?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
insns.dat: fix typos: VCMPORD_SP[SD] entered as VCMPORS_SP[SD]
[nasm/avx512.git] / assemble.c
blobb72e5a6019920fb86a581cb825c693c0852de1c6
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/XOP rather than REX, with the
60 * V field taken from operand 0..3.
61 * \270 - this instruction uses VEX/XOP rather than REX, with the
62 * V field set to 1111b.
63 *
64 * VEX/XOP prefixes are followed by the sequence:
65 * \tmm\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 * t = 0 for VEX (C4/C5), t = 1 for XOP (8F).
73 *
74 * \274..\277 - a signed byte immediate operand, from operand 0..3,
75 * which is to be extended to the operand size.
76 * \310 - indicates fixed 16-bit address size, i.e. optional 0x67.
77 * \311 - indicates fixed 32-bit address size, i.e. optional 0x67.
78 * \312 - (disassembler only) marker on LOOP, LOOPxx instructions.
79 * \313 - indicates fixed 64-bit address size, 0x67 invalid.
80 * \314 - (disassembler only) invalid with REX.B
81 * \315 - (disassembler only) invalid with REX.X
82 * \316 - (disassembler only) invalid with REX.R
83 * \317 - (disassembler only) invalid with REX.W
84 * \320 - indicates fixed 16-bit operand size, i.e. optional 0x66.
85 * \321 - indicates fixed 32-bit operand size, i.e. optional 0x66.
86 * \322 - indicates that this instruction is only valid when the
87 * operand size is the default (instruction to disassembler,
88 * generates no code in the assembler)
89 * \323 - indicates fixed 64-bit operand size, REX on extensions only.
90 * \324 - indicates 64-bit operand size requiring REX prefix.
91 * \330 - a literal byte follows in the code stream, to be added
92 * to the condition code value of the instruction.
93 * \331 - instruction not valid with REP prefix. Hint for
94 * disassembler only; for SSE instructions.
95 * \332 - REP prefix (0xF2 byte) used as opcode extension.
96 * \333 - REP prefix (0xF3 byte) used as opcode extension.
97 * \334 - LOCK prefix used instead of REX.R
98 * \335 - disassemble a rep (0xF3 byte) prefix as repe not rep.
99 * \336 - force a REP(E) prefix (0xF2) even if not specified.
100 * \337 - force a REPNE prefix (0xF3) even if not specified.
101 * \336-\337 are still listed as prefixes in the disassembler.
102 * \340 - reserve <operand 0> bytes of uninitialized storage.
103 * Operand 0 had better be a segmentless constant.
104 * \341 - this instruction needs a WAIT "prefix"
105 * \344,\345 - the PUSH/POP (respectively) codes for CS, DS, ES, SS
106 * (POP is never used for CS) depending on operand 0
107 * \346,\347 - the second byte of PUSH/POP codes for FS, GS, depending
108 * on operand 0
109 * \360 - no SSE prefix (== \364\331)
110 * \361 - 66 SSE prefix (== \366\331)
111 * \362 - F2 SSE prefix (== \364\332)
112 * \363 - F3 SSE prefix (== \364\333)
113 * \364 - operand-size prefix (0x66) not permitted
114 * \365 - address-size prefix (0x67) not permitted
115 * \366 - operand-size prefix (0x66) used as opcode extension
116 * \367 - address-size prefix (0x67) used as opcode extension
117 * \370,\371,\372 - match only if operand 0 meets byte jump criteria.
118 * 370 is used for Jcc, 371 is used for JMP.
119 * \373 - assemble 0x03 if bits==16, 0x05 if bits==32;
120 * used for conditional jump over longer jump
121 */
122
123 #include "compiler.h"
124
125 #include <stdio.h>
126 #include <string.h>
127 #include <inttypes.h>
128
129 #include "nasm.h"
130 #include "nasmlib.h"
131 #include "assemble.h"
132 #include "insns.h"
133 #include "tables.h"
134
135 typedef struct {
136 int sib_present; /* is a SIB byte necessary? */
137 int bytes; /* # of bytes of offset needed */
138 int size; /* lazy - this is sib+bytes+1 */
139 uint8_t modrm, sib, rex, rip; /* the bytes themselves */
140 } ea;
141
142 static uint32_t cpu; /* cpu level received from nasm.c */
143 static efunc errfunc;
144 static struct ofmt *outfmt;
145 static ListGen *list;
146
147 static int64_t calcsize(int32_t, int64_t, int, insn *, const uint8_t *);
148 static void gencode(int32_t segment, int64_t offset, int bits,
149 insn * ins, const struct itemplate *temp,
150 int64_t insn_end);
151 static int matches(const struct itemplate *, insn *, int bits);
152 static int32_t regflag(const operand *);
153 static int32_t regval(const operand *);
154 static int rexflags(int, int32_t, int);
155 static int op_rexflags(const operand *, int);
156 static ea *process_ea(operand *, ea *, int, int, int, int32_t);
157 static void add_asp(insn *, int);
158
159 static int has_prefix(insn * ins, enum prefix_pos pos, enum prefixes prefix)
160 {
161 return ins->prefixes[pos] == prefix;
162 }
163
164 static void assert_no_prefix(insn * ins, enum prefix_pos pos)
165 {
166 if (ins->prefixes[pos])
167 errfunc(ERR_NONFATAL, "invalid %s prefix",
168 prefix_name(ins->prefixes[pos]));
169 }
170
171 static const char *size_name(int size)
172 {
173 switch (size) {
174 case 1:
175 return "byte";
176 case 2:
177 return "word";
178 case 4:
179 return "dword";
180 case 8:
181 return "qword";
182 case 10:
183 return "tword";
184 case 16:
185 return "oword";
186 case 32:
187 return "yword";
188 default:
189 return "???";
190 }
191 }
192
193 static void warn_overflow(int size, const struct operand *o)
194 {
195 if (size < 8 && o->wrt == NO_SEG && o->segment == NO_SEG) {
196 int64_t lim = ((int64_t)1 << (size*8))-1;
197 int64_t data = o->offset;
198
199 if (data < ~lim || data > lim)
200 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
201 "%s data exceeds bounds", size_name(size));
202 }
203 }
204 /*
205 * This routine wrappers the real output format's output routine,
206 * in order to pass a copy of the data off to the listing file
207 * generator at the same time.
208 */
209 static void out(int64_t offset, int32_t segto, const void *data,
210 enum out_type type, uint64_t size,
211 int32_t segment, int32_t wrt)
212 {
213 static int32_t lineno = 0; /* static!!! */
214 static char *lnfname = NULL;
215 uint8_t p[8];
216
217 if (type == OUT_ADDRESS && segment == NO_SEG && wrt == NO_SEG) {
218 /*
219 * This is a non-relocated address, and we're going to
220 * convert it into RAWDATA format.
221 */
222 uint8_t *q = p;
223
224 if (size > 8) {
225 errfunc(ERR_PANIC, "OUT_ADDRESS with size > 8");
226 return;
227 }
228
229 WRITEADDR(q, *(int64_t *)data, size);
230 data = p;
231 type = OUT_RAWDATA;
232 }
233
234 list->output(offset, data, type, size);
235
236 /*
237 * this call to src_get determines when we call the
238 * debug-format-specific "linenum" function
239 * it updates lineno and lnfname to the current values
240 * returning 0 if "same as last time", -2 if lnfname
241 * changed, and the amount by which lineno changed,
242 * if it did. thus, these variables must be static
243 */
244
245 if (src_get(&lineno, &lnfname)) {
246 outfmt->current_dfmt->linenum(lnfname, lineno, segto);
247 }
248
249 outfmt->output(segto, data, type, size, segment, wrt);
250 }
251
252 static bool jmp_match(int32_t segment, int64_t offset, int bits,
253 insn * ins, const uint8_t *code)
254 {
255 int64_t isize;
256 uint8_t c = code[0];
257
258 if ((c != 0370 && c != 0371) || (ins->oprs[0].type & STRICT))
259 return false;
260 if (!optimizing)
261 return false;
262 if (optimizing < 0 && c == 0371)
263 return false;
264
265 isize = calcsize(segment, offset, bits, ins, code);
266
267 if (ins->oprs[0].opflags & OPFLAG_UNKNOWN)
268 /* Be optimistic in pass 1 */
269 return true;
270
271 if (ins->oprs[0].segment != segment)
272 return false;
273
274 isize = ins->oprs[0].offset - offset - isize; /* isize is delta */
275 return (isize >= -128 && isize <= 127); /* is it byte size? */
276 }
277
278 int64_t assemble(int32_t segment, int64_t offset, int bits, uint32_t cp,
279 insn * instruction, struct ofmt *output, efunc error,
280 ListGen * listgen)
281 {
282 const struct itemplate *temp;
283 int j;
284 int size_prob;
285 int64_t insn_end;
286 int32_t itimes;
287 int64_t start = offset;
288 int64_t wsize = 0; /* size for DB etc. */
289
290 errfunc = error; /* to pass to other functions */
291 cpu = cp;
292 outfmt = output; /* likewise */
293 list = listgen; /* and again */
294
295 switch (instruction->opcode) {
296 case -1:
297 return 0;
298 case I_DB:
299 wsize = 1;
300 break;
301 case I_DW:
302 wsize = 2;
303 break;
304 case I_DD:
305 wsize = 4;
306 break;
307 case I_DQ:
308 wsize = 8;
309 break;
310 case I_DT:
311 wsize = 10;
312 break;
313 case I_DO:
314 wsize = 16;
315 break;
316 case I_DY:
317 wsize = 32;
318 break;
319 default:
320 break;
321 }
322
323 if (wsize) {
324 extop *e;
325 int32_t t = instruction->times;
326 if (t < 0)
327 errfunc(ERR_PANIC,
328 "instruction->times < 0 (%ld) in assemble()", t);
329
330 while (t--) { /* repeat TIMES times */
331 for (e = instruction->eops; e; e = e->next) {
332 if (e->type == EOT_DB_NUMBER) {
333 if (wsize == 1) {
334 if (e->segment != NO_SEG)
335 errfunc(ERR_NONFATAL,
336 "one-byte relocation attempted");
337 else {
338 uint8_t out_byte = e->offset;
339 out(offset, segment, &out_byte,
340 OUT_RAWDATA, 1, NO_SEG, NO_SEG);
341 }
342 } else if (wsize > 8) {
343 errfunc(ERR_NONFATAL,
344 "integer supplied to a DT, DO or DY"
345 " instruction");
346 } else
347 out(offset, segment, &e->offset,
348 OUT_ADDRESS, wsize, e->segment, e->wrt);
349 offset += wsize;
350 } else if (e->type == EOT_DB_STRING ||
351 e->type == EOT_DB_STRING_FREE) {
352 int align;
353
354 out(offset, segment, e->stringval,
355 OUT_RAWDATA, e->stringlen, NO_SEG, NO_SEG);
356 align = e->stringlen % wsize;
357
358 if (align) {
359 align = wsize - align;
360 out(offset, segment, zero_buffer,
361 OUT_RAWDATA, align, NO_SEG, NO_SEG);
362 }
363 offset += e->stringlen + align;
364 }
365 }
366 if (t > 0 && t == instruction->times - 1) {
367 /*
368 * Dummy call to list->output to give the offset to the
369 * listing module.
