?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
NASM 2.08rc1
[nasm/avx512.git] / assemble.c
blob7ea5fb9c457adf93bdce422ccb730595f5dfcdc9
1 /* ----------------------------------------------------------------------- *
2 *
3 * Copyright 1996-2009 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 int32_t regflag(const operand *);
203 static int32_t regval(const operand *);
204 static int rexflags(int, int32_t, int);
205 static int op_rexflags(const operand *, int);
206 static ea *process_ea(operand *, ea *, int, int, int, int32_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 (size < 8 && 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
308 outfmt->output(segto, data, type, size, segment, wrt);
309 }
310
311 static bool jmp_match(int32_t segment, int64_t offset, int bits,
312 insn * ins, const uint8_t *code)
313 {
314 int64_t isize;
315 uint8_t c = code[0];
316
317 if ((c != 0370 && c != 0371) || (ins->oprs[0].type & STRICT))
318 return false;
319 if (!optimizing)
320 return false;
321 if (optimizing < 0 && c == 0371)
322 return false;
323
324 isize = calcsize(segment, offset, bits, ins, code);
325
326 if (ins->oprs[0].opflags & OPFLAG_UNKNOWN)
327 /* Be optimistic in pass 1 */
328 return true;
329
330 if (ins->oprs[0].segment != segment)
331 return false;
332
333 isize = ins->oprs[0].offset - offset - isize; /* isize is delta */
334 return (isize >= -128 && isize <= 127); /* is it byte size? */
335 }
336
337 int64_t assemble(int32_t segment, int64_t offset, int bits, uint32_t cp,
338 insn * instruction, struct ofmt *output, efunc error,
339 ListGen * listgen)
340 {
341 const struct itemplate *temp;
342 int j;
343 enum match_result m;
344 int64_t insn_end;
345 int32_t itimes;
346 int64_t start = offset;
347 int64_t wsize = 0; /* size for DB etc. */
348
349 errfunc = error; /* to pass to other functions */
350 cpu = cp;
351 outfmt = output; /* likewise */
352 list = listgen; /* and again */
353
354 switch (instruction->opcode) {
355 case -1:
356 return 0;
357 case I_DB:
358 wsize = 1;
359 break;
360 case I_DW:
361 wsize = 2;
362 break;
363 case I_DD:
364 wsize = 4;
365 break;
366 case I_DQ:
367 wsize = 8;
368 break;
369 case I_DT:
370 wsize = 10;
371 break;
372 case I_DO:
373 wsize = 16;
374 break;
375 case I_DY:
376 wsize = 32;
377 break;
378 default:
379 break;
380 }
381
382 if (wsize) {
383 extop *e;
384 int32_t t = instruction->times;
385 if (t < 0)
386 errfunc(ERR_PANIC,
387 "instruction->times < 0 (%ld) in assemble()", t);
388
389 while (t--) { /* repeat TIMES times */
390 list_for_each(e, instruction->eops) {
391 if (e->type == EOT_DB_NUMBER) {
392 if (wsize == 1) {
393 if (e->segment != NO_SEG)
394 errfunc(ERR_NONFATAL,
395 "one-byte relocation attempted");
396 else {
397 uint8_t out_byte = e->offset;
398 out(offset, segment, &out_byte,
399 OUT_RAWDATA, 1, NO_SEG, NO_SEG);
400 }
401 } else if (wsize > 8) {
402 errfunc(ERR_NONFATAL,
403 "integer supplied to a DT, DO or DY"
404 " instruction");
405 } else
406 out(offset, segment, &e->offset,
407 OUT_ADDRESS, wsize, e->segment, e->wrt);
408 offset += wsize;
409 } else if (e->type == EOT_DB_STRING ||
410 e->type == EOT_DB_STRING_FREE) {
411 int align;
412
413 out(offset, segment, e->stringval,
414 OUT_RAWDATA, e->stringlen, NO_SEG, NO_SEG);
415 align = e->stringlen % wsize;
416
417 if (align) {
418 align = wsize - align;
419 out(offset, segment, zero_buffer,
420 OUT_RAWDATA, align, NO_SEG, NO_SEG);
421 }
422 offset += e->stringlen + align;
423 }
424 }
425 if (t > 0 && t == instruction->times - 1) {
426 /*
427 * Dummy call to list->output to give the offset to the
428 * listing module.
429 */
430 list->output(offset, NULL, OUT_RAWDATA, 0);
431 list->uplevel(LIST_TIMES);
432 }
433 }
434 if (instruction->times > 1)
435 list->downlevel(LIST_TIMES);
436 return offset - start;
437 }
438
439 if (instruction->opcode == I_INCBIN) {
440 const char *fname = instruction->eops->stringval;
441 FILE *fp;
442
443 fp = fopen(fname, "rb");
444 if (!fp) {
445 error(ERR_NONFATAL, "`incbin': unable to open file `%s'",
446 fname);
447 } else if (fseek(fp, 0L, SEEK_END) < 0) {
448 error(ERR_NONFATAL, "`incbin': unable to seek on file `%s'",
449 fname);
450 } else {
451 static char buf[4096];
452 size_t t = instruction->times;
453 size_t base = 0;
454 size_t len;
455
456 len = ftell(fp);
457 if (instruction->eops->next) {
458 base = instruction->eops->next->offset;
459 len -= base;
460 if (instruction->eops->next->next &&
461 len > (size_t)instruction->eops->next->next->offset)
462 len = (size_t)instruction->eops->next->next->offset;
463 }
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_INCBIN);
470 while (t--) {
471 size_t l;
472
473 fseek(fp, base, SEEK_SET);
474 l = len;
475 while (l > 0) {
476 int32_t m;
477 m = fread(buf, 1, l > sizeof(buf) ? sizeof(buf) : l, fp);
478 if (!m) {
479 /*
480 * This shouldn't happen unless the file
481 * actually changes while we are reading
482 * it.
483 */
484 error(ERR_NONFATAL,
485 "`incbin': unexpected EOF while"
486 " reading file `%s'", fname);
487 t = 0; /* Try to exit cleanly */
488 break;
489 }
490 out(offset, segment, buf, OUT_RAWDATA, m,
491 NO_SEG, NO_SEG);
492 l -= m;
493 }
494 }
495 list->downlevel(LIST_INCBIN);
496 if (instruction->times > 1) {
497 /*
498 * Dummy call to list->output to give the offset to the
499 * listing module.
500 */
501 list->output(offset, NULL, OUT_RAWDATA, 0);
502 list->uplevel(LIST_TIMES);
503 list->downlevel(LIST_TIMES);
504 }
505 fclose(fp);
506 return instruction->times * len;
507 }
508 return 0; /* if we're here, there's an error */
509 }
510
511 /* Check to see if we need an address-size prefix */
512 add_asp(instruction, bits);
513
514 m = find_match(&temp, instruction, segment, offset, bits);
515
516 if (m == MOK_GOOD) {
517 /* Matches! */
518 int64_t insn_size = calcsize(segment, offset, bits,
519 instruction, temp->code);
520 itimes = instruction->times;
521 if (insn_size < 0) /* shouldn't be, on pass two */
522 error(ERR_PANIC, "errors made it through from pass one");
523 else
524 while (itimes--) {
525 for (j = 0; j < MAXPREFIX; j++) {
526 uint8_t c = 0;
527 switch (instruction->prefixes[j]) {
528 case P_WAIT:
529 c = 0x9B;
530 break;
531 case P_LOCK:
532 c = 0xF0;
533 break;
534 case P_REPNE:
535 case P_REPNZ:
536 c = 0xF2;
537 break;
538 case P_REPE:
539 case P_REPZ:
540 case P_REP:
541 c = 0xF3;
542 break;
543 case R_CS:
544 if (bits == 64) {
545 error(ERR_WARNING | ERR_PASS2,
546 "cs segment base generated, but will be ignored in 64-bit mode");
547 }
548 c = 0x2E;
549 break;
550 case R_DS:
551 if (bits == 64) {
552 error(ERR_WARNING | ERR_PASS2,
553 "ds segment base generated, but will be ignored in 64-bit mode");
554 }
555 c = 0x3E;
556 break;
557 case R_ES:
558 if (bits == 64) {
559 error(ERR_WARNING | ERR_PASS2,
560 "es segment base generated, but will be ignored in 64-bit mode");
561 }
562 c = 0x26;
563 break;
564 case R_FS:
565 c = 0x64;
566 break;
567 case R_GS:
568 c = 0x65;
569 break;
570 case R_SS:
571 if (bits == 64) {
572 error(ERR_WARNING | ERR_PASS2,
573 "ss segment base generated, but will be ignored in 64-bit mode");
574 }
575 c = 0x36;
576 break;
577 case R_SEGR6:
578 case R_SEGR7:
579 error(ERR_NONFATAL,
580 "segr6 and segr7 cannot be used as prefixes");
581 break;
582 case P_A16:
583 if (bits == 64) {
584 error(ERR_NONFATAL,
585 "16-bit addressing is not supported "
586 "in 64-bit mode");
587 } else if (bits != 16)
588 c = 0x67;
589 break;
590 case P_A32:
591 if (bits != 32)
592 c = 0x67;
593 break;
594 case P_A64:
595 if (bits != 64) {
596 error(ERR_NONFATAL,
597 "64-bit addressing is only supported "
598 "in 64-bit mode");
599 }
600 break;
601 case P_ASP:
602 c = 0x67;
603 break;
604 case P_O16:
605 if (bits != 16)
606 c = 0x66;
607 break;
608 case P_O32:
609 if (bits == 16)
610 c = 0x66;
611 break;
612 case P_O64:
613 /* REX.W */
614 break;
615 case P_OSP:
616 c = 0x66;
617 break;
618 case P_none:
619 break;
620 default:
621 error(ERR_PANIC, "invalid instruction prefix");
622 }
623 if (c != 0) {
624 out(offset, segment, &c, OUT_RAWDATA, 1,
625 NO_SEG, NO_SEG);
626 offset++;
627 }
628 }
629 insn_end = offset + insn_size;
630 gencode(segment, offset, bits, instruction,
631 temp, insn_end);
632 offset += insn_size;
633 if (itimes > 0 && itimes == instruction->times - 1) {
634 /*
635 * Dummy call to list->output to give the offset to the
636 * listing module.
