen_op_ld_v(s, dflag, cpu_T[0], cpu_A0);
6411 /* NOTE: keeping EIP updated is not a problem in case of
6412 exception */
6413 gen_op_jmp_v(cpu_T[0]);
6414 /* pop selector */
6415 gen_op_addl_A0_im(1 << dflag);
6416 gen_op_ld_v(s, dflag, cpu_T[0], cpu_A0);
6417 gen_op_movl_seg_T0_vm(R_CS);
6418 /* add stack offset */
6419 gen_stack_update(s, val + (2 << dflag));
6420 }
6421 gen_eob(s);
6422 break;
6423 case 0xcb: /* lret */
6424 val = 0;
6425 goto do_lret;
6426 case 0xcf: /* iret */
6427 gen_svm_check_intercept(s, pc_start, SVM_EXIT_IRET);
6428 if (!s->pe) {
6429 /* real mode */
6430 gen_helper_iret_real(cpu_env, tcg_const_i32(dflag - 1));
6431 set_cc_op(s, CC_OP_EFLAGS);
6432 } else if (s->vm86) {
6433 if (s->iopl != 3) {
6434 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
6435 } else {
6436 gen_helper_iret_real(cpu_env, tcg_const_i32(dflag - 1));
6437 set_cc_op(s, CC_OP_EFLAGS);
6438 }
6439 } else {
6440 gen_update_cc_op(s);
6441 gen_jmp_im(pc_start - s->cs_base);
6442 gen_helper_iret_protected(cpu_env, tcg_const_i32(dflag - 1),
6443 tcg_const_i32(s->pc - s->cs_base));
6444 set_cc_op(s, CC_OP_EFLAGS);
6445 }
6446 gen_eob(s);
6447 break;
6448 case 0xe8: /* call im */
6449 {
6450 if (dflag != MO_16) {
6451 tval = (int32_t)insn_get(env, s, MO_32);
6452 } else {
6453 tval = (int16_t)insn_get(env, s, MO_16);
6454 }
6455 next_eip = s->pc - s->cs_base;
6456 tval += next_eip;
6457 if (dflag == MO_16) {
6458 tval &= 0xffff;
6459 } else if (!CODE64(s)) {
6460 tval &= 0xffffffff;
6461 }
6462 tcg_gen_movi_tl(cpu_T[0], next_eip);
6463 gen_push_v(s, cpu_T[0]);
6464 gen_jmp(s, tval);
6465 }
6466 break;
6467 case 0x9a: /* lcall im */
6468 {
6469 unsigned int selector, offset;
6470
6471 if (CODE64(s))
6472 goto illegal_op;
6473 ot = dflag;
6474 offset = insn_get(env, s, ot);
6475 selector = insn_get(env, s, MO_16);
6476
6477 tcg_gen_movi_tl(cpu_T[0], selector);
6478 tcg_gen_movi_tl(cpu_T[1], offset);
6479 }
6480 goto do_lcall;
6481 case 0xe9: /* jmp im */
6482 if (dflag != MO_16) {
6483 tval = (int32_t)insn_get(env, s, MO_32);
6484 } else {
6485 tval = (int16_t)insn_get(env, s, MO_16);
6486 }
6487 tval += s->pc - s->cs_base;
6488 if (dflag == MO_16) {
6489 tval &= 0xffff;
6490 } else if (!CODE64(s)) {
6491 tval &= 0xffffffff;
6492 }
6493 gen_jmp(s, tval);
6494 break;
6495 case 0xea: /* ljmp im */
6496 {
6497 unsigned int selector, offset;
6498
6499 if (CODE64(s))
6500 goto illegal_op;
6501 ot = dflag;
6502 offset = insn_get(env, s, ot);
6503 selector = insn_get(env, s, MO_16);
6504
6505 tcg_gen_movi_tl(cpu_T[0], selector);
6506 tcg_gen_movi_tl(cpu_T[1], offset);
6507 }
6508 goto do_ljmp;
6509 case 0xeb: /* jmp Jb */
6510 tval = (int8_t)insn_get(env, s, MO_8);
6511 tval += s->pc - s->cs_base;
6512 if (dflag == MO_16) {
6513 tval &= 0xffff;
6514 }
6515 gen_jmp(s, tval);
6516 break;
6517 case 0x70 ... 0x7f: /* jcc Jb */
6518 tval = (int8_t)insn_get(env, s, MO_8);
6519 goto do_jcc;
6520 case 0x180 ... 0x18f: /* jcc Jv */
6521 if (dflag != MO_16) {
6522 tval = (int32_t)insn_get(env, s, MO_32);
6523 } else {
6524 tval = (int16_t)insn_get(env, s, MO_16);
6525 }
6526 do_jcc:
6527 next_eip = s->pc - s->cs_base;
6528 tval += next_eip;
6529 if (dflag == MO_16) {
6530 tval &= 0xffff;
6531 }
6532 gen_jcc(s, b, tval, next_eip);
6533 break;
6534
6535 case 0x190 ... 0x19f: /* setcc Gv */
6536 modrm = cpu_ldub_code(env, s->pc++);
6537 gen_setcc1(s, b, cpu_T[0]);
6538 gen_ldst_modrm(env, s, modrm, MO_8, OR_TMP0, 1);
6539 break;
6540 case 0x140 ... 0x14f: /* cmov Gv, Ev */
6541 if (!(s->cpuid_features & CPUID_CMOV)) {
6542 goto illegal_op;
6543 }
6544 ot = dflag;
6545 modrm = cpu_ldub_code(env, s->pc++);
6546 reg = ((modrm >> 3) & 7) | rex_r;
6547 gen_cmovcc1(env, s, ot, b, modrm, reg);
6548 break;
6549
6550 /************************/
6551 /* flags */
6552 case 0x9c: /* pushf */
6553 gen_svm_check_intercept(s, pc_start, SVM_EXIT_PUSHF);
6554 if (s->vm86 && s->iopl != 3) {
6555 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
6556 } else {
6557 gen_update_cc_op(s);
6558 gen_helper_read_eflags(cpu_T[0], cpu_env);
6559 gen_push_v(s, cpu_T[0]);
6560 }
6561 break;
6562 case 0x9d: /* popf */
6563 gen_svm_check_intercept(s, pc_start, SVM_EXIT_POPF);
6564 if (s->vm86 && s->iopl != 3) {
6565 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
6566 } else {
6567 ot = gen_pop_T0(s);
6568 if (s->cpl == 0) {
6569 if (dflag != MO_16) {
6570 gen_helper_write_eflags(cpu_env, cpu_T[0],
6571 tcg_const_i32((TF_MASK | AC_MASK |
6572 ID_MASK | NT_MASK |
6573 IF_MASK |
6574 IOPL_MASK)));
6575 } else {
6576 gen_helper_write_eflags(cpu_env, cpu_T[0],
6577 tcg_const_i32((TF_MASK | AC_MASK |
6578 ID_MASK | NT_MASK |
6579 IF_MASK | IOPL_MASK)
6580 & 0xffff));
6581 }
6582 } else {
6583 if (s->cpl <= s->iopl) {
6584 if (dflag != MO_16) {
6585 gen_helper_write_eflags(cpu_env, cpu_T[0],
6586 tcg_const_i32((TF_MASK |
6587 AC_MASK |
6588 ID_MASK |
6589 NT_MASK |
6590 IF_MASK)));
6591 } else {
6592 gen_helper_write_eflags(cpu_env, cpu_T[0],
6593 tcg_const_i32((TF_MASK |
6594 AC_MASK |
6595 ID_MASK |
6596 NT_MASK |
6597 IF_MASK)
6598 & 0xffff));
6599 }
6600 } else {
6601 if (dflag != MO_16) {
6602 gen_helper_write_eflags(cpu_env, cpu_T[0],
6603 tcg_const_i32((TF_MASK | AC_MASK |
6604 ID_MASK | NT_MASK)));
6605 } else {
6606 gen_helper_write_eflags(cpu_env, cpu_T[0],
6607 tcg_const_i32((TF_MASK | AC_MASK |
6608 ID_MASK | NT_MASK)
6609 & 0xffff));
6610 }
6611 }
6612 }
6613 gen_pop_update(s, ot);
6614 set_cc_op(s, CC_OP_EFLAGS);
6615 /* abort translation because TF/AC flag may change */
6616 gen_jmp_im(s->pc - s->cs_base);
6617 gen_eob(s);
6618 }
6619 break;
6620 case 0x9e: /* sahf */
6621 if (CODE64(s) && !(s->cpuid_ext3_features & CPUID_EXT3_LAHF_LM))
6622 goto illegal_op;
6623 gen_op_mov_v_reg(MO_8, cpu_T[0], R_AH);
6624 gen_compute_eflags(s);
6625 tcg_gen_andi_tl(cpu_cc_src, cpu_cc_src, CC_O);
6626 tcg_gen_andi_tl(cpu_T[0], cpu_T[0], CC_S | CC_Z | CC_A | CC_P | CC_C);
6627 tcg_gen_or_tl(cpu_cc_src, cpu_cc_src, cpu_T[0]);
6628 break;