370 */
371 list->output(offset, NULL, OUT_RAWDATA, 0);
372 list->uplevel(LIST_TIMES);
373 }
374 }
375 if (instruction->times > 1)
376 list->downlevel(LIST_TIMES);
377 return offset - start;
378 }
379
380 if (instruction->opcode == I_INCBIN) {
381 const char *fname = instruction->eops->stringval;
382 FILE *fp;
383
384 fp = fopen(fname, "rb");
385 if (!fp) {
386 error(ERR_NONFATAL, "`incbin': unable to open file `%s'",
387 fname);
388 } else if (fseek(fp, 0L, SEEK_END) < 0) {
389 error(ERR_NONFATAL, "`incbin': unable to seek on file `%s'",
390 fname);
391 } else {
392 static char buf[4096];
393 size_t t = instruction->times;
394 size_t base = 0;
395 size_t len;
396
397 len = ftell(fp);
398 if (instruction->eops->next) {
399 base = instruction->eops->next->offset;
400 len -= base;
401 if (instruction->eops->next->next &&
402 len > (size_t)instruction->eops->next->next->offset)
403 len = (size_t)instruction->eops->next->next->offset;
404 }
405 /*
406 * Dummy call to list->output to give the offset to the
407 * listing module.
408 */
409 list->output(offset, NULL, OUT_RAWDATA, 0);
410 list->uplevel(LIST_INCBIN);
411 while (t--) {
412 size_t l;
413
414 fseek(fp, base, SEEK_SET);
415 l = len;
416 while (l > 0) {
417 int32_t m =
418 fread(buf, 1, (l > (int32_t) sizeof(buf) ? (int32_t) sizeof(buf) : l),
419 fp);
420 if (!m) {
421 /*
422 * This shouldn't happen unless the file
423 * actually changes while we are reading
424 * it.
425 */
426 error(ERR_NONFATAL,
427 "`incbin': unexpected EOF while"
428 " reading file `%s'", fname);
429 t = 0; /* Try to exit cleanly */
430 break;
431 }
432 out(offset, segment, buf, OUT_RAWDATA, m,
433 NO_SEG, NO_SEG);
434 l -= m;
435 }
436 }
437 list->downlevel(LIST_INCBIN);
438 if (instruction->times > 1) {
439 /*
440 * Dummy call to list->output to give the offset to the
441 * listing module.
442 */
443 list->output(offset, NULL, OUT_RAWDATA, 0);
444 list->uplevel(LIST_TIMES);
445 list->downlevel(LIST_TIMES);
446 }
447 fclose(fp);
448 return instruction->times * len;
449 }
450 return 0; /* if we're here, there's an error */
451 }
452
453 /* Check to see if we need an address-size prefix */
454 add_asp(instruction, bits);
455
456 size_prob = 0;
457
458 for (temp = nasm_instructions[instruction->opcode]; temp->opcode != -1; temp++){
459 int m = matches(temp, instruction, bits);
460 if (m == 100 ||
461 (m == 99 && jmp_match(segment, offset, bits,
462 instruction, temp->code))) {
463 /* Matches! */
464 int64_t insn_size = calcsize(segment, offset, bits,
465 instruction, temp->code);
466 itimes = instruction->times;
467 if (insn_size < 0) /* shouldn't be, on pass two */
468 error(ERR_PANIC, "errors made it through from pass one");
469 else
470 while (itimes--) {
471 for (j = 0; j < MAXPREFIX; j++) {
472 uint8_t c = 0;
473 switch (instruction->prefixes[j]) {
474 case P_WAIT:
475 c = 0x9B;
476 break;
477 case P_LOCK:
478 c = 0xF0;
479 break;
480 case P_REPNE:
481 case P_REPNZ:
482 c = 0xF2;
483 break;
484 case P_REPE:
485 case P_REPZ:
486 case P_REP:
487 c = 0xF3;
488 break;
489 case R_CS:
490 if (bits == 64) {
491 error(ERR_WARNING | ERR_PASS2,
492 "cs segment base generated, but will be ignored in 64-bit mode");
493 }
494 c = 0x2E;
495 break;
496 case R_DS:
497 if (bits == 64) {
498 error(ERR_WARNING | ERR_PASS2,
499 "ds segment base generated, but will be ignored in 64-bit mode");
500 }
501 c = 0x3E;
502 break;
503 case R_ES:
504 if (bits == 64) {
505 error(ERR_WARNING | ERR_PASS2,
506 "es segment base generated, but will be ignored in 64-bit mode");
507 }
508 c = 0x26;
509 break;
510 case R_FS:
511 c = 0x64;
512 break;
513 case R_GS:
514 c = 0x65;
515 break;
516 case R_SS:
517 if (bits == 64) {
518 error(ERR_WARNING | ERR_PASS2,
519 "ss segment base generated, but will be ignored in 64-bit mode");
520 }
521 c = 0x36;
522 break;
523 case R_SEGR6:
524 case R_SEGR7:
525 error(ERR_NONFATAL,
526 "segr6 and segr7 cannot be used as prefixes");
527 break;
528 case P_A16:
529 if (bits == 64) {
530 error(ERR_NONFATAL,
531 "16-bit addressing is not supported "
532 "in 64-bit mode");
533 } else if (bits != 16)
534 c = 0x67;
535 break;
536 case P_A32:
537 if (bits != 32)
538 c = 0x67;
539 break;
540 case P_A64:
541 if (bits != 64) {
542 error(ERR_NONFATAL,
543 "64-bit addressing is only supported "
544 "in 64-bit mode");
545 }
546 break;
547 case P_ASP:
548 c = 0x67;
549 break;
550 case P_O16:
551 if (bits != 16)
552 c = 0x66;
553 break;
554 case P_O32:
555 if (bits == 16)
556 c = 0x66;
557 break;
558 case P_O64:
559 /* REX.W */
560 break;
561 case P_OSP:
562 c = 0x66;
563 break;
564 case P_none:
565 break;
566 default:
567 error(ERR_PANIC, "invalid instruction prefix");
568 }
569 if (c != 0) {
570 out(offset, segment, &c, OUT_RAWDATA, 1,
571 NO_SEG, NO_SEG);
572 offset++;
573 }
574 }
575 insn_end = offset + insn_size;
576 gencode(segment, offset, bits, instruction,
577 temp, insn_end);
578 offset += insn_size;
579 if (itimes > 0 && itimes == instruction->times - 1) {
580 /*
581 * Dummy call to list->output to give the offset to the
582 * listing module.
