?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
NASM 2.09rc5
[nasm/avx512.git] / assemble.c
blob39d722d01613c9f035dc998f2b014ba557e004e5
1 /* ----------------------------------------------------------------------- *
2 *
3 * Copyright 1996-2010 The NASM Authors - All Rights Reserved
4 * See the file AUTHORS included with the NASM distribution for
5 * the specific copyright holders.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following
9 * conditions are met:
10 *
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above
14 * copyright notice, this list of conditions and the following
15 * disclaimer in the documentation and/or other materials provided
16 * with the distribution.
17 *
18 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND
19 * CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES,
20 * INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
21 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
22 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
23 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
24 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
25 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
26 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
29 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
30 * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
31 *
32 * ----------------------------------------------------------------------- */
33
34 /*
35 * assemble.c code generation for the Netwide Assembler
36 *
37 * the actual codes (C syntax, i.e. octal):
38 * \0 - terminates the code. (Unless it's a literal of course.)
39 * \1..\4 - that many literal bytes follow in the code stream
40 * \5 - add 4 to the primary operand number (b, low octdigit)
41 * \6 - add 4 to the secondary operand number (a, middle octdigit)
42 * \7 - add 4 to both the primary and the secondary operand number
43 * \10..\13 - a literal byte follows in the code stream, to be added
44 * to the register value of operand 0..3
45 * \14..\17 - a signed byte immediate operand, from operand 0..3
46 * \20..\23 - a byte immediate operand, from operand 0..3
47 * \24..\27 - an unsigned byte immediate operand, from operand 0..3
48 * \30..\33 - a word immediate operand, from operand 0..3
49 * \34..\37 - select between \3[0-3] and \4[0-3] depending on 16/32 bit
50 * assembly mode or the operand-size override on the operand
51 * \40..\43 - a long immediate operand, from operand 0..3
52 * \44..\47 - select between \3[0-3], \4[0-3] and \5[4-7]
53 * depending on the address size of the instruction.
54 * \50..\53 - a byte relative operand, from operand 0..3
55 * \54..\57 - a qword immediate operand, from operand 0..3
56 * \60..\63 - a word relative operand, from operand 0..3
57 * \64..\67 - select between \6[0-3] and \7[0-3] depending on 16/32 bit
58 * assembly mode or the operand-size override on the operand
59 * \70..\73 - a long relative operand, from operand 0..3
60 * \74..\77 - a word constant, from the _segment_ part of operand 0..3
61 * \1ab - a ModRM, calculated on EA in operand a, with the spare
62 * field the register value of operand b.
63 * \140..\143 - an immediate word or signed byte for operand 0..3
64 * \144..\147 - or 2 (s-field) into opcode byte if operand 0..3
65 * is a signed byte rather than a word. Opcode byte follows.
66 * \150..\153 - an immediate dword or signed byte for operand 0..3
67 * \154..\157 - or 2 (s-field) into opcode byte if operand 0..3
68 * is a signed byte rather than a dword. Opcode byte follows.
69 * \160..\163 - this instruction uses DREX rather than REX, with the
70 * OC0 field set to 0, and the dest field taken from
71 * operand 0..3.
72 * \164..\167 - this instruction uses DREX rather than REX, with the
73 * OC0 field set to 1, and the dest field taken from
74 * operand 0..3.
75 * \171 - placement of DREX suffix in the absence of an EA
76 * \172\ab - the register number from operand a in bits 7..4, with
77 * the 4-bit immediate from operand b in bits 3..0.
78 * \173\xab - the register number from operand a in bits 7..4, with
79 * the value b in bits 3..0.
80 * \174\a - the register number from operand a in bits 7..4, and
81 * an arbitrary value in bits 3..0 (assembled as zero.)
82 * \2ab - a ModRM, calculated on EA in operand a, with the spare
83 * field equal to digit b.
84 * \250..\253 - same as \150..\153, except warn if the 64-bit operand
85 * is not equal to the truncated and sign-extended 32-bit
86 * operand; used for 32-bit immediates in 64-bit mode.
87 * \254..\257 - a signed 32-bit operand to be extended to 64 bits.
88 * \260..\263 - this instruction uses VEX/XOP rather than REX, with the
89 * V field taken from operand 0..3.
90 * \270 - this instruction uses VEX/XOP rather than REX, with the
91 * V field set to 1111b.
92 *
93 * VEX/XOP prefixes are followed by the sequence:
94 * \tmm\wlp where mm is the M field; and wlp is:
95 * 00 0ww lpp
96 * [w0] ww = 0 for W = 0
97 * [w1] ww = 1 for W = 1
98 * [wx] ww = 2 for W don't care (always assembled as 0)
99 * [ww] ww = 3 for W used as REX.W
100 *
101 * t = 0 for VEX (C4/C5), t = 1 for XOP (8F).
102 *
103 * \274..\277 - a signed byte immediate operand, from operand 0..3,
104 * which is to be extended to the operand size.
105 * \310 - indicates fixed 16-bit address size, i.e. optional 0x67.
106 * \311 - indicates fixed 32-bit address size, i.e. optional 0x67.
107 * \312 - (disassembler only) invalid with non-default address size.
108 * \313 - indicates fixed 64-bit address size, 0x67 invalid.
109 * \314 - (disassembler only) invalid with REX.B
110 * \315 - (disassembler only) invalid with REX.X
111 * \316 - (disassembler only) invalid with REX.R
112 * \317 - (disassembler only) invalid with REX.W
113 * \320 - indicates fixed 16-bit operand size, i.e. optional 0x66.
114 * \321 - indicates fixed 32-bit operand size, i.e. optional 0x66.
115 * \322 - indicates that this instruction is only valid when the
116 * operand size is the default (instruction to disassembler,
117 * generates no code in the assembler)
118 * \323 - indicates fixed 64-bit operand size, REX on extensions only.
119 * \324 - indicates 64-bit operand size requiring REX prefix.
120 * \325 - instruction which always uses spl/bpl/sil/dil
121 * \330 - a literal byte follows in the code stream, to be added
122 * to the condition code value of the instruction.
123 * \331 - instruction not valid with REP prefix. Hint for
124 * disassembler only; for SSE instructions.
125 * \332 - REP prefix (0xF2 byte) used as opcode extension.
126 * \333 - REP prefix (0xF3 byte) used as opcode extension.
127 * \334 - LOCK prefix used as REX.R (used in non-64-bit mode)
128 * \335 - disassemble a rep (0xF3 byte) prefix as repe not rep.
129 * \336 - force a REP(E) prefix (0xF2) even if not specified.
130 * \337 - force a REPNE prefix (0xF3) even if not specified.
131 * \336-\337 are still listed as prefixes in the disassembler.
132 * \340 - reserve <operand 0> bytes of uninitialized storage.
133 * Operand 0 had better be a segmentless constant.
134 * \341 - this instruction needs a WAIT "prefix"
135 * \344,\345 - the PUSH/POP (respectively) codes for CS, DS, ES, SS
136 * (POP is never used for CS) depending on operand 0
137 * \346,\347 - the second byte of PUSH/POP codes for FS, GS, depending
138 * on operand 0
139 * \360 - no SSE prefix (== \364\331)
140 * \361 - 66 SSE prefix (== \366\331)
141 * \362 - F2 SSE prefix (== \364\332)
142 * \363 - F3 SSE prefix (== \364\333)
143 * \364 - operand-size prefix (0x66) not permitted
144 * \365 - address-size prefix (0x67) not permitted
145 * \366 - operand-size prefix (0x66) used as opcode extension
146 * \367 - address-size prefix (0x67) used as opcode extension
147 * \370,\371,\372 - match only if operand 0 meets byte jump criteria.
148 * 370 is used for Jcc, 371 is used for JMP.
149 * \373 - assemble 0x03 if bits==16, 0x05 if bits==32;
150 * used for conditional jump over longer jump
151 */
152
153 #include "compiler.h"
154
155 #include <stdio.h>
156 #include <string.h>
157 #include <inttypes.h>
158
159 #include "nasm.h"
160 #include "nasmlib.h"
161 #include "assemble.h"
162 #include "insns.h"
163 #include "tables.h"
164
165 enum match_result {
166 /*
167 * Matching errors. These should be sorted so that more specific
168 * errors come later in the sequence.
169 */
170 MERR_INVALOP,
171 MERR_OPSIZEMISSING,
172 MERR_OPSIZEMISMATCH,
173 MERR_BADCPU,
174 MERR_BADMODE,
175 /*
176 * Matching success; the conditional ones first
177 */
178 MOK_JUMP, /* Matching OK but needs jmp_match() */
179 MOK_GOOD /* Matching unconditionally OK */
180 };
181
182 typedef struct {
183 int sib_present; /* is a SIB byte necessary? */
184 int bytes; /* # of bytes of offset needed */
185 int size; /* lazy - this is sib+bytes+1 */
186 uint8_t modrm, sib, rex, rip; /* the bytes themselves */
187 } ea;
188
189 static uint32_t cpu; /* cpu level received from nasm.c */
190 static efunc errfunc;
191 static struct ofmt *outfmt;
192 static ListGen *list;
193
194 static int64_t calcsize(int32_t, int64_t, int, insn *, const uint8_t *);
195 static void gencode(int32_t segment, int64_t offset, int bits,
196 insn * ins, const struct itemplate *temp,
197 int64_t insn_end);
198 static enum match_result find_match(const struct itemplate **tempp,
199 insn *instruction,
200 int32_t segment, int64_t offset, int bits);
201 static enum match_result matches(const struct itemplate *, insn *, int bits);
202 static opflags_t regflag(const operand *);
203 static int32_t regval(const operand *);
204 static int rexflags(int, opflags_t, int);
205 static int op_rexflags(const operand *, int);
206 static ea *process_ea(operand *, ea *, int, int, int, opflags_t);
207 static void add_asp(insn *, int);
208
209 static int has_prefix(insn * ins, enum prefix_pos pos, enum prefixes prefix)
210 {
211 return ins->prefixes[pos] == prefix;
212 }
213
214 static void assert_no_prefix(insn * ins, enum prefix_pos pos)
215 {
216 if (ins->prefixes[pos])
217 errfunc(ERR_NONFATAL, "invalid %s prefix",
218 prefix_name(ins->prefixes[pos]));
219 }
220
221 static const char *size_name(int size)
222 {
223 switch (size) {
224 case 1:
225 return "byte";
226 case 2:
227 return "word";
228 case 4:
229 return "dword";
230 case 8:
231 return "qword";
232 case 10:
233 return "tword";
234 case 16:
235 return "oword";
236 case 32:
237 return "yword";
238 default:
239 return "???";
240 }
241 }
242
243 static void warn_overflow(int pass, int size)
244 {
245 errfunc(ERR_WARNING | pass | ERR_WARN_NOV,
246 "%s data exceeds bounds", size_name(size));
247 }
248
249 static void warn_overflow_const(int64_t data, int size)
250 {
251 if (overflow_general(data, size))
252 warn_overflow(ERR_PASS1, size);
253 }
254
255 static void warn_overflow_opd(const struct operand *o, int size)
256 {
257 if (o->wrt == NO_SEG && o->segment == NO_SEG) {
258 if (overflow_general(o->offset, size))
259 warn_overflow(ERR_PASS2, size);
260 }
261 }
262
263 /*
264 * This routine wrappers the real output format's output routine,
265 * in order to pass a copy of the data off to the listing file
266 * generator at the same time.