637 */
638 list->output(offset, NULL, OUT_RAWDATA, 0);
639 list->uplevel(LIST_TIMES);
640 }
641 }
642 if (instruction->times > 1)
643 list->downlevel(LIST_TIMES);
644 return offset - start;
645 } else {
646 /* No match */
647 switch (m) {
648 case MERR_OPSIZEMISSING:
649 error(ERR_NONFATAL, "operation size not specified");
650 break;
651 case MERR_OPSIZEMISMATCH:
652 error(ERR_NONFATAL, "mismatch in operand sizes");
653 break;
654 case MERR_BADCPU:
655 error(ERR_NONFATAL, "no instruction for this cpu level");
656 break;
657 case MERR_BADMODE:
658 error(ERR_NONFATAL, "instruction not supported in %d-bit mode",
659 bits);
660 break;
661 default:
662 error(ERR_NONFATAL,
663 "invalid combination of opcode and operands");
664 break;
665 }
666 }
667 return 0;
668 }
669
670 int64_t insn_size(int32_t segment, int64_t offset, int bits, uint32_t cp,
671 insn * instruction, efunc error)
672 {
673 const struct itemplate *temp;
674 enum match_result m;
675
676 errfunc = error; /* to pass to other functions */
677 cpu = cp;
678
679 if (instruction->opcode == I_none)
680 return 0;
681
682 if (instruction->opcode == I_DB || instruction->opcode == I_DW ||
683 instruction->opcode == I_DD || instruction->opcode == I_DQ ||
684 instruction->opcode == I_DT || instruction->opcode == I_DO ||
685 instruction->opcode == I_DY) {
686 extop *e;
687 int32_t isize, osize, wsize = 0; /* placate gcc */
688
689 isize = 0;
690 switch (instruction->opcode) {
691 case I_DB:
692 wsize = 1;
693 break;
694 case I_DW:
695 wsize = 2;
696 break;
697 case I_DD:
698 wsize = 4;
699 break;
700 case I_DQ:
701 wsize = 8;
702 break;
703 case I_DT:
704 wsize = 10;
705 break;
706 case I_DO:
707 wsize = 16;
708 break;
709 case I_DY:
710 wsize = 32;
711 break;
712 default:
713 break;
714 }
715
716 list_for_each(e, instruction->eops) {
717 int32_t align;
718
719 osize = 0;
720 if (e->type == EOT_DB_NUMBER) {
721 osize = 1;
722 warn_overflow_const(e->offset, wsize);
723 } else if (e->type == EOT_DB_STRING ||
724 e->type == EOT_DB_STRING_FREE)
725 osize = e->stringlen;
726
727 align = (-osize) % wsize;
728 if (align < 0)
729 align += wsize;
730 isize += osize + align;
731 }
732 return isize * instruction->times;
733 }
734
735 if (instruction->opcode == I_INCBIN) {
736 const char *fname = instruction->eops->stringval;
737 FILE *fp;
738 size_t len;
739
740 fp = fopen(fname, "rb");
741 if (!fp)
742 error(ERR_NONFATAL, "`incbin': unable to open file `%s'",
743 fname);
744 else if (fseek(fp, 0L, SEEK_END) < 0)
745 error(ERR_NONFATAL, "`incbin': unable to seek on file `%s'",
746 fname);
747 else {
748 len = ftell(fp);
749 fclose(fp);
750 if (instruction->eops->next) {
751 len -= instruction->eops->next->offset;
752 if (instruction->eops->next->next &&
753 len > (size_t)instruction->eops->next->next->offset) {
754 len = (size_t)instruction->eops->next->next->offset;
755 }
756 }
757 return instruction->times * len;
758 }
759 return 0; /* if we're here, there's an error */
760 }
761
762 /* Check to see if we need an address-size prefix */
763 add_asp(instruction, bits);
764
765 m = find_match(&temp, instruction, segment, offset, bits);
766 if (m == MOK_GOOD) {
767 /* we've matched an instruction. */
768 int64_t isize;
769 const uint8_t *codes = temp->code;
770 int j;
771
772 isize = calcsize(segment, offset, bits, instruction, codes);
773 if (isize < 0)
774 return -1;
775 for (j = 0; j < MAXPREFIX; j++) {
776 switch (instruction->prefixes[j]) {
777 case P_A16:
778 if (bits != 16)
779 isize++;
780 break;
781 case P_A32:
782 if (bits != 32)
783 isize++;
784 break;
785 case P_O16:
786 if (bits != 16)
787 isize++;
788 break;
789 case P_O32:
790 if (bits == 16)
791 isize++;
792 break;
793 case P_A64:
794 case P_O64:
795 case P_none:
796 break;
797 default:
798 isize++;
799 break;
800 }
801 }
802 return isize * instruction->times;
803 } else {
804 return -1; /* didn't match any instruction */
805 }
806 }
807
808 static bool possible_sbyte(operand *o)
809 {
810 return o->wrt == NO_SEG && o->segment == NO_SEG &&
811 !(o->opflags & OPFLAG_UNKNOWN) &&
812 optimizing >= 0 && !(o->type & STRICT);
813 }
814
815 /* check that opn[op] is a signed byte of size 16 or 32 */
816 static bool is_sbyte16(operand *o)
817 {
818 int16_t v;
819
820 if (!possible_sbyte(o))
821 return false;
822
823 v = o->offset;
824 return v >= -128 && v <= 127;
825 }
826
827 static bool is_sbyte32(operand *o)
828 {
829 int32_t v;
830
831 if (!possible_sbyte(o))
832 return false;
833
834 v = o->offset;
835 return v >= -128 && v <= 127;
836 }
837
838 /* Common construct */
839 #define case4(x) case (x): case (x)+1: case (x)+2: case (x)+3
840
841 static int64_t calcsize(int32_t segment, int64_t offset, int bits,
842 insn * ins, const uint8_t *codes)
843 {
844 int64_t length = 0;
845 uint8_t c;
846 int rex_mask = ~0;
847 int op1, op2;
848 struct operand *opx;
849 uint8_t opex = 0;
850
851 ins->rex = 0; /* Ensure REX is reset */
852
853 if (ins->prefixes[PPS_OSIZE] == P_O64)
854 ins->rex |= REX_W;
855
856 (void)segment; /* Don't warn that this parameter is unused */
857 (void)offset; /* Don't warn that this parameter is unused */
858
859 while (*codes) {
860 c = *codes++;
861 op1 = (c & 3) + ((opex & 1) << 2);
862 op2 = ((c >> 3) & 3) + ((opex & 2) << 1);
863 opx = &ins->oprs[op1];
864 opex = 0; /* For the next iteration */
865
866 switch (c) {
867 case 01:
868 case 02:
869 case 03:
870 case 04:
871 codes += c, length += c;
872 break;
873
874 case 05:
875 case 06:
876 case 07:
877 opex = c;
878 break;
879
880 case4(010):
881 ins->rex |=
882 op_rexflags(opx, REX_B|REX_H|REX_P|REX_W);
883 codes++, length++;
884 break;
885
886 case4(014):
887 case4(020):
888 case4(024):
889 length++;
890 break;
891
892 case4(030):
893 length += 2;
894 break;
895
896 case4(034):
897 if (opx->type & (BITS16 | BITS32 | BITS64))
898 length += (opx->type & BITS16) ? 2 : 4;
899 else
900 length += (bits == 16) ? 2 : 4;
901 break;
902
903 case4(040):
904 length += 4;
905 break;
906
907 case4(044):
908 length += ins->addr_size >> 3;
909 break;
910
911 case4(050):
912 length++;
913 break;
914
915 case4(054):
916 length += 8; /* MOV reg64/imm */
917 break;
918
919 case4(060):
920 length += 2;
921 break;
922
923 case4(064):
924 if (opx->type & (BITS16 | BITS32 | BITS64))
925 length += (opx->type & BITS16) ? 2 : 4;
926 else
927 length += (bits == 16) ? 2 : 4;
928 break;
929
930 case4(070):
931 length += 4;
932 break;
933
934 case4(074):
935 length += 2;
936 break;
937
938 case4(0140):
939 length += is_sbyte16(opx) ? 1 : 2;
940 break;
941
942 case4(0144):
943 codes++;
944 length++;
945 break;
946
947 case4(0150):
948 length += is_sbyte32(opx) ? 1 : 4;
949 break;
950
951 case4(0154):
952 codes++;
953 length++;
954 break;
955
956 case4(0160):
957 length++;
958 ins->rex |= REX_D;
959 ins->drexdst = regval(opx);
960 break;
961
962 case4(0164):
963 length++;
964 ins->rex |= REX_D|REX_OC;
965 ins->drexdst = regval(opx);
966 break;
967
968 case 0171:
969 break;
970
971 case 0172:
972 case 0173:
973 case 0174:
974 codes++;
975 length++;
976 break;
977
978 case4(0250):
979 length += is_sbyte32(opx) ? 1 : 4;