583 */
584 list->output(offset, NULL, OUT_RAWDATA, 0);
585 list->uplevel(LIST_TIMES);
586 }
587 }
588 if (instruction->times > 1)
589 list->downlevel(LIST_TIMES);
590 return offset - start;
591 } else if (m > 0 && m > size_prob) {
592 size_prob = m;
593 }
594 }
595
596 if (temp->opcode == -1) { /* didn't match any instruction */
597 switch (size_prob) {
598 case 1:
599 error(ERR_NONFATAL, "operation size not specified");
600 break;
601 case 2:
602 error(ERR_NONFATAL, "mismatch in operand sizes");
603 break;
604 case 3:
605 error(ERR_NONFATAL, "no instruction for this cpu level");
606 break;
607 case 4:
608 error(ERR_NONFATAL, "instruction not supported in %d-bit mode",
609 bits);
610 break;
611 default:
612 error(ERR_NONFATAL,
613 "invalid combination of opcode and operands");
614 break;
615 }
616 }
617 return 0;
618 }
619
620 int64_t insn_size(int32_t segment, int64_t offset, int bits, uint32_t cp,
621 insn * instruction, efunc error)
622 {
623 const struct itemplate *temp;
624
625 errfunc = error; /* to pass to other functions */
626 cpu = cp;
627
628 if (instruction->opcode == -1)
629 return 0;
630
631 if (instruction->opcode == I_DB || instruction->opcode == I_DW ||
632 instruction->opcode == I_DD || instruction->opcode == I_DQ ||
633 instruction->opcode == I_DT || instruction->opcode == I_DO ||
634 instruction->opcode == I_DY) {
635 extop *e;
636 int32_t isize, osize, wsize = 0; /* placate gcc */
637
638 isize = 0;
639 switch (instruction->opcode) {
640 case I_DB:
641 wsize = 1;
642 break;
643 case I_DW:
644 wsize = 2;
645 break;
646 case I_DD:
647 wsize = 4;
648 break;
649 case I_DQ:
650 wsize = 8;
651 break;
652 case I_DT:
653 wsize = 10;
654 break;
655 case I_DO:
656 wsize = 16;
657 break;
658 case I_DY:
659 wsize = 32;
660 break;
661 default:
662 break;
663 }
664
665 for (e = instruction->eops; e; e = e->next) {
666 int32_t align;
667
668 osize = 0;
669 if (e->type == EOT_DB_NUMBER)
670 osize = 1;
671 else if (e->type == EOT_DB_STRING ||
672 e->type == EOT_DB_STRING_FREE)
673 osize = e->stringlen;
674
675 align = (-osize) % wsize;
676 if (align < 0)
677 align += wsize;
678 isize += osize + align;
679 }
680 return isize * instruction->times;
681 }
682
683 if (instruction->opcode == I_INCBIN) {
684 const char *fname = instruction->eops->stringval;
685 FILE *fp;
686 size_t len;
687
688 fp = fopen(fname, "rb");
689 if (!fp)
690 error(ERR_NONFATAL, "`incbin': unable to open file `%s'",
691 fname);
692 else if (fseek(fp, 0L, SEEK_END) < 0)
693 error(ERR_NONFATAL, "`incbin': unable to seek on file `%s'",
694 fname);
695 else {
696 len = ftell(fp);
697 fclose(fp);
698 if (instruction->eops->next) {
699 len -= instruction->eops->next->offset;
700 if (instruction->eops->next->next &&
701 len > (size_t)instruction->eops->next->next->offset) {
702 len = (size_t)instruction->eops->next->next->offset;
703 }
704 }
705 return instruction->times * len;
706 }
707 return 0; /* if we're here, there's an error */
708 }
709
710 /* Check to see if we need an address-size prefix */
711 add_asp(instruction, bits);
712
713 for (temp = nasm_instructions[instruction->opcode]; temp->opcode != -1; temp++) {
714 int m = matches(temp, instruction, bits);
715 if (m == 100 ||
716 (m == 99 && jmp_match(segment, offset, bits,
717 instruction, temp->code))) {
718 /* we've matched an instruction. */
719 int64_t isize;
720 const uint8_t *codes = temp->code;
721 int j;
722
723 isize = calcsize(segment, offset, bits, instruction, codes);
724 if (isize < 0)
725 return -1;
726 for (j = 0; j < MAXPREFIX; j++) {
727 switch (instruction->prefixes[j]) {
728 case P_A16:
729 if (bits != 16)
730 isize++;
731 break;
732 case P_A32:
733 if (bits != 32)
734 isize++;
735 break;
736 case P_O16:
737 if (bits != 16)
738 isize++;
739 break;
740 case P_O32:
741 if (bits == 16)
742 isize++;
743 break;
744 case P_A64:
745 case P_O64:
746 case P_none:
747 break;
748 default:
749 isize++;
750 break;
751 }
752 }
753 return isize * instruction->times;
754 }
755 }
756 return -1; /* didn't match any instruction */
757 }
758
759 static bool possible_sbyte(operand *o)
760 {
761 return o->wrt == NO_SEG && o->segment == NO_SEG &&
762 !(o->opflags & OPFLAG_UNKNOWN) &&
763 optimizing >= 0 && !(o->type & STRICT);
764 }
765
766 /* check that opn[op] is a signed byte of size 16 or 32 */
767 static bool is_sbyte16(operand *o)
768 {
769 int16_t v;
770
771 if (!possible_sbyte(o))
772 return false;
773
774 v = o->offset;
775 return v >= -128 && v <= 127;
776 }
777
778 static bool is_sbyte32(operand *o)
779 {
780 int32_t v;
781
782 if (!possible_sbyte(o))
783 return false;
784
785 v = o->offset;
786 return v >= -128 && v <= 127;
787 }
788
789 /* Common construct */
790 #define case4(x) case (x): case (x)+1: case (x)+2: case (x)+3
791
792 static int64_t calcsize(int32_t segment, int64_t offset, int bits,
793 insn * ins, const uint8_t *codes)
794 {
795 int64_t length = 0;
796 uint8_t c;
797 int rex_mask = ~0;
798 int op1, op2;
799 struct operand *opx;
800 uint8_t opex = 0;
801
802 ins->rex = 0; /* Ensure REX is reset */
803
804 if (ins->prefixes[PPS_OSIZE] == P_O64)
805 ins->rex |= REX_W;
806
807 (void)segment; /* Don't warn that this parameter is unused */
808 (void)offset; /* Don't warn that this parameter is unused */
809
810 while (*codes) {
811 c = *codes++;
812 op1 = (c & 3) + ((opex & 1) << 2);
813 op2 = ((c >> 3) & 3) + ((opex & 2) << 1);
814 opx = &ins->oprs[op1];
815 opex = 0; /* For the next iteration */
816
817 switch (c) {
818 case 01:
819 case 02:
820 case 03:
821 case 04:
822 codes += c, length += c;
823 break;
824
825 case 05:
826 case 06:
827 case 07:
828 opex = c;
829 break;
830
831 case4(010):
832 ins->rex |=
833 op_rexflags(opx, REX_B|REX_H|REX_P|REX_W);
834 codes++, length++;
835 break;
836
837 case4(014):
838 case4(020):
839 case4(024):
840 length++;
841 break;
842
843 case4(030):
844 length += 2;
845 break;
846
847 case4(034):
848 if (opx->type & (BITS16 | BITS32 | BITS64))
849 length += (opx->type & BITS16) ? 2 : 4;
850 else
851 length += (bits == 16) ? 2 : 4;
852 break;
853
854 case4(040):
855 length += 4;
856 break;
857
858 case4(044):
859 length += ins->addr_size >> 3;
860 break;
861
862 case4(050):
863 length++;
864 break;
865
866 case4(054):
867 length += 8; /* MOV reg64/imm */
868 break;
869
870 case4(060):
871 length += 2;
872 break;
873
874 case4(064):
875 if (opx->type & (BITS16 | BITS32 | BITS64))
876 length += (opx->type & BITS16) ? 2 : 4;
877 else
878 length += (bits == 16) ? 2 : 4;
879 break;
880
881 case4(070):
882 length += 4;
883 break;
884
885 case4(074):
886 length += 2;
887 break;
888
889 case4(0140):
890 length += is_sbyte16(opx) ? 1 : 2;
891 break;
892
893 case4(0144):
894 codes++;
895 length++;
896 break;
897
898 case4(0150):
899 length += is_sbyte32(opx) ? 1 : 4;
900 break;
901
902 case4(0154):
903 codes++;
904 length++;
905 break;
906
907 case4(0160):
908 length++;
909 ins->rex |= REX_D;
910 ins->drexdst = regval(opx);
911 break;
912
913 case4(0164):
914 length++;
915 ins->rex |= REX_D|REX_OC;
916 ins->drexdst = regval(opx);
917 break;
918
919 case 0171:
920 break;
921
922 case 0172:
923 case 0173:
924 case 0174:
925 codes++;
926 length++;
927 break;
928
929 case4(0250):
930 length += is_sbyte32(opx) ? 1 : 4;
931 break;
932
933 case4(0254):
934 length += 4;
935 break;
936
937 case4(0260):
938 ins->rex |= REX_V;
939 ins->drexdst = regval(opx);
940 ins->vex_cm = *codes++;
941 ins->vex_wlp = *codes++;
942 break;
943
944 case 0270:
945 ins->rex |= REX_V;
946 ins->drexdst = 0;