267 */
268 static void out(int64_t offset, int32_t segto, const void *data,
269 enum out_type type, uint64_t size,
270 int32_t segment, int32_t wrt)
271 {
272 static int32_t lineno = 0; /* static!!! */
273 static char *lnfname = NULL;
274 uint8_t p[8];
275
276 if (type == OUT_ADDRESS && segment == NO_SEG && wrt == NO_SEG) {
277 /*
278 * This is a non-relocated address, and we're going to
279 * convert it into RAWDATA format.
280 */
281 uint8_t *q = p;
282
283 if (size > 8) {
284 errfunc(ERR_PANIC, "OUT_ADDRESS with size > 8");
285 return;
286 }
287
288 WRITEADDR(q, *(int64_t *)data, size);
289 data = p;
290 type = OUT_RAWDATA;
291 }
292
293 list->output(offset, data, type, size);
294
295 /*
296 * this call to src_get determines when we call the
297 * debug-format-specific "linenum" function
298 * it updates lineno and lnfname to the current values
299 * returning 0 if "same as last time", -2 if lnfname
300 * changed, and the amount by which lineno changed,
301 * if it did. thus, these variables must be static
302 */
303
304 if (src_get(&lineno, &lnfname))
305 outfmt->current_dfmt->linenum(lnfname, lineno, segto);
306
307 outfmt->output(segto, data, type, size, segment, wrt);
308 }
309
310 static bool jmp_match(int32_t segment, int64_t offset, int bits,
311 insn * ins, const uint8_t *code)
312 {
313 int64_t isize;
314 uint8_t c = code[0];
315
316 if ((c != 0370 && c != 0371) || (ins->oprs[0].type & STRICT))
317 return false;
318 if (!optimizing)
319 return false;
320 if (optimizing < 0 && c == 0371)
321 return false;
322
323 isize = calcsize(segment, offset, bits, ins, code);
324
325 if (ins->oprs[0].opflags & OPFLAG_UNKNOWN)
326 /* Be optimistic in pass 1 */
327 return true;
328
329 if (ins->oprs[0].segment != segment)
330 return false;
331
332 isize = ins->oprs[0].offset - offset - isize; /* isize is delta */
333 return (isize >= -128 && isize <= 127); /* is it byte size? */
334 }
335
336 int64_t assemble(int32_t segment, int64_t offset, int bits, uint32_t cp,
337 insn * instruction, struct ofmt *output, efunc error,
338 ListGen * listgen)
339 {
340 const struct itemplate *temp;
341 int j;
342 enum match_result m;
343 int64_t insn_end;
344 int32_t itimes;
345 int64_t start = offset;
346 int64_t wsize; /* size for DB etc. */
347
348 errfunc = error; /* to pass to other functions */
349 cpu = cp;
350 outfmt = output; /* likewise */
351 list = listgen; /* and again */
352
353 wsize = idata_bytes(instruction->opcode);
354 if (wsize == -1)
355 return 0;
356
357 if (wsize) {
358 extop *e;
359 int32_t t = instruction->times;
360 if (t < 0)
361 errfunc(ERR_PANIC,
362 "instruction->times < 0 (%ld) in assemble()", t);
363
364 while (t--) { /* repeat TIMES times */
365 list_for_each(e, instruction->eops) {
366 if (e->type == EOT_DB_NUMBER) {
367 if (wsize > 8) {
368 errfunc(ERR_NONFATAL,
369 "integer supplied to a DT, DO or DY"
370 " instruction");
371 } else {
372 out(offset, segment, &e->offset,
373 OUT_ADDRESS, wsize, e->segment, e->wrt);
374 offset += wsize;
375 }
376 } else if (e->type == EOT_DB_STRING ||
377 e->type == EOT_DB_STRING_FREE) {
378 int align;
379
380 out(offset, segment, e->stringval,
381 OUT_RAWDATA, e->stringlen, NO_SEG, NO_SEG);
382 align = e->stringlen % wsize;
383
384 if (align) {
385 align = wsize - align;
386 out(offset, segment, zero_buffer,
387 OUT_RAWDATA, align, NO_SEG, NO_SEG);
388 }
389 offset += e->stringlen + align;
390 }
391 }
392 if (t > 0 && t == instruction->times - 1) {
393 /*
394 * Dummy call to list->output to give the offset to the
395 * listing module.
396 */
397 list->output(offset, NULL, OUT_RAWDATA, 0);
398 list->uplevel(LIST_TIMES);
399 }
400 }
401 if (instruction->times > 1)
402 list->downlevel(LIST_TIMES);
403 return offset - start;
404 }
405
406 if (instruction->opcode == I_INCBIN) {
407 const char *fname = instruction->eops->stringval;
408 FILE *fp;
409
410 fp = fopen(fname, "rb");
411 if (!fp) {
412 error(ERR_NONFATAL, "`incbin': unable to open file `%s'",
413 fname);
414 } else if (fseek(fp, 0L, SEEK_END) < 0) {
415 error(ERR_NONFATAL, "`incbin': unable to seek on file `%s'",
416 fname);
417 } else {
418 static char buf[4096];
419 size_t t = instruction->times;
420 size_t base = 0;
421 size_t len;
422
423 len = ftell(fp);
424 if (instruction->eops->next) {
425 base = instruction->eops->next->offset;
426 len -= base;
427 if (instruction->eops->next->next &&
428 len > (size_t)instruction->eops->next->next->offset)
429 len = (size_t)instruction->eops->next->next->offset;
430 }
431 /*
432 * Dummy call to list->output to give the offset to the
433 * listing module.
434 */
435 list->output(offset, NULL, OUT_RAWDATA, 0);
436 list->uplevel(LIST_INCBIN);
437 while (t--) {
438 size_t l;
439
440 fseek(fp, base, SEEK_SET);
441 l = len;
442 while (l > 0) {
443 int32_t m;
444 m = fread(buf, 1, l > sizeof(buf) ? sizeof(buf) : l, fp);
445 if (!m) {
446 /*
447 * This shouldn't happen unless the file
448 * actually changes while we are reading
449 * it.
450 */
451 error(ERR_NONFATAL,
452 "`incbin': unexpected EOF while"
453 " reading file `%s'", fname);
454 t = 0; /* Try to exit cleanly */
455 break;
456 }
457 out(offset, segment, buf, OUT_RAWDATA, m,
458 NO_SEG, NO_SEG);
459 l -= m;
460 }
461 }
462 list->downlevel(LIST_INCBIN);
463 if (instruction->times > 1) {
464 /*
465 * Dummy call to list->output to give the offset to the
466 * listing module.
