?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
BR 2557903: fix disassembly of a set of SSE MOV* instructions
[nasm/avx512.git] / assemble.c
blobad7ea70603243f15a563bd6df189473cdb490e3d
1 /* assemble.c code generation for the Netwide Assembler
2 *
3 * The Netwide Assembler is copyright (C) 1996 Simon Tatham and
4 * Julian Hall. All rights reserved. The software is
5 * redistributable under the license given in the file "LICENSE"
6 * distributed in the NASM archive.
7 *
8 * the actual codes (C syntax, i.e. octal):
9 * \0 - terminates the code. (Unless it's a literal of course.)
10 * \1..\4 - that many literal bytes follow in the code stream
11 * \5 - add 4 to the primary operand number (b, low octdigit)
12 * \6 - add 4 to the secondary operand number (a, middle octdigit)
13 * \7 - add 4 to both the primary and the secondary operand number
14 * \10..\13 - a literal byte follows in the code stream, to be added
15 * to the register value of operand 0..3
16 * \14..\17 - a signed byte immediate operand, from operand 0..3
17 * \20..\23 - a byte immediate operand, from operand 0..3
18 * \24..\27 - an unsigned byte immediate operand, from operand 0..3
19 * \30..\33 - a word immediate operand, from operand 0..3
20 * \34..\37 - select between \3[0-3] and \4[0-3] depending on 16/32 bit
21 * assembly mode or the operand-size override on the operand
22 * \40..\43 - a long immediate operand, from operand 0..3
23 * \44..\47 - select between \3[0-3], \4[0-3] and \5[4-7]
24 * depending on the address size of the instruction.
25 * \50..\53 - a byte relative operand, from operand 0..3
26 * \54..\57 - a qword immediate operand, from operand 0..3
27 * \60..\63 - a word relative operand, from operand 0..3
28 * \64..\67 - select between \6[0-3] and \7[0-3] depending on 16/32 bit
29 * assembly mode or the operand-size override on the operand
30 * \70..\73 - a long relative operand, from operand 0..3
31 * \74..\77 - a word constant, from the _segment_ part of operand 0..3
32 * \1ab - a ModRM, calculated on EA in operand a, with the spare
33 * field the register value of operand b.
34 * \140..\143 - an immediate word or signed byte for operand 0..3
35 * \144..\147 - or 2 (s-field) into opcode byte if operand 0..3
36 * is a signed byte rather than a word. Opcode byte follows.
37 * \150..\153 - an immediate dword or signed byte for operand 0..3
38 * \154..\157 - or 2 (s-field) into opcode byte if operand 0..3
39 * is a signed byte rather than a dword. Opcode byte follows.
40 * \160..\163 - this instruction uses DREX rather than REX, with the
41 * OC0 field set to 0, and the dest field taken from
42 * operand 0..3.
43 * \164..\167 - this instruction uses DREX rather than REX, with the
44 * OC0 field set to 1, and the dest field taken from
45 * operand 0..3.
46 * \171 - placement of DREX suffix in the absence of an EA
47 * \172\ab - the register number from operand a in bits 7..4, with
48 * the 4-bit immediate from operand b in bits 3..0.
49 * \173\xab - the register number from operand a in bits 7..4, with
50 * the value b in bits 3..0.
51 * \174\a - the register number from operand a in bits 7..4, and
52 * an arbitrary value in bits 3..0 (assembled as zero.)
53 * \2ab - a ModRM, calculated on EA in operand a, with the spare
54 * field equal to digit b.
55 * \250..\253 - same as \150..\153, except warn if the 64-bit operand
56 * is not equal to the truncated and sign-extended 32-bit
57 * operand; used for 32-bit immediates in 64-bit mode.
58 * \254..\257 - a signed 32-bit operand to be extended to 64 bits.
59 * \260..\263 - this instruction uses VEX rather than REX, with the
60 * V field taken from operand 0..3.
61 * \270 - this instruction uses VEX rather than REX, with the
62 * V field set to 1111b.
63 *
64 * VEX prefixes are followed by the sequence:
65 * \mm\wlp where mm is the M field; and wlp is:
66 * 00 0ww lpp
67 * [w0] ww = 0 for W = 0
68 * [w1] ww = 1 for W = 1
69 * [wx] ww = 2 for W don't care (always assembled as 0)
70 * [ww] ww = 3 for W used as REX.W
71 *
72 *
73 * \274..\277 - a signed byte immediate operand, from operand 0..3,
74 * which is to be extended to the operand size.
75 * \310 - indicates fixed 16-bit address size, i.e. optional 0x67.
76 * \311 - indicates fixed 32-bit address size, i.e. optional 0x67.
77 * \312 - (disassembler only) marker on LOOP, LOOPxx instructions.
78 * \313 - indicates fixed 64-bit address size, 0x67 invalid.
79 * \314 - (disassembler only) invalid with REX.B
80 * \315 - (disassembler only) invalid with REX.X
81 * \316 - (disassembler only) invalid with REX.R
82 * \317 - (disassembler only) invalid with REX.W
83 * \320 - indicates fixed 16-bit operand size, i.e. optional 0x66.
84 * \321 - indicates fixed 32-bit operand size, i.e. optional 0x66.
85 * \322 - indicates that this instruction is only valid when the
86 * operand size is the default (instruction to disassembler,
87 * generates no code in the assembler)
88 * \323 - indicates fixed 64-bit operand size, REX on extensions only.
89 * \324 - indicates 64-bit operand size requiring REX prefix.
90 * \330 - a literal byte follows in the code stream, to be added
91 * to the condition code value of the instruction.
92 * \331 - instruction not valid with REP prefix. Hint for
93 * disassembler only; for SSE instructions.
94 * \332 - REP prefix (0xF2 byte) used as opcode extension.
95 * \333 - REP prefix (0xF3 byte) used as opcode extension.
96 * \334 - LOCK prefix used instead of REX.R
97 * \335 - disassemble a rep (0xF3 byte) prefix as repe not rep.
98 * \336 - force a REP(E) prefix (0xF2) even if not specified.
99 * \337 - force a REPNE prefix (0xF3) even if not specified.
100 * \336-\337 are still listed as prefixes in the disassembler.
101 * \340 - reserve <operand 0> bytes of uninitialized storage.
102 * Operand 0 had better be a segmentless constant.
103 * \341 - this instruction needs a WAIT "prefix"
104 * \344,\345 - the PUSH/POP (respectively) codes for CS, DS, ES, SS
105 * (POP is never used for CS) depending on operand 0
106 * \346,\347 - the second byte of PUSH/POP codes for FS, GS, depending
107 * on operand 0
108 * \360 - no SSE prefix (== \364\331)
109 * \361 - 66 SSE prefix (== \366\331)
110 * \362 - F2 SSE prefix (== \364\332)
111 * \363 - F3 SSE prefix (== \364\333)
112 * \364 - operand-size prefix (0x66) not permitted
113 * \365 - address-size prefix (0x67) not permitted
114 * \366 - operand-size prefix (0x66) used as opcode extension
115 * \367 - address-size prefix (0x67) used as opcode extension
116 * \370,\371,\372 - match only if operand 0 meets byte jump criteria.
117 * 370 is used for Jcc, 371 is used for JMP.
118 * \373 - assemble 0x03 if bits==16, 0x05 if bits==32;
119 * used for conditional jump over longer jump
120 */
121
122 #include "compiler.h"
123
124 #include <stdio.h>
125 #include <string.h>
126 #include <inttypes.h>
127
128 #include "nasm.h"
129 #include "nasmlib.h"
130 #include "assemble.h"
131 #include "insns.h"
132 #include "tables.h"
133
134 typedef struct {
135 int sib_present; /* is a SIB byte necessary? */
136 int bytes; /* # of bytes of offset needed */
137 int size; /* lazy - this is sib+bytes+1 */
138 uint8_t modrm, sib, rex, rip; /* the bytes themselves */
139 } ea;
140
141 static uint32_t cpu; /* cpu level received from nasm.c */
142 static efunc errfunc;
143 static struct ofmt *outfmt;
144 static ListGen *list;
145
146 static int64_t calcsize(int32_t, int64_t, int, insn *, const uint8_t *);
147 static void gencode(int32_t segment, int64_t offset, int bits,
148 insn * ins, const struct itemplate *temp,
149 int64_t insn_end);
150 static int matches(const struct itemplate *, insn *, int bits);
151 static int32_t regflag(const operand *);
152 static int32_t regval(const operand *);
153 static int rexflags(int, int32_t, int);
154 static int op_rexflags(const operand *, int);
155 static ea *process_ea(operand *, ea *, int, int, int, int32_t);
156 static void add_asp(insn *, int);
157
158 static int has_prefix(insn * ins, enum prefix_pos pos, enum prefixes prefix)
159 {
160 return ins->prefixes[pos] == prefix;
161 }
162
163 static void assert_no_prefix(insn * ins, enum prefix_pos pos)
164 {
165 if (ins->prefixes[pos])
166 errfunc(ERR_NONFATAL, "invalid %s prefix",
167 prefix_name(ins->prefixes[pos]));
168 }
169
170 static const char *size_name(int size)
171 {
172 switch (size) {
173 case 1:
174 return "byte";
175 case 2:
176 return "word";
177 case 4:
178 return "dword";
179 case 8:
180 return "qword";
181 case 10:
182 return "tword";
183 case 16:
184 return "oword";
185 case 32:
186 return "yword";
187 default:
188 return "???";
189 }
190 }
191
192 static void warn_overflow(int size, const struct operand *o)
193 {
194 if (size < 8 && o->wrt == NO_SEG && o->segment == NO_SEG) {
195 int64_t lim = ((int64_t)1 << (size*8))-1;
196 int64_t data = o->offset;
197
198 if (data < ~lim || data > lim)
199 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
200 "%s data exceeds bounds", size_name(size));
201 }
202 }
203 /*
204 * This routine wrappers the real output format's output routine,
205 * in order to pass a copy of the data off to the listing file
206 * generator at the same time.