947 ins->vex_cm = *codes++;
948 ins->vex_wlp = *codes++;
949 break;
950
951 case4(0274):
952 length++;
953 break;
954
955 case4(0300):
956 break;
957
958 case 0310:
959 if (bits == 64)
960 return -1;
961 length += (bits != 16) && !has_prefix(ins, PPS_ASIZE, P_A16);
962 break;
963
964 case 0311:
965 length += (bits != 32) && !has_prefix(ins, PPS_ASIZE, P_A32);
966 break;
967
968 case 0312:
969 break;
970
971 case 0313:
972 if (bits != 64 || has_prefix(ins, PPS_ASIZE, P_A16) ||
973 has_prefix(ins, PPS_ASIZE, P_A32))
974 return -1;
975 break;
976
977 case4(0314):
978 break;
979
980 case 0320:
981 length += (bits != 16);
982 break;
983
984 case 0321:
985 length += (bits == 16);
986 break;
987
988 case 0322:
989 break;
990
991 case 0323:
992 rex_mask &= ~REX_W;
993 break;
994
995 case 0324:
996 ins->rex |= REX_W;
997 break;
998
999 case 0330:
1000 codes++, length++;
1001 break;
1002
1003 case 0331:
1004 break;
1005
1006 case 0332:
1007 case 0333:
1008 length++;
1009 break;
1010
1011 case 0334:
1012 ins->rex |= REX_L;
1013 break;
1014
1015 case 0335:
1016 break;
1017
1018 case 0336:
1019 if (!ins->prefixes[PPS_LREP])
1020 ins->prefixes[PPS_LREP] = P_REP;
1021 break;
1022
1023 case 0337:
1024 if (!ins->prefixes[PPS_LREP])
1025 ins->prefixes[PPS_LREP] = P_REPNE;
1026 break;
1027
1028 case 0340:
1029 if (ins->oprs[0].segment != NO_SEG)
1030 errfunc(ERR_NONFATAL, "attempt to reserve non-constant"
1031 " quantity of BSS space");
1032 else
1033 length += ins->oprs[0].offset;
1034 break;
1035
1036 case 0341:
1037 if (!ins->prefixes[PPS_WAIT])
1038 ins->prefixes[PPS_WAIT] = P_WAIT;
1039 break;
1040
1041 case4(0344):
1042 length++;
1043 break;
1044
1045 case 0360:
1046 break;
1047
1048 case 0361:
1049 case 0362:
1050 case 0363:
1051 length++;
1052 break;
1053
1054 case 0364:
1055 case 0365:
1056 break;
1057
1058 case 0366:
1059 case 0367:
1060 length++;
1061 break;
1062
1063 case 0370:
1064 case 0371:
1065 case 0372:
1066 break;
1067
1068 case 0373:
1069 length++;
1070 break;
1071
1072 case4(0100):
1073 case4(0110):
1074 case4(0120):
1075 case4(0130):
1076 case4(0200):
1077 case4(0204):
1078 case4(0210):
1079 case4(0214):
1080 case4(0220):
1081 case4(0224):
1082 case4(0230):
1083 case4(0234):
1084 {
1085 ea ea_data;
1086 int rfield;
1087 int32_t rflags;
1088 struct operand *opy = &ins->oprs[op2];
1089
1090 ea_data.rex = 0; /* Ensure ea.REX is initially 0 */
1091
1092 if (c <= 0177) {
1093 /* pick rfield from operand b (opx) */
1094 rflags = regflag(opx);
1095 rfield = nasm_regvals[opx->basereg];
1096 } else {
1097 rflags = 0;
1098 rfield = c & 7;
1099 }
1100 if (!process_ea(opy, &ea_data, bits,
1101 ins->addr_size, rfield, rflags)) {
1102 errfunc(ERR_NONFATAL, "invalid effective address");
1103 return -1;
1104 } else {
1105 ins->rex |= ea_data.rex;
1106 length += ea_data.size;
1107 }
1108 }
1109 break;
1110
1111 default:
1112 errfunc(ERR_PANIC, "internal instruction table corrupt"
1113 ": instruction code \\%o (0x%02X) given", c, c);
1114 break;
1115 }
1116 }
1117
1118 ins->rex &= rex_mask;
1119
1120 if (ins->rex & REX_V) {
1121 int bad32 = REX_R|REX_W|REX_X|REX_B;
1122
1123 if (ins->rex & REX_H) {
1124 errfunc(ERR_NONFATAL, "cannot use high register in vex instruction");
1125 return -1;
1126 }
1127 switch (ins->vex_wlp & 030) {
1128 case 000:
1129 case 020:
1130 ins->rex &= ~REX_W;
1131 break;
1132 case 010:
1133 ins->rex |= REX_W;
1134 bad32 &= ~REX_W;
1135 break;
1136 case 030:
1137 /* Follow REX_W */
1138 break;
1139 }
1140
1141 if (bits != 64 && ((ins->rex & bad32) || ins->drexdst > 7)) {
1142 errfunc(ERR_NONFATAL, "invalid operands in non-64-bit mode");
1143 return -1;
1144 }
1145 if (ins->vex_cm != 1 || (ins->rex & (REX_W|REX_R|REX_B)))
1146 length += 3;
1147 else
1148 length += 2;
1149 } else if (ins->rex & REX_D) {
1150 if (ins->rex & REX_H) {
1151 errfunc(ERR_NONFATAL, "cannot use high register in drex instruction");
1152 return -1;
1153 }
1154 if (bits != 64 && ((ins->rex & (REX_R|REX_W|REX_X|REX_B)) ||
1155 ins->drexdst > 7)) {
1156 errfunc(ERR_NONFATAL, "invalid operands in non-64-bit mode");
1157 return -1;
1158 }
1159 length++;
1160 } else if (ins->rex & REX_REAL) {
1161 if (ins->rex & REX_H) {
1162 errfunc(ERR_NONFATAL, "cannot use high register in rex instruction");
1163 return -1;
1164 } else if (bits == 64) {
1165 length++;
1166 } else if ((ins->rex & REX_L) &&
1167 !(ins->rex & (REX_P|REX_W|REX_X|REX_B)) &&
1168 cpu >= IF_X86_64) {
1169 /* LOCK-as-REX.R */
1170 assert_no_prefix(ins, PPS_LREP);
1171 length++;
1172 } else {
1173 errfunc(ERR_NONFATAL, "invalid operands in non-64-bit mode");
1174 return -1;
1175 }
1176 }
1177
1178 return length;
1179 }
1180
1181 #define EMIT_REX() \
1182 if (!(ins->rex & (REX_D|REX_V)) && (ins->rex & REX_REAL) && (bits == 64)) { \
1183 ins->rex = (ins->rex & REX_REAL)|REX_P; \
1184 out(offset, segment, &ins->rex, OUT_RAWDATA, 1, NO_SEG, NO_SEG); \
1185 ins->rex = 0; \
1186 offset += 1; \
1187 }
1188
1189 static void gencode(int32_t segment, int64_t offset, int bits,
1190 insn * ins, const struct itemplate *temp,
1191 int64_t insn_end)
1192 {
1193 static char condval[] = { /* conditional opcodes */
1194 0x7, 0x3, 0x2, 0x6, 0x2, 0x4, 0xF, 0xD, 0xC, 0xE, 0x6, 0x2,
1195 0x3, 0x7, 0x3, 0x5, 0xE, 0xC, 0xD, 0xF, 0x1, 0xB, 0x9, 0x5,
1196 0x0, 0xA, 0xA, 0xB, 0x8, 0x4
1197 };
1198 uint8_t c;
1199 uint8_t bytes[4];
1200 int64_t size;
1201 int64_t data;
1202 int op1, op2;
1203 struct operand *opx;
1204 const uint8_t *codes = temp->code;
1205 uint8_t opex = 0;
1206
1207 while (*codes) {
1208 c = *codes++;
1209 op1 = (c & 3) + ((opex & 1) << 2);
1210 op2 = ((c >> 3) & 3) + ((opex & 2) << 1);
1211 opx = &ins->oprs[op1];
1212 opex = 0; /* For the next iteration */
1213
1214 switch (c) {
1215 case 01:
1216 case 02:
1217 case 03:
1218 case 04:
1219 EMIT_REX();
1220 out(offset, segment, codes, OUT_RAWDATA, c, NO_SEG, NO_SEG);
1221 codes += c;
1222 offset += c;
1223 break;
1224
1225 case 05:
1226 case 06:
1227 case 07:
1228 opex = c;
1229 break;
1230
1231 case4(010):
1232 EMIT_REX();
1233 bytes[0] = *codes++ + (regval(opx) & 7);
1234 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1235 offset += 1;
1236 break;
1237
1238 case4(014):
1239 /* The test for BITS8 and SBYTE here is intended to avoid
1240 warning on optimizer actions due to SBYTE, while still
1241 warn on explicit BYTE directives. Also warn, obviously,
1242 if the optimizer isn't enabled. */
1243 if (((opx->type & BITS8) ||
1244 !(opx->type & temp->opd[op1] & BYTENESS)) &&
1245 (opx->offset < -128 || opx->offset > 127)) {
1246 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1247 "signed byte value exceeds bounds");
1248 }
1249 if (opx->segment != NO_SEG) {
1250 data = opx->offset;
1251 out(offset, segment, &data, OUT_ADDRESS, 1,
1252 opx->segment, opx->wrt);
1253 } else {
1254 bytes[0] = opx->offset;
1255 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1256 NO_SEG);
1257 }
1258 offset += 1;
1259 break;
1260
1261 case4(020):
1262 if (opx->offset < -256 || opx->offset > 255) {
1263 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1264 "byte value exceeds bounds");
1265 }
1266 if (opx->segment != NO_SEG) {
1267 data = opx->offset;
1268 out(offset, segment, &data, OUT_ADDRESS, 1,
1269 opx->segment, opx->wrt);
1270 } else {
1271 bytes[0] = opx->offset;
1272 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1273 NO_SEG);
1274 }
1275 offset += 1;
1276 break;
1277
1278 case4(024):
1279 if (opx->offset < 0 || opx->offset > 255)