467 */
468 list->output(offset, NULL, OUT_RAWDATA, 0);
469 list->uplevel(LIST_TIMES);
470 list->downlevel(LIST_TIMES);
471 }
472 fclose(fp);
473 return instruction->times * len;
474 }
475 return 0; /* if we're here, there's an error */
476 }
477
478 /* Check to see if we need an address-size prefix */
479 add_asp(instruction, bits);
480
481 m = find_match(&temp, instruction, segment, offset, bits);
482
483 if (m == MOK_GOOD) {
484 /* Matches! */
485 int64_t insn_size = calcsize(segment, offset, bits,
486 instruction, temp->code);
487 itimes = instruction->times;
488 if (insn_size < 0) /* shouldn't be, on pass two */
489 error(ERR_PANIC, "errors made it through from pass one");
490 else
491 while (itimes--) {
492 for (j = 0; j < MAXPREFIX; j++) {
493 uint8_t c = 0;
494 switch (instruction->prefixes[j]) {
495 case P_WAIT:
496 c = 0x9B;
497 break;
498 case P_LOCK:
499 c = 0xF0;
500 break;
501 case P_REPNE:
502 case P_REPNZ:
503 c = 0xF2;
504 break;
505 case P_REPE:
506 case P_REPZ:
507 case P_REP:
508 c = 0xF3;
509 break;
510 case R_CS:
511 if (bits == 64) {
512 error(ERR_WARNING | ERR_PASS2,
513 "cs segment base generated, but will be ignored in 64-bit mode");
514 }
515 c = 0x2E;
516 break;
517 case R_DS:
518 if (bits == 64) {
519 error(ERR_WARNING | ERR_PASS2,
520 "ds segment base generated, but will be ignored in 64-bit mode");
521 }
522 c = 0x3E;
523 break;
524 case R_ES:
525 if (bits == 64) {
526 error(ERR_WARNING | ERR_PASS2,
527 "es segment base generated, but will be ignored in 64-bit mode");
528 }
529 c = 0x26;
530 break;
531 case R_FS:
532 c = 0x64;
533 break;
534 case R_GS:
535 c = 0x65;
536 break;
537 case R_SS:
538 if (bits == 64) {
539 error(ERR_WARNING | ERR_PASS2,
540 "ss segment base generated, but will be ignored in 64-bit mode");
541 }
542 c = 0x36;
543 break;
544 case R_SEGR6:
545 case R_SEGR7:
546 error(ERR_NONFATAL,
547 "segr6 and segr7 cannot be used as prefixes");
548 break;
549 case P_A16:
550 if (bits == 64) {
551 error(ERR_NONFATAL,
552 "16-bit addressing is not supported "
553 "in 64-bit mode");
554 } else if (bits != 16)
555 c = 0x67;
556 break;
557 case P_A32:
558 if (bits != 32)
559 c = 0x67;
560 break;
561 case P_A64:
562 if (bits != 64) {
563 error(ERR_NONFATAL,
564 "64-bit addressing is only supported "
565 "in 64-bit mode");
566 }
567 break;
568 case P_ASP:
569 c = 0x67;
570 break;
571 case P_O16:
572 if (bits != 16)
573 c = 0x66;
574 break;
575 case P_O32:
576 if (bits == 16)
577 c = 0x66;
578 break;
579 case P_O64:
580 /* REX.W */
581 break;
582 case P_OSP:
583 c = 0x66;
584 break;
585 case P_none:
586 break;
587 default:
588 error(ERR_PANIC, "invalid instruction prefix");
589 }
590 if (c != 0) {
591 out(offset, segment, &c, OUT_RAWDATA, 1,
592 NO_SEG, NO_SEG);
593 offset++;
594 }
595 }
596 insn_end = offset + insn_size;
597 gencode(segment, offset, bits, instruction,
598 temp, insn_end);
599 offset += insn_size;
600 if (itimes > 0 && itimes == instruction->times - 1) {
601 /*
602 * Dummy call to list->output to give the offset to the
603 * listing module.
604 */
605 list->output(offset, NULL, OUT_RAWDATA, 0);
606 list->uplevel(LIST_TIMES);
607 }
608 }
609 if (instruction->times > 1)
610 list->downlevel(LIST_TIMES);
611 return offset - start;
612 } else {
613 /* No match */
614 switch (m) {
615 case MERR_OPSIZEMISSING:
616 error(ERR_NONFATAL, "operation size not specified");
617 break;
618 case MERR_OPSIZEMISMATCH:
619 error(ERR_NONFATAL, "mismatch in operand sizes");
620 break;
621 case MERR_BADCPU:
622 error(ERR_NONFATAL, "no instruction for this cpu level");
623 break;
624 case MERR_BADMODE:
625 error(ERR_NONFATAL, "instruction not supported in %d-bit mode",
626 bits);
627 break;
628 default:
629 error(ERR_NONFATAL,
630 "invalid combination of opcode and operands");
631 break;
632 }
633 }
634 return 0;
635 }
636
637 int64_t insn_size(int32_t segment, int64_t offset, int bits, uint32_t cp,
638 insn * instruction, efunc error)
639 {
640 const struct itemplate *temp;
641 enum match_result m;
642
643 errfunc = error; /* to pass to other functions */
644 cpu = cp;
645
646 if (instruction->opcode == I_none)
647 return 0;
648
649 if (instruction->opcode == I_DB || instruction->opcode == I_DW ||
650 instruction->opcode == I_DD || instruction->opcode == I_DQ ||
651 instruction->opcode == I_DT || instruction->opcode == I_DO ||
652 instruction->opcode == I_DY) {
653 extop *e;
654 int32_t isize, osize, wsize;
655
656 isize = 0;
657 wsize = idata_bytes(instruction->opcode);
658
659 list_for_each(e, instruction->eops) {
660 int32_t align;
661
662 osize = 0;
663 if (e->type == EOT_DB_NUMBER) {
664 osize = 1;
665 warn_overflow_const(e->offset, wsize);
666 } else if (e->type == EOT_DB_STRING ||
667 e->type == EOT_DB_STRING_FREE)
668 osize = e->stringlen;
669
670 align = (-osize) % wsize;
671 if (align < 0)
672 align += wsize;
673 isize += osize + align;
674 }
675 return isize * instruction->times;
676 }
677
678 if (instruction->opcode == I_INCBIN) {
679 const char *fname = instruction->eops->stringval;
680 FILE *fp;
681 int64_t val = 0;
682 size_t len;
683
684 fp = fopen(fname, "rb");
685 if (!fp)
686 error(ERR_NONFATAL, "`incbin': unable to open file `%s'",
687 fname);
688 else if (fseek(fp, 0L, SEEK_END) < 0)
689 error(ERR_NONFATAL, "`incbin': unable to seek on file `%s'",
690 fname);
691 else {
692 len = ftell(fp);
693 if (instruction->eops->next) {
694 len -= instruction->eops->next->offset;
695 if (instruction->eops->next->next &&
696 len > (size_t)instruction->eops->next->next->offset) {
697 len = (size_t)instruction->eops->next->next->offset;
698 }
699 }
700 val = instruction->times * len;
701 }
702 if (fp)
703 fclose(fp);
704 return val;
705 }
706
707 /* Check to see if we need an address-size prefix */
708 add_asp(instruction, bits);
709
710 m = find_match(&temp, instruction, segment, offset, bits);
711 if (m == MOK_GOOD) {
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 } else {
749 return -1; /* didn't match any instruction */
750 }
751 }
752
753 static bool possible_sbyte(operand *o)
754 {
755 return o->wrt == NO_SEG && o->segment == NO_SEG &&
756 !(o->opflags & OPFLAG_UNKNOWN) &&
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_cm = *codes++;
935 ins->vex_wlp = *codes++;
936 break;
937
938 case 0270:
939 ins->rex |= REX_V;
940 ins->drexdst = 0;
941 ins->vex_cm = *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 0325:
994 ins->rex |= REX_NH;
995 break;
996
997 case 0330:
998 codes++, length++;
999 break;
1000
1001 case 0331:
1002 break;
1003
1004 case 0332:
1005 case 0333:
1006 length++;
1007 break;
1008
1009 case 0334:
1010 ins->rex |= REX_L;
1011 break;
1012
1013 case 0335:
1014 break;
1015
1016 case 0336:
1017 if (!ins->prefixes[PPS_LREP])
1018 ins->prefixes[PPS_LREP] = P_REP;
1019 break;
1020
1021 case 0337:
1022 if (!ins->prefixes[PPS_LREP])
1023 ins->prefixes[PPS_LREP] = P_REPNE;
1024 break;
1025
1026 case 0340:
1027 if (ins->oprs[0].segment != NO_SEG)
1028 errfunc(ERR_NONFATAL, "attempt to reserve non-constant"
1029 " quantity of BSS space");
1030 else
1031 length += ins->oprs[0].offset;
1032 break;
1033
1034 case 0341:
1035 if (!ins->prefixes[PPS_WAIT])
1036 ins->prefixes[PPS_WAIT] = P_WAIT;
1037 break;
1038
1039 case4(0344):
1040 length++;
1041 break;
1042
1043 case 0360:
1044 break;
1045
1046 case 0361:
1047 case 0362:
1048 case 0363:
1049 length++;
1050 break;
1051
1052 case 0364:
1053 case 0365:
1054 break;
1055
1056 case 0366:
1057 case 0367:
1058 length++;
1059 break;
1060
1061 case 0370:
1062 case 0371:
1063 case 0372:
1064 break;
1065
1066 case 0373:
1067 length++;
1068 break;
1069
1070 case4(0100):
1071 case4(0110):
1072 case4(0120):
1073 case4(0130):
1074 case4(0200):
1075 case4(0204):
1076 case4(0210):
1077 case4(0214):
1078 case4(0220):
1079 case4(0224):
1080 case4(0230):
1081 case4(0234):
1082 {
1083 ea ea_data;
1084 int rfield;
1085 opflags_t rflags;
1086 struct operand *opy = &ins->oprs[op2];
1087
1088 ea_data.rex = 0; /* Ensure ea.REX is initially 0 */
1089
1090 if (c <= 0177) {
1091 /* pick rfield from operand b (opx) */
1092 rflags = regflag(opx);
1093 rfield = nasm_regvals[opx->basereg];
1094 } else {
1095 rflags = 0;
1096 rfield = c & 7;
1097 }
1098 if (!process_ea(opy, &ea_data, bits,
1099 ins->addr_size, rfield, rflags)) {
1100 errfunc(ERR_NONFATAL, "invalid effective address");