980 break;
981
982 case4(0254):
983 length += 4;
984 break;
985
986 case4(0260):
987 ins->rex |= REX_V;
988 ins->drexdst = regval(opx);
989 ins->vex_cm = *codes++;
990 ins->vex_wlp = *codes++;
991 break;
992
993 case 0270:
994 ins->rex |= REX_V;
995 ins->drexdst = 0;
996 ins->vex_cm = *codes++;
997 ins->vex_wlp = *codes++;
998 break;
999
1000 case4(0274):
1001 length++;
1002 break;
1003
1004 case4(0300):
1005 break;
1006
1007 case 0310:
1008 if (bits == 64)
1009 return -1;
1010 length += (bits != 16) && !has_prefix(ins, PPS_ASIZE, P_A16);
1011 break;
1012
1013 case 0311:
1014 length += (bits != 32) && !has_prefix(ins, PPS_ASIZE, P_A32);
1015 break;
1016
1017 case 0312:
1018 break;
1019
1020 case 0313:
1021 if (bits != 64 || has_prefix(ins, PPS_ASIZE, P_A16) ||
1022 has_prefix(ins, PPS_ASIZE, P_A32))
1023 return -1;
1024 break;
1025
1026 case4(0314):
1027 break;
1028
1029 case 0320:
1030 length += (bits != 16);
1031 break;
1032
1033 case 0321:
1034 length += (bits == 16);
1035 break;
1036
1037 case 0322:
1038 break;
1039
1040 case 0323:
1041 rex_mask &= ~REX_W;
1042 break;
1043
1044 case 0324:
1045 ins->rex |= REX_W;
1046 break;
1047
1048 case 0325:
1049 ins->rex |= REX_NH;
1050 break;
1051
1052 case 0330:
1053 codes++, length++;
1054 break;
1055
1056 case 0331:
1057 break;
1058
1059 case 0332:
1060 case 0333:
1061 length++;
1062 break;
1063
1064 case 0334:
1065 ins->rex |= REX_L;
1066 break;
1067
1068 case 0335:
1069 break;
1070
1071 case 0336:
1072 if (!ins->prefixes[PPS_LREP])
1073 ins->prefixes[PPS_LREP] = P_REP;
1074 break;
1075
1076 case 0337:
1077 if (!ins->prefixes[PPS_LREP])
1078 ins->prefixes[PPS_LREP] = P_REPNE;
1079 break;
1080
1081 case 0340:
1082 if (ins->oprs[0].segment != NO_SEG)
1083 errfunc(ERR_NONFATAL, "attempt to reserve non-constant"
1084 " quantity of BSS space");
1085 else
1086 length += ins->oprs[0].offset;
1087 break;
1088
1089 case 0341:
1090 if (!ins->prefixes[PPS_WAIT])
1091 ins->prefixes[PPS_WAIT] = P_WAIT;
1092 break;
1093
1094 case4(0344):
1095 length++;
1096 break;
1097
1098 case 0360:
1099 break;
1100
1101 case 0361:
1102 case 0362:
1103 case 0363:
1104 length++;
1105 break;
1106
1107 case 0364:
1108 case 0365:
1109 break;
1110
1111 case 0366:
1112 case 0367:
1113 length++;
1114 break;
1115
1116 case 0370:
1117 case 0371:
1118 case 0372:
1119 break;
1120
1121 case 0373:
1122 length++;
1123 break;
1124
1125 case4(0100):
1126 case4(0110):
1127 case4(0120):
1128 case4(0130):
1129 case4(0200):
1130 case4(0204):
1131 case4(0210):
1132 case4(0214):
1133 case4(0220):
1134 case4(0224):
1135 case4(0230):
1136 case4(0234):
1137 {
1138 ea ea_data;
1139 int rfield;
1140 int32_t rflags;
1141 struct operand *opy = &ins->oprs[op2];
1142
1143 ea_data.rex = 0; /* Ensure ea.REX is initially 0 */
1144
1145 if (c <= 0177) {
1146 /* pick rfield from operand b (opx) */
1147 rflags = regflag(opx);
1148 rfield = nasm_regvals[opx->basereg];
1149 } else {
1150 rflags = 0;
1151 rfield = c & 7;
1152 }
1153 if (!process_ea(opy, &ea_data, bits,
1154 ins->addr_size, rfield, rflags)) {
1155 errfunc(ERR_NONFATAL, "invalid effective address");
1156 return -1;
1157 } else {
1158 ins->rex |= ea_data.rex;
1159 length += ea_data.size;
1160 }
1161 }
1162 break;
1163
1164 default:
1165 errfunc(ERR_PANIC, "internal instruction table corrupt"
1166 ": instruction code \\%o (0x%02X) given", c, c);
1167 break;
1168 }
1169 }
1170
1171 ins->rex &= rex_mask;
1172
1173 if (ins->rex & REX_NH) {
1174 if (ins->rex & REX_H) {
1175 errfunc(ERR_NONFATAL, "instruction cannot use high registers");
1176 return -1;
1177 }
1178 ins->rex &= ~REX_P; /* Don't force REX prefix due to high reg */
1179 }
1180
1181 if (ins->rex & REX_V) {
1182 int bad32 = REX_R|REX_W|REX_X|REX_B;
1183
1184 if (ins->rex & REX_H) {
1185 errfunc(ERR_NONFATAL, "cannot use high register in vex instruction");
1186 return -1;
1187 }
1188 switch (ins->vex_wlp & 030) {
1189 case 000:
1190 case 020:
1191 ins->rex &= ~REX_W;
1192 break;
1193 case 010:
1194 ins->rex |= REX_W;
1195 bad32 &= ~REX_W;
1196 break;
1197 case 030:
1198 /* Follow REX_W */
1199 break;
1200 }
1201
1202 if (bits != 64 && ((ins->rex & bad32) || ins->drexdst > 7)) {
1203 errfunc(ERR_NONFATAL, "invalid operands in non-64-bit mode");
1204 return -1;
1205 }
1206 if (ins->vex_cm != 1 || (ins->rex & (REX_W|REX_R|REX_B)))
1207 length += 3;
1208 else
1209 length += 2;
1210 } else if (ins->rex & REX_D) {
1211 if (ins->rex & REX_H) {
1212 errfunc(ERR_NONFATAL, "cannot use high register in drex instruction");
1213 return -1;
1214 }
1215 if (bits != 64 && ((ins->rex & (REX_R|REX_W|REX_X|REX_B)) ||
1216 ins->drexdst > 7)) {
1217 errfunc(ERR_NONFATAL, "invalid operands in non-64-bit mode");
1218 return -1;
1219 }
1220 length++;
1221 } else if (ins->rex & REX_REAL) {
1222 if (ins->rex & REX_H) {
1223 errfunc(ERR_NONFATAL, "cannot use high register in rex instruction");
1224 return -1;
1225 } else if (bits == 64) {
1226 length++;
1227 } else if ((ins->rex & REX_L) &&
1228 !(ins->rex & (REX_P|REX_W|REX_X|REX_B)) &&
1229 cpu >= IF_X86_64) {
1230 /* LOCK-as-REX.R */
1231 assert_no_prefix(ins, PPS_LREP);
1232 length++;
1233 } else {
1234 errfunc(ERR_NONFATAL, "invalid operands in non-64-bit mode");
1235 return -1;
1236 }
1237 }
1238
1239 return length;
1240 }
1241
1242 #define EMIT_REX() \
1243 if (!(ins->rex & (REX_D|REX_V)) && (ins->rex & REX_REAL) && (bits == 64)) { \
1244 ins->rex = (ins->rex & REX_REAL)|REX_P; \
1245 out(offset, segment, &ins->rex, OUT_RAWDATA, 1, NO_SEG, NO_SEG); \
1246 ins->rex = 0; \
1247 offset += 1; \
1248 }
1249
1250 static void gencode(int32_t segment, int64_t offset, int bits,
1251 insn * ins, const struct itemplate *temp,
1252 int64_t insn_end)
1253 {
1254 static char condval[] = { /* conditional opcodes */
1255 0x7, 0x3, 0x2, 0x6, 0x2, 0x4, 0xF, 0xD, 0xC, 0xE, 0x6, 0x2,
1256 0x3, 0x7, 0x3, 0x5, 0xE, 0xC, 0xD, 0xF, 0x1, 0xB, 0x9, 0x5,
1257 0x0, 0xA, 0xA, 0xB, 0x8, 0x4
1258 };
1259 uint8_t c;
1260 uint8_t bytes[4];
1261 int64_t size;
1262 int64_t data;
1263 int op1, op2;
1264 struct operand *opx;
1265 const uint8_t *codes = temp->code;
1266 uint8_t opex = 0;
1267
1268 while (*codes) {
1269 c = *codes++;
1270 op1 = (c & 3) + ((opex & 1) << 2);
1271 op2 = ((c >> 3) & 3) + ((opex & 2) << 1);
1272 opx = &ins->oprs[op1];
1273 opex = 0; /* For the next iteration */
1274
1275 switch (c) {
1276 case 01:
1277 case 02:
1278 case 03:
1279 case 04:
1280 EMIT_REX();
1281 out(offset, segment, codes, OUT_RAWDATA, c, NO_SEG, NO_SEG);
1282 codes += c;
1283 offset += c;
1284 break;
1285
1286 case 05:
1287 case 06:
1288 case 07:
1289 opex = c;
1290 break;
1291
1292 case4(010):
1293 EMIT_REX();
1294 bytes[0] = *codes++ + (regval(opx) & 7);
1295 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1296 offset += 1;
1297 break;
1298
1299 case4(014):
1300 /* The test for BITS8 and SBYTE here is intended to avoid
1301 warning on optimizer actions due to SBYTE, while still
1302 warn on explicit BYTE directives. Also warn, obviously,
1303 if the optimizer isn't enabled. */
1304 if (((opx->type & BITS8) ||
1305 !(opx->type & temp->opd[op1] & BYTENESS)) &&