1280 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1281 "unsigned byte value exceeds bounds");
1282 if (opx->segment != NO_SEG) {
1283 data = opx->offset;
1284 out(offset, segment, &data, OUT_ADDRESS, 1,
1285 opx->segment, opx->wrt);
1286 } else {
1287 bytes[0] = opx->offset;
1288 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1289 NO_SEG);
1290 }
1291 offset += 1;
1292 break;
1293
1294 case4(030):
1295 warn_overflow(2, opx);
1296 data = opx->offset;
1297 out(offset, segment, &data, OUT_ADDRESS, 2,
1298 opx->segment, opx->wrt);
1299 offset += 2;
1300 break;
1301
1302 case4(034):
1303 if (opx->type & (BITS16 | BITS32))
1304 size = (opx->type & BITS16) ? 2 : 4;
1305 else
1306 size = (bits == 16) ? 2 : 4;
1307 warn_overflow(size, opx);
1308 data = opx->offset;
1309 out(offset, segment, &data, OUT_ADDRESS, size,
1310 opx->segment, opx->wrt);
1311 offset += size;
1312 break;
1313
1314 case4(040):
1315 warn_overflow(4, opx);
1316 data = opx->offset;
1317 out(offset, segment, &data, OUT_ADDRESS, 4,
1318 opx->segment, opx->wrt);
1319 offset += 4;
1320 break;
1321
1322 case4(044):
1323 data = opx->offset;
1324 size = ins->addr_size >> 3;
1325 warn_overflow(size, opx);
1326 out(offset, segment, &data, OUT_ADDRESS, size,
1327 opx->segment, opx->wrt);
1328 offset += size;
1329 break;
1330
1331 case4(050):
1332 if (opx->segment != segment)
1333 errfunc(ERR_NONFATAL,
1334 "short relative jump outside segment");
1335 data = opx->offset - insn_end;
1336 if (data > 127 || data < -128)
1337 errfunc(ERR_NONFATAL, "short jump is out of range");
1338 bytes[0] = data;
1339 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1340 offset += 1;
1341 break;
1342
1343 case4(054):
1344 data = (int64_t)opx->offset;
1345 out(offset, segment, &data, OUT_ADDRESS, 8,
1346 opx->segment, opx->wrt);
1347 offset += 8;
1348 break;
1349
1350 case4(060):
1351 if (opx->segment != segment) {
1352 data = opx->offset;
1353 out(offset, segment, &data,
1354 OUT_REL2ADR, insn_end - offset,
1355 opx->segment, opx->wrt);
1356 } else {
1357 data = opx->offset - insn_end;
1358 out(offset, segment, &data,
1359 OUT_ADDRESS, 2, NO_SEG, NO_SEG);
1360 }
1361 offset += 2;
1362 break;
1363
1364 case4(064):
1365 if (opx->type & (BITS16 | BITS32 | BITS64))
1366 size = (opx->type & BITS16) ? 2 : 4;
1367 else
1368 size = (bits == 16) ? 2 : 4;
1369 if (opx->segment != segment) {
1370 data = opx->offset;
1371 out(offset, segment, &data,
1372 size == 2 ? OUT_REL2ADR : OUT_REL4ADR,
1373 insn_end - offset, opx->segment, opx->wrt);
1374 } else {
1375 data = opx->offset - insn_end;
1376 out(offset, segment, &data,
1377 OUT_ADDRESS, size, NO_SEG, NO_SEG);
1378 }
1379 offset += size;
1380 break;
1381
1382 case4(070):
1383 if (opx->segment != segment) {
1384 data = opx->offset;
1385 out(offset, segment, &data,
1386 OUT_REL4ADR, insn_end - offset,
1387 opx->segment, opx->wrt);
1388 } else {
1389 data = opx->offset - insn_end;
1390 out(offset, segment, &data,
1391 OUT_ADDRESS, 4, NO_SEG, NO_SEG);
1392 }
1393 offset += 4;
1394 break;
1395
1396 case4(074):
1397 if (opx->segment == NO_SEG)
1398 errfunc(ERR_NONFATAL, "value referenced by FAR is not"
1399 " relocatable");
1400 data = 0;
1401 out(offset, segment, &data, OUT_ADDRESS, 2,
1402 outfmt->segbase(1 + opx->segment),
1403 opx->wrt);
1404 offset += 2;
1405 break;
1406
1407 case4(0140):
1408 data = opx->offset;
1409 warn_overflow(2, opx);
1410 if (is_sbyte16(opx)) {
1411 bytes[0] = data;
1412 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1413 NO_SEG);
1414 offset++;
1415 } else {
1416 out(offset, segment, &data, OUT_ADDRESS, 2,
1417 opx->segment, opx->wrt);
1418 offset += 2;
1419 }
1420 break;
1421
1422 case4(0144):
1423 EMIT_REX();
1424 bytes[0] = *codes++;
1425 if (is_sbyte16(opx))
1426 bytes[0] |= 2; /* s-bit */
1427 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1428 offset++;
1429 break;
1430
1431 case4(0150):
1432 data = opx->offset;
1433 warn_overflow(4, opx);
1434 if (is_sbyte32(opx)) {
1435 bytes[0] = data;
1436 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1437 NO_SEG);
1438 offset++;
1439 } else {
1440 out(offset, segment, &data, OUT_ADDRESS, 4,
1441 opx->segment, opx->wrt);
1442 offset += 4;
1443 }
1444 break;
1445
1446 case4(0154):
1447 EMIT_REX();
1448 bytes[0] = *codes++;
1449 if (is_sbyte32(opx))
1450 bytes[0] |= 2; /* s-bit */
1451 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1452 offset++;
1453 break;
1454
1455 case4(0160):
1456 case4(0164):
1457 break;
1458
1459 case 0171:
1460 bytes[0] =
1461 (ins->drexdst << 4) |
1462 (ins->rex & REX_OC ? 0x08 : 0) |
1463 (ins->rex & (REX_R|REX_X|REX_B));
1464 ins->rex = 0;
1465 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1466 offset++;
1467 break;
1468
1469 case 0172:
1470 c = *codes++;
1471 opx = &ins->oprs[c >> 3];
1472 bytes[0] = nasm_regvals[opx->basereg] << 4;
1473 opx = &ins->oprs[c & 7];
1474 if (opx->segment != NO_SEG || opx->wrt != NO_SEG) {
1475 errfunc(ERR_NONFATAL,
1476 "non-absolute expression not permitted as argument %d",
1477 c & 7);
1478 } else {
1479 if (opx->offset & ~15) {
1480 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1481 "four-bit argument exceeds bounds");
1482 }
1483 bytes[0] |= opx->offset & 15;
1484 }
1485 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1486 offset++;
1487 break;
1488
1489 case 0173:
1490 c = *codes++;
1491 opx = &ins->oprs[c >> 4];
1492 bytes[0] = nasm_regvals[opx->basereg] << 4;
1493 bytes[0] |= c & 15;
1494 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1495 offset++;
1496 break;
1497
1498 case 0174:
1499 c = *codes++;
1500 opx = &ins->oprs[c];
1501 bytes[0] = nasm_regvals[opx->basereg] << 4;
1502 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1503 offset++;
1504 break;
1505
1506 case4(0250):
1507 data = opx->offset;
1508 if (opx->wrt == NO_SEG && opx->segment == NO_SEG &&
1509 (int32_t)data != (int64_t)data) {
1510 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1511 "signed dword immediate exceeds bounds");
1512 }
1513 if (is_sbyte32(opx)) {
1514 bytes[0] = data;
1515 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1516 NO_SEG);
1517 offset++;
1518 } else {
1519 out(offset, segment, &data, OUT_ADDRESS, 4,
1520 opx->segment, opx->wrt);
1521 offset += 4;
1522 }
1523 break;
1524
1525 case4(0254):
1526 data = opx->offset;
1527 if (opx->wrt == NO_SEG && opx->segment == NO_SEG &&
1528 (int32_t)data != (int64_t)data) {
1529 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1530 "signed dword immediate exceeds bounds");
1531 }
1532 out(offset, segment, &data, OUT_ADDRESS, 4,
1533 opx->segment, opx->wrt);
1534 offset += 4;
1535 break;
1536
1537 case4(0260):
1538 case 0270:
1539 codes += 2;
1540 if (ins->vex_cm != 1 || (ins->rex & (REX_W|REX_X|REX_B))) {
1541 bytes[0] = (ins->vex_cm >> 6) ? 0x8f : 0xc4;
1542 bytes[1] = (ins->vex_cm & 31) | ((~ins->rex & 7) << 5);
1543 bytes[2] = ((ins->rex & REX_W) << (7-3)) |
1544 ((~ins->drexdst & 15)<< 3) | (ins->vex_wlp & 07);
1545 out(offset, segment, &bytes, OUT_RAWDATA, 3, NO_SEG, NO_SEG);
1546 offset += 3;
1547 } else {
1548 bytes[0] = 0xc5;
1549 bytes[1] = ((~ins->rex & REX_R) << (7-2)) |
1550 ((~ins->drexdst & 15) << 3) | (ins->vex_wlp & 07);
1551 out(offset, segment, &bytes, OUT_RAWDATA, 2, NO_SEG, NO_SEG);
1552 offset += 2;
1553 }
1554 break;
1555
1556 case4(0274):
1557 {
1558 uint64_t uv, um;
1559 int s;
1560
1561 if (ins->rex & REX_W)
1562 s = 64;
1563 else if (ins->prefixes[PPS_OSIZE] == P_O16)
1564 s = 16;
1565 else if (ins->prefixes[PPS_OSIZE] == P_O32)