6629 case 0x9f: /* lahf */
6630 if (CODE64(s) && !(s->cpuid_ext3_features & CPUID_EXT3_LAHF_LM))
6631 goto illegal_op;
6632 gen_compute_eflags(s);
6633 /* Note: gen_compute_eflags() only gives the condition codes */
6634 tcg_gen_ori_tl(cpu_T[0], cpu_cc_src, 0x02);
6635 gen_op_mov_reg_v(MO_8, R_AH, cpu_T[0]);
6636 break;
6637 case 0xf5: /* cmc */
6638 gen_compute_eflags(s);
6639 tcg_gen_xori_tl(cpu_cc_src, cpu_cc_src, CC_C);
6640 break;
6641 case 0xf8: /* clc */
6642 gen_compute_eflags(s);
6643 tcg_gen_andi_tl(cpu_cc_src, cpu_cc_src, ~CC_C);
6644 break;
6645 case 0xf9: /* stc */
6646 gen_compute_eflags(s);
6647 tcg_gen_ori_tl(cpu_cc_src, cpu_cc_src, CC_C);
6648 break;
6649 case 0xfc: /* cld */
6650 tcg_gen_movi_i32(cpu_tmp2_i32, 1);
6651 tcg_gen_st_i32(cpu_tmp2_i32, cpu_env, offsetof(CPUX86State, df));
6652 break;
6653 case 0xfd: /* std */
6654 tcg_gen_movi_i32(cpu_tmp2_i32, -1);
6655 tcg_gen_st_i32(cpu_tmp2_i32, cpu_env, offsetof(CPUX86State, df));
6656 break;
6657
6658 /************************/
6659 /* bit operations */
6660 case 0x1ba: /* bt/bts/btr/btc Gv, im */
6661 ot = dflag;
6662 modrm = cpu_ldub_code(env, s->pc++);
6663 op = (modrm >> 3) & 7;
6664 mod = (modrm >> 6) & 3;
6665 rm = (modrm & 7) | REX_B(s);
6666 if (mod != 3) {
6667 s->rip_offset = 1;
6668 gen_lea_modrm(env, s, modrm);
6669 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
6670 } else {
6671 gen_op_mov_v_reg(ot, cpu_T[0], rm);
6672 }
6673 /* load shift */
6674 val = cpu_ldub_code(env, s->pc++);
6675 tcg_gen_movi_tl(cpu_T[1], val);
6676 if (op < 4)
6677 goto illegal_op;
6678 op -= 4;
6679 goto bt_op;
6680 case 0x1a3: /* bt Gv, Ev */
6681 op = 0;
6682 goto do_btx;
6683 case 0x1ab: /* bts */
6684 op = 1;
6685 goto do_btx;
6686 case 0x1b3: /* btr */
6687 op = 2;
6688 goto do_btx;
6689 case 0x1bb: /* btc */
6690 op = 3;
6691 do_btx:
6692 ot = dflag;
6693 modrm = cpu_ldub_code(env, s->pc++);
6694 reg = ((modrm >> 3) & 7) | rex_r;
6695 mod = (modrm >> 6) & 3;
6696 rm = (modrm & 7) | REX_B(s);
6697 gen_op_mov_v_reg(MO_32, cpu_T[1], reg);
6698 if (mod != 3) {
6699 gen_lea_modrm(env, s, modrm);
6700 /* specific case: we need to add a displacement */
6701 gen_exts(ot, cpu_T[1]);
6702 tcg_gen_sari_tl(cpu_tmp0, cpu_T[1], 3 + ot);
6703 tcg_gen_shli_tl(cpu_tmp0, cpu_tmp0, ot);
6704 tcg_gen_add_tl(cpu_A0, cpu_A0, cpu_tmp0);
6705 gen_op_ld_v(s, ot, cpu_T[0], cpu_A0);
6706 } else {
6707 gen_op_mov_v_reg(ot, cpu_T[0], rm);
6708 }
6709 bt_op:
6710 tcg_gen_andi_tl(cpu_T[1], cpu_T[1], (1 << (3 + ot)) - 1);
6711 switch(op) {
6712 case 0:
6713 tcg_gen_shr_tl(cpu_cc_src, cpu_T[0], cpu_T[1]);
6714 tcg_gen_movi_tl(cpu_cc_dst, 0);
6715 break;
6716 case 1:
6717 tcg_gen_shr_tl(cpu_tmp4, cpu_T[0], cpu_T[1]);
6718 tcg_gen_movi_tl(cpu_tmp0, 1);
6719 tcg_gen_shl_tl(cpu_tmp0, cpu_tmp0, cpu_T[1]);
6720 tcg_gen_or_tl(cpu_T[0], cpu_T[0], cpu_tmp0);
6721 break;
6722 case 2:
6723 tcg_gen_shr_tl(cpu_tmp4, cpu_T[0], cpu_T[1]);
6724 tcg_gen_movi_tl(cpu_tmp0, 1);
6725 tcg_gen_shl_tl(cpu_tmp0, cpu_tmp0, cpu_T[1]);
6726 tcg_gen_not_tl(cpu_tmp0, cpu_tmp0);
6727 tcg_gen_and_tl(cpu_T[0], cpu_T[0], cpu_tmp0);
6728 break;
6729 default:
6730 case 3:
6731 tcg_gen_shr_tl(cpu_tmp4, cpu_T[0], cpu_T[1]);
6732 tcg_gen_movi_tl(cpu_tmp0, 1);
6733 tcg_gen_shl_tl(cpu_tmp0, cpu_tmp0, cpu_T[1]);
6734 tcg_gen_xor_tl(cpu_T[0], cpu_T[0], cpu_tmp0);
6735 break;
6736 }
6737 set_cc_op(s, CC_OP_SARB + ot);
6738 if (op != 0) {
6739 if (mod != 3) {
6740 gen_op_st_v(s, ot, cpu_T[0], cpu_A0);
6741 } else {
6742 gen_op_mov_reg_v(ot, rm, cpu_T[0]);
6743 }
6744 tcg_gen_mov_tl(cpu_cc_src, cpu_tmp4);
6745 tcg_gen_movi_tl(cpu_cc_dst, 0);
6746 }
6747 break;
6748 case 0x1bc: /* bsf / tzcnt */
6749 case 0x1bd: /* bsr / lzcnt */
6750 ot = dflag;
6751 modrm = cpu_ldub_code(env, s->pc++);
6752 reg = ((modrm >> 3) & 7) | rex_r;
6753 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
6754 gen_extu(ot, cpu_T[0]);
6755
6756 /* Note that lzcnt and tzcnt are in different extensions. */
6757 if ((prefixes & PREFIX_REPZ)
6758 && (b & 1
6759 ? s->cpuid_ext3_features & CPUID_EXT3_ABM
6760 : s->cpuid_7_0_ebx_features & CPUID_7_0_EBX_BMI1)) {
6761 int size = 8 << ot;
6762 tcg_gen_mov_tl(cpu_cc_src, cpu_T[0]);
6763 if (b & 1) {
6764 /* For lzcnt, reduce the target_ulong result by the
6765 number of zeros that we expect to find at the top. */
6766 gen_helper_clz(cpu_T[0], cpu_T[0]);
6767 tcg_gen_subi_tl(cpu_T[0], cpu_T[0], TARGET_LONG_BITS - size);
6768 } else {
6769 /* For tzcnt, a zero input must return the operand size:
6770 force all bits outside the operand size to 1. */
6771 target_ulong mask = (target_ulong)-2 << (size - 1);
6772 tcg_gen_ori_tl(cpu_T[0], cpu_T[0], mask);
6773 gen_helper_ctz(cpu_T[0], cpu_T[0]);
6774 }
6775 /* For lzcnt/tzcnt, C and Z bits are defined and are
6776 related to the result. */
6777 gen_op_update1_cc();
6778 set_cc_op(s, CC_OP_BMILGB + ot);
6779 } else {
6780 /* For bsr/bsf, only the Z bit is defined and it is related
6781 to the input and not the result. */
6782 tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
6783 set_cc_op(s, CC_OP_LOGICB + ot);
6784 if (b & 1) {
6785 /* For bsr, return the bit index of the first 1 bit,
6786 not the count of leading zeros. */
6787 gen_helper_clz(cpu_T[0], cpu_T[0]);
6788 tcg_gen_xori_tl(cpu_T[0], cpu_T[0], TARGET_LONG_BITS - 1);
6789 } else {
6790 gen_helper_ctz(cpu_T[0], cpu_T[0]);
6791 }
6792 /* ??? The manual says that the output is undefined when the
6793 input is zero, but real hardware leaves it unchanged, and
6794 real programs appear to depend on that. */
6795 tcg_gen_movi_tl(cpu_tmp0, 0);
6796 tcg_gen_movcond_tl(TCG_COND_EQ, cpu_T[0], cpu_cc_dst, cpu_tmp0,
6797 cpu_regs[reg], cpu_T[0]);
6798 }
6799 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
6800 break;
6801 /************************/
6802 /* bcd */
6803 case 0x27: /* daa */
6804 if (CODE64(s))
6805 goto illegal_op;
6806 gen_update_cc_op(s);
6807 gen_helper_daa(cpu_env);
6808 set_cc_op(s, CC_OP_EFLAGS);
6809 break;
6810 case 0x2f: /* das */