207 */
208 static void out(int64_t offset, int32_t segto, const void *data,
209 enum out_type type, uint64_t size,
210 int32_t segment, int32_t wrt)
211 {
212 static int32_t lineno = 0; /* static!!! */
213 static char *lnfname = NULL;
214 uint8_t p[8];
215
216 if (type == OUT_ADDRESS && segment == NO_SEG && wrt == NO_SEG) {
217 /*
218 * This is a non-relocated address, and we're going to
219 * convert it into RAWDATA format.
220 */
221 uint8_t *q = p;
222
223 if (size > 8) {
224 errfunc(ERR_PANIC, "OUT_ADDRESS with size > 8");
225 return;
226 }
227
228 WRITEADDR(q, *(int64_t *)data, size);
229 data = p;
230 type = OUT_RAWDATA;
231 }
232
233 list->output(offset, data, type, size);
234
235 /*
236 * this call to src_get determines when we call the
237 * debug-format-specific "linenum" function
238 * it updates lineno and lnfname to the current values
239 * returning 0 if "same as last time", -2 if lnfname
240 * changed, and the amount by which lineno changed,
241 * if it did. thus, these variables must be static
242 */
243
244 if (src_get(&lineno, &lnfname)) {
245 outfmt->current_dfmt->linenum(lnfname, lineno, segto);
246 }
247
248 outfmt->output(segto, data, type, size, segment, wrt);
249 }
250
251 static bool jmp_match(int32_t segment, int64_t offset, int bits,
252 insn * ins, const uint8_t *code)
253 {
254 int64_t isize;
255 uint8_t c = code[0];
256
257 if ((c != 0370 && c != 0371) || (ins->oprs[0].type & STRICT))
258 return false;
259 if (!optimizing)
260 return false;
261 if (optimizing < 0 && c == 0371)
262 return false;
263
264 isize = calcsize(segment, offset, bits, ins, code);
265 if (ins->oprs[0].segment != segment)
266 return false;
267
268 isize = ins->oprs[0].offset - offset - isize; /* isize is delta */
269 return (isize >= -128 && isize <= 127); /* is it byte size? */
270 }
271
272 int64_t assemble(int32_t segment, int64_t offset, int bits, uint32_t cp,
273 insn * instruction, struct ofmt *output, efunc error,
274 ListGen * listgen)
275 {
276 const struct itemplate *temp;
277 int j;
278 int size_prob;
279 int64_t insn_end;
280 int32_t itimes;
281 int64_t start = offset;
282 int64_t wsize = 0; /* size for DB etc. */
283
284 errfunc = error; /* to pass to other functions */
285 cpu = cp;
286 outfmt = output; /* likewise */
287 list = listgen; /* and again */
288
289 switch (instruction->opcode) {
290 case -1:
291 return 0;
292 case I_DB:
293 wsize = 1;
294 break;
295 case I_DW:
296 wsize = 2;
297 break;
298 case I_DD:
299 wsize = 4;
300 break;
301 case I_DQ:
302 wsize = 8;
303 break;
304 case I_DT:
305 wsize = 10;
306 break;
307 case I_DO:
308 wsize = 16;
309 break;
310 case I_DY:
311 wsize = 32;
312 break;
313 default:
314 break;
315 }
316
317 if (wsize) {
318 extop *e;
319 int32_t t = instruction->times;
320 if (t < 0)
321 errfunc(ERR_PANIC,
322 "instruction->times < 0 (%ld) in assemble()", t);
323
324 while (t--) { /* repeat TIMES times */
325 for (e = instruction->eops; e; e = e->next) {
326 if (e->type == EOT_DB_NUMBER) {
327 if (wsize == 1) {
328 if (e->segment != NO_SEG)
329 errfunc(ERR_NONFATAL,
330 "one-byte relocation attempted");
331 else {
332 uint8_t out_byte = e->offset;
333 out(offset, segment, &out_byte,
334 OUT_RAWDATA, 1, NO_SEG, NO_SEG);
335 }
336 } else if (wsize > 8) {
337 errfunc(ERR_NONFATAL,
338 "integer supplied to a DT, DO or DY"
339 " instruction");
340 } else
341 out(offset, segment, &e->offset,
342 OUT_ADDRESS, wsize, e->segment, e->wrt);
343 offset += wsize;
344 } else if (e->type == EOT_DB_STRING ||
345 e->type == EOT_DB_STRING_FREE) {
346 int align;
347
348 out(offset, segment, e->stringval,
349 OUT_RAWDATA, e->stringlen, NO_SEG, NO_SEG);
350 align = e->stringlen % wsize;
351
352 if (align) {
353 align = wsize - align;
354 out(offset, segment, zero_buffer,
355 OUT_RAWDATA, align, NO_SEG, NO_SEG);
356 }
357 offset += e->stringlen + align;
358 }
359 }
360 if (t > 0 && t == instruction->times - 1) {
361 /*
362 * Dummy call to list->output to give the offset to the
363 * listing module.
364 */
365 list->output(offset, NULL, OUT_RAWDATA, 0);
366 list->uplevel(LIST_TIMES);
367 }
368 }
369 if (instruction->times > 1)
370 list->downlevel(LIST_TIMES);
371 return offset - start;
372 }
373
374 if (instruction->opcode == I_INCBIN) {
375 const char *fname = instruction->eops->stringval;
376 FILE *fp;
377
378 fp = fopen(fname, "rb");
379 if (!fp) {
380 error(ERR_NONFATAL, "`incbin': unable to open file `%s'",
381 fname);
382 } else if (fseek(fp, 0L, SEEK_END) < 0) {
383 error(ERR_NONFATAL, "`incbin': unable to seek on file `%s'",
384 fname);
385 } else {
386 static char buf[4096];
387 size_t t = instruction->times;
388 size_t base = 0;
389 size_t len;
390
391 len = ftell(fp);
392 if (instruction->eops->next) {
393 base = instruction->eops->next->offset;
394 len -= base;
395 if (instruction->eops->next->next &&
396 len > (size_t)instruction->eops->next->next->offset)
397 len = (size_t)instruction->eops->next->next->offset;
398 }
399 /*
400 * Dummy call to list->output to give the offset to the
401 * listing module.
402 */
403 list->output(offset, NULL, OUT_RAWDATA, 0);
404 list->uplevel(LIST_INCBIN);
405 while (t--) {
406 size_t l;
407
408 fseek(fp, base, SEEK_SET);
409 l = len;
410 while (l > 0) {
411 int32_t m =
412 fread(buf, 1, (l > (int32_t) sizeof(buf) ? (int32_t) sizeof(buf) : l),
413 fp);
414 if (!m) {
415 /*
416 * This shouldn't happen unless the file
417 * actually changes while we are reading
418 * it.
419 */
420 error(ERR_NONFATAL,
421 "`incbin': unexpected EOF while"
422 " reading file `%s'", fname);
423 t = 0; /* Try to exit cleanly */
424 break;
425 }
426 out(offset, segment, buf, OUT_RAWDATA, m,
427 NO_SEG, NO_SEG);
428 l -= m;
429 }
430 }
431 list->downlevel(LIST_INCBIN);
432 if (instruction->times > 1) {
433 /*
434 * Dummy call to list->output to give the offset to the
435 * listing module.
436 */
437 list->output(offset, NULL, OUT_RAWDATA, 0);
438 list->uplevel(LIST_TIMES);
439 list->downlevel(LIST_TIMES);
440 }
441 fclose(fp);
442 return instruction->times * len;
443 }
444 return 0; /* if we're here, there's an error */
445 }
446
447 /* Check to see if we need an address-size prefix */
448 add_asp(instruction, bits);
449
450 size_prob = 0;
451
452 for (temp = nasm_instructions[instruction->opcode]; temp->opcode != -1; temp++){
453 int m = matches(temp, instruction, bits);
454 if (m == 100 ||
455 (m == 99 && jmp_match(segment, offset, bits,
456 instruction, temp->code))) {
457 /* Matches! */
458 int64_t insn_size = calcsize(segment, offset, bits,
459 instruction, temp->code);
460 itimes = instruction->times;
461 if (insn_size < 0) /* shouldn't be, on pass two */
462 error(ERR_PANIC, "errors made it through from pass one");
463 else
464 while (itimes--) {
465 for (j = 0; j < MAXPREFIX; j++) {
466 uint8_t c = 0;
467 switch (instruction->prefixes[j]) {
468 case P_WAIT:
469 c = 0x9B;
470 break;
471 case P_LOCK:
472 c = 0xF0;
473 break;
474 case P_REPNE:
475 case P_REPNZ:
476 c = 0xF2;
477 break;
478 case P_REPE:
479 case P_REPZ:
480 case P_REP:
481 c = 0xF3;
482 break;
483 case R_CS:
484 if (bits == 64) {
485 error(ERR_WARNING | ERR_PASS2,
486 "cs segment base generated, but will be ignored in 64-bit mode");
487 }
488 c = 0x2E;
489 break;
490 case R_DS:
491 if (bits == 64) {
492 error(ERR_WARNING | ERR_PASS2,
493 "ds segment base generated, but will be ignored in 64-bit mode");
494 }
495 c = 0x3E;
496 break;
497 case R_ES:
498 if (bits == 64) {
499 error(ERR_WARNING | ERR_PASS2,
500 "es segment base generated, but will be ignored in 64-bit mode");
501 }
502 c = 0x26;
503 break;
504 case R_FS:
505 c = 0x64;
506 break;
507 case R_GS:
508 c = 0x65;
509 break;
510 case R_SS:
511 if (bits == 64) {
512 error(ERR_WARNING | ERR_PASS2,
513 "ss segment base generated, but will be ignored in 64-bit mode");
514 }
515 c = 0x36;
516 break;
517 case R_SEGR6:
518 case R_SEGR7:
519 error(ERR_NONFATAL,
520 "segr6 and segr7 cannot be used as prefixes");
521 break;
522 case P_A16:
523 if (bits == 64) {
524 error(ERR_NONFATAL,
525 "16-bit addressing is not supported "
526 "in 64-bit mode");
527 } else if (bits != 16)
528 c = 0x67;
529 break;
530 case P_A32:
531 if (bits != 32)
532 c = 0x67;
533 break;
534 case P_A64:
535 if (bits != 64) {
536 error(ERR_NONFATAL,
537 "64-bit addressing is only supported "
538 "in 64-bit mode");
539 }
540 break;
541 case P_ASP:
542 c = 0x67;
543 break;
544 case P_O16:
545 if (bits != 16)
546 c = 0x66;
547 break;
548 case P_O32:
549 if (bits == 16)
550 c = 0x66;
551 break;
552 case P_O64:
553 /* REX.W */
554 break;
555 case P_OSP:
556 c = 0x66;
557 break;
558 case P_none:
559 break;
560 default:
561 error(ERR_PANIC, "invalid instruction prefix");
562 }
563 if (c != 0) {
564 out(offset, segment, &c, OUT_RAWDATA, 1,
565 NO_SEG, NO_SEG);
566 offset++;
567 }
568 }
569 insn_end = offset + insn_size;
570 gencode(segment, offset, bits, instruction,
571 temp, insn_end);
572 offset += insn_size;
573 if (itimes > 0 && itimes == instruction->times - 1) {
574 /*
575 * Dummy call to list->output to give the offset to the
576 * listing module.