1566 s = 32;
1567 else
1568 s = bits;
1569
1570 um = (uint64_t)2 << (s-1);
1571 uv = opx->offset;
1572
1573 if (uv > 127 && uv < (uint64_t)-128 &&
1574 (uv < um-128 || uv > um-1)) {
1575 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1576 "signed byte value exceeds bounds");
1577 }
1578 if (opx->segment != NO_SEG) {
1579 data = uv;
1580 out(offset, segment, &data, OUT_ADDRESS, 1,
1581 opx->segment, opx->wrt);
1582 } else {
1583 bytes[0] = uv;
1584 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1585 NO_SEG);
1586 }
1587 offset += 1;
1588 break;
1589 }
1590
1591 case4(0300):
1592 break;
1593
1594 case 0310:
1595 if (bits == 32 && !has_prefix(ins, PPS_ASIZE, P_A16)) {
1596 *bytes = 0x67;
1597 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1598 offset += 1;
1599 } else
1600 offset += 0;
1601 break;
1602
1603 case 0311:
1604 if (bits != 32 && !has_prefix(ins, PPS_ASIZE, P_A32)) {
1605 *bytes = 0x67;
1606 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1607 offset += 1;
1608 } else
1609 offset += 0;
1610 break;
1611
1612 case 0312:
1613 break;
1614
1615 case 0313:
1616 ins->rex = 0;
1617 break;
1618
1619 case4(0314):
1620 break;
1621
1622 case 0320:
1623 if (bits != 16) {
1624 *bytes = 0x66;
1625 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1626 offset += 1;
1627 } else
1628 offset += 0;
1629 break;
1630
1631 case 0321:
1632 if (bits == 16) {
1633 *bytes = 0x66;
1634 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1635 offset += 1;
1636 } else
1637 offset += 0;
1638 break;
1639
1640 case 0322:
1641 case 0323:
1642 break;
1643
1644 case 0324:
1645 ins->rex |= REX_W;
1646 break;
1647
1648 case 0330:
1649 *bytes = *codes++ ^ condval[ins->condition];
1650 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1651 offset += 1;
1652 break;
1653
1654 case 0331:
1655 break;
1656
1657 case 0332:
1658 case 0333:
1659 *bytes = c - 0332 + 0xF2;
1660 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1661 offset += 1;
1662 break;
1663
1664 case 0334:
1665 if (ins->rex & REX_R) {
1666 *bytes = 0xF0;
1667 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1668 offset += 1;
1669 }
1670 ins->rex &= ~(REX_L|REX_R);
1671 break;
1672
1673 case 0335:
1674 break;
1675
1676 case 0336:
1677 case 0337:
1678 break;
1679
1680 case 0340:
1681 if (ins->oprs[0].segment != NO_SEG)
1682 errfunc(ERR_PANIC, "non-constant BSS size in pass two");
1683 else {
1684 int64_t size = ins->oprs[0].offset;
1685 if (size > 0)
1686 out(offset, segment, NULL,
1687 OUT_RESERVE, size, NO_SEG, NO_SEG);
1688 offset += size;
1689 }
1690 break;
1691
1692 case 0341:
1693 break;
1694
1695 case 0344:
1696 case 0345:
1697 bytes[0] = c & 1;
1698 switch (ins->oprs[0].basereg) {
1699 case R_CS:
1700 bytes[0] += 0x0E;
1701 break;
1702 case R_DS:
1703 bytes[0] += 0x1E;
1704 break;
1705 case R_ES:
1706 bytes[0] += 0x06;
1707 break;
1708 case R_SS:
1709 bytes[0] += 0x16;
1710 break;
1711 default:
1712 errfunc(ERR_PANIC,
1713 "bizarre 8086 segment register received");
1714 }
1715 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1716 offset++;
1717 break;
1718
1719 case 0346:
1720 case 0347:
1721 bytes[0] = c & 1;
1722 switch (ins->oprs[0].basereg) {
1723 case R_FS:
1724 bytes[0] += 0xA0;
1725 break;
1726 case R_GS:
1727 bytes[0] += 0xA8;
1728 break;
1729 default:
1730 errfunc(ERR_PANIC,
1731 "bizarre 386 segment register received");
1732 }
1733 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1734 offset++;
1735 break;
1736
1737 case 0360:
1738 break;
1739
1740 case 0361:
1741 bytes[0] = 0x66;
1742 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1743 offset += 1;
1744 break;
1745
1746 case 0362:
1747 case 0363:
1748 bytes[0] = c - 0362 + 0xf2;
1749 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1750 offset += 1;
1751 break;
1752
1753 case 0364:
1754 case 0365:
1755 break;
1756
1757 case 0366:
1758 case 0367:
1759 *bytes = c - 0366 + 0x66;
1760 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1761 offset += 1;
1762 break;
1763
1764 case 0370:
1765 case 0371:
1766 case 0372:
1767 break;
1768
1769 case 0373:
1770 *bytes = bits == 16 ? 3 : 5;
1771 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1772 offset += 1;
1773 break;
1774
1775 case4(0100):
1776 case4(0110):
1777 case4(0120):
1778 case4(0130):
1779 case4(0200):
1780 case4(0204):
1781 case4(0210):
1782 case4(0214):
1783 case4(0220):
1784 case4(0224):
1785 case4(0230):
1786 case4(0234):
1787 {
1788 ea ea_data;
1789 int rfield;
1790 int32_t rflags;
1791 uint8_t *p;
1792 int32_t s;
1793 enum out_type type;
1794 struct operand *opy = &ins->oprs[op2];
1795
1796 if (c <= 0177) {
1797 /* pick rfield from operand b (opx) */
1798 rflags = regflag(opx);
1799 rfield = nasm_regvals[opx->basereg];
1800 } else {
1801 /* rfield is constant */
1802 rflags = 0;
1803 rfield = c & 7;
1804 }
1805
1806 if (!process_ea(opy, &ea_data, bits, ins->addr_size,
1807 rfield, rflags)) {
1808 errfunc(ERR_NONFATAL, "invalid effective address");
1809 }
1810
1811
1812 p = bytes;
1813 *p++ = ea_data.modrm;
1814 if (ea_data.sib_present)
1815 *p++ = ea_data.sib;
1816
1817 /* DREX suffixes come between the SIB and the displacement */
1818 if (ins->rex & REX_D) {
1819 *p++ = (ins->drexdst << 4) |
1820 (ins->rex & REX_OC ? 0x08 : 0) |
1821 (ins->rex & (REX_R|REX_X|REX_B));
1822 ins->rex = 0;
1823 }
1824
1825 s = p - bytes;
1826 out(offset, segment, bytes, OUT_RAWDATA, s, NO_SEG, NO_SEG);
1827
1828 /*
1829 * Make sure the address gets the right offset in case
1830 * the line breaks in the .lst file (BR 1197827)
1831 */
1832 offset += s;
1833 s = 0;
1834
1835 switch (ea_data.bytes) {
1836 case 0:
1837 break;
1838 case 1:
1839 case 2:
1840 case 4:
1841 case 8:
1842 data = opy->offset;
1843 warn_overflow(ea_data.bytes, opy);
1844 s += ea_data.bytes;
1845 if (ea_data.rip) {
1846 if (opy->segment == segment) {
1847 data -= insn_end;
1848 out(offset, segment, &data, OUT_ADDRESS,
1849 ea_data.bytes, NO_SEG, NO_SEG);
1850 } else {
1851 out(offset, segment, &data, OUT_REL4ADR,
1852 insn_end - offset, opy->segment, opy->wrt);
1853 }
1854 } else {
1855 type = OUT_ADDRESS;
1856 out(offset, segment, &data, OUT_ADDRESS,
1857 ea_data.bytes, opy->segment, opy->wrt);
1858 }
1859 break;
1860 default:
1861 /* Impossible! */
1862 errfunc(ERR_PANIC,
1863 "Invalid amount of bytes (%d) for offset?!",
1864 ea_data.bytes);
1865 break;
1866 }
1867 offset += s;
1868 }
1869 break;
1870
1871 default:
1872 errfunc(ERR_PANIC, "internal instruction table corrupt"
1873 ": instruction code \\%o (0x%02X) given", c, c);
1874 break;
1875 }
1876 }
1877 }
1878
1879 static int32_t regflag(const operand * o)
1880 {
1881 if (o->basereg < EXPR_REG_START || o->basereg >= REG_ENUM_LIMIT) {
1882 errfunc(ERR_PANIC, "invalid operand passed to regflag()");
1883 }
1884 return nasm_reg_flags[o->basereg];
1885 }
1886
1887 static int32_t regval(const operand * o)
1888 {
1889 if (o->basereg < EXPR_REG_START || o->basereg >= REG_ENUM_LIMIT) {
1890 errfunc(ERR_PANIC, "invalid operand passed to regval()");
1891 }
1892 return nasm_regvals[o->basereg];
1893 }
1894
1895 static int op_rexflags(const operand * o, int mask)
1896 {
1897 int32_t flags;
1898 int val;
1899
1900 if (o->basereg < EXPR_REG_START || o->basereg >= REG_ENUM_LIMIT) {
1901 errfunc(ERR_PANIC, "invalid operand passed to op_rexflags()");
1902 }
1903
1904 flags = nasm_reg_flags[o->basereg];
1905 val = nasm_regvals[o->basereg];
1906
1907 return rexflags(val, flags, mask);
1908 }
1909
1910 static int rexflags(int val, int32_t flags, int mask)