6811 if (CODE64(s))
6812 goto illegal_op;
6813 gen_update_cc_op(s);
6814 gen_helper_das(cpu_env);
6815 set_cc_op(s, CC_OP_EFLAGS);
6816 break;
6817 case 0x37: /* aaa */
6818 if (CODE64(s))
6819 goto illegal_op;
6820 gen_update_cc_op(s);
6821 gen_helper_aaa(cpu_env);
6822 set_cc_op(s, CC_OP_EFLAGS);
6823 break;
6824 case 0x3f: /* aas */
6825 if (CODE64(s))
6826 goto illegal_op;
6827 gen_update_cc_op(s);
6828 gen_helper_aas(cpu_env);
6829 set_cc_op(s, CC_OP_EFLAGS);
6830 break;
6831 case 0xd4: /* aam */
6832 if (CODE64(s))
6833 goto illegal_op;
6834 val = cpu_ldub_code(env, s->pc++);
6835 if (val == 0) {
6836 gen_exception(s, EXCP00_DIVZ, pc_start - s->cs_base);
6837 } else {
6838 gen_helper_aam(cpu_env, tcg_const_i32(val));
6839 set_cc_op(s, CC_OP_LOGICB);
6840 }
6841 break;
6842 case 0xd5: /* aad */
6843 if (CODE64(s))
6844 goto illegal_op;
6845 val = cpu_ldub_code(env, s->pc++);
6846 gen_helper_aad(cpu_env, tcg_const_i32(val));
6847 set_cc_op(s, CC_OP_LOGICB);
6848 break;
6849 /************************/
6850 /* misc */
6851 case 0x90: /* nop */
6852 /* XXX: correct lock test for all insn */
6853 if (prefixes & PREFIX_LOCK) {
6854 goto illegal_op;
6855 }
6856 /* If REX_B is set, then this is xchg eax, r8d, not a nop. */
6857 if (REX_B(s)) {
6858 goto do_xchg_reg_eax;
6859 }
6860 if (prefixes & PREFIX_REPZ) {
6861 gen_update_cc_op(s);
6862 gen_jmp_im(pc_start - s->cs_base);
6863 gen_helper_pause(cpu_env, tcg_const_i32(s->pc - pc_start));
6864 s->is_jmp = DISAS_TB_JUMP;
6865 }
6866 break;
6867 case 0x9b: /* fwait */
6868 if ((s->flags & (HF_MP_MASK | HF_TS_MASK)) ==
6869 (HF_MP_MASK | HF_TS_MASK)) {
6870 gen_exception(s, EXCP07_PREX, pc_start - s->cs_base);
6871 } else {
6872 gen_update_cc_op(s);
6873 gen_jmp_im(pc_start - s->cs_base);
6874 gen_helper_fwait(cpu_env);
6875 }
6876 break;
6877 case 0xcc: /* int3 */
6878 gen_interrupt(s, EXCP03_INT3, pc_start - s->cs_base, s->pc - s->cs_base);
6879 break;
6880 case 0xcd: /* int N */
6881 val = cpu_ldub_code(env, s->pc++);
6882 if (s->vm86 && s->iopl != 3) {
6883 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
6884 } else {
6885 gen_interrupt(s, val, pc_start - s->cs_base, s->pc - s->cs_base);
6886 }
6887 break;
6888 case 0xce: /* into */
6889 if (CODE64(s))
6890 goto illegal_op;
6891 gen_update_cc_op(s);
6892 gen_jmp_im(pc_start - s->cs_base);
6893 gen_helper_into(cpu_env, tcg_const_i32(s->pc - pc_start));
6894 break;
6895 #ifdef WANT_ICEBP
6896 case 0xf1: /* icebp (undocumented, exits to external debugger) */
6897 gen_svm_check_intercept(s, pc_start, SVM_EXIT_ICEBP);
6898 #if 1
6899 gen_debug(s, pc_start - s->cs_base);
6900 #else
6901 /* start debug */
6902 tb_flush(env);
6903 qemu_set_log(CPU_LOG_INT | CPU_LOG_TB_IN_ASM);
6904 #endif
6905 break;
6906 #endif
6907 case 0xfa: /* cli */
6908 if (!s->vm86) {
6909 if (s->cpl <= s->iopl) {
6910 gen_helper_cli(cpu_env);
6911 } else {
6912 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
6913 }
6914 } else {
6915 if (s->iopl == 3) {
6916 gen_helper_cli(cpu_env);
6917 } else {
6918 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
6919 }
6920 }
6921 break;
6922 case 0xfb: /* sti */
6923 if (!s->vm86) {
6924 if (s->cpl <= s->iopl) {
6925 gen_sti:
6926 gen_helper_sti(cpu_env);
6927 /* interruptions are enabled only the first insn after sti */
6928 /* If several instructions disable interrupts, only the
6929 _first_ does it */
6930 if (!(s->tb->flags & HF_INHIBIT_IRQ_MASK))
6931 gen_helper_set_inhibit_irq(cpu_env);
6932 /* give a chance to handle pending irqs */
6933 gen_jmp_im(s->pc - s->cs_base);
6934 gen_eob(s);
6935 } else {
6936 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
6937 }
6938 } else {
6939 if (s->iopl == 3) {
6940 goto gen_sti;
6941 } else {
6942 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
6943 }
6944 }
6945 break;
6946 case 0x62: /* bound */
6947 if (CODE64(s))
6948 goto illegal_op;
6949 ot = dflag;
6950 modrm = cpu_ldub_code(env, s->pc++);
6951 reg = (modrm >> 3) & 7;
6952 mod = (modrm >> 6) & 3;
6953 if (mod == 3)
6954 goto illegal_op;
6955 gen_op_mov_v_reg(ot, cpu_T[0], reg);
6956 gen_lea_modrm(env, s, modrm);
6957 gen_jmp_im(pc_start - s->cs_base);
6958 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
6959 if (ot == MO_16) {
6960 gen_helper_boundw(cpu_env, cpu_A0, cpu_tmp2_i32);
6961 } else {
6962 gen_helper_boundl(cpu_env, cpu_A0, cpu_tmp2_i32);
6963 }
6964 break;
6965 case 0x1c8 ... 0x1cf: /* bswap reg */
6966 reg = (b & 7) | REX_B(s);
6967 #ifdef TARGET_X86_64
6968 if (dflag == MO_64) {
6969 gen_op_mov_v_reg(MO_64, cpu_T[0], reg);
6970 tcg_gen_bswap64_i64(cpu_T[0], cpu_T[0]);
6971 gen_op_mov_reg_v(MO_64, reg, cpu_T[0]);
6972 } else
6973 #endif
6974 {
6975 gen_op_mov_v_reg(MO_32, cpu_T[0], reg);
6976 tcg_gen_ext32u_tl(cpu_T[0], cpu_T[0]);
6977 tcg_gen_bswap32_tl(cpu_T[0], cpu_T[0]);
6978 gen_op_mov_reg_v(MO_32, reg, cpu_T[0]);
6979 }
6980 break;
6981 case 0xd6: /* salc */
6982 if (CODE64(s))
6983 goto illegal_op;
6984 gen_compute_eflags_c(s, cpu_T[0]);
6985 tcg_gen_neg_tl(cpu_T[0], cpu_T[0]);
6986 gen_op_mov_reg_v(MO_8, R_EAX, cpu_T[0]);
6987 break;
6988 case 0xe0: /* loopnz */
6989 case 0xe1: /* loopz */
6990 case 0xe2: /* loop */
6991 case 0xe3: /* jecxz */
6992 {
6993 int l1, l2, l3;
6994
6995 tval = (int8_t)insn_get(env, s, MO_8);
6996 next_eip = s->pc - s->cs_base;
6997 tval += next_eip;
6998 if (dflag == MO_16) {
6999 tval &= 0xffff;
7000 }
7001
7002 l1 = gen_new_label();
7003 l2 = gen_new_label();
7004 l3 = gen_new_label();
7005 b &= 3;
7006 switch(b) {
7007 case 0: /* loopnz */
7008 case 1: /* loopz */
7009 gen_op_add_reg_im(s->aflag, R_ECX, -1);
7010 gen_op_jz_ecx(s->aflag, l3);
7011 gen_jcc1(s, (JCC_Z << 1) | (b ^ 1), l1);
7012 break;
7013 case 2: /* loop */
7014 gen_op_add_reg_im(s->aflag, R_ECX, -1);
7015 gen_op_jnz_ecx(s->aflag, l1);
7016 break;
7017 default:
7018 case 3: /* jcxz */
7019 gen_op_jz_ecx(s->aflag, l1);
7020 break;
7021 }
7022
7023 gen_set_label(l3);
7024 gen_jmp_im(next_eip);
7025 tcg_gen_br(l2);
7026
7027 gen_set_label(l1);