1101 return -1;
1102 } else {
1103 ins->rex |= ea_data.rex;
1104 length += ea_data.size;
1105 }
1106 }
1107 break;
1108
1109 default:
1110 errfunc(ERR_PANIC, "internal instruction table corrupt"
1111 ": instruction code \\%o (0x%02X) given", c, c);
1112 break;
1113 }
1114 }
1115
1116 ins->rex &= rex_mask;
1117
1118 if (ins->rex & REX_NH) {
1119 if (ins->rex & REX_H) {
1120 errfunc(ERR_NONFATAL, "instruction cannot use high registers");
1121 return -1;
1122 }
1123 ins->rex &= ~REX_P; /* Don't force REX prefix due to high reg */
1124 }
1125
1126 if (ins->rex & REX_V) {
1127 int bad32 = REX_R|REX_W|REX_X|REX_B;
1128
1129 if (ins->rex & REX_H) {
1130 errfunc(ERR_NONFATAL, "cannot use high register in vex instruction");
1131 return -1;
1132 }
1133 switch (ins->vex_wlp & 030) {
1134 case 000:
1135 case 020:
1136 ins->rex &= ~REX_W;
1137 break;
1138 case 010:
1139 ins->rex |= REX_W;
1140 bad32 &= ~REX_W;
1141 break;
1142 case 030:
1143 /* Follow REX_W */
1144 break;
1145 }
1146
1147 if (bits != 64 && ((ins->rex & bad32) || ins->drexdst > 7)) {
1148 errfunc(ERR_NONFATAL, "invalid operands in non-64-bit mode");
1149 return -1;
1150 }
1151 if (ins->vex_cm != 1 || (ins->rex & (REX_W|REX_R|REX_B)))
1152 length += 3;
1153 else
1154 length += 2;
1155 } else if (ins->rex & REX_D) {
1156 if (ins->rex & REX_H) {
1157 errfunc(ERR_NONFATAL, "cannot use high register in drex instruction");
1158 return -1;
1159 }
1160 if (bits != 64 && ((ins->rex & (REX_R|REX_W|REX_X|REX_B)) ||
1161 ins->drexdst > 7)) {
1162 errfunc(ERR_NONFATAL, "invalid operands in non-64-bit mode");
1163 return -1;
1164 }
1165 length++;
1166 } else if (ins->rex & REX_REAL) {
1167 if (ins->rex & REX_H) {
1168 errfunc(ERR_NONFATAL, "cannot use high register in rex instruction");
1169 return -1;
1170 } else if (bits == 64) {
1171 length++;
1172 } else if ((ins->rex & REX_L) &&
1173 !(ins->rex & (REX_P|REX_W|REX_X|REX_B)) &&
1174 cpu >= IF_X86_64) {
1175 /* LOCK-as-REX.R */
1176 assert_no_prefix(ins, PPS_LREP);
1177 length++;
1178 } else {
1179 errfunc(ERR_NONFATAL, "invalid operands in non-64-bit mode");
1180 return -1;
1181 }
1182 }
1183
1184 return length;
1185 }
1186
1187 #define EMIT_REX() \
1188 if (!(ins->rex & (REX_D|REX_V)) && (ins->rex & REX_REAL) && (bits == 64)) { \
1189 ins->rex = (ins->rex & REX_REAL)|REX_P; \
1190 out(offset, segment, &ins->rex, OUT_RAWDATA, 1, NO_SEG, NO_SEG); \
1191 ins->rex = 0; \
1192 offset += 1; \
1193 }
1194
1195 static void gencode(int32_t segment, int64_t offset, int bits,
1196 insn * ins, const struct itemplate *temp,
1197 int64_t insn_end)
1198 {
1199 static char condval[] = { /* conditional opcodes */
1200 0x7, 0x3, 0x2, 0x6, 0x2, 0x4, 0xF, 0xD, 0xC, 0xE, 0x6, 0x2,
1201 0x3, 0x7, 0x3, 0x5, 0xE, 0xC, 0xD, 0xF, 0x1, 0xB, 0x9, 0x5,
1202 0x0, 0xA, 0xA, 0xB, 0x8, 0x4
1203 };
1204 uint8_t c;
1205 uint8_t bytes[4];
1206 int64_t size;
1207 int64_t data;
1208 int op1, op2;
1209 struct operand *opx;
1210 const uint8_t *codes = temp->code;
1211 uint8_t opex = 0;
1212
1213 while (*codes) {
1214 c = *codes++;
1215 op1 = (c & 3) + ((opex & 1) << 2);
1216 op2 = ((c >> 3) & 3) + ((opex & 2) << 1);
1217 opx = &ins->oprs[op1];
1218 opex = 0; /* For the next iteration */
1219
1220 switch (c) {
1221 case 01:
1222 case 02:
1223 case 03:
1224 case 04:
1225 EMIT_REX();
1226 out(offset, segment, codes, OUT_RAWDATA, c, NO_SEG, NO_SEG);
1227 codes += c;
1228 offset += c;
1229 break;
1230
1231 case 05:
1232 case 06:
1233 case 07:
1234 opex = c;
1235 break;
1236
1237 case4(010):
1238 EMIT_REX();
1239 bytes[0] = *codes++ + (regval(opx) & 7);
1240 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1241 offset += 1;
1242 break;
1243
1244 case4(014):
1245 /*
1246 * The test for BITS8 and SBYTE here is intended to avoid
1247 * warning on optimizer actions due to SBYTE, while still
1248 * warn on explicit BYTE directives. Also warn, obviously,
1249 * if the optimizer isn't enabled.
1250 */
1251 if (((opx->type & BITS8) ||
1252 !(opx->type & temp->opd[op1] & BYTENESS)) &&
1253 (opx->offset < -128 || opx->offset > 127)) {
1254 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1255 "signed byte value exceeds bounds");
1256 }
1257 if (opx->segment != NO_SEG) {
1258 data = opx->offset;
1259 out(offset, segment, &data, OUT_ADDRESS, 1,
1260 opx->segment, opx->wrt);
1261 } else {
1262 bytes[0] = opx->offset;
1263 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1264 NO_SEG);
1265 }
1266 offset += 1;
1267 break;
1268
1269 case4(020):
1270 if (opx->offset < -256 || opx->offset > 255) {
1271 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1272 "byte value exceeds bounds");
1273 }
1274 if (opx->segment != NO_SEG) {
1275 data = opx->offset;
1276 out(offset, segment, &data, OUT_ADDRESS, 1,
1277 opx->segment, opx->wrt);
1278 } else {
1279 bytes[0] = opx->offset;
1280 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1281 NO_SEG);
1282 }
1283 offset += 1;
1284 break;
1285
1286 case4(024):
1287 if (opx->offset < 0 || opx->offset > 255)
1288 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1289 "unsigned byte value exceeds bounds");
1290 if (opx->segment != NO_SEG) {
1291 data = opx->offset;
1292 out(offset, segment, &data, OUT_ADDRESS, 1,
1293 opx->segment, opx->wrt);
1294 } else {
1295 bytes[0] = opx->offset;
1296 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1297 NO_SEG);
1298 }
1299 offset += 1;
1300 break;
1301
1302 case4(030):
1303 warn_overflow_opd(opx, 2);
1304 data = opx->offset;
1305 out(offset, segment, &data, OUT_ADDRESS, 2,
1306 opx->segment, opx->wrt);
1307 offset += 2;
1308 break;
1309
1310 case4(034):
1311 if (opx->type & (BITS16 | BITS32))
1312 size = (opx->type & BITS16) ? 2 : 4;
1313 else
1314 size = (bits == 16) ? 2 : 4;
1315 warn_overflow_opd(opx, size);
1316 data = opx->offset;
1317 out(offset, segment, &data, OUT_ADDRESS, size,
1318 opx->segment, opx->wrt);
1319 offset += size;
1320 break;
1321
1322 case4(040):
1323 warn_overflow_opd(opx, 4);
1324 data = opx->offset;
1325 out(offset, segment, &data, OUT_ADDRESS, 4,
1326 opx->segment, opx->wrt);
1327 offset += 4;
1328 break;
1329
1330 case4(044):
1331 data = opx->offset;
1332 size = ins->addr_size >> 3;
1333 warn_overflow_opd(opx, size);
1334 out(offset, segment, &data, OUT_ADDRESS, size,
1335 opx->segment, opx->wrt);
1336 offset += size;
1337 break;
1338
1339 case4(050):
1340 if (opx->segment != segment) {
1341 data = opx->offset;
1342 out(offset, segment, &data,
1343 OUT_REL1ADR, insn_end - offset,
1344 opx->segment, opx->wrt);
1345 } else {
1346 data = opx->offset - insn_end;
1347 if (data > 127 || data < -128)
1348 errfunc(ERR_NONFATAL, "short jump is out of range");
1349 out(offset, segment, &data,
1350 OUT_ADDRESS, 1, NO_SEG, NO_SEG);
1351 }
1352 offset += 1;
1353 break;
1354
1355 case4(054):
1356 data = (int64_t)opx->offset;
1357 out(offset, segment, &data, OUT_ADDRESS, 8,
1358 opx->segment, opx->wrt);
1359 offset += 8;
1360 break;
1361
1362 case4(060):
1363 if (opx->segment != segment) {
1364 data = opx->offset;
1365 out(offset, segment, &data,
1366 OUT_REL2ADR, insn_end - offset,
1367 opx->segment, opx->wrt);
1368 } else {
1369 data = opx->offset - insn_end;
1370 out(offset, segment, &data,
1371 OUT_ADDRESS, 2, NO_SEG, NO_SEG);
1372 }
1373 offset += 2;
1374 break;
1375
1376 case4(064):
1377 if (opx->type & (BITS16 | BITS32 | BITS64))
1378 size = (opx->type & BITS16) ? 2 : 4;
1379 else
1380 size = (bits == 16) ? 2 : 4;
1381 if (opx->segment != segment) {
1382 data = opx->offset;
1383 out(offset, segment, &data,
1384 size == 2 ? OUT_REL2ADR : OUT_REL4ADR,
1385 insn_end - offset, opx->segment, opx->wrt);
1386 } else {
1387 data = opx->offset - insn_end;
1388 out(offset, segment, &data,
1389 OUT_ADDRESS, size, NO_SEG, NO_SEG);
1390 }
1391 offset += size;
1392 break;
1393
1394 case4(070):
1395 if (opx->segment != segment) {
1396 data = opx->offset;
1397 out(offset, segment, &data,
1398 OUT_REL4ADR, insn_end - offset,
1399 opx->segment, opx->wrt);
1400 } else {
1401 data = opx->offset - insn_end;
1402 out(offset, segment, &data,
1403 OUT_ADDRESS, 4, NO_SEG, NO_SEG);
1404 }
1405 offset += 4;
1406 break;
1407
1408 case4(074):
1409 if (opx->segment == NO_SEG)
1410 errfunc(ERR_NONFATAL, "value referenced by FAR is not"
1411 " relocatable");
1412 data = 0;
1413 out(offset, segment, &data, OUT_ADDRESS, 2,