577 */
578 list->output(offset, NULL, OUT_RAWDATA, 0);
579 list->uplevel(LIST_TIMES);
580 }
581 }
582 if (instruction->times > 1)
583 list->downlevel(LIST_TIMES);
584 return offset - start;
585 } else if (m > 0 && m > size_prob) {
586 size_prob = m;
587 }
588 }
589
590 if (temp->opcode == -1) { /* didn't match any instruction */
591 switch (size_prob) {
592 case 1:
593 error(ERR_NONFATAL, "operation size not specified");
594 break;
595 case 2:
596 error(ERR_NONFATAL, "mismatch in operand sizes");
597 break;
598 case 3:
599 error(ERR_NONFATAL, "no instruction for this cpu level");
600 break;
601 case 4:
602 error(ERR_NONFATAL, "instruction not supported in %d-bit mode",
603 bits);
604 break;
605 default:
606 error(ERR_NONFATAL,
607 "invalid combination of opcode and operands");
608 break;
609 }
610 }
611 return 0;
612 }
613
614 int64_t insn_size(int32_t segment, int64_t offset, int bits, uint32_t cp,
615 insn * instruction, efunc error)
616 {
617 const struct itemplate *temp;
618
619 errfunc = error; /* to pass to other functions */
620 cpu = cp;
621
622 if (instruction->opcode == -1)
623 return 0;
624
625 if (instruction->opcode == I_DB || instruction->opcode == I_DW ||
626 instruction->opcode == I_DD || instruction->opcode == I_DQ ||
627 instruction->opcode == I_DT || instruction->opcode == I_DO ||
628 instruction->opcode == I_DY) {
629 extop *e;
630 int32_t isize, osize, wsize = 0; /* placate gcc */
631
632 isize = 0;
633 switch (instruction->opcode) {
634 case I_DB:
635 wsize = 1;
636 break;
637 case I_DW:
638 wsize = 2;
639 break;
640 case I_DD:
641 wsize = 4;
642 break;
643 case I_DQ:
644 wsize = 8;
645 break;
646 case I_DT:
647 wsize = 10;
648 break;
649 case I_DO:
650 wsize = 16;
651 break;
652 case I_DY:
653 wsize = 32;
654 break;
655 default:
656 break;
657 }
658
659 for (e = instruction->eops; e; e = e->next) {
660 int32_t align;
661
662 osize = 0;
663 if (e->type == EOT_DB_NUMBER)
664 osize = 1;
665 else if (e->type == EOT_DB_STRING ||
666 e->type == EOT_DB_STRING_FREE)
667 osize = e->stringlen;
668
669 align = (-osize) % wsize;
670 if (align < 0)
671 align += wsize;
672 isize += osize + align;
673 }
674 return isize * instruction->times;
675 }
676
677 if (instruction->opcode == I_INCBIN) {
678 const char *fname = instruction->eops->stringval;
679 FILE *fp;
680 size_t len;
681
682 fp = fopen(fname, "rb");
683 if (!fp)
684 error(ERR_NONFATAL, "`incbin': unable to open file `%s'",
685 fname);
686 else if (fseek(fp, 0L, SEEK_END) < 0)
687 error(ERR_NONFATAL, "`incbin': unable to seek on file `%s'",
688 fname);
689 else {
690 len = ftell(fp);
691 fclose(fp);
692 if (instruction->eops->next) {
693 len -= instruction->eops->next->offset;
694 if (instruction->eops->next->next &&
695 len > (size_t)instruction->eops->next->next->offset) {
696 len = (size_t)instruction->eops->next->next->offset;
697 }
698 }
699 return instruction->times * len;
700 }
701 return 0; /* if we're here, there's an error */
702 }
703
704 /* Check to see if we need an address-size prefix */
705 add_asp(instruction, bits);
706
707 for (temp = nasm_instructions[instruction->opcode]; temp->opcode != -1; temp++) {
708 int m = matches(temp, instruction, bits);
709 if (m == 100 ||
710 (m == 99 && jmp_match(segment, offset, bits,
711 instruction, temp->code))) {
712 /* we've matched an instruction. */
713 int64_t isize;
714 const uint8_t *codes = temp->code;
715 int j;
716
717 isize = calcsize(segment, offset, bits, instruction, codes);
718 if (isize < 0)
719 return -1;
720 for (j = 0; j < MAXPREFIX; j++) {
721 switch (instruction->prefixes[j]) {
722 case P_A16:
723 if (bits != 16)
724 isize++;
725 break;
726 case P_A32:
727 if (bits != 32)
728 isize++;
729 break;
730 case P_O16:
731 if (bits != 16)
732 isize++;
733 break;
734 case P_O32:
735 if (bits == 16)
736 isize++;
737 break;
738 case P_A64:
739 case P_O64:
740 case P_none:
741 break;
742 default:
743 isize++;
744 break;
745 }
746 }
747 return isize * instruction->times;
748 }
749 }
750 return -1; /* didn't match any instruction */
751 }
752
753 static bool possible_sbyte(operand *o)
754 {
755 return o->wrt == NO_SEG && o->segment == NO_SEG &&
756 !(o->opflags & OPFLAG_FORWARD) &&
757 optimizing >= 0 && !(o->type & STRICT);
758 }
759
760 /* check that opn[op] is a signed byte of size 16 or 32 */
761 static bool is_sbyte16(operand *o)
762 {
763 int16_t v;
764
765 if (!possible_sbyte(o))
766 return false;
767
768 v = o->offset;
769 return v >= -128 && v <= 127;
770 }
771
772 static bool is_sbyte32(operand *o)
773 {
774 int32_t v;
775
776 if (!possible_sbyte(o))
777 return false;
778
779 v = o->offset;
780 return v >= -128 && v <= 127;
781 }
782
783 /* Common construct */
784 #define case4(x) case (x): case (x)+1: case (x)+2: case (x)+3
785
786 static int64_t calcsize(int32_t segment, int64_t offset, int bits,
787 insn * ins, const uint8_t *codes)
788 {
789 int64_t length = 0;
790 uint8_t c;
791 int rex_mask = ~0;
792 int op1, op2;
793 struct operand *opx;
794 uint8_t opex = 0;
795
796 ins->rex = 0; /* Ensure REX is reset */
797
798 if (ins->prefixes[PPS_OSIZE] == P_O64)
799 ins->rex |= REX_W;
800
801 (void)segment; /* Don't warn that this parameter is unused */
802 (void)offset; /* Don't warn that this parameter is unused */
803
804 while (*codes) {
805 c = *codes++;
806 op1 = (c & 3) + ((opex & 1) << 2);
807 op2 = ((c >> 3) & 3) + ((opex & 2) << 1);
808 opx = &ins->oprs[op1];
809 opex = 0; /* For the next iteration */
810
811 switch (c) {
812 case 01:
813 case 02:
814 case 03:
815 case 04:
816 codes += c, length += c;
817 break;
818
819 case 05:
820 case 06:
821 case 07:
822 opex = c;
823 break;
824
825 case4(010):
826 ins->rex |=
827 op_rexflags(opx, REX_B|REX_H|REX_P|REX_W);
828 codes++, length++;
829 break;
830
831 case4(014):
832 case4(020):
833 case4(024):
834 length++;
835 break;
836
837 case4(030):
838 length += 2;
839 break;
840
841 case4(034):
842 if (opx->type & (BITS16 | BITS32 | BITS64))
843 length += (opx->type & BITS16) ? 2 : 4;
844 else
845 length += (bits == 16) ? 2 : 4;
846 break;
847
848 case4(040):
849 length += 4;
850 break;
851
852 case4(044):
853 length += ins->addr_size >> 3;
854 break;
855
856 case4(050):
857 length++;
858 break;
859
860 case4(054):
861 length += 8; /* MOV reg64/imm */
862 break;
863
864 case4(060):
865 length += 2;
866 break;
867
868 case4(064):
869 if (opx->type & (BITS16 | BITS32 | BITS64))
870 length += (opx->type & BITS16) ? 2 : 4;
871 else
872 length += (bits == 16) ? 2 : 4;
873 break;
874
875 case4(070):
876 length += 4;
877 break;
878
879 case4(074):
880 length += 2;
881 break;
882
883 case4(0140):
884 length += is_sbyte16(opx) ? 1 : 2;
885 break;
886
887 case4(0144):
888 codes++;
889 length++;
890 break;
891
892 case4(0150):
893 length += is_sbyte32(opx) ? 1 : 4;
894 break;
895
896 case4(0154):
897 codes++;
898 length++;
899 break;
900
901 case4(0160):
902 length++;
903 ins->rex |= REX_D;
904 ins->drexdst = regval(opx);
905 break;
906
907 case4(0164):
908 length++;
909 ins->rex |= REX_D|REX_OC;
910 ins->drexdst = regval(opx);
911 break;
912
913 case 0171:
914 break;
915
916 case 0172:
917 case 0173:
918 case 0174:
919 codes++;
920 length++;
921 break;
922
923 case4(0250):
924 length += is_sbyte32(opx) ? 1 : 4;
925 break;
926
927 case4(0254):
928 length += 4;
929 break;
930
931 case4(0260):
932 ins->rex |= REX_V;
933 ins->drexdst = regval(opx);
934 ins->vex_m = *codes++;
935 ins->vex_wlp = *codes++;
936 break;
937
938 case 0270:
939 ins->rex |= REX_V;