7028 gen_jmp_im(tval);
7029 gen_set_label(l2);
7030 gen_eob(s);
7031 }
7032 break;
7033 case 0x130: /* wrmsr */
7034 case 0x132: /* rdmsr */
7035 if (s->cpl != 0) {
7036 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7037 } else {
7038 gen_update_cc_op(s);
7039 gen_jmp_im(pc_start - s->cs_base);
7040 if (b & 2) {
7041 gen_helper_rdmsr(cpu_env);
7042 } else {
7043 gen_helper_wrmsr(cpu_env);
7044 }
7045 }
7046 break;
7047 case 0x131: /* rdtsc */
7048 gen_update_cc_op(s);
7049 gen_jmp_im(pc_start - s->cs_base);
7050 if (use_icount)
7051 gen_io_start();
7052 gen_helper_rdtsc(cpu_env);
7053 if (use_icount) {
7054 gen_io_end();
7055 gen_jmp(s, s->pc - s->cs_base);
7056 }
7057 break;
7058 case 0x133: /* rdpmc */
7059 gen_update_cc_op(s);
7060 gen_jmp_im(pc_start - s->cs_base);
7061 gen_helper_rdpmc(cpu_env);
7062 break;
7063 case 0x134: /* sysenter */
7064 /* For Intel SYSENTER is valid on 64-bit */
7065 if (CODE64(s) && env->cpuid_vendor1 != CPUID_VENDOR_INTEL_1)
7066 goto illegal_op;
7067 if (!s->pe) {
7068 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7069 } else {
7070 gen_update_cc_op(s);
7071 gen_jmp_im(pc_start - s->cs_base);
7072 gen_helper_sysenter(cpu_env);
7073 gen_eob(s);
7074 }
7075 break;
7076 case 0x135: /* sysexit */
7077 /* For Intel SYSEXIT is valid on 64-bit */
7078 if (CODE64(s) && env->cpuid_vendor1 != CPUID_VENDOR_INTEL_1)
7079 goto illegal_op;
7080 if (!s->pe) {
7081 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7082 } else {
7083 gen_update_cc_op(s);
7084 gen_jmp_im(pc_start - s->cs_base);
7085 gen_helper_sysexit(cpu_env, tcg_const_i32(dflag - 1));
7086 gen_eob(s);
7087 }
7088 break;
7089 #ifdef TARGET_X86_64
7090 case 0x105: /* syscall */
7091 /* XXX: is it usable in real mode ? */
7092 gen_update_cc_op(s);
7093 gen_jmp_im(pc_start - s->cs_base);
7094 gen_helper_syscall(cpu_env, tcg_const_i32(s->pc - pc_start));
7095 gen_eob(s);
7096 break;
7097 case 0x107: /* sysret */
7098 if (!s->pe) {
7099 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7100 } else {
7101 gen_update_cc_op(s);
7102 gen_jmp_im(pc_start - s->cs_base);
7103 gen_helper_sysret(cpu_env, tcg_const_i32(dflag - 1));
7104 /* condition codes are modified only in long mode */
7105 if (s->lma) {
7106 set_cc_op(s, CC_OP_EFLAGS);
7107 }
7108 gen_eob(s);
7109 }
7110 break;
7111 #endif
7112 case 0x1a2: /* cpuid */
7113 gen_update_cc_op(s);
7114 gen_jmp_im(pc_start - s->cs_base);
7115 gen_helper_cpuid(cpu_env);
7116 break;
7117 case 0xf4: /* hlt */
7118 if (s->cpl != 0) {
7119 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7120 } else {
7121 gen_update_cc_op(s);
7122 gen_jmp_im(pc_start - s->cs_base);
7123 gen_helper_hlt(cpu_env, tcg_const_i32(s->pc - pc_start));
7124 s->is_jmp = DISAS_TB_JUMP;
7125 }
7126 break;
7127 case 0x100:
7128 modrm = cpu_ldub_code(env, s->pc++);
7129 mod = (modrm >> 6) & 3;
7130 op = (modrm >> 3) & 7;
7131 switch(op) {
7132 case 0: /* sldt */
7133 if (!s->pe || s->vm86)
7134 goto illegal_op;
7135 gen_svm_check_intercept(s, pc_start, SVM_EXIT_LDTR_READ);
7136 tcg_gen_ld32u_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,ldt.selector));
7137 ot = mod == 3 ? dflag : MO_16;
7138 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 1);
7139 break;
7140 case 2: /* lldt */
7141 if (!s->pe || s->vm86)
7142 goto illegal_op;
7143 if (s->cpl != 0) {
7144 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7145 } else {
7146 gen_svm_check_intercept(s, pc_start, SVM_EXIT_LDTR_WRITE);
7147 gen_ldst_modrm(env, s, modrm, MO_16, OR_TMP0, 0);
7148 gen_jmp_im(pc_start - s->cs_base);
7149 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
7150 gen_helper_lldt(cpu_env, cpu_tmp2_i32);
7151 }
7152 break;
7153 case 1: /* str */
7154 if (!s->pe || s->vm86)
7155 goto illegal_op;
7156 gen_svm_check_intercept(s, pc_start, SVM_EXIT_TR_READ);
7157 tcg_gen_ld32u_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,tr.selector));
7158 ot = mod == 3 ? dflag : MO_16;
7159 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 1);
7160 break;
7161 case 3: /* ltr */
7162 if (!s->pe || s->vm86)
7163 goto illegal_op;
7164 if (s->cpl != 0) {
7165 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7166 } else {
7167 gen_svm_check_intercept(s, pc_start, SVM_EXIT_TR_WRITE);
7168 gen_ldst_modrm(env, s, modrm, MO_16, OR_TMP0, 0);
7169 gen_jmp_im(pc_start - s->cs_base);
7170 tcg_gen_trunc_tl_i32(cpu_tmp2_i32, cpu_T[0]);
7171 gen_helper_ltr(cpu_env, cpu_tmp2_i32);
7172 }
7173 break;
7174 case 4: /* verr */
7175 case 5: /* verw */
7176 if (!s->pe || s->vm86)
7177 goto illegal_op;
7178 gen_ldst_modrm(env, s, modrm, MO_16, OR_TMP0, 0);
7179 gen_update_cc_op(s);
7180 if (op == 4) {
7181 gen_helper_verr(cpu_env, cpu_T[0]);
7182 } else {
7183 gen_helper_verw(cpu_env, cpu_T[0]);
7184 }
7185 set_cc_op(s, CC_OP_EFLAGS);
7186 break;
7187 default:
7188 goto illegal_op;
7189 }
7190 break;
7191 case 0x101:
7192 modrm = cpu_ldub_code(env, s->pc++);
7193 mod = (modrm >> 6) & 3;
7194 op = (modrm >> 3) & 7;
7195 rm = modrm & 7;
7196 switch(op) {
7197 case 0: /* sgdt */
7198 if (mod == 3)
7199 goto illegal_op;
7200 gen_svm_check_intercept(s, pc_start, SVM_EXIT_GDTR_READ);
7201 gen_lea_modrm(env, s, modrm);
7202 tcg_gen_ld32u_tl(cpu_T[0], cpu_env, offsetof(CPUX86State, gdt.limit));
7203 gen_op_st_v(s, MO_16, cpu_T[0], cpu_A0);
7204 gen_add_A0_im(s, 2);
7205 tcg_gen_ld_tl(cpu_T[0], cpu_env, offsetof(CPUX86State, gdt.base));
7206 if (dflag == MO_16) {
7207 tcg_gen_andi_tl(cpu_T[0], cpu_T[0], 0xffffff);
7208 }
7209 gen_op_st_v(s, CODE64(s) + MO_32, cpu_T[0], cpu_A0);
7210 break;
7211 case 1:
7212 if (mod == 3) {
7213 switch (rm) {
7214 case 0: /* monitor */
7215 if (!(s->cpuid_ext_features & CPUID_EXT_MONITOR) ||
7216 s->cpl != 0)
7217 goto illegal_op;
7218 gen_update_cc_op(s);
7219 gen_jmp_im(pc_start - s->cs_base);
7220 tcg_gen_mov_tl(cpu_A0, cpu_regs[R_EAX]);
7221 gen_extu(s->aflag, cpu_A0);
7222 gen_add_A0_ds_seg(s);
7223 gen_helper_monitor(cpu_env, cpu_A0);
7224 break;
7225 case 1: /* mwait */