1414 outfmt->segbase(1 + opx->segment),
1415 opx->wrt);
1416 offset += 2;
1417 break;
1418
1419 case4(0140):
1420 data = opx->offset;
1421 warn_overflow_opd(opx, 2);
1422 if (is_sbyte16(opx)) {
1423 bytes[0] = data;
1424 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1425 NO_SEG);
1426 offset++;
1427 } else {
1428 out(offset, segment, &data, OUT_ADDRESS, 2,
1429 opx->segment, opx->wrt);
1430 offset += 2;
1431 }
1432 break;
1433
1434 case4(0144):
1435 EMIT_REX();
1436 bytes[0] = *codes++;
1437 if (is_sbyte16(opx))
1438 bytes[0] |= 2; /* s-bit */
1439 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1440 offset++;
1441 break;
1442
1443 case4(0150):
1444 data = opx->offset;
1445 warn_overflow_opd(opx, 4);
1446 if (is_sbyte32(opx)) {
1447 bytes[0] = data;
1448 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1449 NO_SEG);
1450 offset++;
1451 } else {
1452 out(offset, segment, &data, OUT_ADDRESS, 4,
1453 opx->segment, opx->wrt);
1454 offset += 4;
1455 }
1456 break;
1457
1458 case4(0154):
1459 EMIT_REX();
1460 bytes[0] = *codes++;
1461 if (is_sbyte32(opx))
1462 bytes[0] |= 2; /* s-bit */
1463 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1464 offset++;
1465 break;
1466
1467 case4(0160):
1468 case4(0164):
1469 break;
1470
1471 case 0171:
1472 bytes[0] =
1473 (ins->drexdst << 4) |
1474 (ins->rex & REX_OC ? 0x08 : 0) |
1475 (ins->rex & (REX_R|REX_X|REX_B));
1476 ins->rex = 0;
1477 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1478 offset++;
1479 break;
1480
1481 case 0172:
1482 c = *codes++;
1483 opx = &ins->oprs[c >> 3];
1484 bytes[0] = nasm_regvals[opx->basereg] << 4;
1485 opx = &ins->oprs[c & 7];
1486 if (opx->segment != NO_SEG || opx->wrt != NO_SEG) {
1487 errfunc(ERR_NONFATAL,
1488 "non-absolute expression not permitted as argument %d",
1489 c & 7);
1490 } else {
1491 if (opx->offset & ~15) {
1492 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1493 "four-bit argument exceeds bounds");
1494 }
1495 bytes[0] |= opx->offset & 15;
1496 }
1497 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1498 offset++;
1499 break;
1500
1501 case 0173:
1502 c = *codes++;
1503 opx = &ins->oprs[c >> 4];
1504 bytes[0] = nasm_regvals[opx->basereg] << 4;
1505 bytes[0] |= c & 15;
1506 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1507 offset++;
1508 break;
1509
1510 case 0174:
1511 c = *codes++;
1512 opx = &ins->oprs[c];
1513 bytes[0] = nasm_regvals[opx->basereg] << 4;
1514 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1515 offset++;
1516 break;
1517
1518 case4(0250):
1519 data = opx->offset;
1520 if (opx->wrt == NO_SEG && opx->segment == NO_SEG &&
1521 (int32_t)data != (int64_t)data) {
1522 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1523 "signed dword immediate exceeds bounds");
1524 }
1525 if (is_sbyte32(opx)) {
1526 bytes[0] = data;
1527 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1528 NO_SEG);
1529 offset++;
1530 } else {
1531 out(offset, segment, &data, OUT_ADDRESS, 4,
1532 opx->segment, opx->wrt);
1533 offset += 4;
1534 }
1535 break;
1536
1537 case4(0254):
1538 data = opx->offset;
1539 if (opx->wrt == NO_SEG && opx->segment == NO_SEG &&
1540 (int32_t)data != (int64_t)data) {
1541 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1542 "signed dword immediate exceeds bounds");
1543 }
1544 out(offset, segment, &data, OUT_ADDRESS, 4,
1545 opx->segment, opx->wrt);
1546 offset += 4;
1547 break;
1548
1549 case4(0260):
1550 case 0270:
1551 codes += 2;
1552 if (ins->vex_cm != 1 || (ins->rex & (REX_W|REX_X|REX_B))) {
1553 bytes[0] = (ins->vex_cm >> 6) ? 0x8f : 0xc4;
1554 bytes[1] = (ins->vex_cm & 31) | ((~ins->rex & 7) << 5);
1555 bytes[2] = ((ins->rex & REX_W) << (7-3)) |
1556 ((~ins->drexdst & 15)<< 3) | (ins->vex_wlp & 07);
1557 out(offset, segment, &bytes, OUT_RAWDATA, 3, NO_SEG, NO_SEG);
1558 offset += 3;
1559 } else {
1560 bytes[0] = 0xc5;
1561 bytes[1] = ((~ins->rex & REX_R) << (7-2)) |
1562 ((~ins->drexdst & 15) << 3) | (ins->vex_wlp & 07);
1563 out(offset, segment, &bytes, OUT_RAWDATA, 2, NO_SEG, NO_SEG);
1564 offset += 2;
1565 }
1566 break;
1567
1568 case4(0274):
1569 {
1570 uint64_t uv, um;
1571 int s;
1572
1573 if (ins->rex & REX_W)
1574 s = 64;
1575 else if (ins->prefixes[PPS_OSIZE] == P_O16)
1576 s = 16;
1577 else if (ins->prefixes[PPS_OSIZE] == P_O32)
1578 s = 32;
1579 else
1580 s = bits;
1581
1582 um = (uint64_t)2 << (s-1);
1583 uv = opx->offset;
1584
1585 if (uv > 127 && uv < (uint64_t)-128 &&
1586 (uv < um-128 || uv > um-1)) {
1587 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1588 "signed byte value exceeds bounds");
1589 }
1590 if (opx->segment != NO_SEG) {
1591 data = uv;
1592 out(offset, segment, &data, OUT_ADDRESS, 1,
1593 opx->segment, opx->wrt);
1594 } else {
1595 bytes[0] = uv;
1596 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1597 NO_SEG);
1598 }
1599 offset += 1;
1600 break;
1601 }
1602
1603 case4(0300):
1604 break;
1605
1606 case 0310:
1607 if (bits == 32 && !has_prefix(ins, PPS_ASIZE, P_A16)) {
1608 *bytes = 0x67;
1609 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1610 offset += 1;
1611 } else
1612 offset += 0;
1613 break;
1614
1615 case 0311:
1616 if (bits != 32 && !has_prefix(ins, PPS_ASIZE, P_A32)) {
1617 *bytes = 0x67;
1618 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1619 offset += 1;
1620 } else
1621 offset += 0;
1622 break;
1623
1624 case 0312:
1625 break;
1626
1627 case 0313:
1628 ins->rex = 0;
1629 break;
1630
1631 case4(0314):
1632 break;
1633
1634 case 0320:
1635 if (bits != 16) {
1636 *bytes = 0x66;
1637 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1638 offset += 1;
1639 } else
1640 offset += 0;
1641 break;
1642
1643 case 0321:
1644 if (bits == 16) {
1645 *bytes = 0x66;
1646 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1647 offset += 1;
1648 } else
1649 offset += 0;
1650 break;
1651
1652 case 0322:
1653 case 0323:
1654 break;
1655
1656 case 0324:
1657 ins->rex |= REX_W;
1658 break;
1659
1660 case 0325:
1661 break;
1662
1663 case 0330:
1664 *bytes = *codes++ ^ condval[ins->condition];
1665 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1666 offset += 1;
1667 break;
1668
1669 case 0331:
1670 break;
1671
1672 case 0332:
1673 case 0333:
1674 *bytes = c - 0332 + 0xF2;
1675 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1676 offset += 1;
1677 break;
1678
1679 case 0334:
1680 if (ins->rex & REX_R) {
1681 *bytes = 0xF0;
1682 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1683 offset += 1;
1684 }
1685 ins->rex &= ~(REX_L|REX_R);
1686 break;
1687
1688 case 0335:
1689 break;
1690
1691 case 0336:
1692 case 0337:
1693 break;
1694
1695 case 0340:
1696 if (ins->oprs[0].segment != NO_SEG)
1697 errfunc(ERR_PANIC, "non-constant BSS size in pass two");
1698 else {
1699 int64_t size = ins->oprs[0].offset;
1700 if (size > 0)
1701 out(offset, segment, NULL,
1702 OUT_RESERVE, size, NO_SEG, NO_SEG);
1703 offset += size;
1704 }
1705 break;
1706
1707 case 0341:
1708 break;
1709
1710 case 0344:
1711 case 0345:
1712 bytes[0] = c & 1;
1713 switch (ins->oprs[0].basereg) {
1714 case R_CS:
1715 bytes[0] += 0x0E;
1716 break;
1717 case R_DS:
1718 bytes[0] += 0x1E;
1719 break;
1720 case R_ES:
1721 bytes[0] += 0x06;
1722 break;
1723 case R_SS:
1724 bytes[0] += 0x16;
1725 break;
1726 default:
1727 errfunc(ERR_PANIC,
1728 "bizarre 8086 segment register received");
1729 }
1730 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1731 offset++;
1732 break;
1733
1734 case 0346:
1735 case 0347:
1736 bytes[0] = c & 1;
1737 switch (ins->oprs[0].basereg) {
1738 case R_FS:
1739 bytes[0] += 0xA0;
1740 break;
1741 case R_GS:
1742 bytes[0] += 0xA8;
1743 break;
1744 default:
1745 errfunc(ERR_PANIC,
1746 "bizarre 386 segment register received");
1747 }
1748 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1749 offset++;
1750 break;
1751
1752 case 0360:
1753 break;
1754
1755 case 0361:
1756 bytes[0] = 0x66;
1757 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1758 offset += 1;
1759 break;
1760
1761 case 0362:
1762 case 0363:
1763 bytes[0] = c - 0362 + 0xf2;
1764 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1765 offset += 1;
1766 break;
1767
1768 case 0364:
1769 case 0365:
1770 break;
1771
1772 case 0366:
1773 case 0367:
1774 *bytes = c - 0366 + 0x66;
1775 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1776 offset += 1;
1777 break;