1306 (opx->offset < -128 || opx->offset > 127)) {
1307 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1308 "signed byte value exceeds bounds");
1309 }
1310 if (opx->segment != NO_SEG) {
1311 data = opx->offset;
1312 out(offset, segment, &data, OUT_ADDRESS, 1,
1313 opx->segment, opx->wrt);
1314 } else {
1315 bytes[0] = opx->offset;
1316 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1317 NO_SEG);
1318 }
1319 offset += 1;
1320 break;
1321
1322 case4(020):
1323 if (opx->offset < -256 || opx->offset > 255) {
1324 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1325 "byte value exceeds bounds");
1326 }
1327 if (opx->segment != NO_SEG) {
1328 data = opx->offset;
1329 out(offset, segment, &data, OUT_ADDRESS, 1,
1330 opx->segment, opx->wrt);
1331 } else {
1332 bytes[0] = opx->offset;
1333 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1334 NO_SEG);
1335 }
1336 offset += 1;
1337 break;
1338
1339 case4(024):
1340 if (opx->offset < 0 || opx->offset > 255)
1341 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1342 "unsigned byte value exceeds bounds");
1343 if (opx->segment != NO_SEG) {
1344 data = opx->offset;
1345 out(offset, segment, &data, OUT_ADDRESS, 1,
1346 opx->segment, opx->wrt);
1347 } else {
1348 bytes[0] = opx->offset;
1349 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1350 NO_SEG);
1351 }
1352 offset += 1;
1353 break;
1354
1355 case4(030):
1356 warn_overflow_opd(opx, 2);
1357 data = opx->offset;
1358 out(offset, segment, &data, OUT_ADDRESS, 2,
1359 opx->segment, opx->wrt);
1360 offset += 2;
1361 break;
1362
1363 case4(034):
1364 if (opx->type & (BITS16 | BITS32))
1365 size = (opx->type & BITS16) ? 2 : 4;
1366 else
1367 size = (bits == 16) ? 2 : 4;
1368 warn_overflow_opd(opx, size);
1369 data = opx->offset;
1370 out(offset, segment, &data, OUT_ADDRESS, size,
1371 opx->segment, opx->wrt);
1372 offset += size;
1373 break;
1374
1375 case4(040):
1376 warn_overflow_opd(opx, 4);
1377 data = opx->offset;
1378 out(offset, segment, &data, OUT_ADDRESS, 4,
1379 opx->segment, opx->wrt);
1380 offset += 4;
1381 break;
1382
1383 case4(044):
1384 data = opx->offset;
1385 size = ins->addr_size >> 3;
1386 warn_overflow_opd(opx, size);
1387 out(offset, segment, &data, OUT_ADDRESS, size,
1388 opx->segment, opx->wrt);
1389 offset += size;
1390 break;
1391
1392 case4(050):
1393 if (opx->segment != segment)
1394 errfunc(ERR_NONFATAL,
1395 "short relative jump outside segment");
1396 data = opx->offset - insn_end;
1397 if (data > 127 || data < -128)
1398 errfunc(ERR_NONFATAL, "short jump is out of range");
1399 bytes[0] = data;
1400 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1401 offset += 1;
1402 break;
1403
1404 case4(054):
1405 data = (int64_t)opx->offset;
1406 out(offset, segment, &data, OUT_ADDRESS, 8,
1407 opx->segment, opx->wrt);
1408 offset += 8;
1409 break;
1410
1411 case4(060):
1412 if (opx->segment != segment) {
1413 data = opx->offset;
1414 out(offset, segment, &data,
1415 OUT_REL2ADR, insn_end - offset,
1416 opx->segment, opx->wrt);
1417 } else {
1418 data = opx->offset - insn_end;
1419 out(offset, segment, &data,
1420 OUT_ADDRESS, 2, NO_SEG, NO_SEG);
1421 }
1422 offset += 2;
1423 break;
1424
1425 case4(064):
1426 if (opx->type & (BITS16 | BITS32 | BITS64))
1427 size = (opx->type & BITS16) ? 2 : 4;
1428 else
1429 size = (bits == 16) ? 2 : 4;
1430 if (opx->segment != segment) {
1431 data = opx->offset;
1432 out(offset, segment, &data,
1433 size == 2 ? OUT_REL2ADR : OUT_REL4ADR,
1434 insn_end - offset, opx->segment, opx->wrt);
1435 } else {
1436 data = opx->offset - insn_end;
1437 out(offset, segment, &data,
1438 OUT_ADDRESS, size, NO_SEG, NO_SEG);
1439 }
1440 offset += size;
1441 break;
1442
1443 case4(070):
1444 if (opx->segment != segment) {
1445 data = opx->offset;
1446 out(offset, segment, &data,
1447 OUT_REL4ADR, insn_end - offset,
1448 opx->segment, opx->wrt);
1449 } else {
1450 data = opx->offset - insn_end;
1451 out(offset, segment, &data,
1452 OUT_ADDRESS, 4, NO_SEG, NO_SEG);
1453 }
1454 offset += 4;
1455 break;
1456
1457 case4(074):
1458 if (opx->segment == NO_SEG)
1459 errfunc(ERR_NONFATAL, "value referenced by FAR is not"
1460 " relocatable");
1461 data = 0;
1462 out(offset, segment, &data, OUT_ADDRESS, 2,
1463 outfmt->segbase(1 + opx->segment),
1464 opx->wrt);
1465 offset += 2;
1466 break;
1467
1468 case4(0140):
1469 data = opx->offset;
1470 warn_overflow_opd(opx, 2);
1471 if (is_sbyte16(opx)) {
1472 bytes[0] = data;
1473 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1474 NO_SEG);
1475 offset++;
1476 } else {
1477 out(offset, segment, &data, OUT_ADDRESS, 2,
1478 opx->segment, opx->wrt);
1479 offset += 2;
1480 }
1481 break;
1482
1483 case4(0144):
1484 EMIT_REX();
1485 bytes[0] = *codes++;
1486 if (is_sbyte16(opx))
1487 bytes[0] |= 2; /* s-bit */
1488 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1489 offset++;
1490 break;
1491
1492 case4(0150):
1493 data = opx->offset;
1494 warn_overflow_opd(opx, 4);
1495 if (is_sbyte32(opx)) {
1496 bytes[0] = data;
1497 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1498 NO_SEG);
1499 offset++;
1500 } else {
1501 out(offset, segment, &data, OUT_ADDRESS, 4,
1502 opx->segment, opx->wrt);
1503 offset += 4;
1504 }
1505 break;
1506
1507 case4(0154):
1508 EMIT_REX();
1509 bytes[0] = *codes++;
1510 if (is_sbyte32(opx))
1511 bytes[0] |= 2; /* s-bit */
1512 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1513 offset++;
1514 break;
1515
1516 case4(0160):
1517 case4(0164):
1518 break;
1519
1520 case 0171:
1521 bytes[0] =
1522 (ins->drexdst << 4) |
1523 (ins->rex & REX_OC ? 0x08 : 0) |
1524 (ins->rex & (REX_R|REX_X|REX_B));
1525 ins->rex = 0;
1526 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1527 offset++;
1528 break;
1529
1530 case 0172:
1531 c = *codes++;
1532 opx = &ins->oprs[c >> 3];
1533 bytes[0] = nasm_regvals[opx->basereg] << 4;
1534 opx = &ins->oprs[c & 7];
1535 if (opx->segment != NO_SEG || opx->wrt != NO_SEG) {
1536 errfunc(ERR_NONFATAL,
1537 "non-absolute expression not permitted as argument %d",
1538 c & 7);
1539 } else {
1540 if (opx->offset & ~15) {
1541 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1542 "four-bit argument exceeds bounds");
1543 }
1544 bytes[0] |= opx->offset & 15;
1545 }
1546 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1547 offset++;
1548 break;
1549
1550 case 0173:
1551 c = *codes++;
1552 opx = &ins->oprs[c >> 4];
1553 bytes[0] = nasm_regvals[opx->basereg] << 4;
1554 bytes[0] |= c & 15;
1555 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1556 offset++;
1557 break;
1558
1559 case 0174:
1560 c = *codes++;
1561 opx = &ins->oprs[c];
1562 bytes[0] = nasm_regvals[opx->basereg] << 4;
1563 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1564 offset++;
1565 break;
1566
1567 case4(0250):
1568 data = opx->offset;
1569 if (opx->wrt == NO_SEG && opx->segment == NO_SEG &&
1570 (int32_t)data != (int64_t)data) {
1571 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1572 "signed dword immediate exceeds bounds");
1573 }
1574 if (is_sbyte32(opx)) {
1575 bytes[0] = data;
1576 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1577 NO_SEG);