7226 if (!(s->cpuid_ext_features & CPUID_EXT_MONITOR) ||
7227 s->cpl != 0)
7228 goto illegal_op;
7229 gen_update_cc_op(s);
7230 gen_jmp_im(pc_start - s->cs_base);
7231 gen_helper_mwait(cpu_env, tcg_const_i32(s->pc - pc_start));
7232 gen_eob(s);
7233 break;
7234 case 2: /* clac */
7235 if (!(s->cpuid_7_0_ebx_features & CPUID_7_0_EBX_SMAP) ||
7236 s->cpl != 0) {
7237 goto illegal_op;
7238 }
7239 gen_helper_clac(cpu_env);
7240 gen_jmp_im(s->pc - s->cs_base);
7241 gen_eob(s);
7242 break;
7243 case 3: /* stac */
7244 if (!(s->cpuid_7_0_ebx_features & CPUID_7_0_EBX_SMAP) ||
7245 s->cpl != 0) {
7246 goto illegal_op;
7247 }
7248 gen_helper_stac(cpu_env);
7249 gen_jmp_im(s->pc - s->cs_base);
7250 gen_eob(s);
7251 break;
7252 default:
7253 goto illegal_op;
7254 }
7255 } else { /* sidt */
7256 gen_svm_check_intercept(s, pc_start, SVM_EXIT_IDTR_READ);
7257 gen_lea_modrm(env, s, modrm);
7258 tcg_gen_ld32u_tl(cpu_T[0], cpu_env, offsetof(CPUX86State, idt.limit));
7259 gen_op_st_v(s, MO_16, cpu_T[0], cpu_A0);
7260 gen_add_A0_im(s, 2);
7261 tcg_gen_ld_tl(cpu_T[0], cpu_env, offsetof(CPUX86State, idt.base));
7262 if (dflag == MO_16) {
7263 tcg_gen_andi_tl(cpu_T[0], cpu_T[0], 0xffffff);
7264 }
7265 gen_op_st_v(s, CODE64(s) + MO_32, cpu_T[0], cpu_A0);
7266 }
7267 break;
7268 case 2: /* lgdt */
7269 case 3: /* lidt */
7270 if (mod == 3) {
7271 gen_update_cc_op(s);
7272 gen_jmp_im(pc_start - s->cs_base);
7273 switch(rm) {
7274 case 0: /* VMRUN */
7275 if (!(s->flags & HF_SVME_MASK) || !s->pe)
7276 goto illegal_op;
7277 if (s->cpl != 0) {
7278 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7279 break;
7280 } else {
7281 gen_helper_vmrun(cpu_env, tcg_const_i32(s->aflag - 1),
7282 tcg_const_i32(s->pc - pc_start));
7283 tcg_gen_exit_tb(0);
7284 s->is_jmp = DISAS_TB_JUMP;
7285 }
7286 break;
7287 case 1: /* VMMCALL */
7288 if (!(s->flags & HF_SVME_MASK))
7289 goto illegal_op;
7290 gen_helper_vmmcall(cpu_env);
7291 break;
7292 case 2: /* VMLOAD */
7293 if (!(s->flags & HF_SVME_MASK) || !s->pe)
7294 goto illegal_op;
7295 if (s->cpl != 0) {
7296 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7297 break;
7298 } else {
7299 gen_helper_vmload(cpu_env, tcg_const_i32(s->aflag - 1));
7300 }
7301 break;
7302 case 3: /* VMSAVE */
7303 if (!(s->flags & HF_SVME_MASK) || !s->pe)
7304 goto illegal_op;
7305 if (s->cpl != 0) {
7306 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7307 break;
7308 } else {
7309 gen_helper_vmsave(cpu_env, tcg_const_i32(s->aflag - 1));
7310 }
7311 break;
7312 case 4: /* STGI */
7313 if ((!(s->flags & HF_SVME_MASK) &&
7314 !(s->cpuid_ext3_features & CPUID_EXT3_SKINIT)) ||
7315 !s->pe)
7316 goto illegal_op;
7317 if (s->cpl != 0) {
7318 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7319 break;
7320 } else {
7321 gen_helper_stgi(cpu_env);
7322 }
7323 break;
7324 case 5: /* CLGI */
7325 if (!(s->flags & HF_SVME_MASK) || !s->pe)
7326 goto illegal_op;
7327 if (s->cpl != 0) {
7328 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7329 break;
7330 } else {
7331 gen_helper_clgi(cpu_env);
7332 }
7333 break;
7334 case 6: /* SKINIT */
7335 if ((!(s->flags & HF_SVME_MASK) &&
7336 !(s->cpuid_ext3_features & CPUID_EXT3_SKINIT)) ||
7337 !s->pe)
7338 goto illegal_op;
7339 gen_helper_skinit(cpu_env);
7340 break;
7341 case 7: /* INVLPGA */
7342 if (!(s->flags & HF_SVME_MASK) || !s->pe)
7343 goto illegal_op;
7344 if (s->cpl != 0) {
7345 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7346 break;
7347 } else {
7348 gen_helper_invlpga(cpu_env,
7349 tcg_const_i32(s->aflag - 1));
7350 }
7351 break;
7352 default:
7353 goto illegal_op;
7354 }
7355 } else if (s->cpl != 0) {
7356 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7357 } else {
7358 gen_svm_check_intercept(s, pc_start,
7359 op==2 ? SVM_EXIT_GDTR_WRITE : SVM_EXIT_IDTR_WRITE);
7360 gen_lea_modrm(env, s, modrm);
7361 gen_op_ld_v(s, MO_16, cpu_T[1], cpu_A0);
7362 gen_add_A0_im(s, 2);
7363 gen_op_ld_v(s, CODE64(s) + MO_32, cpu_T[0], cpu_A0);
7364 if (dflag == MO_16) {
7365 tcg_gen_andi_tl(cpu_T[0], cpu_T[0], 0xffffff);
7366 }
7367 if (op == 2) {
7368 tcg_gen_st_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,gdt.base));
7369 tcg_gen_st32_tl(cpu_T[1], cpu_env, offsetof(CPUX86State,gdt.limit));
7370 } else {
7371 tcg_gen_st_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,idt.base));
7372 tcg_gen_st32_tl(cpu_T[1], cpu_env, offsetof(CPUX86State,idt.limit));
7373 }
7374 }
7375 break;
7376 case 4: /* smsw */
7377 gen_svm_check_intercept(s, pc_start, SVM_EXIT_READ_CR0);
7378 #if defined TARGET_X86_64 && defined HOST_WORDS_BIGENDIAN
7379 tcg_gen_ld32u_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,cr[0]) + 4);
7380 #else
7381 tcg_gen_ld32u_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,cr[0]));
7382 #endif
7383 gen_ldst_modrm(env, s, modrm, MO_16, OR_TMP0, 1);
7384 break;
7385 case 6: /* lmsw */
7386 if (s->cpl != 0) {
7387 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7388 } else {
7389 gen_svm_check_intercept(s, pc_start, SVM_EXIT_WRITE_CR0);
7390 gen_ldst_modrm(env, s, modrm, MO_16, OR_TMP0, 0);
7391 gen_helper_lmsw(cpu_env, cpu_T[0]);
7392 gen_jmp_im(s->pc - s->cs_base);
7393 gen_eob(s);
7394 }
7395 break;
7396 case 7:
7397 if (mod != 3) { /* invlpg */
7398 if (s->cpl != 0) {
7399 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7400 } else {
7401 gen_update_cc_op(s);
7402 gen_jmp_im(pc_start - s->cs_base);
7403 gen_lea_modrm(env, s, modrm);
7404 gen_helper_invlpg(cpu_env, cpu_A0);
7405 gen_jmp_im(s->pc - s->cs_base);
7406 gen_eob(s);
7407 }
7408 } else {
7409 switch (rm) {
7410 case 0: /* swapgs */
7411 #ifdef TARGET_X86_64
7412 if (CODE64(s)) {
7413 if (s->cpl != 0) {
7414 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7415 } else {
7416 tcg_gen_ld_tl(cpu_T[0], cpu_env,
7417 offsetof(CPUX86State,segs[R_GS].base));
7418 tcg_gen_ld_tl(cpu_T[1], cpu_env,
7419 offsetof(CPUX86State,kernelgsbase));
7420 tcg_gen_st_tl(cpu_T[1], cpu_env,
7421 offsetof(CPUX86State,segs[R_GS].base));