1778
1779 case 0370:
1780 case 0371:
1781 case 0372:
1782 break;
1783
1784 case 0373:
1785 *bytes = bits == 16 ? 3 : 5;
1786 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1787 offset += 1;
1788 break;
1789
1790 case4(0100):
1791 case4(0110):
1792 case4(0120):
1793 case4(0130):
1794 case4(0200):
1795 case4(0204):
1796 case4(0210):
1797 case4(0214):
1798 case4(0220):
1799 case4(0224):
1800 case4(0230):
1801 case4(0234):
1802 {
1803 ea ea_data;
1804 int rfield;
1805 opflags_t rflags;
1806 uint8_t *p;
1807 int32_t s;
1808 enum out_type type;
1809 struct operand *opy = &ins->oprs[op2];
1810
1811 if (c <= 0177) {
1812 /* pick rfield from operand b (opx) */
1813 rflags = regflag(opx);
1814 rfield = nasm_regvals[opx->basereg];
1815 } else {
1816 /* rfield is constant */
1817 rflags = 0;
1818 rfield = c & 7;
1819 }
1820
1821 if (!process_ea(opy, &ea_data, bits, ins->addr_size,
1822 rfield, rflags)) {
1823 errfunc(ERR_NONFATAL, "invalid effective address");
1824 }
1825
1826
1827 p = bytes;
1828 *p++ = ea_data.modrm;
1829 if (ea_data.sib_present)
1830 *p++ = ea_data.sib;
1831
1832 /* DREX suffixes come between the SIB and the displacement */
1833 if (ins->rex & REX_D) {
1834 *p++ = (ins->drexdst << 4) |
1835 (ins->rex & REX_OC ? 0x08 : 0) |
1836 (ins->rex & (REX_R|REX_X|REX_B));
1837 ins->rex = 0;
1838 }
1839
1840 s = p - bytes;
1841 out(offset, segment, bytes, OUT_RAWDATA, s, NO_SEG, NO_SEG);
1842
1843 /*
1844 * Make sure the address gets the right offset in case
1845 * the line breaks in the .lst file (BR 1197827)
1846 */
1847 offset += s;
1848 s = 0;
1849
1850 switch (ea_data.bytes) {
1851 case 0:
1852 break;
1853 case 1:
1854 case 2:
1855 case 4:
1856 case 8:
1857 data = opy->offset;
1858 s += ea_data.bytes;
1859 if (ea_data.rip) {
1860 if (opy->segment == segment) {
1861 data -= insn_end;
1862 if (overflow_signed(data, ea_data.bytes))
1863 warn_overflow(ERR_PASS2, ea_data.bytes);
1864 out(offset, segment, &data, OUT_ADDRESS,
1865 ea_data.bytes, NO_SEG, NO_SEG);
1866 } else {
1867 /* overflow check in output/linker? */
1868 out(offset, segment, &data, OUT_REL4ADR,
1869 insn_end - offset, opy->segment, opy->wrt);
1870 }
1871 } else {
1872 if (overflow_general(opy->offset, ins->addr_size >> 3) ||
1873 signed_bits(opy->offset, ins->addr_size) !=
1874 signed_bits(opy->offset, ea_data.bytes * 8))
1875 warn_overflow(ERR_PASS2, ea_data.bytes);
1876
1877 type = OUT_ADDRESS;
1878 out(offset, segment, &data, OUT_ADDRESS,
1879 ea_data.bytes, opy->segment, opy->wrt);
1880 }
1881 break;
1882 default:
1883 /* Impossible! */
1884 errfunc(ERR_PANIC,
1885 "Invalid amount of bytes (%d) for offset?!",
1886 ea_data.bytes);
1887 break;
1888 }
1889 offset += s;
1890 }
1891 break;
1892
1893 default:
1894 errfunc(ERR_PANIC, "internal instruction table corrupt"
1895 ": instruction code \\%o (0x%02X) given", c, c);
1896 break;
1897 }
1898 }
1899 }
1900
1901 static opflags_t regflag(const operand * o)
1902 {
1903 if (!is_register(o->basereg))
1904 errfunc(ERR_PANIC, "invalid operand passed to regflag()");
1905 return nasm_reg_flags[o->basereg];
1906 }
1907
1908 static int32_t regval(const operand * o)
1909 {
1910 if (!is_register(o->basereg))
1911 errfunc(ERR_PANIC, "invalid operand passed to regval()");
1912 return nasm_regvals[o->basereg];
1913 }
1914
1915 static int op_rexflags(const operand * o, int mask)
1916 {
1917 opflags_t flags;
1918 int val;
1919
1920 if (!is_register(o->basereg))
1921 errfunc(ERR_PANIC, "invalid operand passed to op_rexflags()");
1922
1923 flags = nasm_reg_flags[o->basereg];
1924 val = nasm_regvals[o->basereg];
1925
1926 return rexflags(val, flags, mask);
1927 }
1928
1929 static int rexflags(int val, opflags_t flags, int mask)
1930 {
1931 int rex = 0;
1932
1933 if (val >= 8)
1934 rex |= REX_B|REX_X|REX_R;
1935 if (flags & BITS64)
1936 rex |= REX_W;
1937 if (!(REG_HIGH & ~flags)) /* AH, CH, DH, BH */
1938 rex |= REX_H;
1939 else if (!(REG8 & ~flags) && val >= 4) /* SPL, BPL, SIL, DIL */
1940 rex |= REX_P;
1941
1942 return rex & mask;
1943 }
1944
1945 static enum match_result find_match(const struct itemplate **tempp,
1946 insn *instruction,
1947 int32_t segment, int64_t offset, int bits)
1948 {
1949 const struct itemplate *temp;
1950 enum match_result m, merr;
1951 opflags_t xsizeflags[MAX_OPERANDS];
1952 bool opsizemissing = false;
1953 int i;
1954
1955 for (i = 0; i < instruction->operands; i++)
1956 xsizeflags[i] = instruction->oprs[i].type & SIZE_MASK;
1957
1958 merr = MERR_INVALOP;
1959
1960 for (temp = nasm_instructions[instruction->opcode];
1961 temp->opcode != I_none; temp++) {
1962 m = matches(temp, instruction, bits);
1963 if (m == MOK_JUMP) {
1964 if (jmp_match(segment, offset, bits, instruction, temp->code))
1965 m = MOK_GOOD;
1966 else
1967 m = MERR_INVALOP;
1968 } else if (m == MERR_OPSIZEMISSING &&
1969 (temp->flags & IF_SMASK) != IF_SX) {
1970 /*
1971 * Missing operand size and a candidate for fuzzy matching...
1972 */
1973 for (i = 0; i < temp->operands; i++) {
1974 if ((temp->opd[i] & SAME_AS) == 0)
1975 xsizeflags[i] |= temp->opd[i] & SIZE_MASK;
1976 }
1977 opsizemissing = true;
1978 }
1979 if (m > merr)
1980 merr = m;
1981 if (merr == MOK_GOOD)
1982 goto done;
1983 }
1984
1985 /* No match, but see if we can get a fuzzy operand size match... */
1986 if (!opsizemissing)
1987 goto done;
1988
1989 for (i = 0; i < instruction->operands; i++) {
1990 /*
1991 * We ignore extrinsic operand sizes on registers, so we should
1992 * never try to fuzzy-match on them. This also resolves the case
1993 * when we have e.g. "xmmrm128" in two different positions.
1994 */
1995 if (is_class(REGISTER, instruction->oprs[i].type))
1996 continue;
1997
1998 /* This tests if xsizeflags[i] has more than one bit set */
1999 if ((xsizeflags[i] & (xsizeflags[i]-1)))
2000 goto done; /* No luck */
2001
2002 instruction->oprs[i].type |= xsizeflags[i]; /* Set the size */
2003 }
2004
2005 /* Try matching again... */
2006 for (temp = nasm_instructions[instruction->opcode];
2007 temp->opcode != I_none; temp++) {
2008 m = matches(temp, instruction, bits);
2009 if (m == MOK_JUMP) {
2010 if (jmp_match(segment, offset, bits, instruction, temp->code))
2011 m = MOK_GOOD;
2012 else
2013 m = MERR_INVALOP;
2014 }
2015 if (m > merr)
2016 merr = m;
2017 if (merr == MOK_GOOD)
2018 goto done;
2019 }
2020
2021 done:
2022 *tempp = temp;
2023 return merr;
2024 }
2025
2026 static enum match_result matches(const struct itemplate *itemp,
2027 insn *instruction, int bits)
2028 {
2029 int i, size[MAX_OPERANDS], asize, oprs;
2030 bool opsizemissing = false;
2031
2032 /*
2033 * Check the opcode
2034 */
2035 if (itemp->opcode != instruction->opcode)
2036 return MERR_INVALOP;
2037
2038 /*
2039 * Count the operands
2040 */
2041 if (itemp->operands != instruction->operands)
2042 return MERR_INVALOP;
2043
2044 /*
2045 * Check that no spurious colons or TOs are present
2046 */
2047 for (i = 0; i < itemp->operands; i++)
2048 if (instruction->oprs[i].type & ~itemp->opd[i] & (COLON | TO))
2049 return MERR_INVALOP;
2050
2051 /*
2052 * Process size flags
2053 */
2054 switch (itemp->flags & IF_SMASK) {
2055 case IF_SB:
2056 asize = BITS8;
2057 break;
2058 case IF_SW:
2059 asize = BITS16;
2060 break;
2061 case IF_SD:
2062 asize = BITS32;
2063 break;
2064 case IF_SQ:
2065 asize = BITS64;
2066 break;
2067 case IF_SO:
2068 asize = BITS128;
2069 break;
2070 case IF_SY:
2071 asize = BITS256;
2072 break;
2073 case IF_SZ:
2074 switch (bits) {
2075 case 16:
2076 asize = BITS16;
2077 break;
2078 case 32:
2079 asize = BITS32;
2080 break;
2081 case 64:
2082 asize = BITS64;
2083 break;
2084 default:
2085 asize = 0;
2086 break;
2087 }
2088 break;
2089 default:
2090 asize = 0;
2091 break;
2092 }
2093
2094 if (itemp->flags & IF_ARMASK) {
2095 /* S- flags only apply to a specific operand */
2096 i = ((itemp->flags & IF_ARMASK) >> IF_ARSHFT) - 1;
2097 memset(size, 0, sizeof size);
2098 size[i] = asize;
2099 } else {
2100 /* S- flags apply to all operands */
2101 for (i = 0; i < MAX_OPERANDS; i++)
2102 size[i] = asize;
2103 }
2104
2105 /*
2106 * Check that the operand flags all match up,
2107 * it's a bit tricky so lets be verbose:
2108 *
2109 * 1) Find out the size of operand. If instruction
2110 * doesn't have one specified -- we're trying to
2111 * guess it either from template (IF_S* flag) or
2112 * from code bits.