1578 offset++;
1579 } else {
1580 out(offset, segment, &data, OUT_ADDRESS, 4,
1581 opx->segment, opx->wrt);
1582 offset += 4;
1583 }
1584 break;
1585
1586 case4(0254):
1587 data = opx->offset;
1588 if (opx->wrt == NO_SEG && opx->segment == NO_SEG &&
1589 (int32_t)data != (int64_t)data) {
1590 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1591 "signed dword immediate exceeds bounds");
1592 }
1593 out(offset, segment, &data, OUT_ADDRESS, 4,
1594 opx->segment, opx->wrt);
1595 offset += 4;
1596 break;
1597
1598 case4(0260):
1599 case 0270:
1600 codes += 2;
1601 if (ins->vex_cm != 1 || (ins->rex & (REX_W|REX_X|REX_B))) {
1602 bytes[0] = (ins->vex_cm >> 6) ? 0x8f : 0xc4;
1603 bytes[1] = (ins->vex_cm & 31) | ((~ins->rex & 7) << 5);
1604 bytes[2] = ((ins->rex & REX_W) << (7-3)) |
1605 ((~ins->drexdst & 15)<< 3) | (ins->vex_wlp & 07);
1606 out(offset, segment, &bytes, OUT_RAWDATA, 3, NO_SEG, NO_SEG);
1607 offset += 3;
1608 } else {
1609 bytes[0] = 0xc5;
1610 bytes[1] = ((~ins->rex & REX_R) << (7-2)) |
1611 ((~ins->drexdst & 15) << 3) | (ins->vex_wlp & 07);
1612 out(offset, segment, &bytes, OUT_RAWDATA, 2, NO_SEG, NO_SEG);
1613 offset += 2;
1614 }
1615 break;
1616
1617 case4(0274):
1618 {
1619 uint64_t uv, um;
1620 int s;
1621
1622 if (ins->rex & REX_W)
1623 s = 64;
1624 else if (ins->prefixes[PPS_OSIZE] == P_O16)
1625 s = 16;
1626 else if (ins->prefixes[PPS_OSIZE] == P_O32)
1627 s = 32;
1628 else
1629 s = bits;
1630
1631 um = (uint64_t)2 << (s-1);
1632 uv = opx->offset;
1633
1634 if (uv > 127 && uv < (uint64_t)-128 &&
1635 (uv < um-128 || uv > um-1)) {
1636 errfunc(ERR_WARNING | ERR_PASS2 | ERR_WARN_NOV,
1637 "signed byte value exceeds bounds");
1638 }
1639 if (opx->segment != NO_SEG) {
1640 data = uv;
1641 out(offset, segment, &data, OUT_ADDRESS, 1,
1642 opx->segment, opx->wrt);
1643 } else {
1644 bytes[0] = uv;
1645 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG,
1646 NO_SEG);
1647 }
1648 offset += 1;
1649 break;
1650 }
1651
1652 case4(0300):
1653 break;
1654
1655 case 0310:
1656 if (bits == 32 && !has_prefix(ins, PPS_ASIZE, P_A16)) {
1657 *bytes = 0x67;
1658 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1659 offset += 1;
1660 } else
1661 offset += 0;
1662 break;
1663
1664 case 0311:
1665 if (bits != 32 && !has_prefix(ins, PPS_ASIZE, P_A32)) {
1666 *bytes = 0x67;
1667 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1668 offset += 1;
1669 } else
1670 offset += 0;
1671 break;
1672
1673 case 0312:
1674 break;
1675
1676 case 0313:
1677 ins->rex = 0;
1678 break;
1679
1680 case4(0314):
1681 break;
1682
1683 case 0320:
1684 if (bits != 16) {
1685 *bytes = 0x66;
1686 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1687 offset += 1;
1688 } else
1689 offset += 0;
1690 break;
1691
1692 case 0321:
1693 if (bits == 16) {
1694 *bytes = 0x66;
1695 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1696 offset += 1;
1697 } else
1698 offset += 0;
1699 break;
1700
1701 case 0322:
1702 case 0323:
1703 break;
1704
1705 case 0324:
1706 ins->rex |= REX_W;
1707 break;
1708
1709 case 0325:
1710 break;
1711
1712 case 0330:
1713 *bytes = *codes++ ^ condval[ins->condition];
1714 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1715 offset += 1;
1716 break;
1717
1718 case 0331:
1719 break;
1720
1721 case 0332:
1722 case 0333:
1723 *bytes = c - 0332 + 0xF2;
1724 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1725 offset += 1;
1726 break;
1727
1728 case 0334:
1729 if (ins->rex & REX_R) {
1730 *bytes = 0xF0;
1731 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1732 offset += 1;
1733 }
1734 ins->rex &= ~(REX_L|REX_R);
1735 break;
1736
1737 case 0335:
1738 break;
1739
1740 case 0336:
1741 case 0337:
1742 break;
1743
1744 case 0340:
1745 if (ins->oprs[0].segment != NO_SEG)
1746 errfunc(ERR_PANIC, "non-constant BSS size in pass two");
1747 else {
1748 int64_t size = ins->oprs[0].offset;
1749 if (size > 0)
1750 out(offset, segment, NULL,
1751 OUT_RESERVE, size, NO_SEG, NO_SEG);
1752 offset += size;
1753 }
1754 break;
1755
1756 case 0341:
1757 break;
1758
1759 case 0344:
1760 case 0345:
1761 bytes[0] = c & 1;
1762 switch (ins->oprs[0].basereg) {
1763 case R_CS:
1764 bytes[0] += 0x0E;
1765 break;
1766 case R_DS:
1767 bytes[0] += 0x1E;
1768 break;
1769 case R_ES:
1770 bytes[0] += 0x06;
1771 break;
1772 case R_SS:
1773 bytes[0] += 0x16;
1774 break;
1775 default:
1776 errfunc(ERR_PANIC,
1777 "bizarre 8086 segment register received");
1778 }
1779 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1780 offset++;
1781 break;
1782
1783 case 0346:
1784 case 0347:
1785 bytes[0] = c & 1;
1786 switch (ins->oprs[0].basereg) {
1787 case R_FS:
1788 bytes[0] += 0xA0;
1789 break;
1790 case R_GS:
1791 bytes[0] += 0xA8;
1792 break;
1793 default:
1794 errfunc(ERR_PANIC,
1795 "bizarre 386 segment register received");
1796 }
1797 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1798 offset++;
1799 break;
1800
1801 case 0360:
1802 break;
1803
1804 case 0361:
1805 bytes[0] = 0x66;
1806 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1807 offset += 1;
1808 break;
1809
1810 case 0362:
1811 case 0363:
1812 bytes[0] = c - 0362 + 0xf2;
1813 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1814 offset += 1;
1815 break;
1816
1817 case 0364:
1818 case 0365:
1819 break;
1820
1821 case 0366:
1822 case 0367:
1823 *bytes = c - 0366 + 0x66;
1824 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1825 offset += 1;
1826 break;
1827
1828 case 0370:
1829 case 0371:
1830 case 0372:
1831 break;
1832
1833 case 0373:
1834 *bytes = bits == 16 ? 3 : 5;
1835 out(offset, segment, bytes, OUT_RAWDATA, 1, NO_SEG, NO_SEG);
1836 offset += 1;
1837 break;
1838
1839 case4(0100):
1840 case4(0110):
1841 case4(0120):
1842 case4(0130):
1843 case4(0200):
1844 case4(0204):
1845 case4(0210):
1846 case4(0214):
1847 case4(0220):
1848 case4(0224):
1849 case4(0230):
1850 case4(0234):
1851 {
1852 ea ea_data;
1853 int rfield;
1854 int32_t rflags;
1855 uint8_t *p;
1856 int32_t s;
1857 enum out_type type;
1858 struct operand *opy = &ins->oprs[op2];
1859
1860 if (c <= 0177) {
1861 /* pick rfield from operand b (opx) */
1862 rflags = regflag(opx);
1863 rfield = nasm_regvals[opx->basereg];
1864 } else {
1865 /* rfield is constant */
1866 rflags = 0;
1867 rfield = c & 7;
1868 }
1869
1870 if (!process_ea(opy, &ea_data, bits, ins->addr_size,
1871 rfield, rflags)) {
1872 errfunc(ERR_NONFATAL, "invalid effective address");
1873 }
1874
1875
1876 p = bytes;
1877 *p++ = ea_data.modrm;
1878 if (ea_data.sib_present)
1879 *p++ = ea_data.sib;
1880
1881 /* DREX suffixes come between the SIB and the displacement */
1882 if (ins->rex & REX_D) {
1883 *p++ = (ins->drexdst << 4) |
1884 (ins->rex & REX_OC ? 0x08 : 0) |
1885 (ins->rex & (REX_R|REX_X|REX_B));
1886 ins->rex = 0;
1887 }
1888
1889 s = p - bytes;
1890 out(offset, segment, bytes, OUT_RAWDATA, s, NO_SEG, NO_SEG);
1891
1892 /*
1893 * Make sure the address gets the right offset in case
1894 * the line breaks in the .lst file (BR 1197827)
1895 */
1896 offset += s;
1897 s = 0;
1898
1899 switch (ea_data.bytes) {
1900 case 0:
1901 break;