7422 tcg_gen_st_tl(cpu_T[0], cpu_env,
7423 offsetof(CPUX86State,kernelgsbase));
7424 }
7425 } else
7426 #endif
7427 {
7428 goto illegal_op;
7429 }
7430 break;
7431 case 1: /* rdtscp */
7432 if (!(s->cpuid_ext2_features & CPUID_EXT2_RDTSCP))
7433 goto illegal_op;
7434 gen_update_cc_op(s);
7435 gen_jmp_im(pc_start - s->cs_base);
7436 if (use_icount)
7437 gen_io_start();
7438 gen_helper_rdtscp(cpu_env);
7439 if (use_icount) {
7440 gen_io_end();
7441 gen_jmp(s, s->pc - s->cs_base);
7442 }
7443 break;
7444 default:
7445 goto illegal_op;
7446 }
7447 }
7448 break;
7449 default:
7450 goto illegal_op;
7451 }
7452 break;
7453 case 0x108: /* invd */
7454 case 0x109: /* wbinvd */
7455 if (s->cpl != 0) {
7456 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7457 } else {
7458 gen_svm_check_intercept(s, pc_start, (b & 2) ? SVM_EXIT_INVD : SVM_EXIT_WBINVD);
7459 /* nothing to do */
7460 }
7461 break;
7462 case 0x63: /* arpl or movslS (x86_64) */
7463 #ifdef TARGET_X86_64
7464 if (CODE64(s)) {
7465 int d_ot;
7466 /* d_ot is the size of destination */
7467 d_ot = dflag;
7468
7469 modrm = cpu_ldub_code(env, s->pc++);
7470 reg = ((modrm >> 3) & 7) | rex_r;
7471 mod = (modrm >> 6) & 3;
7472 rm = (modrm & 7) | REX_B(s);
7473
7474 if (mod == 3) {
7475 gen_op_mov_v_reg(MO_32, cpu_T[0], rm);
7476 /* sign extend */
7477 if (d_ot == MO_64) {
7478 tcg_gen_ext32s_tl(cpu_T[0], cpu_T[0]);
7479 }
7480 gen_op_mov_reg_v(d_ot, reg, cpu_T[0]);
7481 } else {
7482 gen_lea_modrm(env, s, modrm);
7483 gen_op_ld_v(s, MO_32 | MO_SIGN, cpu_T[0], cpu_A0);
7484 gen_op_mov_reg_v(d_ot, reg, cpu_T[0]);
7485 }
7486 } else
7487 #endif
7488 {
7489 int label1;
7490 TCGv t0, t1, t2, a0;
7491
7492 if (!s->pe || s->vm86)
7493 goto illegal_op;
7494 t0 = tcg_temp_local_new();
7495 t1 = tcg_temp_local_new();
7496 t2 = tcg_temp_local_new();
7497 ot = MO_16;
7498 modrm = cpu_ldub_code(env, s->pc++);
7499 reg = (modrm >> 3) & 7;
7500 mod = (modrm >> 6) & 3;
7501 rm = modrm & 7;
7502 if (mod != 3) {
7503 gen_lea_modrm(env, s, modrm);
7504 gen_op_ld_v(s, ot, t0, cpu_A0);
7505 a0 = tcg_temp_local_new();
7506 tcg_gen_mov_tl(a0, cpu_A0);
7507 } else {
7508 gen_op_mov_v_reg(ot, t0, rm);
7509 TCGV_UNUSED(a0);
7510 }
7511 gen_op_mov_v_reg(ot, t1, reg);
7512 tcg_gen_andi_tl(cpu_tmp0, t0, 3);
7513 tcg_gen_andi_tl(t1, t1, 3);
7514 tcg_gen_movi_tl(t2, 0);
7515 label1 = gen_new_label();
7516 tcg_gen_brcond_tl(TCG_COND_GE, cpu_tmp0, t1, label1);
7517 tcg_gen_andi_tl(t0, t0, ~3);
7518 tcg_gen_or_tl(t0, t0, t1);
7519 tcg_gen_movi_tl(t2, CC_Z);
7520 gen_set_label(label1);
7521 if (mod != 3) {
7522 gen_op_st_v(s, ot, t0, a0);
7523 tcg_temp_free(a0);
7524 } else {
7525 gen_op_mov_reg_v(ot, rm, t0);
7526 }
7527 gen_compute_eflags(s);
7528 tcg_gen_andi_tl(cpu_cc_src, cpu_cc_src, ~CC_Z);
7529 tcg_gen_or_tl(cpu_cc_src, cpu_cc_src, t2);
7530 tcg_temp_free(t0);
7531 tcg_temp_free(t1);
7532 tcg_temp_free(t2);
7533 }
7534 break;
7535 case 0x102: /* lar */
7536 case 0x103: /* lsl */
7537 {
7538 int label1;
7539 TCGv t0;
7540 if (!s->pe || s->vm86)
7541 goto illegal_op;
7542 ot = dflag != MO_16 ? MO_32 : MO_16;
7543 modrm = cpu_ldub_code(env, s->pc++);
7544 reg = ((modrm >> 3) & 7) | rex_r;
7545 gen_ldst_modrm(env, s, modrm, MO_16, OR_TMP0, 0);
7546 t0 = tcg_temp_local_new();
7547 gen_update_cc_op(s);
7548 if (b == 0x102) {
7549 gen_helper_lar(t0, cpu_env, cpu_T[0]);
7550 } else {
7551 gen_helper_lsl(t0, cpu_env, cpu_T[0]);
7552 }
7553 tcg_gen_andi_tl(cpu_tmp0, cpu_cc_src, CC_Z);
7554 label1 = gen_new_label();
7555 tcg_gen_brcondi_tl(TCG_COND_EQ, cpu_tmp0, 0, label1);
7556 gen_op_mov_reg_v(ot, reg, t0);
7557 gen_set_label(label1);
7558 set_cc_op(s, CC_OP_EFLAGS);
7559 tcg_temp_free(t0);
7560 }
7561 break;
7562 case 0x118:
7563 modrm = cpu_ldub_code(env, s->pc++);
7564 mod = (modrm >> 6) & 3;
7565 op = (modrm >> 3) & 7;
7566 switch(op) {
7567 case 0: /* prefetchnta */
7568 case 1: /* prefetchnt0 */
7569 case 2: /* prefetchnt0 */
7570 case 3: /* prefetchnt0 */
7571 if (mod == 3)
7572 goto illegal_op;
7573 gen_lea_modrm(env, s, modrm);
7574 /* nothing more to do */
7575 break;
7576 default: /* nop (multi byte) */
7577 gen_nop_modrm(env, s, modrm);
7578 break;
7579 }
7580 break;
7581 case 0x119 ... 0x11f: /* nop (multi byte) */
7582 modrm = cpu_ldub_code(env, s->pc++);
7583 gen_nop_modrm(env, s, modrm);
7584 break;
7585 case 0x120: /* mov reg, crN */
7586 case 0x122: /* mov crN, reg */
7587 if (s->cpl != 0) {
7588 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7589 } else {
7590 modrm = cpu_ldub_code(env, s->pc++);
7591 /* Ignore the mod bits (assume (modrm&0xc0)==0xc0).
7592 * AMD documentation (24594.pdf) and testing of
7593 * intel 386 and 486 processors all show that the mod bits
7594 * are assumed to be 1's, regardless of actual values.
7595 */
7596 rm = (modrm & 7) | REX_B(s);
7597 reg = ((modrm >> 3) & 7) | rex_r;
7598 if (CODE64(s))
7599 ot = MO_64;
7600 else
7601 ot = MO_32;
7602 if ((prefixes & PREFIX_LOCK) && (reg == 0) &&
7603 (s->cpuid_ext3_features & CPUID_EXT3_CR8LEG)) {
7604 reg = 8;
7605 }
7606 switch(reg) {
7607 case 0:
7608 case 2:
7609 case 3:
7610 case 4:
7611 case 8:
7612 gen_update_cc_op(s);
7613 gen_jmp_im(pc_start - s->cs_base);
7614 if (b & 2) {
7615 gen_op_mov_v_reg(ot, cpu_T[0], rm);
7616 gen_helper_write_crN(cpu_env, tcg_const_i32(reg),
7617 cpu_T[0]);
7618 gen_jmp_im(s->pc - s->cs_base);
7619 gen_eob(s);
7620 } else {
7621 gen_helper_read_crN(cpu_T[0], cpu_env, tcg_const_i32(reg));
7622 gen_op_mov_reg_v(ot, rm, cpu_T[0]);
7623 }
7624 break;
7625 default:
7626 goto illegal_op;
7627 }
7628 }
7629 break;
7630 case 0x121: /* mov reg, drN */
7631 case 0x123: /* mov drN, reg */
7632 if (s->cpl != 0) {
7633 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7634 } else {
7635 modrm = cpu_ldub_code(env, s->pc++);
7636 /* Ignore the mod bits (assume (modrm&0xc0)==0xc0).
7637 * AMD documentation (24594.pdf) and testing of
7638 * intel 386 and 486 processors all show that the mod bits
7639 * are assumed to be 1's, regardless of actual values.