940 ins->drexdst = 0;
941 ins->vex_m = *codes++;
942 ins->vex_wlp = *codes++;
943 break;
944
945 case4(0274):
946 length++;
947 break;
948
949 case4(0300):
950 break;
951
952 case 0310:
953 if (bits == 64)
954 return -1;
955 length += (bits != 16) && !has_prefix(ins, PPS_ASIZE, P_A16);
956 break;
957
958 case 0311:
959 length += (bits != 32) && !has_prefix(ins, PPS_ASIZE, P_A32);
960 break;
961
962 case 0312:
963 break;
964
965 case 0313:
966 if (bits != 64 || has_prefix(ins, PPS_ASIZE, P_A16) ||
967 has_prefix(ins, PPS_ASIZE, P_A32))
968 return -1;
969 break;
970
971 case4(0314):
972 break;
973
974 case 0320:
975 length += (bits != 16);
976 break;
977
978 case 0321:
979 length += (bits == 16);
980 break;
981
982 case 0322:
983 break;
984
985 case 0323:
986 rex_mask &= ~REX_W;
987 break;
988
989 case 0324:
990 ins->rex |= REX_W;
991 break;
992
993 case 0330:
994 codes++, length++;
995 break;
996
997 case 0331:
998 break;
999
1000 case 0332:
1001 case 0333:
1002 length++;
1003 break;
1004
1005 case 0334:
1006 ins->rex |= REX_L;
1007 break;
1008
1009 case 0335:
1010 break;
1011
1012 case 0336:
1013 if (!ins->prefixes[PPS_LREP])
1014 ins->prefixes[PPS_LREP] = P_REP;
1015 break;
1016
1017 case 0337:
1018 if (!ins->prefixes[PPS_LREP])
1019 ins->prefixes[PPS_LREP] = P_REPNE;
1020 break;
1021
1022 case 0340:
1023 if (ins->oprs[0].segment != NO_SEG)
1024 errfunc(ERR_NONFATAL, "attempt to reserve non-constant"
1025 " quantity of BSS space");
1026 else
1027 length += ins->oprs[0].offset;
1028 break;
1029
1030 case 0341:
1031 if (!ins->prefixes[PPS_WAIT])
1032 ins->prefixes[PPS_WAIT] = P_WAIT;
1033 break;
1034
1035 case4(0344):
1036 length++;
1037 break;
1038
1039 case 0360:
1040 break;
1041
1042 case 0361:
1043 case 0362:
1044 case 0363:
1045 length++;
1046 break;
1047
1048 case 0364:
1049 case 0365:
1050 break;
1051
1052 case 0366:
1053 case 0367:
1054 length++;
1055 break;
1056
1057 case 0370:
1058 case 0371:
1059 case 0372:
1060 break;
1061
1062 case 0373:
1063 length++;
1064 break;
1065
1066 case4(0100):
1067 case4(0110):
1068 case4(0120):
1069 case4(0130):
1070 case4(0200):
1071 case4(0204):
1072 case4(0210):
1073 case4(0214):
1074 case4(0220):
1075 case4(0224):
1076 case4(0230):
1077 case4(0234):
1078 {
1079 ea ea_data;
1080 int rfield;
1081 int32_t rflags;
1082 struct operand *opy = &ins->oprs[op2];
1083
1084 ea_data.rex = 0; /* Ensure ea.REX is initially 0 */
1085
1086 if (c <= 0177) {
1087 /* pick rfield from operand b (opx) */
1088 rflags = regflag(opx);
1089 rfield = nasm_regvals[opx->basereg];
1090 } else {
1091 rflags = 0;
1092 rfield = c & 7;
1093 }
1094 if (!process_ea(opy, &ea_data, bits,
1095 ins->addr_size, rfield, rflags)) {
1096 errfunc(ERR_NONFATAL, "invalid effective address");
1097 return -1;
1098 } else {
1099 ins->rex |= ea_data.rex;
1100 length += ea_data.size;
1101 }
1102 }
1103 break;
1104
1105 default:
1106 errfunc(ERR_PANIC, "internal instruction table corrupt"
1107 ": instruction code 0x%02X given", c);
1108 break;
1109 }
1110 }
1111
1112 ins->rex &= rex_mask;
1113
1114 if (ins->rex & REX_V) {
1115 int bad32 = REX_R|REX_W|REX_X|REX_B;
1116
1117 if (ins->rex & REX_H) {
1118 errfunc(ERR_NONFATAL, "cannot use high register in vex instruction");
1119 return -1;
1120 }
1121 switch (ins->vex_wlp & 030) {
1122 case 000:
1123 case 020:
1124 ins->rex &= ~REX_W;
1125 break;
1126 case 010:
1127 ins->rex |= REX_W;
1128 bad32 &= ~REX_W;
1129 break;
1130 case 030:
1131 /* Follow REX_W */
1132 break;
1133 }
1134
1135 if (bits != 64 && ((ins->rex & bad32) || ins->drexdst > 7)) {
1136 errfunc(ERR_NONFATAL, "invalid operands in non-64-bit mode");
1137 return -1;
1138 }
1139 if (ins->vex_m != 1 || (ins->rex & (REX_W|REX_R|REX_B)))
1140 length += 3;
1141 else
1142 length += 2;
1143 } else if (ins->rex & REX_D) {
1144 if (ins->rex & REX_H) {
1145 errfunc(ERR_NONFATAL, "cannot use high register in drex instruction");
1146 return -1;
1147 }
1148 if (bits != 64 && ((ins->rex & (REX_R|REX_W|REX_X|REX_B)) ||
1149 ins->drexdst > 7)) {
1150 errfunc(ERR_NONFATAL, "invalid operands in non-64-bit mode");
1151 return -1;
1152 }
1153 length++;
1154 } else if (ins->rex & REX_REAL) {
1155 if (ins->rex & REX_H) {
1156 errfunc(ERR_NONFATAL, "cannot use high register in rex instruction");
1157 return -1;
1158 } else if (bits == 64) {
1159 length++;
1160 } else if ((ins->rex & REX_L) &&
1161 !(ins->rex & (REX_P|REX_W|REX_X|REX_B)) &&
1162 cpu >= IF_X86_64) {
1163 /* LOCK-as-REX.R */
1164 assert_no_prefix(ins, PPS_LREP);
1165 length++;
1166 } else {
1167 errfunc(ERR_NONFATAL, "invalid operands in non-64-bit mode");
1168 return -1;
1169 }
1170 }
1171
1172 return length;
1173 }
1174
1175 #define EMIT_REX() \
1176 if (!(ins->rex & (REX_D|REX_V)) && (ins->rex & REX_REAL) && (bits == 64)) { \
1177 ins->rex = (ins->rex & REX_REAL)|REX_P; \
1178 out(offset, segment, &ins->rex, OUT_RAWDATA, 1, NO_SEG, NO_SEG); \
1179 ins->rex = 0; \
1180 offset += 1; \
1181 }
1182
1183 static void gencode(int32_t segment, int64_t offset, int bits,
1184 insn * ins, const struct itemplate *temp,
1185 int64_t insn_end)
1186 {
1187 static char condval[] = { /* conditional opcodes */
1188 0x7, 0x3, 0x2, 0x6, 0x2, 0x4, 0xF, 0xD, 0xC, 0xE, 0x6, 0x2,
1189 0x3, 0x7, 0x3, 0x5, 0xE, 0xC, 0xD, 0xF, 0x1, 0xB, 0x9, 0x5,
1190 0x0, 0xA, 0xA, 0xB, 0x8, 0x4
1191 };
1192 uint8_t c;
1193 uint8_t bytes[4];
1194 int64_t size;
1195 int64_t data;
1196 int op1, op2;
1197 struct operand *opx;
1198 const uint8_t *codes = temp->code;
1199 uint8_t opex = 0;
1200
1201 while (*codes) {
1202 c = *codes++;
1203 op1 = (c & 3) + ((opex & 1) << 2);
1204 op2 = ((c >> 3) & 3) + ((opex & 2) << 1);
1205 opx = &ins->oprs[op1];
1206 opex = 0; /* For the next iteration */
1207
1208 switch (c) {
1209 case 01:
1210 case 02:
1211 case 03:
1212 case 04:
1213 EMIT_REX();
1214 out(offset, segment, codes, OUT_RAWDATA, c, NO_SEG, NO_SEG);
1215 codes += c;
1216 offset += c;
1217 break;
1218
1219 case 05:
1220 case 06:
1221 case 07:
1222 opex = c;
1223 break;
1224
1225 case4(010):
1226 EMIT_REX();
1227 bytes[0] = *codes++ + (regval(opx) & 7);
1228 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1229 offset += 1;
1230 break;
1231
1232 case4(014):
1233 /* The test for BITS8 and SBYTE here is intended to avoid
1234 warning on optimizer actions due to SBYTE, while still
1235 warn on explicit BYTE directives. Also warn, obviously,
1236 if the optimizer isn't enabled. */
1237 if (((opx->type & BITS8) ||
1238 !(opx->type & temp->opd[op1] & BYTENESS)) &&
1239 (opx->offset < -128 || opx->offset > 127)) {
1240 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1241 "signed byte value exceeds bounds");
1242 }
1243 if (opx->segment != NO_SEG) {
1244 data = opx->offset;
1245 out(offset, segment, &data, OUT_ADDRESS, 1,
1246 opx->segment, opx->wrt);
1247 } else {
1248 bytes[0] = opx->offset;
1249 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1250 NO_SEG);
1251 }
1252 offset += 1;
1253 break;
1254
1255 case4(020):
1256 if (opx->offset < -256 || opx->offset > 255) {
1257 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1258 "byte value exceeds bounds");
1259 }
1260 if (opx->segment != NO_SEG) {
1261 data = opx->offset;
1262 out(offset, segment, &data, OUT_ADDRESS, 1,
1263 opx->segment, opx->wrt);
1264 } else {
1265 bytes[0] = opx->offset;
1266 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1267 NO_SEG);
1268 }
1269 offset += 1;
1270 break;
1271
1272 case4(024):
1273 if (opx->offset < 0 || opx->offset > 255)