1902 case 1:
1903 case 2:
1904 case 4:
1905 case 8:
1906 data = opy->offset;
1907 warn_overflow_opd(opy, ea_data.bytes);
1908 s += ea_data.bytes;
1909 if (ea_data.rip) {
1910 if (opy->segment == segment) {
1911 data -= insn_end;
1912 out(offset, segment, &data, OUT_ADDRESS,
1913 ea_data.bytes, NO_SEG, NO_SEG);
1914 } else {
1915 out(offset, segment, &data, OUT_REL4ADR,
1916 insn_end - offset, opy->segment, opy->wrt);
1917 }
1918 } else {
1919 type = OUT_ADDRESS;
1920 out(offset, segment, &data, OUT_ADDRESS,
1921 ea_data.bytes, opy->segment, opy->wrt);
1922 }
1923 break;
1924 default:
1925 /* Impossible! */
1926 errfunc(ERR_PANIC,
1927 "Invalid amount of bytes (%d) for offset?!",
1928 ea_data.bytes);
1929 break;
1930 }
1931 offset += s;
1932 }
1933 break;
1934
1935 default:
1936 errfunc(ERR_PANIC, "internal instruction table corrupt"
1937 ": instruction code \\%o (0x%02X) given", c, c);
1938 break;
1939 }
1940 }
1941 }
1942
1943 static int32_t regflag(const operand * o)
1944 {
1945 if (o->basereg < EXPR_REG_START || o->basereg >= REG_ENUM_LIMIT) {
1946 errfunc(ERR_PANIC, "invalid operand passed to regflag()");
1947 }
1948 return nasm_reg_flags[o->basereg];
1949 }
1950
1951 static int32_t regval(const operand * o)
1952 {
1953 if (o->basereg < EXPR_REG_START || o->basereg >= REG_ENUM_LIMIT) {
1954 errfunc(ERR_PANIC, "invalid operand passed to regval()");
1955 }
1956 return nasm_regvals[o->basereg];
1957 }
1958
1959 static int op_rexflags(const operand * o, int mask)
1960 {
1961 int32_t flags;
1962 int val;
1963
1964 if (o->basereg < EXPR_REG_START || o->basereg >= REG_ENUM_LIMIT) {
1965 errfunc(ERR_PANIC, "invalid operand passed to op_rexflags()");
1966 }
1967
1968 flags = nasm_reg_flags[o->basereg];
1969 val = nasm_regvals[o->basereg];
1970
1971 return rexflags(val, flags, mask);
1972 }
1973
1974 static int rexflags(int val, int32_t flags, int mask)
1975 {
1976 int rex = 0;
1977
1978 if (val >= 8)
1979 rex |= REX_B|REX_X|REX_R;
1980 if (flags & BITS64)
1981 rex |= REX_W;
1982 if (!(REG_HIGH & ~flags)) /* AH, CH, DH, BH */
1983 rex |= REX_H;
1984 else if (!(REG8 & ~flags) && val >= 4) /* SPL, BPL, SIL, DIL */
1985 rex |= REX_P;
1986
1987 return rex & mask;
1988 }
1989
1990 static enum match_result find_match(const struct itemplate **tempp,
1991 insn *instruction,
1992 int32_t segment, int64_t offset, int bits)
1993 {
1994 const struct itemplate *temp;
1995 enum match_result m, merr;
1996 int32_t xsizeflags[MAX_OPERANDS];
1997 bool opsizemissing = false;
1998 int i;
1999
2000 for (i = 0; i < instruction->operands; i++)
2001 xsizeflags[i] = instruction->oprs[i].type & SIZE_MASK;
2002
2003 merr = MERR_INVALOP;
2004
2005 for (temp = nasm_instructions[instruction->opcode];
2006 temp->opcode != I_none; temp++) {
2007 m = matches(temp, instruction, bits);
2008 if (m == MOK_JUMP) {
2009 if (jmp_match(segment, offset, bits, instruction, temp->code))
2010 m = MOK_GOOD;
2011 else
2012 m = MERR_INVALOP;
2013 } else if (m == MERR_OPSIZEMISSING &&
2014 (temp->flags & IF_SMASK) != IF_SX) {
2015 /*
2016 * Missing operand size and a candidate for fuzzy matching...
2017 */
2018 for (i = 0; i < temp->operands; i++) {
2019 if ((temp->opd[i] & SAME_AS) == 0)
2020 xsizeflags[i] |= temp->opd[i] & SIZE_MASK;
2021 }
2022 opsizemissing = true;
2023 }
2024 if (m > merr)
2025 merr = m;
2026 if (merr == MOK_GOOD)
2027 goto done;
2028 }
2029
2030 /* No match, but see if we can get a fuzzy operand size match... */
2031 if (!opsizemissing)
2032 goto done;
2033
2034 for (i = 0; i < instruction->operands; i++) {
2035 /*
2036 * We ignore extrinsic operand sizes on registers, so we should
2037 * never try to fuzzy-match on them. This also resolves the case
2038 * when we have e.g. "xmmrm128" in two different positions.
2039 */
2040 if (is_class(REGISTER, instruction->oprs[i].type))
2041 continue;
2042
2043 /* This tests if xsizeflags[i] has more than one bit set */
2044 if ((xsizeflags[i] & (xsizeflags[i]-1)))
2045 goto done; /* No luck */
2046
2047 instruction->oprs[i].type |= xsizeflags[i]; /* Set the size */
2048 }
2049
2050 /* Try matching again... */
2051 for (temp = nasm_instructions[instruction->opcode];
2052 temp->opcode != I_none; temp++) {
2053 m = matches(temp, instruction, bits);
2054 if (m == MOK_JUMP) {
2055 if (jmp_match(segment, offset, bits, instruction, temp->code))
2056 m = MOK_GOOD;
2057 else
2058 m = MERR_INVALOP;
2059 }
2060 if (m > merr)
2061 merr = m;
2062 if (merr == MOK_GOOD)
2063 goto done;
2064 }
2065
2066 done:
2067 *tempp = temp;
2068 return merr;
2069 }
2070
2071 static enum match_result matches(const struct itemplate *itemp,
2072 insn *instruction, int bits)
2073 {
2074 int i, size[MAX_OPERANDS], asize, oprs;
2075 bool opsizemissing = false;
2076
2077 /*
2078 * Check the opcode
2079 */
2080 if (itemp->opcode != instruction->opcode)
2081 return MERR_INVALOP;
2082
2083 /*
2084 * Count the operands
2085 */
2086 if (itemp->operands != instruction->operands)
2087 return MERR_INVALOP;
2088
2089 /*
2090 * Check that no spurious colons or TOs are present
2091 */
2092 for (i = 0; i < itemp->operands; i++)
2093 if (instruction->oprs[i].type & ~itemp->opd[i] & (COLON | TO))
2094 return MERR_INVALOP;
2095
2096 /*
2097 * Process size flags
2098 */
2099 switch (itemp->flags & IF_SMASK) {
2100 case IF_SB:
2101 asize = BITS8;
2102 break;
2103 case IF_SW:
2104 asize = BITS16;
2105 break;
2106 case IF_SD:
2107 asize = BITS32;
2108 break;
2109 case IF_SQ:
2110 asize = BITS64;
2111 break;
2112 case IF_SO:
2113 asize = BITS128;
2114 break;
2115 case IF_SY:
2116 asize = BITS256;
2117 break;
2118 case IF_SZ:
2119 switch (bits) {
2120 case 16:
2121 asize = BITS16;
2122 break;
2123 case 32:
2124 asize = BITS32;
2125 break;
2126 case 64:
2127 asize = BITS64;
2128 break;
2129 default:
2130 asize = 0;
2131 break;
2132 }
2133 break;
2134 default:
2135 asize = 0;
2136 break;
2137 }
2138
2139 if (itemp->flags & IF_ARMASK) {
2140 /* S- flags only apply to a specific operand */
2141 i = ((itemp->flags & IF_ARMASK) >> IF_ARSHFT) - 1;
2142 memset(size, 0, sizeof size);
2143 size[i] = asize;
2144 } else {
2145 /* S- flags apply to all operands */
2146 for (i = 0; i < MAX_OPERANDS; i++)
2147 size[i] = asize;
2148 }
2149
2150 /*
2151 * Check that the operand flags all match up
2152 */
2153 for (i = 0; i < itemp->operands; i++) {
2154 opflags_t type = instruction->oprs[i].type;
2155 if (!(type & SIZE_MASK))
2156 type |= size[i];
2157
2158 if (itemp->opd[i] & SAME_AS) {
2159 int j = itemp->opd[i] & ~SAME_AS;
2160 if (type != instruction->oprs[j].type ||
2161 instruction->oprs[i].basereg != instruction->oprs[j].basereg)
2162 return MERR_INVALOP;
2163 } else if (itemp->opd[i] & ~type ||
2164 ((itemp->opd[i] & SIZE_MASK) &&
2165 ((itemp->opd[i] ^ type) & SIZE_MASK))) {
2166 if ((itemp->opd[i] & ~type & ~SIZE_MASK) || (type & SIZE_MASK)) {
2167 return MERR_INVALOP;
2168 } else if (!is_class(REGISTER, type)) {
2169 /*
2170 * Note: we don't honor extrinsic operand sizes for registers,
2171 * so "missing operand size" for a register should be
2172 * considered a wildcard match rather than an error.