2113 *
2114 * 2) If template operand (i) has SAME_AS flag [used for registers only]
2115 * (ie the same operand as was specified somewhere in template, and
2116 * this referred operand index is being achieved via ~SAME_AS)
2117 * we are to be sure that both registers (in template and instruction)
2118 * do exactly match.
2119 *
2120 * 3) If template operand do not match the instruction OR
2121 * template has an operand size specified AND this size differ
2122 * from which instruction has (perhaps we got it from code bits)
2123 * we are:
2124 * a) Check that only size of instruction and operand is differ
2125 * other characteristics do match
2126 * b) Perhaps it's a register specified in instruction so
2127 * for such a case we just mark that operand as "size
2128 * missing" and this will turn on fuzzy operand size
2129 * logic facility (handled by a caller)
2130 */
2131 for (i = 0; i < itemp->operands; i++) {
2132 opflags_t type = instruction->oprs[i].type;
2133 if (!(type & SIZE_MASK))
2134 type |= size[i];
2135
2136 if (itemp->opd[i] & SAME_AS) {
2137 int j = itemp->opd[i] & ~SAME_AS;
2138 if (type != instruction->oprs[j].type ||
2139 instruction->oprs[i].basereg != instruction->oprs[j].basereg)
2140 return MERR_INVALOP;
2141 } else if (itemp->opd[i] & ~type ||
2142 ((itemp->opd[i] & SIZE_MASK) &&
2143 ((itemp->opd[i] ^ type) & SIZE_MASK))) {
2144 if ((itemp->opd[i] & ~type & ~SIZE_MASK) || (type & SIZE_MASK)) {
2145 return MERR_INVALOP;
2146 } else if (!is_class(REGISTER, type)) {
2147 /*
2148 * Note: we don't honor extrinsic operand sizes for registers,
2149 * so "missing operand size" for a register should be
2150 * considered a wildcard match rather than an error.
2151 */
2152 opsizemissing = true;
2153 }
2154 }
2155 }
2156
2157 if (opsizemissing)
2158 return MERR_OPSIZEMISSING;
2159
2160 /*
2161 * Check operand sizes
2162 */
2163 if (itemp->flags & (IF_SM | IF_SM2)) {
2164 oprs = (itemp->flags & IF_SM2 ? 2 : itemp->operands);
2165 for (i = 0; i < oprs; i++) {
2166 asize = itemp->opd[i] & SIZE_MASK;
2167 if (asize) {
2168 for (i = 0; i < oprs; i++)
2169 size[i] = asize;
2170 break;
2171 }
2172 }
2173 } else {
2174 oprs = itemp->operands;
2175 }
2176
2177 for (i = 0; i < itemp->operands; i++) {
2178 if (!(itemp->opd[i] & SIZE_MASK) &&
2179 (instruction->oprs[i].type & SIZE_MASK & ~size[i]))
2180 return MERR_OPSIZEMISMATCH;
2181 }
2182
2183 /*
2184 * Check template is okay at the set cpu level
2185 */
2186 if (((itemp->flags & IF_PLEVEL) > cpu))
2187 return MERR_BADCPU;
2188
2189 /*
2190 * Verify the appropriate long mode flag.
2191 */
2192 if ((itemp->flags & (bits == 64 ? IF_NOLONG : IF_LONG)))
2193 return MERR_BADMODE;
2194
2195 /*
2196 * Check if special handling needed for Jumps
2197 */
2198 if ((itemp->code[0] & 0374) == 0370)
2199 return MOK_JUMP;
2200
2201 return MOK_GOOD;
2202 }
2203
2204 static ea *process_ea(operand * input, ea * output, int bits,
2205 int addrbits, int rfield, opflags_t rflags)
2206 {
2207 bool forw_ref = !!(input->opflags & OPFLAG_UNKNOWN);
2208
2209 output->rip = false;
2210
2211 /* REX flags for the rfield operand */
2212 output->rex |= rexflags(rfield, rflags, REX_R | REX_P | REX_W | REX_H);
2213
2214 if (is_class(REGISTER, input->type)) { /* register direct */
2215 int i;
2216 opflags_t f;
2217
2218 if (!is_register(input->basereg))
2219 return NULL;
2220 f = regflag(input);
2221 i = nasm_regvals[input->basereg];
2222
2223 if (REG_EA & ~f)
2224 return NULL; /* Invalid EA register */
2225
2226 output->rex |= op_rexflags(input, REX_B | REX_P | REX_W | REX_H);
2227
2228 output->sib_present = false; /* no SIB necessary */
2229 output->bytes = 0; /* no offset necessary either */
2230 output->modrm = 0xC0 | ((rfield & 7) << 3) | (i & 7);
2231 } else { /* it's a memory reference */
2232 if (input->basereg == -1 &&
2233 (input->indexreg == -1 || input->scale == 0)) {
2234 /* it's a pure offset */
2235
2236 if (bits == 64 && ((input->type & IP_REL) == IP_REL) &&
2237 input->segment == NO_SEG) {
2238 nasm_error(ERR_WARNING | ERR_PASS1, "absolute address can not be RIP-relative");
2239 input->type &= ~IP_REL;
2240 input->type |= MEMORY;
2241 }
2242
2243 if (input->eaflags & EAF_BYTEOFFS ||
2244 (input->eaflags & EAF_WORDOFFS &&
2245 input->disp_size != (addrbits != 16 ? 32 : 16))) {
2246 nasm_error(ERR_WARNING | ERR_PASS1, "displacement size ignored on absolute address");
2247 }
2248
2249 if (bits == 64 && (~input->type & IP_REL)) {
2250 int scale, index, base;
2251 output->sib_present = true;
2252 scale = 0;
2253 index = 4;
2254 base = 5;
2255 output->sib = (scale << 6) | (index << 3) | base;
2256 output->bytes = 4;
2257 output->modrm = 4 | ((rfield & 7) << 3);
2258 output->rip = false;
2259 } else {
2260 output->sib_present = false;
2261 output->bytes = (addrbits != 16 ? 4 : 2);
2262 output->modrm = (addrbits != 16 ? 5 : 6) | ((rfield & 7) << 3);
2263 output->rip = bits == 64;
2264 }
2265 } else { /* it's an indirection */
2266 int i = input->indexreg, b = input->basereg, s = input->scale;
2267 int32_t seg = input->segment;
2268 int hb = input->hintbase, ht = input->hinttype;
2269 int t, it, bt; /* register numbers */
2270 opflags_t x, ix, bx; /* register flags */
2271
2272 if (s == 0)
2273 i = -1; /* make this easy, at least */
2274
2275 if (is_register(i)) {
2276 it = nasm_regvals[i];
2277 ix = nasm_reg_flags[i];
2278 } else {
2279 it = -1;
2280 ix = 0;
2281 }
2282
2283 if (is_register(b)) {
2284 bt = nasm_regvals[b];
2285 bx = nasm_reg_flags[b];
2286 } else {
2287 bt = -1;
2288 bx = 0;
2289 }
2290
2291 /* check for a 32/64-bit memory reference... */
2292 if ((ix|bx) & (BITS32|BITS64)) {
2293 /*
2294 * it must be a 32/64-bit memory reference. Firstly we have
2295 * to check that all registers involved are type E/Rxx.