1911 {
1912 int rex = 0;
1913
1914 if (val >= 8)
1915 rex |= REX_B|REX_X|REX_R;
1916 if (flags & BITS64)
1917 rex |= REX_W;
1918 if (!(REG_HIGH & ~flags)) /* AH, CH, DH, BH */
1919 rex |= REX_H;
1920 else if (!(REG8 & ~flags) && val >= 4) /* SPL, BPL, SIL, DIL */
1921 rex |= REX_P;
1922
1923 return rex & mask;
1924 }
1925
1926 static int matches(const struct itemplate *itemp, insn * instruction, int bits)
1927 {
1928 int i, size[MAX_OPERANDS], asize, oprs, ret;
1929
1930 ret = 100;
1931
1932 /*
1933 * Check the opcode
1934 */
1935 if (itemp->opcode != instruction->opcode)
1936 return 0;
1937
1938 /*
1939 * Count the operands
1940 */
1941 if (itemp->operands != instruction->operands)
1942 return 0;
1943
1944 /*
1945 * Check that no spurious colons or TOs are present
1946 */
1947 for (i = 0; i < itemp->operands; i++)
1948 if (instruction->oprs[i].type & ~itemp->opd[i] & (COLON | TO))
1949 return 0;
1950
1951 /*
1952 * Process size flags
1953 */
1954 if (itemp->flags & IF_ARMASK) {
1955 memset(size, 0, sizeof size);
1956
1957 i = ((itemp->flags & IF_ARMASK) >> IF_ARSHFT) - 1;
1958
1959 switch (itemp->flags & IF_SMASK) {
1960 case IF_SB:
1961 size[i] = BITS8;
1962 break;
1963 case IF_SW:
1964 size[i] = BITS16;
1965 break;
1966 case IF_SD:
1967 size[i] = BITS32;
1968 break;
1969 case IF_SQ:
1970 size[i] = BITS64;
1971 break;
1972 case IF_SO:
1973 size[i] = BITS128;
1974 break;
1975 case IF_SY:
1976 size[i] = BITS256;
1977 break;
1978 case IF_SZ:
1979 switch (bits) {
1980 case 16:
1981 size[i] = BITS16;
1982 break;
1983 case 32:
1984 size[i] = BITS32;
1985 break;
1986 case 64:
1987 size[i] = BITS64;
1988 break;
1989 }
1990 break;
1991 default:
1992 break;
1993 }
1994 } else {
1995 asize = 0;
1996 switch (itemp->flags & IF_SMASK) {
1997 case IF_SB:
1998 asize = BITS8;
1999 break;
2000 case IF_SW:
2001 asize = BITS16;
2002 break;
2003 case IF_SD:
2004 asize = BITS32;
2005 break;
2006 case IF_SQ:
2007 asize = BITS64;
2008 break;
2009 case IF_SO:
2010 asize = BITS128;
2011 break;
2012 case IF_SY:
2013 asize = BITS256;
2014 break;
2015 case IF_SZ:
2016 switch (bits) {
2017 case 16:
2018 asize = BITS16;
2019 break;
2020 case 32:
2021 asize = BITS32;
2022 break;
2023 case 64:
2024 asize = BITS64;
2025 break;
2026 }
2027 break;
2028 default:
2029 break;
2030 }
2031 for (i = 0; i < MAX_OPERANDS; i++)
2032 size[i] = asize;
2033 }
2034
2035 /*
2036 * Check that the operand flags all match up
2037 */
2038 for (i = 0; i < itemp->operands; i++) {
2039 int32_t type = instruction->oprs[i].type;
2040 if (!(type & SIZE_MASK))
2041 type |= size[i];
2042
2043 if (itemp->opd[i] & SAME_AS) {
2044 int j = itemp->opd[i] & ~SAME_AS;
2045 if (type != instruction->oprs[j].type ||
2046 instruction->oprs[i].basereg != instruction->oprs[j].basereg)
2047 return 0;
2048 } else if (itemp->opd[i] & ~type ||
2049 ((itemp->opd[i] & SIZE_MASK) &&
2050 ((itemp->opd[i] ^ type) & SIZE_MASK))) {
2051 if ((itemp->opd[i] & ~type & ~SIZE_MASK) ||
2052 (type & SIZE_MASK))
2053 return 0;
2054 else
2055 return 1;
2056 }
2057 }
2058
2059 /*
2060 * Check operand sizes
2061 */
2062 if (itemp->flags & (IF_SM | IF_SM2)) {
2063 oprs = (itemp->flags & IF_SM2 ? 2 : itemp->operands);
2064 asize = 0;
2065 for (i = 0; i < oprs; i++) {
2066 if ((asize = itemp->opd[i] & SIZE_MASK) != 0) {
2067 int j;
2068 for (j = 0; j < oprs; j++)
2069 size[j] = asize;
2070 break;
2071 }
2072 }
2073 } else {
2074 oprs = itemp->operands;
2075 }
2076
2077 for (i = 0; i < itemp->operands; i++) {
2078 if (!(itemp->opd[i] & SIZE_MASK) &&
2079 (instruction->oprs[i].type & SIZE_MASK & ~size[i]))
2080 return 2;
2081 }
2082
2083 /*
2084 * Check template is okay at the set cpu level
2085 */
2086 if (((itemp->flags & IF_PLEVEL) > cpu))
2087 return 3;
2088
2089 /*
2090 * Verify the appropriate long mode flag.
2091 */
2092 if ((itemp->flags & (bits == 64 ? IF_NOLONG : IF_LONG)))
2093 return 4;
2094
2095 /*
2096 * Check if special handling needed for Jumps
2097 */
2098 if ((uint8_t)(itemp->code[0]) >= 0370)
2099 return 99;
2100
2101 return ret;
2102 }
2103
2104 static ea *process_ea(operand * input, ea * output, int bits,
2105 int addrbits, int rfield, int32_t rflags)
2106 {
2107 bool forw_ref = !!(input->opflags & OPFLAG_UNKNOWN);
2108
2109 output->rip = false;
2110
2111 /* REX flags for the rfield operand */
2112 output->rex |= rexflags(rfield, rflags, REX_R|REX_P|REX_W|REX_H);
2113
2114 if (!(REGISTER & ~input->type)) { /* register direct */
2115 int i;
2116 int32_t f;
2117
2118 if (input->basereg < EXPR_REG_START /* Verify as Register */
2119 || input->basereg >= REG_ENUM_LIMIT)
2120 return NULL;
2121 f = regflag(input);
2122 i = nasm_regvals[input->basereg];
2123
2124 if (REG_EA & ~f)
2125 return NULL; /* Invalid EA register */
2126
2127 output->rex |= op_rexflags(input, REX_B|REX_P|REX_W|REX_H);
2128
2129 output->sib_present = false; /* no SIB necessary */
2130 output->bytes = 0; /* no offset necessary either */
2131 output->modrm = 0xC0 | ((rfield & 7) << 3) | (i & 7);
2132 } else { /* it's a memory reference */
2133 if (input->basereg == -1
2134 && (input->indexreg == -1 || input->scale == 0)) {
2135 /* it's a pure offset */
2136 if (bits == 64 && (~input->type & IP_REL)) {
2137 int scale, index, base;
2138 output->sib_present = true;
2139 scale = 0;
2140 index = 4;
2141 base = 5;
2142 output->sib = (scale << 6) | (index << 3) | base;
2143 output->bytes = 4;
2144 output->modrm = 4 | ((rfield & 7) << 3);
2145 output->rip = false;
2146 } else {
2147 output->sib_present = false;
2148 output->bytes = (addrbits != 16 ? 4 : 2);
2149 output->modrm = (addrbits != 16 ? 5 : 6) | ((rfield & 7) << 3);
2150 output->rip = bits == 64;
2151 }
2152 } else { /* it's an indirection */
2153 int i = input->indexreg, b = input->basereg, s = input->scale;
2154 int32_t o = input->offset, seg = input->segment;
2155 int hb = input->hintbase, ht = input->hinttype;
2156 int t;
2157 int it, bt;
2158 int32_t ix, bx; /* register flags */
2159
2160 if (s == 0)
2161 i = -1; /* make this easy, at least */
2162
2163 if (i >= EXPR_REG_START && i < REG_ENUM_LIMIT) {
2164 it = nasm_regvals[i];
2165 ix = nasm_reg_flags[i];
2166 } else {
2167 it = -1;
2168 ix = 0;
2169 }
2170
2171 if (b >= EXPR_REG_START && b < REG_ENUM_LIMIT) {
2172 bt = nasm_regvals[b];
2173 bx = nasm_reg_flags[b];
2174 } else {
2175 bt = -1;
2176 bx = 0;
2177 }
2178
2179 /* check for a 32/64-bit memory reference... */
2180 if ((ix|bx) & (BITS32|BITS64)) {
2181 /* it must be a 32/64-bit memory reference. Firstly we have
2182 * to check that all registers involved are type E/Rxx. */
2183 int32_t sok = BITS32|BITS64;
2184
2185 if (it != -1) {
2186 if (!(REG64 & ~ix) || !(REG32 & ~ix))
2187 sok &= ix;
2188 else
2189 return NULL;
2190 }
2191
2192 if (bt != -1) {
2193 if (REG_GPR & ~bx)
2194 return NULL; /* Invalid register */
2195 if (~sok & bx & SIZE_MASK)
2196 return NULL; /* Invalid size */
2197 sok &= bx;
2198 }
2199
2200 /* While we're here, ensure the user didn't specify
2201 WORD or QWORD. */
2202 if (input->disp_size == 16 || input->disp_size == 64)
2203 return NULL;
2204
2205 if (addrbits == 16 ||
2206 (addrbits == 32 && !(sok & BITS32)) ||
2207 (addrbits == 64 && !(sok & BITS64)))
2208 return NULL;
2209
2210 /* now reorganize base/index */
2211 if (s == 1 && bt != it && bt != -1 && it != -1 &&
2212 ((hb == b && ht == EAH_NOTBASE)
2213 || (hb == i && ht == EAH_MAKEBASE))) {
2214 /* swap if hints say so */
2215 t = bt, bt = it, it = t;
2216 t = bx, bx = ix, ix = t;