1274 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1275 "unsigned byte value exceeds bounds");
1276 if (opx->segment != NO_SEG) {
1277 data = opx->offset;
1278 out(offset, segment, &data, OUT_ADDRESS, 1,
1279 opx->segment, opx->wrt);
1280 } else {
1281 bytes[0] = opx->offset;
1282 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1283 NO_SEG);
1284 }
1285 offset += 1;
1286 break;
1287
1288 case4(030):
1289 warn_overflow(2, opx);
1290 data = opx->offset;
1291 out(offset, segment, &data, OUT_ADDRESS, 2,
1292 opx->segment, opx->wrt);
1293 offset += 2;
1294 break;
1295
1296 case4(034):
1297 if (opx->type & (BITS16 | BITS32))
1298 size = (opx->type & BITS16) ? 2 : 4;
1299 else
1300 size = (bits == 16) ? 2 : 4;
1301 warn_overflow(size, opx);
1302 data = opx->offset;
1303 out(offset, segment, &data, OUT_ADDRESS, size,
1304 opx->segment, opx->wrt);
1305 offset += size;
1306 break;
1307
1308 case4(040):
1309 warn_overflow(4, opx);
1310 data = opx->offset;
1311 out(offset, segment, &data, OUT_ADDRESS, 4,
1312 opx->segment, opx->wrt);
1313 offset += 4;
1314 break;
1315
1316 case4(044):
1317 data = opx->offset;
1318 size = ins->addr_size >> 3;
1319 warn_overflow(size, opx);
1320 out(offset, segment, &data, OUT_ADDRESS, size,
1321 opx->segment, opx->wrt);
1322 offset += size;
1323 break;
1324
1325 case4(050):
1326 if (opx->segment != segment)
1327 errfunc(ERR_NONFATAL,
1328 "short relative jump outside segment");
1329 data = opx->offset - insn_end;
1330 if (data > 127 || data < -128)
1331 errfunc(ERR_NONFATAL, "short jump is out of range");
1332 bytes[0] = data;
1333 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1334 offset += 1;
1335 break;
1336
1337 case4(054):
1338 data = (int64_t)opx->offset;
1339 out(offset, segment, &data, OUT_ADDRESS, 8,
1340 opx->segment, opx->wrt);
1341 offset += 8;
1342 break;
1343
1344 case4(060):
1345 if (opx->segment != segment) {
1346 data = opx->offset;
1347 out(offset, segment, &data,
1348 OUT_REL2ADR, insn_end - offset,
1349 opx->segment, opx->wrt);
1350 } else {
1351 data = opx->offset - insn_end;
1352 out(offset, segment, &data,
1353 OUT_ADDRESS, 2, NO_SEG, NO_SEG);
1354 }
1355 offset += 2;
1356 break;
1357
1358 case4(064):
1359 if (opx->type & (BITS16 | BITS32 | BITS64))
1360 size = (opx->type & BITS16) ? 2 : 4;
1361 else
1362 size = (bits == 16) ? 2 : 4;
1363 if (opx->segment != segment) {
1364 data = opx->offset;
1365 out(offset, segment, &data,
1366 size == 2 ? OUT_REL2ADR : OUT_REL4ADR,
1367 insn_end - offset, opx->segment, opx->wrt);
1368 } else {
1369 data = opx->offset - insn_end;
1370 out(offset, segment, &data,
1371 OUT_ADDRESS, size, NO_SEG, NO_SEG);
1372 }
1373 offset += size;
1374 break;
1375
1376 case4(070):
1377 if (opx->segment != segment) {
1378 data = opx->offset;
1379 out(offset, segment, &data,
1380 OUT_REL4ADR, insn_end - offset,
1381 opx->segment, opx->wrt);
1382 } else {
1383 data = opx->offset - insn_end;
1384 out(offset, segment, &data,
1385 OUT_ADDRESS, 4, NO_SEG, NO_SEG);
1386 }
1387 offset += 4;
1388 break;
1389
1390 case4(074):
1391 if (opx->segment == NO_SEG)
1392 errfunc(ERR_NONFATAL, "value referenced by FAR is not"
1393 " relocatable");
1394 data = 0;
1395 out(offset, segment, &data, OUT_ADDRESS, 2,
1396 outfmt->segbase(1 + opx->segment),
1397 opx->wrt);
1398 offset += 2;
1399 break;
1400
1401 case4(0140):
1402 data = opx->offset;
1403 warn_overflow(2, opx);
1404 if (is_sbyte16(opx)) {
1405 bytes[0] = data;
1406 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1407 NO_SEG);
1408 offset++;
1409 } else {
1410 out(offset, segment, &data, OUT_ADDRESS, 2,
1411 opx->segment, opx->wrt);
1412 offset += 2;
1413 }
1414 break;
1415
1416 case4(0144):
1417 EMIT_REX();
1418 bytes[0] = *codes++;
1419 if (is_sbyte16(opx))
1420 bytes[0] |= 2; /* s-bit */
1421 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1422 offset++;
1423 break;
1424
1425 case4(0150):
1426 data = opx->offset;
1427 warn_overflow(4, opx);
1428 if (is_sbyte32(opx)) {
1429 bytes[0] = data;
1430 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1431 NO_SEG);
1432 offset++;
1433 } else {
1434 out(offset, segment, &data, OUT_ADDRESS, 4,
1435 opx->segment, opx->wrt);
1436 offset += 4;
1437 }
1438 break;
1439
1440 case4(0154):
1441 EMIT_REX();
1442 bytes[0] = *codes++;
1443 if (is_sbyte32(opx))
1444 bytes[0] |= 2; /* s-bit */
1445 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1446 offset++;
1447 break;
1448
1449 case4(0160):
1450 case4(0164):
1451 break;
1452
1453 case 0171:
1454 bytes[0] =
1455 (ins->drexdst << 4) |
1456 (ins->rex & REX_OC ? 0x08 : 0) |
1457 (ins->rex & (REX_R|REX_X|REX_B));
1458 ins->rex = 0;
1459 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1460 offset++;
1461 break;
1462
1463 case 0172:
1464 c = *codes++;
1465 opx = &ins->oprs[c >> 3];
1466 bytes[0] = nasm_regvals[opx->basereg] << 4;
1467 opx = &ins->oprs[c & 7];
1468 if (opx->segment != NO_SEG || opx->wrt != NO_SEG) {
1469 errfunc(ERR_NONFATAL,
1470 "non-absolute expression not permitted as argument %d",
1471 c & 7);
1472 } else {
1473 if (opx->offset & ~15) {
1474 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1475 "four-bit argument exceeds bounds");
1476 }
1477 bytes[0] |= opx->offset & 15;
1478 }
1479 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1480 offset++;
1481 break;
1482
1483 case 0173:
1484 c = *codes++;
1485 opx = &ins->oprs[c >> 4];
1486 bytes[0] = nasm_regvals[opx->basereg] << 4;
1487 bytes[0] |= c & 15;
1488 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1489 offset++;
1490 break;
1491
1492 case 0174:
1493 c = *codes++;
1494 opx = &ins->oprs[c];
1495 bytes[0] = nasm_regvals[opx->basereg] << 4;
1496 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1497 offset++;
1498 break;
1499
1500 case4(0250):
1501 data = opx->offset;
1502 if (opx->wrt == NO_SEG && opx->segment == NO_SEG &&
1503 (int32_t)data != (int64_t)data) {
1504 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1505 "signed dword immediate exceeds bounds");
1506 }
1507 if (is_sbyte32(opx)) {
1508 bytes[0] = data;
1509 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1510 NO_SEG);
1511 offset++;
1512 } else {
1513 out(offset, segment, &data, OUT_ADDRESS, 4,
1514 opx->segment, opx->wrt);
1515 offset += 4;
1516 }
1517 break;
1518
1519 case4(0254):
1520 data = opx->offset;
1521 if (opx->wrt == NO_SEG && opx->segment == NO_SEG &&
1522 (int32_t)data != (int64_t)data) {
1523 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1524 "signed dword immediate exceeds bounds");
1525 }
1526 out(offset, segment, &data, OUT_ADDRESS, 4,
1527 opx->segment, opx->wrt);
1528 offset += 4;
1529 break;
1530
1531 case4(0260):
1532 case 0270:
1533 codes += 2;
1534 if (ins->vex_m != 1 || (ins->rex & (REX_W|REX_X|REX_B))) {
1535 bytes[0] = 0xc4;
1536 bytes[1] = ins->vex_m | ((~ins->rex & 7) << 5);
1537 bytes[2] = ((ins->rex & REX_W) << (7-3)) |
1538 ((~ins->drexdst & 15)<< 3) | (ins->vex_wlp & 07);
1539 out(offset, segment, &bytes, OUT_RAWDATA, 3, NO_SEG, NO_SEG);
1540 offset += 3;
1541 } else {
1542 bytes[0] = 0xc5;
1543 bytes[1] = ((~ins->rex & REX_R) << (7-2)) |
1544 ((~ins->drexdst & 15) << 3) | (ins->vex_wlp & 07);
1545 out(offset, segment, &bytes, OUT_RAWDATA, 2, NO_SEG, NO_SEG);
1546 offset += 2;
1547 }
1548 break;
1549
1550 case4(0274):
1551 {
1552 uint64_t uv, um;
1553 int s;
1554
1555 if (ins->rex & REX_W)
1556 s = 64;
1557 else if (ins->prefixes[PPS_OSIZE] == P_O16)
1558 s = 16;
1559 else if (ins->prefixes[PPS_OSIZE] == P_O32)
1560 s = 32;