2296 */
2297 int32_t sok = BITS32 | BITS64, o = input->offset;
2298
2299 if (it != -1) {
2300 if (!(REG64 & ~ix) || !(REG32 & ~ix))
2301 sok &= ix;
2302 else
2303 return NULL;
2304 }
2305
2306 if (bt != -1) {
2307 if (REG_GPR & ~bx)
2308 return NULL; /* Invalid register */
2309 if (~sok & bx & SIZE_MASK)
2310 return NULL; /* Invalid size */
2311 sok &= bx;
2312 }
2313
2314 /*
2315 * While we're here, ensure the user didn't specify
2316 * WORD or QWORD
2317 */
2318 if (input->disp_size == 16 || input->disp_size == 64)
2319 return NULL;
2320
2321 if (addrbits == 16 ||
2322 (addrbits == 32 && !(sok & BITS32)) ||
2323 (addrbits == 64 && !(sok & BITS64)))
2324 return NULL;
2325
2326 /* now reorganize base/index */
2327 if (s == 1 && bt != it && bt != -1 && it != -1 &&
2328 ((hb == b && ht == EAH_NOTBASE) ||
2329 (hb == i && ht == EAH_MAKEBASE))) {
2330 /* swap if hints say so */
2331 t = bt, bt = it, it = t;
2332 x = bx, bx = ix, ix = x;
2333 }
2334 if (bt == it) /* convert EAX+2*EAX to 3*EAX */
2335 bt = -1, bx = 0, s++;
2336 if (bt == -1 && s == 1 && !(hb == it && ht == EAH_NOTBASE)) {
2337 /* make single reg base, unless hint */
2338 bt = it, bx = ix, it = -1, ix = 0;
2339 }
2340 if (((s == 2 && it != REG_NUM_ESP && !(input->eaflags & EAF_TIMESTWO)) ||
2341 s == 3 || s == 5 || s == 9) && bt == -1)
2342 bt = it, bx = ix, s--; /* convert 3*EAX to EAX+2*EAX */
2343 if (it == -1 && (bt & 7) != REG_NUM_ESP &&
2344 (input->eaflags & EAF_TIMESTWO))
2345 it = bt, ix = bx, bt = -1, bx = 0, s = 1;
2346 /* convert [NOSPLIT EAX] to sib format with 0x0 displacement */
2347 if (s == 1 && it == REG_NUM_ESP) {
2348 /* swap ESP into base if scale is 1 */
2349 t = it, it = bt, bt = t;
2350 x = ix, ix = bx, bx = x;
2351 }
2352 if (it == REG_NUM_ESP ||
2353 (s != 1 && s != 2 && s != 4 && s != 8 && it != -1))
2354 return NULL; /* wrong, for various reasons */
2355
2356 output->rex |= rexflags(it, ix, REX_X);
2357 output->rex |= rexflags(bt, bx, REX_B);
2358
2359 if (it == -1 && (bt & 7) != REG_NUM_ESP) {
2360 /* no SIB needed */
2361 int mod, rm;
2362
2363 if (bt == -1) {
2364 rm = 5;
2365 mod = 0;
2366 } else {
2367 rm = (bt & 7);
2368 if (rm != REG_NUM_EBP && o == 0 &&
2369 seg == NO_SEG && !forw_ref &&
2370 !(input->eaflags & (EAF_BYTEOFFS | EAF_WORDOFFS)))
2371 mod = 0;
2372 else if (input->eaflags & EAF_BYTEOFFS ||
2373 (o >= -128 && o <= 127 &&
2374 seg == NO_SEG && !forw_ref &&
2375 !(input->eaflags & EAF_WORDOFFS)))
2376 mod = 1;
2377 else
2378 mod = 2;
2379 }
2380
2381 output->sib_present = false;
2382 output->bytes = (bt == -1 || mod == 2 ? 4 : mod);
2383 output->modrm = (mod << 6) | ((rfield & 7) << 3) | rm;
2384 } else {
2385 /* we need a SIB */
2386 int mod, scale, index, base;
2387
2388 if (it == -1)
2389 index = 4, s = 1;
2390 else
2391 index = (it & 7);
2392
2393 switch (s) {
2394 case 1:
2395 scale = 0;
2396 break;
2397 case 2:
2398 scale = 1;
2399 break;
2400 case 4:
2401 scale = 2;
2402 break;
2403 case 8:
2404 scale = 3;
2405 break;
2406 default: /* then what the smeg is it? */
2407 return NULL; /* panic */
2408 }
2409
2410 if (bt == -1) {
2411 base = 5;
2412 mod = 0;
2413 } else {
2414 base = (bt & 7);
2415 if (base != REG_NUM_EBP && o == 0 &&
2416 seg == NO_SEG && !forw_ref &&
2417 !(input->eaflags & (EAF_BYTEOFFS | EAF_WORDOFFS)))
2418 mod = 0;
2419 else if (input->eaflags & EAF_BYTEOFFS ||
2420 (o >= -128 && o <= 127 &&
2421 seg == NO_SEG && !forw_ref &&
2422 !(input->eaflags & EAF_WORDOFFS)))
2423 mod = 1;
2424 else
2425 mod = 2;
2426 }
2427
2428 output->sib_present = true;
2429 output->bytes = (bt == -1 || mod == 2 ? 4 : mod);
2430 output->modrm = (mod << 6) | ((rfield & 7) << 3) | 4;
2431 output->sib = (scale << 6) | (index << 3) | base;
2432 }
2433 } else { /* it's 16-bit */
2434 int mod, rm;
2435 int16_t o = input->offset;
2436
2437 /* check for 64-bit long mode */
2438 if (addrbits == 64)
2439 return NULL;
2440
2441 /* check all registers are BX, BP, SI or DI */
2442 if ((b != -1 && b != R_BP && b != R_BX && b != R_SI && b != R_DI) ||
2443 (i != -1 && i != R_BP && i != R_BX && i != R_SI && i != R_DI))
2444 return NULL;
2445
2446 /* ensure the user didn't specify DWORD/QWORD */
2447 if (input->disp_size == 32 || input->disp_size == 64)
2448 return NULL;
2449
2450 if (s != 1 && i != -1)
2451 return NULL; /* no can do, in 16-bit EA */
2452 if (b == -1 && i != -1) {
2453 int tmp = b;
2454 b = i;
2455 i = tmp;
2456 } /* swap */
2457 if ((b == R_SI || b == R_DI) && i != -1) {
2458 int tmp = b;
2459 b = i;
2460 i = tmp;
2461 }
2462 /* have BX/BP as base, SI/DI index */
2463 if (b == i)
2464 return NULL; /* shouldn't ever happen, in theory */
2465 if (i != -1 && b != -1 &&
2466 (i == R_BP || i == R_BX || b == R_SI || b == R_DI))
2467 return NULL; /* invalid combinations */
2468 if (b == -1) /* pure offset: handled above */
2469 return NULL; /* so if it gets to here, panic! */
2470
2471 rm = -1;
2472 if (i != -1)
2473 switch (i * 256 + b) {
2474 case R_SI * 256 + R_BX:
2475 rm = 0;
2476 break;
2477 case R_DI * 256 + R_BX:
2478 rm = 1;
2479 break;
2480 case R_SI * 256 + R_BP:
2481 rm = 2;
2482 break;
2483 case R_DI * 256 + R_BP:
2484 rm = 3;
2485 break;
2486 } else
2487 switch (b) {
2488 case R_SI:
2489 rm = 4;
2490 break;
2491 case R_DI:
2492 rm = 5;
2493 break;
2494 case R_BP:
2495 rm = 6;
2496 break;
2497 case R_BX:
2498 rm = 7;
2499 break;
2500 }
2501 if (rm == -1) /* can't happen, in theory */
2502 return NULL; /* so panic if it does */
2503
2504 if (o == 0 && seg == NO_SEG && !forw_ref && rm != 6 &&
2505 !(input->eaflags & (EAF_BYTEOFFS | EAF_WORDOFFS)))
2506 mod = 0;
2507 else if (input->eaflags & EAF_BYTEOFFS ||
2508 (o >= -128 && o <= 127 && seg == NO_SEG &&
2509 !forw_ref && !(input->eaflags & EAF_WORDOFFS)))
2510 mod = 1;
2511 else
2512 mod = 2;
2513
2514 output->sib_present = false; /* no SIB - it's 16-bit */
2515 output->bytes = mod; /* bytes of offset needed */
2516 output->modrm = (mod << 6) | ((rfield & 7) << 3) | rm;
2517 }
2518 }
2519 }
2520
2521 output->size = 1 + output->sib_present + output->bytes;
2522 return output;
2523 }
2524
2525 static void add_asp(insn *ins, int addrbits)
2526 {
2527 int j, valid;
2528 int defdisp;
2529
2530 valid = (addrbits == 64) ? 64|32 : 32|16;
2531
2532 switch (ins->prefixes[PPS_ASIZE]) {
2533 case P_A16:
2534 valid &= 16;
2535 break;
2536 case P_A32:
2537 valid &= 32;
2538 break;
2539 case P_A64:
2540 valid &= 64;
2541 break;
2542 case P_ASP:
2543 valid &= (addrbits == 32) ? 16 : 32;
2544 break;
2545 default:
2546 break;
2547 }
2548
2549 for (j = 0; j < ins->operands; j++) {
2550 if (is_class(MEMORY, ins->oprs[j].type)) {
2551 opflags_t i, b;
2552
2553 /* Verify as Register */
2554 if (!is_register(ins->oprs[j].indexreg))
2555 i = 0;
2556 else
2557 i = nasm_reg_flags[ins->oprs[j].indexreg];
2558
2559 /* Verify as Register */
2560 if (!is_register(ins->oprs[j].basereg))
2561 b = 0;
2562 else
2563 b = nasm_reg_flags[ins->oprs[j].basereg];
2564
2565 if (ins->oprs[j].scale == 0)
2566 i = 0;
2567
2568 if (!i && !b) {
2569 int ds = ins->oprs[j].disp_size;
2570 if ((addrbits != 64 && ds > 8) ||
2571 (addrbits == 64 && ds == 16))
2572 valid &= ds;
2573 } else {
2574 if (!(REG16 & ~b))
2575 valid &= 16;
2576 if (!(REG32 & ~b))
2577 valid &= 32;
2578 if (!(REG64 & ~b))
2579 valid &= 64;
2580
2581 if (!(REG16 & ~i))
2582 valid &= 16;
2583 if (!(REG32 & ~i))
2584 valid &= 32;
2585 if (!(REG64 & ~i))
2586 valid &= 64;
2587 }
2588 }
2589 }
2590
2591 if (valid & addrbits) {
2592 ins->addr_size = addrbits;
2593 } else if (valid & ((addrbits == 32) ? 16 : 32)) {
2594 /* Add an address size prefix */
2595 enum prefixes pref = (addrbits == 32) ? P_A16 : P_A32;
2596 ins->prefixes[PPS_ASIZE] = pref;
2597 ins->addr_size = (addrbits == 32) ? 16 : 32;
2598 } else {
2599 /* Impossible... */
2600 errfunc(ERR_NONFATAL, "impossible combination of address sizes");
2601 ins->addr_size = addrbits; /* Error recovery */
2602 }
2603
2604 defdisp = ins->addr_size == 16 ? 16 : 32;
2605
2606 for (j = 0; j < ins->operands; j++) {
2607 if (!(MEM_OFFS & ~ins->oprs[j].type) &&
2608 (ins->oprs[j].disp_size ? ins->oprs[j].disp_size : defdisp) != ins->addr_size) {
2609 /*
2610 * mem_offs sizes must match the address size; if not,
2611 * strip the MEM_OFFS bit and match only EA instructions
2612 */
2613 ins->oprs[j].type &= ~(MEM_OFFS & ~MEMORY);
2614 }
2615 }
2616 }
Adding AVX-512 features