2217 }
2218 if (bt == it) /* convert EAX+2*EAX to 3*EAX */
2219 bt = -1, bx = 0, s++;
2220 if (bt == -1 && s == 1 && !(hb == it && ht == EAH_NOTBASE)) {
2221 /* make single reg base, unless hint */
2222 bt = it, bx = ix, it = -1, ix = 0;
2223 }
2224 if (((s == 2 && it != REG_NUM_ESP
2225 && !(input->eaflags & EAF_TIMESTWO)) || s == 3
2226 || s == 5 || s == 9) && bt == -1)
2227 bt = it, bx = ix, s--; /* convert 3*EAX to EAX+2*EAX */
2228 if (it == -1 && (bt & 7) != REG_NUM_ESP
2229 && (input->eaflags & EAF_TIMESTWO))
2230 it = bt, ix = bx, bt = -1, bx = 0, s = 1;
2231 /* convert [NOSPLIT EAX] to sib format with 0x0 displacement */
2232 if (s == 1 && it == REG_NUM_ESP) {
2233 /* swap ESP into base if scale is 1 */
2234 t = it, it = bt, bt = t;
2235 t = ix, ix = bx, bx = t;
2236 }
2237 if (it == REG_NUM_ESP
2238 || (s != 1 && s != 2 && s != 4 && s != 8 && it != -1))
2239 return NULL; /* wrong, for various reasons */
2240
2241 output->rex |= rexflags(it, ix, REX_X);
2242 output->rex |= rexflags(bt, bx, REX_B);
2243
2244 if (it == -1 && (bt & 7) != REG_NUM_ESP) {
2245 /* no SIB needed */
2246 int mod, rm;
2247
2248 if (bt == -1) {
2249 rm = 5;
2250 mod = 0;
2251 } else {
2252 rm = (bt & 7);
2253 if (rm != REG_NUM_EBP && o == 0 &&
2254 seg == NO_SEG && !forw_ref &&
2255 !(input->eaflags &
2256 (EAF_BYTEOFFS | EAF_WORDOFFS)))
2257 mod = 0;
2258 else if (input->eaflags & EAF_BYTEOFFS ||
2259 (o >= -128 && o <= 127 && seg == NO_SEG
2260 && !forw_ref
2261 && !(input->eaflags & EAF_WORDOFFS)))
2262 mod = 1;
2263 else
2264 mod = 2;
2265 }
2266
2267 output->sib_present = false;
2268 output->bytes = (bt == -1 || mod == 2 ? 4 : mod);
2269 output->modrm = (mod << 6) | ((rfield & 7) << 3) | rm;
2270 } else {
2271 /* we need a SIB */
2272 int mod, scale, index, base;
2273
2274 if (it == -1)
2275 index = 4, s = 1;
2276 else
2277 index = (it & 7);
2278
2279 switch (s) {
2280 case 1:
2281 scale = 0;
2282 break;
2283 case 2:
2284 scale = 1;
2285 break;
2286 case 4:
2287 scale = 2;
2288 break;
2289 case 8:
2290 scale = 3;
2291 break;
2292 default: /* then what the smeg is it? */
2293 return NULL; /* panic */
2294 }
2295
2296 if (bt == -1) {
2297 base = 5;
2298 mod = 0;
2299 } else {
2300 base = (bt & 7);
2301 if (base != REG_NUM_EBP && o == 0 &&
2302 seg == NO_SEG && !forw_ref &&
2303 !(input->eaflags &
2304 (EAF_BYTEOFFS | EAF_WORDOFFS)))
2305 mod = 0;
2306 else if (input->eaflags & EAF_BYTEOFFS ||
2307 (o >= -128 && o <= 127 && seg == NO_SEG
2308 && !forw_ref
2309 && !(input->eaflags & EAF_WORDOFFS)))
2310 mod = 1;
2311 else
2312 mod = 2;
2313 }
2314
2315 output->sib_present = true;
2316 output->bytes = (bt == -1 || mod == 2 ? 4 : mod);
2317 output->modrm = (mod << 6) | ((rfield & 7) << 3) | 4;
2318 output->sib = (scale << 6) | (index << 3) | base;
2319 }
2320 } else { /* it's 16-bit */
2321 int mod, rm;
2322
2323 /* check for 64-bit long mode */
2324 if (addrbits == 64)
2325 return NULL;
2326
2327 /* check all registers are BX, BP, SI or DI */
2328 if ((b != -1 && b != R_BP && b != R_BX && b != R_SI
2329 && b != R_DI) || (i != -1 && i != R_BP && i != R_BX
2330 && i != R_SI && i != R_DI))
2331 return NULL;
2332
2333 /* ensure the user didn't specify DWORD/QWORD */
2334 if (input->disp_size == 32 || input->disp_size == 64)
2335 return NULL;
2336
2337 if (s != 1 && i != -1)
2338 return NULL; /* no can do, in 16-bit EA */
2339 if (b == -1 && i != -1) {
2340 int tmp = b;
2341 b = i;
2342 i = tmp;
2343 } /* swap */
2344 if ((b == R_SI || b == R_DI) && i != -1) {
2345 int tmp = b;
2346 b = i;
2347 i = tmp;
2348 }
2349 /* have BX/BP as base, SI/DI index */
2350 if (b == i)
2351 return NULL; /* shouldn't ever happen, in theory */
2352 if (i != -1 && b != -1 &&
2353 (i == R_BP || i == R_BX || b == R_SI || b == R_DI))
2354 return NULL; /* invalid combinations */
2355 if (b == -1) /* pure offset: handled above */
2356 return NULL; /* so if it gets to here, panic! */
2357
2358 rm = -1;
2359 if (i != -1)
2360 switch (i * 256 + b) {
2361 case R_SI * 256 + R_BX:
2362 rm = 0;
2363 break;
2364 case R_DI * 256 + R_BX:
2365 rm = 1;
2366 break;
2367 case R_SI * 256 + R_BP:
2368 rm = 2;
2369 break;
2370 case R_DI * 256 + R_BP:
2371 rm = 3;
2372 break;
2373 } else
2374 switch (b) {
2375 case R_SI:
2376 rm = 4;
2377 break;
2378 case R_DI:
2379 rm = 5;
2380 break;
2381 case R_BP:
2382 rm = 6;
2383 break;
2384 case R_BX:
2385 rm = 7;
2386 break;
2387 }
2388 if (rm == -1) /* can't happen, in theory */
2389 return NULL; /* so panic if it does */
2390
2391 if (o == 0 && seg == NO_SEG && !forw_ref && rm != 6 &&
2392 !(input->eaflags & (EAF_BYTEOFFS | EAF_WORDOFFS)))
2393 mod = 0;
2394 else if (input->eaflags & EAF_BYTEOFFS ||
2395 (o >= -128 && o <= 127 && seg == NO_SEG
2396 && !forw_ref
2397 && !(input->eaflags & EAF_WORDOFFS)))
2398 mod = 1;
2399 else
2400 mod = 2;
2401
2402 output->sib_present = false; /* no SIB - it's 16-bit */
2403 output->bytes = mod; /* bytes of offset needed */
2404 output->modrm = (mod << 6) | ((rfield & 7) << 3) | rm;
2405 }
2406 }
2407 }
2408
2409 output->size = 1 + output->sib_present + output->bytes;
2410 return output;
2411 }
2412
2413 static void add_asp(insn *ins, int addrbits)
2414 {
2415 int j, valid;
2416 int defdisp;
2417
2418 valid = (addrbits == 64) ? 64|32 : 32|16;
2419
2420 switch (ins->prefixes[PPS_ASIZE]) {
2421 case P_A16:
2422 valid &= 16;
2423 break;
2424 case P_A32:
2425 valid &= 32;
2426 break;
2427 case P_A64:
2428 valid &= 64;
2429 break;
2430 case P_ASP:
2431 valid &= (addrbits == 32) ? 16 : 32;
2432 break;
2433 default:
2434 break;
2435 }
2436
2437 for (j = 0; j < ins->operands; j++) {
2438 if (!(MEMORY & ~ins->oprs[j].type)) {
2439 int32_t i, b;
2440
2441 /* Verify as Register */
2442 if (ins->oprs[j].indexreg < EXPR_REG_START
2443 || ins->oprs[j].indexreg >= REG_ENUM_LIMIT)
2444 i = 0;
2445 else
2446 i = nasm_reg_flags[ins->oprs[j].indexreg];
2447
2448 /* Verify as Register */
2449 if (ins->oprs[j].basereg < EXPR_REG_START
2450 || ins->oprs[j].basereg >= REG_ENUM_LIMIT)
2451 b = 0;
2452 else
2453 b = nasm_reg_flags[ins->oprs[j].basereg];
2454
2455 if (ins->oprs[j].scale == 0)
2456 i = 0;
2457
2458 if (!i && !b) {
2459 int ds = ins->oprs[j].disp_size;
2460 if ((addrbits != 64 && ds > 8) ||
2461 (addrbits == 64 && ds == 16))
2462 valid &= ds;
2463 } else {
2464 if (!(REG16 & ~b))
2465 valid &= 16;
2466 if (!(REG32 & ~b))
2467 valid &= 32;
2468 if (!(REG64 & ~b))
2469 valid &= 64;
2470
2471 if (!(REG16 & ~i))
2472 valid &= 16;
2473 if (!(REG32 & ~i))
2474 valid &= 32;
2475 if (!(REG64 & ~i))
2476 valid &= 64;
2477 }
2478 }
2479 }
2480
2481 if (valid & addrbits) {
2482 ins->addr_size = addrbits;
2483 } else if (valid & ((addrbits == 32) ? 16 : 32)) {
2484 /* Add an address size prefix */
2485 enum prefixes pref = (addrbits == 32) ? P_A16 : P_A32;
2486 ins->prefixes[PPS_ASIZE] = pref;
2487 ins->addr_size = (addrbits == 32) ? 16 : 32;
2488 } else {
2489 /* Impossible... */
2490 errfunc(ERR_NONFATAL, "impossible combination of address sizes");
2491 ins->addr_size = addrbits; /* Error recovery */
2492 }
2493
2494 defdisp = ins->addr_size == 16 ? 16 : 32;
2495
2496 for (j = 0; j < ins->operands; j++) {
2497 if (!(MEM_OFFS & ~ins->oprs[j].type) &&
2498 (ins->oprs[j].disp_size ? ins->oprs[j].disp_size : defdisp)
2499 != ins->addr_size) {
2500 /* mem_offs sizes must match the address size; if not,
2501 strip the MEM_OFFS bit and match only EA instructions */
2502 ins->oprs[j].type &= ~(MEM_OFFS & ~MEMORY);
2503 }
2504 }
2505 }
Adding AVX-512 features