1561 else
1562 s = bits;
1563
1564 um = (uint64_t)2 << (s-1);
1565 uv = opx->offset;
1566
1567 if (uv > 127 && uv < (uint64_t)-128 &&
1568 (uv < um-128 || uv > um-1)) {
1569 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1570 "signed byte value exceeds bounds");
1571 }
1572 if (opx->segment != NO_SEG) {
1573 data = uv;
1574 out(offset, segment, &data, OUT_ADDRESS, 1,
1575 opx->segment, opx->wrt);
1576 } else {
1577 bytes[0] = uv;
1578 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1579 NO_SEG);
1580 }
1581 offset += 1;
1582 break;
1583 }
1584
1585 case4(0300):
1586 break;
1587
1588 case 0310:
1589 if (bits == 32 && !has_prefix(ins, PPS_ASIZE, P_A16)) {
1590 *bytes = 0x67;
1591 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1592 offset += 1;
1593 } else
1594 offset += 0;
1595 break;
1596
1597 case 0311:
1598 if (bits != 32 && !has_prefix(ins, PPS_ASIZE, P_A32)) {
1599 *bytes = 0x67;
1600 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1601 offset += 1;
1602 } else
1603 offset += 0;
1604 break;
1605
1606 case 0312:
1607 break;
1608
1609 case 0313:
1610 ins->rex = 0;
1611 break;
1612
1613 case4(0314):
1614 break;
1615
1616 case 0320:
1617 if (bits != 16) {
1618 *bytes = 0x66;
1619 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1620 offset += 1;
1621 } else
1622 offset += 0;
1623 break;
1624
1625 case 0321:
1626 if (bits == 16) {
1627 *bytes = 0x66;
1628 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1629 offset += 1;
1630 } else
1631 offset += 0;
1632 break;
1633
1634 case 0322:
1635 case 0323:
1636 break;
1637
1638 case 0324:
1639 ins->rex |= REX_W;
1640 break;
1641
1642 case 0330:
1643 *bytes = *codes++ ^ condval[ins->condition];
1644 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1645 offset += 1;
1646 break;
1647
1648 case 0331:
1649 break;
1650
1651 case 0332:
1652 case 0333:
1653 *bytes = c - 0332 + 0xF2;
1654 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1655 offset += 1;
1656 break;
1657
1658 case 0334:
1659 if (ins->rex & REX_R) {
1660 *bytes = 0xF0;
1661 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1662 offset += 1;
1663 }
1664 ins->rex &= ~(REX_L|REX_R);
1665 break;
1666
1667 case 0335:
1668 break;
1669
1670 case 0336:
1671 case 0337:
1672 break;
1673
1674 case 0340:
1675 if (ins->oprs[0].segment != NO_SEG)
1676 errfunc(ERR_PANIC, "non-constant BSS size in pass two");
1677 else {
1678 int64_t size = ins->oprs[0].offset;
1679 if (size > 0)
1680 out(offset, segment, NULL,
1681 OUT_RESERVE, size, NO_SEG, NO_SEG);
1682 offset += size;
1683 }
1684 break;
1685
1686 case 0341:
1687 break;
1688
1689 case 0344:
1690 case 0345:
1691 bytes[0] = c & 1;
1692 switch (ins->oprs[0].basereg) {
1693 case R_CS:
1694 bytes[0] += 0x0E;
1695 break;
1696 case R_DS:
1697 bytes[0] += 0x1E;
1698 break;
1699 case R_ES:
1700 bytes[0] += 0x06;
1701 break;
1702 case R_SS:
1703 bytes[0] += 0x16;
1704 break;
1705 default:
1706 errfunc(ERR_PANIC,
1707 "bizarre 8086 segment register received");
1708 }
1709 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1710 offset++;
1711 break;
1712
1713 case 0346:
1714 case 0347:
1715 bytes[0] = c & 1;
1716 switch (ins->oprs[0].basereg) {
1717 case R_FS:
1718 bytes[0] += 0xA0;
1719 break;
1720 case R_GS:
1721 bytes[0] += 0xA8;
1722 break;
1723 default:
1724 errfunc(ERR_PANIC,
1725 "bizarre 386 segment register received");
1726 }
1727 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1728 offset++;
1729 break;
1730
1731 case 0360:
1732 break;
1733
1734 case 0361:
1735 bytes[0] = 0x66;
1736 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1737 offset += 1;
1738 break;
1739
1740 case 0362:
1741 case 0363:
1742 bytes[0] = c - 0362 + 0xf2;
1743 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1744 offset += 1;
1745 break;
1746
1747 case 0364:
1748 case 0365:
1749 break;
1750
1751 case 0366:
1752 case 0367:
1753 *bytes = c - 0366 + 0x66;
1754 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1755 offset += 1;
1756 break;
1757
1758 case 0370:
1759 case 0371:
1760 case 0372:
1761 break;
1762
1763 case 0373:
1764 *bytes = bits == 16 ? 3 : 5;
1765 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1766 offset += 1;
1767 break;
1768
1769 case4(0100):
1770 case4(0110):
1771 case4(0120):
1772 case4(0130):
1773 case4(0200):
1774 case4(0204):
1775 case4(0210):
1776 case4(0214):
1777 case4(0220):
1778 case4(0224):
1779 case4(0230):
1780 case4(0234):
1781 {
1782 ea ea_data;
1783 int rfield;
1784 int32_t rflags;
1785 uint8_t *p;
1786 int32_t s;
1787 enum out_type type;
1788 struct operand *opy = &ins->oprs[op2];
1789
1790 if (c <= 0177) {
1791 /* pick rfield from operand b (opx) */
1792 rflags = regflag(opx);
1793 rfield = nasm_regvals[opx->basereg];
1794 } else {
1795 /* rfield is constant */
1796 rflags = 0;
1797 rfield = c & 7;
1798 }
1799
1800 if (!process_ea(opy, &ea_data, bits, ins->addr_size,
1801 rfield, rflags)) {
1802 errfunc(ERR_NONFATAL, "invalid effective address");
1803 }
1804
1805
1806 p = bytes;
1807 *p++ = ea_data.modrm;
1808 if (ea_data.sib_present)
1809 *p++ = ea_data.sib;
1810
1811 /* DREX suffixes come between the SIB and the displacement */
1812 if (ins->rex & REX_D) {
1813 *p++ = (ins->drexdst << 4) |
1814 (ins->rex & REX_OC ? 0x08 : 0) |
1815 (ins->rex & (REX_R|REX_X|REX_B));
1816 ins->rex = 0;
1817 }
1818
1819 s = p - bytes;
1820 out(offset, segment, bytes, OUT_RAWDATA, s, NO_SEG, NO_SEG);
1821
1822 /*
1823 * Make sure the address gets the right offset in case
1824 * the line breaks in the .lst file (BR 1197827)
1825 */
1826 offset += s;
1827 s = 0;
1828
1829 switch (ea_data.bytes) {
1830 case 0:
1831 break;
1832 case 1:
1833 case 2:
1834 case 4:
1835 case 8:
1836 data = opy->offset;
1837 warn_overflow(ea_data.bytes, opy);
1838 s += ea_data.bytes;
1839 if (ea_data.rip) {
1840 if (opy->segment == segment) {
1841 data -= insn_end;
1842 out(offset, segment, &data, OUT_ADDRESS,
1843 ea_data.bytes, NO_SEG, NO_SEG);
1844 } else {
1845 out(offset, segment, &data, OUT_REL4ADR,
1846 insn_end - offset, opy->segment, opy->wrt);
1847 }
1848 } else {
1849 type = OUT_ADDRESS;
1850 out(offset, segment, &data, OUT_ADDRESS,
1851 ea_data.bytes, opy->segment, opy->wrt);
1852 }
1853 break;
1854 default:
1855 /* Impossible! */
1856 errfunc(ERR_PANIC,
1857 "Invalid amount of bytes (%d) for offset?!",
1858 ea_data.bytes);
1859 break;
1860 }
1861 offset += s;
1862 }
1863 break;
1864
1865 default:
1866 errfunc(ERR_PANIC, "internal instruction table corrupt"
1867 ": instruction code 0x%02X given", c);
1868 break;
1869 }
1870 }
1871 }
1872
1873 static int32_t regflag(const operand * o)
1874 {
1875 if (o->basereg < EXPR_REG_START || o->basereg >= REG_ENUM_LIMIT) {
1876 errfunc(ERR_PANIC, "invalid operand passed to regflag()");
1877 }
1878 return nasm_reg_flags[o->basereg];
1879 }
1880
1881 static int32_t regval(const operand * o)
1882 {
1883 if (o->basereg < EXPR_REG_START || o->basereg >= REG_ENUM_LIMIT) {
1884 errfunc(ERR_PANIC, "invalid operand passed to regval()");
1885 }
1886 return nasm_regvals[o->basereg];
1887 }
1888
1889 static int op_rexflags(const operand * o, int mask)
1890 {
1891 int32_t flags;
1892 int val;
1893
1894 if (o->basereg < EXPR_REG_START || o->basereg >= REG_ENUM_LIMIT) {
1895 errfunc(ERR_PANIC, "invalid operand passed to op_rexflags()");
1896 }
1897
1898 flags = nasm_reg_flags[o->basereg];
1899 val = nasm_regvals[o->basereg];
1900
1901 return rexflags(val, flags, mask);
1902 }
1903
1904 static int rexflags(int val, int32_t flags, int mask)