2173 */
2174 opsizemissing = true;
2175 }
2176 }
2177 }
2178
2179 if (opsizemissing)
2180 return MERR_OPSIZEMISSING;
2181
2182 /*
2183 * Check operand sizes
2184 */
2185 if (itemp->flags & (IF_SM | IF_SM2)) {
2186 oprs = (itemp->flags & IF_SM2 ? 2 : itemp->operands);
2187 asize = 0;
2188 for (i = 0; i < oprs; i++) {
2189 if ((asize = itemp->opd[i] & SIZE_MASK) != 0) {
2190 int j;
2191 for (j = 0; j < oprs; j++)
2192 size[j] = asize;
2193 break;
2194 }
2195 }
2196 } else {
2197 oprs = itemp->operands;
2198 }
2199
2200 for (i = 0; i < itemp->operands; i++) {
2201 if (!(itemp->opd[i] & SIZE_MASK) &&
2202 (instruction->oprs[i].type & SIZE_MASK & ~size[i]))
2203 return MERR_OPSIZEMISMATCH;
2204 }
2205
2206 /*
2207 * Check template is okay at the set cpu level
2208 */
2209 if (((itemp->flags & IF_PLEVEL) > cpu))
2210 return MERR_BADCPU;
2211
2212 /*
2213 * Verify the appropriate long mode flag.
2214 */
2215 if ((itemp->flags & (bits == 64 ? IF_NOLONG : IF_LONG)))
2216 return MERR_BADMODE;
2217
2218 /*
2219 * Check if special handling needed for Jumps
2220 */
2221 if ((itemp->code[0] & 0374) == 0370)
2222 return MOK_JUMP;
2223
2224 return MOK_GOOD;
2225 }
2226
2227 static ea *process_ea(operand * input, ea * output, int bits,
2228 int addrbits, int rfield, int32_t rflags)
2229 {
2230 bool forw_ref = !!(input->opflags & OPFLAG_UNKNOWN);
2231
2232 output->rip = false;
2233
2234 /* REX flags for the rfield operand */
2235 output->rex |= rexflags(rfield, rflags, REX_R|REX_P|REX_W|REX_H);
2236
2237 if (is_class(REGISTER, input->type)) { /* register direct */
2238 int i;
2239 int32_t f;
2240
2241 if (input->basereg < EXPR_REG_START /* Verify as Register */
2242 || input->basereg >= REG_ENUM_LIMIT)
2243 return NULL;
2244 f = regflag(input);
2245 i = nasm_regvals[input->basereg];
2246
2247 if (REG_EA & ~f)
2248 return NULL; /* Invalid EA register */
2249
2250 output->rex |= op_rexflags(input, REX_B|REX_P|REX_W|REX_H);
2251
2252 output->sib_present = false; /* no SIB necessary */
2253 output->bytes = 0; /* no offset necessary either */
2254 output->modrm = 0xC0 | ((rfield & 7) << 3) | (i & 7);
2255 } else { /* it's a memory reference */
2256 if (input->basereg == -1
2257 && (input->indexreg == -1 || input->scale == 0)) {
2258 /* it's a pure offset */
2259 if (bits == 64 && (~input->type & IP_REL)) {
2260 int scale, index, base;
2261 output->sib_present = true;
2262 scale = 0;
2263 index = 4;
2264 base = 5;
2265 output->sib = (scale << 6) | (index << 3) | base;
2266 output->bytes = 4;
2267 output->modrm = 4 | ((rfield & 7) << 3);
2268 output->rip = false;
2269 } else {
2270 output->sib_present = false;
2271 output->bytes = (addrbits != 16 ? 4 : 2);
2272 output->modrm = (addrbits != 16 ? 5 : 6) | ((rfield & 7) << 3);
2273 output->rip = bits == 64;
2274 }
2275 } else { /* it's an indirection */
2276 int i = input->indexreg, b = input->basereg, s = input->scale;
2277 int32_t o = input->offset, seg = input->segment;
2278 int hb = input->hintbase, ht = input->hinttype;
2279 int t;
2280 int it, bt;
2281 int32_t ix, bx; /* register flags */
2282
2283 if (s == 0)
2284 i = -1; /* make this easy, at least */
2285
2286 if (i >= EXPR_REG_START && i < REG_ENUM_LIMIT) {
2287 it = nasm_regvals[i];
2288 ix = nasm_reg_flags[i];
2289 } else {
2290 it = -1;
2291 ix = 0;
2292 }
2293
2294 if (b >= EXPR_REG_START && b < REG_ENUM_LIMIT) {
2295 bt = nasm_regvals[b];
2296 bx = nasm_reg_flags[b];
2297 } else {
2298 bt = -1;
2299 bx = 0;
2300 }
2301
2302 /* check for a 32/64-bit memory reference... */
2303 if ((ix|bx) & (BITS32|BITS64)) {
2304 /* it must be a 32/64-bit memory reference. Firstly we have
2305 * to check that all registers involved are type E/Rxx. */
2306 int32_t sok = BITS32|BITS64;
2307
2308 if (it != -1) {
2309 if (!(REG64 & ~ix) || !(REG32 & ~ix))
2310 sok &= ix;
2311 else
2312 return NULL;
2313 }
2314
2315 if (bt != -1) {
2316 if (REG_GPR & ~bx)
2317 return NULL; /* Invalid register */
2318 if (~sok & bx & SIZE_MASK)
2319 return NULL; /* Invalid size */
2320 sok &= bx;
2321 }
2322
2323 /* While we're here, ensure the user didn't specify
2324 WORD or QWORD. */
2325 if (input->disp_size == 16 || input->disp_size == 64)
2326 return NULL;
2327
2328 if (addrbits == 16 ||
2329 (addrbits == 32 && !(sok & BITS32)) ||
2330 (addrbits == 64 && !(sok & BITS64)))
2331 return NULL;
2332
2333 /* now reorganize base/index */
2334 if (s == 1 && bt != it && bt != -1 && it != -1 &&
2335 ((hb == b && ht == EAH_NOTBASE)
2336 || (hb == i && ht == EAH_MAKEBASE))) {
2337 /* swap if hints say so */
2338 t = bt, bt = it, it = t;
2339 t = bx, bx = ix, ix = t;
2340 }
2341 if (bt == it) /* convert EAX+2*EAX to 3*EAX */
2342 bt = -1, bx = 0, s++;
2343 if (bt == -1 && s == 1 && !(hb == it && ht == EAH_NOTBASE)) {
2344 /* make single reg base, unless hint */
2345 bt = it, bx = ix, it = -1, ix = 0;
2346 }
2347 if (((s == 2 && it != REG_NUM_ESP
2348 && !(input->eaflags & EAF_TIMESTWO)) || s == 3
2349 || s == 5 || s == 9) && bt == -1)
2350 bt = it, bx = ix, s--; /* convert 3*EAX to EAX+2*EAX */
2351 if (it == -1 && (bt & 7) != REG_NUM_ESP
2352 && (input->eaflags & EAF_TIMESTWO))
2353 it = bt, ix = bx, bt = -1, bx = 0, s = 1;
2354 /* convert [NOSPLIT EAX] to sib format with 0x0 displacement */
2355 if (s == 1 && it == REG_NUM_ESP) {
2356 /* swap ESP into base if scale is 1 */
2357 t = it, it = bt, bt = t;
2358 t = ix, ix = bx, bx = t;
2359 }
2360 if (it == REG_NUM_ESP
2361 || (s != 1 && s != 2 && s != 4 && s != 8 && it != -1))
2362 return NULL; /* wrong, for various reasons */
2363
2364 output->rex |= rexflags(it, ix, REX_X);
2365 output->rex |= rexflags(bt, bx, REX_B);
2366
2367 if (it == -1 && (bt & 7) != REG_NUM_ESP) {
2368 /* no SIB needed */
2369 int mod, rm;
2370
2371 if (bt == -1) {
2372 rm = 5;
2373 mod = 0;
2374 } else {
2375 rm = (bt & 7);
2376 if (rm != REG_NUM_EBP && o == 0 &&
2377 seg == NO_SEG && !forw_ref &&
2378 !(input->eaflags &
2379 (EAF_BYTEOFFS | EAF_WORDOFFS)))
2380 mod = 0;
2381 else if (input->eaflags & EAF_BYTEOFFS ||
2382 (o >= -128 && o <= 127 && seg == NO_SEG