7640 */
7641 rm = (modrm & 7) | REX_B(s);
7642 reg = ((modrm >> 3) & 7) | rex_r;
7643 if (CODE64(s))
7644 ot = MO_64;
7645 else
7646 ot = MO_32;
7647 /* XXX: do it dynamically with CR4.DE bit */
7648 if (reg == 4 || reg == 5 || reg >= 8)
7649 goto illegal_op;
7650 if (b & 2) {
7651 gen_svm_check_intercept(s, pc_start, SVM_EXIT_WRITE_DR0 + reg);
7652 gen_op_mov_v_reg(ot, cpu_T[0], rm);
7653 gen_helper_movl_drN_T0(cpu_env, tcg_const_i32(reg), cpu_T[0]);
7654 gen_jmp_im(s->pc - s->cs_base);
7655 gen_eob(s);
7656 } else {
7657 gen_svm_check_intercept(s, pc_start, SVM_EXIT_READ_DR0 + reg);
7658 tcg_gen_ld_tl(cpu_T[0], cpu_env, offsetof(CPUX86State,dr[reg]));
7659 gen_op_mov_reg_v(ot, rm, cpu_T[0]);
7660 }
7661 }
7662 break;
7663 case 0x106: /* clts */
7664 if (s->cpl != 0) {
7665 gen_exception(s, EXCP0D_GPF, pc_start - s->cs_base);
7666 } else {
7667 gen_svm_check_intercept(s, pc_start, SVM_EXIT_WRITE_CR0);
7668 gen_helper_clts(cpu_env);
7669 /* abort block because static cpu state changed */
7670 gen_jmp_im(s->pc - s->cs_base);
7671 gen_eob(s);
7672 }
7673 break;
7674 /* MMX/3DNow!/SSE/SSE2/SSE3/SSSE3/SSE4 support */
7675 case 0x1c3: /* MOVNTI reg, mem */
7676 if (!(s->cpuid_features & CPUID_SSE2))
7677 goto illegal_op;
7678 ot = mo_64_32(dflag);
7679 modrm = cpu_ldub_code(env, s->pc++);
7680 mod = (modrm >> 6) & 3;
7681 if (mod == 3)
7682 goto illegal_op;
7683 reg = ((modrm >> 3) & 7) | rex_r;
7684 /* generate a generic store */
7685 gen_ldst_modrm(env, s, modrm, ot, reg, 1);
7686 break;
7687 case 0x1ae:
7688 modrm = cpu_ldub_code(env, s->pc++);
7689 mod = (modrm >> 6) & 3;
7690 op = (modrm >> 3) & 7;
7691 switch(op) {
7692 case 0: /* fxsave */
7693 if (mod == 3 || !(s->cpuid_features & CPUID_FXSR) ||
7694 (s->prefix & PREFIX_LOCK))
7695 goto illegal_op;
7696 if ((s->flags & HF_EM_MASK) || (s->flags & HF_TS_MASK)) {
7697 gen_exception(s, EXCP07_PREX, pc_start - s->cs_base);
7698 break;
7699 }
7700 gen_lea_modrm(env, s, modrm);
7701 gen_update_cc_op(s);
7702 gen_jmp_im(pc_start - s->cs_base);
7703 gen_helper_fxsave(cpu_env, cpu_A0, tcg_const_i32(dflag == MO_64));
7704 break;
7705 case 1: /* fxrstor */
7706 if (mod == 3 || !(s->cpuid_features & CPUID_FXSR) ||
7707 (s->prefix & PREFIX_LOCK))
7708 goto illegal_op;
7709 if ((s->flags & HF_EM_MASK) || (s->flags & HF_TS_MASK)) {
7710 gen_exception(s, EXCP07_PREX, pc_start - s->cs_base);
7711 break;
7712 }
7713 gen_lea_modrm(env, s, modrm);
7714 gen_update_cc_op(s);
7715 gen_jmp_im(pc_start - s->cs_base);
7716 gen_helper_fxrstor(cpu_env, cpu_A0, tcg_const_i32(dflag == MO_64));
7717 break;
7718 case 2: /* ldmxcsr */
7719 case 3: /* stmxcsr */
7720 if (s->flags & HF_TS_MASK) {
7721 gen_exception(s, EXCP07_PREX, pc_start - s->cs_base);
7722 break;
7723 }
7724 if ((s->flags & HF_EM_MASK) || !(s->flags & HF_OSFXSR_MASK) ||
7725 mod == 3)
7726 goto illegal_op;
7727 gen_lea_modrm(env, s, modrm);
7728 if (op == 2) {
7729 tcg_gen_qemu_ld_i32(cpu_tmp2_i32, cpu_A0,
7730 s->mem_index, MO_LEUL);
7731 gen_helper_ldmxcsr(cpu_env, cpu_tmp2_i32);
7732 } else {
7733 tcg_gen_ld32u_tl(cpu_T[0], cpu_env, offsetof(CPUX86State, mxcsr));
7734 gen_op_st_v(s, MO_32, cpu_T[0], cpu_A0);
7735 }
7736 break;
7737 case 5: /* lfence */
7738 case 6: /* mfence */
7739 if ((modrm & 0xc7) != 0xc0 || !(s->cpuid_features & CPUID_SSE2))
7740 goto illegal_op;
7741 break;
7742 case 7: /* sfence / clflush */
7743 if ((modrm & 0xc7) == 0xc0) {
7744 /* sfence */
7745 /* XXX: also check for cpuid_ext2_features & CPUID_EXT2_EMMX */
7746 if (!(s->cpuid_features & CPUID_SSE))
7747 goto illegal_op;
7748 } else {
7749 /* clflush */
7750 if (!(s->cpuid_features & CPUID_CLFLUSH))
7751 goto illegal_op;
7752 gen_lea_modrm(env, s, modrm);
7753 }
7754 break;
7755 default:
7756 goto illegal_op;
7757 }
7758 break;
7759 case 0x10d: /* 3DNow! prefetch(w) */
7760 modrm = cpu_ldub_code(env, s->pc++);
7761 mod = (modrm >> 6) & 3;
7762 if (mod == 3)
7763 goto illegal_op;
7764 gen_lea_modrm(env, s, modrm);
7765 /* ignore for now */
7766 break;
7767 case 0x1aa: /* rsm */
7768 gen_svm_check_intercept(s, pc_start, SVM_EXIT_RSM);
7769 if (!(s->flags & HF_SMM_MASK))
7770 goto illegal_op;
7771 gen_update_cc_op(s);
7772 gen_jmp_im(s->pc - s->cs_base);
7773 gen_helper_rsm(cpu_env);
7774 gen_eob(s);
7775 break;
7776 case 0x1b8: /* SSE4.2 popcnt */
7777 if ((prefixes & (PREFIX_REPZ | PREFIX_LOCK | PREFIX_REPNZ)) !=
7778 PREFIX_REPZ)
7779 goto illegal_op;
7780 if (!(s->cpuid_ext_features & CPUID_EXT_POPCNT))
7781 goto illegal_op;
7782
7783 modrm = cpu_ldub_code(env, s->pc++);
7784 reg = ((modrm >> 3) & 7) | rex_r;
7785
7786 if (s->prefix & PREFIX_DATA) {
7787 ot = MO_16;
7788 } else {
7789 ot = mo_64_32(dflag);
7790 }
7791
7792 gen_ldst_modrm(env, s, modrm, ot, OR_TMP0, 0);
7793 gen_helper_popcnt(cpu_T[0], cpu_env, cpu_T[0], tcg_const_i32(ot));
7794 gen_op_mov_reg_v(ot, reg, cpu_T[0]);
7795
7796 set_cc_op(s, CC_OP_EFLAGS);
7797 break;
7798 case 0x10e ... 0x10f:
7799 /* 3DNow! instructions, ignore prefixes */
7800 s->prefix &= ~(PREFIX_REPZ | PREFIX_REPNZ | PREFIX_DATA);
7801 case 0x110 ... 0x117:
7802 case 0x128 ... 0x12f:
7803 case 0x138 ... 0x13a:
7804 case 0x150 ... 0x179:
7805 case 0x17c ... 0x17f:
7806 case 0x1c2:
7807 case 0x1c4 ... 0x1c6:
7808 case 0x1d0 ... 0x1fe:
7809 gen_sse(env, s, b, pc_start, rex_r);
7810 break;
7811 default:
7812 goto illegal_op;
7813 }
7814 /* lock generation */
7815 if (s->prefix & PREFIX_LOCK)
7816 gen_helper_unlock();
7817 return s->pc;
7818 illegal_op:
7819 if (s->prefix & PREFIX_LOCK)
7820 gen_helper_unlock();
7821 /* XXX: ensure that no lock was generated */
7822 gen_exception(s, EXCP06_ILLOP, pc_start - s->cs_base);
7823 return s->pc;
7824 }
7825
7826 void optimize_flags_init(void)
7827 {
7828 static const char reg_names[CPU_NB_REGS][4] = {
7829 #ifdef TARGET_X86_64
7830 [R_EAX] = "rax",
7831 [R_EBX] = "rbx",
7832 [R_ECX] = "rcx",
7833 [R_EDX] = "rdx",
7834 [R_ESI] = "rsi",
7835 [R_EDI] = "rdi",
7836 [R_EBP] = "rbp",
7837 [R_ESP] = "rsp",
7838 [8] = "r8",
7839 [9] = "r9",
7840 [10] = "r10",
7841 [11] = "r11",
7842 [12] = "r12",
7843 [13] = "r13",
7844 [14] = "r14",
7845 [15] = "r15",
7846 #else
7847 [R_EAX] = "eax",
7848 [R_EBX] = "ebx",
7849 [R_ECX] = "ecx",
7850 [R_EDX] = "edx",
7851 [R_ESI] = "esi",
7852 [R_EDI] = "edi",
7853 [R_EBP] = "ebp",
7854 [R_ESP] = "esp",
7855 #endif
7856 };
7857 int i;
7858
7859 cpu_env = tcg_global_reg_new_ptr(TCG_AREG0, "env");
7860 cpu_cc_op = tcg_global_mem_new_i32(TCG_AREG0,
7861 offsetof(CPUX86State, cc_op), "cc_op");
7862 cpu_cc_dst = tcg_global_mem_new(TCG_AREG0, offsetof(CPUX86State, cc_dst),
7863 "cc_dst");
7864 cpu_cc_src = tcg_global_mem_new(TCG_AREG0, offsetof(CPUX86State, cc_src),
7865 "cc_src");