1905 {
1906 int rex = 0;
1907
1908 if (val >= 8)
1909 rex |= REX_B|REX_X|REX_R;
1910 if (flags & BITS64)
1911 rex |= REX_W;
1912 if (!(REG_HIGH & ~flags)) /* AH, CH, DH, BH */
1913 rex |= REX_H;
1914 else if (!(REG8 & ~flags) && val >= 4) /* SPL, BPL, SIL, DIL */
1915 rex |= REX_P;
1916
1917 return rex & mask;
1918 }
1919
1920 static int matches(const struct itemplate *itemp, insn * instruction, int bits)
1921 {
1922 int i, size[MAX_OPERANDS], asize, oprs, ret;
1923
1924 ret = 100;
1925
1926 /*
1927 * Check the opcode
1928 */
1929 if (itemp->opcode != instruction->opcode)
1930 return 0;
1931
1932 /*
1933 * Count the operands
1934 */
1935 if (itemp->operands != instruction->operands)
1936 return 0;
1937
1938 /*
1939 * Check that no spurious colons or TOs are present
1940 */
1941 for (i = 0; i < itemp->operands; i++)
1942 if (instruction->oprs[i].type & ~itemp->opd[i] & (COLON | TO))
1943 return 0;
1944
1945 /*
1946 * Process size flags
1947 */
1948 if (itemp->flags & IF_ARMASK) {
1949 memset(size, 0, sizeof size);
1950
1951 i = ((itemp->flags & IF_ARMASK) >> IF_ARSHFT) - 1;
1952
1953 switch (itemp->flags & IF_SMASK) {
1954 case IF_SB:
1955 size[i] = BITS8;
1956 break;
1957 case IF_SW:
1958 size[i] = BITS16;
1959 break;
1960 case IF_SD:
1961 size[i] = BITS32;
1962 break;
1963 case IF_SQ:
1964 size[i] = BITS64;
1965 break;
1966 case IF_SO:
1967 size[i] = BITS128;
1968 break;
1969 case IF_SY:
1970 size[i] = BITS256;
1971 break;
1972 case IF_SZ:
1973 switch (bits) {
1974 case 16:
1975 size[i] = BITS16;
1976 break;
1977 case 32:
1978 size[i] = BITS32;
1979 break;
1980 case 64:
1981 size[i] = BITS64;
1982 break;
1983 }
1984 break;
1985 default:
1986 break;
1987 }
1988 } else {
1989 asize = 0;
1990 switch (itemp->flags & IF_SMASK) {
1991 case IF_SB:
1992 asize = BITS8;
1993 break;
1994 case IF_SW:
1995 asize = BITS16;
1996 break;
1997 case IF_SD:
1998 asize = BITS32;
1999 break;
2000 case IF_SQ:
2001 asize = BITS64;
2002 break;
2003 case IF_SO:
2004 asize = BITS128;
2005 break;
2006 case IF_SY:
2007 asize = BITS256;
2008 break;
2009 case IF_SZ:
2010 switch (bits) {
2011 case 16:
2012 asize = BITS16;
2013 break;
2014 case 32:
2015 asize = BITS32;
2016 break;
2017 case 64:
2018 asize = BITS64;
2019 break;
2020 }
2021 break;
2022 default:
2023 break;
2024 }
2025 for (i = 0; i < MAX_OPERANDS; i++)
2026 size[i] = asize;
2027 }
2028
2029 /*
2030 * Check that the operand flags all match up
2031 */
2032 for (i = 0; i < itemp->operands; i++) {
2033 int32_t type = instruction->oprs[i].type;
2034 if (!(type & SIZE_MASK))
2035 type |= size[i];
2036
2037 if (itemp->opd[i] & SAME_AS) {
2038 int j = itemp->opd[i] & ~SAME_AS;
2039 if (type != instruction->oprs[j].type ||
2040 instruction->oprs[i].basereg != instruction->oprs[j].basereg)
2041 return 0;
2042 } else if (itemp->opd[i] & ~type ||
2043 ((itemp->opd[i] & SIZE_MASK) &&
2044 ((itemp->opd[i] ^ type) & SIZE_MASK))) {
2045 if ((itemp->opd[i] & ~type & ~SIZE_MASK) ||
2046 (type & SIZE_MASK))
2047 return 0;
2048 else
2049 return 1;
2050 }
2051 }
2052
2053 /*
2054 * Check operand sizes
2055 */
2056 if (itemp->flags & (IF_SM | IF_SM2)) {
2057 oprs = (itemp->flags & IF_SM2 ? 2 : itemp->operands);
2058 asize = 0;
2059 for (i = 0; i < oprs; i++) {
2060 if ((asize = itemp->opd[i] & SIZE_MASK) != 0) {
2061 int j;
2062 for (j = 0; j < oprs; j++)
2063 size[j] = asize;
2064 break;
2065 }
2066 }
2067 } else {
2068 oprs = itemp->operands;
2069 }
2070
2071 for (i = 0; i < itemp->operands; i++) {
2072 if (!(itemp->opd[i] & SIZE_MASK) &&
2073 (instruction->oprs[i].type & SIZE_MASK & ~size[i]))
2074 return 2;
2075 }
2076
2077 /*
2078 * Check template is okay at the set cpu level
2079 */
2080 if (((itemp->flags & IF_PLEVEL) > cpu))
2081 return 3;
2082
2083 /*
2084 * Verify the appropriate long mode flag.
2085 */
2086 if ((itemp->flags & (bits == 64 ? IF_NOLONG : IF_LONG)))
2087 return 4;
2088
2089 /*
2090 * Check if special handling needed for Jumps
2091 */
2092 if ((uint8_t)(itemp->code[0]) >= 0370)
2093 return 99;
2094
2095 return ret;
2096 }
2097
2098 static ea *process_ea(operand * input, ea * output, int bits,
2099 int addrbits, int rfield, int32_t rflags)
2100 {
2101 bool forw_ref = !!(input->opflags & OPFLAG_FORWARD);
2102
2103 output->rip = false;
2104
2105 /* REX flags for the rfield operand */
2106 output->rex |= rexflags(rfield, rflags, REX_R|REX_P|REX_W|REX_H);
2107
2108 if (!(REGISTER & ~input->type)) { /* register direct */
2109 int i;
2110 int32_t f;
2111
2112 if (input->basereg < EXPR_REG_START /* Verify as Register */
2113 || input->basereg >= REG_ENUM_LIMIT)
2114 return NULL;
2115 f = regflag(input);
2116 i = nasm_regvals[input->basereg];
2117
2118 if (REG_EA & ~f)
2119 return NULL; /* Invalid EA register */
2120
2121 output->rex |= op_rexflags(input, REX_B|REX_P|REX_W|REX_H);
2122
2123 output->sib_present = false; /* no SIB necessary */
2124 output->bytes = 0; /* no offset necessary either */
2125 output->modrm = 0xC0 | ((rfield & 7) << 3) | (i & 7);
2126 } else { /* it's a memory reference */
2127 if (input->basereg == -1
2128 && (input->indexreg == -1 || input->scale == 0)) {
2129 /* it's a pure offset */
2130 if (bits == 64 && (~input->type & IP_REL)) {
2131 int scale, index, base;
2132 output->sib_present = true;
2133 scale = 0;
2134 index = 4;
2135 base = 5;
2136 output->sib = (scale << 6) | (index << 3) | base;
2137 output->bytes = 4;
2138 output->modrm = 4 | ((rfield & 7) << 3);
2139 output->rip = false;
2140 } else {
2141 output->sib_present = false;
2142 output->bytes = (addrbits != 16 ? 4 : 2);
2143 output->modrm = (addrbits != 16 ? 5 : 6) | ((rfield & 7) << 3);
2144 output->rip = bits == 64;
2145 }
2146 } else { /* it's an indirection */
2147 int i = input->indexreg, b = input->basereg, s = input->scale;
2148 int32_t o = input->offset, seg = input->segment;
2149 int hb = input->hintbase, ht = input->hinttype;
2150 int t;
2151 int it, bt;
2152 int32_t ix, bx; /* register flags */
2153
2154 if (s == 0)
2155 i = -1; /* make this easy, at least */
2156
2157 if (i >= EXPR_REG_START && i < REG_ENUM_LIMIT) {
2158 it = nasm_regvals[i];
2159 ix = nasm_reg_flags[i];
2160 } else {
2161 it = -1;
2162 ix = 0;
2163 }
2164
2165 if (b >= EXPR_REG_START && b < REG_ENUM_LIMIT) {
2166 bt = nasm_regvals[b];
2167 bx = nasm_reg_flags[b];
2168 } else {
2169 bt = -1;
2170 bx = 0;
2171 }
2172
2173 /* check for a 32/64-bit memory reference... */
2174 if ((ix|bx) & (BITS32|BITS64)) {
2175 /* it must be a 32/64-bit memory reference. Firstly we have
2176 * to check that all registers involved are type E/Rxx. */
2177 int32_t sok = BITS32|BITS64;
2178
2179 if (it != -1) {
2180 if (!(REG64 & ~ix) || !(REG32 & ~ix))
2181 sok &= ix;
2182 else
2183 return NULL;
2184 }
2185
2186 if (bt != -1) {
2187 if (REG_GPR & ~bx)
2188 return NULL; /* Invalid register */
2189 if (~sok & bx & SIZE_MASK)
2190 return NULL; /* Invalid size */
2191 sok &= bx;
2192 }
2193
2194 /* While we're here, ensure the user didn't specify
2195 WORD or QWORD. */
2196 if (input->disp_size == 16 || input->disp_size == 64)
2197 return NULL;
2198
2199 if (addrbits == 16 ||
2200 (addrbits == 32 && !(sok & BITS32)) ||
2201 (addrbits == 64 && !(sok & BITS64)))
2202 return NULL;
2203
2204 /* now reorganize base/index */
2205 if (s == 1 && bt != it && bt != -1 && it != -1 &&
2206 ((hb == b && ht == EAH_NOTBASE)
2207 || (hb == i && ht == EAH_MAKEBASE))) {
2208 /* swap if hints say so */
2209 t = bt, bt = it, it = t;
2210 t = bx, bx = ix, ix = t;
2211 }