2383 && !forw_ref
2384 && !(input->eaflags & EAF_WORDOFFS)))
2385 mod = 1;
2386 else
2387 mod = 2;
2388 }
2389
2390 output->sib_present = false;
2391 output->bytes = (bt == -1 || mod == 2 ? 4 : mod);
2392 output->modrm = (mod << 6) | ((rfield & 7) << 3) | rm;
2393 } else {
2394 /* we need a SIB */
2395 int mod, scale, index, base;
2396
2397 if (it == -1)
2398 index = 4, s = 1;
2399 else
2400 index = (it & 7);
2401
2402 switch (s) {
2403 case 1:
2404 scale = 0;
2405 break;
2406 case 2:
2407 scale = 1;
2408 break;
2409 case 4:
2410 scale = 2;
2411 break;
2412 case 8:
2413 scale = 3;
2414 break;
2415 default: /* then what the smeg is it? */
2416 return NULL; /* panic */
2417 }
2418
2419 if (bt == -1) {
2420 base = 5;
2421 mod = 0;
2422 } else {
2423 base = (bt & 7);
2424 if (base != REG_NUM_EBP && o == 0 &&
2425 seg == NO_SEG && !forw_ref &&
2426 !(input->eaflags &
2427 (EAF_BYTEOFFS | EAF_WORDOFFS)))
2428 mod = 0;
2429 else if (input->eaflags & EAF_BYTEOFFS ||
2430 (o >= -128 && o <= 127 && seg == NO_SEG
2431 && !forw_ref
2432 && !(input->eaflags & EAF_WORDOFFS)))
2433 mod = 1;
2434 else
2435 mod = 2;
2436 }
2437
2438 output->sib_present = true;
2439 output->bytes = (bt == -1 || mod == 2 ? 4 : mod);
2440 output->modrm = (mod << 6) | ((rfield & 7) << 3) | 4;
2441 output->sib = (scale << 6) | (index << 3) | base;
2442 }
2443 } else { /* it's 16-bit */
2444 int mod, rm;
2445
2446 /* check for 64-bit long mode */
2447 if (addrbits == 64)
2448 return NULL;
2449
2450 /* check all registers are BX, BP, SI or DI */
2451 if ((b != -1 && b != R_BP && b != R_BX && b != R_SI
2452 && b != R_DI) || (i != -1 && i != R_BP && i != R_BX
2453 && i != R_SI && i != R_DI))
2454 return NULL;
2455
2456 /* ensure the user didn't specify DWORD/QWORD */
2457 if (input->disp_size == 32 || input->disp_size == 64)
2458 return NULL;
2459
2460 if (s != 1 && i != -1)
2461 return NULL; /* no can do, in 16-bit EA */
2462 if (b == -1 && i != -1) {
2463 int tmp = b;
2464 b = i;
2465 i = tmp;
2466 } /* swap */
2467 if ((b == R_SI || b == R_DI) && i != -1) {
2468 int tmp = b;
2469 b = i;
2470 i = tmp;
2471 }
2472 /* have BX/BP as base, SI/DI index */
2473 if (b == i)
2474 return NULL; /* shouldn't ever happen, in theory */
2475 if (i != -1 && b != -1 &&
2476 (i == R_BP || i == R_BX || b == R_SI || b == R_DI))
2477 return NULL; /* invalid combinations */
2478 if (b == -1) /* pure offset: handled above */
2479 return NULL; /* so if it gets to here, panic! */
2480
2481 rm = -1;
2482 if (i != -1)
2483 switch (i * 256 + b) {
2484 case R_SI * 256 + R_BX:
2485 rm = 0;
2486 break;
2487 case R_DI * 256 + R_BX:
2488 rm = 1;
2489 break;
2490 case R_SI * 256 + R_BP:
2491 rm = 2;
2492 break;
2493 case R_DI * 256 + R_BP:
2494 rm = 3;
2495 break;
2496 } else
2497 switch (b) {
2498 case R_SI:
2499 rm = 4;
2500 break;
2501 case R_DI:
2502 rm = 5;
2503 break;
2504 case R_BP:
2505 rm = 6;
2506 break;
2507 case R_BX:
2508 rm = 7;
2509 break;
2510 }
2511 if (rm == -1) /* can't happen, in theory */
2512 return NULL; /* so panic if it does */
2513
2514 if (o == 0 && seg == NO_SEG && !forw_ref && rm != 6 &&
2515 !(input->eaflags & (EAF_BYTEOFFS | EAF_WORDOFFS)))
2516 mod = 0;
2517 else if (input->eaflags & EAF_BYTEOFFS ||
2518 (o >= -128 && o <= 127 && seg == NO_SEG
2519 && !forw_ref
2520 && !(input->eaflags & EAF_WORDOFFS)))
2521 mod = 1;
2522 else
2523 mod = 2;
2524
2525 output->sib_present = false; /* no SIB - it's 16-bit */
2526 output->bytes = mod; /* bytes of offset needed */
2527 output->modrm = (mod << 6) | ((rfield & 7) << 3) | rm;
2528 }
2529 }
2530 }
2531
2532 output->size = 1 + output->sib_present + output->bytes;
2533 return output;
2534 }
2535
2536 static void add_asp(insn *ins, int addrbits)
2537 {
2538 int j, valid;
2539 int defdisp;
2540
2541 valid = (addrbits == 64) ? 64|32 : 32|16;
2542
2543 switch (ins->prefixes[PPS_ASIZE]) {
2544 case P_A16:
2545 valid &= 16;
2546 break;
2547 case P_A32:
2548 valid &= 32;
2549 break;
2550 case P_A64:
2551 valid &= 64;
2552 break;
2553 case P_ASP:
2554 valid &= (addrbits == 32) ? 16 : 32;
2555 break;
2556 default:
2557 break;
2558 }
2559
2560 for (j = 0; j < ins->operands; j++) {
2561 if (!(MEMORY & ~ins->oprs[j].type)) {
2562 int32_t i, b;
2563
2564 /* Verify as Register */
2565 if (ins->oprs[j].indexreg < EXPR_REG_START
2566 || ins->oprs[j].indexreg >= REG_ENUM_LIMIT)
2567 i = 0;
2568 else
2569 i = nasm_reg_flags[ins->oprs[j].indexreg];
2570
2571 /* Verify as Register */
2572 if (ins->oprs[j].basereg < EXPR_REG_START
2573 || ins->oprs[j].basereg >= REG_ENUM_LIMIT)
2574 b = 0;
2575 else
2576 b = nasm_reg_flags[ins->oprs[j].basereg];
2577
2578 if (ins->oprs[j].scale == 0)
2579 i = 0;
2580
2581 if (!i && !b) {
2582 int ds = ins->oprs[j].disp_size;
2583 if ((addrbits != 64 && ds > 8) ||
2584 (addrbits == 64 && ds == 16))
2585 valid &= ds;
2586 } else {
2587 if (!(REG16 & ~b))
2588 valid &= 16;
2589 if (!(REG32 & ~b))
2590 valid &= 32;
2591 if (!(REG64 & ~b))
2592 valid &= 64;
2593
2594 if (!(REG16 & ~i))
2595 valid &= 16;
2596 if (!(REG32 & ~i))
2597 valid &= 32;
2598 if (!(REG64 & ~i))
2599 valid &= 64;
2600 }
2601 }
2602 }
2603
2604 if (valid & addrbits) {
2605 ins->addr_size = addrbits;
2606 } else if (valid & ((addrbits == 32) ? 16 : 32)) {
2607 /* Add an address size prefix */
2608 enum prefixes pref = (addrbits == 32) ? P_A16 : P_A32;
2609 ins->prefixes[PPS_ASIZE] = pref;
2610 ins->addr_size = (addrbits == 32) ? 16 : 32;
2611 } else {
2612 /* Impossible... */
2613 errfunc(ERR_NONFATAL, "impossible combination of address sizes");
2614 ins->addr_size = addrbits; /* Error recovery */
2615 }
2616
2617 defdisp = ins->addr_size == 16 ? 16 : 32;
2618
2619 for (j = 0; j < ins->operands; j++) {
2620 if (!(MEM_OFFS & ~ins->oprs[j].type) &&
2621 (ins->oprs[j].disp_size ? ins->oprs[j].disp_size : defdisp)
2622 != ins->addr_size) {
2623 /* mem_offs sizes must match the address size; if not,
2624 strip the MEM_OFFS bit and match only EA instructions */
2625 ins->oprs[j].type &= ~(MEM_OFFS & ~MEMORY);
2626 }
2627 }
2628 }
Adding AVX-512 features