7866 cpu_cc_src2 = tcg_global_mem_new(TCG_AREG0, offsetof(CPUX86State, cc_src2),
7867 "cc_src2");
7868
7869 for (i = 0; i < CPU_NB_REGS; ++i) {
7870 cpu_regs[i] = tcg_global_mem_new(TCG_AREG0,
7871 offsetof(CPUX86State, regs[i]),
7872 reg_names[i]);
7873 }
7874 }
7875
7876 /* generate intermediate code in gen_opc_buf and gen_opparam_buf for
7877 basic block 'tb'. If search_pc is TRUE, also generate PC
7878 information for each intermediate instruction. */
7879 static inline void gen_intermediate_code_internal(X86CPU *cpu,
7880 TranslationBlock *tb,
7881 bool search_pc)
7882 {
7883 CPUState *cs = CPU(cpu);
7884 CPUX86State *env = &cpu->env;
7885 DisasContext dc1, *dc = &dc1;
7886 target_ulong pc_ptr;
7887 uint16_t *gen_opc_end;
7888 CPUBreakpoint *bp;
7889 int j, lj;
7890 uint64_t flags;
7891 target_ulong pc_start;
7892 target_ulong cs_base;
7893 int num_insns;
7894 int max_insns;
7895
7896 /* generate intermediate code */
7897 pc_start = tb->pc;
7898 cs_base = tb->cs_base;
7899 flags = tb->flags;
7900
7901 dc->pe = (flags >> HF_PE_SHIFT) & 1;
7902 dc->code32 = (flags >> HF_CS32_SHIFT) & 1;
7903 dc->ss32 = (flags >> HF_SS32_SHIFT) & 1;
7904 dc->addseg = (flags >> HF_ADDSEG_SHIFT) & 1;
7905 dc->f_st = 0;
7906 dc->vm86 = (flags >> VM_SHIFT) & 1;
7907 dc->cpl = (flags >> HF_CPL_SHIFT) & 3;
7908 dc->iopl = (flags >> IOPL_SHIFT) & 3;
7909 dc->tf = (flags >> TF_SHIFT) & 1;
7910 dc->singlestep_enabled = cs->singlestep_enabled;
7911 dc->cc_op = CC_OP_DYNAMIC;
7912 dc->cc_op_dirty = false;
7913 dc->cs_base = cs_base;
7914 dc->tb = tb;
7915 dc->popl_esp_hack = 0;
7916 /* select memory access functions */
7917 dc->mem_index = 0;
7918 if (flags & HF_SOFTMMU_MASK) {
7919 dc->mem_index = cpu_mmu_index(env);
7920 }
7921 dc->cpuid_features = env->features[FEAT_1_EDX];
7922 dc->cpuid_ext_features = env->features[FEAT_1_ECX];
7923 dc->cpuid_ext2_features = env->features[FEAT_8000_0001_EDX];
7924 dc->cpuid_ext3_features = env->features[FEAT_8000_0001_ECX];
7925 dc->cpuid_7_0_ebx_features = env->features[FEAT_7_0_EBX];
7926 #ifdef TARGET_X86_64
7927 dc->lma = (flags >> HF_LMA_SHIFT) & 1;
7928 dc->code64 = (flags >> HF_CS64_SHIFT) & 1;
7929 #endif
7930 dc->flags = flags;
7931 dc->jmp_opt = !(dc->tf || cs->singlestep_enabled ||
7932 (flags & HF_INHIBIT_IRQ_MASK)
7933 #ifndef CONFIG_SOFTMMU
7934 || (flags & HF_SOFTMMU_MASK)
7935 #endif
7936 );
7937 #if 0
7938 /* check addseg logic */
7939 if (!dc->addseg && (dc->vm86 || !dc->pe || !dc->code32))
7940 printf("ERROR addseg\n");
7941 #endif
7942
7943 cpu_T[0] = tcg_temp_new();
7944 cpu_T[1] = tcg_temp_new();
7945 cpu_A0 = tcg_temp_new();
7946
7947 cpu_tmp0 = tcg_temp_new();
7948 cpu_tmp1_i64 = tcg_temp_new_i64();
7949 cpu_tmp2_i32 = tcg_temp_new_i32();
7950 cpu_tmp3_i32 = tcg_temp_new_i32();
7951 cpu_tmp4 = tcg_temp_new();
7952 cpu_ptr0 = tcg_temp_new_ptr();
7953 cpu_ptr1 = tcg_temp_new_ptr();
7954 cpu_cc_srcT = tcg_temp_local_new();
7955
7956 gen_opc_end = tcg_ctx.gen_opc_buf + OPC_MAX_SIZE;
7957
7958 dc->is_jmp = DISAS_NEXT;
7959 pc_ptr = pc_start;
7960 lj = -1;
7961 num_insns = 0;
7962 max_insns = tb->cflags & CF_COUNT_MASK;
7963 if (max_insns == 0)
7964 max_insns = CF_COUNT_MASK;
7965
7966 gen_tb_start();
7967 for(;;) {
7968 if (unlikely(!QTAILQ_EMPTY(&env->breakpoints))) {
7969 QTAILQ_FOREACH(bp, &env->breakpoints, entry) {
7970 if (bp->pc == pc_ptr &&
7971 !((bp->flags & BP_CPU) && (tb->flags & HF_RF_MASK))) {
7972 gen_debug(dc, pc_ptr - dc->cs_base);
7973 break;
7974 }
7975 }
7976 }
7977 if (search_pc) {
7978 j = tcg_ctx.gen_opc_ptr - tcg_ctx.gen_opc_buf;
7979 if (lj < j) {
7980 lj++;
7981 while (lj < j)
7982 tcg_ctx.gen_opc_instr_start[lj++] = 0;
7983 }
7984 tcg_ctx.gen_opc_pc[lj] = pc_ptr;
7985 gen_opc_cc_op[lj] = dc->cc_op;
7986 tcg_ctx.gen_opc_instr_start[lj] = 1;
7987 tcg_ctx.gen_opc_icount[lj] = num_insns;
7988 }
7989 if (num_insns + 1 == max_insns && (tb->cflags & CF_LAST_IO))
7990 gen_io_start();
7991
7992 pc_ptr = disas_insn(env, dc, pc_ptr);
7993 num_insns++;
7994 /* stop translation if indicated */
7995 if (dc->is_jmp)
7996 break;
7997 /* if single step mode, we generate only one instruction and
7998 generate an exception */
7999 /* if irq were inhibited with HF_INHIBIT_IRQ_MASK, we clear
8000 the flag and abort the translation to give the irqs a
8001 change to be happen */
8002 if (dc->tf || dc->singlestep_enabled ||
8003 (flags & HF_INHIBIT_IRQ_MASK)) {
8004 gen_jmp_im(pc_ptr - dc->cs_base);
8005 gen_eob(dc);
8006 break;
8007 }
8008 /* if too long translation, stop generation too */
8009 if (tcg_ctx.gen_opc_ptr >= gen_opc_end ||
8010 (pc_ptr - pc_start) >= (TARGET_PAGE_SIZE - 32) ||
8011 num_insns >= max_insns) {
8012 gen_jmp_im(pc_ptr - dc->cs_base);
8013 gen_eob(dc);
8014 break;
8015 }
8016 if (singlestep) {
8017 gen_jmp_im(pc_ptr - dc->cs_base);
8018 gen_eob(dc);
8019 break;
8020 }
8021 }
8022 if (tb->cflags & CF_LAST_IO)
8023 gen_io_end();
8024 gen_tb_end(tb, num_insns);
8025 *tcg_ctx.gen_opc_ptr = INDEX_op_end;
8026 /* we don't forget to fill the last values */
8027 if (search_pc) {
8028 j = tcg_ctx.gen_opc_ptr - tcg_ctx.gen_opc_buf;
8029 lj++;
8030 while (lj <= j)
8031 tcg_ctx.gen_opc_instr_start[lj++] = 0;
8032 }
8033
8034 #ifdef DEBUG_DISAS
8035 if (qemu_loglevel_mask(CPU_LOG_TB_IN_ASM)) {
8036 int disas_flags;
8037 qemu_log("----------------\n");
8038 qemu_log("IN: %s\n", lookup_symbol(pc_start));
8039 #ifdef TARGET_X86_64
8040 if (dc->code64)
8041 disas_flags = 2;
8042 else
8043 #endif
8044 disas_flags = !dc->code32;
8045 log_target_disas(env, pc_start, pc_ptr - pc_start, disas_flags);
8046 qemu_log("\n");
8047 }
8048 #endif
8049
8050 if (!search_pc) {
8051 tb->size = pc_ptr - pc_start;
8052 tb->icount = num_insns;
8053 }
8054 }
8055
8056 void gen_intermediate_code(CPUX86State *env, TranslationBlock *tb)
8057 {
8058 gen_intermediate_code_internal(x86_env_get_cpu(env), tb, false);
8059 }
8060
8061 void gen_intermediate_code_pc(CPUX86State *env, TranslationBlock *tb)
8062 {
8063 gen_intermediate_code_internal(x86_env_get_cpu(env), tb, true);
8064 }
8065
8066 void restore_state_to_opc(CPUX86State *env, TranslationBlock *tb, int pc_pos)
8067 {
8068 int cc_op;
8069 #ifdef DEBUG_DISAS
8070 if (qemu_loglevel_mask(CPU_LOG_TB_OP)) {
8071 int i;
8072 qemu_log("RESTORE:\n");
8073 for(i = 0;i <= pc_pos; i++) {
8074 if (tcg_ctx.gen_opc_instr_start[i]) {
8075 qemu_log("0x%04x: " TARGET_FMT_lx "\n", i,
8076 tcg_ctx.gen_opc_pc[i]);
8077 }
8078 }
8079 qemu_log("pc_pos=0x%x eip=" TARGET_FMT_lx " cs_base=%x\n",
8080 pc_pos, tcg_ctx.gen_opc_pc[pc_pos] - tb->cs_base,
8081 (uint32_t)tb->cs_base);
8082 }
8083 #endif
8084 env->eip = tcg_ctx.gen_opc_pc[pc_pos] - tb->cs_base;
8085 cc_op = gen_opc_cc_op[pc_pos];
8086 if (cc_op != CC_OP_DYNAMIC)
8087 env->cc_op = cc_op;
8088 }
QEmu