2212 if (bt == it) /* convert EAX+2*EAX to 3*EAX */
2213 bt = -1, bx = 0, s++;
2214 if (bt == -1 && s == 1 && !(hb == it && ht == EAH_NOTBASE)) {
2215 /* make single reg base, unless hint */
2216 bt = it, bx = ix, it = -1, ix = 0;
2217 }
2218 if (((s == 2 && it != REG_NUM_ESP
2219 && !(input->eaflags & EAF_TIMESTWO)) || s == 3
2220 || s == 5 || s == 9) && bt == -1)
2221 bt = it, bx = ix, s--; /* convert 3*EAX to EAX+2*EAX */
2222 if (it == -1 && (bt & 7) != REG_NUM_ESP
2223 && (input->eaflags & EAF_TIMESTWO))
2224 it = bt, ix = bx, bt = -1, bx = 0, s = 1;
2225 /* convert [NOSPLIT EAX] to sib format with 0x0 displacement */
2226 if (s == 1 && it == REG_NUM_ESP) {
2227 /* swap ESP into base if scale is 1 */
2228 t = it, it = bt, bt = t;
2229 t = ix, ix = bx, bx = t;
2230 }
2231 if (it == REG_NUM_ESP
2232 || (s != 1 && s != 2 && s != 4 && s != 8 && it != -1))
2233 return NULL; /* wrong, for various reasons */
2234
2235 output->rex |= rexflags(it, ix, REX_X);
2236 output->rex |= rexflags(bt, bx, REX_B);
2237
2238 if (it == -1 && (bt & 7) != REG_NUM_ESP) {
2239 /* no SIB needed */
2240 int mod, rm;
2241
2242 if (bt == -1) {
2243 rm = 5;
2244 mod = 0;
2245 } else {
2246 rm = (bt & 7);
2247 if (rm != REG_NUM_EBP && o == 0 &&
2248 seg == NO_SEG && !forw_ref &&
2249 !(input->eaflags &
2250 (EAF_BYTEOFFS | EAF_WORDOFFS)))
2251 mod = 0;
2252 else if (input->eaflags & EAF_BYTEOFFS ||
2253 (o >= -128 && o <= 127 && seg == NO_SEG
2254 && !forw_ref
2255 && !(input->eaflags & EAF_WORDOFFS)))
2256 mod = 1;
2257 else
2258 mod = 2;
2259 }
2260
2261 output->sib_present = false;
2262 output->bytes = (bt == -1 || mod == 2 ? 4 : mod);
2263 output->modrm = (mod << 6) | ((rfield & 7) << 3) | rm;
2264 } else {
2265 /* we need a SIB */
2266 int mod, scale, index, base;
2267
2268 if (it == -1)
2269 index = 4, s = 1;
2270 else
2271 index = (it & 7);
2272
2273 switch (s) {
2274 case 1:
2275 scale = 0;
2276 break;
2277 case 2:
2278 scale = 1;
2279 break;
2280 case 4:
2281 scale = 2;
2282 break;
2283 case 8:
2284 scale = 3;
2285 break;
2286 default: /* then what the smeg is it? */
2287 return NULL; /* panic */
2288 }
2289
2290 if (bt == -1) {
2291 base = 5;
2292 mod = 0;
2293 } else {
2294 base = (bt & 7);
2295 if (base != REG_NUM_EBP && o == 0 &&
2296 seg == NO_SEG && !forw_ref &&
2297 !(input->eaflags &
2298 (EAF_BYTEOFFS | EAF_WORDOFFS)))
2299 mod = 0;
2300 else if (input->eaflags & EAF_BYTEOFFS ||
2301 (o >= -128 && o <= 127 && seg == NO_SEG
2302 && !forw_ref
2303 && !(input->eaflags & EAF_WORDOFFS)))
2304 mod = 1;
2305 else
2306 mod = 2;
2307 }
2308
2309 output->sib_present = true;
2310 output->bytes = (bt == -1 || mod == 2 ? 4 : mod);
2311 output->modrm = (mod << 6) | ((rfield & 7) << 3) | 4;
2312 output->sib = (scale << 6) | (index << 3) | base;
2313 }
2314 } else { /* it's 16-bit */
2315 int mod, rm;
2316
2317 /* check for 64-bit long mode */
2318 if (addrbits == 64)
2319 return NULL;
2320
2321 /* check all registers are BX, BP, SI or DI */
2322 if ((b != -1 && b != R_BP && b != R_BX && b != R_SI
2323 && b != R_DI) || (i != -1 && i != R_BP && i != R_BX
2324 && i != R_SI && i != R_DI))
2325 return NULL;
2326
2327 /* ensure the user didn't specify DWORD/QWORD */
2328 if (input->disp_size == 32 || input->disp_size == 64)
2329 return NULL;
2330
2331 if (s != 1 && i != -1)
2332 return NULL; /* no can do, in 16-bit EA */
2333 if (b == -1 && i != -1) {
2334 int tmp = b;
2335 b = i;
2336 i = tmp;
2337 } /* swap */
2338 if ((b == R_SI || b == R_DI) && i != -1) {
2339 int tmp = b;
2340 b = i;
2341 i = tmp;
2342 }
2343 /* have BX/BP as base, SI/DI index */
2344 if (b == i)
2345 return NULL; /* shouldn't ever happen, in theory */
2346 if (i != -1 && b != -1 &&
2347 (i == R_BP || i == R_BX || b == R_SI || b == R_DI))
2348 return NULL; /* invalid combinations */
2349 if (b == -1) /* pure offset: handled above */
2350 return NULL; /* so if it gets to here, panic! */
2351
2352 rm = -1;
2353 if (i != -1)
2354 switch (i * 256 + b) {
2355 case R_SI * 256 + R_BX:
2356 rm = 0;
2357 break;
2358 case R_DI * 256 + R_BX:
2359 rm = 1;
2360 break;
2361 case R_SI * 256 + R_BP:
2362 rm = 2;
2363 break;
2364 case R_DI * 256 + R_BP:
2365 rm = 3;
2366 break;
2367 } else
2368 switch (b) {
2369 case R_SI:
2370 rm = 4;
2371 break;
2372 case R_DI:
2373 rm = 5;
2374 break;
2375 case R_BP:
2376 rm = 6;
2377 break;
2378 case R_BX:
2379 rm = 7;
2380 break;
2381 }
2382 if (rm == -1) /* can't happen, in theory */
2383 return NULL; /* so panic if it does */
2384
2385 if (o == 0 && seg == NO_SEG && !forw_ref && rm != 6 &&
2386 !(input->eaflags & (EAF_BYTEOFFS | EAF_WORDOFFS)))
2387 mod = 0;
2388 else if (input->eaflags & EAF_BYTEOFFS ||
2389 (o >= -128 && o <= 127 && seg == NO_SEG
2390 && !forw_ref
2391 && !(input->eaflags & EAF_WORDOFFS)))
2392 mod = 1;
2393 else
2394 mod = 2;
2395
2396 output->sib_present = false; /* no SIB - it's 16-bit */
2397 output->bytes = mod; /* bytes of offset needed */
2398 output->modrm = (mod << 6) | ((rfield & 7) << 3) | rm;
2399 }
2400 }
2401 }
2402
2403 output->size = 1 + output->sib_present + output->bytes;
2404 return output;
2405 }
2406
2407 static void add_asp(insn *ins, int addrbits)
2408 {
2409 int j, valid;
2410 int defdisp;
2411
2412 valid = (addrbits == 64) ? 64|32 : 32|16;
2413
2414 switch (ins->prefixes[PPS_ASIZE]) {
2415 case P_A16:
2416 valid &= 16;
2417 break;
2418 case P_A32:
2419 valid &= 32;
2420 break;
2421 case P_A64:
2422 valid &= 64;
2423 break;
2424 case P_ASP:
2425 valid &= (addrbits == 32) ? 16 : 32;
2426 break;
2427 default:
2428 break;
2429 }
2430
2431 for (j = 0; j < ins->operands; j++) {
2432 if (!(MEMORY & ~ins->oprs[j].type)) {
2433 int32_t i, b;
2434
2435 /* Verify as Register */
2436 if (ins->oprs[j].indexreg < EXPR_REG_START
2437 || ins->oprs[j].indexreg >= REG_ENUM_LIMIT)
2438 i = 0;
2439 else
2440 i = nasm_reg_flags[ins->oprs[j].indexreg];
2441
2442 /* Verify as Register */
2443 if (ins->oprs[j].basereg < EXPR_REG_START
2444 || ins->oprs[j].basereg >= REG_ENUM_LIMIT)
2445 b = 0;
2446 else
2447 b = nasm_reg_flags[ins->oprs[j].basereg];
2448
2449 if (ins->oprs[j].scale == 0)
2450 i = 0;
2451
2452 if (!i && !b) {
2453 int ds = ins->oprs[j].disp_size;
2454 if ((addrbits != 64 && ds > 8) ||
2455 (addrbits == 64 && ds == 16))
2456 valid &= ds;
2457 } else {
2458 if (!(REG16 & ~b))
2459 valid &= 16;
2460 if (!(REG32 & ~b))
2461 valid &= 32;
2462 if (!(REG64 & ~b))
2463 valid &= 64;
2464
2465 if (!(REG16 & ~i))
2466 valid &= 16;
2467 if (!(REG32 & ~i))
2468 valid &= 32;
2469 if (!(REG64 & ~i))
2470 valid &= 64;
2471 }
2472 }
2473 }
2474
2475 if (valid & addrbits) {
2476 ins->addr_size = addrbits;
2477 } else if (valid & ((addrbits == 32) ? 16 : 32)) {
2478 /* Add an address size prefix */
2479 enum prefixes pref = (addrbits == 32) ? P_A16 : P_A32;
2480 ins->prefixes[PPS_ASIZE] = pref;
2481 ins->addr_size = (addrbits == 32) ? 16 : 32;
2482 } else {
2483 /* Impossible... */
2484 errfunc(ERR_NONFATAL, "impossible combination of address sizes");
2485 ins->addr_size = addrbits; /* Error recovery */
2486 }
2487
2488 defdisp = ins->addr_size == 16 ? 16 : 32;
2489
2490 for (j = 0; j < ins->operands; j++) {
2491 if (!(MEM_OFFS & ~ins->oprs[j].type) &&
2492 (ins->oprs[j].disp_size ? ins->oprs[j].disp_size : defdisp)
2493 != ins->addr_size) {
2494 /* mem_offs sizes must match the address size; if not,
2495 strip the MEM_OFFS bit and match only EA instructions */
2496 ins->oprs[j].type &= ~(MEM_OFFS & ~MEMORY);
2497 }
2498 }
2499 }
Adding AVX-512 features